//===- DAGCombiner.cpp - Implement a DAG node combiner --------------------===// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===----------------------------------------------------------------------===// // // This pass combines dag nodes to form fewer, simpler DAG nodes. It can be run // both before and after the DAG is legalized. // // This pass is not a substitute for the LLVM IR instcombine pass. This pass is // primarily intended to handle simplification opportunities that are implicit // in the LLVM IR and exposed by the various codegen lowering phases. // //===----------------------------------------------------------------------===// #include "llvm/ADT/APFloat.h" #include "llvm/ADT/APInt.h" #include "llvm/ADT/ArrayRef.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/IntervalMap.h" #include "llvm/ADT/None.h" #include "llvm/ADT/Optional.h" #include "llvm/ADT/STLExtras.h" #include "llvm/ADT/SetVector.h" #include "llvm/ADT/SmallBitVector.h" #include "llvm/ADT/SmallPtrSet.h" #include "llvm/ADT/SmallSet.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/Statistic.h" #include "llvm/Analysis/AliasAnalysis.h" #include "llvm/Analysis/MemoryLocation.h" #include "llvm/Analysis/TargetLibraryInfo.h" #include "llvm/Analysis/VectorUtils.h" #include "llvm/CodeGen/DAGCombine.h" #include "llvm/CodeGen/ISDOpcodes.h" #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/MachineMemOperand.h" #include "llvm/CodeGen/RuntimeLibcalls.h" #include "llvm/CodeGen/SelectionDAG.h" #include "llvm/CodeGen/SelectionDAGAddressAnalysis.h" #include "llvm/CodeGen/SelectionDAGNodes.h" #include "llvm/CodeGen/SelectionDAGTargetInfo.h" #include "llvm/CodeGen/TargetLowering.h" #include "llvm/CodeGen/TargetRegisterInfo.h" #include "llvm/CodeGen/TargetSubtargetInfo.h" #include "llvm/CodeGen/ValueTypes.h" #include "llvm/IR/Attributes.h" #include "llvm/IR/Constant.h" #include "llvm/IR/DataLayout.h" #include "llvm/IR/DerivedTypes.h" #include "llvm/IR/Function.h" #include "llvm/IR/Metadata.h" #include "llvm/Support/Casting.h" #include "llvm/Support/CodeGen.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Compiler.h" #include "llvm/Support/Debug.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/KnownBits.h" #include "llvm/Support/MachineValueType.h" #include "llvm/Support/MathExtras.h" #include "llvm/Support/raw_ostream.h" #include "llvm/Target/TargetMachine.h" #include "llvm/Target/TargetOptions.h" #include #include #include #include #include #include #include #include using namespace llvm; #define DEBUG_TYPE "dagcombine" STATISTIC(NodesCombined , "Number of dag nodes combined"); STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created"); STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created"); STATISTIC(OpsNarrowed , "Number of load/op/store narrowed"); STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int"); STATISTIC(SlicedLoads, "Number of load sliced"); STATISTIC(NumFPLogicOpsConv, "Number of logic ops converted to fp ops"); static cl::opt CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden, cl::desc("Enable DAG combiner's use of IR alias analysis")); static cl::opt UseTBAA("combiner-use-tbaa", cl::Hidden, cl::init(true), cl::desc("Enable DAG combiner's use of TBAA")); #ifndef NDEBUG static cl::opt CombinerAAOnlyFunc("combiner-aa-only-func", cl::Hidden, cl::desc("Only use DAG-combiner alias analysis in this" " function")); #endif /// Hidden option to stress test load slicing, i.e., when this option /// is enabled, load slicing bypasses most of its profitability guards. static cl::opt StressLoadSlicing("combiner-stress-load-slicing", cl::Hidden, cl::desc("Bypass the profitability model of load slicing"), cl::init(false)); static cl::opt MaySplitLoadIndex("combiner-split-load-index", cl::Hidden, cl::init(true), cl::desc("DAG combiner may split indexing from loads")); static cl::opt EnableStoreMerging("combiner-store-merging", cl::Hidden, cl::init(true), cl::desc("DAG combiner enable merging multiple stores " "into a wider store")); static cl::opt TokenFactorInlineLimit( "combiner-tokenfactor-inline-limit", cl::Hidden, cl::init(2048), cl::desc("Limit the number of operands to inline for Token Factors")); static cl::opt StoreMergeDependenceLimit( "combiner-store-merge-dependence-limit", cl::Hidden, cl::init(10), cl::desc("Limit the number of times for the same StoreNode and RootNode " "to bail out in store merging dependence check")); static cl::opt EnableReduceLoadOpStoreWidth( "combiner-reduce-load-op-store-width", cl::Hidden, cl::init(true), cl::desc("DAG combiner enable reducing the width of load/op/store " "sequence")); static cl::opt EnableShrinkLoadReplaceStoreWithStore( "combiner-shrink-load-replace-store-with-store", cl::Hidden, cl::init(true), cl::desc("DAG combiner enable load//store with " "a narrower store")); namespace { class DAGCombiner { SelectionDAG &DAG; const TargetLowering &TLI; const SelectionDAGTargetInfo *STI; CombineLevel Level = BeforeLegalizeTypes; CodeGenOpt::Level OptLevel; bool LegalDAG = false; bool LegalOperations = false; bool LegalTypes = false; bool ForCodeSize; bool DisableGenericCombines; /// Worklist of all of the nodes that need to be simplified. /// /// This must behave as a stack -- new nodes to process are pushed onto the /// back and when processing we pop off of the back. /// /// The worklist will not contain duplicates but may contain null entries /// due to nodes being deleted from the underlying DAG. SmallVector Worklist; /// Mapping from an SDNode to its position on the worklist. /// /// This is used to find and remove nodes from the worklist (by nulling /// them) when they are deleted from the underlying DAG. It relies on /// stable indices of nodes within the worklist. DenseMap WorklistMap; /// This records all nodes attempted to add to the worklist since we /// considered a new worklist entry. As we keep do not add duplicate nodes /// in the worklist, this is different from the tail of the worklist. SmallSetVector PruningList; /// Set of nodes which have been combined (at least once). /// /// This is used to allow us to reliably add any operands of a DAG node /// which have not yet been combined to the worklist. SmallPtrSet CombinedNodes; /// Map from candidate StoreNode to the pair of RootNode and count. /// The count is used to track how many times we have seen the StoreNode /// with the same RootNode bail out in dependence check. If we have seen /// the bail out for the same pair many times over a limit, we won't /// consider the StoreNode with the same RootNode as store merging /// candidate again. DenseMap> StoreRootCountMap; // AA - Used for DAG load/store alias analysis. AliasAnalysis *AA; /// When an instruction is simplified, add all users of the instruction to /// the work lists because they might get more simplified now. void AddUsersToWorklist(SDNode *N) { for (SDNode *Node : N->uses()) AddToWorklist(Node); } /// Convenient shorthand to add a node and all of its user to the worklist. void AddToWorklistWithUsers(SDNode *N) { AddUsersToWorklist(N); AddToWorklist(N); } // Prune potentially dangling nodes. This is called after // any visit to a node, but should also be called during a visit after any // failed combine which may have created a DAG node. void clearAddedDanglingWorklistEntries() { // Check any nodes added to the worklist to see if they are prunable. while (!PruningList.empty()) { auto *N = PruningList.pop_back_val(); if (N->use_empty()) recursivelyDeleteUnusedNodes(N); } } SDNode *getNextWorklistEntry() { // Before we do any work, remove nodes that are not in use. clearAddedDanglingWorklistEntries(); SDNode *N = nullptr; // The Worklist holds the SDNodes in order, but it may contain null // entries. while (!N && !Worklist.empty()) { N = Worklist.pop_back_val(); } if (N) { bool GoodWorklistEntry = WorklistMap.erase(N); (void)GoodWorklistEntry; assert(GoodWorklistEntry && "Found a worklist entry without a corresponding map entry!"); } return N; } /// Call the node-specific routine that folds each particular type of node. SDValue visit(SDNode *N); public: DAGCombiner(SelectionDAG &D, AliasAnalysis *AA, CodeGenOpt::Level OL) : DAG(D), TLI(D.getTargetLoweringInfo()), STI(D.getSubtarget().getSelectionDAGInfo()), OptLevel(OL), AA(AA) { ForCodeSize = DAG.shouldOptForSize(); DisableGenericCombines = STI && STI->disableGenericCombines(OptLevel); MaximumLegalStoreInBits = 0; // We use the minimum store size here, since that's all we can guarantee // for the scalable vector types. for (MVT VT : MVT::all_valuetypes()) if (EVT(VT).isSimple() && VT != MVT::Other && TLI.isTypeLegal(EVT(VT)) && VT.getSizeInBits().getKnownMinSize() >= MaximumLegalStoreInBits) MaximumLegalStoreInBits = VT.getSizeInBits().getKnownMinSize(); } void ConsiderForPruning(SDNode *N) { // Mark this for potential pruning. PruningList.insert(N); } /// Add to the worklist making sure its instance is at the back (next to be /// processed.) void AddToWorklist(SDNode *N) { assert(N->getOpcode() != ISD::DELETED_NODE && "Deleted Node added to Worklist"); // Skip handle nodes as they can't usefully be combined and confuse the // zero-use deletion strategy. if (N->getOpcode() == ISD::HANDLENODE) return; ConsiderForPruning(N); if (WorklistMap.insert(std::make_pair(N, Worklist.size())).second) Worklist.push_back(N); } /// Remove all instances of N from the worklist. void removeFromWorklist(SDNode *N) { CombinedNodes.erase(N); PruningList.remove(N); StoreRootCountMap.erase(N); auto It = WorklistMap.find(N); if (It == WorklistMap.end()) return; // Not in the worklist. // Null out the entry rather than erasing it to avoid a linear operation. Worklist[It->second] = nullptr; WorklistMap.erase(It); } void deleteAndRecombine(SDNode *N); bool recursivelyDeleteUnusedNodes(SDNode *N); /// Replaces all uses of the results of one DAG node with new values. SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo, bool AddTo = true); /// Replaces all uses of the results of one DAG node with new values. SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) { return CombineTo(N, &Res, 1, AddTo); } /// Replaces all uses of the results of one DAG node with new values. SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo = true) { SDValue To[] = { Res0, Res1 }; return CombineTo(N, To, 2, AddTo); } void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO); private: unsigned MaximumLegalStoreInBits; /// Check the specified integer node value to see if it can be simplified or /// if things it uses can be simplified by bit propagation. /// If so, return true. bool SimplifyDemandedBits(SDValue Op) { unsigned BitWidth = Op.getScalarValueSizeInBits(); APInt DemandedBits = APInt::getAllOnes(BitWidth); return SimplifyDemandedBits(Op, DemandedBits); } bool SimplifyDemandedBits(SDValue Op, const APInt &DemandedBits) { TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations); KnownBits Known; if (!TLI.SimplifyDemandedBits(Op, DemandedBits, Known, TLO, 0, false)) return false; // Revisit the node. AddToWorklist(Op.getNode()); CommitTargetLoweringOpt(TLO); return true; } /// Check the specified vector node value to see if it can be simplified or /// if things it uses can be simplified as it only uses some of the /// elements. If so, return true. bool SimplifyDemandedVectorElts(SDValue Op) { // TODO: For now just pretend it cannot be simplified. if (Op.getValueType().isScalableVector()) return false; unsigned NumElts = Op.getValueType().getVectorNumElements(); APInt DemandedElts = APInt::getAllOnes(NumElts); return SimplifyDemandedVectorElts(Op, DemandedElts); } bool SimplifyDemandedBits(SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts, bool AssumeSingleUse = false); bool SimplifyDemandedVectorElts(SDValue Op, const APInt &DemandedElts, bool AssumeSingleUse = false); bool CombineToPreIndexedLoadStore(SDNode *N); bool CombineToPostIndexedLoadStore(SDNode *N); SDValue SplitIndexingFromLoad(LoadSDNode *LD); bool SliceUpLoad(SDNode *N); // Scalars have size 0 to distinguish from singleton vectors. SDValue ForwardStoreValueToDirectLoad(LoadSDNode *LD); bool getTruncatedStoreValue(StoreSDNode *ST, SDValue &Val); bool extendLoadedValueToExtension(LoadSDNode *LD, SDValue &Val); /// Replace an ISD::EXTRACT_VECTOR_ELT of a load with a narrowed /// load. /// /// \param EVE ISD::EXTRACT_VECTOR_ELT to be replaced. /// \param InVecVT type of the input vector to EVE with bitcasts resolved. /// \param EltNo index of the vector element to load. /// \param OriginalLoad load that EVE came from to be replaced. /// \returns EVE on success SDValue() on failure. SDValue scalarizeExtractedVectorLoad(SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad); void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad); SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace); SDValue SExtPromoteOperand(SDValue Op, EVT PVT); SDValue ZExtPromoteOperand(SDValue Op, EVT PVT); SDValue PromoteIntBinOp(SDValue Op); SDValue PromoteIntShiftOp(SDValue Op); SDValue PromoteExtend(SDValue Op); bool PromoteLoad(SDValue Op); /// Call the node-specific routine that knows how to fold each /// particular type of node. If that doesn't do anything, try the /// target-specific DAG combines. SDValue combine(SDNode *N); // Visitation implementation - Implement dag node combining for different // node types. The semantics are as follows: // Return Value: // SDValue.getNode() == 0 - No change was made // SDValue.getNode() == N - N was replaced, is dead and has been handled. // otherwise - N should be replaced by the returned Operand. // SDValue visitTokenFactor(SDNode *N); SDValue visitMERGE_VALUES(SDNode *N); SDValue visitADD(SDNode *N); SDValue visitADDLike(SDNode *N); SDValue visitADDLikeCommutative(SDValue N0, SDValue N1, SDNode *LocReference); SDValue visitSUB(SDNode *N); SDValue visitADDSAT(SDNode *N); SDValue visitSUBSAT(SDNode *N); SDValue visitADDC(SDNode *N); SDValue visitADDO(SDNode *N); SDValue visitUADDOLike(SDValue N0, SDValue N1, SDNode *N); SDValue visitSUBC(SDNode *N); SDValue visitSUBO(SDNode *N); SDValue visitADDE(SDNode *N); SDValue visitADDCARRY(SDNode *N); SDValue visitSADDO_CARRY(SDNode *N); SDValue visitADDCARRYLike(SDValue N0, SDValue N1, SDValue CarryIn, SDNode *N); SDValue visitSUBE(SDNode *N); SDValue visitSUBCARRY(SDNode *N); SDValue visitSSUBO_CARRY(SDNode *N); SDValue visitMUL(SDNode *N); SDValue visitMULFIX(SDNode *N); SDValue useDivRem(SDNode *N); SDValue visitSDIV(SDNode *N); SDValue visitSDIVLike(SDValue N0, SDValue N1, SDNode *N); SDValue visitUDIV(SDNode *N); SDValue visitUDIVLike(SDValue N0, SDValue N1, SDNode *N); SDValue visitREM(SDNode *N); SDValue visitMULHU(SDNode *N); SDValue visitMULHS(SDNode *N); SDValue visitAVG(SDNode *N); SDValue visitSMUL_LOHI(SDNode *N); SDValue visitUMUL_LOHI(SDNode *N); SDValue visitMULO(SDNode *N); SDValue visitIMINMAX(SDNode *N); SDValue visitAND(SDNode *N); SDValue visitANDLike(SDValue N0, SDValue N1, SDNode *N); SDValue visitOR(SDNode *N); SDValue visitORLike(SDValue N0, SDValue N1, SDNode *N); SDValue visitXOR(SDNode *N); SDValue SimplifyVBinOp(SDNode *N, const SDLoc &DL); SDValue visitSHL(SDNode *N); SDValue visitSRA(SDNode *N); SDValue visitSRL(SDNode *N); SDValue visitFunnelShift(SDNode *N); SDValue visitSHLSAT(SDNode *N); SDValue visitRotate(SDNode *N); SDValue visitABS(SDNode *N); SDValue visitBSWAP(SDNode *N); SDValue visitBITREVERSE(SDNode *N); SDValue visitCTLZ(SDNode *N); SDValue visitCTLZ_ZERO_UNDEF(SDNode *N); SDValue visitCTTZ(SDNode *N); SDValue visitCTTZ_ZERO_UNDEF(SDNode *N); SDValue visitCTPOP(SDNode *N); SDValue visitSELECT(SDNode *N); SDValue visitVSELECT(SDNode *N); SDValue visitSELECT_CC(SDNode *N); SDValue visitSETCC(SDNode *N); SDValue visitSETCCCARRY(SDNode *N); SDValue visitSIGN_EXTEND(SDNode *N); SDValue visitZERO_EXTEND(SDNode *N); SDValue visitANY_EXTEND(SDNode *N); SDValue visitAssertExt(SDNode *N); SDValue visitAssertAlign(SDNode *N); SDValue visitSIGN_EXTEND_INREG(SDNode *N); SDValue visitEXTEND_VECTOR_INREG(SDNode *N); SDValue visitTRUNCATE(SDNode *N); SDValue visitBITCAST(SDNode *N); SDValue visitFREEZE(SDNode *N); SDValue visitBUILD_PAIR(SDNode *N); SDValue visitFADD(SDNode *N); SDValue visitSTRICT_FADD(SDNode *N); SDValue visitFSUB(SDNode *N); SDValue visitFMUL(SDNode *N); SDValue visitFMA(SDNode *N); SDValue visitFDIV(SDNode *N); SDValue visitFREM(SDNode *N); SDValue visitFSQRT(SDNode *N); SDValue visitFCOPYSIGN(SDNode *N); SDValue visitFPOW(SDNode *N); SDValue visitSINT_TO_FP(SDNode *N); SDValue visitUINT_TO_FP(SDNode *N); SDValue visitFP_TO_SINT(SDNode *N); SDValue visitFP_TO_UINT(SDNode *N); SDValue visitFP_ROUND(SDNode *N); SDValue visitFP_EXTEND(SDNode *N); SDValue visitFNEG(SDNode *N); SDValue visitFABS(SDNode *N); SDValue visitFCEIL(SDNode *N); SDValue visitFTRUNC(SDNode *N); SDValue visitFFLOOR(SDNode *N); SDValue visitFMinMax(SDNode *N); SDValue visitBRCOND(SDNode *N); SDValue visitBR_CC(SDNode *N); SDValue visitLOAD(SDNode *N); SDValue replaceStoreChain(StoreSDNode *ST, SDValue BetterChain); SDValue replaceStoreOfFPConstant(StoreSDNode *ST); SDValue visitSTORE(SDNode *N); SDValue visitLIFETIME_END(SDNode *N); SDValue visitINSERT_VECTOR_ELT(SDNode *N); SDValue visitEXTRACT_VECTOR_ELT(SDNode *N); SDValue visitBUILD_VECTOR(SDNode *N); SDValue visitCONCAT_VECTORS(SDNode *N); SDValue visitEXTRACT_SUBVECTOR(SDNode *N); SDValue visitVECTOR_SHUFFLE(SDNode *N); SDValue visitSCALAR_TO_VECTOR(SDNode *N); SDValue visitINSERT_SUBVECTOR(SDNode *N); SDValue visitMLOAD(SDNode *N); SDValue visitMSTORE(SDNode *N); SDValue visitMGATHER(SDNode *N); SDValue visitMSCATTER(SDNode *N); SDValue visitFP_TO_FP16(SDNode *N); SDValue visitFP16_TO_FP(SDNode *N); SDValue visitFP_TO_BF16(SDNode *N); SDValue visitVECREDUCE(SDNode *N); SDValue visitVPOp(SDNode *N); SDValue visitFADDForFMACombine(SDNode *N); SDValue visitFSUBForFMACombine(SDNode *N); SDValue visitFMULForFMADistributiveCombine(SDNode *N); SDValue XformToShuffleWithZero(SDNode *N); bool reassociationCanBreakAddressingModePattern(unsigned Opc, const SDLoc &DL, SDNode *N, SDValue N0, SDValue N1); SDValue reassociateOpsCommutative(unsigned Opc, const SDLoc &DL, SDValue N0, SDValue N1); SDValue reassociateOps(unsigned Opc, const SDLoc &DL, SDValue N0, SDValue N1, SDNodeFlags Flags); SDValue visitShiftByConstant(SDNode *N); SDValue foldSelectOfConstants(SDNode *N); SDValue foldVSelectOfConstants(SDNode *N); SDValue foldBinOpIntoSelect(SDNode *BO); bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS); SDValue hoistLogicOpWithSameOpcodeHands(SDNode *N); SDValue SimplifySelect(const SDLoc &DL, SDValue N0, SDValue N1, SDValue N2); SDValue SimplifySelectCC(const SDLoc &DL, SDValue N0, SDValue N1, SDValue N2, SDValue N3, ISD::CondCode CC, bool NotExtCompare = false); SDValue convertSelectOfFPConstantsToLoadOffset( const SDLoc &DL, SDValue N0, SDValue N1, SDValue N2, SDValue N3, ISD::CondCode CC); SDValue foldSignChangeInBitcast(SDNode *N); SDValue foldSelectCCToShiftAnd(const SDLoc &DL, SDValue N0, SDValue N1, SDValue N2, SDValue N3, ISD::CondCode CC); SDValue foldSelectOfBinops(SDNode *N); SDValue foldSextSetcc(SDNode *N); SDValue foldLogicOfSetCCs(bool IsAnd, SDValue N0, SDValue N1, const SDLoc &DL); SDValue foldSubToUSubSat(EVT DstVT, SDNode *N); SDValue unfoldMaskedMerge(SDNode *N); SDValue unfoldExtremeBitClearingToShifts(SDNode *N); SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond, const SDLoc &DL, bool foldBooleans); SDValue rebuildSetCC(SDValue N); bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS, SDValue &CC, bool MatchStrict = false) const; bool isOneUseSetCC(SDValue N) const; SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp, unsigned HiOp); SDValue CombineConsecutiveLoads(SDNode *N, EVT VT); SDValue CombineExtLoad(SDNode *N); SDValue CombineZExtLogicopShiftLoad(SDNode *N); SDValue combineRepeatedFPDivisors(SDNode *N); SDValue combineInsertEltToShuffle(SDNode *N, unsigned InsIndex); SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT); SDValue BuildSDIV(SDNode *N); SDValue BuildSDIVPow2(SDNode *N); SDValue BuildUDIV(SDNode *N); SDValue BuildSREMPow2(SDNode *N); SDValue buildOptimizedSREM(SDValue N0, SDValue N1, SDNode *N); SDValue BuildLogBase2(SDValue V, const SDLoc &DL); SDValue BuildDivEstimate(SDValue N, SDValue Op, SDNodeFlags Flags); SDValue buildRsqrtEstimate(SDValue Op, SDNodeFlags Flags); SDValue buildSqrtEstimate(SDValue Op, SDNodeFlags Flags); SDValue buildSqrtEstimateImpl(SDValue Op, SDNodeFlags Flags, bool Recip); SDValue buildSqrtNROneConst(SDValue Arg, SDValue Est, unsigned Iterations, SDNodeFlags Flags, bool Reciprocal); SDValue buildSqrtNRTwoConst(SDValue Arg, SDValue Est, unsigned Iterations, SDNodeFlags Flags, bool Reciprocal); SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1, bool DemandHighBits = true); SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1); SDValue MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg, SDValue InnerPos, SDValue InnerNeg, bool HasPos, unsigned PosOpcode, unsigned NegOpcode, const SDLoc &DL); SDValue MatchFunnelPosNeg(SDValue N0, SDValue N1, SDValue Pos, SDValue Neg, SDValue InnerPos, SDValue InnerNeg, bool HasPos, unsigned PosOpcode, unsigned NegOpcode, const SDLoc &DL); SDValue MatchRotate(SDValue LHS, SDValue RHS, const SDLoc &DL); SDValue MatchLoadCombine(SDNode *N); SDValue mergeTruncStores(StoreSDNode *N); SDValue reduceLoadWidth(SDNode *N); SDValue ReduceLoadOpStoreWidth(SDNode *N); SDValue splitMergedValStore(StoreSDNode *ST); SDValue TransformFPLoadStorePair(SDNode *N); SDValue convertBuildVecZextToZext(SDNode *N); SDValue reduceBuildVecExtToExtBuildVec(SDNode *N); SDValue reduceBuildVecTruncToBitCast(SDNode *N); SDValue reduceBuildVecToShuffle(SDNode *N); SDValue createBuildVecShuffle(const SDLoc &DL, SDNode *N, ArrayRef VectorMask, SDValue VecIn1, SDValue VecIn2, unsigned LeftIdx, bool DidSplitVec); SDValue matchVSelectOpSizesWithSetCC(SDNode *Cast); /// Walk up chain skipping non-aliasing memory nodes, /// looking for aliasing nodes and adding them to the Aliases vector. void GatherAllAliases(SDNode *N, SDValue OriginalChain, SmallVectorImpl &Aliases); /// Return true if there is any possibility that the two addresses overlap. bool mayAlias(SDNode *Op0, SDNode *Op1) const; /// Walk up chain skipping non-aliasing memory nodes, looking for a better /// chain (aliasing node.) SDValue FindBetterChain(SDNode *N, SDValue Chain); /// Try to replace a store and any possibly adjacent stores on /// consecutive chains with better chains. Return true only if St is /// replaced. /// /// Notice that other chains may still be replaced even if the function /// returns false. bool findBetterNeighborChains(StoreSDNode *St); // Helper for findBetterNeighborChains. Walk up store chain add additional // chained stores that do not overlap and can be parallelized. bool parallelizeChainedStores(StoreSDNode *St); /// Holds a pointer to an LSBaseSDNode as well as information on where it /// is located in a sequence of memory operations connected by a chain. struct MemOpLink { // Ptr to the mem node. LSBaseSDNode *MemNode; // Offset from the base ptr. int64_t OffsetFromBase; MemOpLink(LSBaseSDNode *N, int64_t Offset) : MemNode(N), OffsetFromBase(Offset) {} }; // Classify the origin of a stored value. enum class StoreSource { Unknown, Constant, Extract, Load }; StoreSource getStoreSource(SDValue StoreVal) { switch (StoreVal.getOpcode()) { case ISD::Constant: case ISD::ConstantFP: return StoreSource::Constant; case ISD::EXTRACT_VECTOR_ELT: case ISD::EXTRACT_SUBVECTOR: return StoreSource::Extract; case ISD::LOAD: return StoreSource::Load; default: return StoreSource::Unknown; } } /// This is a helper function for visitMUL to check the profitability /// of folding (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2). /// MulNode is the original multiply, AddNode is (add x, c1), /// and ConstNode is c2. bool isMulAddWithConstProfitable(SDNode *MulNode, SDValue AddNode, SDValue ConstNode); /// This is a helper function for visitAND and visitZERO_EXTEND. Returns /// true if the (and (load x) c) pattern matches an extload. ExtVT returns /// the type of the loaded value to be extended. bool isAndLoadExtLoad(ConstantSDNode *AndC, LoadSDNode *LoadN, EVT LoadResultTy, EVT &ExtVT); /// Helper function to calculate whether the given Load/Store can have its /// width reduced to ExtVT. bool isLegalNarrowLdSt(LSBaseSDNode *LDSTN, ISD::LoadExtType ExtType, EVT &MemVT, unsigned ShAmt = 0); /// Used by BackwardsPropagateMask to find suitable loads. bool SearchForAndLoads(SDNode *N, SmallVectorImpl &Loads, SmallPtrSetImpl &NodesWithConsts, ConstantSDNode *Mask, SDNode *&NodeToMask); /// Attempt to propagate a given AND node back to load leaves so that they /// can be combined into narrow loads. bool BackwardsPropagateMask(SDNode *N); /// Helper function for mergeConsecutiveStores which merges the component /// store chains. SDValue getMergeStoreChains(SmallVectorImpl &StoreNodes, unsigned NumStores); /// This is a helper function for mergeConsecutiveStores. When the source /// elements of the consecutive stores are all constants or all extracted /// vector elements, try to merge them into one larger store introducing /// bitcasts if necessary. \return True if a merged store was created. bool mergeStoresOfConstantsOrVecElts(SmallVectorImpl &StoreNodes, EVT MemVT, unsigned NumStores, bool IsConstantSrc, bool UseVector, bool UseTrunc); /// This is a helper function for mergeConsecutiveStores. Stores that /// potentially may be merged with St are placed in StoreNodes. RootNode is /// a chain predecessor to all store candidates. void getStoreMergeCandidates(StoreSDNode *St, SmallVectorImpl &StoreNodes, SDNode *&Root); /// Helper function for mergeConsecutiveStores. Checks if candidate stores /// have indirect dependency through their operands. RootNode is the /// predecessor to all stores calculated by getStoreMergeCandidates and is /// used to prune the dependency check. \return True if safe to merge. bool checkMergeStoreCandidatesForDependencies( SmallVectorImpl &StoreNodes, unsigned NumStores, SDNode *RootNode); /// This is a helper function for mergeConsecutiveStores. Given a list of /// store candidates, find the first N that are consecutive in memory. /// Returns 0 if there are not at least 2 consecutive stores to try merging. unsigned getConsecutiveStores(SmallVectorImpl &StoreNodes, int64_t ElementSizeBytes) const; /// This is a helper function for mergeConsecutiveStores. It is used for /// store chains that are composed entirely of constant values. bool tryStoreMergeOfConstants(SmallVectorImpl &StoreNodes, unsigned NumConsecutiveStores, EVT MemVT, SDNode *Root, bool AllowVectors); /// This is a helper function for mergeConsecutiveStores. It is used for /// store chains that are composed entirely of extracted vector elements. /// When extracting multiple vector elements, try to store them in one /// vector store rather than a sequence of scalar stores. bool tryStoreMergeOfExtracts(SmallVectorImpl &StoreNodes, unsigned NumConsecutiveStores, EVT MemVT, SDNode *Root); /// This is a helper function for mergeConsecutiveStores. It is used for /// store chains that are composed entirely of loaded values. bool tryStoreMergeOfLoads(SmallVectorImpl &StoreNodes, unsigned NumConsecutiveStores, EVT MemVT, SDNode *Root, bool AllowVectors, bool IsNonTemporalStore, bool IsNonTemporalLoad); /// Merge consecutive store operations into a wide store. /// This optimization uses wide integers or vectors when possible. /// \return true if stores were merged. bool mergeConsecutiveStores(StoreSDNode *St); /// Try to transform a truncation where C is a constant: /// (trunc (and X, C)) -> (and (trunc X), (trunc C)) /// /// \p N needs to be a truncation and its first operand an AND. Other /// requirements are checked by the function (e.g. that trunc is /// single-use) and if missed an empty SDValue is returned. SDValue distributeTruncateThroughAnd(SDNode *N); /// Helper function to determine whether the target supports operation /// given by \p Opcode for type \p VT, that is, whether the operation /// is legal or custom before legalizing operations, and whether is /// legal (but not custom) after legalization. bool hasOperation(unsigned Opcode, EVT VT) { return TLI.isOperationLegalOrCustom(Opcode, VT, LegalOperations); } public: /// Runs the dag combiner on all nodes in the work list void Run(CombineLevel AtLevel); SelectionDAG &getDAG() const { return DAG; } /// Returns a type large enough to hold any valid shift amount - before type /// legalization these can be huge. EVT getShiftAmountTy(EVT LHSTy) { assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); return TLI.getShiftAmountTy(LHSTy, DAG.getDataLayout(), LegalTypes); } /// This method returns true if we are running before type legalization or /// if the specified VT is legal. bool isTypeLegal(const EVT &VT) { if (!LegalTypes) return true; return TLI.isTypeLegal(VT); } /// Convenience wrapper around TargetLowering::getSetCCResultType EVT getSetCCResultType(EVT VT) const { return TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); } void ExtendSetCCUses(const SmallVectorImpl &SetCCs, SDValue OrigLoad, SDValue ExtLoad, ISD::NodeType ExtType); }; /// This class is a DAGUpdateListener that removes any deleted /// nodes from the worklist. class WorklistRemover : public SelectionDAG::DAGUpdateListener { DAGCombiner &DC; public: explicit WorklistRemover(DAGCombiner &dc) : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {} void NodeDeleted(SDNode *N, SDNode *E) override { DC.removeFromWorklist(N); } }; class WorklistInserter : public SelectionDAG::DAGUpdateListener { DAGCombiner &DC; public: explicit WorklistInserter(DAGCombiner &dc) : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {} // FIXME: Ideally we could add N to the worklist, but this causes exponential // compile time costs in large DAGs, e.g. Halide. void NodeInserted(SDNode *N) override { DC.ConsiderForPruning(N); } }; } // end anonymous namespace //===----------------------------------------------------------------------===// // TargetLowering::DAGCombinerInfo implementation //===----------------------------------------------------------------------===// void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) { ((DAGCombiner*)DC)->AddToWorklist(N); } SDValue TargetLowering::DAGCombinerInfo:: CombineTo(SDNode *N, ArrayRef To, bool AddTo) { return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo); } SDValue TargetLowering::DAGCombinerInfo:: CombineTo(SDNode *N, SDValue Res, bool AddTo) { return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo); } SDValue TargetLowering::DAGCombinerInfo:: CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) { return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo); } bool TargetLowering::DAGCombinerInfo:: recursivelyDeleteUnusedNodes(SDNode *N) { return ((DAGCombiner*)DC)->recursivelyDeleteUnusedNodes(N); } void TargetLowering::DAGCombinerInfo:: CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) { return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO); } //===----------------------------------------------------------------------===// // Helper Functions //===----------------------------------------------------------------------===// void DAGCombiner::deleteAndRecombine(SDNode *N) { removeFromWorklist(N); // If the operands of this node are only used by the node, they will now be // dead. Make sure to re-visit them and recursively delete dead nodes. for (const SDValue &Op : N->ops()) // For an operand generating multiple values, one of the values may // become dead allowing further simplification (e.g. split index // arithmetic from an indexed load). if (Op->hasOneUse() || Op->getNumValues() > 1) AddToWorklist(Op.getNode()); DAG.DeleteNode(N); } // APInts must be the same size for most operations, this helper // function zero extends the shorter of the pair so that they match. // We provide an Offset so that we can create bitwidths that won't overflow. static void zeroExtendToMatch(APInt &LHS, APInt &RHS, unsigned Offset = 0) { unsigned Bits = Offset + std::max(LHS.getBitWidth(), RHS.getBitWidth()); LHS = LHS.zext(Bits); RHS = RHS.zext(Bits); } // Return true if this node is a setcc, or is a select_cc // that selects between the target values used for true and false, making it // equivalent to a setcc. Also, set the incoming LHS, RHS, and CC references to // the appropriate nodes based on the type of node we are checking. This // simplifies life a bit for the callers. bool DAGCombiner::isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS, SDValue &CC, bool MatchStrict) const { if (N.getOpcode() == ISD::SETCC) { LHS = N.getOperand(0); RHS = N.getOperand(1); CC = N.getOperand(2); return true; } if (MatchStrict && (N.getOpcode() == ISD::STRICT_FSETCC || N.getOpcode() == ISD::STRICT_FSETCCS)) { LHS = N.getOperand(1); RHS = N.getOperand(2); CC = N.getOperand(3); return true; } if (N.getOpcode() != ISD::SELECT_CC || !TLI.isConstTrueVal(N.getOperand(2)) || !TLI.isConstFalseVal(N.getOperand(3))) return false; if (TLI.getBooleanContents(N.getValueType()) == TargetLowering::UndefinedBooleanContent) return false; LHS = N.getOperand(0); RHS = N.getOperand(1); CC = N.getOperand(4); return true; } /// Return true if this is a SetCC-equivalent operation with only one use. /// If this is true, it allows the users to invert the operation for free when /// it is profitable to do so. bool DAGCombiner::isOneUseSetCC(SDValue N) const { SDValue N0, N1, N2; if (isSetCCEquivalent(N, N0, N1, N2) && N->hasOneUse()) return true; return false; } static bool isConstantSplatVectorMaskForType(SDNode *N, EVT ScalarTy) { if (!ScalarTy.isSimple()) return false; uint64_t MaskForTy = 0ULL; switch (ScalarTy.getSimpleVT().SimpleTy) { case MVT::i8: MaskForTy = 0xFFULL; break; case MVT::i16: MaskForTy = 0xFFFFULL; break; case MVT::i32: MaskForTy = 0xFFFFFFFFULL; break; default: return false; break; } APInt Val; if (ISD::isConstantSplatVector(N, Val)) return Val.getLimitedValue() == MaskForTy; return false; } // Determines if it is a constant integer or a splat/build vector of constant // integers (and undefs). // Do not permit build vector implicit truncation. static bool isConstantOrConstantVector(SDValue N, bool NoOpaques = false) { if (ConstantSDNode *Const = dyn_cast(N)) return !(Const->isOpaque() && NoOpaques); if (N.getOpcode() != ISD::BUILD_VECTOR && N.getOpcode() != ISD::SPLAT_VECTOR) return false; unsigned BitWidth = N.getScalarValueSizeInBits(); for (const SDValue &Op : N->op_values()) { if (Op.isUndef()) continue; ConstantSDNode *Const = dyn_cast(Op); if (!Const || Const->getAPIntValue().getBitWidth() != BitWidth || (Const->isOpaque() && NoOpaques)) return false; } return true; } // Determines if a BUILD_VECTOR is composed of all-constants possibly mixed with // undef's. static bool isAnyConstantBuildVector(SDValue V, bool NoOpaques = false) { if (V.getOpcode() != ISD::BUILD_VECTOR) return false; return isConstantOrConstantVector(V, NoOpaques) || ISD::isBuildVectorOfConstantFPSDNodes(V.getNode()); } // Determine if this an indexed load with an opaque target constant index. static bool canSplitIdx(LoadSDNode *LD) { return MaySplitLoadIndex && (LD->getOperand(2).getOpcode() != ISD::TargetConstant || !cast(LD->getOperand(2))->isOpaque()); } bool DAGCombiner::reassociationCanBreakAddressingModePattern(unsigned Opc, const SDLoc &DL, SDNode *N, SDValue N0, SDValue N1) { // Currently this only tries to ensure we don't undo the GEP splits done by // CodeGenPrepare when shouldConsiderGEPOffsetSplit is true. To ensure this, // we check if the following transformation would be problematic: // (load/store (add, (add, x, offset1), offset2)) -> // (load/store (add, x, offset1+offset2)). // (load/store (add, (add, x, y), offset2)) -> // (load/store (add, (add, x, offset2), y)). if (Opc != ISD::ADD || N0.getOpcode() != ISD::ADD) return false; auto *C2 = dyn_cast(N1); if (!C2) return false; const APInt &C2APIntVal = C2->getAPIntValue(); if (C2APIntVal.getSignificantBits() > 64) return false; if (auto *C1 = dyn_cast(N0.getOperand(1))) { if (N0.hasOneUse()) return false; const APInt &C1APIntVal = C1->getAPIntValue(); const APInt CombinedValueIntVal = C1APIntVal + C2APIntVal; if (CombinedValueIntVal.getSignificantBits() > 64) return false; const int64_t CombinedValue = CombinedValueIntVal.getSExtValue(); for (SDNode *Node : N->uses()) { if (auto *LoadStore = dyn_cast(Node)) { // Is x[offset2] already not a legal addressing mode? If so then // reassociating the constants breaks nothing (we test offset2 because // that's the one we hope to fold into the load or store). TargetLoweringBase::AddrMode AM; AM.HasBaseReg = true; AM.BaseOffs = C2APIntVal.getSExtValue(); EVT VT = LoadStore->getMemoryVT(); unsigned AS = LoadStore->getAddressSpace(); Type *AccessTy = VT.getTypeForEVT(*DAG.getContext()); if (!TLI.isLegalAddressingMode(DAG.getDataLayout(), AM, AccessTy, AS)) continue; // Would x[offset1+offset2] still be a legal addressing mode? AM.BaseOffs = CombinedValue; if (!TLI.isLegalAddressingMode(DAG.getDataLayout(), AM, AccessTy, AS)) return true; } } } else { if (auto *GA = dyn_cast(N0.getOperand(1))) if (GA->getOpcode() == ISD::GlobalAddress && TLI.isOffsetFoldingLegal(GA)) return false; for (SDNode *Node : N->uses()) { auto *LoadStore = dyn_cast(Node); if (!LoadStore) return false; // Is x[offset2] a legal addressing mode? If so then // reassociating the constants breaks address pattern TargetLoweringBase::AddrMode AM; AM.HasBaseReg = true; AM.BaseOffs = C2APIntVal.getSExtValue(); EVT VT = LoadStore->getMemoryVT(); unsigned AS = LoadStore->getAddressSpace(); Type *AccessTy = VT.getTypeForEVT(*DAG.getContext()); if (!TLI.isLegalAddressingMode(DAG.getDataLayout(), AM, AccessTy, AS)) return false; } return true; } return false; } // Helper for DAGCombiner::reassociateOps. Try to reassociate an expression // such as (Opc N0, N1), if \p N0 is the same kind of operation as \p Opc. SDValue DAGCombiner::reassociateOpsCommutative(unsigned Opc, const SDLoc &DL, SDValue N0, SDValue N1) { EVT VT = N0.getValueType(); if (N0.getOpcode() != Opc) return SDValue(); SDValue N00 = N0.getOperand(0); SDValue N01 = N0.getOperand(1); if (DAG.isConstantIntBuildVectorOrConstantInt(peekThroughBitcasts(N01))) { if (DAG.isConstantIntBuildVectorOrConstantInt(peekThroughBitcasts(N1))) { // Reassociate: (op (op x, c1), c2) -> (op x, (op c1, c2)) if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, DL, VT, {N01, N1})) return DAG.getNode(Opc, DL, VT, N00, OpNode); return SDValue(); } if (TLI.isReassocProfitable(DAG, N0, N1)) { // Reassociate: (op (op x, c1), y) -> (op (op x, y), c1) // iff (op x, c1) has one use SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N00, N1); return DAG.getNode(Opc, DL, VT, OpNode, N01); } } // Check for repeated operand logic simplifications. if (Opc == ISD::AND || Opc == ISD::OR) { // (N00 & N01) & N00 --> N00 & N01 // (N00 & N01) & N01 --> N00 & N01 // (N00 | N01) | N00 --> N00 | N01 // (N00 | N01) | N01 --> N00 | N01 if (N1 == N00 || N1 == N01) return N0; } if (Opc == ISD::XOR) { // (N00 ^ N01) ^ N00 --> N01 if (N1 == N00) return N01; // (N00 ^ N01) ^ N01 --> N00 if (N1 == N01) return N00; } if (TLI.isReassocProfitable(DAG, N0, N1)) { if (N1 != N01) { // Reassociate if (op N00, N1) already exist if (SDNode *NE = DAG.getNodeIfExists(Opc, DAG.getVTList(VT), {N00, N1})) { // if Op (Op N00, N1), N01 already exist // we need to stop reassciate to avoid dead loop if (!DAG.doesNodeExist(Opc, DAG.getVTList(VT), {SDValue(NE, 0), N01})) return DAG.getNode(Opc, DL, VT, SDValue(NE, 0), N01); } } if (N1 != N00) { // Reassociate if (op N01, N1) already exist if (SDNode *NE = DAG.getNodeIfExists(Opc, DAG.getVTList(VT), {N01, N1})) { // if Op (Op N01, N1), N00 already exist // we need to stop reassciate to avoid dead loop if (!DAG.doesNodeExist(Opc, DAG.getVTList(VT), {SDValue(NE, 0), N00})) return DAG.getNode(Opc, DL, VT, SDValue(NE, 0), N00); } } } return SDValue(); } // Try to reassociate commutative binops. SDValue DAGCombiner::reassociateOps(unsigned Opc, const SDLoc &DL, SDValue N0, SDValue N1, SDNodeFlags Flags) { assert(TLI.isCommutativeBinOp(Opc) && "Operation not commutative."); // Floating-point reassociation is not allowed without loose FP math. if (N0.getValueType().isFloatingPoint() || N1.getValueType().isFloatingPoint()) if (!Flags.hasAllowReassociation() || !Flags.hasNoSignedZeros()) return SDValue(); if (SDValue Combined = reassociateOpsCommutative(Opc, DL, N0, N1)) return Combined; if (SDValue Combined = reassociateOpsCommutative(Opc, DL, N1, N0)) return Combined; return SDValue(); } SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo, bool AddTo) { assert(N->getNumValues() == NumTo && "Broken CombineTo call!"); ++NodesCombined; LLVM_DEBUG(dbgs() << "\nReplacing.1 "; N->dump(&DAG); dbgs() << "\nWith: "; To[0].dump(&DAG); dbgs() << " and " << NumTo - 1 << " other values\n"); for (unsigned i = 0, e = NumTo; i != e; ++i) assert((!To[i].getNode() || N->getValueType(i) == To[i].getValueType()) && "Cannot combine value to value of different type!"); WorklistRemover DeadNodes(*this); DAG.ReplaceAllUsesWith(N, To); if (AddTo) { // Push the new nodes and any users onto the worklist for (unsigned i = 0, e = NumTo; i != e; ++i) { if (To[i].getNode()) AddToWorklistWithUsers(To[i].getNode()); } } // Finally, if the node is now dead, remove it from the graph. The node // may not be dead if the replacement process recursively simplified to // something else needing this node. if (N->use_empty()) deleteAndRecombine(N); return SDValue(N, 0); } void DAGCombiner:: CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) { // Replace the old value with the new one. ++NodesCombined; LLVM_DEBUG(dbgs() << "\nReplacing.2 "; TLO.Old.dump(&DAG); dbgs() << "\nWith: "; TLO.New.dump(&DAG); dbgs() << '\n'); // Replace all uses. If any nodes become isomorphic to other nodes and // are deleted, make sure to remove them from our worklist. WorklistRemover DeadNodes(*this); DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New); // Push the new node and any (possibly new) users onto the worklist. AddToWorklistWithUsers(TLO.New.getNode()); // Finally, if the node is now dead, remove it from the graph. The node // may not be dead if the replacement process recursively simplified to // something else needing this node. if (TLO.Old->use_empty()) deleteAndRecombine(TLO.Old.getNode()); } /// Check the specified integer node value to see if it can be simplified or if /// things it uses can be simplified by bit propagation. If so, return true. bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts, bool AssumeSingleUse) { TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations); KnownBits Known; if (!TLI.SimplifyDemandedBits(Op, DemandedBits, DemandedElts, Known, TLO, 0, AssumeSingleUse)) return false; // Revisit the node. AddToWorklist(Op.getNode()); CommitTargetLoweringOpt(TLO); return true; } /// Check the specified vector node value to see if it can be simplified or /// if things it uses can be simplified as it only uses some of the elements. /// If so, return true. bool DAGCombiner::SimplifyDemandedVectorElts(SDValue Op, const APInt &DemandedElts, bool AssumeSingleUse) { TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations); APInt KnownUndef, KnownZero; if (!TLI.SimplifyDemandedVectorElts(Op, DemandedElts, KnownUndef, KnownZero, TLO, 0, AssumeSingleUse)) return false; // Revisit the node. AddToWorklist(Op.getNode()); CommitTargetLoweringOpt(TLO); return true; } void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) { SDLoc DL(Load); EVT VT = Load->getValueType(0); SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, VT, SDValue(ExtLoad, 0)); LLVM_DEBUG(dbgs() << "\nReplacing.9 "; Load->dump(&DAG); dbgs() << "\nWith: "; Trunc.dump(&DAG); dbgs() << '\n'); WorklistRemover DeadNodes(*this); DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc); DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1)); deleteAndRecombine(Load); AddToWorklist(Trunc.getNode()); } SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) { Replace = false; SDLoc DL(Op); if (ISD::isUNINDEXEDLoad(Op.getNode())) { LoadSDNode *LD = cast(Op); EVT MemVT = LD->getMemoryVT(); ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD) ? ISD::EXTLOAD : LD->getExtensionType(); Replace = true; return DAG.getExtLoad(ExtType, DL, PVT, LD->getChain(), LD->getBasePtr(), MemVT, LD->getMemOperand()); } unsigned Opc = Op.getOpcode(); switch (Opc) { default: break; case ISD::AssertSext: if (SDValue Op0 = SExtPromoteOperand(Op.getOperand(0), PVT)) return DAG.getNode(ISD::AssertSext, DL, PVT, Op0, Op.getOperand(1)); break; case ISD::AssertZext: if (SDValue Op0 = ZExtPromoteOperand(Op.getOperand(0), PVT)) return DAG.getNode(ISD::AssertZext, DL, PVT, Op0, Op.getOperand(1)); break; case ISD::Constant: { unsigned ExtOpc = Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; return DAG.getNode(ExtOpc, DL, PVT, Op); } } if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT)) return SDValue(); return DAG.getNode(ISD::ANY_EXTEND, DL, PVT, Op); } SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) { if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT)) return SDValue(); EVT OldVT = Op.getValueType(); SDLoc DL(Op); bool Replace = false; SDValue NewOp = PromoteOperand(Op, PVT, Replace); if (!NewOp.getNode()) return SDValue(); AddToWorklist(NewOp.getNode()); if (Replace) ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode()); return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, NewOp.getValueType(), NewOp, DAG.getValueType(OldVT)); } SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) { EVT OldVT = Op.getValueType(); SDLoc DL(Op); bool Replace = false; SDValue NewOp = PromoteOperand(Op, PVT, Replace); if (!NewOp.getNode()) return SDValue(); AddToWorklist(NewOp.getNode()); if (Replace) ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode()); return DAG.getZeroExtendInReg(NewOp, DL, OldVT); } /// Promote the specified integer binary operation if the target indicates it is /// beneficial. e.g. On x86, it's usually better to promote i16 operations to /// i32 since i16 instructions are longer. SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) { if (!LegalOperations) return SDValue(); EVT VT = Op.getValueType(); if (VT.isVector() || !VT.isInteger()) return SDValue(); // If operation type is 'undesirable', e.g. i16 on x86, consider // promoting it. unsigned Opc = Op.getOpcode(); if (TLI.isTypeDesirableForOp(Opc, VT)) return SDValue(); EVT PVT = VT; // Consult target whether it is a good idea to promote this operation and // what's the right type to promote it to. if (TLI.IsDesirableToPromoteOp(Op, PVT)) { assert(PVT != VT && "Don't know what type to promote to!"); LLVM_DEBUG(dbgs() << "\nPromoting "; Op.dump(&DAG)); bool Replace0 = false; SDValue N0 = Op.getOperand(0); SDValue NN0 = PromoteOperand(N0, PVT, Replace0); bool Replace1 = false; SDValue N1 = Op.getOperand(1); SDValue NN1 = PromoteOperand(N1, PVT, Replace1); SDLoc DL(Op); SDValue RV = DAG.getNode(ISD::TRUNCATE, DL, VT, DAG.getNode(Opc, DL, PVT, NN0, NN1)); // We are always replacing N0/N1's use in N and only need additional // replacements if there are additional uses. // Note: We are checking uses of the *nodes* (SDNode) rather than values // (SDValue) here because the node may reference multiple values // (for example, the chain value of a load node). Replace0 &= !N0->hasOneUse(); Replace1 &= (N0 != N1) && !N1->hasOneUse(); // Combine Op here so it is preserved past replacements. CombineTo(Op.getNode(), RV); // If operands have a use ordering, make sure we deal with // predecessor first. if (Replace0 && Replace1 && N0->isPredecessorOf(N1.getNode())) { std::swap(N0, N1); std::swap(NN0, NN1); } if (Replace0) { AddToWorklist(NN0.getNode()); ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode()); } if (Replace1) { AddToWorklist(NN1.getNode()); ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode()); } return Op; } return SDValue(); } /// Promote the specified integer shift operation if the target indicates it is /// beneficial. e.g. On x86, it's usually better to promote i16 operations to /// i32 since i16 instructions are longer. SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) { if (!LegalOperations) return SDValue(); EVT VT = Op.getValueType(); if (VT.isVector() || !VT.isInteger()) return SDValue(); // If operation type is 'undesirable', e.g. i16 on x86, consider // promoting it. unsigned Opc = Op.getOpcode(); if (TLI.isTypeDesirableForOp(Opc, VT)) return SDValue(); EVT PVT = VT; // Consult target whether it is a good idea to promote this operation and // what's the right type to promote it to. if (TLI.IsDesirableToPromoteOp(Op, PVT)) { assert(PVT != VT && "Don't know what type to promote to!"); LLVM_DEBUG(dbgs() << "\nPromoting "; Op.dump(&DAG)); bool Replace = false; SDValue N0 = Op.getOperand(0); if (Opc == ISD::SRA) N0 = SExtPromoteOperand(N0, PVT); else if (Opc == ISD::SRL) N0 = ZExtPromoteOperand(N0, PVT); else N0 = PromoteOperand(N0, PVT, Replace); if (!N0.getNode()) return SDValue(); SDLoc DL(Op); SDValue N1 = Op.getOperand(1); SDValue RV = DAG.getNode(ISD::TRUNCATE, DL, VT, DAG.getNode(Opc, DL, PVT, N0, N1)); if (Replace) ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode()); // Deal with Op being deleted. if (Op && Op.getOpcode() != ISD::DELETED_NODE) return RV; } return SDValue(); } SDValue DAGCombiner::PromoteExtend(SDValue Op) { if (!LegalOperations) return SDValue(); EVT VT = Op.getValueType(); if (VT.isVector() || !VT.isInteger()) return SDValue(); // If operation type is 'undesirable', e.g. i16 on x86, consider // promoting it. unsigned Opc = Op.getOpcode(); if (TLI.isTypeDesirableForOp(Opc, VT)) return SDValue(); EVT PVT = VT; // Consult target whether it is a good idea to promote this operation and // what's the right type to promote it to. if (TLI.IsDesirableToPromoteOp(Op, PVT)) { assert(PVT != VT && "Don't know what type to promote to!"); // fold (aext (aext x)) -> (aext x) // fold (aext (zext x)) -> (zext x) // fold (aext (sext x)) -> (sext x) LLVM_DEBUG(dbgs() << "\nPromoting "; Op.dump(&DAG)); return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0)); } return SDValue(); } bool DAGCombiner::PromoteLoad(SDValue Op) { if (!LegalOperations) return false; if (!ISD::isUNINDEXEDLoad(Op.getNode())) return false; EVT VT = Op.getValueType(); if (VT.isVector() || !VT.isInteger()) return false; // If operation type is 'undesirable', e.g. i16 on x86, consider // promoting it. unsigned Opc = Op.getOpcode(); if (TLI.isTypeDesirableForOp(Opc, VT)) return false; EVT PVT = VT; // Consult target whether it is a good idea to promote this operation and // what's the right type to promote it to. if (TLI.IsDesirableToPromoteOp(Op, PVT)) { assert(PVT != VT && "Don't know what type to promote to!"); SDLoc DL(Op); SDNode *N = Op.getNode(); LoadSDNode *LD = cast(N); EVT MemVT = LD->getMemoryVT(); ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD) ? ISD::EXTLOAD : LD->getExtensionType(); SDValue NewLD = DAG.getExtLoad(ExtType, DL, PVT, LD->getChain(), LD->getBasePtr(), MemVT, LD->getMemOperand()); SDValue Result = DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD); LLVM_DEBUG(dbgs() << "\nPromoting "; N->dump(&DAG); dbgs() << "\nTo: "; Result.dump(&DAG); dbgs() << '\n'); WorklistRemover DeadNodes(*this); DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result); DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1)); deleteAndRecombine(N); AddToWorklist(Result.getNode()); return true; } return false; } /// Recursively delete a node which has no uses and any operands for /// which it is the only use. /// /// Note that this both deletes the nodes and removes them from the worklist. /// It also adds any nodes who have had a user deleted to the worklist as they /// may now have only one use and subject to other combines. bool DAGCombiner::recursivelyDeleteUnusedNodes(SDNode *N) { if (!N->use_empty()) return false; SmallSetVector Nodes; Nodes.insert(N); do { N = Nodes.pop_back_val(); if (!N) continue; if (N->use_empty()) { for (const SDValue &ChildN : N->op_values()) Nodes.insert(ChildN.getNode()); removeFromWorklist(N); DAG.DeleteNode(N); } else { AddToWorklist(N); } } while (!Nodes.empty()); return true; } //===----------------------------------------------------------------------===// // Main DAG Combiner implementation //===----------------------------------------------------------------------===// void DAGCombiner::Run(CombineLevel AtLevel) { // set the instance variables, so that the various visit routines may use it. Level = AtLevel; LegalDAG = Level >= AfterLegalizeDAG; LegalOperations = Level >= AfterLegalizeVectorOps; LegalTypes = Level >= AfterLegalizeTypes; WorklistInserter AddNodes(*this); // Add all the dag nodes to the worklist. for (SDNode &Node : DAG.allnodes()) AddToWorklist(&Node); // Create a dummy node (which is not added to allnodes), that adds a reference // to the root node, preventing it from being deleted, and tracking any // changes of the root. HandleSDNode Dummy(DAG.getRoot()); // While we have a valid worklist entry node, try to combine it. while (SDNode *N = getNextWorklistEntry()) { // If N has no uses, it is dead. Make sure to revisit all N's operands once // N is deleted from the DAG, since they too may now be dead or may have a // reduced number of uses, allowing other xforms. if (recursivelyDeleteUnusedNodes(N)) continue; WorklistRemover DeadNodes(*this); // If this combine is running after legalizing the DAG, re-legalize any // nodes pulled off the worklist. if (LegalDAG) { SmallSetVector UpdatedNodes; bool NIsValid = DAG.LegalizeOp(N, UpdatedNodes); for (SDNode *LN : UpdatedNodes) AddToWorklistWithUsers(LN); if (!NIsValid) continue; } LLVM_DEBUG(dbgs() << "\nCombining: "; N->dump(&DAG)); // Add any operands of the new node which have not yet been combined to the // worklist as well. Because the worklist uniques things already, this // won't repeatedly process the same operand. CombinedNodes.insert(N); for (const SDValue &ChildN : N->op_values()) if (!CombinedNodes.count(ChildN.getNode())) AddToWorklist(ChildN.getNode()); SDValue RV = combine(N); if (!RV.getNode()) continue; ++NodesCombined; // If we get back the same node we passed in, rather than a new node or // zero, we know that the node must have defined multiple values and // CombineTo was used. Since CombineTo takes care of the worklist // mechanics for us, we have no work to do in this case. if (RV.getNode() == N) continue; assert(N->getOpcode() != ISD::DELETED_NODE && RV.getOpcode() != ISD::DELETED_NODE && "Node was deleted but visit returned new node!"); LLVM_DEBUG(dbgs() << " ... into: "; RV.dump(&DAG)); if (N->getNumValues() == RV->getNumValues()) DAG.ReplaceAllUsesWith(N, RV.getNode()); else { assert(N->getValueType(0) == RV.getValueType() && N->getNumValues() == 1 && "Type mismatch"); DAG.ReplaceAllUsesWith(N, &RV); } // Push the new node and any users onto the worklist. Omit this if the // new node is the EntryToken (e.g. if a store managed to get optimized // out), because re-visiting the EntryToken and its users will not uncover // any additional opportunities, but there may be a large number of such // users, potentially causing compile time explosion. if (RV.getOpcode() != ISD::EntryToken) { AddToWorklist(RV.getNode()); AddUsersToWorklist(RV.getNode()); } // Finally, if the node is now dead, remove it from the graph. The node // may not be dead if the replacement process recursively simplified to // something else needing this node. This will also take care of adding any // operands which have lost a user to the worklist. recursivelyDeleteUnusedNodes(N); } // If the root changed (e.g. it was a dead load, update the root). DAG.setRoot(Dummy.getValue()); DAG.RemoveDeadNodes(); } SDValue DAGCombiner::visit(SDNode *N) { switch (N->getOpcode()) { default: break; case ISD::TokenFactor: return visitTokenFactor(N); case ISD::MERGE_VALUES: return visitMERGE_VALUES(N); case ISD::ADD: return visitADD(N); case ISD::SUB: return visitSUB(N); case ISD::SADDSAT: case ISD::UADDSAT: return visitADDSAT(N); case ISD::SSUBSAT: case ISD::USUBSAT: return visitSUBSAT(N); case ISD::ADDC: return visitADDC(N); case ISD::SADDO: case ISD::UADDO: return visitADDO(N); case ISD::SUBC: return visitSUBC(N); case ISD::SSUBO: case ISD::USUBO: return visitSUBO(N); case ISD::ADDE: return visitADDE(N); case ISD::ADDCARRY: return visitADDCARRY(N); case ISD::SADDO_CARRY: return visitSADDO_CARRY(N); case ISD::SUBE: return visitSUBE(N); case ISD::SUBCARRY: return visitSUBCARRY(N); case ISD::SSUBO_CARRY: return visitSSUBO_CARRY(N); case ISD::SMULFIX: case ISD::SMULFIXSAT: case ISD::UMULFIX: case ISD::UMULFIXSAT: return visitMULFIX(N); case ISD::MUL: return visitMUL(N); case ISD::SDIV: return visitSDIV(N); case ISD::UDIV: return visitUDIV(N); case ISD::SREM: case ISD::UREM: return visitREM(N); case ISD::MULHU: return visitMULHU(N); case ISD::MULHS: return visitMULHS(N); case ISD::AVGFLOORS: case ISD::AVGFLOORU: case ISD::AVGCEILS: case ISD::AVGCEILU: return visitAVG(N); case ISD::SMUL_LOHI: return visitSMUL_LOHI(N); case ISD::UMUL_LOHI: return visitUMUL_LOHI(N); case ISD::SMULO: case ISD::UMULO: return visitMULO(N); case ISD::SMIN: case ISD::SMAX: case ISD::UMIN: case ISD::UMAX: return visitIMINMAX(N); case ISD::AND: return visitAND(N); case ISD::OR: return visitOR(N); case ISD::XOR: return visitXOR(N); case ISD::SHL: return visitSHL(N); case ISD::SRA: return visitSRA(N); case ISD::SRL: return visitSRL(N); case ISD::ROTR: case ISD::ROTL: return visitRotate(N); case ISD::FSHL: case ISD::FSHR: return visitFunnelShift(N); case ISD::SSHLSAT: case ISD::USHLSAT: return visitSHLSAT(N); case ISD::ABS: return visitABS(N); case ISD::BSWAP: return visitBSWAP(N); case ISD::BITREVERSE: return visitBITREVERSE(N); case ISD::CTLZ: return visitCTLZ(N); case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N); case ISD::CTTZ: return visitCTTZ(N); case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N); case ISD::CTPOP: return visitCTPOP(N); case ISD::SELECT: return visitSELECT(N); case ISD::VSELECT: return visitVSELECT(N); case ISD::SELECT_CC: return visitSELECT_CC(N); case ISD::SETCC: return visitSETCC(N); case ISD::SETCCCARRY: return visitSETCCCARRY(N); case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N); case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N); case ISD::ANY_EXTEND: return visitANY_EXTEND(N); case ISD::AssertSext: case ISD::AssertZext: return visitAssertExt(N); case ISD::AssertAlign: return visitAssertAlign(N); case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N); case ISD::SIGN_EXTEND_VECTOR_INREG: case ISD::ZERO_EXTEND_VECTOR_INREG: return visitEXTEND_VECTOR_INREG(N); case ISD::TRUNCATE: return visitTRUNCATE(N); case ISD::BITCAST: return visitBITCAST(N); case ISD::BUILD_PAIR: return visitBUILD_PAIR(N); case ISD::FADD: return visitFADD(N); case ISD::STRICT_FADD: return visitSTRICT_FADD(N); case ISD::FSUB: return visitFSUB(N); case ISD::FMUL: return visitFMUL(N); case ISD::FMA: return visitFMA(N); case ISD::FDIV: return visitFDIV(N); case ISD::FREM: return visitFREM(N); case ISD::FSQRT: return visitFSQRT(N); case ISD::FCOPYSIGN: return visitFCOPYSIGN(N); case ISD::FPOW: return visitFPOW(N); case ISD::SINT_TO_FP: return visitSINT_TO_FP(N); case ISD::UINT_TO_FP: return visitUINT_TO_FP(N); case ISD::FP_TO_SINT: return visitFP_TO_SINT(N); case ISD::FP_TO_UINT: return visitFP_TO_UINT(N); case ISD::FP_ROUND: return visitFP_ROUND(N); case ISD::FP_EXTEND: return visitFP_EXTEND(N); case ISD::FNEG: return visitFNEG(N); case ISD::FABS: return visitFABS(N); case ISD::FFLOOR: return visitFFLOOR(N); case ISD::FMINNUM: case ISD::FMAXNUM: case ISD::FMINIMUM: case ISD::FMAXIMUM: return visitFMinMax(N); case ISD::FCEIL: return visitFCEIL(N); case ISD::FTRUNC: return visitFTRUNC(N); case ISD::BRCOND: return visitBRCOND(N); case ISD::BR_CC: return visitBR_CC(N); case ISD::LOAD: return visitLOAD(N); case ISD::STORE: return visitSTORE(N); case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N); case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N); case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N); case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N); case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N); case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N); case ISD::SCALAR_TO_VECTOR: return visitSCALAR_TO_VECTOR(N); case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N); case ISD::MGATHER: return visitMGATHER(N); case ISD::MLOAD: return visitMLOAD(N); case ISD::MSCATTER: return visitMSCATTER(N); case ISD::MSTORE: return visitMSTORE(N); case ISD::LIFETIME_END: return visitLIFETIME_END(N); case ISD::FP_TO_FP16: return visitFP_TO_FP16(N); case ISD::FP16_TO_FP: return visitFP16_TO_FP(N); case ISD::FP_TO_BF16: return visitFP_TO_BF16(N); case ISD::FREEZE: return visitFREEZE(N); case ISD::VECREDUCE_FADD: case ISD::VECREDUCE_FMUL: case ISD::VECREDUCE_ADD: case ISD::VECREDUCE_MUL: case ISD::VECREDUCE_AND: case ISD::VECREDUCE_OR: case ISD::VECREDUCE_XOR: case ISD::VECREDUCE_SMAX: case ISD::VECREDUCE_SMIN: case ISD::VECREDUCE_UMAX: case ISD::VECREDUCE_UMIN: case ISD::VECREDUCE_FMAX: case ISD::VECREDUCE_FMIN: return visitVECREDUCE(N); #define BEGIN_REGISTER_VP_SDNODE(SDOPC, ...) case ISD::SDOPC: #include "llvm/IR/VPIntrinsics.def" return visitVPOp(N); } return SDValue(); } SDValue DAGCombiner::combine(SDNode *N) { SDValue RV; if (!DisableGenericCombines) RV = visit(N); // If nothing happened, try a target-specific DAG combine. if (!RV.getNode()) { assert(N->getOpcode() != ISD::DELETED_NODE && "Node was deleted but visit returned NULL!"); if (N->getOpcode() >= ISD::BUILTIN_OP_END || TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) { // Expose the DAG combiner to the target combiner impls. TargetLowering::DAGCombinerInfo DagCombineInfo(DAG, Level, false, this); RV = TLI.PerformDAGCombine(N, DagCombineInfo); } } // If nothing happened still, try promoting the operation. if (!RV.getNode()) { switch (N->getOpcode()) { default: break; case ISD::ADD: case ISD::SUB: case ISD::MUL: case ISD::AND: case ISD::OR: case ISD::XOR: RV = PromoteIntBinOp(SDValue(N, 0)); break; case ISD::SHL: case ISD::SRA: case ISD::SRL: RV = PromoteIntShiftOp(SDValue(N, 0)); break; case ISD::SIGN_EXTEND: case ISD::ZERO_EXTEND: case ISD::ANY_EXTEND: RV = PromoteExtend(SDValue(N, 0)); break; case ISD::LOAD: if (PromoteLoad(SDValue(N, 0))) RV = SDValue(N, 0); break; } } // If N is a commutative binary node, try to eliminate it if the commuted // version is already present in the DAG. if (!RV.getNode() && TLI.isCommutativeBinOp(N->getOpcode())) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); // Constant operands are canonicalized to RHS. if (N0 != N1 && (isa(N0) || !isa(N1))) { SDValue Ops[] = {N1, N0}; SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops, N->getFlags()); if (CSENode) return SDValue(CSENode, 0); } } return RV; } /// Given a node, return its input chain if it has one, otherwise return a null /// sd operand. static SDValue getInputChainForNode(SDNode *N) { if (unsigned NumOps = N->getNumOperands()) { if (N->getOperand(0).getValueType() == MVT::Other) return N->getOperand(0); if (N->getOperand(NumOps-1).getValueType() == MVT::Other) return N->getOperand(NumOps-1); for (unsigned i = 1; i < NumOps-1; ++i) if (N->getOperand(i).getValueType() == MVT::Other) return N->getOperand(i); } return SDValue(); } SDValue DAGCombiner::visitTokenFactor(SDNode *N) { // If N has two operands, where one has an input chain equal to the other, // the 'other' chain is redundant. if (N->getNumOperands() == 2) { if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1)) return N->getOperand(0); if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0)) return N->getOperand(1); } // Don't simplify token factors if optnone. if (OptLevel == CodeGenOpt::None) return SDValue(); // Don't simplify the token factor if the node itself has too many operands. if (N->getNumOperands() > TokenFactorInlineLimit) return SDValue(); // If the sole user is a token factor, we should make sure we have a // chance to merge them together. This prevents TF chains from inhibiting // optimizations. if (N->hasOneUse() && N->use_begin()->getOpcode() == ISD::TokenFactor) AddToWorklist(*(N->use_begin())); SmallVector TFs; // List of token factors to visit. SmallVector Ops; // Ops for replacing token factor. SmallPtrSet SeenOps; bool Changed = false; // If we should replace this token factor. // Start out with this token factor. TFs.push_back(N); // Iterate through token factors. The TFs grows when new token factors are // encountered. for (unsigned i = 0; i < TFs.size(); ++i) { // Limit number of nodes to inline, to avoid quadratic compile times. // We have to add the outstanding Token Factors to Ops, otherwise we might // drop Ops from the resulting Token Factors. if (Ops.size() > TokenFactorInlineLimit) { for (unsigned j = i; j < TFs.size(); j++) Ops.emplace_back(TFs[j], 0); // Drop unprocessed Token Factors from TFs, so we do not add them to the // combiner worklist later. TFs.resize(i); break; } SDNode *TF = TFs[i]; // Check each of the operands. for (const SDValue &Op : TF->op_values()) { switch (Op.getOpcode()) { case ISD::EntryToken: // Entry tokens don't need to be added to the list. They are // redundant. Changed = true; break; case ISD::TokenFactor: if (Op.hasOneUse() && !is_contained(TFs, Op.getNode())) { // Queue up for processing. TFs.push_back(Op.getNode()); Changed = true; break; } LLVM_FALLTHROUGH; default: // Only add if it isn't already in the list. if (SeenOps.insert(Op.getNode()).second) Ops.push_back(Op); else Changed = true; break; } } } // Re-visit inlined Token Factors, to clean them up in case they have been // removed. Skip the first Token Factor, as this is the current node. for (unsigned i = 1, e = TFs.size(); i < e; i++) AddToWorklist(TFs[i]); // Remove Nodes that are chained to another node in the list. Do so // by walking up chains breath-first stopping when we've seen // another operand. In general we must climb to the EntryNode, but we can exit // early if we find all remaining work is associated with just one operand as // no further pruning is possible. // List of nodes to search through and original Ops from which they originate. SmallVector, 8> Worklist; SmallVector OpWorkCount; // Count of work for each Op. SmallPtrSet SeenChains; bool DidPruneOps = false; unsigned NumLeftToConsider = 0; for (const SDValue &Op : Ops) { Worklist.push_back(std::make_pair(Op.getNode(), NumLeftToConsider++)); OpWorkCount.push_back(1); } auto AddToWorklist = [&](unsigned CurIdx, SDNode *Op, unsigned OpNumber) { // If this is an Op, we can remove the op from the list. Remark any // search associated with it as from the current OpNumber. if (SeenOps.contains(Op)) { Changed = true; DidPruneOps = true; unsigned OrigOpNumber = 0; while (OrigOpNumber < Ops.size() && Ops[OrigOpNumber].getNode() != Op) OrigOpNumber++; assert((OrigOpNumber != Ops.size()) && "expected to find TokenFactor Operand"); // Re-mark worklist from OrigOpNumber to OpNumber for (unsigned i = CurIdx + 1; i < Worklist.size(); ++i) { if (Worklist[i].second == OrigOpNumber) { Worklist[i].second = OpNumber; } } OpWorkCount[OpNumber] += OpWorkCount[OrigOpNumber]; OpWorkCount[OrigOpNumber] = 0; NumLeftToConsider--; } // Add if it's a new chain if (SeenChains.insert(Op).second) { OpWorkCount[OpNumber]++; Worklist.push_back(std::make_pair(Op, OpNumber)); } }; for (unsigned i = 0; i < Worklist.size() && i < 1024; ++i) { // We need at least be consider at least 2 Ops to prune. if (NumLeftToConsider <= 1) break; auto CurNode = Worklist[i].first; auto CurOpNumber = Worklist[i].second; assert((OpWorkCount[CurOpNumber] > 0) && "Node should not appear in worklist"); switch (CurNode->getOpcode()) { case ISD::EntryToken: // Hitting EntryToken is the only way for the search to terminate without // hitting // another operand's search. Prevent us from marking this operand // considered. NumLeftToConsider++; break; case ISD::TokenFactor: for (const SDValue &Op : CurNode->op_values()) AddToWorklist(i, Op.getNode(), CurOpNumber); break; case ISD::LIFETIME_START: case ISD::LIFETIME_END: case ISD::CopyFromReg: case ISD::CopyToReg: AddToWorklist(i, CurNode->getOperand(0).getNode(), CurOpNumber); break; default: if (auto *MemNode = dyn_cast(CurNode)) AddToWorklist(i, MemNode->getChain().getNode(), CurOpNumber); break; } OpWorkCount[CurOpNumber]--; if (OpWorkCount[CurOpNumber] == 0) NumLeftToConsider--; } // If we've changed things around then replace token factor. if (Changed) { SDValue Result; if (Ops.empty()) { // The entry token is the only possible outcome. Result = DAG.getEntryNode(); } else { if (DidPruneOps) { SmallVector PrunedOps; // for (const SDValue &Op : Ops) { if (SeenChains.count(Op.getNode()) == 0) PrunedOps.push_back(Op); } Result = DAG.getTokenFactor(SDLoc(N), PrunedOps); } else { Result = DAG.getTokenFactor(SDLoc(N), Ops); } } return Result; } return SDValue(); } /// MERGE_VALUES can always be eliminated. SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) { WorklistRemover DeadNodes(*this); // Replacing results may cause a different MERGE_VALUES to suddenly // be CSE'd with N, and carry its uses with it. Iterate until no // uses remain, to ensure that the node can be safely deleted. // First add the users of this node to the work list so that they // can be tried again once they have new operands. AddUsersToWorklist(N); do { // Do as a single replacement to avoid rewalking use lists. SmallVector Ops; for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) Ops.push_back(N->getOperand(i)); DAG.ReplaceAllUsesWith(N, Ops.data()); } while (!N->use_empty()); deleteAndRecombine(N); return SDValue(N, 0); // Return N so it doesn't get rechecked! } /// If \p N is a ConstantSDNode with isOpaque() == false return it casted to a /// ConstantSDNode pointer else nullptr. static ConstantSDNode *getAsNonOpaqueConstant(SDValue N) { ConstantSDNode *Const = dyn_cast(N); return Const != nullptr && !Const->isOpaque() ? Const : nullptr; } /// Return true if 'Use' is a load or a store that uses N as its base pointer /// and that N may be folded in the load / store addressing mode. static bool canFoldInAddressingMode(SDNode *N, SDNode *Use, SelectionDAG &DAG, const TargetLowering &TLI) { EVT VT; unsigned AS; if (LoadSDNode *LD = dyn_cast(Use)) { if (LD->isIndexed() || LD->getBasePtr().getNode() != N) return false; VT = LD->getMemoryVT(); AS = LD->getAddressSpace(); } else if (StoreSDNode *ST = dyn_cast(Use)) { if (ST->isIndexed() || ST->getBasePtr().getNode() != N) return false; VT = ST->getMemoryVT(); AS = ST->getAddressSpace(); } else if (MaskedLoadSDNode *LD = dyn_cast(Use)) { if (LD->isIndexed() || LD->getBasePtr().getNode() != N) return false; VT = LD->getMemoryVT(); AS = LD->getAddressSpace(); } else if (MaskedStoreSDNode *ST = dyn_cast(Use)) { if (ST->isIndexed() || ST->getBasePtr().getNode() != N) return false; VT = ST->getMemoryVT(); AS = ST->getAddressSpace(); } else { return false; } TargetLowering::AddrMode AM; if (N->getOpcode() == ISD::ADD) { AM.HasBaseReg = true; ConstantSDNode *Offset = dyn_cast(N->getOperand(1)); if (Offset) // [reg +/- imm] AM.BaseOffs = Offset->getSExtValue(); else // [reg +/- reg] AM.Scale = 1; } else if (N->getOpcode() == ISD::SUB) { AM.HasBaseReg = true; ConstantSDNode *Offset = dyn_cast(N->getOperand(1)); if (Offset) // [reg +/- imm] AM.BaseOffs = -Offset->getSExtValue(); else // [reg +/- reg] AM.Scale = 1; } else { return false; } return TLI.isLegalAddressingMode(DAG.getDataLayout(), AM, VT.getTypeForEVT(*DAG.getContext()), AS); } /// This inverts a canonicalization in IR that replaces a variable select arm /// with an identity constant. Codegen improves if we re-use the variable /// operand rather than load a constant. This can also be converted into a /// masked vector operation if the target supports it. static SDValue foldSelectWithIdentityConstant(SDNode *N, SelectionDAG &DAG, bool ShouldCommuteOperands) { // Match a select as operand 1. The identity constant that we are looking for // is only valid as operand 1 of a non-commutative binop. SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); if (ShouldCommuteOperands) std::swap(N0, N1); // TODO: Should this apply to scalar select too? if (!N1.hasOneUse() || N1.getOpcode() != ISD::VSELECT) return SDValue(); unsigned Opcode = N->getOpcode(); EVT VT = N->getValueType(0); SDValue Cond = N1.getOperand(0); SDValue TVal = N1.getOperand(1); SDValue FVal = N1.getOperand(2); // TODO: The cases should match with IR's ConstantExpr::getBinOpIdentity(). // TODO: Target-specific opcodes could be added. Ex: "isCommutativeBinOp()". // TODO: With fast-math (NSZ), allow the opposite-sign form of zero? auto isIdentityConstantForOpcode = [](unsigned Opcode, SDValue V) { if (ConstantFPSDNode *C = isConstOrConstSplatFP(V)) { switch (Opcode) { case ISD::FADD: // X + -0.0 --> X return C->isZero() && C->isNegative(); case ISD::FSUB: // X - 0.0 --> X return C->isZero() && !C->isNegative(); case ISD::FMUL: // X * 1.0 --> X case ISD::FDIV: // X / 1.0 --> X return C->isExactlyValue(1.0); } } if (ConstantSDNode *C = isConstOrConstSplat(V)) { switch (Opcode) { case ISD::ADD: // X + 0 --> X case ISD::SUB: // X - 0 --> X case ISD::SHL: // X << 0 --> X case ISD::SRA: // X s>> 0 --> X case ISD::SRL: // X u>> 0 --> X return C->isZero(); case ISD::MUL: // X * 1 --> X return C->isOne(); } } return false; }; // This transform increases uses of N0, so freeze it to be safe. // binop N0, (vselect Cond, IDC, FVal) --> vselect Cond, N0, (binop N0, FVal) if (isIdentityConstantForOpcode(Opcode, TVal)) { SDValue F0 = DAG.getFreeze(N0); SDValue NewBO = DAG.getNode(Opcode, SDLoc(N), VT, F0, FVal, N->getFlags()); return DAG.getSelect(SDLoc(N), VT, Cond, F0, NewBO); } // binop N0, (vselect Cond, TVal, IDC) --> vselect Cond, (binop N0, TVal), N0 if (isIdentityConstantForOpcode(Opcode, FVal)) { SDValue F0 = DAG.getFreeze(N0); SDValue NewBO = DAG.getNode(Opcode, SDLoc(N), VT, F0, TVal, N->getFlags()); return DAG.getSelect(SDLoc(N), VT, Cond, NewBO, F0); } return SDValue(); } SDValue DAGCombiner::foldBinOpIntoSelect(SDNode *BO) { assert(TLI.isBinOp(BO->getOpcode()) && BO->getNumValues() == 1 && "Unexpected binary operator"); const TargetLowering &TLI = DAG.getTargetLoweringInfo(); auto BinOpcode = BO->getOpcode(); EVT VT = BO->getValueType(0); if (TLI.shouldFoldSelectWithIdentityConstant(BinOpcode, VT)) { if (SDValue Sel = foldSelectWithIdentityConstant(BO, DAG, false)) return Sel; if (TLI.isCommutativeBinOp(BO->getOpcode())) if (SDValue Sel = foldSelectWithIdentityConstant(BO, DAG, true)) return Sel; } // Don't do this unless the old select is going away. We want to eliminate the // binary operator, not replace a binop with a select. // TODO: Handle ISD::SELECT_CC. unsigned SelOpNo = 0; SDValue Sel = BO->getOperand(0); if (Sel.getOpcode() != ISD::SELECT || !Sel.hasOneUse()) { SelOpNo = 1; Sel = BO->getOperand(1); } if (Sel.getOpcode() != ISD::SELECT || !Sel.hasOneUse()) return SDValue(); SDValue CT = Sel.getOperand(1); if (!isConstantOrConstantVector(CT, true) && !DAG.isConstantFPBuildVectorOrConstantFP(CT)) return SDValue(); SDValue CF = Sel.getOperand(2); if (!isConstantOrConstantVector(CF, true) && !DAG.isConstantFPBuildVectorOrConstantFP(CF)) return SDValue(); // Bail out if any constants are opaque because we can't constant fold those. // The exception is "and" and "or" with either 0 or -1 in which case we can // propagate non constant operands into select. I.e.: // and (select Cond, 0, -1), X --> select Cond, 0, X // or X, (select Cond, -1, 0) --> select Cond, -1, X bool CanFoldNonConst = (BinOpcode == ISD::AND || BinOpcode == ISD::OR) && (isNullOrNullSplat(CT) || isAllOnesOrAllOnesSplat(CT)) && (isNullOrNullSplat(CF) || isAllOnesOrAllOnesSplat(CF)); SDValue CBO = BO->getOperand(SelOpNo ^ 1); if (!CanFoldNonConst && !isConstantOrConstantVector(CBO, true) && !DAG.isConstantFPBuildVectorOrConstantFP(CBO)) return SDValue(); // We have a select-of-constants followed by a binary operator with a // constant. Eliminate the binop by pulling the constant math into the select. // Example: add (select Cond, CT, CF), CBO --> select Cond, CT + CBO, CF + CBO SDLoc DL(Sel); SDValue NewCT = SelOpNo ? DAG.getNode(BinOpcode, DL, VT, CBO, CT) : DAG.getNode(BinOpcode, DL, VT, CT, CBO); if (!CanFoldNonConst && !NewCT.isUndef() && !isConstantOrConstantVector(NewCT, true) && !DAG.isConstantFPBuildVectorOrConstantFP(NewCT)) return SDValue(); SDValue NewCF = SelOpNo ? DAG.getNode(BinOpcode, DL, VT, CBO, CF) : DAG.getNode(BinOpcode, DL, VT, CF, CBO); if (!CanFoldNonConst && !NewCF.isUndef() && !isConstantOrConstantVector(NewCF, true) && !DAG.isConstantFPBuildVectorOrConstantFP(NewCF)) return SDValue(); SDValue SelectOp = DAG.getSelect(DL, VT, Sel.getOperand(0), NewCT, NewCF); SelectOp->setFlags(BO->getFlags()); return SelectOp; } static SDValue foldAddSubBoolOfMaskedVal(SDNode *N, SelectionDAG &DAG) { assert((N->getOpcode() == ISD::ADD || N->getOpcode() == ISD::SUB) && "Expecting add or sub"); // Match a constant operand and a zext operand for the math instruction: // add Z, C // sub C, Z bool IsAdd = N->getOpcode() == ISD::ADD; SDValue C = IsAdd ? N->getOperand(1) : N->getOperand(0); SDValue Z = IsAdd ? N->getOperand(0) : N->getOperand(1); auto *CN = dyn_cast(C); if (!CN || Z.getOpcode() != ISD::ZERO_EXTEND) return SDValue(); // Match the zext operand as a setcc of a boolean. if (Z.getOperand(0).getOpcode() != ISD::SETCC || Z.getOperand(0).getValueType() != MVT::i1) return SDValue(); // Match the compare as: setcc (X & 1), 0, eq. SDValue SetCC = Z.getOperand(0); ISD::CondCode CC = cast(SetCC->getOperand(2))->get(); if (CC != ISD::SETEQ || !isNullConstant(SetCC.getOperand(1)) || SetCC.getOperand(0).getOpcode() != ISD::AND || !isOneConstant(SetCC.getOperand(0).getOperand(1))) return SDValue(); // We are adding/subtracting a constant and an inverted low bit. Turn that // into a subtract/add of the low bit with incremented/decremented constant: // add (zext i1 (seteq (X & 1), 0)), C --> sub C+1, (zext (X & 1)) // sub C, (zext i1 (seteq (X & 1), 0)) --> add C-1, (zext (X & 1)) EVT VT = C.getValueType(); SDLoc DL(N); SDValue LowBit = DAG.getZExtOrTrunc(SetCC.getOperand(0), DL, VT); SDValue C1 = IsAdd ? DAG.getConstant(CN->getAPIntValue() + 1, DL, VT) : DAG.getConstant(CN->getAPIntValue() - 1, DL, VT); return DAG.getNode(IsAdd ? ISD::SUB : ISD::ADD, DL, VT, C1, LowBit); } /// Try to fold a 'not' shifted sign-bit with add/sub with constant operand into /// a shift and add with a different constant. static SDValue foldAddSubOfSignBit(SDNode *N, SelectionDAG &DAG) { assert((N->getOpcode() == ISD::ADD || N->getOpcode() == ISD::SUB) && "Expecting add or sub"); // We need a constant operand for the add/sub, and the other operand is a // logical shift right: add (srl), C or sub C, (srl). bool IsAdd = N->getOpcode() == ISD::ADD; SDValue ConstantOp = IsAdd ? N->getOperand(1) : N->getOperand(0); SDValue ShiftOp = IsAdd ? N->getOperand(0) : N->getOperand(1); if (!DAG.isConstantIntBuildVectorOrConstantInt(ConstantOp) || ShiftOp.getOpcode() != ISD::SRL) return SDValue(); // The shift must be of a 'not' value. SDValue Not = ShiftOp.getOperand(0); if (!Not.hasOneUse() || !isBitwiseNot(Not)) return SDValue(); // The shift must be moving the sign bit to the least-significant-bit. EVT VT = ShiftOp.getValueType(); SDValue ShAmt = ShiftOp.getOperand(1); ConstantSDNode *ShAmtC = isConstOrConstSplat(ShAmt); if (!ShAmtC || ShAmtC->getAPIntValue() != (VT.getScalarSizeInBits() - 1)) return SDValue(); // Eliminate the 'not' by adjusting the shift and add/sub constant: // add (srl (not X), 31), C --> add (sra X, 31), (C + 1) // sub C, (srl (not X), 31) --> add (srl X, 31), (C - 1) SDLoc DL(N); if (SDValue NewC = DAG.FoldConstantArithmetic( IsAdd ? ISD::ADD : ISD::SUB, DL, VT, {ConstantOp, DAG.getConstant(1, DL, VT)})) { SDValue NewShift = DAG.getNode(IsAdd ? ISD::SRA : ISD::SRL, DL, VT, Not.getOperand(0), ShAmt); return DAG.getNode(ISD::ADD, DL, VT, NewShift, NewC); } return SDValue(); } static bool isADDLike(SDValue V, const SelectionDAG &DAG) { unsigned Opcode = V.getOpcode(); if (Opcode == ISD::OR) return DAG.haveNoCommonBitsSet(V.getOperand(0), V.getOperand(1)); if (Opcode == ISD::XOR) return isMinSignedConstant(V.getOperand(1)); return false; } /// Try to fold a node that behaves like an ADD (note that N isn't necessarily /// an ISD::ADD here, it could for example be an ISD::OR if we know that there /// are no common bits set in the operands). SDValue DAGCombiner::visitADDLike(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N0.getValueType(); SDLoc DL(N); // fold (add x, undef) -> undef if (N0.isUndef()) return N0; if (N1.isUndef()) return N1; // fold (add c1, c2) -> c1+c2 if (SDValue C = DAG.FoldConstantArithmetic(ISD::ADD, DL, VT, {N0, N1})) return C; // canonicalize constant to RHS if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && !DAG.isConstantIntBuildVectorOrConstantInt(N1)) return DAG.getNode(ISD::ADD, DL, VT, N1, N0); // fold vector ops if (VT.isVector()) { if (SDValue FoldedVOp = SimplifyVBinOp(N, DL)) return FoldedVOp; // fold (add x, 0) -> x, vector edition if (ISD::isConstantSplatVectorAllZeros(N1.getNode())) return N0; } // fold (add x, 0) -> x if (isNullConstant(N1)) return N0; if (N0.getOpcode() == ISD::SUB) { SDValue N00 = N0.getOperand(0); SDValue N01 = N0.getOperand(1); // fold ((A-c1)+c2) -> (A+(c2-c1)) if (SDValue Sub = DAG.FoldConstantArithmetic(ISD::SUB, DL, VT, {N1, N01})) return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), Sub); // fold ((c1-A)+c2) -> (c1+c2)-A if (SDValue Add = DAG.FoldConstantArithmetic(ISD::ADD, DL, VT, {N1, N00})) return DAG.getNode(ISD::SUB, DL, VT, Add, N0.getOperand(1)); } // add (sext i1 X), 1 -> zext (not i1 X) // We don't transform this pattern: // add (zext i1 X), -1 -> sext (not i1 X) // because most (?) targets generate better code for the zext form. if (N0.getOpcode() == ISD::SIGN_EXTEND && N0.hasOneUse() && isOneOrOneSplat(N1)) { SDValue X = N0.getOperand(0); if ((!LegalOperations || (TLI.isOperationLegal(ISD::XOR, X.getValueType()) && TLI.isOperationLegal(ISD::ZERO_EXTEND, VT))) && X.getScalarValueSizeInBits() == 1) { SDValue Not = DAG.getNOT(DL, X, X.getValueType()); return DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Not); } } // Fold (add (or x, c0), c1) -> (add x, (c0 + c1)) // iff (or x, c0) is equivalent to (add x, c0). // Fold (add (xor x, c0), c1) -> (add x, (c0 + c1)) // iff (xor x, c0) is equivalent to (add x, c0). if (isADDLike(N0, DAG)) { SDValue N01 = N0.getOperand(1); if (SDValue Add = DAG.FoldConstantArithmetic(ISD::ADD, DL, VT, {N1, N01})) return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), Add); } if (SDValue NewSel = foldBinOpIntoSelect(N)) return NewSel; // reassociate add if (!reassociationCanBreakAddressingModePattern(ISD::ADD, DL, N, N0, N1)) { if (SDValue RADD = reassociateOps(ISD::ADD, DL, N0, N1, N->getFlags())) return RADD; // Reassociate (add (or x, c), y) -> (add add(x, y), c)) if (or x, c) is // equivalent to (add x, c). // Reassociate (add (xor x, c), y) -> (add add(x, y), c)) if (xor x, c) is // equivalent to (add x, c). auto ReassociateAddOr = [&](SDValue N0, SDValue N1) { if (isADDLike(N0, DAG) && N0.hasOneUse() && isConstantOrConstantVector(N0.getOperand(1), /* NoOpaque */ true)) { return DAG.getNode(ISD::ADD, DL, VT, DAG.getNode(ISD::ADD, DL, VT, N1, N0.getOperand(0)), N0.getOperand(1)); } return SDValue(); }; if (SDValue Add = ReassociateAddOr(N0, N1)) return Add; if (SDValue Add = ReassociateAddOr(N1, N0)) return Add; } // fold ((0-A) + B) -> B-A if (N0.getOpcode() == ISD::SUB && isNullOrNullSplat(N0.getOperand(0))) return DAG.getNode(ISD::SUB, DL, VT, N1, N0.getOperand(1)); // fold (A + (0-B)) -> A-B if (N1.getOpcode() == ISD::SUB && isNullOrNullSplat(N1.getOperand(0))) return DAG.getNode(ISD::SUB, DL, VT, N0, N1.getOperand(1)); // fold (A+(B-A)) -> B if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1)) return N1.getOperand(0); // fold ((B-A)+A) -> B if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1)) return N0.getOperand(0); // fold ((A-B)+(C-A)) -> (C-B) if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB && N0.getOperand(0) == N1.getOperand(1)) return DAG.getNode(ISD::SUB, DL, VT, N1.getOperand(0), N0.getOperand(1)); // fold ((A-B)+(B-C)) -> (A-C) if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB && N0.getOperand(1) == N1.getOperand(0)) return DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), N1.getOperand(1)); // fold (A+(B-(A+C))) to (B-C) if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD && N0 == N1.getOperand(1).getOperand(0)) return DAG.getNode(ISD::SUB, DL, VT, N1.getOperand(0), N1.getOperand(1).getOperand(1)); // fold (A+(B-(C+A))) to (B-C) if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD && N0 == N1.getOperand(1).getOperand(1)) return DAG.getNode(ISD::SUB, DL, VT, N1.getOperand(0), N1.getOperand(1).getOperand(0)); // fold (A+((B-A)+or-C)) to (B+or-C) if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) && N1.getOperand(0).getOpcode() == ISD::SUB && N0 == N1.getOperand(0).getOperand(1)) return DAG.getNode(N1.getOpcode(), DL, VT, N1.getOperand(0).getOperand(0), N1.getOperand(1)); // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB && N0->hasOneUse() && N1->hasOneUse()) { SDValue N00 = N0.getOperand(0); SDValue N01 = N0.getOperand(1); SDValue N10 = N1.getOperand(0); SDValue N11 = N1.getOperand(1); if (isConstantOrConstantVector(N00) || isConstantOrConstantVector(N10)) return DAG.getNode(ISD::SUB, DL, VT, DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10), DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11)); } // fold (add (umax X, C), -C) --> (usubsat X, C) if (N0.getOpcode() == ISD::UMAX && hasOperation(ISD::USUBSAT, VT)) { auto MatchUSUBSAT = [](ConstantSDNode *Max, ConstantSDNode *Op) { return (!Max && !Op) || (Max && Op && Max->getAPIntValue() == (-Op->getAPIntValue())); }; if (ISD::matchBinaryPredicate(N0.getOperand(1), N1, MatchUSUBSAT, /*AllowUndefs*/ true)) return DAG.getNode(ISD::USUBSAT, DL, VT, N0.getOperand(0), N0.getOperand(1)); } if (SimplifyDemandedBits(SDValue(N, 0))) return SDValue(N, 0); if (isOneOrOneSplat(N1)) { // fold (add (xor a, -1), 1) -> (sub 0, a) if (isBitwiseNot(N0)) return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), N0.getOperand(0)); // fold (add (add (xor a, -1), b), 1) -> (sub b, a) if (N0.getOpcode() == ISD::ADD) { SDValue A, Xor; if (isBitwiseNot(N0.getOperand(0))) { A = N0.getOperand(1); Xor = N0.getOperand(0); } else if (isBitwiseNot(N0.getOperand(1))) { A = N0.getOperand(0); Xor = N0.getOperand(1); } if (Xor) return DAG.getNode(ISD::SUB, DL, VT, A, Xor.getOperand(0)); } // Look for: // add (add x, y), 1 // And if the target does not like this form then turn into: // sub y, (xor x, -1) if (!TLI.preferIncOfAddToSubOfNot(VT) && N0.getOpcode() == ISD::ADD && N0.hasOneUse()) { SDValue Not = DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(0), DAG.getAllOnesConstant(DL, VT)); return DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(1), Not); } } // (x - y) + -1 -> add (xor y, -1), x if (N0.getOpcode() == ISD::SUB && N0.hasOneUse() && isAllOnesOrAllOnesSplat(N1)) { SDValue Xor = DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(1), N1); return DAG.getNode(ISD::ADD, DL, VT, Xor, N0.getOperand(0)); } if (SDValue Combined = visitADDLikeCommutative(N0, N1, N)) return Combined; if (SDValue Combined = visitADDLikeCommutative(N1, N0, N)) return Combined; return SDValue(); } SDValue DAGCombiner::visitADD(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N0.getValueType(); SDLoc DL(N); if (SDValue Combined = visitADDLike(N)) return Combined; if (SDValue V = foldAddSubBoolOfMaskedVal(N, DAG)) return V; if (SDValue V = foldAddSubOfSignBit(N, DAG)) return V; // fold (a+b) -> (a|b) iff a and b share no bits. if ((!LegalOperations || TLI.isOperationLegal(ISD::OR, VT)) && DAG.haveNoCommonBitsSet(N0, N1)) return DAG.getNode(ISD::OR, DL, VT, N0, N1); // Fold (add (vscale * C0), (vscale * C1)) to (vscale * (C0 + C1)). if (N0.getOpcode() == ISD::VSCALE && N1.getOpcode() == ISD::VSCALE) { const APInt &C0 = N0->getConstantOperandAPInt(0); const APInt &C1 = N1->getConstantOperandAPInt(0); return DAG.getVScale(DL, VT, C0 + C1); } // fold a+vscale(c1)+vscale(c2) -> a+vscale(c1+c2) if ((N0.getOpcode() == ISD::ADD) && (N0.getOperand(1).getOpcode() == ISD::VSCALE) && (N1.getOpcode() == ISD::VSCALE)) { const APInt &VS0 = N0.getOperand(1)->getConstantOperandAPInt(0); const APInt &VS1 = N1->getConstantOperandAPInt(0); SDValue VS = DAG.getVScale(DL, VT, VS0 + VS1); return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), VS); } // Fold (add step_vector(c1), step_vector(c2) to step_vector(c1+c2)) if (N0.getOpcode() == ISD::STEP_VECTOR && N1.getOpcode() == ISD::STEP_VECTOR) { const APInt &C0 = N0->getConstantOperandAPInt(0); const APInt &C1 = N1->getConstantOperandAPInt(0); APInt NewStep = C0 + C1; return DAG.getStepVector(DL, VT, NewStep); } // Fold a + step_vector(c1) + step_vector(c2) to a + step_vector(c1+c2) if ((N0.getOpcode() == ISD::ADD) && (N0.getOperand(1).getOpcode() == ISD::STEP_VECTOR) && (N1.getOpcode() == ISD::STEP_VECTOR)) { const APInt &SV0 = N0.getOperand(1)->getConstantOperandAPInt(0); const APInt &SV1 = N1->getConstantOperandAPInt(0); APInt NewStep = SV0 + SV1; SDValue SV = DAG.getStepVector(DL, VT, NewStep); return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), SV); } return SDValue(); } SDValue DAGCombiner::visitADDSAT(SDNode *N) { unsigned Opcode = N->getOpcode(); SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N0.getValueType(); SDLoc DL(N); // fold (add_sat x, undef) -> -1 if (N0.isUndef() || N1.isUndef()) return DAG.getAllOnesConstant(DL, VT); // fold (add_sat c1, c2) -> c3 if (SDValue C = DAG.FoldConstantArithmetic(Opcode, DL, VT, {N0, N1})) return C; // canonicalize constant to RHS if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && !DAG.isConstantIntBuildVectorOrConstantInt(N1)) return DAG.getNode(Opcode, DL, VT, N1, N0); // fold vector ops if (VT.isVector()) { if (SDValue FoldedVOp = SimplifyVBinOp(N, DL)) return FoldedVOp; // fold (add_sat x, 0) -> x, vector edition if (ISD::isConstantSplatVectorAllZeros(N1.getNode())) return N0; } // fold (add_sat x, 0) -> x if (isNullConstant(N1)) return N0; // If it cannot overflow, transform into an add. if (Opcode == ISD::UADDSAT) if (DAG.computeOverflowKind(N0, N1) == SelectionDAG::OFK_Never) return DAG.getNode(ISD::ADD, DL, VT, N0, N1); return SDValue(); } static SDValue getAsCarry(const TargetLowering &TLI, SDValue V) { bool Masked = false; // First, peel away TRUNCATE/ZERO_EXTEND/AND nodes due to legalization. while (true) { if (V.getOpcode() == ISD::TRUNCATE || V.getOpcode() == ISD::ZERO_EXTEND) { V = V.getOperand(0); continue; } if (V.getOpcode() == ISD::AND && isOneConstant(V.getOperand(1))) { Masked = true; V = V.getOperand(0); continue; } break; } // If this is not a carry, return. if (V.getResNo() != 1) return SDValue(); if (V.getOpcode() != ISD::ADDCARRY && V.getOpcode() != ISD::SUBCARRY && V.getOpcode() != ISD::UADDO && V.getOpcode() != ISD::USUBO) return SDValue(); EVT VT = V->getValueType(0); if (!TLI.isOperationLegalOrCustom(V.getOpcode(), VT)) return SDValue(); // If the result is masked, then no matter what kind of bool it is we can // return. If it isn't, then we need to make sure the bool type is either 0 or // 1 and not other values. if (Masked || TLI.getBooleanContents(V.getValueType()) == TargetLoweringBase::ZeroOrOneBooleanContent) return V; return SDValue(); } /// Given the operands of an add/sub operation, see if the 2nd operand is a /// masked 0/1 whose source operand is actually known to be 0/-1. If so, invert /// the opcode and bypass the mask operation. static SDValue foldAddSubMasked1(bool IsAdd, SDValue N0, SDValue N1, SelectionDAG &DAG, const SDLoc &DL) { if (N1.getOpcode() != ISD::AND || !isOneOrOneSplat(N1->getOperand(1))) return SDValue(); EVT VT = N0.getValueType(); if (DAG.ComputeNumSignBits(N1.getOperand(0)) != VT.getScalarSizeInBits()) return SDValue(); // add N0, (and (AssertSext X, i1), 1) --> sub N0, X // sub N0, (and (AssertSext X, i1), 1) --> add N0, X return DAG.getNode(IsAdd ? ISD::SUB : ISD::ADD, DL, VT, N0, N1.getOperand(0)); } /// Helper for doing combines based on N0 and N1 being added to each other. SDValue DAGCombiner::visitADDLikeCommutative(SDValue N0, SDValue N1, SDNode *LocReference) { EVT VT = N0.getValueType(); SDLoc DL(LocReference); // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n)) if (N1.getOpcode() == ISD::SHL && N1.getOperand(0).getOpcode() == ISD::SUB && isNullOrNullSplat(N1.getOperand(0).getOperand(0))) return DAG.getNode(ISD::SUB, DL, VT, N0, DAG.getNode(ISD::SHL, DL, VT, N1.getOperand(0).getOperand(1), N1.getOperand(1))); if (SDValue V = foldAddSubMasked1(true, N0, N1, DAG, DL)) return V; // Look for: // add (add x, 1), y // And if the target does not like this form then turn into: // sub y, (xor x, -1) if (!TLI.preferIncOfAddToSubOfNot(VT) && N0.getOpcode() == ISD::ADD && N0.hasOneUse() && isOneOrOneSplat(N0.getOperand(1))) { SDValue Not = DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(0), DAG.getAllOnesConstant(DL, VT)); return DAG.getNode(ISD::SUB, DL, VT, N1, Not); } if (N0.getOpcode() == ISD::SUB && N0.hasOneUse()) { // Hoist one-use subtraction by non-opaque constant: // (x - C) + y -> (x + y) - C // This is necessary because SUB(X,C) -> ADD(X,-C) doesn't work for vectors. if (isConstantOrConstantVector(N0.getOperand(1), /*NoOpaques=*/true)) { SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), N1); return DAG.getNode(ISD::SUB, DL, VT, Add, N0.getOperand(1)); } // Hoist one-use subtraction from non-opaque constant: // (C - x) + y -> (y - x) + C if (isConstantOrConstantVector(N0.getOperand(0), /*NoOpaques=*/true)) { SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N1, N0.getOperand(1)); return DAG.getNode(ISD::ADD, DL, VT, Sub, N0.getOperand(0)); } } // If the target's bool is represented as 0/1, prefer to make this 'sub 0/1' // rather than 'add 0/-1' (the zext should get folded). // add (sext i1 Y), X --> sub X, (zext i1 Y) if (N0.getOpcode() == ISD::SIGN_EXTEND && N0.getOperand(0).getScalarValueSizeInBits() == 1 && TLI.getBooleanContents(VT) == TargetLowering::ZeroOrOneBooleanContent) { SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)); return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt); } // add X, (sextinreg Y i1) -> sub X, (and Y 1) if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) { VTSDNode *TN = cast(N1.getOperand(1)); if (TN->getVT() == MVT::i1) { SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0), DAG.getConstant(1, DL, VT)); return DAG.getNode(ISD::SUB, DL, VT, N0, ZExt); } } // (add X, (addcarry Y, 0, Carry)) -> (addcarry X, Y, Carry) if (N1.getOpcode() == ISD::ADDCARRY && isNullConstant(N1.getOperand(1)) && N1.getResNo() == 0) return DAG.getNode(ISD::ADDCARRY, DL, N1->getVTList(), N0, N1.getOperand(0), N1.getOperand(2)); // (add X, Carry) -> (addcarry X, 0, Carry) if (TLI.isOperationLegalOrCustom(ISD::ADDCARRY, VT)) if (SDValue Carry = getAsCarry(TLI, N1)) return DAG.getNode(ISD::ADDCARRY, DL, DAG.getVTList(VT, Carry.getValueType()), N0, DAG.getConstant(0, DL, VT), Carry); return SDValue(); } SDValue DAGCombiner::visitADDC(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N0.getValueType(); SDLoc DL(N); // If the flag result is dead, turn this into an ADD. if (!N->hasAnyUseOfValue(1)) return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1), DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); // canonicalize constant to RHS. ConstantSDNode *N0C = dyn_cast(N0); ConstantSDNode *N1C = dyn_cast(N1); if (N0C && !N1C) return DAG.getNode(ISD::ADDC, DL, N->getVTList(), N1, N0); // fold (addc x, 0) -> x + no carry out if (isNullConstant(N1)) return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); // If it cannot overflow, transform into an add. if (DAG.computeOverflowKind(N0, N1) == SelectionDAG::OFK_Never) return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1), DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); return SDValue(); } /** * Flips a boolean if it is cheaper to compute. If the Force parameters is set, * then the flip also occurs if computing the inverse is the same cost. * This function returns an empty SDValue in case it cannot flip the boolean * without increasing the cost of the computation. If you want to flip a boolean * no matter what, use DAG.getLogicalNOT. */ static SDValue extractBooleanFlip(SDValue V, SelectionDAG &DAG, const TargetLowering &TLI, bool Force) { if (Force && isa(V)) return DAG.getLogicalNOT(SDLoc(V), V, V.getValueType()); if (V.getOpcode() != ISD::XOR) return SDValue(); ConstantSDNode *Const = isConstOrConstSplat(V.getOperand(1), false); if (!Const) return SDValue(); EVT VT = V.getValueType(); bool IsFlip = false; switch(TLI.getBooleanContents(VT)) { case TargetLowering::ZeroOrOneBooleanContent: IsFlip = Const->isOne(); break; case TargetLowering::ZeroOrNegativeOneBooleanContent: IsFlip = Const->isAllOnes(); break; case TargetLowering::UndefinedBooleanContent: IsFlip = (Const->getAPIntValue() & 0x01) == 1; break; } if (IsFlip) return V.getOperand(0); if (Force) return DAG.getLogicalNOT(SDLoc(V), V, V.getValueType()); return SDValue(); } SDValue DAGCombiner::visitADDO(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N0.getValueType(); bool IsSigned = (ISD::SADDO == N->getOpcode()); EVT CarryVT = N->getValueType(1); SDLoc DL(N); // If the flag result is dead, turn this into an ADD. if (!N->hasAnyUseOfValue(1)) return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1), DAG.getUNDEF(CarryVT)); // canonicalize constant to RHS. if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && !DAG.isConstantIntBuildVectorOrConstantInt(N1)) return DAG.getNode(N->getOpcode(), DL, N->getVTList(), N1, N0); // fold (addo x, 0) -> x + no carry out if (isNullOrNullSplat(N1)) return CombineTo(N, N0, DAG.getConstant(0, DL, CarryVT)); if (!IsSigned) { // If it cannot overflow, transform into an add. if (DAG.computeOverflowKind(N0, N1) == SelectionDAG::OFK_Never) return CombineTo(N, DAG.getNode(ISD::ADD, DL, VT, N0, N1), DAG.getConstant(0, DL, CarryVT)); // fold (uaddo (xor a, -1), 1) -> (usub 0, a) and flip carry. if (isBitwiseNot(N0) && isOneOrOneSplat(N1)) { SDValue Sub = DAG.getNode(ISD::USUBO, DL, N->getVTList(), DAG.getConstant(0, DL, VT), N0.getOperand(0)); return CombineTo( N, Sub, DAG.getLogicalNOT(DL, Sub.getValue(1), Sub->getValueType(1))); } if (SDValue Combined = visitUADDOLike(N0, N1, N)) return Combined; if (SDValue Combined = visitUADDOLike(N1, N0, N)) return Combined; } return SDValue(); } SDValue DAGCombiner::visitUADDOLike(SDValue N0, SDValue N1, SDNode *N) { EVT VT = N0.getValueType(); if (VT.isVector()) return SDValue(); // (uaddo X, (addcarry Y, 0, Carry)) -> (addcarry X, Y, Carry) // If Y + 1 cannot overflow. if (N1.getOpcode() == ISD::ADDCARRY && isNullConstant(N1.getOperand(1))) { SDValue Y = N1.getOperand(0); SDValue One = DAG.getConstant(1, SDLoc(N), Y.getValueType()); if (DAG.computeOverflowKind(Y, One) == SelectionDAG::OFK_Never) return DAG.getNode(ISD::ADDCARRY, SDLoc(N), N->getVTList(), N0, Y, N1.getOperand(2)); } // (uaddo X, Carry) -> (addcarry X, 0, Carry) if (TLI.isOperationLegalOrCustom(ISD::ADDCARRY, VT)) if (SDValue Carry = getAsCarry(TLI, N1)) return DAG.getNode(ISD::ADDCARRY, SDLoc(N), N->getVTList(), N0, DAG.getConstant(0, SDLoc(N), VT), Carry); return SDValue(); } SDValue DAGCombiner::visitADDE(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); SDValue CarryIn = N->getOperand(2); // canonicalize constant to RHS ConstantSDNode *N0C = dyn_cast(N0); ConstantSDNode *N1C = dyn_cast(N1); if (N0C && !N1C) return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(), N1, N0, CarryIn); // fold (adde x, y, false) -> (addc x, y) if (CarryIn.getOpcode() == ISD::CARRY_FALSE) return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1); return SDValue(); } SDValue DAGCombiner::visitADDCARRY(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); SDValue CarryIn = N->getOperand(2); SDLoc DL(N); // canonicalize constant to RHS ConstantSDNode *N0C = dyn_cast(N0); ConstantSDNode *N1C = dyn_cast(N1); if (N0C && !N1C) return DAG.getNode(ISD::ADDCARRY, DL, N->getVTList(), N1, N0, CarryIn); // fold (addcarry x, y, false) -> (uaddo x, y) if (isNullConstant(CarryIn)) { if (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::UADDO, N->getValueType(0))) return DAG.getNode(ISD::UADDO, DL, N->getVTList(), N0, N1); } // fold (addcarry 0, 0, X) -> (and (ext/trunc X), 1) and no carry. if (isNullConstant(N0) && isNullConstant(N1)) { EVT VT = N0.getValueType(); EVT CarryVT = CarryIn.getValueType(); SDValue CarryExt = DAG.getBoolExtOrTrunc(CarryIn, DL, VT, CarryVT); AddToWorklist(CarryExt.getNode()); return CombineTo(N, DAG.getNode(ISD::AND, DL, VT, CarryExt, DAG.getConstant(1, DL, VT)), DAG.getConstant(0, DL, CarryVT)); } if (SDValue Combined = visitADDCARRYLike(N0, N1, CarryIn, N)) return Combined; if (SDValue Combined = visitADDCARRYLike(N1, N0, CarryIn, N)) return Combined; return SDValue(); } SDValue DAGCombiner::visitSADDO_CARRY(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); SDValue CarryIn = N->getOperand(2); SDLoc DL(N); // canonicalize constant to RHS ConstantSDNode *N0C = dyn_cast(N0); ConstantSDNode *N1C = dyn_cast(N1); if (N0C && !N1C) return DAG.getNode(ISD::SADDO_CARRY, DL, N->getVTList(), N1, N0, CarryIn); // fold (saddo_carry x, y, false) -> (saddo x, y) if (isNullConstant(CarryIn)) { if (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::SADDO, N->getValueType(0))) return DAG.getNode(ISD::SADDO, DL, N->getVTList(), N0, N1); } return SDValue(); } /** * If we are facing some sort of diamond carry propapagtion pattern try to * break it up to generate something like: * (addcarry X, 0, (addcarry A, B, Z):Carry) * * The end result is usually an increase in operation required, but because the * carry is now linearized, other tranforms can kick in and optimize the DAG. * * Patterns typically look something like * (uaddo A, B) * / \ * Carry Sum * | \ * | (addcarry *, 0, Z) * | / * \ Carry * | / * (addcarry X, *, *) * * But numerous variation exist. Our goal is to identify A, B, X and Z and * produce a combine with a single path for carry propagation. */ static SDValue combineADDCARRYDiamond(DAGCombiner &Combiner, SelectionDAG &DAG, SDValue X, SDValue Carry0, SDValue Carry1, SDNode *N) { if (Carry1.getResNo() != 1 || Carry0.getResNo() != 1) return SDValue(); if (Carry1.getOpcode() != ISD::UADDO) return SDValue(); SDValue Z; /** * First look for a suitable Z. It will present itself in the form of * (addcarry Y, 0, Z) or its equivalent (uaddo Y, 1) for Z=true */ if (Carry0.getOpcode() == ISD::ADDCARRY && isNullConstant(Carry0.getOperand(1))) { Z = Carry0.getOperand(2); } else if (Carry0.getOpcode() == ISD::UADDO && isOneConstant(Carry0.getOperand(1))) { EVT VT = Combiner.getSetCCResultType(Carry0.getValueType()); Z = DAG.getConstant(1, SDLoc(Carry0.getOperand(1)), VT); } else { // We couldn't find a suitable Z. return SDValue(); } auto cancelDiamond = [&](SDValue A,SDValue B) { SDLoc DL(N); SDValue NewY = DAG.getNode(ISD::ADDCARRY, DL, Carry0->getVTList(), A, B, Z); Combiner.AddToWorklist(NewY.getNode()); return DAG.getNode(ISD::ADDCARRY, DL, N->getVTList(), X, DAG.getConstant(0, DL, X.getValueType()), NewY.getValue(1)); }; /** * (uaddo A, B) * | * Sum * | * (addcarry *, 0, Z) */ if (Carry0.getOperand(0) == Carry1.getValue(0)) { return cancelDiamond(Carry1.getOperand(0), Carry1.getOperand(1)); } /** * (addcarry A, 0, Z) * | * Sum * | * (uaddo *, B) */ if (Carry1.getOperand(0) == Carry0.getValue(0)) { return cancelDiamond(Carry0.getOperand(0), Carry1.getOperand(1)); } if (Carry1.getOperand(1) == Carry0.getValue(0)) { return cancelDiamond(Carry1.getOperand(0), Carry0.getOperand(0)); } return SDValue(); } // If we are facing some sort of diamond carry/borrow in/out pattern try to // match patterns like: // // (uaddo A, B) CarryIn // | \ | // | \ | // PartialSum PartialCarryOutX / // | | / // | ____|____________/ // | / | // (uaddo *, *) \________ // | \ \ // | \ | // | PartialCarryOutY | // | \ | // | \ / // AddCarrySum | ______/ // | / // CarryOut = (or *, *) // // And generate ADDCARRY (or SUBCARRY) with two result values: // // {AddCarrySum, CarryOut} = (addcarry A, B, CarryIn) // // Our goal is to identify A, B, and CarryIn and produce ADDCARRY/SUBCARRY with // a single path for carry/borrow out propagation: static SDValue combineCarryDiamond(SelectionDAG &DAG, const TargetLowering &TLI, SDValue N0, SDValue N1, SDNode *N) { SDValue Carry0 = getAsCarry(TLI, N0); if (!Carry0) return SDValue(); SDValue Carry1 = getAsCarry(TLI, N1); if (!Carry1) return SDValue(); unsigned Opcode = Carry0.getOpcode(); if (Opcode != Carry1.getOpcode()) return SDValue(); if (Opcode != ISD::UADDO && Opcode != ISD::USUBO) return SDValue(); // Canonicalize the add/sub of A and B (the top node in the above ASCII art) // as Carry0 and the add/sub of the carry in as Carry1 (the middle node). if (Carry1.getNode()->isOperandOf(Carry0.getNode())) std::swap(Carry0, Carry1); // Check if nodes are connected in expected way. if (Carry1.getOperand(0) != Carry0.getValue(0) && Carry1.getOperand(1) != Carry0.getValue(0)) return SDValue(); // The carry in value must be on the righthand side for subtraction. unsigned CarryInOperandNum = Carry1.getOperand(0) == Carry0.getValue(0) ? 1 : 0; if (Opcode == ISD::USUBO && CarryInOperandNum != 1) return SDValue(); SDValue CarryIn = Carry1.getOperand(CarryInOperandNum); unsigned NewOp = Opcode == ISD::UADDO ? ISD::ADDCARRY : ISD::SUBCARRY; if (!TLI.isOperationLegalOrCustom(NewOp, Carry0.getValue(0).getValueType())) return SDValue(); // Verify that the carry/borrow in is plausibly a carry/borrow bit. // TODO: make getAsCarry() aware of how partial carries are merged. if (CarryIn.getOpcode() != ISD::ZERO_EXTEND) return SDValue(); CarryIn = CarryIn.getOperand(0); if (CarryIn.getValueType() != MVT::i1) return SDValue(); SDLoc DL(N); SDValue Merged = DAG.getNode(NewOp, DL, Carry1->getVTList(), Carry0.getOperand(0), Carry0.getOperand(1), CarryIn); // Please note that because we have proven that the result of the UADDO/USUBO // of A and B feeds into the UADDO/USUBO that does the carry/borrow in, we can // therefore prove that if the first UADDO/USUBO overflows, the second // UADDO/USUBO cannot. For example consider 8-bit numbers where 0xFF is the // maximum value. // // 0xFF + 0xFF == 0xFE with carry but 0xFE + 1 does not carry // 0x00 - 0xFF == 1 with a carry/borrow but 1 - 1 == 0 (no carry/borrow) // // This is important because it means that OR and XOR can be used to merge // carry flags; and that AND can return a constant zero. // // TODO: match other operations that can merge flags (ADD, etc) DAG.ReplaceAllUsesOfValueWith(Carry1.getValue(0), Merged.getValue(0)); if (N->getOpcode() == ISD::AND) return DAG.getConstant(0, DL, MVT::i1); return Merged.getValue(1); } SDValue DAGCombiner::visitADDCARRYLike(SDValue N0, SDValue N1, SDValue CarryIn, SDNode *N) { // fold (addcarry (xor a, -1), b, c) -> (subcarry b, a, !c) and flip carry. if (isBitwiseNot(N0)) if (SDValue NotC = extractBooleanFlip(CarryIn, DAG, TLI, true)) { SDLoc DL(N); SDValue Sub = DAG.getNode(ISD::SUBCARRY, DL, N->getVTList(), N1, N0.getOperand(0), NotC); return CombineTo( N, Sub, DAG.getLogicalNOT(DL, Sub.getValue(1), Sub->getValueType(1))); } // Iff the flag result is dead: // (addcarry (add|uaddo X, Y), 0, Carry) -> (addcarry X, Y, Carry) // Don't do this if the Carry comes from the uaddo. It won't remove the uaddo // or the dependency between the instructions. if ((N0.getOpcode() == ISD::ADD || (N0.getOpcode() == ISD::UADDO && N0.getResNo() == 0 && N0.getValue(1) != CarryIn)) && isNullConstant(N1) && !N->hasAnyUseOfValue(1)) return DAG.getNode(ISD::ADDCARRY, SDLoc(N), N->getVTList(), N0.getOperand(0), N0.getOperand(1), CarryIn); /** * When one of the addcarry argument is itself a carry, we may be facing * a diamond carry propagation. In which case we try to transform the DAG * to ensure linear carry propagation if that is possible. */ if (auto Y = getAsCarry(TLI, N1)) { // Because both are carries, Y and Z can be swapped. if (auto R = combineADDCARRYDiamond(*this, DAG, N0, Y, CarryIn, N)) return R; if (auto R = combineADDCARRYDiamond(*this, DAG, N0, CarryIn, Y, N)) return R; } return SDValue(); } // Attempt to create a USUBSAT(LHS, RHS) node with DstVT, performing a // clamp/truncation if necessary. static SDValue getTruncatedUSUBSAT(EVT DstVT, EVT SrcVT, SDValue LHS, SDValue RHS, SelectionDAG &DAG, const SDLoc &DL) { assert(DstVT.getScalarSizeInBits() <= SrcVT.getScalarSizeInBits() && "Illegal truncation"); if (DstVT == SrcVT) return DAG.getNode(ISD::USUBSAT, DL, DstVT, LHS, RHS); // If the LHS is zero-extended then we can perform the USUBSAT as DstVT by // clamping RHS. APInt UpperBits = APInt::getBitsSetFrom(SrcVT.getScalarSizeInBits(), DstVT.getScalarSizeInBits()); if (!DAG.MaskedValueIsZero(LHS, UpperBits)) return SDValue(); SDValue SatLimit = DAG.getConstant(APInt::getLowBitsSet(SrcVT.getScalarSizeInBits(), DstVT.getScalarSizeInBits()), DL, SrcVT); RHS = DAG.getNode(ISD::UMIN, DL, SrcVT, RHS, SatLimit); RHS = DAG.getNode(ISD::TRUNCATE, DL, DstVT, RHS); LHS = DAG.getNode(ISD::TRUNCATE, DL, DstVT, LHS); return DAG.getNode(ISD::USUBSAT, DL, DstVT, LHS, RHS); } // Try to find umax(a,b) - b or a - umin(a,b) patterns that may be converted to // usubsat(a,b), optionally as a truncated type. SDValue DAGCombiner::foldSubToUSubSat(EVT DstVT, SDNode *N) { if (N->getOpcode() != ISD::SUB || !(!LegalOperations || hasOperation(ISD::USUBSAT, DstVT))) return SDValue(); EVT SubVT = N->getValueType(0); SDValue Op0 = N->getOperand(0); SDValue Op1 = N->getOperand(1); // Try to find umax(a,b) - b or a - umin(a,b) patterns // they may be converted to usubsat(a,b). if (Op0.getOpcode() == ISD::UMAX && Op0.hasOneUse()) { SDValue MaxLHS = Op0.getOperand(0); SDValue MaxRHS = Op0.getOperand(1); if (MaxLHS == Op1) return getTruncatedUSUBSAT(DstVT, SubVT, MaxRHS, Op1, DAG, SDLoc(N)); if (MaxRHS == Op1) return getTruncatedUSUBSAT(DstVT, SubVT, MaxLHS, Op1, DAG, SDLoc(N)); } if (Op1.getOpcode() == ISD::UMIN && Op1.hasOneUse()) { SDValue MinLHS = Op1.getOperand(0); SDValue MinRHS = Op1.getOperand(1); if (MinLHS == Op0) return getTruncatedUSUBSAT(DstVT, SubVT, Op0, MinRHS, DAG, SDLoc(N)); if (MinRHS == Op0) return getTruncatedUSUBSAT(DstVT, SubVT, Op0, MinLHS, DAG, SDLoc(N)); } // sub(a,trunc(umin(zext(a),b))) -> usubsat(a,trunc(umin(b,SatLimit))) if (Op1.getOpcode() == ISD::TRUNCATE && Op1.getOperand(0).getOpcode() == ISD::UMIN && Op1.getOperand(0).hasOneUse()) { SDValue MinLHS = Op1.getOperand(0).getOperand(0); SDValue MinRHS = Op1.getOperand(0).getOperand(1); if (MinLHS.getOpcode() == ISD::ZERO_EXTEND && MinLHS.getOperand(0) == Op0) return getTruncatedUSUBSAT(DstVT, MinLHS.getValueType(), MinLHS, MinRHS, DAG, SDLoc(N)); if (MinRHS.getOpcode() == ISD::ZERO_EXTEND && MinRHS.getOperand(0) == Op0) return getTruncatedUSUBSAT(DstVT, MinLHS.getValueType(), MinRHS, MinLHS, DAG, SDLoc(N)); } return SDValue(); } // Since it may not be valid to emit a fold to zero for vector initializers // check if we can before folding. static SDValue tryFoldToZero(const SDLoc &DL, const TargetLowering &TLI, EVT VT, SelectionDAG &DAG, bool LegalOperations) { if (!VT.isVector()) return DAG.getConstant(0, DL, VT); if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) return DAG.getConstant(0, DL, VT); return SDValue(); } SDValue DAGCombiner::visitSUB(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N0.getValueType(); SDLoc DL(N); auto PeekThroughFreeze = [](SDValue N) { if (N->getOpcode() == ISD::FREEZE && N.hasOneUse()) return N->getOperand(0); return N; }; // fold (sub x, x) -> 0 // FIXME: Refactor this and xor and other similar operations together. if (PeekThroughFreeze(N0) == PeekThroughFreeze(N1)) return tryFoldToZero(DL, TLI, VT, DAG, LegalOperations); // fold (sub c1, c2) -> c3 if (SDValue C = DAG.FoldConstantArithmetic(ISD::SUB, DL, VT, {N0, N1})) return C; // fold vector ops if (VT.isVector()) { if (SDValue FoldedVOp = SimplifyVBinOp(N, DL)) return FoldedVOp; // fold (sub x, 0) -> x, vector edition if (ISD::isConstantSplatVectorAllZeros(N1.getNode())) return N0; } if (SDValue NewSel = foldBinOpIntoSelect(N)) return NewSel; ConstantSDNode *N1C = getAsNonOpaqueConstant(N1); // fold (sub x, c) -> (add x, -c) if (N1C) { return DAG.getNode(ISD::ADD, DL, VT, N0, DAG.getConstant(-N1C->getAPIntValue(), DL, VT)); } if (isNullOrNullSplat(N0)) { unsigned BitWidth = VT.getScalarSizeInBits(); // Right-shifting everything out but the sign bit followed by negation is // the same as flipping arithmetic/logical shift type without the negation: // -(X >>u 31) -> (X >>s 31) // -(X >>s 31) -> (X >>u 31) if (N1->getOpcode() == ISD::SRA || N1->getOpcode() == ISD::SRL) { ConstantSDNode *ShiftAmt = isConstOrConstSplat(N1.getOperand(1)); if (ShiftAmt && ShiftAmt->getAPIntValue() == (BitWidth - 1)) { auto NewSh = N1->getOpcode() == ISD::SRA ? ISD::SRL : ISD::SRA; if (!LegalOperations || TLI.isOperationLegal(NewSh, VT)) return DAG.getNode(NewSh, DL, VT, N1.getOperand(0), N1.getOperand(1)); } } // 0 - X --> 0 if the sub is NUW. if (N->getFlags().hasNoUnsignedWrap()) return N0; if (DAG.MaskedValueIsZero(N1, ~APInt::getSignMask(BitWidth))) { // N1 is either 0 or the minimum signed value. If the sub is NSW, then // N1 must be 0 because negating the minimum signed value is undefined. if (N->getFlags().hasNoSignedWrap()) return N0; // 0 - X --> X if X is 0 or the minimum signed value. return N1; } // Convert 0 - abs(x). if (N1.getOpcode() == ISD::ABS && N1.hasOneUse() && !TLI.isOperationLegalOrCustom(ISD::ABS, VT)) if (SDValue Result = TLI.expandABS(N1.getNode(), DAG, true)) return Result; // Fold neg(splat(neg(x)) -> splat(x) if (VT.isVector()) { SDValue N1S = DAG.getSplatValue(N1, true); if (N1S && N1S.getOpcode() == ISD::SUB && isNullConstant(N1S.getOperand(0))) { if (VT.isScalableVector()) return DAG.getSplatVector(VT, DL, N1S.getOperand(1)); return DAG.getSplatBuildVector(VT, DL, N1S.getOperand(1)); } } } // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) if (isAllOnesOrAllOnesSplat(N0)) return DAG.getNode(ISD::XOR, DL, VT, N1, N0); // fold (A - (0-B)) -> A+B if (N1.getOpcode() == ISD::SUB && isNullOrNullSplat(N1.getOperand(0))) return DAG.getNode(ISD::ADD, DL, VT, N0, N1.getOperand(1)); // fold A-(A-B) -> B if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0)) return N1.getOperand(1); // fold (A+B)-A -> B if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1) return N0.getOperand(1); // fold (A+B)-B -> A if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1) return N0.getOperand(0); // fold (A+C1)-C2 -> A+(C1-C2) if (N0.getOpcode() == ISD::ADD) { SDValue N01 = N0.getOperand(1); if (SDValue NewC = DAG.FoldConstantArithmetic(ISD::SUB, DL, VT, {N01, N1})) return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), NewC); } // fold C2-(A+C1) -> (C2-C1)-A if (N1.getOpcode() == ISD::ADD) { SDValue N11 = N1.getOperand(1); if (SDValue NewC = DAG.FoldConstantArithmetic(ISD::SUB, DL, VT, {N0, N11})) return DAG.getNode(ISD::SUB, DL, VT, NewC, N1.getOperand(0)); } // fold (A-C1)-C2 -> A-(C1+C2) if (N0.getOpcode() == ISD::SUB) { SDValue N01 = N0.getOperand(1); if (SDValue NewC = DAG.FoldConstantArithmetic(ISD::ADD, DL, VT, {N01, N1})) return DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), NewC); } // fold (c1-A)-c2 -> (c1-c2)-A if (N0.getOpcode() == ISD::SUB) { SDValue N00 = N0.getOperand(0); if (SDValue NewC = DAG.FoldConstantArithmetic(ISD::SUB, DL, VT, {N00, N1})) return DAG.getNode(ISD::SUB, DL, VT, NewC, N0.getOperand(1)); } // fold ((A+(B+or-C))-B) -> A+or-C if (N0.getOpcode() == ISD::ADD && (N0.getOperand(1).getOpcode() == ISD::SUB || N0.getOperand(1).getOpcode() == ISD::ADD) && N0.getOperand(1).getOperand(0) == N1) return DAG.getNode(N0.getOperand(1).getOpcode(), DL, VT, N0.getOperand(0), N0.getOperand(1).getOperand(1)); // fold ((A+(C+B))-B) -> A+C if (N0.getOpcode() == ISD::ADD && N0.getOperand(1).getOpcode() == ISD::ADD && N0.getOperand(1).getOperand(1) == N1) return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), N0.getOperand(1).getOperand(0)); // fold ((A-(B-C))-C) -> A-B if (N0.getOpcode() == ISD::SUB && N0.getOperand(1).getOpcode() == ISD::SUB && N0.getOperand(1).getOperand(1) == N1) return DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), N0.getOperand(1).getOperand(0)); // fold (A-(B-C)) -> A+(C-B) if (N1.getOpcode() == ISD::SUB && N1.hasOneUse()) return DAG.getNode(ISD::ADD, DL, VT, N0, DAG.getNode(ISD::SUB, DL, VT, N1.getOperand(1), N1.getOperand(0))); // A - (A & B) -> A & (~B) if (N1.getOpcode() == ISD::AND) { SDValue A = N1.getOperand(0); SDValue B = N1.getOperand(1); if (A != N0) std::swap(A, B); if (A == N0 && (N1.hasOneUse() || isConstantOrConstantVector(B, /*NoOpaques=*/true))) { SDValue InvB = DAG.getNode(ISD::XOR, DL, VT, B, DAG.getAllOnesConstant(DL, VT)); return DAG.getNode(ISD::AND, DL, VT, A, InvB); } } // fold (X - (-Y * Z)) -> (X + (Y * Z)) if (N1.getOpcode() == ISD::MUL && N1.hasOneUse()) { if (N1.getOperand(0).getOpcode() == ISD::SUB && isNullOrNullSplat(N1.getOperand(0).getOperand(0))) { SDValue Mul = DAG.getNode(ISD::MUL, DL, VT, N1.getOperand(0).getOperand(1), N1.getOperand(1)); return DAG.getNode(ISD::ADD, DL, VT, N0, Mul); } if (N1.getOperand(1).getOpcode() == ISD::SUB && isNullOrNullSplat(N1.getOperand(1).getOperand(0))) { SDValue Mul = DAG.getNode(ISD::MUL, DL, VT, N1.getOperand(0), N1.getOperand(1).getOperand(1)); return DAG.getNode(ISD::ADD, DL, VT, N0, Mul); } } // If either operand of a sub is undef, the result is undef if (N0.isUndef()) return N0; if (N1.isUndef()) return N1; if (SDValue V = foldAddSubBoolOfMaskedVal(N, DAG)) return V; if (SDValue V = foldAddSubOfSignBit(N, DAG)) return V; if (SDValue V = foldAddSubMasked1(false, N0, N1, DAG, SDLoc(N))) return V; if (SDValue V = foldSubToUSubSat(VT, N)) return V; // (x - y) - 1 -> add (xor y, -1), x if (N0.hasOneUse() && N0.getOpcode() == ISD::SUB && isOneOrOneSplat(N1)) { SDValue Xor = DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(1), DAG.getAllOnesConstant(DL, VT)); return DAG.getNode(ISD::ADD, DL, VT, Xor, N0.getOperand(0)); } // Look for: // sub y, (xor x, -1) // And if the target does not like this form then turn into: // add (add x, y), 1 if (TLI.preferIncOfAddToSubOfNot(VT) && N1.hasOneUse() && isBitwiseNot(N1)) { SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, N1.getOperand(0)); return DAG.getNode(ISD::ADD, DL, VT, Add, DAG.getConstant(1, DL, VT)); } // Hoist one-use addition by non-opaque constant: // (x + C) - y -> (x - y) + C if (N0.hasOneUse() && N0.getOpcode() == ISD::ADD && isConstantOrConstantVector(N0.getOperand(1), /*NoOpaques=*/true)) { SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), N1); return DAG.getNode(ISD::ADD, DL, VT, Sub, N0.getOperand(1)); } // y - (x + C) -> (y - x) - C if (N1.hasOneUse() && N1.getOpcode() == ISD::ADD && isConstantOrConstantVector(N1.getOperand(1), /*NoOpaques=*/true)) { SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0, N1.getOperand(0)); return DAG.getNode(ISD::SUB, DL, VT, Sub, N1.getOperand(1)); } // (x - C) - y -> (x - y) - C // This is necessary because SUB(X,C) -> ADD(X,-C) doesn't work for vectors. if (N0.hasOneUse() && N0.getOpcode() == ISD::SUB && isConstantOrConstantVector(N0.getOperand(1), /*NoOpaques=*/true)) { SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), N1); return DAG.getNode(ISD::SUB, DL, VT, Sub, N0.getOperand(1)); } // (C - x) - y -> C - (x + y) if (N0.hasOneUse() && N0.getOpcode() == ISD::SUB && isConstantOrConstantVector(N0.getOperand(0), /*NoOpaques=*/true)) { SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(1), N1); return DAG.getNode(ISD::SUB, DL, VT, N0.getOperand(0), Add); } // If the target's bool is represented as 0/-1, prefer to make this 'add 0/-1' // rather than 'sub 0/1' (the sext should get folded). // sub X, (zext i1 Y) --> add X, (sext i1 Y) if (N1.getOpcode() == ISD::ZERO_EXTEND && N1.getOperand(0).getScalarValueSizeInBits() == 1 && TLI.getBooleanContents(VT) == TargetLowering::ZeroOrNegativeOneBooleanContent) { SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, VT, N1.getOperand(0)); return DAG.getNode(ISD::ADD, DL, VT, N0, SExt); } // fold Y = sra (X, size(X)-1); sub (xor (X, Y), Y) -> (abs X) if (TLI.isOperationLegalOrCustom(ISD::ABS, VT)) { if (N0.getOpcode() == ISD::XOR && N1.getOpcode() == ISD::SRA) { SDValue X0 = N0.getOperand(0), X1 = N0.getOperand(1); SDValue S0 = N1.getOperand(0); if ((X0 == S0 && X1 == N1) || (X0 == N1 && X1 == S0)) if (ConstantSDNode *C = isConstOrConstSplat(N1.getOperand(1))) if (C->getAPIntValue() == (VT.getScalarSizeInBits() - 1)) return DAG.getNode(ISD::ABS, SDLoc(N), VT, S0); } } // If the relocation model supports it, consider symbol offsets. if (GlobalAddressSDNode *GA = dyn_cast(N0)) if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) { // fold (sub Sym, c) -> Sym-c if (N1C && GA->getOpcode() == ISD::GlobalAddress) return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT, GA->getOffset() - (uint64_t)N1C->getSExtValue()); // fold (sub Sym+c1, Sym+c2) -> c1-c2 if (GlobalAddressSDNode *GB = dyn_cast(N1)) if (GA->getGlobal() == GB->getGlobal()) return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(), DL, VT); } // sub X, (sextinreg Y i1) -> add X, (and Y 1) if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) { VTSDNode *TN = cast(N1.getOperand(1)); if (TN->getVT() == MVT::i1) { SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0), DAG.getConstant(1, DL, VT)); return DAG.getNode(ISD::ADD, DL, VT, N0, ZExt); } } // canonicalize (sub X, (vscale * C)) to (add X, (vscale * -C)) if (N1.getOpcode() == ISD::VSCALE) { const APInt &IntVal = N1.getConstantOperandAPInt(0); return DAG.getNode(ISD::ADD, DL, VT, N0, DAG.getVScale(DL, VT, -IntVal)); } // canonicalize (sub X, step_vector(C)) to (add X, step_vector(-C)) if (N1.getOpcode() == ISD::STEP_VECTOR && N1.hasOneUse()) { APInt NewStep = -N1.getConstantOperandAPInt(0); return DAG.getNode(ISD::ADD, DL, VT, N0, DAG.getStepVector(DL, VT, NewStep)); } // Prefer an add for more folding potential and possibly better codegen: // sub N0, (lshr N10, width-1) --> add N0, (ashr N10, width-1) if (!LegalOperations && N1.getOpcode() == ISD::SRL && N1.hasOneUse()) { SDValue ShAmt = N1.getOperand(1); ConstantSDNode *ShAmtC = isConstOrConstSplat(ShAmt); if (ShAmtC && ShAmtC->getAPIntValue() == (N1.getScalarValueSizeInBits() - 1)) { SDValue SRA = DAG.getNode(ISD::SRA, DL, VT, N1.getOperand(0), ShAmt); return DAG.getNode(ISD::ADD, DL, VT, N0, SRA); } } // As with the previous fold, prefer add for more folding potential. // Subtracting SMIN/0 is the same as adding SMIN/0: // N0 - (X << BW-1) --> N0 + (X << BW-1) if (N1.getOpcode() == ISD::SHL) { ConstantSDNode *ShlC = isConstOrConstSplat(N1.getOperand(1)); if (ShlC && ShlC->getAPIntValue() == VT.getScalarSizeInBits() - 1) return DAG.getNode(ISD::ADD, DL, VT, N1, N0); } if (TLI.isOperationLegalOrCustom(ISD::ADDCARRY, VT)) { // (sub Carry, X) -> (addcarry (sub 0, X), 0, Carry) if (SDValue Carry = getAsCarry(TLI, N0)) { SDValue X = N1; SDValue Zero = DAG.getConstant(0, DL, VT); SDValue NegX = DAG.getNode(ISD::SUB, DL, VT, Zero, X); return DAG.getNode(ISD::ADDCARRY, DL, DAG.getVTList(VT, Carry.getValueType()), NegX, Zero, Carry); } } // If there's no chance of borrowing from adjacent bits, then sub is xor: // sub C0, X --> xor X, C0 if (ConstantSDNode *C0 = isConstOrConstSplat(N0)) { if (!C0->isOpaque()) { const APInt &C0Val = C0->getAPIntValue(); const APInt &MaybeOnes = ~DAG.computeKnownBits(N1).Zero; if ((C0Val - MaybeOnes) == (C0Val ^ MaybeOnes)) return DAG.getNode(ISD::XOR, DL, VT, N1, N0); } } return SDValue(); } SDValue DAGCombiner::visitSUBSAT(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N0.getValueType(); SDLoc DL(N); // fold (sub_sat x, undef) -> 0 if (N0.isUndef() || N1.isUndef()) return DAG.getConstant(0, DL, VT); // fold (sub_sat x, x) -> 0 if (N0 == N1) return DAG.getConstant(0, DL, VT); // fold (sub_sat c1, c2) -> c3 if (SDValue C = DAG.FoldConstantArithmetic(N->getOpcode(), DL, VT, {N0, N1})) return C; // fold vector ops if (VT.isVector()) { if (SDValue FoldedVOp = SimplifyVBinOp(N, DL)) return FoldedVOp; // fold (sub_sat x, 0) -> x, vector edition if (ISD::isConstantSplatVectorAllZeros(N1.getNode())) return N0; } // fold (sub_sat x, 0) -> x if (isNullConstant(N1)) return N0; return SDValue(); } SDValue DAGCombiner::visitSUBC(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N0.getValueType(); SDLoc DL(N); // If the flag result is dead, turn this into an SUB. if (!N->hasAnyUseOfValue(1)) return CombineTo(N, DAG.getNode(ISD::SUB, DL, VT, N0, N1), DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); // fold (subc x, x) -> 0 + no borrow if (N0 == N1) return CombineTo(N, DAG.getConstant(0, DL, VT), DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); // fold (subc x, 0) -> x + no borrow if (isNullConstant(N1)) return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow if (isAllOnesConstant(N0)) return CombineTo(N, DAG.getNode(ISD::XOR, DL, VT, N1, N0), DAG.getNode(ISD::CARRY_FALSE, DL, MVT::Glue)); return SDValue(); } SDValue DAGCombiner::visitSUBO(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N0.getValueType(); bool IsSigned = (ISD::SSUBO == N->getOpcode()); EVT CarryVT = N->getValueType(1); SDLoc DL(N); // If the flag result is dead, turn this into an SUB. if (!N->hasAnyUseOfValue(1)) return CombineTo(N, DAG.getNode(ISD::SUB, DL, VT, N0, N1), DAG.getUNDEF(CarryVT)); // fold (subo x, x) -> 0 + no borrow if (N0 == N1) return CombineTo(N, DAG.getConstant(0, DL, VT), DAG.getConstant(0, DL, CarryVT)); ConstantSDNode *N1C = getAsNonOpaqueConstant(N1); // fold (subox, c) -> (addo x, -c) if (IsSigned && N1C && !N1C->getAPIntValue().isMinSignedValue()) { return DAG.getNode(ISD::SADDO, DL, N->getVTList(), N0, DAG.getConstant(-N1C->getAPIntValue(), DL, VT)); } // fold (subo x, 0) -> x + no borrow if (isNullOrNullSplat(N1)) return CombineTo(N, N0, DAG.getConstant(0, DL, CarryVT)); // Canonicalize (usubo -1, x) -> ~x, i.e. (xor x, -1) + no borrow if (!IsSigned && isAllOnesOrAllOnesSplat(N0)) return CombineTo(N, DAG.getNode(ISD::XOR, DL, VT, N1, N0), DAG.getConstant(0, DL, CarryVT)); return SDValue(); } SDValue DAGCombiner::visitSUBE(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); SDValue CarryIn = N->getOperand(2); // fold (sube x, y, false) -> (subc x, y) if (CarryIn.getOpcode() == ISD::CARRY_FALSE) return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1); return SDValue(); } SDValue DAGCombiner::visitSUBCARRY(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); SDValue CarryIn = N->getOperand(2); // fold (subcarry x, y, false) -> (usubo x, y) if (isNullConstant(CarryIn)) { if (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::USUBO, N->getValueType(0))) return DAG.getNode(ISD::USUBO, SDLoc(N), N->getVTList(), N0, N1); } return SDValue(); } SDValue DAGCombiner::visitSSUBO_CARRY(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); SDValue CarryIn = N->getOperand(2); // fold (ssubo_carry x, y, false) -> (ssubo x, y) if (isNullConstant(CarryIn)) { if (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::SSUBO, N->getValueType(0))) return DAG.getNode(ISD::SSUBO, SDLoc(N), N->getVTList(), N0, N1); } return SDValue(); } // Notice that "mulfix" can be any of SMULFIX, SMULFIXSAT, UMULFIX and // UMULFIXSAT here. SDValue DAGCombiner::visitMULFIX(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); SDValue Scale = N->getOperand(2); EVT VT = N0.getValueType(); // fold (mulfix x, undef, scale) -> 0 if (N0.isUndef() || N1.isUndef()) return DAG.getConstant(0, SDLoc(N), VT); // Canonicalize constant to RHS (vector doesn't have to splat) if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && !DAG.isConstantIntBuildVectorOrConstantInt(N1)) return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N1, N0, Scale); // fold (mulfix x, 0, scale) -> 0 if (isNullConstant(N1)) return DAG.getConstant(0, SDLoc(N), VT); return SDValue(); } SDValue DAGCombiner::visitMUL(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N0.getValueType(); SDLoc DL(N); // fold (mul x, undef) -> 0 if (N0.isUndef() || N1.isUndef()) return DAG.getConstant(0, DL, VT); // fold (mul c1, c2) -> c1*c2 if (SDValue C = DAG.FoldConstantArithmetic(ISD::MUL, DL, VT, {N0, N1})) return C; // canonicalize constant to RHS (vector doesn't have to splat) if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && !DAG.isConstantIntBuildVectorOrConstantInt(N1)) return DAG.getNode(ISD::MUL, DL, VT, N1, N0); bool N1IsConst = false; bool N1IsOpaqueConst = false; APInt ConstValue1; // fold vector ops if (VT.isVector()) { if (SDValue FoldedVOp = SimplifyVBinOp(N, DL)) return FoldedVOp; N1IsConst = ISD::isConstantSplatVector(N1.getNode(), ConstValue1); assert((!N1IsConst || ConstValue1.getBitWidth() == VT.getScalarSizeInBits()) && "Splat APInt should be element width"); } else { N1IsConst = isa(N1); if (N1IsConst) { ConstValue1 = cast(N1)->getAPIntValue(); N1IsOpaqueConst = cast(N1)->isOpaque(); } } // fold (mul x, 0) -> 0 if (N1IsConst && ConstValue1.isZero()) return N1; // fold (mul x, 1) -> x if (N1IsConst && ConstValue1.isOne()) return N0; if (SDValue NewSel = foldBinOpIntoSelect(N)) return NewSel; // fold (mul x, -1) -> 0-x if (N1IsConst && ConstValue1.isAllOnes()) return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), N0); // fold (mul x, (1 << c)) -> x << c if (isConstantOrConstantVector(N1, /*NoOpaques*/ true) && DAG.isKnownToBeAPowerOfTwo(N1) && (!VT.isVector() || Level <= AfterLegalizeVectorOps)) { SDValue LogBase2 = BuildLogBase2(N1, DL); EVT ShiftVT = getShiftAmountTy(N0.getValueType()); SDValue Trunc = DAG.getZExtOrTrunc(LogBase2, DL, ShiftVT); return DAG.getNode(ISD::SHL, DL, VT, N0, Trunc); } // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c if (N1IsConst && !N1IsOpaqueConst && ConstValue1.isNegatedPowerOf2()) { unsigned Log2Val = (-ConstValue1).logBase2(); // FIXME: If the input is something that is easily negated (e.g. a // single-use add), we should put the negate there. return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), DAG.getNode(ISD::SHL, DL, VT, N0, DAG.getConstant(Log2Val, DL, getShiftAmountTy(N0.getValueType())))); } // Try to transform: // (1) multiply-by-(power-of-2 +/- 1) into shift and add/sub. // mul x, (2^N + 1) --> add (shl x, N), x // mul x, (2^N - 1) --> sub (shl x, N), x // Examples: x * 33 --> (x << 5) + x // x * 15 --> (x << 4) - x // x * -33 --> -((x << 5) + x) // x * -15 --> -((x << 4) - x) ; this reduces --> x - (x << 4) // (2) multiply-by-(power-of-2 +/- power-of-2) into shifts and add/sub. // mul x, (2^N + 2^M) --> (add (shl x, N), (shl x, M)) // mul x, (2^N - 2^M) --> (sub (shl x, N), (shl x, M)) // Examples: x * 0x8800 --> (x << 15) + (x << 11) // x * 0xf800 --> (x << 16) - (x << 11) // x * -0x8800 --> -((x << 15) + (x << 11)) // x * -0xf800 --> -((x << 16) - (x << 11)) ; (x << 11) - (x << 16) if (N1IsConst && TLI.decomposeMulByConstant(*DAG.getContext(), VT, N1)) { // TODO: We could handle more general decomposition of any constant by // having the target set a limit on number of ops and making a // callback to determine that sequence (similar to sqrt expansion). unsigned MathOp = ISD::DELETED_NODE; APInt MulC = ConstValue1.abs(); // The constant `2` should be treated as (2^0 + 1). unsigned TZeros = MulC == 2 ? 0 : MulC.countTrailingZeros(); MulC.lshrInPlace(TZeros); if ((MulC - 1).isPowerOf2()) MathOp = ISD::ADD; else if ((MulC + 1).isPowerOf2()) MathOp = ISD::SUB; if (MathOp != ISD::DELETED_NODE) { unsigned ShAmt = MathOp == ISD::ADD ? (MulC - 1).logBase2() : (MulC + 1).logBase2(); ShAmt += TZeros; assert(ShAmt < VT.getScalarSizeInBits() && "multiply-by-constant generated out of bounds shift"); SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, N0, DAG.getConstant(ShAmt, DL, VT)); SDValue R = TZeros ? DAG.getNode(MathOp, DL, VT, Shl, DAG.getNode(ISD::SHL, DL, VT, N0, DAG.getConstant(TZeros, DL, VT))) : DAG.getNode(MathOp, DL, VT, Shl, N0); if (ConstValue1.isNegative()) R = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), R); return R; } } // (mul (shl X, c1), c2) -> (mul X, c2 << c1) if (N0.getOpcode() == ISD::SHL) { SDValue N01 = N0.getOperand(1); if (SDValue C3 = DAG.FoldConstantArithmetic(ISD::SHL, DL, VT, {N1, N01})) return DAG.getNode(ISD::MUL, DL, VT, N0.getOperand(0), C3); } // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one // use. { SDValue Sh, Y; // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)). if (N0.getOpcode() == ISD::SHL && isConstantOrConstantVector(N0.getOperand(1)) && N0->hasOneUse()) { Sh = N0; Y = N1; } else if (N1.getOpcode() == ISD::SHL && isConstantOrConstantVector(N1.getOperand(1)) && N1->hasOneUse()) { Sh = N1; Y = N0; } if (Sh.getNode()) { SDValue Mul = DAG.getNode(ISD::MUL, DL, VT, Sh.getOperand(0), Y); return DAG.getNode(ISD::SHL, DL, VT, Mul, Sh.getOperand(1)); } } // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2) if (DAG.isConstantIntBuildVectorOrConstantInt(N1) && N0.getOpcode() == ISD::ADD && DAG.isConstantIntBuildVectorOrConstantInt(N0.getOperand(1)) && isMulAddWithConstProfitable(N, N0, N1)) return DAG.getNode( ISD::ADD, DL, VT, DAG.getNode(ISD::MUL, SDLoc(N0), VT, N0.getOperand(0), N1), DAG.getNode(ISD::MUL, SDLoc(N1), VT, N0.getOperand(1), N1)); // Fold (mul (vscale * C0), C1) to (vscale * (C0 * C1)). if (N0.getOpcode() == ISD::VSCALE) if (ConstantSDNode *NC1 = isConstOrConstSplat(N1)) { const APInt &C0 = N0.getConstantOperandAPInt(0); const APInt &C1 = NC1->getAPIntValue(); return DAG.getVScale(DL, VT, C0 * C1); } // Fold (mul step_vector(C0), C1) to (step_vector(C0 * C1)). APInt MulVal; if (N0.getOpcode() == ISD::STEP_VECTOR) if (ISD::isConstantSplatVector(N1.getNode(), MulVal)) { const APInt &C0 = N0.getConstantOperandAPInt(0); APInt NewStep = C0 * MulVal; return DAG.getStepVector(DL, VT, NewStep); } // Fold ((mul x, 0/undef) -> 0, // (mul x, 1) -> x) -> x) // -> and(x, mask) // We can replace vectors with '0' and '1' factors with a clearing mask. if (VT.isFixedLengthVector()) { unsigned NumElts = VT.getVectorNumElements(); SmallBitVector ClearMask; ClearMask.reserve(NumElts); auto IsClearMask = [&ClearMask](ConstantSDNode *V) { if (!V || V->isZero()) { ClearMask.push_back(true); return true; } ClearMask.push_back(false); return V->isOne(); }; if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::AND, VT)) && ISD::matchUnaryPredicate(N1, IsClearMask, /*AllowUndefs*/ true)) { assert(N1.getOpcode() == ISD::BUILD_VECTOR && "Unknown constant vector"); EVT LegalSVT = N1.getOperand(0).getValueType(); SDValue Zero = DAG.getConstant(0, DL, LegalSVT); SDValue AllOnes = DAG.getAllOnesConstant(DL, LegalSVT); SmallVector Mask(NumElts, AllOnes); for (unsigned I = 0; I != NumElts; ++I) if (ClearMask[I]) Mask[I] = Zero; return DAG.getNode(ISD::AND, DL, VT, N0, DAG.getBuildVector(VT, DL, Mask)); } } // reassociate mul if (SDValue RMUL = reassociateOps(ISD::MUL, DL, N0, N1, N->getFlags())) return RMUL; // Simplify the operands using demanded-bits information. if (SimplifyDemandedBits(SDValue(N, 0))) return SDValue(N, 0); return SDValue(); } /// Return true if divmod libcall is available. static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned, const TargetLowering &TLI) { RTLIB::Libcall LC; EVT NodeType = Node->getValueType(0); if (!NodeType.isSimple()) return false; switch (NodeType.getSimpleVT().SimpleTy) { default: return false; // No libcall for vector types. case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break; case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break; case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break; case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break; case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break; } return TLI.getLibcallName(LC) != nullptr; } /// Issue divrem if both quotient and remainder are needed. SDValue DAGCombiner::useDivRem(SDNode *Node) { if (Node->use_empty()) return SDValue(); // This is a dead node, leave it alone. unsigned Opcode = Node->getOpcode(); bool isSigned = (Opcode == ISD::SDIV) || (Opcode == ISD::SREM); unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM; // DivMod lib calls can still work on non-legal types if using lib-calls. EVT VT = Node->getValueType(0); if (VT.isVector() || !VT.isInteger()) return SDValue(); if (!TLI.isTypeLegal(VT) && !TLI.isOperationCustom(DivRemOpc, VT)) return SDValue(); // If DIVREM is going to get expanded into a libcall, // but there is no libcall available, then don't combine. if (!TLI.isOperationLegalOrCustom(DivRemOpc, VT) && !isDivRemLibcallAvailable(Node, isSigned, TLI)) return SDValue(); // If div is legal, it's better to do the normal expansion unsigned OtherOpcode = 0; if ((Opcode == ISD::SDIV) || (Opcode == ISD::UDIV)) { OtherOpcode = isSigned ? ISD::SREM : ISD::UREM; if (TLI.isOperationLegalOrCustom(Opcode, VT)) return SDValue(); } else { OtherOpcode = isSigned ? ISD::SDIV : ISD::UDIV; if (TLI.isOperationLegalOrCustom(OtherOpcode, VT)) return SDValue(); } SDValue Op0 = Node->getOperand(0); SDValue Op1 = Node->getOperand(1); SDValue combined; for (SDNode *User : Op0->uses()) { if (User == Node || User->getOpcode() == ISD::DELETED_NODE || User->use_empty()) continue; // Convert the other matching node(s), too; // otherwise, the DIVREM may get target-legalized into something // target-specific that we won't be able to recognize. unsigned UserOpc = User->getOpcode(); if ((UserOpc == Opcode || UserOpc == OtherOpcode || UserOpc == DivRemOpc) && User->getOperand(0) == Op0 && User->getOperand(1) == Op1) { if (!combined) { if (UserOpc == OtherOpcode) { SDVTList VTs = DAG.getVTList(VT, VT); combined = DAG.getNode(DivRemOpc, SDLoc(Node), VTs, Op0, Op1); } else if (UserOpc == DivRemOpc) { combined = SDValue(User, 0); } else { assert(UserOpc == Opcode); continue; } } if (UserOpc == ISD::SDIV || UserOpc == ISD::UDIV) CombineTo(User, combined); else if (UserOpc == ISD::SREM || UserOpc == ISD::UREM) CombineTo(User, combined.getValue(1)); } } return combined; } static SDValue simplifyDivRem(SDNode *N, SelectionDAG &DAG) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); SDLoc DL(N); unsigned Opc = N->getOpcode(); bool IsDiv = (ISD::SDIV == Opc) || (ISD::UDIV == Opc); ConstantSDNode *N1C = isConstOrConstSplat(N1); // X / undef -> undef // X % undef -> undef // X / 0 -> undef // X % 0 -> undef // NOTE: This includes vectors where any divisor element is zero/undef. if (DAG.isUndef(Opc, {N0, N1})) return DAG.getUNDEF(VT); // undef / X -> 0 // undef % X -> 0 if (N0.isUndef()) return DAG.getConstant(0, DL, VT); // 0 / X -> 0 // 0 % X -> 0 ConstantSDNode *N0C = isConstOrConstSplat(N0); if (N0C && N0C->isZero()) return N0; // X / X -> 1 // X % X -> 0 if (N0 == N1) return DAG.getConstant(IsDiv ? 1 : 0, DL, VT); // X / 1 -> X // X % 1 -> 0 // If this is a boolean op (single-bit element type), we can't have // division-by-zero or remainder-by-zero, so assume the divisor is 1. // TODO: Similarly, if we're zero-extending a boolean divisor, then assume // it's a 1. if ((N1C && N1C->isOne()) || (VT.getScalarType() == MVT::i1)) return IsDiv ? N0 : DAG.getConstant(0, DL, VT); return SDValue(); } SDValue DAGCombiner::visitSDIV(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); EVT CCVT = getSetCCResultType(VT); SDLoc DL(N); // fold (sdiv c1, c2) -> c1/c2 if (SDValue C = DAG.FoldConstantArithmetic(ISD::SDIV, DL, VT, {N0, N1})) return C; // fold vector ops if (VT.isVector()) if (SDValue FoldedVOp = SimplifyVBinOp(N, DL)) return FoldedVOp; // fold (sdiv X, -1) -> 0-X ConstantSDNode *N1C = isConstOrConstSplat(N1); if (N1C && N1C->isAllOnes()) return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), N0); // fold (sdiv X, MIN_SIGNED) -> select(X == MIN_SIGNED, 1, 0) if (N1C && N1C->getAPIntValue().isMinSignedValue()) return DAG.getSelect(DL, VT, DAG.getSetCC(DL, CCVT, N0, N1, ISD::SETEQ), DAG.getConstant(1, DL, VT), DAG.getConstant(0, DL, VT)); if (SDValue V = simplifyDivRem(N, DAG)) return V; if (SDValue NewSel = foldBinOpIntoSelect(N)) return NewSel; // If we know the sign bits of both operands are zero, strength reduce to a // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0)) return DAG.getNode(ISD::UDIV, DL, N1.getValueType(), N0, N1); if (SDValue V = visitSDIVLike(N0, N1, N)) { // If the corresponding remainder node exists, update its users with // (Dividend - (Quotient * Divisor). if (SDNode *RemNode = DAG.getNodeIfExists(ISD::SREM, N->getVTList(), { N0, N1 })) { SDValue Mul = DAG.getNode(ISD::MUL, DL, VT, V, N1); SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0, Mul); AddToWorklist(Mul.getNode()); AddToWorklist(Sub.getNode()); CombineTo(RemNode, Sub); } return V; } // sdiv, srem -> sdivrem // If the divisor is constant, then return DIVREM only if isIntDivCheap() is // true. Otherwise, we break the simplification logic in visitREM(). AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); if (!N1C || TLI.isIntDivCheap(N->getValueType(0), Attr)) if (SDValue DivRem = useDivRem(N)) return DivRem; return SDValue(); } static bool isDivisorPowerOfTwo(SDValue Divisor) { // Helper for determining whether a value is a power-2 constant scalar or a // vector of such elements. auto IsPowerOfTwo = [](ConstantSDNode *C) { if (C->isZero() || C->isOpaque()) return false; if (C->getAPIntValue().isPowerOf2()) return true; if (C->getAPIntValue().isNegatedPowerOf2()) return true; return false; }; return ISD::matchUnaryPredicate(Divisor, IsPowerOfTwo); } SDValue DAGCombiner::visitSDIVLike(SDValue N0, SDValue N1, SDNode *N) { SDLoc DL(N); EVT VT = N->getValueType(0); EVT CCVT = getSetCCResultType(VT); unsigned BitWidth = VT.getScalarSizeInBits(); // fold (sdiv X, pow2) -> simple ops after legalize // FIXME: We check for the exact bit here because the generic lowering gives // better results in that case. The target-specific lowering should learn how // to handle exact sdivs efficiently. if (!N->getFlags().hasExact() && isDivisorPowerOfTwo(N1)) { // Target-specific implementation of sdiv x, pow2. if (SDValue Res = BuildSDIVPow2(N)) return Res; // Create constants that are functions of the shift amount value. EVT ShiftAmtTy = getShiftAmountTy(N0.getValueType()); SDValue Bits = DAG.getConstant(BitWidth, DL, ShiftAmtTy); SDValue C1 = DAG.getNode(ISD::CTTZ, DL, VT, N1); C1 = DAG.getZExtOrTrunc(C1, DL, ShiftAmtTy); SDValue Inexact = DAG.getNode(ISD::SUB, DL, ShiftAmtTy, Bits, C1); if (!isConstantOrConstantVector(Inexact)) return SDValue(); // Splat the sign bit into the register SDValue Sign = DAG.getNode(ISD::SRA, DL, VT, N0, DAG.getConstant(BitWidth - 1, DL, ShiftAmtTy)); AddToWorklist(Sign.getNode()); // Add (N0 < 0) ? abs2 - 1 : 0; SDValue Srl = DAG.getNode(ISD::SRL, DL, VT, Sign, Inexact); AddToWorklist(Srl.getNode()); SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N0, Srl); AddToWorklist(Add.getNode()); SDValue Sra = DAG.getNode(ISD::SRA, DL, VT, Add, C1); AddToWorklist(Sra.getNode()); // Special case: (sdiv X, 1) -> X // Special Case: (sdiv X, -1) -> 0-X SDValue One = DAG.getConstant(1, DL, VT); SDValue AllOnes = DAG.getAllOnesConstant(DL, VT); SDValue IsOne = DAG.getSetCC(DL, CCVT, N1, One, ISD::SETEQ); SDValue IsAllOnes = DAG.getSetCC(DL, CCVT, N1, AllOnes, ISD::SETEQ); SDValue IsOneOrAllOnes = DAG.getNode(ISD::OR, DL, CCVT, IsOne, IsAllOnes); Sra = DAG.getSelect(DL, VT, IsOneOrAllOnes, N0, Sra); // If dividing by a positive value, we're done. Otherwise, the result must // be negated. SDValue Zero = DAG.getConstant(0, DL, VT); SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, Zero, Sra); // FIXME: Use SELECT_CC once we improve SELECT_CC constant-folding. SDValue IsNeg = DAG.getSetCC(DL, CCVT, N1, Zero, ISD::SETLT); SDValue Res = DAG.getSelect(DL, VT, IsNeg, Sub, Sra); return Res; } // If integer divide is expensive and we satisfy the requirements, emit an // alternate sequence. Targets may check function attributes for size/speed // trade-offs. AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); if (isConstantOrConstantVector(N1) && !TLI.isIntDivCheap(N->getValueType(0), Attr)) if (SDValue Op = BuildSDIV(N)) return Op; return SDValue(); } SDValue DAGCombiner::visitUDIV(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); EVT CCVT = getSetCCResultType(VT); SDLoc DL(N); // fold (udiv c1, c2) -> c1/c2 if (SDValue C = DAG.FoldConstantArithmetic(ISD::UDIV, DL, VT, {N0, N1})) return C; // fold vector ops if (VT.isVector()) if (SDValue FoldedVOp = SimplifyVBinOp(N, DL)) return FoldedVOp; // fold (udiv X, -1) -> select(X == -1, 1, 0) ConstantSDNode *N1C = isConstOrConstSplat(N1); if (N1C && N1C->isAllOnes()) return DAG.getSelect(DL, VT, DAG.getSetCC(DL, CCVT, N0, N1, ISD::SETEQ), DAG.getConstant(1, DL, VT), DAG.getConstant(0, DL, VT)); if (SDValue V = simplifyDivRem(N, DAG)) return V; if (SDValue NewSel = foldBinOpIntoSelect(N)) return NewSel; if (SDValue V = visitUDIVLike(N0, N1, N)) { // If the corresponding remainder node exists, update its users with // (Dividend - (Quotient * Divisor). if (SDNode *RemNode = DAG.getNodeIfExists(ISD::UREM, N->getVTList(), { N0, N1 })) { SDValue Mul = DAG.getNode(ISD::MUL, DL, VT, V, N1); SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0, Mul); AddToWorklist(Mul.getNode()); AddToWorklist(Sub.getNode()); CombineTo(RemNode, Sub); } return V; } // sdiv, srem -> sdivrem // If the divisor is constant, then return DIVREM only if isIntDivCheap() is // true. Otherwise, we break the simplification logic in visitREM(). AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); if (!N1C || TLI.isIntDivCheap(N->getValueType(0), Attr)) if (SDValue DivRem = useDivRem(N)) return DivRem; return SDValue(); } SDValue DAGCombiner::visitUDIVLike(SDValue N0, SDValue N1, SDNode *N) { SDLoc DL(N); EVT VT = N->getValueType(0); // fold (udiv x, (1 << c)) -> x >>u c if (isConstantOrConstantVector(N1, /*NoOpaques*/ true) && DAG.isKnownToBeAPowerOfTwo(N1)) { SDValue LogBase2 = BuildLogBase2(N1, DL); AddToWorklist(LogBase2.getNode()); EVT ShiftVT = getShiftAmountTy(N0.getValueType()); SDValue Trunc = DAG.getZExtOrTrunc(LogBase2, DL, ShiftVT); AddToWorklist(Trunc.getNode()); return DAG.getNode(ISD::SRL, DL, VT, N0, Trunc); } // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2 if (N1.getOpcode() == ISD::SHL) { SDValue N10 = N1.getOperand(0); if (isConstantOrConstantVector(N10, /*NoOpaques*/ true) && DAG.isKnownToBeAPowerOfTwo(N10)) { SDValue LogBase2 = BuildLogBase2(N10, DL); AddToWorklist(LogBase2.getNode()); EVT ADDVT = N1.getOperand(1).getValueType(); SDValue Trunc = DAG.getZExtOrTrunc(LogBase2, DL, ADDVT); AddToWorklist(Trunc.getNode()); SDValue Add = DAG.getNode(ISD::ADD, DL, ADDVT, N1.getOperand(1), Trunc); AddToWorklist(Add.getNode()); return DAG.getNode(ISD::SRL, DL, VT, N0, Add); } } // fold (udiv x, c) -> alternate AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); if (isConstantOrConstantVector(N1) && !TLI.isIntDivCheap(N->getValueType(0), Attr)) if (SDValue Op = BuildUDIV(N)) return Op; return SDValue(); } SDValue DAGCombiner::buildOptimizedSREM(SDValue N0, SDValue N1, SDNode *N) { if (!N->getFlags().hasExact() && isDivisorPowerOfTwo(N1) && !DAG.doesNodeExist(ISD::SDIV, N->getVTList(), {N0, N1})) { // Target-specific implementation of srem x, pow2. if (SDValue Res = BuildSREMPow2(N)) return Res; } return SDValue(); } // handles ISD::SREM and ISD::UREM SDValue DAGCombiner::visitREM(SDNode *N) { unsigned Opcode = N->getOpcode(); SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); EVT CCVT = getSetCCResultType(VT); bool isSigned = (Opcode == ISD::SREM); SDLoc DL(N); // fold (rem c1, c2) -> c1%c2 if (SDValue C = DAG.FoldConstantArithmetic(Opcode, DL, VT, {N0, N1})) return C; // fold (urem X, -1) -> select(FX == -1, 0, FX) // Freeze the numerator to avoid a miscompile with an undefined value. if (!isSigned && llvm::isAllOnesOrAllOnesSplat(N1, /*AllowUndefs*/ false)) { SDValue F0 = DAG.getFreeze(N0); SDValue EqualsNeg1 = DAG.getSetCC(DL, CCVT, F0, N1, ISD::SETEQ); return DAG.getSelect(DL, VT, EqualsNeg1, DAG.getConstant(0, DL, VT), F0); } if (SDValue V = simplifyDivRem(N, DAG)) return V; if (SDValue NewSel = foldBinOpIntoSelect(N)) return NewSel; if (isSigned) { // If we know the sign bits of both operands are zero, strength reduce to a // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0)) return DAG.getNode(ISD::UREM, DL, VT, N0, N1); } else { if (DAG.isKnownToBeAPowerOfTwo(N1)) { // fold (urem x, pow2) -> (and x, pow2-1) SDValue NegOne = DAG.getAllOnesConstant(DL, VT); SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N1, NegOne); AddToWorklist(Add.getNode()); return DAG.getNode(ISD::AND, DL, VT, N0, Add); } // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1)) // fold (urem x, (lshr pow2, y)) -> (and x, (add (lshr pow2, y), -1)) // TODO: We should sink the following into isKnownToBePowerOfTwo // using a OrZero parameter analogous to our handling in ValueTracking. if ((N1.getOpcode() == ISD::SHL || N1.getOpcode() == ISD::SRL) && DAG.isKnownToBeAPowerOfTwo(N1.getOperand(0))) { SDValue NegOne = DAG.getAllOnesConstant(DL, VT); SDValue Add = DAG.getNode(ISD::ADD, DL, VT, N1, NegOne); AddToWorklist(Add.getNode()); return DAG.getNode(ISD::AND, DL, VT, N0, Add); } } AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); // If X/C can be simplified by the division-by-constant logic, lower // X%C to the equivalent of X-X/C*C. // Reuse the SDIVLike/UDIVLike combines - to avoid mangling nodes, the // speculative DIV must not cause a DIVREM conversion. We guard against this // by skipping the simplification if isIntDivCheap(). When div is not cheap, // combine will not return a DIVREM. Regardless, checking cheapness here // makes sense since the simplification results in fatter code. if (DAG.isKnownNeverZero(N1) && !TLI.isIntDivCheap(VT, Attr)) { if (isSigned) { // check if we can build faster implementation for srem if (SDValue OptimizedRem = buildOptimizedSREM(N0, N1, N)) return OptimizedRem; } SDValue OptimizedDiv = isSigned ? visitSDIVLike(N0, N1, N) : visitUDIVLike(N0, N1, N); if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != N) { // If the equivalent Div node also exists, update its users. unsigned DivOpcode = isSigned ? ISD::SDIV : ISD::UDIV; if (SDNode *DivNode = DAG.getNodeIfExists(DivOpcode, N->getVTList(), { N0, N1 })) CombineTo(DivNode, OptimizedDiv); SDValue Mul = DAG.getNode(ISD::MUL, DL, VT, OptimizedDiv, N1); SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, N0, Mul); AddToWorklist(OptimizedDiv.getNode()); AddToWorklist(Mul.getNode()); return Sub; } } // sdiv, srem -> sdivrem if (SDValue DivRem = useDivRem(N)) return DivRem.getValue(1); return SDValue(); } SDValue DAGCombiner::visitMULHS(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); SDLoc DL(N); // fold (mulhs c1, c2) if (SDValue C = DAG.FoldConstantArithmetic(ISD::MULHS, DL, VT, {N0, N1})) return C; // canonicalize constant to RHS. if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && !DAG.isConstantIntBuildVectorOrConstantInt(N1)) return DAG.getNode(ISD::MULHS, DL, N->getVTList(), N1, N0); if (VT.isVector()) { if (SDValue FoldedVOp = SimplifyVBinOp(N, DL)) return FoldedVOp; // fold (mulhs x, 0) -> 0 // do not return N1, because undef node may exist. if (ISD::isConstantSplatVectorAllZeros(N1.getNode())) return DAG.getConstant(0, DL, VT); } // fold (mulhs x, 0) -> 0 if (isNullConstant(N1)) return N1; // fold (mulhs x, 1) -> (sra x, size(x)-1) if (isOneConstant(N1)) return DAG.getNode(ISD::SRA, DL, N0.getValueType(), N0, DAG.getConstant(N0.getScalarValueSizeInBits() - 1, DL, getShiftAmountTy(N0.getValueType()))); // fold (mulhs x, undef) -> 0 if (N0.isUndef() || N1.isUndef()) return DAG.getConstant(0, DL, VT); // If the type twice as wide is legal, transform the mulhs to a wider multiply // plus a shift. if (!TLI.isOperationLegalOrCustom(ISD::MULHS, VT) && VT.isSimple() && !VT.isVector()) { MVT Simple = VT.getSimpleVT(); unsigned SimpleSize = Simple.getSizeInBits(); EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); if (TLI.isOperationLegal(ISD::MUL, NewVT)) { N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0); N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1); N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1); N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1, DAG.getConstant(SimpleSize, DL, getShiftAmountTy(N1.getValueType()))); return DAG.getNode(ISD::TRUNCATE, DL, VT, N1); } } return SDValue(); } SDValue DAGCombiner::visitMULHU(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); SDLoc DL(N); // fold (mulhu c1, c2) if (SDValue C = DAG.FoldConstantArithmetic(ISD::MULHU, DL, VT, {N0, N1})) return C; // canonicalize constant to RHS. if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && !DAG.isConstantIntBuildVectorOrConstantInt(N1)) return DAG.getNode(ISD::MULHU, DL, N->getVTList(), N1, N0); if (VT.isVector()) { if (SDValue FoldedVOp = SimplifyVBinOp(N, DL)) return FoldedVOp; // fold (mulhu x, 0) -> 0 // do not return N1, because undef node may exist. if (ISD::isConstantSplatVectorAllZeros(N1.getNode())) return DAG.getConstant(0, DL, VT); } // fold (mulhu x, 0) -> 0 if (isNullConstant(N1)) return N1; // fold (mulhu x, 1) -> 0 if (isOneConstant(N1)) return DAG.getConstant(0, DL, N0.getValueType()); // fold (mulhu x, undef) -> 0 if (N0.isUndef() || N1.isUndef()) return DAG.getConstant(0, DL, VT); // fold (mulhu x, (1 << c)) -> x >> (bitwidth - c) if (isConstantOrConstantVector(N1, /*NoOpaques*/ true) && DAG.isKnownToBeAPowerOfTwo(N1) && hasOperation(ISD::SRL, VT)) { unsigned NumEltBits = VT.getScalarSizeInBits(); SDValue LogBase2 = BuildLogBase2(N1, DL); SDValue SRLAmt = DAG.getNode( ISD::SUB, DL, VT, DAG.getConstant(NumEltBits, DL, VT), LogBase2); EVT ShiftVT = getShiftAmountTy(N0.getValueType()); SDValue Trunc = DAG.getZExtOrTrunc(SRLAmt, DL, ShiftVT); return DAG.getNode(ISD::SRL, DL, VT, N0, Trunc); } // If the type twice as wide is legal, transform the mulhu to a wider multiply // plus a shift. if (!TLI.isOperationLegalOrCustom(ISD::MULHU, VT) && VT.isSimple() && !VT.isVector()) { MVT Simple = VT.getSimpleVT(); unsigned SimpleSize = Simple.getSizeInBits(); EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); if (TLI.isOperationLegal(ISD::MUL, NewVT)) { N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0); N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1); N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1); N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1, DAG.getConstant(SimpleSize, DL, getShiftAmountTy(N1.getValueType()))); return DAG.getNode(ISD::TRUNCATE, DL, VT, N1); } } // Simplify the operands using demanded-bits information. // We don't have demanded bits support for MULHU so this just enables constant // folding based on known bits. if (SimplifyDemandedBits(SDValue(N, 0))) return SDValue(N, 0); return SDValue(); } SDValue DAGCombiner::visitAVG(SDNode *N) { unsigned Opcode = N->getOpcode(); SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); SDLoc DL(N); // fold (avg c1, c2) if (SDValue C = DAG.FoldConstantArithmetic(Opcode, DL, VT, {N0, N1})) return C; // canonicalize constant to RHS. if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && !DAG.isConstantIntBuildVectorOrConstantInt(N1)) return DAG.getNode(Opcode, DL, N->getVTList(), N1, N0); if (VT.isVector()) { if (SDValue FoldedVOp = SimplifyVBinOp(N, DL)) return FoldedVOp; // fold (avgfloor x, 0) -> x >> 1 if (ISD::isConstantSplatVectorAllZeros(N1.getNode())) { if (Opcode == ISD::AVGFLOORS) return DAG.getNode(ISD::SRA, DL, VT, N0, DAG.getConstant(1, DL, VT)); if (Opcode == ISD::AVGFLOORU) return DAG.getNode(ISD::SRL, DL, VT, N0, DAG.getConstant(1, DL, VT)); } } // fold (avg x, undef) -> x if (N0.isUndef()) return N1; if (N1.isUndef()) return N0; // TODO If we use avg for scalars anywhere, we can add (avgfl x, 0) -> x >> 1 return SDValue(); } /// Perform optimizations common to nodes that compute two values. LoOp and HiOp /// give the opcodes for the two computations that are being performed. Return /// true if a simplification was made. SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp, unsigned HiOp) { // If the high half is not needed, just compute the low half. bool HiExists = N->hasAnyUseOfValue(1); if (!HiExists && (!LegalOperations || TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) { SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops()); return CombineTo(N, Res, Res); } // If the low half is not needed, just compute the high half. bool LoExists = N->hasAnyUseOfValue(0); if (!LoExists && (!LegalOperations || TLI.isOperationLegalOrCustom(HiOp, N->getValueType(1)))) { SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops()); return CombineTo(N, Res, Res); } // If both halves are used, return as it is. if (LoExists && HiExists) return SDValue(); // If the two computed results can be simplified separately, separate them. if (LoExists) { SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops()); AddToWorklist(Lo.getNode()); SDValue LoOpt = combine(Lo.getNode()); if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() && (!LegalOperations || TLI.isOperationLegalOrCustom(LoOpt.getOpcode(), LoOpt.getValueType()))) return CombineTo(N, LoOpt, LoOpt); } if (HiExists) { SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops()); AddToWorklist(Hi.getNode()); SDValue HiOpt = combine(Hi.getNode()); if (HiOpt.getNode() && HiOpt != Hi && (!LegalOperations || TLI.isOperationLegalOrCustom(HiOpt.getOpcode(), HiOpt.getValueType()))) return CombineTo(N, HiOpt, HiOpt); } return SDValue(); } SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) { if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS)) return Res; SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); SDLoc DL(N); // canonicalize constant to RHS (vector doesn't have to splat) if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && !DAG.isConstantIntBuildVectorOrConstantInt(N1)) return DAG.getNode(ISD::SMUL_LOHI, DL, N->getVTList(), N1, N0); // If the type is twice as wide is legal, transform the mulhu to a wider // multiply plus a shift. if (VT.isSimple() && !VT.isVector()) { MVT Simple = VT.getSimpleVT(); unsigned SimpleSize = Simple.getSizeInBits(); EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); if (TLI.isOperationLegal(ISD::MUL, NewVT)) { SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0); SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1); Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi); // Compute the high part as N1. Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo, DAG.getConstant(SimpleSize, DL, getShiftAmountTy(Lo.getValueType()))); Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi); // Compute the low part as N0. Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo); return CombineTo(N, Lo, Hi); } } return SDValue(); } SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) { if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU)) return Res; SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); SDLoc DL(N); // canonicalize constant to RHS (vector doesn't have to splat) if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && !DAG.isConstantIntBuildVectorOrConstantInt(N1)) return DAG.getNode(ISD::UMUL_LOHI, DL, N->getVTList(), N1, N0); // (umul_lohi N0, 0) -> (0, 0) if (isNullConstant(N1)) { SDValue Zero = DAG.getConstant(0, DL, VT); return CombineTo(N, Zero, Zero); } // (umul_lohi N0, 1) -> (N0, 0) if (isOneConstant(N1)) { SDValue Zero = DAG.getConstant(0, DL, VT); return CombineTo(N, N0, Zero); } // If the type is twice as wide is legal, transform the mulhu to a wider // multiply plus a shift. if (VT.isSimple() && !VT.isVector()) { MVT Simple = VT.getSimpleVT(); unsigned SimpleSize = Simple.getSizeInBits(); EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); if (TLI.isOperationLegal(ISD::MUL, NewVT)) { SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0); SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1); Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi); // Compute the high part as N1. Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo, DAG.getConstant(SimpleSize, DL, getShiftAmountTy(Lo.getValueType()))); Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi); // Compute the low part as N0. Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo); return CombineTo(N, Lo, Hi); } } return SDValue(); } SDValue DAGCombiner::visitMULO(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N0.getValueType(); bool IsSigned = (ISD::SMULO == N->getOpcode()); EVT CarryVT = N->getValueType(1); SDLoc DL(N); ConstantSDNode *N0C = isConstOrConstSplat(N0); ConstantSDNode *N1C = isConstOrConstSplat(N1); // fold operation with constant operands. // TODO: Move this to FoldConstantArithmetic when it supports nodes with // multiple results. if (N0C && N1C) { bool Overflow; APInt Result = IsSigned ? N0C->getAPIntValue().smul_ov(N1C->getAPIntValue(), Overflow) : N0C->getAPIntValue().umul_ov(N1C->getAPIntValue(), Overflow); return CombineTo(N, DAG.getConstant(Result, DL, VT), DAG.getBoolConstant(Overflow, DL, CarryVT, CarryVT)); } // canonicalize constant to RHS. if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && !DAG.isConstantIntBuildVectorOrConstantInt(N1)) return DAG.getNode(N->getOpcode(), DL, N->getVTList(), N1, N0); // fold (mulo x, 0) -> 0 + no carry out if (isNullOrNullSplat(N1)) return CombineTo(N, DAG.getConstant(0, DL, VT), DAG.getConstant(0, DL, CarryVT)); // (mulo x, 2) -> (addo x, x) // FIXME: This needs a freeze. if (N1C && N1C->getAPIntValue() == 2 && (!IsSigned || VT.getScalarSizeInBits() > 2)) return DAG.getNode(IsSigned ? ISD::SADDO : ISD::UADDO, DL, N->getVTList(), N0, N0); if (IsSigned) { // A 1 bit SMULO overflows if both inputs are 1. if (VT.getScalarSizeInBits() == 1) { SDValue And = DAG.getNode(ISD::AND, DL, VT, N0, N1); return CombineTo(N, And, DAG.getSetCC(DL, CarryVT, And, DAG.getConstant(0, DL, VT), ISD::SETNE)); } // Multiplying n * m significant bits yields a result of n + m significant // bits. If the total number of significant bits does not exceed the // result bit width (minus 1), there is no overflow. unsigned SignBits = DAG.ComputeNumSignBits(N0); if (SignBits > 1) SignBits += DAG.ComputeNumSignBits(N1); if (SignBits > VT.getScalarSizeInBits() + 1) return CombineTo(N, DAG.getNode(ISD::MUL, DL, VT, N0, N1), DAG.getConstant(0, DL, CarryVT)); } else { KnownBits N1Known = DAG.computeKnownBits(N1); KnownBits N0Known = DAG.computeKnownBits(N0); bool Overflow; (void)N0Known.getMaxValue().umul_ov(N1Known.getMaxValue(), Overflow); if (!Overflow) return CombineTo(N, DAG.getNode(ISD::MUL, DL, VT, N0, N1), DAG.getConstant(0, DL, CarryVT)); } return SDValue(); } // Function to calculate whether the Min/Max pair of SDNodes (potentially // swapped around) make a signed saturate pattern, clamping to between a signed // saturate of -2^(BW-1) and 2^(BW-1)-1, or an unsigned saturate of 0 and 2^BW. // Returns the node being clamped and the bitwidth of the clamp in BW. Should // work with both SMIN/SMAX nodes and setcc/select combo. The operands are the // same as SimplifySelectCC. N0getAPIntValue(); const APInt &C2 = N3C->getAPIntValue(); if (C1.getBitWidth() < C2.getBitWidth() || C1 != C2.sext(C1.getBitWidth())) return 0; return CC == ISD::SETLT ? ISD::SMIN : (CC == ISD::SETGT ? ISD::SMAX : 0); }; // Check the initial value is a SMIN/SMAX equivalent. unsigned Opcode0 = isSignedMinMax(N0, N1, N2, N3, CC); if (!Opcode0) return SDValue(); SDValue N00, N01, N02, N03; ISD::CondCode N0CC; switch (N0.getOpcode()) { case ISD::SMIN: case ISD::SMAX: N00 = N02 = N0.getOperand(0); N01 = N03 = N0.getOperand(1); N0CC = N0.getOpcode() == ISD::SMIN ? ISD::SETLT : ISD::SETGT; break; case ISD::SELECT_CC: N00 = N0.getOperand(0); N01 = N0.getOperand(1); N02 = N0.getOperand(2); N03 = N0.getOperand(3); N0CC = cast(N0.getOperand(4))->get(); break; case ISD::SELECT: case ISD::VSELECT: if (N0.getOperand(0).getOpcode() != ISD::SETCC) return SDValue(); N00 = N0.getOperand(0).getOperand(0); N01 = N0.getOperand(0).getOperand(1); N02 = N0.getOperand(1); N03 = N0.getOperand(2); N0CC = cast(N0.getOperand(0).getOperand(2))->get(); break; default: return SDValue(); } unsigned Opcode1 = isSignedMinMax(N00, N01, N02, N03, N0CC); if (!Opcode1 || Opcode0 == Opcode1) return SDValue(); ConstantSDNode *MinCOp = isConstOrConstSplat(Opcode0 == ISD::SMIN ? N1 : N01); ConstantSDNode *MaxCOp = isConstOrConstSplat(Opcode0 == ISD::SMIN ? N01 : N1); if (!MinCOp || !MaxCOp || MinCOp->getValueType(0) != MaxCOp->getValueType(0)) return SDValue(); const APInt &MinC = MinCOp->getAPIntValue(); const APInt &MaxC = MaxCOp->getAPIntValue(); APInt MinCPlus1 = MinC + 1; if (-MaxC == MinCPlus1 && MinCPlus1.isPowerOf2()) { BW = MinCPlus1.exactLogBase2() + 1; Unsigned = false; return N02; } if (MaxC == 0 && MinCPlus1.isPowerOf2()) { BW = MinCPlus1.exactLogBase2(); Unsigned = true; return N02; } return SDValue(); } static SDValue PerformMinMaxFpToSatCombine(SDValue N0, SDValue N1, SDValue N2, SDValue N3, ISD::CondCode CC, SelectionDAG &DAG) { unsigned BW; bool Unsigned; SDValue Fp = isSaturatingMinMax(N0, N1, N2, N3, CC, BW, Unsigned); if (!Fp || Fp.getOpcode() != ISD::FP_TO_SINT) return SDValue(); EVT FPVT = Fp.getOperand(0).getValueType(); EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), BW); if (FPVT.isVector()) NewVT = EVT::getVectorVT(*DAG.getContext(), NewVT, FPVT.getVectorElementCount()); unsigned NewOpc = Unsigned ? ISD::FP_TO_UINT_SAT : ISD::FP_TO_SINT_SAT; if (!DAG.getTargetLoweringInfo().shouldConvertFpToSat(NewOpc, FPVT, NewVT)) return SDValue(); SDLoc DL(Fp); SDValue Sat = DAG.getNode(NewOpc, DL, NewVT, Fp.getOperand(0), DAG.getValueType(NewVT.getScalarType())); return Unsigned ? DAG.getZExtOrTrunc(Sat, DL, N2->getValueType(0)) : DAG.getSExtOrTrunc(Sat, DL, N2->getValueType(0)); } static SDValue PerformUMinFpToSatCombine(SDValue N0, SDValue N1, SDValue N2, SDValue N3, ISD::CondCode CC, SelectionDAG &DAG) { // We are looking for UMIN(FPTOUI(X), (2^n)-1), which may have come via a // select/vselect/select_cc. The two operands pairs for the select (N2/N3) may // be truncated versions of the the setcc (N0/N1). if ((N0 != N2 && (N2.getOpcode() != ISD::TRUNCATE || N0 != N2.getOperand(0))) || N0.getOpcode() != ISD::FP_TO_UINT || CC != ISD::SETULT) return SDValue(); ConstantSDNode *N1C = isConstOrConstSplat(N1); ConstantSDNode *N3C = isConstOrConstSplat(N3); if (!N1C || !N3C) return SDValue(); const APInt &C1 = N1C->getAPIntValue(); const APInt &C3 = N3C->getAPIntValue(); if (!(C1 + 1).isPowerOf2() || C1.getBitWidth() < C3.getBitWidth() || C1 != C3.zext(C1.getBitWidth())) return SDValue(); unsigned BW = (C1 + 1).exactLogBase2(); EVT FPVT = N0.getOperand(0).getValueType(); EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), BW); if (FPVT.isVector()) NewVT = EVT::getVectorVT(*DAG.getContext(), NewVT, FPVT.getVectorElementCount()); if (!DAG.getTargetLoweringInfo().shouldConvertFpToSat(ISD::FP_TO_UINT_SAT, FPVT, NewVT)) return SDValue(); SDValue Sat = DAG.getNode(ISD::FP_TO_UINT_SAT, SDLoc(N0), NewVT, N0.getOperand(0), DAG.getValueType(NewVT.getScalarType())); return DAG.getZExtOrTrunc(Sat, SDLoc(N0), N3.getValueType()); } SDValue DAGCombiner::visitIMINMAX(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N0.getValueType(); unsigned Opcode = N->getOpcode(); SDLoc DL(N); // fold operation with constant operands. if (SDValue C = DAG.FoldConstantArithmetic(Opcode, DL, VT, {N0, N1})) return C; // If the operands are the same, this is a no-op. if (N0 == N1) return N0; // canonicalize constant to RHS if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && !DAG.isConstantIntBuildVectorOrConstantInt(N1)) return DAG.getNode(Opcode, DL, VT, N1, N0); // fold vector ops if (VT.isVector()) if (SDValue FoldedVOp = SimplifyVBinOp(N, DL)) return FoldedVOp; // Is sign bits are zero, flip between UMIN/UMAX and SMIN/SMAX. // Only do this if the current op isn't legal and the flipped is. if (!TLI.isOperationLegal(Opcode, VT) && (N0.isUndef() || DAG.SignBitIsZero(N0)) && (N1.isUndef() || DAG.SignBitIsZero(N1))) { unsigned AltOpcode; switch (Opcode) { case ISD::SMIN: AltOpcode = ISD::UMIN; break; case ISD::SMAX: AltOpcode = ISD::UMAX; break; case ISD::UMIN: AltOpcode = ISD::SMIN; break; case ISD::UMAX: AltOpcode = ISD::SMAX; break; default: llvm_unreachable("Unknown MINMAX opcode"); } if (TLI.isOperationLegal(AltOpcode, VT)) return DAG.getNode(AltOpcode, DL, VT, N0, N1); } if (Opcode == ISD::SMIN || Opcode == ISD::SMAX) if (SDValue S = PerformMinMaxFpToSatCombine( N0, N1, N0, N1, Opcode == ISD::SMIN ? ISD::SETLT : ISD::SETGT, DAG)) return S; if (Opcode == ISD::UMIN) if (SDValue S = PerformUMinFpToSatCombine(N0, N1, N0, N1, ISD::SETULT, DAG)) return S; // Simplify the operands using demanded-bits information. if (SimplifyDemandedBits(SDValue(N, 0))) return SDValue(N, 0); return SDValue(); } /// If this is a bitwise logic instruction and both operands have the same /// opcode, try to sink the other opcode after the logic instruction. SDValue DAGCombiner::hoistLogicOpWithSameOpcodeHands(SDNode *N) { SDValue N0 = N->getOperand(0), N1 = N->getOperand(1); EVT VT = N0.getValueType(); unsigned LogicOpcode = N->getOpcode(); unsigned HandOpcode = N0.getOpcode(); assert((LogicOpcode == ISD::AND || LogicOpcode == ISD::OR || LogicOpcode == ISD::XOR) && "Expected logic opcode"); assert(HandOpcode == N1.getOpcode() && "Bad input!"); // Bail early if none of these transforms apply. if (N0.getNumOperands() == 0) return SDValue(); // FIXME: We should check number of uses of the operands to not increase // the instruction count for all transforms. // Handle size-changing casts. SDValue X = N0.getOperand(0); SDValue Y = N1.getOperand(0); EVT XVT = X.getValueType(); SDLoc DL(N); if (HandOpcode == ISD::ANY_EXTEND || HandOpcode == ISD::ZERO_EXTEND || HandOpcode == ISD::SIGN_EXTEND) { // If both operands have other uses, this transform would create extra // instructions without eliminating anything. if (!N0.hasOneUse() && !N1.hasOneUse()) return SDValue(); // We need matching integer source types. if (XVT != Y.getValueType()) return SDValue(); // Don't create an illegal op during or after legalization. Don't ever // create an unsupported vector op. if ((VT.isVector() || LegalOperations) && !TLI.isOperationLegalOrCustom(LogicOpcode, XVT)) return SDValue(); // Avoid infinite looping with PromoteIntBinOp. // TODO: Should we apply desirable/legal constraints to all opcodes? if (HandOpcode == ISD::ANY_EXTEND && LegalTypes && !TLI.isTypeDesirableForOp(LogicOpcode, XVT)) return SDValue(); // logic_op (hand_op X), (hand_op Y) --> hand_op (logic_op X, Y) SDValue Logic = DAG.getNode(LogicOpcode, DL, XVT, X, Y); return DAG.getNode(HandOpcode, DL, VT, Logic); } // logic_op (truncate x), (truncate y) --> truncate (logic_op x, y) if (HandOpcode == ISD::TRUNCATE) { // If both operands have other uses, this transform would create extra // instructions without eliminating anything. if (!N0.hasOneUse() && !N1.hasOneUse()) return SDValue(); // We need matching source types. if (XVT != Y.getValueType()) return SDValue(); // Don't create an illegal op during or after legalization. if (LegalOperations && !TLI.isOperationLegal(LogicOpcode, XVT)) return SDValue(); // Be extra careful sinking truncate. If it's free, there's no benefit in // widening a binop. Also, don't create a logic op on an illegal type. if (TLI.isZExtFree(VT, XVT) && TLI.isTruncateFree(XVT, VT)) return SDValue(); if (!TLI.isTypeLegal(XVT)) return SDValue(); SDValue Logic = DAG.getNode(LogicOpcode, DL, XVT, X, Y); return DAG.getNode(HandOpcode, DL, VT, Logic); } // For binops SHL/SRL/SRA/AND: // logic_op (OP x, z), (OP y, z) --> OP (logic_op x, y), z if ((HandOpcode == ISD::SHL || HandOpcode == ISD::SRL || HandOpcode == ISD::SRA || HandOpcode == ISD::AND) && N0.getOperand(1) == N1.getOperand(1)) { // If either operand has other uses, this transform is not an improvement. if (!N0.hasOneUse() || !N1.hasOneUse()) return SDValue(); SDValue Logic = DAG.getNode(LogicOpcode, DL, XVT, X, Y); return DAG.getNode(HandOpcode, DL, VT, Logic, N0.getOperand(1)); } // Unary ops: logic_op (bswap x), (bswap y) --> bswap (logic_op x, y) if (HandOpcode == ISD::BSWAP) { // If either operand has other uses, this transform is not an improvement. if (!N0.hasOneUse() || !N1.hasOneUse()) return SDValue(); SDValue Logic = DAG.getNode(LogicOpcode, DL, XVT, X, Y); return DAG.getNode(HandOpcode, DL, VT, Logic); } // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B)) // Only perform this optimization up until type legalization, before // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and // we don't want to undo this promotion. // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper // on scalars. if ((HandOpcode == ISD::BITCAST || HandOpcode == ISD::SCALAR_TO_VECTOR) && Level <= AfterLegalizeTypes) { // Input types must be integer and the same. if (XVT.isInteger() && XVT == Y.getValueType() && !(VT.isVector() && TLI.isTypeLegal(VT) && !XVT.isVector() && !TLI.isTypeLegal(XVT))) { SDValue Logic = DAG.getNode(LogicOpcode, DL, XVT, X, Y); return DAG.getNode(HandOpcode, DL, VT, Logic); } } // Xor/and/or are indifferent to the swizzle operation (shuffle of one value). // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B)) // If both shuffles use the same mask, and both shuffle within a single // vector, then it is worthwhile to move the swizzle after the operation. // The type-legalizer generates this pattern when loading illegal // vector types from memory. In many cases this allows additional shuffle // optimizations. // There are other cases where moving the shuffle after the xor/and/or // is profitable even if shuffles don't perform a swizzle. // If both shuffles use the same mask, and both shuffles have the same first // or second operand, then it might still be profitable to move the shuffle // after the xor/and/or operation. if (HandOpcode == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) { auto *SVN0 = cast(N0); auto *SVN1 = cast(N1); assert(X.getValueType() == Y.getValueType() && "Inputs to shuffles are not the same type"); // Check that both shuffles use the same mask. The masks are known to be of // the same length because the result vector type is the same. // Check also that shuffles have only one use to avoid introducing extra // instructions. if (!SVN0->hasOneUse() || !SVN1->hasOneUse() || !SVN0->getMask().equals(SVN1->getMask())) return SDValue(); // Don't try to fold this node if it requires introducing a // build vector of all zeros that might be illegal at this stage. SDValue ShOp = N0.getOperand(1); if (LogicOpcode == ISD::XOR && !ShOp.isUndef()) ShOp = tryFoldToZero(DL, TLI, VT, DAG, LegalOperations); // (logic_op (shuf (A, C), shuf (B, C))) --> shuf (logic_op (A, B), C) if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) { SDValue Logic = DAG.getNode(LogicOpcode, DL, VT, N0.getOperand(0), N1.getOperand(0)); return DAG.getVectorShuffle(VT, DL, Logic, ShOp, SVN0->getMask()); } // Don't try to fold this node if it requires introducing a // build vector of all zeros that might be illegal at this stage. ShOp = N0.getOperand(0); if (LogicOpcode == ISD::XOR && !ShOp.isUndef()) ShOp = tryFoldToZero(DL, TLI, VT, DAG, LegalOperations); // (logic_op (shuf (C, A), shuf (C, B))) --> shuf (C, logic_op (A, B)) if (N0.getOperand(0) == N1.getOperand(0) && ShOp.getNode()) { SDValue Logic = DAG.getNode(LogicOpcode, DL, VT, N0.getOperand(1), N1.getOperand(1)); return DAG.getVectorShuffle(VT, DL, ShOp, Logic, SVN0->getMask()); } } return SDValue(); } /// Try to make (and/or setcc (LL, LR), setcc (RL, RR)) more efficient. SDValue DAGCombiner::foldLogicOfSetCCs(bool IsAnd, SDValue N0, SDValue N1, const SDLoc &DL) { SDValue LL, LR, RL, RR, N0CC, N1CC; if (!isSetCCEquivalent(N0, LL, LR, N0CC) || !isSetCCEquivalent(N1, RL, RR, N1CC)) return SDValue(); assert(N0.getValueType() == N1.getValueType() && "Unexpected operand types for bitwise logic op"); assert(LL.getValueType() == LR.getValueType() && RL.getValueType() == RR.getValueType() && "Unexpected operand types for setcc"); // If we're here post-legalization or the logic op type is not i1, the logic // op type must match a setcc result type. Also, all folds require new // operations on the left and right operands, so those types must match. EVT VT = N0.getValueType(); EVT OpVT = LL.getValueType(); if (LegalOperations || VT.getScalarType() != MVT::i1) if (VT != getSetCCResultType(OpVT)) return SDValue(); if (OpVT != RL.getValueType()) return SDValue(); ISD::CondCode CC0 = cast(N0CC)->get(); ISD::CondCode CC1 = cast(N1CC)->get(); bool IsInteger = OpVT.isInteger(); if (LR == RR && CC0 == CC1 && IsInteger) { bool IsZero = isNullOrNullSplat(LR); bool IsNeg1 = isAllOnesOrAllOnesSplat(LR); // All bits clear? bool AndEqZero = IsAnd && CC1 == ISD::SETEQ && IsZero; // All sign bits clear? bool AndGtNeg1 = IsAnd && CC1 == ISD::SETGT && IsNeg1; // Any bits set? bool OrNeZero = !IsAnd && CC1 == ISD::SETNE && IsZero; // Any sign bits set? bool OrLtZero = !IsAnd && CC1 == ISD::SETLT && IsZero; // (and (seteq X, 0), (seteq Y, 0)) --> (seteq (or X, Y), 0) // (and (setgt X, -1), (setgt Y, -1)) --> (setgt (or X, Y), -1) // (or (setne X, 0), (setne Y, 0)) --> (setne (or X, Y), 0) // (or (setlt X, 0), (setlt Y, 0)) --> (setlt (or X, Y), 0) if (AndEqZero || AndGtNeg1 || OrNeZero || OrLtZero) { SDValue Or = DAG.getNode(ISD::OR, SDLoc(N0), OpVT, LL, RL); AddToWorklist(Or.getNode()); return DAG.getSetCC(DL, VT, Or, LR, CC1); } // All bits set? bool AndEqNeg1 = IsAnd && CC1 == ISD::SETEQ && IsNeg1; // All sign bits set? bool AndLtZero = IsAnd && CC1 == ISD::SETLT && IsZero; // Any bits clear? bool OrNeNeg1 = !IsAnd && CC1 == ISD::SETNE && IsNeg1; // Any sign bits clear? bool OrGtNeg1 = !IsAnd && CC1 == ISD::SETGT && IsNeg1; // (and (seteq X, -1), (seteq Y, -1)) --> (seteq (and X, Y), -1) // (and (setlt X, 0), (setlt Y, 0)) --> (setlt (and X, Y), 0) // (or (setne X, -1), (setne Y, -1)) --> (setne (and X, Y), -1) // (or (setgt X, -1), (setgt Y -1)) --> (setgt (and X, Y), -1) if (AndEqNeg1 || AndLtZero || OrNeNeg1 || OrGtNeg1) { SDValue And = DAG.getNode(ISD::AND, SDLoc(N0), OpVT, LL, RL); AddToWorklist(And.getNode()); return DAG.getSetCC(DL, VT, And, LR, CC1); } } // TODO: What is the 'or' equivalent of this fold? // (and (setne X, 0), (setne X, -1)) --> (setuge (add X, 1), 2) if (IsAnd && LL == RL && CC0 == CC1 && OpVT.getScalarSizeInBits() > 1 && IsInteger && CC0 == ISD::SETNE && ((isNullConstant(LR) && isAllOnesConstant(RR)) || (isAllOnesConstant(LR) && isNullConstant(RR)))) { SDValue One = DAG.getConstant(1, DL, OpVT); SDValue Two = DAG.getConstant(2, DL, OpVT); SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0), OpVT, LL, One); AddToWorklist(Add.getNode()); return DAG.getSetCC(DL, VT, Add, Two, ISD::SETUGE); } // Try more general transforms if the predicates match and the only user of // the compares is the 'and' or 'or'. if (IsInteger && TLI.convertSetCCLogicToBitwiseLogic(OpVT) && CC0 == CC1 && N0.hasOneUse() && N1.hasOneUse()) { // and (seteq A, B), (seteq C, D) --> seteq (or (xor A, B), (xor C, D)), 0 // or (setne A, B), (setne C, D) --> setne (or (xor A, B), (xor C, D)), 0 if ((IsAnd && CC1 == ISD::SETEQ) || (!IsAnd && CC1 == ISD::SETNE)) { SDValue XorL = DAG.getNode(ISD::XOR, SDLoc(N0), OpVT, LL, LR); SDValue XorR = DAG.getNode(ISD::XOR, SDLoc(N1), OpVT, RL, RR); SDValue Or = DAG.getNode(ISD::OR, DL, OpVT, XorL, XorR); SDValue Zero = DAG.getConstant(0, DL, OpVT); return DAG.getSetCC(DL, VT, Or, Zero, CC1); } // Turn compare of constants whose difference is 1 bit into add+and+setcc. if ((IsAnd && CC1 == ISD::SETNE) || (!IsAnd && CC1 == ISD::SETEQ)) { // Match a shared variable operand and 2 non-opaque constant operands. auto MatchDiffPow2 = [&](ConstantSDNode *C0, ConstantSDNode *C1) { // The difference of the constants must be a single bit. const APInt &CMax = APIntOps::umax(C0->getAPIntValue(), C1->getAPIntValue()); const APInt &CMin = APIntOps::umin(C0->getAPIntValue(), C1->getAPIntValue()); return !C0->isOpaque() && !C1->isOpaque() && (CMax - CMin).isPowerOf2(); }; if (LL == RL && ISD::matchBinaryPredicate(LR, RR, MatchDiffPow2)) { // and/or (setcc X, CMax, ne), (setcc X, CMin, ne/eq) --> // setcc ((sub X, CMin), ~(CMax - CMin)), 0, ne/eq SDValue Max = DAG.getNode(ISD::UMAX, DL, OpVT, LR, RR); SDValue Min = DAG.getNode(ISD::UMIN, DL, OpVT, LR, RR); SDValue Offset = DAG.getNode(ISD::SUB, DL, OpVT, LL, Min); SDValue Diff = DAG.getNode(ISD::SUB, DL, OpVT, Max, Min); SDValue Mask = DAG.getNOT(DL, Diff, OpVT); SDValue And = DAG.getNode(ISD::AND, DL, OpVT, Offset, Mask); SDValue Zero = DAG.getConstant(0, DL, OpVT); return DAG.getSetCC(DL, VT, And, Zero, CC0); } } } // Canonicalize equivalent operands to LL == RL. if (LL == RR && LR == RL) { CC1 = ISD::getSetCCSwappedOperands(CC1); std::swap(RL, RR); } // (and (setcc X, Y, CC0), (setcc X, Y, CC1)) --> (setcc X, Y, NewCC) // (or (setcc X, Y, CC0), (setcc X, Y, CC1)) --> (setcc X, Y, NewCC) if (LL == RL && LR == RR) { ISD::CondCode NewCC = IsAnd ? ISD::getSetCCAndOperation(CC0, CC1, OpVT) : ISD::getSetCCOrOperation(CC0, CC1, OpVT); if (NewCC != ISD::SETCC_INVALID && (!LegalOperations || (TLI.isCondCodeLegal(NewCC, LL.getSimpleValueType()) && TLI.isOperationLegal(ISD::SETCC, OpVT)))) return DAG.getSetCC(DL, VT, LL, LR, NewCC); } return SDValue(); } /// This contains all DAGCombine rules which reduce two values combined by /// an And operation to a single value. This makes them reusable in the context /// of visitSELECT(). Rules involving constants are not included as /// visitSELECT() already handles those cases. SDValue DAGCombiner::visitANDLike(SDValue N0, SDValue N1, SDNode *N) { EVT VT = N1.getValueType(); SDLoc DL(N); // fold (and x, undef) -> 0 if (N0.isUndef() || N1.isUndef()) return DAG.getConstant(0, DL, VT); if (SDValue V = foldLogicOfSetCCs(true, N0, N1, DL)) return V; // TODO: Rewrite this to return a new 'AND' instead of using CombineTo. if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL && VT.getSizeInBits() <= 64 && N0->hasOneUse()) { if (ConstantSDNode *ADDI = dyn_cast(N0.getOperand(1))) { if (ConstantSDNode *SRLI = dyn_cast(N1.getOperand(1))) { // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal // immediate for an add, but it is legal if its top c2 bits are set, // transform the ADD so the immediate doesn't need to be materialized // in a register. APInt ADDC = ADDI->getAPIntValue(); APInt SRLC = SRLI->getAPIntValue(); if (ADDC.getMinSignedBits() <= 64 && SRLC.ult(VT.getSizeInBits()) && !TLI.isLegalAddImmediate(ADDC.getSExtValue())) { APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(), SRLC.getZExtValue()); if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) { ADDC |= Mask; if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) { SDLoc DL0(N0); SDValue NewAdd = DAG.getNode(ISD::ADD, DL0, VT, N0.getOperand(0), DAG.getConstant(ADDC, DL, VT)); CombineTo(N0.getNode(), NewAdd); // Return N so it doesn't get rechecked! return SDValue(N, 0); } } } } } } // Reduce bit extract of low half of an integer to the narrower type. // (and (srl i64:x, K), KMask) -> // (i64 zero_extend (and (srl (i32 (trunc i64:x)), K)), KMask) if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) { if (ConstantSDNode *CAnd = dyn_cast(N1)) { if (ConstantSDNode *CShift = dyn_cast(N0.getOperand(1))) { unsigned Size = VT.getSizeInBits(); const APInt &AndMask = CAnd->getAPIntValue(); unsigned ShiftBits = CShift->getZExtValue(); // Bail out, this node will probably disappear anyway. if (ShiftBits == 0) return SDValue(); unsigned MaskBits = AndMask.countTrailingOnes(); EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), Size / 2); if (AndMask.isMask() && // Required bits must not span the two halves of the integer and // must fit in the half size type. (ShiftBits + MaskBits <= Size / 2) && TLI.isNarrowingProfitable(VT, HalfVT) && TLI.isTypeDesirableForOp(ISD::AND, HalfVT) && TLI.isTypeDesirableForOp(ISD::SRL, HalfVT) && TLI.isTruncateFree(VT, HalfVT) && TLI.isZExtFree(HalfVT, VT)) { // The isNarrowingProfitable is to avoid regressions on PPC and // AArch64 which match a few 64-bit bit insert / bit extract patterns // on downstream users of this. Those patterns could probably be // extended to handle extensions mixed in. SDValue SL(N0); assert(MaskBits <= Size); // Extracting the highest bit of the low half. EVT ShiftVT = TLI.getShiftAmountTy(HalfVT, DAG.getDataLayout()); SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SL, HalfVT, N0.getOperand(0)); SDValue NewMask = DAG.getConstant(AndMask.trunc(Size / 2), SL, HalfVT); SDValue ShiftK = DAG.getConstant(ShiftBits, SL, ShiftVT); SDValue Shift = DAG.getNode(ISD::SRL, SL, HalfVT, Trunc, ShiftK); SDValue And = DAG.getNode(ISD::AND, SL, HalfVT, Shift, NewMask); return DAG.getNode(ISD::ZERO_EXTEND, SL, VT, And); } } } } return SDValue(); } bool DAGCombiner::isAndLoadExtLoad(ConstantSDNode *AndC, LoadSDNode *LoadN, EVT LoadResultTy, EVT &ExtVT) { if (!AndC->getAPIntValue().isMask()) return false; unsigned ActiveBits = AndC->getAPIntValue().countTrailingOnes(); ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits); EVT LoadedVT = LoadN->getMemoryVT(); if (ExtVT == LoadedVT && (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy, ExtVT))) { // ZEXTLOAD will match without needing to change the size of the value being // loaded. return true; } // Do not change the width of a volatile or atomic loads. if (!LoadN->isSimple()) return false; // Do not generate loads of non-round integer types since these can // be expensive (and would be wrong if the type is not byte sized). if (!LoadedVT.bitsGT(ExtVT) || !ExtVT.isRound()) return false; if (LegalOperations && !TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy, ExtVT)) return false; if (!TLI.shouldReduceLoadWidth(LoadN, ISD::ZEXTLOAD, ExtVT)) return false; return true; } bool DAGCombiner::isLegalNarrowLdSt(LSBaseSDNode *LDST, ISD::LoadExtType ExtType, EVT &MemVT, unsigned ShAmt) { if (!LDST) return false; // Only allow byte offsets. if (ShAmt % 8) return false; // Do not generate loads of non-round integer types since these can // be expensive (and would be wrong if the type is not byte sized). if (!MemVT.isRound()) return false; // Don't change the width of a volatile or atomic loads. if (!LDST->isSimple()) return false; EVT LdStMemVT = LDST->getMemoryVT(); // Bail out when changing the scalable property, since we can't be sure that // we're actually narrowing here. if (LdStMemVT.isScalableVector() != MemVT.isScalableVector()) return false; // Verify that we are actually reducing a load width here. if (LdStMemVT.bitsLT(MemVT)) return false; // Ensure that this isn't going to produce an unsupported memory access. if (ShAmt) { assert(ShAmt % 8 == 0 && "ShAmt is byte offset"); const unsigned ByteShAmt = ShAmt / 8; const Align LDSTAlign = LDST->getAlign(); const Align NarrowAlign = commonAlignment(LDSTAlign, ByteShAmt); if (!TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), MemVT, LDST->getAddressSpace(), NarrowAlign, LDST->getMemOperand()->getFlags())) return false; } // It's not possible to generate a constant of extended or untyped type. EVT PtrType = LDST->getBasePtr().getValueType(); if (PtrType == MVT::Untyped || PtrType.isExtended()) return false; if (isa(LDST)) { LoadSDNode *Load = cast(LDST); // Don't transform one with multiple uses, this would require adding a new // load. if (!SDValue(Load, 0).hasOneUse()) return false; if (LegalOperations && !TLI.isLoadExtLegal(ExtType, Load->getValueType(0), MemVT)) return false; // For the transform to be legal, the load must produce only two values // (the value loaded and the chain). Don't transform a pre-increment // load, for example, which produces an extra value. Otherwise the // transformation is not equivalent, and the downstream logic to replace // uses gets things wrong. if (Load->getNumValues() > 2) return false; // If the load that we're shrinking is an extload and we're not just // discarding the extension we can't simply shrink the load. Bail. // TODO: It would be possible to merge the extensions in some cases. if (Load->getExtensionType() != ISD::NON_EXTLOAD && Load->getMemoryVT().getSizeInBits() < MemVT.getSizeInBits() + ShAmt) return false; if (!TLI.shouldReduceLoadWidth(Load, ExtType, MemVT)) return false; } else { assert(isa(LDST) && "It is not a Load nor a Store SDNode"); StoreSDNode *Store = cast(LDST); // Can't write outside the original store if (Store->getMemoryVT().getSizeInBits() < MemVT.getSizeInBits() + ShAmt) return false; if (LegalOperations && !TLI.isTruncStoreLegal(Store->getValue().getValueType(), MemVT)) return false; } return true; } bool DAGCombiner::SearchForAndLoads(SDNode *N, SmallVectorImpl &Loads, SmallPtrSetImpl &NodesWithConsts, ConstantSDNode *Mask, SDNode *&NodeToMask) { // Recursively search for the operands, looking for loads which can be // narrowed. for (SDValue Op : N->op_values()) { if (Op.getValueType().isVector()) return false; // Some constants may need fixing up later if they are too large. if (auto *C = dyn_cast(Op)) { if ((N->getOpcode() == ISD::OR || N->getOpcode() == ISD::XOR) && (Mask->getAPIntValue() & C->getAPIntValue()) != C->getAPIntValue()) NodesWithConsts.insert(N); continue; } if (!Op.hasOneUse()) return false; switch(Op.getOpcode()) { case ISD::LOAD: { auto *Load = cast(Op); EVT ExtVT; if (isAndLoadExtLoad(Mask, Load, Load->getValueType(0), ExtVT) && isLegalNarrowLdSt(Load, ISD::ZEXTLOAD, ExtVT)) { // ZEXTLOAD is already small enough. if (Load->getExtensionType() == ISD::ZEXTLOAD && ExtVT.bitsGE(Load->getMemoryVT())) continue; // Use LE to convert equal sized loads to zext. if (ExtVT.bitsLE(Load->getMemoryVT())) Loads.push_back(Load); continue; } return false; } case ISD::ZERO_EXTEND: case ISD::AssertZext: { unsigned ActiveBits = Mask->getAPIntValue().countTrailingOnes(); EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits); EVT VT = Op.getOpcode() == ISD::AssertZext ? cast(Op.getOperand(1))->getVT() : Op.getOperand(0).getValueType(); // We can accept extending nodes if the mask is wider or an equal // width to the original type. if (ExtVT.bitsGE(VT)) continue; break; } case ISD::OR: case ISD::XOR: case ISD::AND: if (!SearchForAndLoads(Op.getNode(), Loads, NodesWithConsts, Mask, NodeToMask)) return false; continue; } // Allow one node which will masked along with any loads found. if (NodeToMask) return false; // Also ensure that the node to be masked only produces one data result. NodeToMask = Op.getNode(); if (NodeToMask->getNumValues() > 1) { bool HasValue = false; for (unsigned i = 0, e = NodeToMask->getNumValues(); i < e; ++i) { MVT VT = SDValue(NodeToMask, i).getSimpleValueType(); if (VT != MVT::Glue && VT != MVT::Other) { if (HasValue) { NodeToMask = nullptr; return false; } HasValue = true; } } assert(HasValue && "Node to be masked has no data result?"); } } return true; } bool DAGCombiner::BackwardsPropagateMask(SDNode *N) { auto *Mask = dyn_cast(N->getOperand(1)); if (!Mask) return false; if (!Mask->getAPIntValue().isMask()) return false; // No need to do anything if the and directly uses a load. if (isa(N->getOperand(0))) return false; SmallVector Loads; SmallPtrSet NodesWithConsts; SDNode *FixupNode = nullptr; if (SearchForAndLoads(N, Loads, NodesWithConsts, Mask, FixupNode)) { if (Loads.size() == 0) return false; LLVM_DEBUG(dbgs() << "Backwards propagate AND: "; N->dump()); SDValue MaskOp = N->getOperand(1); // If it exists, fixup the single node we allow in the tree that needs // masking. if (FixupNode) { LLVM_DEBUG(dbgs() << "First, need to fix up: "; FixupNode->dump()); SDValue And = DAG.getNode(ISD::AND, SDLoc(FixupNode), FixupNode->getValueType(0), SDValue(FixupNode, 0), MaskOp); DAG.ReplaceAllUsesOfValueWith(SDValue(FixupNode, 0), And); if (And.getOpcode() == ISD ::AND) DAG.UpdateNodeOperands(And.getNode(), SDValue(FixupNode, 0), MaskOp); } // Narrow any constants that need it. for (auto *LogicN : NodesWithConsts) { SDValue Op0 = LogicN->getOperand(0); SDValue Op1 = LogicN->getOperand(1); if (isa(Op0)) std::swap(Op0, Op1); SDValue And = DAG.getNode(ISD::AND, SDLoc(Op1), Op1.getValueType(), Op1, MaskOp); DAG.UpdateNodeOperands(LogicN, Op0, And); } // Create narrow loads. for (auto *Load : Loads) { LLVM_DEBUG(dbgs() << "Propagate AND back to: "; Load->dump()); SDValue And = DAG.getNode(ISD::AND, SDLoc(Load), Load->getValueType(0), SDValue(Load, 0), MaskOp); DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), And); if (And.getOpcode() == ISD ::AND) And = SDValue( DAG.UpdateNodeOperands(And.getNode(), SDValue(Load, 0), MaskOp), 0); SDValue NewLoad = reduceLoadWidth(And.getNode()); assert(NewLoad && "Shouldn't be masking the load if it can't be narrowed"); CombineTo(Load, NewLoad, NewLoad.getValue(1)); } DAG.ReplaceAllUsesWith(N, N->getOperand(0).getNode()); return true; } return false; } // Unfold // x & (-1 'logical shift' y) // To // (x 'opposite logical shift' y) 'logical shift' y // if it is better for performance. SDValue DAGCombiner::unfoldExtremeBitClearingToShifts(SDNode *N) { assert(N->getOpcode() == ISD::AND); SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); // Do we actually prefer shifts over mask? if (!TLI.shouldFoldMaskToVariableShiftPair(N0)) return SDValue(); // Try to match (-1 '[outer] logical shift' y) unsigned OuterShift; unsigned InnerShift; // The opposite direction to the OuterShift. SDValue Y; // Shift amount. auto matchMask = [&OuterShift, &InnerShift, &Y](SDValue M) -> bool { if (!M.hasOneUse()) return false; OuterShift = M->getOpcode(); if (OuterShift == ISD::SHL) InnerShift = ISD::SRL; else if (OuterShift == ISD::SRL) InnerShift = ISD::SHL; else return false; if (!isAllOnesConstant(M->getOperand(0))) return false; Y = M->getOperand(1); return true; }; SDValue X; if (matchMask(N1)) X = N0; else if (matchMask(N0)) X = N1; else return SDValue(); SDLoc DL(N); EVT VT = N->getValueType(0); // tmp = x 'opposite logical shift' y SDValue T0 = DAG.getNode(InnerShift, DL, VT, X, Y); // ret = tmp 'logical shift' y SDValue T1 = DAG.getNode(OuterShift, DL, VT, T0, Y); return T1; } /// Try to replace shift/logic that tests if a bit is clear with mask + setcc. /// For a target with a bit test, this is expected to become test + set and save /// at least 1 instruction. static SDValue combineShiftAnd1ToBitTest(SDNode *And, SelectionDAG &DAG) { assert(And->getOpcode() == ISD::AND && "Expected an 'and' op"); // This is probably not worthwhile without a supported type. EVT VT = And->getValueType(0); const TargetLowering &TLI = DAG.getTargetLoweringInfo(); if (!TLI.isTypeLegal(VT)) return SDValue(); // Look through an optional extension. SDValue And0 = And->getOperand(0), And1 = And->getOperand(1); if (And0.getOpcode() == ISD::ANY_EXTEND && And0.hasOneUse()) And0 = And0.getOperand(0); if (!isOneConstant(And1) || !And0.hasOneUse()) return SDValue(); SDValue Src = And0; // Attempt to find a 'not' op. // TODO: Should we favor test+set even without the 'not' op? bool FoundNot = false; if (isBitwiseNot(Src)) { FoundNot = true; Src = Src.getOperand(0); // Look though an optional truncation. The source operand may not be the // same type as the original 'and', but that is ok because we are masking // off everything but the low bit. if (Src.getOpcode() == ISD::TRUNCATE && Src.hasOneUse()) Src = Src.getOperand(0); } // Match a shift-right by constant. if (Src.getOpcode() != ISD::SRL || !Src.hasOneUse()) return SDValue(); // We might have looked through casts that make this transform invalid. // TODO: If the source type is wider than the result type, do the mask and // compare in the source type. unsigned VTBitWidth = VT.getScalarSizeInBits(); SDValue ShiftAmt = Src.getOperand(1); auto *ShiftAmtC = dyn_cast(ShiftAmt); if (!ShiftAmtC || !ShiftAmtC->getAPIntValue().ult(VTBitWidth)) return SDValue(); // Set source to shift source. Src = Src.getOperand(0); // Try again to find a 'not' op. // TODO: Should we favor test+set even with two 'not' ops? if (!FoundNot) { if (!isBitwiseNot(Src)) return SDValue(); Src = Src.getOperand(0); } if (!TLI.hasBitTest(Src, ShiftAmt)) return SDValue(); // Turn this into a bit-test pattern using mask op + setcc: // and (not (srl X, C)), 1 --> (and X, 1< (and X, 1<getZExtValue()), DL, VT); SDValue NewAnd = DAG.getNode(ISD::AND, DL, VT, X, Mask); SDValue Zero = DAG.getConstant(0, DL, VT); SDValue Setcc = DAG.getSetCC(DL, CCVT, NewAnd, Zero, ISD::SETEQ); return DAG.getZExtOrTrunc(Setcc, DL, VT); } /// For targets that support usubsat, match a bit-hack form of that operation /// that ends in 'and' and convert it. static SDValue foldAndToUsubsat(SDNode *N, SelectionDAG &DAG) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N1.getValueType(); // Canonicalize SRA as operand 1. if (N0.getOpcode() == ISD::SRA) std::swap(N0, N1); // xor/add with SMIN (signmask) are logically equivalent. if (N0.getOpcode() != ISD::XOR && N0.getOpcode() != ISD::ADD) return SDValue(); if (N1.getOpcode() != ISD::SRA || !N0.hasOneUse() || !N1.hasOneUse() || N0.getOperand(0) != N1.getOperand(0)) return SDValue(); unsigned BitWidth = VT.getScalarSizeInBits(); ConstantSDNode *XorC = isConstOrConstSplat(N0.getOperand(1), true); ConstantSDNode *SraC = isConstOrConstSplat(N1.getOperand(1), true); if (!XorC || !XorC->getAPIntValue().isSignMask() || !SraC || SraC->getAPIntValue() != BitWidth - 1) return SDValue(); // (i8 X ^ 128) & (i8 X s>> 7) --> usubsat X, 128 // (i8 X + 128) & (i8 X s>> 7) --> usubsat X, 128 SDLoc DL(N); SDValue SignMask = DAG.getConstant(XorC->getAPIntValue(), DL, VT); return DAG.getNode(ISD::USUBSAT, DL, VT, N0.getOperand(0), SignMask); } /// Given a bitwise logic operation N with a matching bitwise logic operand, /// fold a pattern where 2 of the source operands are identically shifted /// values. For example: /// ((X0 << Y) | Z) | (X1 << Y) --> ((X0 | X1) << Y) | Z static SDValue foldLogicOfShifts(SDNode *N, SDValue LogicOp, SDValue ShiftOp, SelectionDAG &DAG) { unsigned LogicOpcode = N->getOpcode(); assert((LogicOpcode == ISD::AND || LogicOpcode == ISD::OR || LogicOpcode == ISD::XOR) && "Expected bitwise logic operation"); if (!LogicOp.hasOneUse() || !ShiftOp.hasOneUse()) return SDValue(); // Match another bitwise logic op and a shift. unsigned ShiftOpcode = ShiftOp.getOpcode(); if (LogicOp.getOpcode() != LogicOpcode || !(ShiftOpcode == ISD::SHL || ShiftOpcode == ISD::SRL || ShiftOpcode == ISD::SRA)) return SDValue(); // Match another shift op inside the first logic operand. Handle both commuted // possibilities. // LOGIC (LOGIC (SH X0, Y), Z), (SH X1, Y) --> LOGIC (SH (LOGIC X0, X1), Y), Z // LOGIC (LOGIC Z, (SH X0, Y)), (SH X1, Y) --> LOGIC (SH (LOGIC X0, X1), Y), Z SDValue X1 = ShiftOp.getOperand(0); SDValue Y = ShiftOp.getOperand(1); SDValue X0, Z; if (LogicOp.getOperand(0).getOpcode() == ShiftOpcode && LogicOp.getOperand(0).getOperand(1) == Y) { X0 = LogicOp.getOperand(0).getOperand(0); Z = LogicOp.getOperand(1); } else if (LogicOp.getOperand(1).getOpcode() == ShiftOpcode && LogicOp.getOperand(1).getOperand(1) == Y) { X0 = LogicOp.getOperand(1).getOperand(0); Z = LogicOp.getOperand(0); } else { return SDValue(); } EVT VT = N->getValueType(0); SDLoc DL(N); SDValue LogicX = DAG.getNode(LogicOpcode, DL, VT, X0, X1); SDValue NewShift = DAG.getNode(ShiftOpcode, DL, VT, LogicX, Y); return DAG.getNode(LogicOpcode, DL, VT, NewShift, Z); } SDValue DAGCombiner::visitAND(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N1.getValueType(); // x & x --> x if (N0 == N1) return N0; // fold (and c1, c2) -> c1&c2 if (SDValue C = DAG.FoldConstantArithmetic(ISD::AND, SDLoc(N), VT, {N0, N1})) return C; // canonicalize constant to RHS if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && !DAG.isConstantIntBuildVectorOrConstantInt(N1)) return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0); // fold vector ops if (VT.isVector()) { if (SDValue FoldedVOp = SimplifyVBinOp(N, SDLoc(N))) return FoldedVOp; // fold (and x, 0) -> 0, vector edition if (ISD::isConstantSplatVectorAllZeros(N1.getNode())) // do not return N1, because undef node may exist in N1 return DAG.getConstant(APInt::getZero(N1.getScalarValueSizeInBits()), SDLoc(N), N1.getValueType()); // fold (and x, -1) -> x, vector edition if (ISD::isConstantSplatVectorAllOnes(N1.getNode())) return N0; // fold (and (masked_load) (splat_vec (x, ...))) to zext_masked_load auto *MLoad = dyn_cast(N0); ConstantSDNode *Splat = isConstOrConstSplat(N1, true, true); if (MLoad && MLoad->getExtensionType() == ISD::EXTLOAD && N0.hasOneUse() && Splat && N1.hasOneUse()) { EVT LoadVT = MLoad->getMemoryVT(); EVT ExtVT = VT; if (TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT, LoadVT)) { // For this AND to be a zero extension of the masked load the elements // of the BuildVec must mask the bottom bits of the extended element // type uint64_t ElementSize = LoadVT.getVectorElementType().getScalarSizeInBits(); if (Splat->getAPIntValue().isMask(ElementSize)) { return DAG.getMaskedLoad( ExtVT, SDLoc(N), MLoad->getChain(), MLoad->getBasePtr(), MLoad->getOffset(), MLoad->getMask(), MLoad->getPassThru(), LoadVT, MLoad->getMemOperand(), MLoad->getAddressingMode(), ISD::ZEXTLOAD, MLoad->isExpandingLoad()); } } } } // fold (and x, -1) -> x if (isAllOnesConstant(N1)) return N0; // if (and x, c) is known to be zero, return 0 unsigned BitWidth = VT.getScalarSizeInBits(); ConstantSDNode *N1C = isConstOrConstSplat(N1); if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0), APInt::getAllOnes(BitWidth))) return DAG.getConstant(0, SDLoc(N), VT); if (SDValue NewSel = foldBinOpIntoSelect(N)) return NewSel; // reassociate and if (SDValue RAND = reassociateOps(ISD::AND, SDLoc(N), N0, N1, N->getFlags())) return RAND; // Try to convert a constant mask AND into a shuffle clear mask. if (VT.isVector()) if (SDValue Shuffle = XformToShuffleWithZero(N)) return Shuffle; if (SDValue Combined = combineCarryDiamond(DAG, TLI, N0, N1, N)) return Combined; // fold (and (or x, C), D) -> D if (C & D) == D auto MatchSubset = [](ConstantSDNode *LHS, ConstantSDNode *RHS) { return RHS->getAPIntValue().isSubsetOf(LHS->getAPIntValue()); }; if (N0.getOpcode() == ISD::OR && ISD::matchBinaryPredicate(N0.getOperand(1), N1, MatchSubset)) return N1; // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits. if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) { SDValue N0Op0 = N0.getOperand(0); APInt Mask = ~N1C->getAPIntValue(); Mask = Mask.trunc(N0Op0.getScalarValueSizeInBits()); if (DAG.MaskedValueIsZero(N0Op0, Mask)) { SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N0.getValueType(), N0Op0); // Replace uses of the AND with uses of the Zero extend node. CombineTo(N, Zext); // We actually want to replace all uses of the any_extend with the // zero_extend, to avoid duplicating things. This will later cause this // AND to be folded. CombineTo(N0.getNode(), Zext); return SDValue(N, 0); // Return N so it doesn't get rechecked! } } // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) -> // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must // already be zero by virtue of the width of the base type of the load. // // the 'X' node here can either be nothing or an extract_vector_elt to catch // more cases. if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT && N0.getValueSizeInBits() == N0.getOperand(0).getScalarValueSizeInBits() && N0.getOperand(0).getOpcode() == ISD::LOAD && N0.getOperand(0).getResNo() == 0) || (N0.getOpcode() == ISD::LOAD && N0.getResNo() == 0)) { LoadSDNode *Load = cast( (N0.getOpcode() == ISD::LOAD) ? N0 : N0.getOperand(0) ); // Get the constant (if applicable) the zero'th operand is being ANDed with. // This can be a pure constant or a vector splat, in which case we treat the // vector as a scalar and use the splat value. APInt Constant = APInt::getZero(1); if (const ConstantSDNode *C = isConstOrConstSplat( N1, /*AllowUndef=*/false, /*AllowTruncation=*/true)) { Constant = C->getAPIntValue(); } else if (BuildVectorSDNode *Vector = dyn_cast(N1)) { APInt SplatValue, SplatUndef; unsigned SplatBitSize; bool HasAnyUndefs; bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, HasAnyUndefs); if (IsSplat) { // Undef bits can contribute to a possible optimisation if set, so // set them. SplatValue |= SplatUndef; // The splat value may be something like "0x00FFFFFF", which means 0 for // the first vector value and FF for the rest, repeating. We need a mask // that will apply equally to all members of the vector, so AND all the // lanes of the constant together. unsigned EltBitWidth = Vector->getValueType(0).getScalarSizeInBits(); // If the splat value has been compressed to a bitlength lower // than the size of the vector lane, we need to re-expand it to // the lane size. if (EltBitWidth > SplatBitSize) for (SplatValue = SplatValue.zextOrTrunc(EltBitWidth); SplatBitSize < EltBitWidth; SplatBitSize = SplatBitSize * 2) SplatValue |= SplatValue.shl(SplatBitSize); // Make sure that variable 'Constant' is only set if 'SplatBitSize' is a // multiple of 'BitWidth'. Otherwise, we could propagate a wrong value. if ((SplatBitSize % EltBitWidth) == 0) { Constant = APInt::getAllOnes(EltBitWidth); for (unsigned i = 0, n = (SplatBitSize / EltBitWidth); i < n; ++i) Constant &= SplatValue.extractBits(EltBitWidth, i * EltBitWidth); } } } // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is // actually legal and isn't going to get expanded, else this is a false // optimisation. bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD, Load->getValueType(0), Load->getMemoryVT()); // Resize the constant to the same size as the original memory access before // extension. If it is still the AllOnesValue then this AND is completely // unneeded. Constant = Constant.zextOrTrunc(Load->getMemoryVT().getScalarSizeInBits()); bool B; switch (Load->getExtensionType()) { default: B = false; break; case ISD::EXTLOAD: B = CanZextLoadProfitably; break; case ISD::ZEXTLOAD: case ISD::NON_EXTLOAD: B = true; break; } if (B && Constant.isAllOnes()) { // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to // preserve semantics once we get rid of the AND. SDValue NewLoad(Load, 0); // Fold the AND away. NewLoad may get replaced immediately. CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0); if (Load->getExtensionType() == ISD::EXTLOAD) { NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD, Load->getValueType(0), SDLoc(Load), Load->getChain(), Load->getBasePtr(), Load->getOffset(), Load->getMemoryVT(), Load->getMemOperand()); // Replace uses of the EXTLOAD with the new ZEXTLOAD. if (Load->getNumValues() == 3) { // PRE/POST_INC loads have 3 values. SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1), NewLoad.getValue(2) }; CombineTo(Load, To, 3, true); } else { CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1)); } } return SDValue(N, 0); // Return N so it doesn't get rechecked! } } if (N0.getOpcode() == ISD::EXTRACT_SUBVECTOR && N0.hasOneUse() && N1C && ISD::isExtOpcode(N0.getOperand(0).getOpcode())) { SDValue Ext = N0.getOperand(0); EVT ExtVT = Ext->getValueType(0); SDValue Extendee = Ext->getOperand(0); unsigned ScalarWidth = Extendee.getValueType().getScalarSizeInBits(); if (N1C->getAPIntValue().isMask(ScalarWidth) && (!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, ExtVT))) { // (and (extract_subvector (zext|anyext|sext v) _) iN_mask) // => (extract_subvector (iN_zeroext v)) SDValue ZeroExtExtendee = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), ExtVT, Extendee); return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N), VT, ZeroExtExtendee, N0.getOperand(1)); } } // fold (and (masked_gather x)) -> (zext_masked_gather x) if (auto *GN0 = dyn_cast(N0)) { EVT MemVT = GN0->getMemoryVT(); EVT ScalarVT = MemVT.getScalarType(); if (SDValue(GN0, 0).hasOneUse() && isConstantSplatVectorMaskForType(N1.getNode(), ScalarVT) && TLI.isVectorLoadExtDesirable(SDValue(SDValue(GN0, 0)))) { SDValue Ops[] = {GN0->getChain(), GN0->getPassThru(), GN0->getMask(), GN0->getBasePtr(), GN0->getIndex(), GN0->getScale()}; SDValue ZExtLoad = DAG.getMaskedGather( DAG.getVTList(VT, MVT::Other), MemVT, SDLoc(N), Ops, GN0->getMemOperand(), GN0->getIndexType(), ISD::ZEXTLOAD); CombineTo(N, ZExtLoad); AddToWorklist(ZExtLoad.getNode()); // Avoid recheck of N. return SDValue(N, 0); } } // fold (and (load x), 255) -> (zextload x, i8) // fold (and (extload x, i16), 255) -> (zextload x, i8) if (N1C && N0.getOpcode() == ISD::LOAD && !VT.isVector()) if (SDValue Res = reduceLoadWidth(N)) return Res; if (LegalTypes) { // Attempt to propagate the AND back up to the leaves which, if they're // loads, can be combined to narrow loads and the AND node can be removed. // Perform after legalization so that extend nodes will already be // combined into the loads. if (BackwardsPropagateMask(N)) return SDValue(N, 0); } if (SDValue Combined = visitANDLike(N0, N1, N)) return Combined; // Simplify: (and (op x...), (op y...)) -> (op (and x, y)) if (N0.getOpcode() == N1.getOpcode()) if (SDValue V = hoistLogicOpWithSameOpcodeHands(N)) return V; if (SDValue R = foldLogicOfShifts(N, N0, N1, DAG)) return R; if (SDValue R = foldLogicOfShifts(N, N1, N0, DAG)) return R; // Masking the negated extension of a boolean is just the zero-extended // boolean: // and (sub 0, zext(bool X)), 1 --> zext(bool X) // and (sub 0, sext(bool X)), 1 --> zext(bool X) // // Note: the SimplifyDemandedBits fold below can make an information-losing // transform, and then we have no way to find this better fold. if (N1C && N1C->isOne() && N0.getOpcode() == ISD::SUB) { if (isNullOrNullSplat(N0.getOperand(0))) { SDValue SubRHS = N0.getOperand(1); if (SubRHS.getOpcode() == ISD::ZERO_EXTEND && SubRHS.getOperand(0).getScalarValueSizeInBits() == 1) return SubRHS; if (SubRHS.getOpcode() == ISD::SIGN_EXTEND && SubRHS.getOperand(0).getScalarValueSizeInBits() == 1) return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, SubRHS.getOperand(0)); } } // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1) // fold (and (sra)) -> (and (srl)) when possible. if (SimplifyDemandedBits(SDValue(N, 0))) return SDValue(N, 0); // fold (zext_inreg (extload x)) -> (zextload x) // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use if (ISD::isUNINDEXEDLoad(N0.getNode()) && (ISD::isEXTLoad(N0.getNode()) || (ISD::isSEXTLoad(N0.getNode()) && N0.hasOneUse()))) { LoadSDNode *LN0 = cast(N0); EVT MemVT = LN0->getMemoryVT(); // If we zero all the possible extended bits, then we can turn this into // a zextload if we are running before legalize or the operation is legal. unsigned ExtBitSize = N1.getScalarValueSizeInBits(); unsigned MemBitSize = MemVT.getScalarSizeInBits(); APInt ExtBits = APInt::getHighBitsSet(ExtBitSize, ExtBitSize - MemBitSize); if (DAG.MaskedValueIsZero(N1, ExtBits) && ((!LegalOperations && LN0->isSimple()) || TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) { SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT, LN0->getChain(), LN0->getBasePtr(), MemVT, LN0->getMemOperand()); AddToWorklist(N); CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); return SDValue(N, 0); // Return N so it doesn't get rechecked! } } // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const) if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) { if (SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0), N0.getOperand(1), false)) return BSwap; } if (SDValue Shifts = unfoldExtremeBitClearingToShifts(N)) return Shifts; if (SDValue V = combineShiftAnd1ToBitTest(N, DAG)) return V; // Recognize the following pattern: // // AndVT = (and (sign_extend NarrowVT to AndVT) #bitmask) // // where bitmask is a mask that clears the upper bits of AndVT. The // number of bits in bitmask must be a power of two. auto IsAndZeroExtMask = [](SDValue LHS, SDValue RHS) { if (LHS->getOpcode() != ISD::SIGN_EXTEND) return false; auto *C = dyn_cast(RHS); if (!C) return false; if (!C->getAPIntValue().isMask( LHS.getOperand(0).getValueType().getFixedSizeInBits())) return false; return true; }; // Replace (and (sign_extend ...) #bitmask) with (zero_extend ...). if (IsAndZeroExtMask(N0, N1)) return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0.getOperand(0)); if (hasOperation(ISD::USUBSAT, VT)) if (SDValue V = foldAndToUsubsat(N, DAG)) return V; return SDValue(); } /// Match (a >> 8) | (a << 8) as (bswap a) >> 16. SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1, bool DemandHighBits) { if (!LegalOperations) return SDValue(); EVT VT = N->getValueType(0); if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16) return SDValue(); if (!TLI.isOperationLegalOrCustom(ISD::BSWAP, VT)) return SDValue(); // Recognize (and (shl a, 8), 0xff00), (and (srl a, 8), 0xff) bool LookPassAnd0 = false; bool LookPassAnd1 = false; if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL) std::swap(N0, N1); if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL) std::swap(N0, N1); if (N0.getOpcode() == ISD::AND) { if (!N0->hasOneUse()) return SDValue(); ConstantSDNode *N01C = dyn_cast(N0.getOperand(1)); // Also handle 0xffff since the LHS is guaranteed to have zeros there. // This is needed for X86. if (!N01C || (N01C->getZExtValue() != 0xFF00 && N01C->getZExtValue() != 0xFFFF)) return SDValue(); N0 = N0.getOperand(0); LookPassAnd0 = true; } if (N1.getOpcode() == ISD::AND) { if (!N1->hasOneUse()) return SDValue(); ConstantSDNode *N11C = dyn_cast(N1.getOperand(1)); if (!N11C || N11C->getZExtValue() != 0xFF) return SDValue(); N1 = N1.getOperand(0); LookPassAnd1 = true; } if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL) std::swap(N0, N1); if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL) return SDValue(); if (!N0->hasOneUse() || !N1->hasOneUse()) return SDValue(); ConstantSDNode *N01C = dyn_cast(N0.getOperand(1)); ConstantSDNode *N11C = dyn_cast(N1.getOperand(1)); if (!N01C || !N11C) return SDValue(); if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8) return SDValue(); // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8) SDValue N00 = N0->getOperand(0); if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) { if (!N00->hasOneUse()) return SDValue(); ConstantSDNode *N001C = dyn_cast(N00.getOperand(1)); if (!N001C || N001C->getZExtValue() != 0xFF) return SDValue(); N00 = N00.getOperand(0); LookPassAnd0 = true; } SDValue N10 = N1->getOperand(0); if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) { if (!N10->hasOneUse()) return SDValue(); ConstantSDNode *N101C = dyn_cast(N10.getOperand(1)); // Also allow 0xFFFF since the bits will be shifted out. This is needed // for X86. if (!N101C || (N101C->getZExtValue() != 0xFF00 && N101C->getZExtValue() != 0xFFFF)) return SDValue(); N10 = N10.getOperand(0); LookPassAnd1 = true; } if (N00 != N10) return SDValue(); // Make sure everything beyond the low halfword gets set to zero since the SRL // 16 will clear the top bits. unsigned OpSizeInBits = VT.getSizeInBits(); if (OpSizeInBits > 16) { // If the left-shift isn't masked out then the only way this is a bswap is // if all bits beyond the low 8 are 0. In that case the entire pattern // reduces to a left shift anyway: leave it for other parts of the combiner. if (DemandHighBits && !LookPassAnd0) return SDValue(); // However, if the right shift isn't masked out then it might be because // it's not needed. See if we can spot that too. If the high bits aren't // demanded, we only need bits 23:16 to be zero. Otherwise, we need all // upper bits to be zero. if (!LookPassAnd1) { unsigned HighBit = DemandHighBits ? OpSizeInBits : 24; if (!DAG.MaskedValueIsZero(N10, APInt::getBitsSet(OpSizeInBits, 16, HighBit))) return SDValue(); } } SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00); if (OpSizeInBits > 16) { SDLoc DL(N); Res = DAG.getNode(ISD::SRL, DL, VT, Res, DAG.getConstant(OpSizeInBits - 16, DL, getShiftAmountTy(VT))); } return Res; } /// Return true if the specified node is an element that makes up a 32-bit /// packed halfword byteswap. /// ((x & 0x000000ff) << 8) | /// ((x & 0x0000ff00) >> 8) | /// ((x & 0x00ff0000) << 8) | /// ((x & 0xff000000) >> 8) static bool isBSwapHWordElement(SDValue N, MutableArrayRef Parts) { if (!N->hasOneUse()) return false; unsigned Opc = N.getOpcode(); if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL) return false; SDValue N0 = N.getOperand(0); unsigned Opc0 = N0.getOpcode(); if (Opc0 != ISD::AND && Opc0 != ISD::SHL && Opc0 != ISD::SRL) return false; ConstantSDNode *N1C = nullptr; // SHL or SRL: look upstream for AND mask operand if (Opc == ISD::AND) N1C = dyn_cast(N.getOperand(1)); else if (Opc0 == ISD::AND) N1C = dyn_cast(N0.getOperand(1)); if (!N1C) return false; unsigned MaskByteOffset; switch (N1C->getZExtValue()) { default: return false; case 0xFF: MaskByteOffset = 0; break; case 0xFF00: MaskByteOffset = 1; break; case 0xFFFF: // In case demanded bits didn't clear the bits that will be shifted out. // This is needed for X86. if (Opc == ISD::SRL || (Opc == ISD::AND && Opc0 == ISD::SHL)) { MaskByteOffset = 1; break; } return false; case 0xFF0000: MaskByteOffset = 2; break; case 0xFF000000: MaskByteOffset = 3; break; } // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00). if (Opc == ISD::AND) { if (MaskByteOffset == 0 || MaskByteOffset == 2) { // (x >> 8) & 0xff // (x >> 8) & 0xff0000 if (Opc0 != ISD::SRL) return false; ConstantSDNode *C = dyn_cast(N0.getOperand(1)); if (!C || C->getZExtValue() != 8) return false; } else { // (x << 8) & 0xff00 // (x << 8) & 0xff000000 if (Opc0 != ISD::SHL) return false; ConstantSDNode *C = dyn_cast(N0.getOperand(1)); if (!C || C->getZExtValue() != 8) return false; } } else if (Opc == ISD::SHL) { // (x & 0xff) << 8 // (x & 0xff0000) << 8 if (MaskByteOffset != 0 && MaskByteOffset != 2) return false; ConstantSDNode *C = dyn_cast(N.getOperand(1)); if (!C || C->getZExtValue() != 8) return false; } else { // Opc == ISD::SRL // (x & 0xff00) >> 8 // (x & 0xff000000) >> 8 if (MaskByteOffset != 1 && MaskByteOffset != 3) return false; ConstantSDNode *C = dyn_cast(N.getOperand(1)); if (!C || C->getZExtValue() != 8) return false; } if (Parts[MaskByteOffset]) return false; Parts[MaskByteOffset] = N0.getOperand(0).getNode(); return true; } // Match 2 elements of a packed halfword bswap. static bool isBSwapHWordPair(SDValue N, MutableArrayRef Parts) { if (N.getOpcode() == ISD::OR) return isBSwapHWordElement(N.getOperand(0), Parts) && isBSwapHWordElement(N.getOperand(1), Parts); if (N.getOpcode() == ISD::SRL && N.getOperand(0).getOpcode() == ISD::BSWAP) { ConstantSDNode *C = isConstOrConstSplat(N.getOperand(1)); if (!C || C->getAPIntValue() != 16) return false; Parts[0] = Parts[1] = N.getOperand(0).getOperand(0).getNode(); return true; } return false; } // Match this pattern: // (or (and (shl (A, 8)), 0xff00ff00), (and (srl (A, 8)), 0x00ff00ff)) // And rewrite this to: // (rotr (bswap A), 16) static SDValue matchBSwapHWordOrAndAnd(const TargetLowering &TLI, SelectionDAG &DAG, SDNode *N, SDValue N0, SDValue N1, EVT VT, EVT ShiftAmountTy) { assert(N->getOpcode() == ISD::OR && VT == MVT::i32 && "MatchBSwapHWordOrAndAnd: expecting i32"); if (!TLI.isOperationLegalOrCustom(ISD::ROTR, VT)) return SDValue(); if (N0.getOpcode() != ISD::AND || N1.getOpcode() != ISD::AND) return SDValue(); // TODO: this is too restrictive; lifting this restriction requires more tests if (!N0->hasOneUse() || !N1->hasOneUse()) return SDValue(); ConstantSDNode *Mask0 = isConstOrConstSplat(N0.getOperand(1)); ConstantSDNode *Mask1 = isConstOrConstSplat(N1.getOperand(1)); if (!Mask0 || !Mask1) return SDValue(); if (Mask0->getAPIntValue() != 0xff00ff00 || Mask1->getAPIntValue() != 0x00ff00ff) return SDValue(); SDValue Shift0 = N0.getOperand(0); SDValue Shift1 = N1.getOperand(0); if (Shift0.getOpcode() != ISD::SHL || Shift1.getOpcode() != ISD::SRL) return SDValue(); ConstantSDNode *ShiftAmt0 = isConstOrConstSplat(Shift0.getOperand(1)); ConstantSDNode *ShiftAmt1 = isConstOrConstSplat(Shift1.getOperand(1)); if (!ShiftAmt0 || !ShiftAmt1) return SDValue(); if (ShiftAmt0->getAPIntValue() != 8 || ShiftAmt1->getAPIntValue() != 8) return SDValue(); if (Shift0.getOperand(0) != Shift1.getOperand(0)) return SDValue(); SDLoc DL(N); SDValue BSwap = DAG.getNode(ISD::BSWAP, DL, VT, Shift0.getOperand(0)); SDValue ShAmt = DAG.getConstant(16, DL, ShiftAmountTy); return DAG.getNode(ISD::ROTR, DL, VT, BSwap, ShAmt); } /// Match a 32-bit packed halfword bswap. That is /// ((x & 0x000000ff) << 8) | /// ((x & 0x0000ff00) >> 8) | /// ((x & 0x00ff0000) << 8) | /// ((x & 0xff000000) >> 8) /// => (rotl (bswap x), 16) SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) { if (!LegalOperations) return SDValue(); EVT VT = N->getValueType(0); if (VT != MVT::i32) return SDValue(); if (!TLI.isOperationLegalOrCustom(ISD::BSWAP, VT)) return SDValue(); if (SDValue BSwap = matchBSwapHWordOrAndAnd(TLI, DAG, N, N0, N1, VT, getShiftAmountTy(VT))) return BSwap; // Try again with commuted operands. if (SDValue BSwap = matchBSwapHWordOrAndAnd(TLI, DAG, N, N1, N0, VT, getShiftAmountTy(VT))) return BSwap; // Look for either // (or (bswaphpair), (bswaphpair)) // (or (or (bswaphpair), (and)), (and)) // (or (or (and), (bswaphpair)), (and)) SDNode *Parts[4] = {}; if (isBSwapHWordPair(N0, Parts)) { // (or (or (and), (and)), (or (and), (and))) if (!isBSwapHWordPair(N1, Parts)) return SDValue(); } else if (N0.getOpcode() == ISD::OR) { // (or (or (or (and), (and)), (and)), (and)) if (!isBSwapHWordElement(N1, Parts)) return SDValue(); SDValue N00 = N0.getOperand(0); SDValue N01 = N0.getOperand(1); if (!(isBSwapHWordElement(N01, Parts) && isBSwapHWordPair(N00, Parts)) && !(isBSwapHWordElement(N00, Parts) && isBSwapHWordPair(N01, Parts))) return SDValue(); } else { return SDValue(); } // Make sure the parts are all coming from the same node. if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3]) return SDValue(); SDLoc DL(N); SDValue BSwap = DAG.getNode(ISD::BSWAP, DL, VT, SDValue(Parts[0], 0)); // Result of the bswap should be rotated by 16. If it's not legal, then // do (x << 16) | (x >> 16). SDValue ShAmt = DAG.getConstant(16, DL, getShiftAmountTy(VT)); if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT)) return DAG.getNode(ISD::ROTL, DL, VT, BSwap, ShAmt); if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT)) return DAG.getNode(ISD::ROTR, DL, VT, BSwap, ShAmt); return DAG.getNode(ISD::OR, DL, VT, DAG.getNode(ISD::SHL, DL, VT, BSwap, ShAmt), DAG.getNode(ISD::SRL, DL, VT, BSwap, ShAmt)); } /// This contains all DAGCombine rules which reduce two values combined by /// an Or operation to a single value \see visitANDLike(). SDValue DAGCombiner::visitORLike(SDValue N0, SDValue N1, SDNode *N) { EVT VT = N1.getValueType(); SDLoc DL(N); // fold (or x, undef) -> -1 if (!LegalOperations && (N0.isUndef() || N1.isUndef())) return DAG.getAllOnesConstant(DL, VT); if (SDValue V = foldLogicOfSetCCs(false, N0, N1, DL)) return V; // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible. if (N0.getOpcode() == ISD::AND && N1.getOpcode() == ISD::AND && // Don't increase # computations. (N0->hasOneUse() || N1->hasOneUse())) { // We can only do this xform if we know that bits from X that are set in C2 // but not in C1 are already zero. Likewise for Y. if (const ConstantSDNode *N0O1C = getAsNonOpaqueConstant(N0.getOperand(1))) { if (const ConstantSDNode *N1O1C = getAsNonOpaqueConstant(N1.getOperand(1))) { // We can only do this xform if we know that bits from X that are set in // C2 but not in C1 are already zero. Likewise for Y. const APInt &LHSMask = N0O1C->getAPIntValue(); const APInt &RHSMask = N1O1C->getAPIntValue(); if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) && DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) { SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT, N0.getOperand(0), N1.getOperand(0)); return DAG.getNode(ISD::AND, DL, VT, X, DAG.getConstant(LHSMask | RHSMask, DL, VT)); } } } } // (or (and X, M), (and X, N)) -> (and X, (or M, N)) if (N0.getOpcode() == ISD::AND && N1.getOpcode() == ISD::AND && N0.getOperand(0) == N1.getOperand(0) && // Don't increase # computations. (N0->hasOneUse() || N1->hasOneUse())) { SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT, N0.getOperand(1), N1.getOperand(1)); return DAG.getNode(ISD::AND, DL, VT, N0.getOperand(0), X); } return SDValue(); } /// OR combines for which the commuted variant will be tried as well. static SDValue visitORCommutative(SelectionDAG &DAG, SDValue N0, SDValue N1, SDNode *N) { EVT VT = N0.getValueType(); if (N0.getOpcode() == ISD::AND) { SDValue N00 = N0.getOperand(0); SDValue N01 = N0.getOperand(1); // fold (or (and X, (xor Y, -1)), Y) -> (or X, Y) // TODO: Set AllowUndefs = true. if (getBitwiseNotOperand(N01, N00, /* AllowUndefs */ false) == N1) return DAG.getNode(ISD::OR, SDLoc(N), VT, N00, N1); // fold (or (and (xor Y, -1), X), Y) -> (or X, Y) if (getBitwiseNotOperand(N00, N01, /* AllowUndefs */ false) == N1) return DAG.getNode(ISD::OR, SDLoc(N), VT, N01, N1); } if (SDValue R = foldLogicOfShifts(N, N0, N1, DAG)) return R; auto peekThroughZext = [](SDValue V) { if (V->getOpcode() == ISD::ZERO_EXTEND) return V->getOperand(0); return V; }; // (fshl X, ?, Y) | (shl X, Y) --> fshl X, ?, Y if (N0.getOpcode() == ISD::FSHL && N1.getOpcode() == ISD::SHL && N0.getOperand(0) == N1.getOperand(0) && peekThroughZext(N0.getOperand(2)) == peekThroughZext(N1.getOperand(1))) return N0; // (fshr ?, X, Y) | (srl X, Y) --> fshr ?, X, Y if (N0.getOpcode() == ISD::FSHR && N1.getOpcode() == ISD::SRL && N0.getOperand(1) == N1.getOperand(0) && peekThroughZext(N0.getOperand(2)) == peekThroughZext(N1.getOperand(1))) return N0; return SDValue(); } SDValue DAGCombiner::visitOR(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N1.getValueType(); // x | x --> x if (N0 == N1) return N0; // fold (or c1, c2) -> c1|c2 if (SDValue C = DAG.FoldConstantArithmetic(ISD::OR, SDLoc(N), VT, {N0, N1})) return C; // canonicalize constant to RHS if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && !DAG.isConstantIntBuildVectorOrConstantInt(N1)) return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0); // fold vector ops if (VT.isVector()) { if (SDValue FoldedVOp = SimplifyVBinOp(N, SDLoc(N))) return FoldedVOp; // fold (or x, 0) -> x, vector edition if (ISD::isConstantSplatVectorAllZeros(N1.getNode())) return N0; // fold (or x, -1) -> -1, vector edition if (ISD::isConstantSplatVectorAllOnes(N1.getNode())) // do not return N1, because undef node may exist in N1 return DAG.getAllOnesConstant(SDLoc(N), N1.getValueType()); // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask) // Do this only if the resulting type / shuffle is legal. auto *SV0 = dyn_cast(N0); auto *SV1 = dyn_cast(N1); if (SV0 && SV1 && TLI.isTypeLegal(VT)) { bool ZeroN00 = ISD::isBuildVectorAllZeros(N0.getOperand(0).getNode()); bool ZeroN01 = ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode()); bool ZeroN10 = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode()); bool ZeroN11 = ISD::isBuildVectorAllZeros(N1.getOperand(1).getNode()); // Ensure both shuffles have a zero input. if ((ZeroN00 != ZeroN01) && (ZeroN10 != ZeroN11)) { assert((!ZeroN00 || !ZeroN01) && "Both inputs zero!"); assert((!ZeroN10 || !ZeroN11) && "Both inputs zero!"); bool CanFold = true; int NumElts = VT.getVectorNumElements(); SmallVector Mask(NumElts, -1); for (int i = 0; i != NumElts; ++i) { int M0 = SV0->getMaskElt(i); int M1 = SV1->getMaskElt(i); // Determine if either index is pointing to a zero vector. bool M0Zero = M0 < 0 || (ZeroN00 == (M0 < NumElts)); bool M1Zero = M1 < 0 || (ZeroN10 == (M1 < NumElts)); // If one element is zero and the otherside is undef, keep undef. // This also handles the case that both are undef. if ((M0Zero && M1 < 0) || (M1Zero && M0 < 0)) continue; // Make sure only one of the elements is zero. if (M0Zero == M1Zero) { CanFold = false; break; } assert((M0 >= 0 || M1 >= 0) && "Undef index!"); // We have a zero and non-zero element. If the non-zero came from // SV0 make the index a LHS index. If it came from SV1, make it // a RHS index. We need to mod by NumElts because we don't care // which operand it came from in the original shuffles. Mask[i] = M1Zero ? M0 % NumElts : (M1 % NumElts) + NumElts; } if (CanFold) { SDValue NewLHS = ZeroN00 ? N0.getOperand(1) : N0.getOperand(0); SDValue NewRHS = ZeroN10 ? N1.getOperand(1) : N1.getOperand(0); SDValue LegalShuffle = TLI.buildLegalVectorShuffle(VT, SDLoc(N), NewLHS, NewRHS, Mask, DAG); if (LegalShuffle) return LegalShuffle; } } } } // fold (or x, 0) -> x if (isNullConstant(N1)) return N0; // fold (or x, -1) -> -1 if (isAllOnesConstant(N1)) return N1; if (SDValue NewSel = foldBinOpIntoSelect(N)) return NewSel; // fold (or x, c) -> c iff (x & ~c) == 0 ConstantSDNode *N1C = dyn_cast(N1); if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue())) return N1; if (SDValue Combined = visitORLike(N0, N1, N)) return Combined; if (SDValue Combined = combineCarryDiamond(DAG, TLI, N0, N1, N)) return Combined; // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16) if (SDValue BSwap = MatchBSwapHWord(N, N0, N1)) return BSwap; if (SDValue BSwap = MatchBSwapHWordLow(N, N0, N1)) return BSwap; // reassociate or if (SDValue ROR = reassociateOps(ISD::OR, SDLoc(N), N0, N1, N->getFlags())) return ROR; // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2) // iff (c1 & c2) != 0 or c1/c2 are undef. auto MatchIntersect = [](ConstantSDNode *C1, ConstantSDNode *C2) { return !C1 || !C2 || C1->getAPIntValue().intersects(C2->getAPIntValue()); }; if (N0.getOpcode() == ISD::AND && N0->hasOneUse() && ISD::matchBinaryPredicate(N0.getOperand(1), N1, MatchIntersect, true)) { if (SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, SDLoc(N1), VT, {N1, N0.getOperand(1)})) { SDValue IOR = DAG.getNode(ISD::OR, SDLoc(N0), VT, N0.getOperand(0), N1); AddToWorklist(IOR.getNode()); return DAG.getNode(ISD::AND, SDLoc(N), VT, COR, IOR); } } if (SDValue Combined = visitORCommutative(DAG, N0, N1, N)) return Combined; if (SDValue Combined = visitORCommutative(DAG, N1, N0, N)) return Combined; // Simplify: (or (op x...), (op y...)) -> (op (or x, y)) if (N0.getOpcode() == N1.getOpcode()) if (SDValue V = hoistLogicOpWithSameOpcodeHands(N)) return V; // See if this is some rotate idiom. if (SDValue Rot = MatchRotate(N0, N1, SDLoc(N))) return Rot; if (SDValue Load = MatchLoadCombine(N)) return Load; // Simplify the operands using demanded-bits information. if (SimplifyDemandedBits(SDValue(N, 0))) return SDValue(N, 0); // If OR can be rewritten into ADD, try combines based on ADD. if ((!LegalOperations || TLI.isOperationLegal(ISD::ADD, VT)) && DAG.haveNoCommonBitsSet(N0, N1)) if (SDValue Combined = visitADDLike(N)) return Combined; return SDValue(); } static SDValue stripConstantMask(SelectionDAG &DAG, SDValue Op, SDValue &Mask) { if (Op.getOpcode() == ISD::AND && DAG.isConstantIntBuildVectorOrConstantInt(Op.getOperand(1))) { Mask = Op.getOperand(1); return Op.getOperand(0); } return Op; } /// Match "(X shl/srl V1) & V2" where V2 may not be present. static bool matchRotateHalf(SelectionDAG &DAG, SDValue Op, SDValue &Shift, SDValue &Mask) { Op = stripConstantMask(DAG, Op, Mask); if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) { Shift = Op; return true; } return false; } /// Helper function for visitOR to extract the needed side of a rotate idiom /// from a shl/srl/mul/udiv. This is meant to handle cases where /// InstCombine merged some outside op with one of the shifts from /// the rotate pattern. /// \returns An empty \c SDValue if the needed shift couldn't be extracted. /// Otherwise, returns an expansion of \p ExtractFrom based on the following /// patterns: /// /// (or (add v v) (shrl v bitwidth-1)): /// expands (add v v) -> (shl v 1) /// /// (or (mul v c0) (shrl (mul v c1) c2)): /// expands (mul v c0) -> (shl (mul v c1) c3) /// /// (or (udiv v c0) (shl (udiv v c1) c2)): /// expands (udiv v c0) -> (shrl (udiv v c1) c3) /// /// (or (shl v c0) (shrl (shl v c1) c2)): /// expands (shl v c0) -> (shl (shl v c1) c3) /// /// (or (shrl v c0) (shl (shrl v c1) c2)): /// expands (shrl v c0) -> (shrl (shrl v c1) c3) /// /// Such that in all cases, c3+c2==bitwidth(op v c1). static SDValue extractShiftForRotate(SelectionDAG &DAG, SDValue OppShift, SDValue ExtractFrom, SDValue &Mask, const SDLoc &DL) { assert(OppShift && ExtractFrom && "Empty SDValue"); if (OppShift.getOpcode() != ISD::SHL && OppShift.getOpcode() != ISD::SRL) return SDValue(); ExtractFrom = stripConstantMask(DAG, ExtractFrom, Mask); // Value and Type of the shift. SDValue OppShiftLHS = OppShift.getOperand(0); EVT ShiftedVT = OppShiftLHS.getValueType(); // Amount of the existing shift. ConstantSDNode *OppShiftCst = isConstOrConstSplat(OppShift.getOperand(1)); // (add v v) -> (shl v 1) // TODO: Should this be a general DAG canonicalization? if (OppShift.getOpcode() == ISD::SRL && OppShiftCst && ExtractFrom.getOpcode() == ISD::ADD && ExtractFrom.getOperand(0) == ExtractFrom.getOperand(1) && ExtractFrom.getOperand(0) == OppShiftLHS && OppShiftCst->getAPIntValue() == ShiftedVT.getScalarSizeInBits() - 1) return DAG.getNode(ISD::SHL, DL, ShiftedVT, OppShiftLHS, DAG.getShiftAmountConstant(1, ShiftedVT, DL)); // Preconditions: // (or (op0 v c0) (shiftl/r (op0 v c1) c2)) // // Find opcode of the needed shift to be extracted from (op0 v c0). unsigned Opcode = ISD::DELETED_NODE; bool IsMulOrDiv = false; // Set Opcode and IsMulOrDiv if the extract opcode matches the needed shift // opcode or its arithmetic (mul or udiv) variant. auto SelectOpcode = [&](unsigned NeededShift, unsigned MulOrDivVariant) { IsMulOrDiv = ExtractFrom.getOpcode() == MulOrDivVariant; if (!IsMulOrDiv && ExtractFrom.getOpcode() != NeededShift) return false; Opcode = NeededShift; return true; }; // op0 must be either the needed shift opcode or the mul/udiv equivalent // that the needed shift can be extracted from. if ((OppShift.getOpcode() != ISD::SRL || !SelectOpcode(ISD::SHL, ISD::MUL)) && (OppShift.getOpcode() != ISD::SHL || !SelectOpcode(ISD::SRL, ISD::UDIV))) return SDValue(); // op0 must be the same opcode on both sides, have the same LHS argument, // and produce the same value type. if (OppShiftLHS.getOpcode() != ExtractFrom.getOpcode() || OppShiftLHS.getOperand(0) != ExtractFrom.getOperand(0) || ShiftedVT != ExtractFrom.getValueType()) return SDValue(); // Constant mul/udiv/shift amount from the RHS of the shift's LHS op. ConstantSDNode *OppLHSCst = isConstOrConstSplat(OppShiftLHS.getOperand(1)); // Constant mul/udiv/shift amount from the RHS of the ExtractFrom op. ConstantSDNode *ExtractFromCst = isConstOrConstSplat(ExtractFrom.getOperand(1)); // TODO: We should be able to handle non-uniform constant vectors for these values // Check that we have constant values. if (!OppShiftCst || !OppShiftCst->getAPIntValue() || !OppLHSCst || !OppLHSCst->getAPIntValue() || !ExtractFromCst || !ExtractFromCst->getAPIntValue()) return SDValue(); // Compute the shift amount we need to extract to complete the rotate. const unsigned VTWidth = ShiftedVT.getScalarSizeInBits(); if (OppShiftCst->getAPIntValue().ugt(VTWidth)) return SDValue(); APInt NeededShiftAmt = VTWidth - OppShiftCst->getAPIntValue(); // Normalize the bitwidth of the two mul/udiv/shift constant operands. APInt ExtractFromAmt = ExtractFromCst->getAPIntValue(); APInt OppLHSAmt = OppLHSCst->getAPIntValue(); zeroExtendToMatch(ExtractFromAmt, OppLHSAmt); // Now try extract the needed shift from the ExtractFrom op and see if the // result matches up with the existing shift's LHS op. if (IsMulOrDiv) { // Op to extract from is a mul or udiv by a constant. // Check: // c2 / (1 << (bitwidth(op0 v c0) - c1)) == c0 // c2 % (1 << (bitwidth(op0 v c0) - c1)) == 0 const APInt ExtractDiv = APInt::getOneBitSet(ExtractFromAmt.getBitWidth(), NeededShiftAmt.getZExtValue()); APInt ResultAmt; APInt Rem; APInt::udivrem(ExtractFromAmt, ExtractDiv, ResultAmt, Rem); if (Rem != 0 || ResultAmt != OppLHSAmt) return SDValue(); } else { // Op to extract from is a shift by a constant. // Check: // c2 - (bitwidth(op0 v c0) - c1) == c0 if (OppLHSAmt != ExtractFromAmt - NeededShiftAmt.zextOrTrunc( ExtractFromAmt.getBitWidth())) return SDValue(); } // Return the expanded shift op that should allow a rotate to be formed. EVT ShiftVT = OppShift.getOperand(1).getValueType(); EVT ResVT = ExtractFrom.getValueType(); SDValue NewShiftNode = DAG.getConstant(NeededShiftAmt, DL, ShiftVT); return DAG.getNode(Opcode, DL, ResVT, OppShiftLHS, NewShiftNode); } // Return true if we can prove that, whenever Neg and Pos are both in the // range [0, EltSize), Neg == (Pos == 0 ? 0 : EltSize - Pos). This means that // for two opposing shifts shift1 and shift2 and a value X with OpBits bits: // // (or (shift1 X, Neg), (shift2 X, Pos)) // // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate // in direction shift1 by Neg. The range [0, EltSize) means that we only need // to consider shift amounts with defined behavior. // // The IsRotate flag should be set when the LHS of both shifts is the same. // Otherwise if matching a general funnel shift, it should be clear. static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned EltSize, SelectionDAG &DAG, bool IsRotate) { const auto &TLI = DAG.getTargetLoweringInfo(); // If EltSize is a power of 2 then: // // (a) (Pos == 0 ? 0 : EltSize - Pos) == (EltSize - Pos) & (EltSize - 1) // (b) Neg == Neg & (EltSize - 1) whenever Neg is in [0, EltSize). // // So if EltSize is a power of 2 and Neg is (and Neg', EltSize-1), we check // for the stronger condition: // // Neg & (EltSize - 1) == (EltSize - Pos) & (EltSize - 1) [A] // // for all Neg and Pos. Since Neg & (EltSize - 1) == Neg' & (EltSize - 1) // we can just replace Neg with Neg' for the rest of the function. // // In other cases we check for the even stronger condition: // // Neg == EltSize - Pos [B] // // for all Neg and Pos. Note that the (or ...) then invokes undefined // behavior if Pos == 0 (and consequently Neg == EltSize). // // We could actually use [A] whenever EltSize is a power of 2, but the // only extra cases that it would match are those uninteresting ones // where Neg and Pos are never in range at the same time. E.g. for // EltSize == 32, using [A] would allow a Neg of the form (sub 64, Pos) // as well as (sub 32, Pos), but: // // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos)) // // always invokes undefined behavior for 32-bit X. // // Below, Mask == EltSize - 1 when using [A] and is all-ones otherwise. // This allows us to peek through any operations that only affect Mask's // un-demanded bits. // // NOTE: We can only do this when matching operations which won't modify the // least Log2(EltSize) significant bits and not a general funnel shift. unsigned MaskLoBits = 0; if (IsRotate && isPowerOf2_64(EltSize)) { unsigned Bits = Log2_64(EltSize); unsigned NegBits = Neg.getScalarValueSizeInBits(); if (NegBits >= Bits) { APInt DemandedBits = APInt::getLowBitsSet(NegBits, Bits); if (SDValue Inner = TLI.SimplifyMultipleUseDemandedBits(Neg, DemandedBits, DAG)) { Neg = Inner; MaskLoBits = Bits; } } } // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1. if (Neg.getOpcode() != ISD::SUB) return false; ConstantSDNode *NegC = isConstOrConstSplat(Neg.getOperand(0)); if (!NegC) return false; SDValue NegOp1 = Neg.getOperand(1); // On the RHS of [A], if Pos is the result of operation on Pos' that won't // affect Mask's demanded bits, just replace Pos with Pos'. These operations // are redundant for the purpose of the equality. if (MaskLoBits) { unsigned PosBits = Pos.getScalarValueSizeInBits(); if (PosBits >= MaskLoBits) { APInt DemandedBits = APInt::getLowBitsSet(PosBits, MaskLoBits); if (SDValue Inner = TLI.SimplifyMultipleUseDemandedBits(Pos, DemandedBits, DAG)) { Pos = Inner; } } } // The condition we need is now: // // (NegC - NegOp1) & Mask == (EltSize - Pos) & Mask // // If NegOp1 == Pos then we need: // // EltSize & Mask == NegC & Mask // // (because "x & Mask" is a truncation and distributes through subtraction). // // We also need to account for a potential truncation of NegOp1 if the amount // has already been legalized to a shift amount type. APInt Width; if ((Pos == NegOp1) || (NegOp1.getOpcode() == ISD::TRUNCATE && Pos == NegOp1.getOperand(0))) Width = NegC->getAPIntValue(); // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC. // Then the condition we want to prove becomes: // // (NegC - NegOp1) & Mask == (EltSize - (NegOp1 + PosC)) & Mask // // which, again because "x & Mask" is a truncation, becomes: // // NegC & Mask == (EltSize - PosC) & Mask // EltSize & Mask == (NegC + PosC) & Mask else if (Pos.getOpcode() == ISD::ADD && Pos.getOperand(0) == NegOp1) { if (ConstantSDNode *PosC = isConstOrConstSplat(Pos.getOperand(1))) Width = PosC->getAPIntValue() + NegC->getAPIntValue(); else return false; } else return false; // Now we just need to check that EltSize & Mask == Width & Mask. if (MaskLoBits) // EltSize & Mask is 0 since Mask is EltSize - 1. return Width.getLoBits(MaskLoBits) == 0; return Width == EltSize; } // A subroutine of MatchRotate used once we have found an OR of two opposite // shifts of Shifted. If Neg == - Pos then the OR reduces // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the // former being preferred if supported. InnerPos and InnerNeg are Pos and // Neg with outer conversions stripped away. SDValue DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg, SDValue InnerPos, SDValue InnerNeg, bool HasPos, unsigned PosOpcode, unsigned NegOpcode, const SDLoc &DL) { // fold (or (shl x, (*ext y)), // (srl x, (*ext (sub 32, y)))) -> // (rotl x, y) or (rotr x, (sub 32, y)) // // fold (or (shl x, (*ext (sub 32, y))), // (srl x, (*ext y))) -> // (rotr x, y) or (rotl x, (sub 32, y)) EVT VT = Shifted.getValueType(); if (matchRotateSub(InnerPos, InnerNeg, VT.getScalarSizeInBits(), DAG, /*IsRotate*/ true)) { return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted, HasPos ? Pos : Neg); } return SDValue(); } // A subroutine of MatchRotate used once we have found an OR of two opposite // shifts of N0 + N1. If Neg == - Pos then the OR reduces // to both (PosOpcode N0, N1, Pos) and (NegOpcode N0, N1, Neg), with the // former being preferred if supported. InnerPos and InnerNeg are Pos and // Neg with outer conversions stripped away. // TODO: Merge with MatchRotatePosNeg. SDValue DAGCombiner::MatchFunnelPosNeg(SDValue N0, SDValue N1, SDValue Pos, SDValue Neg, SDValue InnerPos, SDValue InnerNeg, bool HasPos, unsigned PosOpcode, unsigned NegOpcode, const SDLoc &DL) { EVT VT = N0.getValueType(); unsigned EltBits = VT.getScalarSizeInBits(); // fold (or (shl x0, (*ext y)), // (srl x1, (*ext (sub 32, y)))) -> // (fshl x0, x1, y) or (fshr x0, x1, (sub 32, y)) // // fold (or (shl x0, (*ext (sub 32, y))), // (srl x1, (*ext y))) -> // (fshr x0, x1, y) or (fshl x0, x1, (sub 32, y)) if (matchRotateSub(InnerPos, InnerNeg, EltBits, DAG, /*IsRotate*/ N0 == N1)) { return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, N0, N1, HasPos ? Pos : Neg); } // Matching the shift+xor cases, we can't easily use the xor'd shift amount // so for now just use the PosOpcode case if its legal. // TODO: When can we use the NegOpcode case? if (PosOpcode == ISD::FSHL && isPowerOf2_32(EltBits)) { auto IsBinOpImm = [](SDValue Op, unsigned BinOpc, unsigned Imm) { if (Op.getOpcode() != BinOpc) return false; ConstantSDNode *Cst = isConstOrConstSplat(Op.getOperand(1)); return Cst && (Cst->getAPIntValue() == Imm); }; // fold (or (shl x0, y), (srl (srl x1, 1), (xor y, 31))) // -> (fshl x0, x1, y) if (IsBinOpImm(N1, ISD::SRL, 1) && IsBinOpImm(InnerNeg, ISD::XOR, EltBits - 1) && InnerPos == InnerNeg.getOperand(0) && TLI.isOperationLegalOrCustom(ISD::FSHL, VT)) { return DAG.getNode(ISD::FSHL, DL, VT, N0, N1.getOperand(0), Pos); } // fold (or (shl (shl x0, 1), (xor y, 31)), (srl x1, y)) // -> (fshr x0, x1, y) if (IsBinOpImm(N0, ISD::SHL, 1) && IsBinOpImm(InnerPos, ISD::XOR, EltBits - 1) && InnerNeg == InnerPos.getOperand(0) && TLI.isOperationLegalOrCustom(ISD::FSHR, VT)) { return DAG.getNode(ISD::FSHR, DL, VT, N0.getOperand(0), N1, Neg); } // fold (or (shl (add x0, x0), (xor y, 31)), (srl x1, y)) // -> (fshr x0, x1, y) // TODO: Should add(x,x) -> shl(x,1) be a general DAG canonicalization? if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N0.getOperand(1) && IsBinOpImm(InnerPos, ISD::XOR, EltBits - 1) && InnerNeg == InnerPos.getOperand(0) && TLI.isOperationLegalOrCustom(ISD::FSHR, VT)) { return DAG.getNode(ISD::FSHR, DL, VT, N0.getOperand(0), N1, Neg); } } return SDValue(); } // MatchRotate - Handle an 'or' of two operands. If this is one of the many // idioms for rotate, and if the target supports rotation instructions, generate // a rot[lr]. This also matches funnel shift patterns, similar to rotation but // with different shifted sources. SDValue DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, const SDLoc &DL) { EVT VT = LHS.getValueType(); // The target must have at least one rotate/funnel flavor. // We still try to match rotate by constant pre-legalization. // TODO: Support pre-legalization funnel-shift by constant. bool HasROTL = hasOperation(ISD::ROTL, VT); bool HasROTR = hasOperation(ISD::ROTR, VT); bool HasFSHL = hasOperation(ISD::FSHL, VT); bool HasFSHR = hasOperation(ISD::FSHR, VT); // If the type is going to be promoted and the target has enabled custom // lowering for rotate, allow matching rotate by non-constants. Only allow // this for scalar types. if (VT.isScalarInteger() && TLI.getTypeAction(*DAG.getContext(), VT) == TargetLowering::TypePromoteInteger) { HasROTL |= TLI.getOperationAction(ISD::ROTL, VT) == TargetLowering::Custom; HasROTR |= TLI.getOperationAction(ISD::ROTR, VT) == TargetLowering::Custom; } if (LegalOperations && !HasROTL && !HasROTR && !HasFSHL && !HasFSHR) return SDValue(); // Check for truncated rotate. if (LHS.getOpcode() == ISD::TRUNCATE && RHS.getOpcode() == ISD::TRUNCATE && LHS.getOperand(0).getValueType() == RHS.getOperand(0).getValueType()) { assert(LHS.getValueType() == RHS.getValueType()); if (SDValue Rot = MatchRotate(LHS.getOperand(0), RHS.getOperand(0), DL)) { return DAG.getNode(ISD::TRUNCATE, SDLoc(LHS), LHS.getValueType(), Rot); } } // Match "(X shl/srl V1) & V2" where V2 may not be present. SDValue LHSShift; // The shift. SDValue LHSMask; // AND value if any. matchRotateHalf(DAG, LHS, LHSShift, LHSMask); SDValue RHSShift; // The shift. SDValue RHSMask; // AND value if any. matchRotateHalf(DAG, RHS, RHSShift, RHSMask); // If neither side matched a rotate half, bail if (!LHSShift && !RHSShift) return SDValue(); // InstCombine may have combined a constant shl, srl, mul, or udiv with one // side of the rotate, so try to handle that here. In all cases we need to // pass the matched shift from the opposite side to compute the opcode and // needed shift amount to extract. We still want to do this if both sides // matched a rotate half because one half may be a potential overshift that // can be broken down (ie if InstCombine merged two shl or srl ops into a // single one). // Have LHS side of the rotate, try to extract the needed shift from the RHS. if (LHSShift) if (SDValue NewRHSShift = extractShiftForRotate(DAG, LHSShift, RHS, RHSMask, DL)) RHSShift = NewRHSShift; // Have RHS side of the rotate, try to extract the needed shift from the LHS. if (RHSShift) if (SDValue NewLHSShift = extractShiftForRotate(DAG, RHSShift, LHS, LHSMask, DL)) LHSShift = NewLHSShift; // If a side is still missing, nothing else we can do. if (!RHSShift || !LHSShift) return SDValue(); // At this point we've matched or extracted a shift op on each side. if (LHSShift.getOpcode() == RHSShift.getOpcode()) return SDValue(); // Shifts must disagree. // Canonicalize shl to left side in a shl/srl pair. if (RHSShift.getOpcode() == ISD::SHL) { std::swap(LHS, RHS); std::swap(LHSShift, RHSShift); std::swap(LHSMask, RHSMask); } // Something has gone wrong - we've lost the shl/srl pair - bail. if (LHSShift.getOpcode() != ISD::SHL || RHSShift.getOpcode() != ISD::SRL) return SDValue(); unsigned EltSizeInBits = VT.getScalarSizeInBits(); SDValue LHSShiftArg = LHSShift.getOperand(0); SDValue LHSShiftAmt = LHSShift.getOperand(1); SDValue RHSShiftArg = RHSShift.getOperand(0); SDValue RHSShiftAmt = RHSShift.getOperand(1); auto MatchRotateSum = [EltSizeInBits](ConstantSDNode *LHS, ConstantSDNode *RHS) { return (LHS->getAPIntValue() + RHS->getAPIntValue()) == EltSizeInBits; }; auto ApplyMasks = [&](SDValue Res) { // If there is an AND of either shifted operand, apply it to the result. if (LHSMask.getNode() || RHSMask.getNode()) { SDValue AllOnes = DAG.getAllOnesConstant(DL, VT); SDValue Mask = AllOnes; if (LHSMask.getNode()) { SDValue RHSBits = DAG.getNode(ISD::SRL, DL, VT, AllOnes, RHSShiftAmt); Mask = DAG.getNode(ISD::AND, DL, VT, Mask, DAG.getNode(ISD::OR, DL, VT, LHSMask, RHSBits)); } if (RHSMask.getNode()) { SDValue LHSBits = DAG.getNode(ISD::SHL, DL, VT, AllOnes, LHSShiftAmt); Mask = DAG.getNode(ISD::AND, DL, VT, Mask, DAG.getNode(ISD::OR, DL, VT, RHSMask, LHSBits)); } Res = DAG.getNode(ISD::AND, DL, VT, Res, Mask); } return Res; }; // TODO: Support pre-legalization funnel-shift by constant. bool IsRotate = LHSShift.getOperand(0) == RHSShift.getOperand(0); if (!IsRotate && !(HasFSHL || HasFSHR)) { if (TLI.isTypeLegal(VT) && LHS.hasOneUse() && RHS.hasOneUse() && ISD::matchBinaryPredicate(LHSShiftAmt, RHSShiftAmt, MatchRotateSum)) { // Look for a disguised rotate by constant. // The common shifted operand X may be hidden inside another 'or'. SDValue X, Y; auto matchOr = [&X, &Y](SDValue Or, SDValue CommonOp) { if (!Or.hasOneUse() || Or.getOpcode() != ISD::OR) return false; if (CommonOp == Or.getOperand(0)) { X = CommonOp; Y = Or.getOperand(1); return true; } if (CommonOp == Or.getOperand(1)) { X = CommonOp; Y = Or.getOperand(0); return true; } return false; }; SDValue Res; if (matchOr(LHSShiftArg, RHSShiftArg)) { // (shl (X | Y), C1) | (srl X, C2) --> (rotl X, C1) | (shl Y, C1) SDValue RotX = DAG.getNode(ISD::ROTL, DL, VT, X, LHSShiftAmt); SDValue ShlY = DAG.getNode(ISD::SHL, DL, VT, Y, LHSShiftAmt); Res = DAG.getNode(ISD::OR, DL, VT, RotX, ShlY); } else if (matchOr(RHSShiftArg, LHSShiftArg)) { // (shl X, C1) | (srl (X | Y), C2) --> (rotl X, C1) | (srl Y, C2) SDValue RotX = DAG.getNode(ISD::ROTL, DL, VT, X, LHSShiftAmt); SDValue SrlY = DAG.getNode(ISD::SRL, DL, VT, Y, RHSShiftAmt); Res = DAG.getNode(ISD::OR, DL, VT, RotX, SrlY); } else { return SDValue(); } return ApplyMasks(Res); } return SDValue(); // Requires funnel shift support. } // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1) // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2) // fold (or (shl x, C1), (srl y, C2)) -> (fshl x, y, C1) // fold (or (shl x, C1), (srl y, C2)) -> (fshr x, y, C2) // iff C1+C2 == EltSizeInBits if (ISD::matchBinaryPredicate(LHSShiftAmt, RHSShiftAmt, MatchRotateSum)) { SDValue Res; if (IsRotate && (HasROTL || HasROTR || !(HasFSHL || HasFSHR))) { bool UseROTL = !LegalOperations || HasROTL; Res = DAG.getNode(UseROTL ? ISD::ROTL : ISD::ROTR, DL, VT, LHSShiftArg, UseROTL ? LHSShiftAmt : RHSShiftAmt); } else { bool UseFSHL = !LegalOperations || HasFSHL; Res = DAG.getNode(UseFSHL ? ISD::FSHL : ISD::FSHR, DL, VT, LHSShiftArg, RHSShiftArg, UseFSHL ? LHSShiftAmt : RHSShiftAmt); } return ApplyMasks(Res); } // Even pre-legalization, we can't easily rotate/funnel-shift by a variable // shift. if (!HasROTL && !HasROTR && !HasFSHL && !HasFSHR) return SDValue(); // If there is a mask here, and we have a variable shift, we can't be sure // that we're masking out the right stuff. if (LHSMask.getNode() || RHSMask.getNode()) return SDValue(); // If the shift amount is sign/zext/any-extended just peel it off. SDValue LExtOp0 = LHSShiftAmt; SDValue RExtOp0 = RHSShiftAmt; if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND || LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND || LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND || LHSShiftAmt.getOpcode() == ISD::TRUNCATE) && (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND || RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND || RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND || RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) { LExtOp0 = LHSShiftAmt.getOperand(0); RExtOp0 = RHSShiftAmt.getOperand(0); } if (IsRotate && (HasROTL || HasROTR)) { SDValue TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt, LExtOp0, RExtOp0, HasROTL, ISD::ROTL, ISD::ROTR, DL); if (TryL) return TryL; SDValue TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt, RExtOp0, LExtOp0, HasROTR, ISD::ROTR, ISD::ROTL, DL); if (TryR) return TryR; } SDValue TryL = MatchFunnelPosNeg(LHSShiftArg, RHSShiftArg, LHSShiftAmt, RHSShiftAmt, LExtOp0, RExtOp0, HasFSHL, ISD::FSHL, ISD::FSHR, DL); if (TryL) return TryL; SDValue TryR = MatchFunnelPosNeg(LHSShiftArg, RHSShiftArg, RHSShiftAmt, LHSShiftAmt, RExtOp0, LExtOp0, HasFSHR, ISD::FSHR, ISD::FSHL, DL); if (TryR) return TryR; return SDValue(); } namespace { /// Represents known origin of an individual byte in load combine pattern. The /// value of the byte is either constant zero or comes from memory. struct ByteProvider { // For constant zero providers Load is set to nullptr. For memory providers // Load represents the node which loads the byte from memory. // ByteOffset is the offset of the byte in the value produced by the load. LoadSDNode *Load = nullptr; unsigned ByteOffset = 0; ByteProvider() = default; static ByteProvider getMemory(LoadSDNode *Load, unsigned ByteOffset) { return ByteProvider(Load, ByteOffset); } static ByteProvider getConstantZero() { return ByteProvider(nullptr, 0); } bool isConstantZero() const { return !Load; } bool isMemory() const { return Load; } bool operator==(const ByteProvider &Other) const { return Other.Load == Load && Other.ByteOffset == ByteOffset; } private: ByteProvider(LoadSDNode *Load, unsigned ByteOffset) : Load(Load), ByteOffset(ByteOffset) {} }; } // end anonymous namespace /// Recursively traverses the expression calculating the origin of the requested /// byte of the given value. Returns None if the provider can't be calculated. /// /// For all the values except the root of the expression verifies that the value /// has exactly one use and if it's not true return None. This way if the origin /// of the byte is returned it's guaranteed that the values which contribute to /// the byte are not used outside of this expression. /// /// Because the parts of the expression are not allowed to have more than one /// use this function iterates over trees, not DAGs. So it never visits the same /// node more than once. static const Optional calculateByteProvider(SDValue Op, unsigned Index, unsigned Depth, bool Root = false) { // Typical i64 by i8 pattern requires recursion up to 8 calls depth if (Depth == 10) return None; if (!Root && !Op.hasOneUse()) return None; assert(Op.getValueType().isScalarInteger() && "can't handle other types"); unsigned BitWidth = Op.getValueSizeInBits(); if (BitWidth % 8 != 0) return None; unsigned ByteWidth = BitWidth / 8; assert(Index < ByteWidth && "invalid index requested"); (void) ByteWidth; switch (Op.getOpcode()) { case ISD::OR: { auto LHS = calculateByteProvider(Op->getOperand(0), Index, Depth + 1); if (!LHS) return None; auto RHS = calculateByteProvider(Op->getOperand(1), Index, Depth + 1); if (!RHS) return None; if (LHS->isConstantZero()) return RHS; if (RHS->isConstantZero()) return LHS; return None; } case ISD::SHL: { auto ShiftOp = dyn_cast(Op->getOperand(1)); if (!ShiftOp) return None; uint64_t BitShift = ShiftOp->getZExtValue(); if (BitShift % 8 != 0) return None; uint64_t ByteShift = BitShift / 8; return Index < ByteShift ? ByteProvider::getConstantZero() : calculateByteProvider(Op->getOperand(0), Index - ByteShift, Depth + 1); } case ISD::ANY_EXTEND: case ISD::SIGN_EXTEND: case ISD::ZERO_EXTEND: { SDValue NarrowOp = Op->getOperand(0); unsigned NarrowBitWidth = NarrowOp.getScalarValueSizeInBits(); if (NarrowBitWidth % 8 != 0) return None; uint64_t NarrowByteWidth = NarrowBitWidth / 8; if (Index >= NarrowByteWidth) return Op.getOpcode() == ISD::ZERO_EXTEND ? Optional(ByteProvider::getConstantZero()) : None; return calculateByteProvider(NarrowOp, Index, Depth + 1); } case ISD::BSWAP: return calculateByteProvider(Op->getOperand(0), ByteWidth - Index - 1, Depth + 1); case ISD::LOAD: { auto L = cast(Op.getNode()); if (!L->isSimple() || L->isIndexed()) return None; unsigned NarrowBitWidth = L->getMemoryVT().getSizeInBits(); if (NarrowBitWidth % 8 != 0) return None; uint64_t NarrowByteWidth = NarrowBitWidth / 8; if (Index >= NarrowByteWidth) return L->getExtensionType() == ISD::ZEXTLOAD ? Optional(ByteProvider::getConstantZero()) : None; return ByteProvider::getMemory(L, Index); } } return None; } static unsigned littleEndianByteAt(unsigned BW, unsigned i) { return i; } static unsigned bigEndianByteAt(unsigned BW, unsigned i) { return BW - i - 1; } // Check if the bytes offsets we are looking at match with either big or // little endian value loaded. Return true for big endian, false for little // endian, and None if match failed. static Optional isBigEndian(const ArrayRef ByteOffsets, int64_t FirstOffset) { // The endian can be decided only when it is 2 bytes at least. unsigned Width = ByteOffsets.size(); if (Width < 2) return None; bool BigEndian = true, LittleEndian = true; for (unsigned i = 0; i < Width; i++) { int64_t CurrentByteOffset = ByteOffsets[i] - FirstOffset; LittleEndian &= CurrentByteOffset == littleEndianByteAt(Width, i); BigEndian &= CurrentByteOffset == bigEndianByteAt(Width, i); if (!BigEndian && !LittleEndian) return None; } assert((BigEndian != LittleEndian) && "It should be either big endian or" "little endian"); return BigEndian; } static SDValue stripTruncAndExt(SDValue Value) { switch (Value.getOpcode()) { case ISD::TRUNCATE: case ISD::ZERO_EXTEND: case ISD::SIGN_EXTEND: case ISD::ANY_EXTEND: return stripTruncAndExt(Value.getOperand(0)); } return Value; } /// Match a pattern where a wide type scalar value is stored by several narrow /// stores. Fold it into a single store or a BSWAP and a store if the targets /// supports it. /// /// Assuming little endian target: /// i8 *p = ... /// i32 val = ... /// p[0] = (val >> 0) & 0xFF; /// p[1] = (val >> 8) & 0xFF; /// p[2] = (val >> 16) & 0xFF; /// p[3] = (val >> 24) & 0xFF; /// => /// *((i32)p) = val; /// /// i8 *p = ... /// i32 val = ... /// p[0] = (val >> 24) & 0xFF; /// p[1] = (val >> 16) & 0xFF; /// p[2] = (val >> 8) & 0xFF; /// p[3] = (val >> 0) & 0xFF; /// => /// *((i32)p) = BSWAP(val); SDValue DAGCombiner::mergeTruncStores(StoreSDNode *N) { // The matching looks for "store (trunc x)" patterns that appear early but are // likely to be replaced by truncating store nodes during combining. // TODO: If there is evidence that running this later would help, this // limitation could be removed. Legality checks may need to be added // for the created store and optional bswap/rotate. if (LegalOperations || OptLevel == CodeGenOpt::None) return SDValue(); // We only handle merging simple stores of 1-4 bytes. // TODO: Allow unordered atomics when wider type is legal (see D66309) EVT MemVT = N->getMemoryVT(); if (!(MemVT == MVT::i8 || MemVT == MVT::i16 || MemVT == MVT::i32) || !N->isSimple() || N->isIndexed()) return SDValue(); // Collect all of the stores in the chain. SDValue Chain = N->getChain(); SmallVector Stores = {N}; while (auto *Store = dyn_cast(Chain)) { // All stores must be the same size to ensure that we are writing all of the // bytes in the wide value. // TODO: We could allow multiple sizes by tracking each stored byte. if (Store->getMemoryVT() != MemVT || !Store->isSimple() || Store->isIndexed()) return SDValue(); Stores.push_back(Store); Chain = Store->getChain(); } // There is no reason to continue if we do not have at least a pair of stores. if (Stores.size() < 2) return SDValue(); // Handle simple types only. LLVMContext &Context = *DAG.getContext(); unsigned NumStores = Stores.size(); unsigned NarrowNumBits = N->getMemoryVT().getScalarSizeInBits(); unsigned WideNumBits = NumStores * NarrowNumBits; EVT WideVT = EVT::getIntegerVT(Context, WideNumBits); if (WideVT != MVT::i16 && WideVT != MVT::i32 && WideVT != MVT::i64) return SDValue(); // Check if all bytes of the source value that we are looking at are stored // to the same base address. Collect offsets from Base address into OffsetMap. SDValue SourceValue; SmallVector OffsetMap(NumStores, INT64_MAX); int64_t FirstOffset = INT64_MAX; StoreSDNode *FirstStore = nullptr; Optional Base; for (auto *Store : Stores) { // All the stores store different parts of the CombinedValue. A truncate is // required to get the partial value. SDValue Trunc = Store->getValue(); if (Trunc.getOpcode() != ISD::TRUNCATE) return SDValue(); // Other than the first/last part, a shift operation is required to get the // offset. int64_t Offset = 0; SDValue WideVal = Trunc.getOperand(0); if ((WideVal.getOpcode() == ISD::SRL || WideVal.getOpcode() == ISD::SRA) && isa(WideVal.getOperand(1))) { // The shift amount must be a constant multiple of the narrow type. // It is translated to the offset address in the wide source value "y". // // x = srl y, ShiftAmtC // i8 z = trunc x // store z, ... uint64_t ShiftAmtC = WideVal.getConstantOperandVal(1); if (ShiftAmtC % NarrowNumBits != 0) return SDValue(); Offset = ShiftAmtC / NarrowNumBits; WideVal = WideVal.getOperand(0); } // Stores must share the same source value with different offsets. // Truncate and extends should be stripped to get the single source value. if (!SourceValue) SourceValue = WideVal; else if (stripTruncAndExt(SourceValue) != stripTruncAndExt(WideVal)) return SDValue(); else if (SourceValue.getValueType() != WideVT) { if (WideVal.getValueType() == WideVT || WideVal.getScalarValueSizeInBits() > SourceValue.getScalarValueSizeInBits()) SourceValue = WideVal; // Give up if the source value type is smaller than the store size. if (SourceValue.getScalarValueSizeInBits() < WideVT.getScalarSizeInBits()) return SDValue(); } // Stores must share the same base address. BaseIndexOffset Ptr = BaseIndexOffset::match(Store, DAG); int64_t ByteOffsetFromBase = 0; if (!Base) Base = Ptr; else if (!Base->equalBaseIndex(Ptr, DAG, ByteOffsetFromBase)) return SDValue(); // Remember the first store. if (ByteOffsetFromBase < FirstOffset) { FirstStore = Store; FirstOffset = ByteOffsetFromBase; } // Map the offset in the store and the offset in the combined value, and // early return if it has been set before. if (Offset < 0 || Offset >= NumStores || OffsetMap[Offset] != INT64_MAX) return SDValue(); OffsetMap[Offset] = ByteOffsetFromBase; } assert(FirstOffset != INT64_MAX && "First byte offset must be set"); assert(FirstStore && "First store must be set"); // Check that a store of the wide type is both allowed and fast on the target const DataLayout &Layout = DAG.getDataLayout(); bool Fast = false; bool Allowed = TLI.allowsMemoryAccess(Context, Layout, WideVT, *FirstStore->getMemOperand(), &Fast); if (!Allowed || !Fast) return SDValue(); // Check if the pieces of the value are going to the expected places in memory // to merge the stores. auto checkOffsets = [&](bool MatchLittleEndian) { if (MatchLittleEndian) { for (unsigned i = 0; i != NumStores; ++i) if (OffsetMap[i] != i * (NarrowNumBits / 8) + FirstOffset) return false; } else { // MatchBigEndian by reversing loop counter. for (unsigned i = 0, j = NumStores - 1; i != NumStores; ++i, --j) if (OffsetMap[j] != i * (NarrowNumBits / 8) + FirstOffset) return false; } return true; }; // Check if the offsets line up for the native data layout of this target. bool NeedBswap = false; bool NeedRotate = false; if (!checkOffsets(Layout.isLittleEndian())) { // Special-case: check if byte offsets line up for the opposite endian. if (NarrowNumBits == 8 && checkOffsets(Layout.isBigEndian())) NeedBswap = true; else if (NumStores == 2 && checkOffsets(Layout.isBigEndian())) NeedRotate = true; else return SDValue(); } SDLoc DL(N); if (WideVT != SourceValue.getValueType()) { assert(SourceValue.getValueType().getScalarSizeInBits() > WideNumBits && "Unexpected store value to merge"); SourceValue = DAG.getNode(ISD::TRUNCATE, DL, WideVT, SourceValue); } // Before legalize we can introduce illegal bswaps/rotates which will be later // converted to an explicit bswap sequence. This way we end up with a single // store and byte shuffling instead of several stores and byte shuffling. if (NeedBswap) { SourceValue = DAG.getNode(ISD::BSWAP, DL, WideVT, SourceValue); } else if (NeedRotate) { assert(WideNumBits % 2 == 0 && "Unexpected type for rotate"); SDValue RotAmt = DAG.getConstant(WideNumBits / 2, DL, WideVT); SourceValue = DAG.getNode(ISD::ROTR, DL, WideVT, SourceValue, RotAmt); } SDValue NewStore = DAG.getStore(Chain, DL, SourceValue, FirstStore->getBasePtr(), FirstStore->getPointerInfo(), FirstStore->getAlign()); // Rely on other DAG combine rules to remove the other individual stores. DAG.ReplaceAllUsesWith(N, NewStore.getNode()); return NewStore; } /// Match a pattern where a wide type scalar value is loaded by several narrow /// loads and combined by shifts and ors. Fold it into a single load or a load /// and a BSWAP if the targets supports it. /// /// Assuming little endian target: /// i8 *a = ... /// i32 val = a[0] | (a[1] << 8) | (a[2] << 16) | (a[3] << 24) /// => /// i32 val = *((i32)a) /// /// i8 *a = ... /// i32 val = (a[0] << 24) | (a[1] << 16) | (a[2] << 8) | a[3] /// => /// i32 val = BSWAP(*((i32)a)) /// /// TODO: This rule matches complex patterns with OR node roots and doesn't /// interact well with the worklist mechanism. When a part of the pattern is /// updated (e.g. one of the loads) its direct users are put into the worklist, /// but the root node of the pattern which triggers the load combine is not /// necessarily a direct user of the changed node. For example, once the address /// of t28 load is reassociated load combine won't be triggered: /// t25: i32 = add t4, Constant:i32<2> /// t26: i64 = sign_extend t25 /// t27: i64 = add t2, t26 /// t28: i8,ch = load t0, t27, undef:i64 /// t29: i32 = zero_extend t28 /// t32: i32 = shl t29, Constant:i8<8> /// t33: i32 = or t23, t32 /// As a possible fix visitLoad can check if the load can be a part of a load /// combine pattern and add corresponding OR roots to the worklist. SDValue DAGCombiner::MatchLoadCombine(SDNode *N) { assert(N->getOpcode() == ISD::OR && "Can only match load combining against OR nodes"); // Handles simple types only EVT VT = N->getValueType(0); if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64) return SDValue(); unsigned ByteWidth = VT.getSizeInBits() / 8; bool IsBigEndianTarget = DAG.getDataLayout().isBigEndian(); auto MemoryByteOffset = [&] (ByteProvider P) { assert(P.isMemory() && "Must be a memory byte provider"); unsigned LoadBitWidth = P.Load->getMemoryVT().getSizeInBits(); assert(LoadBitWidth % 8 == 0 && "can only analyze providers for individual bytes not bit"); unsigned LoadByteWidth = LoadBitWidth / 8; return IsBigEndianTarget ? bigEndianByteAt(LoadByteWidth, P.ByteOffset) : littleEndianByteAt(LoadByteWidth, P.ByteOffset); }; Optional Base; SDValue Chain; SmallPtrSet Loads; Optional FirstByteProvider; int64_t FirstOffset = INT64_MAX; // Check if all the bytes of the OR we are looking at are loaded from the same // base address. Collect bytes offsets from Base address in ByteOffsets. SmallVector ByteOffsets(ByteWidth); unsigned ZeroExtendedBytes = 0; for (int i = ByteWidth - 1; i >= 0; --i) { auto P = calculateByteProvider(SDValue(N, 0), i, 0, /*Root=*/true); if (!P) return SDValue(); if (P->isConstantZero()) { // It's OK for the N most significant bytes to be 0, we can just // zero-extend the load. if (++ZeroExtendedBytes != (ByteWidth - static_cast(i))) return SDValue(); continue; } assert(P->isMemory() && "provenance should either be memory or zero"); LoadSDNode *L = P->Load; assert(L->hasNUsesOfValue(1, 0) && L->isSimple() && !L->isIndexed() && "Must be enforced by calculateByteProvider"); assert(L->getOffset().isUndef() && "Unindexed load must have undef offset"); // All loads must share the same chain SDValue LChain = L->getChain(); if (!Chain) Chain = LChain; else if (Chain != LChain) return SDValue(); // Loads must share the same base address BaseIndexOffset Ptr = BaseIndexOffset::match(L, DAG); int64_t ByteOffsetFromBase = 0; if (!Base) Base = Ptr; else if (!Base->equalBaseIndex(Ptr, DAG, ByteOffsetFromBase)) return SDValue(); // Calculate the offset of the current byte from the base address ByteOffsetFromBase += MemoryByteOffset(*P); ByteOffsets[i] = ByteOffsetFromBase; // Remember the first byte load if (ByteOffsetFromBase < FirstOffset) { FirstByteProvider = P; FirstOffset = ByteOffsetFromBase; } Loads.insert(L); } assert(!Loads.empty() && "All the bytes of the value must be loaded from " "memory, so there must be at least one load which produces the value"); assert(Base && "Base address of the accessed memory location must be set"); assert(FirstOffset != INT64_MAX && "First byte offset must be set"); bool NeedsZext = ZeroExtendedBytes > 0; EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), (ByteWidth - ZeroExtendedBytes) * 8); if (!MemVT.isSimple()) return SDValue(); // Before legalize we can introduce too wide illegal loads which will be later // split into legal sized loads. This enables us to combine i64 load by i8 // patterns to a couple of i32 loads on 32 bit targets. if (LegalOperations && !TLI.isOperationLegal(NeedsZext ? ISD::ZEXTLOAD : ISD::NON_EXTLOAD, MemVT)) return SDValue(); // Check if the bytes of the OR we are looking at match with either big or // little endian value load Optional IsBigEndian = isBigEndian( makeArrayRef(ByteOffsets).drop_back(ZeroExtendedBytes), FirstOffset); if (!IsBigEndian) return SDValue(); assert(FirstByteProvider && "must be set"); // Ensure that the first byte is loaded from zero offset of the first load. // So the combined value can be loaded from the first load address. if (MemoryByteOffset(*FirstByteProvider) != 0) return SDValue(); LoadSDNode *FirstLoad = FirstByteProvider->Load; // The node we are looking at matches with the pattern, check if we can // replace it with a single (possibly zero-extended) load and bswap + shift if // needed. // If the load needs byte swap check if the target supports it bool NeedsBswap = IsBigEndianTarget != *IsBigEndian; // Before legalize we can introduce illegal bswaps which will be later // converted to an explicit bswap sequence. This way we end up with a single // load and byte shuffling instead of several loads and byte shuffling. // We do not introduce illegal bswaps when zero-extending as this tends to // introduce too many arithmetic instructions. if (NeedsBswap && (LegalOperations || NeedsZext) && !TLI.isOperationLegal(ISD::BSWAP, VT)) return SDValue(); // If we need to bswap and zero extend, we have to insert a shift. Check that // it is legal. if (NeedsBswap && NeedsZext && LegalOperations && !TLI.isOperationLegal(ISD::SHL, VT)) return SDValue(); // Check that a load of the wide type is both allowed and fast on the target bool Fast = false; bool Allowed = TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), MemVT, *FirstLoad->getMemOperand(), &Fast); if (!Allowed || !Fast) return SDValue(); SDValue NewLoad = DAG.getExtLoad(NeedsZext ? ISD::ZEXTLOAD : ISD::NON_EXTLOAD, SDLoc(N), VT, Chain, FirstLoad->getBasePtr(), FirstLoad->getPointerInfo(), MemVT, FirstLoad->getAlign()); // Transfer chain users from old loads to the new load. for (LoadSDNode *L : Loads) DAG.ReplaceAllUsesOfValueWith(SDValue(L, 1), SDValue(NewLoad.getNode(), 1)); if (!NeedsBswap) return NewLoad; SDValue ShiftedLoad = NeedsZext ? DAG.getNode(ISD::SHL, SDLoc(N), VT, NewLoad, DAG.getShiftAmountConstant(ZeroExtendedBytes * 8, VT, SDLoc(N), LegalOperations)) : NewLoad; return DAG.getNode(ISD::BSWAP, SDLoc(N), VT, ShiftedLoad); } // If the target has andn, bsl, or a similar bit-select instruction, // we want to unfold masked merge, with canonical pattern of: // | A | |B| // ((x ^ y) & m) ^ y // | D | // Into: // (x & m) | (y & ~m) // If y is a constant, m is not a 'not', and the 'andn' does not work with // immediates, we unfold into a different pattern: // ~(~x & m) & (m | y) // If x is a constant, m is a 'not', and the 'andn' does not work with // immediates, we unfold into a different pattern: // (x | ~m) & ~(~m & ~y) // NOTE: we don't unfold the pattern if 'xor' is actually a 'not', because at // the very least that breaks andnpd / andnps patterns, and because those // patterns are simplified in IR and shouldn't be created in the DAG SDValue DAGCombiner::unfoldMaskedMerge(SDNode *N) { assert(N->getOpcode() == ISD::XOR); // Don't touch 'not' (i.e. where y = -1). if (isAllOnesOrAllOnesSplat(N->getOperand(1))) return SDValue(); EVT VT = N->getValueType(0); // There are 3 commutable operators in the pattern, // so we have to deal with 8 possible variants of the basic pattern. SDValue X, Y, M; auto matchAndXor = [&X, &Y, &M](SDValue And, unsigned XorIdx, SDValue Other) { if (And.getOpcode() != ISD::AND || !And.hasOneUse()) return false; SDValue Xor = And.getOperand(XorIdx); if (Xor.getOpcode() != ISD::XOR || !Xor.hasOneUse()) return false; SDValue Xor0 = Xor.getOperand(0); SDValue Xor1 = Xor.getOperand(1); // Don't touch 'not' (i.e. where y = -1). if (isAllOnesOrAllOnesSplat(Xor1)) return false; if (Other == Xor0) std::swap(Xor0, Xor1); if (Other != Xor1) return false; X = Xor0; Y = Xor1; M = And.getOperand(XorIdx ? 0 : 1); return true; }; SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); if (!matchAndXor(N0, 0, N1) && !matchAndXor(N0, 1, N1) && !matchAndXor(N1, 0, N0) && !matchAndXor(N1, 1, N0)) return SDValue(); // Don't do anything if the mask is constant. This should not be reachable. // InstCombine should have already unfolded this pattern, and DAGCombiner // probably shouldn't produce it, too. if (isa(M.getNode())) return SDValue(); // We can transform if the target has AndNot if (!TLI.hasAndNot(M)) return SDValue(); SDLoc DL(N); // If Y is a constant, check that 'andn' works with immediates. Unless M is // a bitwise not that would already allow ANDN to be used. if (!TLI.hasAndNot(Y) && !isBitwiseNot(M)) { assert(TLI.hasAndNot(X) && "Only mask is a variable? Unreachable."); // If not, we need to do a bit more work to make sure andn is still used. SDValue NotX = DAG.getNOT(DL, X, VT); SDValue LHS = DAG.getNode(ISD::AND, DL, VT, NotX, M); SDValue NotLHS = DAG.getNOT(DL, LHS, VT); SDValue RHS = DAG.getNode(ISD::OR, DL, VT, M, Y); return DAG.getNode(ISD::AND, DL, VT, NotLHS, RHS); } // If X is a constant and M is a bitwise not, check that 'andn' works with // immediates. if (!TLI.hasAndNot(X) && isBitwiseNot(M)) { assert(TLI.hasAndNot(Y) && "Only mask is a variable? Unreachable."); // If not, we need to do a bit more work to make sure andn is still used. SDValue NotM = M.getOperand(0); SDValue LHS = DAG.getNode(ISD::OR, DL, VT, X, NotM); SDValue NotY = DAG.getNOT(DL, Y, VT); SDValue RHS = DAG.getNode(ISD::AND, DL, VT, NotM, NotY); SDValue NotRHS = DAG.getNOT(DL, RHS, VT); return DAG.getNode(ISD::AND, DL, VT, LHS, NotRHS); } SDValue LHS = DAG.getNode(ISD::AND, DL, VT, X, M); SDValue NotM = DAG.getNOT(DL, M, VT); SDValue RHS = DAG.getNode(ISD::AND, DL, VT, Y, NotM); return DAG.getNode(ISD::OR, DL, VT, LHS, RHS); } SDValue DAGCombiner::visitXOR(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N0.getValueType(); SDLoc DL(N); // fold (xor undef, undef) -> 0. This is a common idiom (misuse). if (N0.isUndef() && N1.isUndef()) return DAG.getConstant(0, DL, VT); // fold (xor x, undef) -> undef if (N0.isUndef()) return N0; if (N1.isUndef()) return N1; // fold (xor c1, c2) -> c1^c2 if (SDValue C = DAG.FoldConstantArithmetic(ISD::XOR, DL, VT, {N0, N1})) return C; // canonicalize constant to RHS if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && !DAG.isConstantIntBuildVectorOrConstantInt(N1)) return DAG.getNode(ISD::XOR, DL, VT, N1, N0); // fold vector ops if (VT.isVector()) { if (SDValue FoldedVOp = SimplifyVBinOp(N, DL)) return FoldedVOp; // fold (xor x, 0) -> x, vector edition if (ISD::isConstantSplatVectorAllZeros(N1.getNode())) return N0; } // fold (xor x, 0) -> x if (isNullConstant(N1)) return N0; if (SDValue NewSel = foldBinOpIntoSelect(N)) return NewSel; // reassociate xor if (SDValue RXOR = reassociateOps(ISD::XOR, DL, N0, N1, N->getFlags())) return RXOR; // look for 'add-like' folds: // XOR(N0,MIN_SIGNED_VALUE) == ADD(N0,MIN_SIGNED_VALUE) if ((!LegalOperations || TLI.isOperationLegal(ISD::ADD, VT)) && isMinSignedConstant(N1)) if (SDValue Combined = visitADDLike(N)) return Combined; // fold !(x cc y) -> (x !cc y) unsigned N0Opcode = N0.getOpcode(); SDValue LHS, RHS, CC; if (TLI.isConstTrueVal(N1) && isSetCCEquivalent(N0, LHS, RHS, CC, /*MatchStrict*/ true)) { ISD::CondCode NotCC = ISD::getSetCCInverse(cast(CC)->get(), LHS.getValueType()); if (!LegalOperations || TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) { switch (N0Opcode) { default: llvm_unreachable("Unhandled SetCC Equivalent!"); case ISD::SETCC: return DAG.getSetCC(SDLoc(N0), VT, LHS, RHS, NotCC); case ISD::SELECT_CC: return DAG.getSelectCC(SDLoc(N0), LHS, RHS, N0.getOperand(2), N0.getOperand(3), NotCC); case ISD::STRICT_FSETCC: case ISD::STRICT_FSETCCS: { if (N0.hasOneUse()) { // FIXME Can we handle multiple uses? Could we token factor the chain // results from the new/old setcc? SDValue SetCC = DAG.getSetCC(SDLoc(N0), VT, LHS, RHS, NotCC, N0.getOperand(0), N0Opcode == ISD::STRICT_FSETCCS); CombineTo(N, SetCC); DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), SetCC.getValue(1)); recursivelyDeleteUnusedNodes(N0.getNode()); return SDValue(N, 0); // Return N so it doesn't get rechecked! } break; } } } } // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y))) if (isOneConstant(N1) && N0Opcode == ISD::ZERO_EXTEND && N0.hasOneUse() && isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){ SDValue V = N0.getOperand(0); SDLoc DL0(N0); V = DAG.getNode(ISD::XOR, DL0, V.getValueType(), V, DAG.getConstant(1, DL0, V.getValueType())); AddToWorklist(V.getNode()); return DAG.getNode(ISD::ZERO_EXTEND, DL, VT, V); } // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc if (isOneConstant(N1) && VT == MVT::i1 && N0.hasOneUse() && (N0Opcode == ISD::OR || N0Opcode == ISD::AND)) { SDValue N00 = N0.getOperand(0), N01 = N0.getOperand(1); if (isOneUseSetCC(N01) || isOneUseSetCC(N00)) { unsigned NewOpcode = N0Opcode == ISD::AND ? ISD::OR : ISD::AND; N00 = DAG.getNode(ISD::XOR, SDLoc(N00), VT, N00, N1); // N00 = ~N00 N01 = DAG.getNode(ISD::XOR, SDLoc(N01), VT, N01, N1); // N01 = ~N01 AddToWorklist(N00.getNode()); AddToWorklist(N01.getNode()); return DAG.getNode(NewOpcode, DL, VT, N00, N01); } } // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants if (isAllOnesConstant(N1) && N0.hasOneUse() && (N0Opcode == ISD::OR || N0Opcode == ISD::AND)) { SDValue N00 = N0.getOperand(0), N01 = N0.getOperand(1); if (isa(N01) || isa(N00)) { unsigned NewOpcode = N0Opcode == ISD::AND ? ISD::OR : ISD::AND; N00 = DAG.getNode(ISD::XOR, SDLoc(N00), VT, N00, N1); // N00 = ~N00 N01 = DAG.getNode(ISD::XOR, SDLoc(N01), VT, N01, N1); // N01 = ~N01 AddToWorklist(N00.getNode()); AddToWorklist(N01.getNode()); return DAG.getNode(NewOpcode, DL, VT, N00, N01); } } // fold (not (neg x)) -> (add X, -1) // FIXME: This can be generalized to (not (sub Y, X)) -> (add X, ~Y) if // Y is a constant or the subtract has a single use. if (isAllOnesConstant(N1) && N0.getOpcode() == ISD::SUB && isNullConstant(N0.getOperand(0))) { return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(1), DAG.getAllOnesConstant(DL, VT)); } // fold (not (add X, -1)) -> (neg X) if (isAllOnesConstant(N1) && N0.getOpcode() == ISD::ADD && isAllOnesOrAllOnesSplat(N0.getOperand(1))) { return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), N0.getOperand(0)); } // fold (xor (and x, y), y) -> (and (not x), y) if (N0Opcode == ISD::AND && N0.hasOneUse() && N0->getOperand(1) == N1) { SDValue X = N0.getOperand(0); SDValue NotX = DAG.getNOT(SDLoc(X), X, VT); AddToWorklist(NotX.getNode()); return DAG.getNode(ISD::AND, DL, VT, NotX, N1); } // fold Y = sra (X, size(X)-1); xor (add (X, Y), Y) -> (abs X) if (TLI.isOperationLegalOrCustom(ISD::ABS, VT)) { SDValue A = N0Opcode == ISD::ADD ? N0 : N1; SDValue S = N0Opcode == ISD::SRA ? N0 : N1; if (A.getOpcode() == ISD::ADD && S.getOpcode() == ISD::SRA) { SDValue A0 = A.getOperand(0), A1 = A.getOperand(1); SDValue S0 = S.getOperand(0); if ((A0 == S && A1 == S0) || (A1 == S && A0 == S0)) if (ConstantSDNode *C = isConstOrConstSplat(S.getOperand(1))) if (C->getAPIntValue() == (VT.getScalarSizeInBits() - 1)) return DAG.getNode(ISD::ABS, DL, VT, S0); } } // fold (xor x, x) -> 0 if (N0 == N1) return tryFoldToZero(DL, TLI, VT, DAG, LegalOperations); // fold (xor (shl 1, x), -1) -> (rotl ~1, x) // Here is a concrete example of this equivalence: // i16 x == 14 // i16 shl == 1 << 14 == 16384 == 0b0100000000000000 // i16 xor == ~(1 << 14) == 49151 == 0b1011111111111111 // // => // // i16 ~1 == 0b1111111111111110 // i16 rol(~1, 14) == 0b1011111111111111 // // Some additional tips to help conceptualize this transform: // - Try to see the operation as placing a single zero in a value of all ones. // - There exists no value for x which would allow the result to contain zero. // - Values of x larger than the bitwidth are undefined and do not require a // consistent result. // - Pushing the zero left requires shifting one bits in from the right. // A rotate left of ~1 is a nice way of achieving the desired result. if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT) && N0Opcode == ISD::SHL && isAllOnesConstant(N1) && isOneConstant(N0.getOperand(0))) { return DAG.getNode(ISD::ROTL, DL, VT, DAG.getConstant(~1, DL, VT), N0.getOperand(1)); } // Simplify: xor (op x...), (op y...) -> (op (xor x, y)) if (N0Opcode == N1.getOpcode()) if (SDValue V = hoistLogicOpWithSameOpcodeHands(N)) return V; if (SDValue R = foldLogicOfShifts(N, N0, N1, DAG)) return R; if (SDValue R = foldLogicOfShifts(N, N1, N0, DAG)) return R; // Unfold ((x ^ y) & m) ^ y into (x & m) | (y & ~m) if profitable if (SDValue MM = unfoldMaskedMerge(N)) return MM; // Simplify the expression using non-local knowledge. if (SimplifyDemandedBits(SDValue(N, 0))) return SDValue(N, 0); if (SDValue Combined = combineCarryDiamond(DAG, TLI, N0, N1, N)) return Combined; return SDValue(); } /// If we have a shift-by-constant of a bitwise logic op that itself has a /// shift-by-constant operand with identical opcode, we may be able to convert /// that into 2 independent shifts followed by the logic op. This is a /// throughput improvement. static SDValue combineShiftOfShiftedLogic(SDNode *Shift, SelectionDAG &DAG) { // Match a one-use bitwise logic op. SDValue LogicOp = Shift->getOperand(0); if (!LogicOp.hasOneUse()) return SDValue(); unsigned LogicOpcode = LogicOp.getOpcode(); if (LogicOpcode != ISD::AND && LogicOpcode != ISD::OR && LogicOpcode != ISD::XOR) return SDValue(); // Find a matching one-use shift by constant. unsigned ShiftOpcode = Shift->getOpcode(); SDValue C1 = Shift->getOperand(1); ConstantSDNode *C1Node = isConstOrConstSplat(C1); assert(C1Node && "Expected a shift with constant operand"); const APInt &C1Val = C1Node->getAPIntValue(); auto matchFirstShift = [&](SDValue V, SDValue &ShiftOp, const APInt *&ShiftAmtVal) { if (V.getOpcode() != ShiftOpcode || !V.hasOneUse()) return false; ConstantSDNode *ShiftCNode = isConstOrConstSplat(V.getOperand(1)); if (!ShiftCNode) return false; // Capture the shifted operand and shift amount value. ShiftOp = V.getOperand(0); ShiftAmtVal = &ShiftCNode->getAPIntValue(); // Shift amount types do not have to match their operand type, so check that // the constants are the same width. if (ShiftAmtVal->getBitWidth() != C1Val.getBitWidth()) return false; // The fold is not valid if the sum of the shift values exceeds bitwidth. if ((*ShiftAmtVal + C1Val).uge(V.getScalarValueSizeInBits())) return false; return true; }; // Logic ops are commutative, so check each operand for a match. SDValue X, Y; const APInt *C0Val; if (matchFirstShift(LogicOp.getOperand(0), X, C0Val)) Y = LogicOp.getOperand(1); else if (matchFirstShift(LogicOp.getOperand(1), X, C0Val)) Y = LogicOp.getOperand(0); else return SDValue(); // shift (logic (shift X, C0), Y), C1 -> logic (shift X, C0+C1), (shift Y, C1) SDLoc DL(Shift); EVT VT = Shift->getValueType(0); EVT ShiftAmtVT = Shift->getOperand(1).getValueType(); SDValue ShiftSumC = DAG.getConstant(*C0Val + C1Val, DL, ShiftAmtVT); SDValue NewShift1 = DAG.getNode(ShiftOpcode, DL, VT, X, ShiftSumC); SDValue NewShift2 = DAG.getNode(ShiftOpcode, DL, VT, Y, C1); return DAG.getNode(LogicOpcode, DL, VT, NewShift1, NewShift2); } /// Handle transforms common to the three shifts, when the shift amount is a /// constant. /// We are looking for: (shift being one of shl/sra/srl) /// shift (binop X, C0), C1 /// And want to transform into: /// binop (shift X, C1), (shift C0, C1) SDValue DAGCombiner::visitShiftByConstant(SDNode *N) { assert(isConstOrConstSplat(N->getOperand(1)) && "Expected constant operand"); // Do not turn a 'not' into a regular xor. if (isBitwiseNot(N->getOperand(0))) return SDValue(); // The inner binop must be one-use, since we want to replace it. SDValue LHS = N->getOperand(0); if (!LHS.hasOneUse() || !TLI.isDesirableToCommuteWithShift(N, Level)) return SDValue(); // TODO: This is limited to early combining because it may reveal regressions // otherwise. But since we just checked a target hook to see if this is // desirable, that should have filtered out cases where this interferes // with some other pattern matching. if (!LegalTypes) if (SDValue R = combineShiftOfShiftedLogic(N, DAG)) return R; // We want to pull some binops through shifts, so that we have (and (shift)) // instead of (shift (and)), likewise for add, or, xor, etc. This sort of // thing happens with address calculations, so it's important to canonicalize // it. switch (LHS.getOpcode()) { default: return SDValue(); case ISD::OR: case ISD::XOR: case ISD::AND: break; case ISD::ADD: if (N->getOpcode() != ISD::SHL) return SDValue(); // only shl(add) not sr[al](add). break; } // We require the RHS of the binop to be a constant and not opaque as well. ConstantSDNode *BinOpCst = getAsNonOpaqueConstant(LHS.getOperand(1)); if (!BinOpCst) return SDValue(); // FIXME: disable this unless the input to the binop is a shift by a constant // or is copy/select. Enable this in other cases when figure out it's exactly // profitable. SDValue BinOpLHSVal = LHS.getOperand(0); bool IsShiftByConstant = (BinOpLHSVal.getOpcode() == ISD::SHL || BinOpLHSVal.getOpcode() == ISD::SRA || BinOpLHSVal.getOpcode() == ISD::SRL) && isa(BinOpLHSVal.getOperand(1)); bool IsCopyOrSelect = BinOpLHSVal.getOpcode() == ISD::CopyFromReg || BinOpLHSVal.getOpcode() == ISD::SELECT; if (!IsShiftByConstant && !IsCopyOrSelect) return SDValue(); if (IsCopyOrSelect && N->hasOneUse()) return SDValue(); // Fold the constants, shifting the binop RHS by the shift amount. SDLoc DL(N); EVT VT = N->getValueType(0); SDValue NewRHS = DAG.getNode(N->getOpcode(), DL, VT, LHS.getOperand(1), N->getOperand(1)); assert(isa(NewRHS) && "Folding was not successful!"); SDValue NewShift = DAG.getNode(N->getOpcode(), DL, VT, LHS.getOperand(0), N->getOperand(1)); return DAG.getNode(LHS.getOpcode(), DL, VT, NewShift, NewRHS); } SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) { assert(N->getOpcode() == ISD::TRUNCATE); assert(N->getOperand(0).getOpcode() == ISD::AND); // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC) EVT TruncVT = N->getValueType(0); if (N->hasOneUse() && N->getOperand(0).hasOneUse() && TLI.isTypeDesirableForOp(ISD::AND, TruncVT)) { SDValue N01 = N->getOperand(0).getOperand(1); if (isConstantOrConstantVector(N01, /* NoOpaques */ true)) { SDLoc DL(N); SDValue N00 = N->getOperand(0).getOperand(0); SDValue Trunc00 = DAG.getNode(ISD::TRUNCATE, DL, TruncVT, N00); SDValue Trunc01 = DAG.getNode(ISD::TRUNCATE, DL, TruncVT, N01); AddToWorklist(Trunc00.getNode()); AddToWorklist(Trunc01.getNode()); return DAG.getNode(ISD::AND, DL, TruncVT, Trunc00, Trunc01); } } return SDValue(); } SDValue DAGCombiner::visitRotate(SDNode *N) { SDLoc dl(N); SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); unsigned Bitsize = VT.getScalarSizeInBits(); // fold (rot x, 0) -> x if (isNullOrNullSplat(N1)) return N0; // fold (rot x, c) -> x iff (c % BitSize) == 0 if (isPowerOf2_32(Bitsize) && Bitsize > 1) { APInt ModuloMask(N1.getScalarValueSizeInBits(), Bitsize - 1); if (DAG.MaskedValueIsZero(N1, ModuloMask)) return N0; } // fold (rot x, c) -> (rot x, c % BitSize) bool OutOfRange = false; auto MatchOutOfRange = [Bitsize, &OutOfRange](ConstantSDNode *C) { OutOfRange |= C->getAPIntValue().uge(Bitsize); return true; }; if (ISD::matchUnaryPredicate(N1, MatchOutOfRange) && OutOfRange) { EVT AmtVT = N1.getValueType(); SDValue Bits = DAG.getConstant(Bitsize, dl, AmtVT); if (SDValue Amt = DAG.FoldConstantArithmetic(ISD::UREM, dl, AmtVT, {N1, Bits})) return DAG.getNode(N->getOpcode(), dl, VT, N0, Amt); } // rot i16 X, 8 --> bswap X auto *RotAmtC = isConstOrConstSplat(N1); if (RotAmtC && RotAmtC->getAPIntValue() == 8 && VT.getScalarSizeInBits() == 16 && hasOperation(ISD::BSWAP, VT)) return DAG.getNode(ISD::BSWAP, dl, VT, N0); // Simplify the operands using demanded-bits information. if (SimplifyDemandedBits(SDValue(N, 0))) return SDValue(N, 0); // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))). if (N1.getOpcode() == ISD::TRUNCATE && N1.getOperand(0).getOpcode() == ISD::AND) { if (SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode())) return DAG.getNode(N->getOpcode(), dl, VT, N0, NewOp1); } unsigned NextOp = N0.getOpcode(); // fold (rot* (rot* x, c2), c1) // -> (rot* x, ((c1 % bitsize) +- (c2 % bitsize)) % bitsize) if (NextOp == ISD::ROTL || NextOp == ISD::ROTR) { SDNode *C1 = DAG.isConstantIntBuildVectorOrConstantInt(N1); SDNode *C2 = DAG.isConstantIntBuildVectorOrConstantInt(N0.getOperand(1)); if (C1 && C2 && C1->getValueType(0) == C2->getValueType(0)) { EVT ShiftVT = C1->getValueType(0); bool SameSide = (N->getOpcode() == NextOp); unsigned CombineOp = SameSide ? ISD::ADD : ISD::SUB; SDValue BitsizeC = DAG.getConstant(Bitsize, dl, ShiftVT); SDValue Norm1 = DAG.FoldConstantArithmetic(ISD::UREM, dl, ShiftVT, {N1, BitsizeC}); SDValue Norm2 = DAG.FoldConstantArithmetic(ISD::UREM, dl, ShiftVT, {N0.getOperand(1), BitsizeC}); if (Norm1 && Norm2) if (SDValue CombinedShift = DAG.FoldConstantArithmetic( CombineOp, dl, ShiftVT, {Norm1, Norm2})) { SDValue CombinedShiftNorm = DAG.FoldConstantArithmetic( ISD::UREM, dl, ShiftVT, {CombinedShift, BitsizeC}); return DAG.getNode(N->getOpcode(), dl, VT, N0->getOperand(0), CombinedShiftNorm); } } } return SDValue(); } SDValue DAGCombiner::visitSHL(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); if (SDValue V = DAG.simplifyShift(N0, N1)) return V; EVT VT = N0.getValueType(); EVT ShiftVT = N1.getValueType(); unsigned OpSizeInBits = VT.getScalarSizeInBits(); // fold (shl c1, c2) -> c1<(N1); // If setcc produces all-one true value then: // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<isConstant()) { if (N0.getOpcode() == ISD::AND) { SDValue N00 = N0->getOperand(0); SDValue N01 = N0->getOperand(1); BuildVectorSDNode *N01CV = dyn_cast(N01); if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC && TLI.getBooleanContents(N00.getOperand(0).getValueType()) == TargetLowering::ZeroOrNegativeOneBooleanContent) { if (SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, SDLoc(N), VT, {N01, N1})) return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C); } } } } if (SDValue NewSel = foldBinOpIntoSelect(N)) return NewSel; // if (shl x, c) is known to be zero, return 0 if (DAG.MaskedValueIsZero(SDValue(N, 0), APInt::getAllOnes(OpSizeInBits))) return DAG.getConstant(0, SDLoc(N), VT); // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))). if (N1.getOpcode() == ISD::TRUNCATE && N1.getOperand(0).getOpcode() == ISD::AND) { if (SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode())) return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1); } if (SimplifyDemandedBits(SDValue(N, 0))) return SDValue(N, 0); // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2)) if (N0.getOpcode() == ISD::SHL) { auto MatchOutOfRange = [OpSizeInBits](ConstantSDNode *LHS, ConstantSDNode *RHS) { APInt c1 = LHS->getAPIntValue(); APInt c2 = RHS->getAPIntValue(); zeroExtendToMatch(c1, c2, 1 /* Overflow Bit */); return (c1 + c2).uge(OpSizeInBits); }; if (ISD::matchBinaryPredicate(N1, N0.getOperand(1), MatchOutOfRange)) return DAG.getConstant(0, SDLoc(N), VT); auto MatchInRange = [OpSizeInBits](ConstantSDNode *LHS, ConstantSDNode *RHS) { APInt c1 = LHS->getAPIntValue(); APInt c2 = RHS->getAPIntValue(); zeroExtendToMatch(c1, c2, 1 /* Overflow Bit */); return (c1 + c2).ult(OpSizeInBits); }; if (ISD::matchBinaryPredicate(N1, N0.getOperand(1), MatchInRange)) { SDLoc DL(N); SDValue Sum = DAG.getNode(ISD::ADD, DL, ShiftVT, N1, N0.getOperand(1)); return DAG.getNode(ISD::SHL, DL, VT, N0.getOperand(0), Sum); } } // fold (shl (ext (shl x, c1)), c2) -> (shl (ext x), (add c1, c2)) // For this to be valid, the second form must not preserve any of the bits // that are shifted out by the inner shift in the first form. This means // the outer shift size must be >= the number of bits added by the ext. // As a corollary, we don't care what kind of ext it is. if ((N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND || N0.getOpcode() == ISD::SIGN_EXTEND) && N0.getOperand(0).getOpcode() == ISD::SHL) { SDValue N0Op0 = N0.getOperand(0); SDValue InnerShiftAmt = N0Op0.getOperand(1); EVT InnerVT = N0Op0.getValueType(); uint64_t InnerBitwidth = InnerVT.getScalarSizeInBits(); auto MatchOutOfRange = [OpSizeInBits, InnerBitwidth](ConstantSDNode *LHS, ConstantSDNode *RHS) { APInt c1 = LHS->getAPIntValue(); APInt c2 = RHS->getAPIntValue(); zeroExtendToMatch(c1, c2, 1 /* Overflow Bit */); return c2.uge(OpSizeInBits - InnerBitwidth) && (c1 + c2).uge(OpSizeInBits); }; if (ISD::matchBinaryPredicate(InnerShiftAmt, N1, MatchOutOfRange, /*AllowUndefs*/ false, /*AllowTypeMismatch*/ true)) return DAG.getConstant(0, SDLoc(N), VT); auto MatchInRange = [OpSizeInBits, InnerBitwidth](ConstantSDNode *LHS, ConstantSDNode *RHS) { APInt c1 = LHS->getAPIntValue(); APInt c2 = RHS->getAPIntValue(); zeroExtendToMatch(c1, c2, 1 /* Overflow Bit */); return c2.uge(OpSizeInBits - InnerBitwidth) && (c1 + c2).ult(OpSizeInBits); }; if (ISD::matchBinaryPredicate(InnerShiftAmt, N1, MatchInRange, /*AllowUndefs*/ false, /*AllowTypeMismatch*/ true)) { SDLoc DL(N); SDValue Ext = DAG.getNode(N0.getOpcode(), DL, VT, N0Op0.getOperand(0)); SDValue Sum = DAG.getZExtOrTrunc(InnerShiftAmt, DL, ShiftVT); Sum = DAG.getNode(ISD::ADD, DL, ShiftVT, Sum, N1); return DAG.getNode(ISD::SHL, DL, VT, Ext, Sum); } } // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C)) // Only fold this if the inner zext has no other uses to avoid increasing // the total number of instructions. if (N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() && N0.getOperand(0).getOpcode() == ISD::SRL) { SDValue N0Op0 = N0.getOperand(0); SDValue InnerShiftAmt = N0Op0.getOperand(1); auto MatchEqual = [VT](ConstantSDNode *LHS, ConstantSDNode *RHS) { APInt c1 = LHS->getAPIntValue(); APInt c2 = RHS->getAPIntValue(); zeroExtendToMatch(c1, c2); return c1.ult(VT.getScalarSizeInBits()) && (c1 == c2); }; if (ISD::matchBinaryPredicate(InnerShiftAmt, N1, MatchEqual, /*AllowUndefs*/ false, /*AllowTypeMismatch*/ true)) { SDLoc DL(N); EVT InnerShiftAmtVT = N0Op0.getOperand(1).getValueType(); SDValue NewSHL = DAG.getZExtOrTrunc(N1, DL, InnerShiftAmtVT); NewSHL = DAG.getNode(ISD::SHL, DL, N0Op0.getValueType(), N0Op0, NewSHL); AddToWorklist(NewSHL.getNode()); return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL); } } if (N0.getOpcode() == ISD::SRL || N0.getOpcode() == ISD::SRA) { auto MatchShiftAmount = [OpSizeInBits](ConstantSDNode *LHS, ConstantSDNode *RHS) { const APInt &LHSC = LHS->getAPIntValue(); const APInt &RHSC = RHS->getAPIntValue(); return LHSC.ult(OpSizeInBits) && RHSC.ult(OpSizeInBits) && LHSC.getZExtValue() <= RHSC.getZExtValue(); }; SDLoc DL(N); // fold (shl (sr[la] exact X, C1), C2) -> (shl X, (C2-C1)) if C1 <= C2 // fold (shl (sr[la] exact X, C1), C2) -> (sr[la] X, (C2-C1)) if C1 >= C2 if (N0->getFlags().hasExact()) { if (ISD::matchBinaryPredicate(N0.getOperand(1), N1, MatchShiftAmount, /*AllowUndefs*/ false, /*AllowTypeMismatch*/ true)) { SDValue N01 = DAG.getZExtOrTrunc(N0.getOperand(1), DL, ShiftVT); SDValue Diff = DAG.getNode(ISD::SUB, DL, ShiftVT, N1, N01); return DAG.getNode(ISD::SHL, DL, VT, N0.getOperand(0), Diff); } if (ISD::matchBinaryPredicate(N1, N0.getOperand(1), MatchShiftAmount, /*AllowUndefs*/ false, /*AllowTypeMismatch*/ true)) { SDValue N01 = DAG.getZExtOrTrunc(N0.getOperand(1), DL, ShiftVT); SDValue Diff = DAG.getNode(ISD::SUB, DL, ShiftVT, N01, N1); return DAG.getNode(N0.getOpcode(), DL, VT, N0.getOperand(0), Diff); } } // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or // (and (srl x, (sub c1, c2), MASK) // Only fold this if the inner shift has no other uses -- if it does, // folding this will increase the total number of instructions. if (N0.getOpcode() == ISD::SRL && (N0.getOperand(1) == N1 || N0.hasOneUse()) && TLI.shouldFoldConstantShiftPairToMask(N, Level)) { if (ISD::matchBinaryPredicate(N1, N0.getOperand(1), MatchShiftAmount, /*AllowUndefs*/ false, /*AllowTypeMismatch*/ true)) { SDValue N01 = DAG.getZExtOrTrunc(N0.getOperand(1), DL, ShiftVT); SDValue Diff = DAG.getNode(ISD::SUB, DL, ShiftVT, N01, N1); SDValue Mask = DAG.getAllOnesConstant(DL, VT); Mask = DAG.getNode(ISD::SHL, DL, VT, Mask, N01); Mask = DAG.getNode(ISD::SRL, DL, VT, Mask, Diff); SDValue Shift = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0), Diff); return DAG.getNode(ISD::AND, DL, VT, Shift, Mask); } if (ISD::matchBinaryPredicate(N0.getOperand(1), N1, MatchShiftAmount, /*AllowUndefs*/ false, /*AllowTypeMismatch*/ true)) { SDValue N01 = DAG.getZExtOrTrunc(N0.getOperand(1), DL, ShiftVT); SDValue Diff = DAG.getNode(ISD::SUB, DL, ShiftVT, N1, N01); SDValue Mask = DAG.getAllOnesConstant(DL, VT); Mask = DAG.getNode(ISD::SHL, DL, VT, Mask, N1); SDValue Shift = DAG.getNode(ISD::SHL, DL, VT, N0.getOperand(0), Diff); return DAG.getNode(ISD::AND, DL, VT, Shift, Mask); } } } // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1)) if (N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1) && isConstantOrConstantVector(N1, /* No Opaques */ true)) { SDLoc DL(N); SDValue AllBits = DAG.getAllOnesConstant(DL, VT); SDValue HiBitsMask = DAG.getNode(ISD::SHL, DL, VT, AllBits, N1); return DAG.getNode(ISD::AND, DL, VT, N0.getOperand(0), HiBitsMask); } // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) // fold (shl (or x, c1), c2) -> (or (shl x, c2), c1 << c2) // Variant of version done on multiply, except mul by a power of 2 is turned // into a shift. if ((N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::OR) && N0->hasOneUse() && isConstantOrConstantVector(N1, /* No Opaques */ true) && isConstantOrConstantVector(N0.getOperand(1), /* No Opaques */ true) && TLI.isDesirableToCommuteWithShift(N, Level)) { SDValue Shl0 = DAG.getNode(ISD::SHL, SDLoc(N0), VT, N0.getOperand(0), N1); SDValue Shl1 = DAG.getNode(ISD::SHL, SDLoc(N1), VT, N0.getOperand(1), N1); AddToWorklist(Shl0.getNode()); AddToWorklist(Shl1.getNode()); return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, Shl0, Shl1); } // fold (shl (mul x, c1), c2) -> (mul x, c1 << c2) if (N0.getOpcode() == ISD::MUL && N0->hasOneUse()) { SDValue N01 = N0.getOperand(1); if (SDValue Shl = DAG.FoldConstantArithmetic(ISD::SHL, SDLoc(N1), VT, {N01, N1})) return DAG.getNode(ISD::MUL, SDLoc(N), VT, N0.getOperand(0), Shl); } ConstantSDNode *N1C = isConstOrConstSplat(N1); if (N1C && !N1C->isOpaque()) if (SDValue NewSHL = visitShiftByConstant(N)) return NewSHL; // Fold (shl (vscale * C0), C1) to (vscale * (C0 << C1)). if (N0.getOpcode() == ISD::VSCALE) if (ConstantSDNode *NC1 = isConstOrConstSplat(N->getOperand(1))) { const APInt &C0 = N0.getConstantOperandAPInt(0); const APInt &C1 = NC1->getAPIntValue(); return DAG.getVScale(SDLoc(N), VT, C0 << C1); } // Fold (shl step_vector(C0), C1) to (step_vector(C0 << C1)). APInt ShlVal; if (N0.getOpcode() == ISD::STEP_VECTOR) if (ISD::isConstantSplatVector(N1.getNode(), ShlVal)) { const APInt &C0 = N0.getConstantOperandAPInt(0); if (ShlVal.ult(C0.getBitWidth())) { APInt NewStep = C0 << ShlVal; return DAG.getStepVector(SDLoc(N), VT, NewStep); } } return SDValue(); } // Transform a right shift of a multiply into a multiply-high. // Examples: // (srl (mul (zext i32:$a to i64), (zext i32:$a to i64)), 32) -> (mulhu $a, $b) // (sra (mul (sext i32:$a to i64), (sext i32:$a to i64)), 32) -> (mulhs $a, $b) static SDValue combineShiftToMULH(SDNode *N, SelectionDAG &DAG, const TargetLowering &TLI) { assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) && "SRL or SRA node is required here!"); // Check the shift amount. Proceed with the transformation if the shift // amount is constant. ConstantSDNode *ShiftAmtSrc = isConstOrConstSplat(N->getOperand(1)); if (!ShiftAmtSrc) return SDValue(); SDLoc DL(N); // The operation feeding into the shift must be a multiply. SDValue ShiftOperand = N->getOperand(0); if (ShiftOperand.getOpcode() != ISD::MUL) return SDValue(); // Both operands must be equivalent extend nodes. SDValue LeftOp = ShiftOperand.getOperand(0); SDValue RightOp = ShiftOperand.getOperand(1); bool IsSignExt = LeftOp.getOpcode() == ISD::SIGN_EXTEND; bool IsZeroExt = LeftOp.getOpcode() == ISD::ZERO_EXTEND; if (!IsSignExt && !IsZeroExt) return SDValue(); EVT NarrowVT = LeftOp.getOperand(0).getValueType(); unsigned NarrowVTSize = NarrowVT.getScalarSizeInBits(); SDValue MulhRightOp; if (ConstantSDNode *Constant = isConstOrConstSplat(RightOp)) { unsigned ActiveBits = IsSignExt ? Constant->getAPIntValue().getMinSignedBits() : Constant->getAPIntValue().getActiveBits(); if (ActiveBits > NarrowVTSize) return SDValue(); MulhRightOp = DAG.getConstant( Constant->getAPIntValue().trunc(NarrowVT.getScalarSizeInBits()), DL, NarrowVT); } else { if (LeftOp.getOpcode() != RightOp.getOpcode()) return SDValue(); // Check that the two extend nodes are the same type. if (NarrowVT != RightOp.getOperand(0).getValueType()) return SDValue(); MulhRightOp = RightOp.getOperand(0); } EVT WideVT = LeftOp.getValueType(); // Proceed with the transformation if the wide types match. assert((WideVT == RightOp.getValueType()) && "Cannot have a multiply node with two different operand types."); // Proceed with the transformation if the wide type is twice as large // as the narrow type. if (WideVT.getScalarSizeInBits() != 2 * NarrowVTSize) return SDValue(); // Check the shift amount with the narrow type size. // Proceed with the transformation if the shift amount is the width // of the narrow type. unsigned ShiftAmt = ShiftAmtSrc->getZExtValue(); if (ShiftAmt != NarrowVTSize) return SDValue(); // If the operation feeding into the MUL is a sign extend (sext), // we use mulhs. Othewise, zero extends (zext) use mulhu. unsigned MulhOpcode = IsSignExt ? ISD::MULHS : ISD::MULHU; // Combine to mulh if mulh is legal/custom for the narrow type on the target. if (!TLI.isOperationLegalOrCustom(MulhOpcode, NarrowVT)) return SDValue(); SDValue Result = DAG.getNode(MulhOpcode, DL, NarrowVT, LeftOp.getOperand(0), MulhRightOp); return (N->getOpcode() == ISD::SRA ? DAG.getSExtOrTrunc(Result, DL, WideVT) : DAG.getZExtOrTrunc(Result, DL, WideVT)); } SDValue DAGCombiner::visitSRA(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); if (SDValue V = DAG.simplifyShift(N0, N1)) return V; EVT VT = N0.getValueType(); unsigned OpSizeInBits = VT.getScalarSizeInBits(); // fold (sra c1, c2) -> (sra c1, c2) if (SDValue C = DAG.FoldConstantArithmetic(ISD::SRA, SDLoc(N), VT, {N0, N1})) return C; // Arithmetic shifting an all-sign-bit value is a no-op. // fold (sra 0, x) -> 0 // fold (sra -1, x) -> -1 if (DAG.ComputeNumSignBits(N0) == OpSizeInBits) return N0; // fold vector ops if (VT.isVector()) if (SDValue FoldedVOp = SimplifyVBinOp(N, SDLoc(N))) return FoldedVOp; if (SDValue NewSel = foldBinOpIntoSelect(N)) return NewSel; // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports // sext_inreg. ConstantSDNode *N1C = isConstOrConstSplat(N1); if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) { unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue(); EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits); if (VT.isVector()) ExtVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, VT.getVectorElementCount()); if (!LegalOperations || TLI.getOperationAction(ISD::SIGN_EXTEND_INREG, ExtVT) == TargetLowering::Legal) return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0.getOperand(0), DAG.getValueType(ExtVT)); // Even if we can't convert to sext_inreg, we might be able to remove // this shift pair if the input is already sign extended. if (DAG.ComputeNumSignBits(N0.getOperand(0)) > N1C->getZExtValue()) return N0.getOperand(0); } // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2)) // clamp (add c1, c2) to max shift. if (N0.getOpcode() == ISD::SRA) { SDLoc DL(N); EVT ShiftVT = N1.getValueType(); EVT ShiftSVT = ShiftVT.getScalarType(); SmallVector ShiftValues; auto SumOfShifts = [&](ConstantSDNode *LHS, ConstantSDNode *RHS) { APInt c1 = LHS->getAPIntValue(); APInt c2 = RHS->getAPIntValue(); zeroExtendToMatch(c1, c2, 1 /* Overflow Bit */); APInt Sum = c1 + c2; unsigned ShiftSum = Sum.uge(OpSizeInBits) ? (OpSizeInBits - 1) : Sum.getZExtValue(); ShiftValues.push_back(DAG.getConstant(ShiftSum, DL, ShiftSVT)); return true; }; if (ISD::matchBinaryPredicate(N1, N0.getOperand(1), SumOfShifts)) { SDValue ShiftValue; if (N1.getOpcode() == ISD::BUILD_VECTOR) ShiftValue = DAG.getBuildVector(ShiftVT, DL, ShiftValues); else if (N1.getOpcode() == ISD::SPLAT_VECTOR) { assert(ShiftValues.size() == 1 && "Expected matchBinaryPredicate to return one element for " "SPLAT_VECTORs"); ShiftValue = DAG.getSplatVector(ShiftVT, DL, ShiftValues[0]); } else ShiftValue = ShiftValues[0]; return DAG.getNode(ISD::SRA, DL, VT, N0.getOperand(0), ShiftValue); } } // fold (sra (shl X, m), (sub result_size, n)) // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for // result_size - n != m. // If truncate is free for the target sext(shl) is likely to result in better // code. if (N0.getOpcode() == ISD::SHL && N1C) { // Get the two constanst of the shifts, CN0 = m, CN = n. const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1)); if (N01C) { LLVMContext &Ctx = *DAG.getContext(); // Determine what the truncate's result bitsize and type would be. EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue()); if (VT.isVector()) TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorElementCount()); // Determine the residual right-shift amount. int ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue(); // If the shift is not a no-op (in which case this should be just a sign // extend already), the truncated to type is legal, sign_extend is legal // on that type, and the truncate to that type is both legal and free, // perform the transform. if ((ShiftAmt > 0) && TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) && TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) && TLI.isTruncateFree(VT, TruncVT)) { SDLoc DL(N); SDValue Amt = DAG.getConstant(ShiftAmt, DL, getShiftAmountTy(N0.getOperand(0).getValueType())); SDValue Shift = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0), Amt); SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, TruncVT, Shift); return DAG.getNode(ISD::SIGN_EXTEND, DL, N->getValueType(0), Trunc); } } } // We convert trunc/ext to opposing shifts in IR, but casts may be cheaper. // sra (add (shl X, N1C), AddC), N1C --> // sext (add (trunc X to (width - N1C)), AddC') // sra (sub AddC, (shl X, N1C)), N1C --> // sext (sub AddC1',(trunc X to (width - N1C))) if ((N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB) && N1C && N0.hasOneUse()) { bool IsAdd = N0.getOpcode() == ISD::ADD; SDValue Shl = N0.getOperand(IsAdd ? 0 : 1); if (Shl.getOpcode() == ISD::SHL && Shl.getOperand(1) == N1 && Shl.hasOneUse()) { // TODO: AddC does not need to be a splat. if (ConstantSDNode *AddC = isConstOrConstSplat(N0.getOperand(IsAdd ? 1 : 0))) { // Determine what the truncate's type would be and ask the target if // that is a free operation. LLVMContext &Ctx = *DAG.getContext(); unsigned ShiftAmt = N1C->getZExtValue(); EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - ShiftAmt); if (VT.isVector()) TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorElementCount()); // TODO: The simple type check probably belongs in the default hook // implementation and/or target-specific overrides (because // non-simple types likely require masking when legalized), but // that restriction may conflict with other transforms. if (TruncVT.isSimple() && isTypeLegal(TruncVT) && TLI.isTruncateFree(VT, TruncVT)) { SDLoc DL(N); SDValue Trunc = DAG.getZExtOrTrunc(Shl.getOperand(0), DL, TruncVT); SDValue ShiftC = DAG.getConstant(AddC->getAPIntValue().lshr(ShiftAmt).trunc( TruncVT.getScalarSizeInBits()), DL, TruncVT); SDValue Add; if (IsAdd) Add = DAG.getNode(ISD::ADD, DL, TruncVT, Trunc, ShiftC); else Add = DAG.getNode(ISD::SUB, DL, TruncVT, ShiftC, Trunc); return DAG.getSExtOrTrunc(Add, DL, VT); } } } } // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))). if (N1.getOpcode() == ISD::TRUNCATE && N1.getOperand(0).getOpcode() == ISD::AND) { if (SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode())) return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1); } // fold (sra (trunc (sra x, c1)), c2) -> (trunc (sra x, c1 + c2)) // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2)) // if c1 is equal to the number of bits the trunc removes // TODO - support non-uniform vector shift amounts. if (N0.getOpcode() == ISD::TRUNCATE && (N0.getOperand(0).getOpcode() == ISD::SRL || N0.getOperand(0).getOpcode() == ISD::SRA) && N0.getOperand(0).hasOneUse() && N0.getOperand(0).getOperand(1).hasOneUse() && N1C) { SDValue N0Op0 = N0.getOperand(0); if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) { EVT LargeVT = N0Op0.getValueType(); unsigned TruncBits = LargeVT.getScalarSizeInBits() - OpSizeInBits; if (LargeShift->getAPIntValue() == TruncBits) { SDLoc DL(N); EVT LargeShiftVT = getShiftAmountTy(LargeVT); SDValue Amt = DAG.getZExtOrTrunc(N1, DL, LargeShiftVT); Amt = DAG.getNode(ISD::ADD, DL, LargeShiftVT, Amt, DAG.getConstant(TruncBits, DL, LargeShiftVT)); SDValue SRA = DAG.getNode(ISD::SRA, DL, LargeVT, N0Op0.getOperand(0), Amt); return DAG.getNode(ISD::TRUNCATE, DL, VT, SRA); } } } // Simplify, based on bits shifted out of the LHS. if (SimplifyDemandedBits(SDValue(N, 0))) return SDValue(N, 0); // If the sign bit is known to be zero, switch this to a SRL. if (DAG.SignBitIsZero(N0)) return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1); if (N1C && !N1C->isOpaque()) if (SDValue NewSRA = visitShiftByConstant(N)) return NewSRA; // Try to transform this shift into a multiply-high if // it matches the appropriate pattern detected in combineShiftToMULH. if (SDValue MULH = combineShiftToMULH(N, DAG, TLI)) return MULH; // Attempt to convert a sra of a load into a narrower sign-extending load. if (SDValue NarrowLoad = reduceLoadWidth(N)) return NarrowLoad; return SDValue(); } SDValue DAGCombiner::visitSRL(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); if (SDValue V = DAG.simplifyShift(N0, N1)) return V; EVT VT = N0.getValueType(); EVT ShiftVT = N1.getValueType(); unsigned OpSizeInBits = VT.getScalarSizeInBits(); // fold (srl c1, c2) -> c1 >>u c2 if (SDValue C = DAG.FoldConstantArithmetic(ISD::SRL, SDLoc(N), VT, {N0, N1})) return C; // fold vector ops if (VT.isVector()) if (SDValue FoldedVOp = SimplifyVBinOp(N, SDLoc(N))) return FoldedVOp; if (SDValue NewSel = foldBinOpIntoSelect(N)) return NewSel; // if (srl x, c) is known to be zero, return 0 ConstantSDNode *N1C = isConstOrConstSplat(N1); if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0), APInt::getAllOnes(OpSizeInBits))) return DAG.getConstant(0, SDLoc(N), VT); // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2)) if (N0.getOpcode() == ISD::SRL) { auto MatchOutOfRange = [OpSizeInBits](ConstantSDNode *LHS, ConstantSDNode *RHS) { APInt c1 = LHS->getAPIntValue(); APInt c2 = RHS->getAPIntValue(); zeroExtendToMatch(c1, c2, 1 /* Overflow Bit */); return (c1 + c2).uge(OpSizeInBits); }; if (ISD::matchBinaryPredicate(N1, N0.getOperand(1), MatchOutOfRange)) return DAG.getConstant(0, SDLoc(N), VT); auto MatchInRange = [OpSizeInBits](ConstantSDNode *LHS, ConstantSDNode *RHS) { APInt c1 = LHS->getAPIntValue(); APInt c2 = RHS->getAPIntValue(); zeroExtendToMatch(c1, c2, 1 /* Overflow Bit */); return (c1 + c2).ult(OpSizeInBits); }; if (ISD::matchBinaryPredicate(N1, N0.getOperand(1), MatchInRange)) { SDLoc DL(N); SDValue Sum = DAG.getNode(ISD::ADD, DL, ShiftVT, N1, N0.getOperand(1)); return DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0), Sum); } } if (N1C && N0.getOpcode() == ISD::TRUNCATE && N0.getOperand(0).getOpcode() == ISD::SRL) { SDValue InnerShift = N0.getOperand(0); // TODO - support non-uniform vector shift amounts. if (auto *N001C = isConstOrConstSplat(InnerShift.getOperand(1))) { uint64_t c1 = N001C->getZExtValue(); uint64_t c2 = N1C->getZExtValue(); EVT InnerShiftVT = InnerShift.getValueType(); EVT ShiftAmtVT = InnerShift.getOperand(1).getValueType(); uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits(); // srl (trunc (srl x, c1)), c2 --> 0 or (trunc (srl x, (add c1, c2))) // This is only valid if the OpSizeInBits + c1 = size of inner shift. if (c1 + OpSizeInBits == InnerShiftSize) { SDLoc DL(N); if (c1 + c2 >= InnerShiftSize) return DAG.getConstant(0, DL, VT); SDValue NewShiftAmt = DAG.getConstant(c1 + c2, DL, ShiftAmtVT); SDValue NewShift = DAG.getNode(ISD::SRL, DL, InnerShiftVT, InnerShift.getOperand(0), NewShiftAmt); return DAG.getNode(ISD::TRUNCATE, DL, VT, NewShift); } // In the more general case, we can clear the high bits after the shift: // srl (trunc (srl x, c1)), c2 --> trunc (and (srl x, (c1+c2)), Mask) if (N0.hasOneUse() && InnerShift.hasOneUse() && c1 + c2 < InnerShiftSize) { SDLoc DL(N); SDValue NewShiftAmt = DAG.getConstant(c1 + c2, DL, ShiftAmtVT); SDValue NewShift = DAG.getNode(ISD::SRL, DL, InnerShiftVT, InnerShift.getOperand(0), NewShiftAmt); SDValue Mask = DAG.getConstant(APInt::getLowBitsSet(InnerShiftSize, OpSizeInBits - c2), DL, InnerShiftVT); SDValue And = DAG.getNode(ISD::AND, DL, InnerShiftVT, NewShift, Mask); return DAG.getNode(ISD::TRUNCATE, DL, VT, And); } } } // fold (srl (shl x, c1), c2) -> (and (shl x, (sub c1, c2), MASK) or // (and (srl x, (sub c2, c1), MASK) if (N0.getOpcode() == ISD::SHL && (N0.getOperand(1) == N1 || N0->hasOneUse()) && TLI.shouldFoldConstantShiftPairToMask(N, Level)) { auto MatchShiftAmount = [OpSizeInBits](ConstantSDNode *LHS, ConstantSDNode *RHS) { const APInt &LHSC = LHS->getAPIntValue(); const APInt &RHSC = RHS->getAPIntValue(); return LHSC.ult(OpSizeInBits) && RHSC.ult(OpSizeInBits) && LHSC.getZExtValue() <= RHSC.getZExtValue(); }; if (ISD::matchBinaryPredicate(N1, N0.getOperand(1), MatchShiftAmount, /*AllowUndefs*/ false, /*AllowTypeMismatch*/ true)) { SDLoc DL(N); SDValue N01 = DAG.getZExtOrTrunc(N0.getOperand(1), DL, ShiftVT); SDValue Diff = DAG.getNode(ISD::SUB, DL, ShiftVT, N01, N1); SDValue Mask = DAG.getAllOnesConstant(DL, VT); Mask = DAG.getNode(ISD::SRL, DL, VT, Mask, N01); Mask = DAG.getNode(ISD::SHL, DL, VT, Mask, Diff); SDValue Shift = DAG.getNode(ISD::SHL, DL, VT, N0.getOperand(0), Diff); return DAG.getNode(ISD::AND, DL, VT, Shift, Mask); } if (ISD::matchBinaryPredicate(N0.getOperand(1), N1, MatchShiftAmount, /*AllowUndefs*/ false, /*AllowTypeMismatch*/ true)) { SDLoc DL(N); SDValue N01 = DAG.getZExtOrTrunc(N0.getOperand(1), DL, ShiftVT); SDValue Diff = DAG.getNode(ISD::SUB, DL, ShiftVT, N1, N01); SDValue Mask = DAG.getAllOnesConstant(DL, VT); Mask = DAG.getNode(ISD::SRL, DL, VT, Mask, N1); SDValue Shift = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0), Diff); return DAG.getNode(ISD::AND, DL, VT, Shift, Mask); } } // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask) // TODO - support non-uniform vector shift amounts. if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) { // Shifting in all undef bits? EVT SmallVT = N0.getOperand(0).getValueType(); unsigned BitSize = SmallVT.getScalarSizeInBits(); if (N1C->getAPIntValue().uge(BitSize)) return DAG.getUNDEF(VT); if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) { uint64_t ShiftAmt = N1C->getZExtValue(); SDLoc DL0(N0); SDValue SmallShift = DAG.getNode(ISD::SRL, DL0, SmallVT, N0.getOperand(0), DAG.getConstant(ShiftAmt, DL0, getShiftAmountTy(SmallVT))); AddToWorklist(SmallShift.getNode()); APInt Mask = APInt::getLowBitsSet(OpSizeInBits, OpSizeInBits - ShiftAmt); SDLoc DL(N); return DAG.getNode(ISD::AND, DL, VT, DAG.getNode(ISD::ANY_EXTEND, DL, VT, SmallShift), DAG.getConstant(Mask, DL, VT)); } } // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign // bit, which is unmodified by sra. if (N1C && N1C->getAPIntValue() == (OpSizeInBits - 1)) { if (N0.getOpcode() == ISD::SRA) return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1); } // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit). if (N1C && N0.getOpcode() == ISD::CTLZ && N1C->getAPIntValue() == Log2_32(OpSizeInBits)) { KnownBits Known = DAG.computeKnownBits(N0.getOperand(0)); // If any of the input bits are KnownOne, then the input couldn't be all // zeros, thus the result of the srl will always be zero. if (Known.One.getBoolValue()) return DAG.getConstant(0, SDLoc(N0), VT); // If all of the bits input the to ctlz node are known to be zero, then // the result of the ctlz is "32" and the result of the shift is one. APInt UnknownBits = ~Known.Zero; if (UnknownBits == 0) return DAG.getConstant(1, SDLoc(N0), VT); // Otherwise, check to see if there is exactly one bit input to the ctlz. if (UnknownBits.isPowerOf2()) { // Okay, we know that only that the single bit specified by UnknownBits // could be set on input to the CTLZ node. If this bit is set, the SRL // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair // to an SRL/XOR pair, which is likely to simplify more. unsigned ShAmt = UnknownBits.countTrailingZeros(); SDValue Op = N0.getOperand(0); if (ShAmt) { SDLoc DL(N0); Op = DAG.getNode(ISD::SRL, DL, VT, Op, DAG.getConstant(ShAmt, DL, getShiftAmountTy(Op.getValueType()))); AddToWorklist(Op.getNode()); } SDLoc DL(N); return DAG.getNode(ISD::XOR, DL, VT, Op, DAG.getConstant(1, DL, VT)); } } // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))). if (N1.getOpcode() == ISD::TRUNCATE && N1.getOperand(0).getOpcode() == ISD::AND) { if (SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode())) return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1); } // fold operands of srl based on knowledge that the low bits are not // demanded. if (SimplifyDemandedBits(SDValue(N, 0))) return SDValue(N, 0); if (N1C && !N1C->isOpaque()) if (SDValue NewSRL = visitShiftByConstant(N)) return NewSRL; // Attempt to convert a srl of a load into a narrower zero-extending load. if (SDValue NarrowLoad = reduceLoadWidth(N)) return NarrowLoad; // Here is a common situation. We want to optimize: // // %a = ... // %b = and i32 %a, 2 // %c = srl i32 %b, 1 // brcond i32 %c ... // // into // // %a = ... // %b = and %a, 2 // %c = setcc eq %b, 0 // brcond %c ... // // However when after the source operand of SRL is optimized into AND, the SRL // itself may not be optimized further. Look for it and add the BRCOND into // the worklist. if (N->hasOneUse()) { SDNode *Use = *N->use_begin(); if (Use->getOpcode() == ISD::BRCOND) AddToWorklist(Use); else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) { // Also look pass the truncate. Use = *Use->use_begin(); if (Use->getOpcode() == ISD::BRCOND) AddToWorklist(Use); } } // Try to transform this shift into a multiply-high if // it matches the appropriate pattern detected in combineShiftToMULH. if (SDValue MULH = combineShiftToMULH(N, DAG, TLI)) return MULH; return SDValue(); } SDValue DAGCombiner::visitFunnelShift(SDNode *N) { EVT VT = N->getValueType(0); SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); SDValue N2 = N->getOperand(2); bool IsFSHL = N->getOpcode() == ISD::FSHL; unsigned BitWidth = VT.getScalarSizeInBits(); // fold (fshl N0, N1, 0) -> N0 // fold (fshr N0, N1, 0) -> N1 if (isPowerOf2_32(BitWidth)) if (DAG.MaskedValueIsZero( N2, APInt(N2.getScalarValueSizeInBits(), BitWidth - 1))) return IsFSHL ? N0 : N1; auto IsUndefOrZero = [](SDValue V) { return V.isUndef() || isNullOrNullSplat(V, /*AllowUndefs*/ true); }; // TODO - support non-uniform vector shift amounts. if (ConstantSDNode *Cst = isConstOrConstSplat(N2)) { EVT ShAmtTy = N2.getValueType(); // fold (fsh* N0, N1, c) -> (fsh* N0, N1, c % BitWidth) if (Cst->getAPIntValue().uge(BitWidth)) { uint64_t RotAmt = Cst->getAPIntValue().urem(BitWidth); return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N0, N1, DAG.getConstant(RotAmt, SDLoc(N), ShAmtTy)); } unsigned ShAmt = Cst->getZExtValue(); if (ShAmt == 0) return IsFSHL ? N0 : N1; // fold fshl(undef_or_zero, N1, C) -> lshr(N1, BW-C) // fold fshr(undef_or_zero, N1, C) -> lshr(N1, C) // fold fshl(N0, undef_or_zero, C) -> shl(N0, C) // fold fshr(N0, undef_or_zero, C) -> shl(N0, BW-C) if (IsUndefOrZero(N0)) return DAG.getNode(ISD::SRL, SDLoc(N), VT, N1, DAG.getConstant(IsFSHL ? BitWidth - ShAmt : ShAmt, SDLoc(N), ShAmtTy)); if (IsUndefOrZero(N1)) return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, DAG.getConstant(IsFSHL ? ShAmt : BitWidth - ShAmt, SDLoc(N), ShAmtTy)); // fold (fshl ld1, ld0, c) -> (ld0[ofs]) iff ld0 and ld1 are consecutive. // fold (fshr ld1, ld0, c) -> (ld0[ofs]) iff ld0 and ld1 are consecutive. // TODO - bigendian support once we have test coverage. // TODO - can we merge this with CombineConseutiveLoads/MatchLoadCombine? // TODO - permit LHS EXTLOAD if extensions are shifted out. if ((BitWidth % 8) == 0 && (ShAmt % 8) == 0 && !VT.isVector() && !DAG.getDataLayout().isBigEndian()) { auto *LHS = dyn_cast(N0); auto *RHS = dyn_cast(N1); if (LHS && RHS && LHS->isSimple() && RHS->isSimple() && LHS->getAddressSpace() == RHS->getAddressSpace() && (LHS->hasOneUse() || RHS->hasOneUse()) && ISD::isNON_EXTLoad(RHS) && ISD::isNON_EXTLoad(LHS)) { if (DAG.areNonVolatileConsecutiveLoads(LHS, RHS, BitWidth / 8, 1)) { SDLoc DL(RHS); uint64_t PtrOff = IsFSHL ? (((BitWidth - ShAmt) % BitWidth) / 8) : (ShAmt / 8); Align NewAlign = commonAlignment(RHS->getAlign(), PtrOff); bool Fast = false; if (TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT, RHS->getAddressSpace(), NewAlign, RHS->getMemOperand()->getFlags(), &Fast) && Fast) { SDValue NewPtr = DAG.getMemBasePlusOffset( RHS->getBasePtr(), TypeSize::Fixed(PtrOff), DL); AddToWorklist(NewPtr.getNode()); SDValue Load = DAG.getLoad( VT, DL, RHS->getChain(), NewPtr, RHS->getPointerInfo().getWithOffset(PtrOff), NewAlign, RHS->getMemOperand()->getFlags(), RHS->getAAInfo()); // Replace the old load's chain with the new load's chain. WorklistRemover DeadNodes(*this); DAG.ReplaceAllUsesOfValueWith(N1.getValue(1), Load.getValue(1)); return Load; } } } } } // fold fshr(undef_or_zero, N1, N2) -> lshr(N1, N2) // fold fshl(N0, undef_or_zero, N2) -> shl(N0, N2) // iff We know the shift amount is in range. // TODO: when is it worth doing SUB(BW, N2) as well? if (isPowerOf2_32(BitWidth)) { APInt ModuloBits(N2.getScalarValueSizeInBits(), BitWidth - 1); if (IsUndefOrZero(N0) && !IsFSHL && DAG.MaskedValueIsZero(N2, ~ModuloBits)) return DAG.getNode(ISD::SRL, SDLoc(N), VT, N1, N2); if (IsUndefOrZero(N1) && IsFSHL && DAG.MaskedValueIsZero(N2, ~ModuloBits)) return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, N2); } // fold (fshl N0, N0, N2) -> (rotl N0, N2) // fold (fshr N0, N0, N2) -> (rotr N0, N2) // TODO: Investigate flipping this rotate if only one is legal, if funnel shift // is legal as well we might be better off avoiding non-constant (BW - N2). unsigned RotOpc = IsFSHL ? ISD::ROTL : ISD::ROTR; if (N0 == N1 && hasOperation(RotOpc, VT)) return DAG.getNode(RotOpc, SDLoc(N), VT, N0, N2); // Simplify, based on bits shifted out of N0/N1. if (SimplifyDemandedBits(SDValue(N, 0))) return SDValue(N, 0); return SDValue(); } SDValue DAGCombiner::visitSHLSAT(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); if (SDValue V = DAG.simplifyShift(N0, N1)) return V; EVT VT = N0.getValueType(); // fold (*shlsat c1, c2) -> c1<getOpcode(), SDLoc(N), VT, {N0, N1})) return C; ConstantSDNode *N1C = isConstOrConstSplat(N1); if (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::SHL, VT)) { // fold (sshlsat x, c) -> (shl x, c) if (N->getOpcode() == ISD::SSHLSAT && N1C && N1C->getAPIntValue().ult(DAG.ComputeNumSignBits(N0))) return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, N1); // fold (ushlsat x, c) -> (shl x, c) if (N->getOpcode() == ISD::USHLSAT && N1C && N1C->getAPIntValue().ule( DAG.computeKnownBits(N0).countMinLeadingZeros())) return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, N1); } return SDValue(); } // Given a ABS node, detect the following pattern: // (ABS (SUB (EXTEND a), (EXTEND b))). // Generates UABD/SABD instruction. static SDValue combineABSToABD(SDNode *N, SelectionDAG &DAG, const TargetLowering &TLI) { SDValue AbsOp1 = N->getOperand(0); SDValue Op0, Op1; if (AbsOp1.getOpcode() != ISD::SUB) return SDValue(); Op0 = AbsOp1.getOperand(0); Op1 = AbsOp1.getOperand(1); unsigned Opc0 = Op0.getOpcode(); // Check if the operands of the sub are (zero|sign)-extended. if (Opc0 != Op1.getOpcode() || (Opc0 != ISD::ZERO_EXTEND && Opc0 != ISD::SIGN_EXTEND)) return SDValue(); EVT VT = N->getValueType(0); EVT VT1 = Op0.getOperand(0).getValueType(); EVT VT2 = Op1.getOperand(0).getValueType(); unsigned ABDOpcode = (Opc0 == ISD::SIGN_EXTEND) ? ISD::ABDS : ISD::ABDU; // fold abs(sext(x) - sext(y)) -> zext(abds(x, y)) // fold abs(zext(x) - zext(y)) -> zext(abdu(x, y)) // NOTE: Extensions must be equivalent. if (VT1 == VT2 && TLI.isOperationLegalOrCustom(ABDOpcode, VT1)) { Op0 = Op0.getOperand(0); Op1 = Op1.getOperand(0); SDValue ABD = DAG.getNode(ABDOpcode, SDLoc(N), VT1, Op0, Op1); return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, ABD); } // fold abs(sext(x) - sext(y)) -> abds(sext(x), sext(y)) // fold abs(zext(x) - zext(y)) -> abdu(zext(x), zext(y)) if (TLI.isOperationLegalOrCustom(ABDOpcode, VT)) return DAG.getNode(ABDOpcode, SDLoc(N), VT, Op0, Op1); return SDValue(); } SDValue DAGCombiner::visitABS(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); // fold (abs c1) -> c2 if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) return DAG.getNode(ISD::ABS, SDLoc(N), VT, N0); // fold (abs (abs x)) -> (abs x) if (N0.getOpcode() == ISD::ABS) return N0; // fold (abs x) -> x iff not-negative if (DAG.SignBitIsZero(N0)) return N0; if (SDValue ABD = combineABSToABD(N, DAG, TLI)) return ABD; return SDValue(); } SDValue DAGCombiner::visitBSWAP(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); SDLoc DL(N); // fold (bswap c1) -> c2 if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) return DAG.getNode(ISD::BSWAP, DL, VT, N0); // fold (bswap (bswap x)) -> x if (N0.getOpcode() == ISD::BSWAP) return N0.getOperand(0); // Canonicalize bswap(bitreverse(x)) -> bitreverse(bswap(x)). If bitreverse // isn't supported, it will be expanded to bswap followed by a manual reversal // of bits in each byte. By placing bswaps before bitreverse, we can remove // the two bswaps if the bitreverse gets expanded. if (N0.getOpcode() == ISD::BITREVERSE && N0.hasOneUse()) { SDValue BSwap = DAG.getNode(ISD::BSWAP, DL, VT, N0.getOperand(0)); return DAG.getNode(ISD::BITREVERSE, DL, VT, BSwap); } // fold (bswap shl(x,c)) -> (zext(bswap(trunc(shl(x,sub(c,bw/2)))))) // iff x >= bw/2 (i.e. lower half is known zero) unsigned BW = VT.getScalarSizeInBits(); if (BW >= 32 && N0.getOpcode() == ISD::SHL && N0.hasOneUse()) { auto *ShAmt = dyn_cast(N0.getOperand(1)); EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), BW / 2); if (ShAmt && ShAmt->getAPIntValue().ult(BW) && ShAmt->getZExtValue() >= (BW / 2) && (ShAmt->getZExtValue() % 16) == 0 && TLI.isTypeLegal(HalfVT) && TLI.isTruncateFree(VT, HalfVT) && (!LegalOperations || hasOperation(ISD::BSWAP, HalfVT))) { SDValue Res = N0.getOperand(0); if (uint64_t NewShAmt = (ShAmt->getZExtValue() - (BW / 2))) Res = DAG.getNode(ISD::SHL, DL, VT, Res, DAG.getConstant(NewShAmt, DL, getShiftAmountTy(VT))); Res = DAG.getZExtOrTrunc(Res, DL, HalfVT); Res = DAG.getNode(ISD::BSWAP, DL, HalfVT, Res); return DAG.getZExtOrTrunc(Res, DL, VT); } } // Try to canonicalize bswap-of-logical-shift-by-8-bit-multiple as // inverse-shift-of-bswap: // bswap (X u<< C) --> (bswap X) u>> C // bswap (X u>> C) --> (bswap X) u<< C if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) && N0.hasOneUse()) { auto *ShAmt = dyn_cast(N0.getOperand(1)); if (ShAmt && ShAmt->getAPIntValue().ult(BW) && ShAmt->getZExtValue() % 8 == 0) { SDValue NewSwap = DAG.getNode(ISD::BSWAP, DL, VT, N0.getOperand(0)); unsigned InverseShift = N0.getOpcode() == ISD::SHL ? ISD::SRL : ISD::SHL; return DAG.getNode(InverseShift, DL, VT, NewSwap, N0.getOperand(1)); } } return SDValue(); } SDValue DAGCombiner::visitBITREVERSE(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); // fold (bitreverse c1) -> c2 if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) return DAG.getNode(ISD::BITREVERSE, SDLoc(N), VT, N0); // fold (bitreverse (bitreverse x)) -> x if (N0.getOpcode() == ISD::BITREVERSE) return N0.getOperand(0); return SDValue(); } SDValue DAGCombiner::visitCTLZ(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); // fold (ctlz c1) -> c2 if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0); // If the value is known never to be zero, switch to the undef version. if (!LegalOperations || TLI.isOperationLegal(ISD::CTLZ_ZERO_UNDEF, VT)) { if (DAG.isKnownNeverZero(N0)) return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0); } return SDValue(); } SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); // fold (ctlz_zero_undef c1) -> c2 if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0); return SDValue(); } SDValue DAGCombiner::visitCTTZ(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); // fold (cttz c1) -> c2 if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0); // If the value is known never to be zero, switch to the undef version. if (!LegalOperations || TLI.isOperationLegal(ISD::CTTZ_ZERO_UNDEF, VT)) { if (DAG.isKnownNeverZero(N0)) return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0); } return SDValue(); } SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); // fold (cttz_zero_undef c1) -> c2 if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0); return SDValue(); } SDValue DAGCombiner::visitCTPOP(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); // fold (ctpop c1) -> c2 if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0); return SDValue(); } // FIXME: This should be checking for no signed zeros on individual operands, as // well as no nans. static bool isLegalToCombineMinNumMaxNum(SelectionDAG &DAG, SDValue LHS, SDValue RHS, const TargetLowering &TLI) { const TargetOptions &Options = DAG.getTarget().Options; EVT VT = LHS.getValueType(); return Options.NoSignedZerosFPMath && VT.isFloatingPoint() && TLI.isProfitableToCombineMinNumMaxNum(VT) && DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS); } /// Generate Min/Max node static SDValue combineMinNumMaxNum(const SDLoc &DL, EVT VT, SDValue LHS, SDValue RHS, SDValue True, SDValue False, ISD::CondCode CC, const TargetLowering &TLI, SelectionDAG &DAG) { if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True)) return SDValue(); EVT TransformVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT); switch (CC) { case ISD::SETOLT: case ISD::SETOLE: case ISD::SETLT: case ISD::SETLE: case ISD::SETULT: case ISD::SETULE: { // Since it's known never nan to get here already, either fminnum or // fminnum_ieee are OK. Try the ieee version first, since it's fminnum is // expanded in terms of it. unsigned IEEEOpcode = (LHS == True) ? ISD::FMINNUM_IEEE : ISD::FMAXNUM_IEEE; if (TLI.isOperationLegalOrCustom(IEEEOpcode, VT)) return DAG.getNode(IEEEOpcode, DL, VT, LHS, RHS); unsigned Opcode = (LHS == True) ? ISD::FMINNUM : ISD::FMAXNUM; if (TLI.isOperationLegalOrCustom(Opcode, TransformVT)) return DAG.getNode(Opcode, DL, VT, LHS, RHS); return SDValue(); } case ISD::SETOGT: case ISD::SETOGE: case ISD::SETGT: case ISD::SETGE: case ISD::SETUGT: case ISD::SETUGE: { unsigned IEEEOpcode = (LHS == True) ? ISD::FMAXNUM_IEEE : ISD::FMINNUM_IEEE; if (TLI.isOperationLegalOrCustom(IEEEOpcode, VT)) return DAG.getNode(IEEEOpcode, DL, VT, LHS, RHS); unsigned Opcode = (LHS == True) ? ISD::FMAXNUM : ISD::FMINNUM; if (TLI.isOperationLegalOrCustom(Opcode, TransformVT)) return DAG.getNode(Opcode, DL, VT, LHS, RHS); return SDValue(); } default: return SDValue(); } } /// If a (v)select has a condition value that is a sign-bit test, try to smear /// the condition operand sign-bit across the value width and use it as a mask. static SDValue foldSelectOfConstantsUsingSra(SDNode *N, SelectionDAG &DAG) { SDValue Cond = N->getOperand(0); SDValue C1 = N->getOperand(1); SDValue C2 = N->getOperand(2); if (!isConstantOrConstantVector(C1) || !isConstantOrConstantVector(C2)) return SDValue(); EVT VT = N->getValueType(0); if (Cond.getOpcode() != ISD::SETCC || !Cond.hasOneUse() || VT != Cond.getOperand(0).getValueType()) return SDValue(); // The inverted-condition + commuted-select variants of these patterns are // canonicalized to these forms in IR. SDValue X = Cond.getOperand(0); SDValue CondC = Cond.getOperand(1); ISD::CondCode CC = cast(Cond.getOperand(2))->get(); if (CC == ISD::SETGT && isAllOnesOrAllOnesSplat(CondC) && isAllOnesOrAllOnesSplat(C2)) { // i32 X > -1 ? C1 : -1 --> (X >>s 31) | C1 SDLoc DL(N); SDValue ShAmtC = DAG.getConstant(X.getScalarValueSizeInBits() - 1, DL, VT); SDValue Sra = DAG.getNode(ISD::SRA, DL, VT, X, ShAmtC); return DAG.getNode(ISD::OR, DL, VT, Sra, C1); } if (CC == ISD::SETLT && isNullOrNullSplat(CondC) && isNullOrNullSplat(C2)) { // i8 X < 0 ? C1 : 0 --> (X >>s 7) & C1 SDLoc DL(N); SDValue ShAmtC = DAG.getConstant(X.getScalarValueSizeInBits() - 1, DL, VT); SDValue Sra = DAG.getNode(ISD::SRA, DL, VT, X, ShAmtC); return DAG.getNode(ISD::AND, DL, VT, Sra, C1); } return SDValue(); } SDValue DAGCombiner::foldSelectOfConstants(SDNode *N) { SDValue Cond = N->getOperand(0); SDValue N1 = N->getOperand(1); SDValue N2 = N->getOperand(2); EVT VT = N->getValueType(0); EVT CondVT = Cond.getValueType(); SDLoc DL(N); if (!VT.isInteger()) return SDValue(); auto *C1 = dyn_cast(N1); auto *C2 = dyn_cast(N2); if (!C1 || !C2) return SDValue(); // Only do this before legalization to avoid conflicting with target-specific // transforms in the other direction (create a select from a zext/sext). There // is also a target-independent combine here in DAGCombiner in the other // direction for (select Cond, -1, 0) when the condition is not i1. if (CondVT == MVT::i1 && !LegalOperations) { if (C1->isZero() && C2->isOne()) { // select Cond, 0, 1 --> zext (!Cond) SDValue NotCond = DAG.getNOT(DL, Cond, MVT::i1); if (VT != MVT::i1) NotCond = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, NotCond); return NotCond; } if (C1->isZero() && C2->isAllOnes()) { // select Cond, 0, -1 --> sext (!Cond) SDValue NotCond = DAG.getNOT(DL, Cond, MVT::i1); if (VT != MVT::i1) NotCond = DAG.getNode(ISD::SIGN_EXTEND, DL, VT, NotCond); return NotCond; } if (C1->isOne() && C2->isZero()) { // select Cond, 1, 0 --> zext (Cond) if (VT != MVT::i1) Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Cond); return Cond; } if (C1->isAllOnes() && C2->isZero()) { // select Cond, -1, 0 --> sext (Cond) if (VT != MVT::i1) Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, VT, Cond); return Cond; } // Use a target hook because some targets may prefer to transform in the // other direction. if (TLI.convertSelectOfConstantsToMath(VT)) { // For any constants that differ by 1, we can transform the select into an // extend and add. const APInt &C1Val = C1->getAPIntValue(); const APInt &C2Val = C2->getAPIntValue(); if (C1Val - 1 == C2Val) { // select Cond, C1, C1-1 --> add (zext Cond), C1-1 if (VT != MVT::i1) Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Cond); return DAG.getNode(ISD::ADD, DL, VT, Cond, N2); } if (C1Val + 1 == C2Val) { // select Cond, C1, C1+1 --> add (sext Cond), C1+1 if (VT != MVT::i1) Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, VT, Cond); return DAG.getNode(ISD::ADD, DL, VT, Cond, N2); } // select Cond, Pow2, 0 --> (zext Cond) << log2(Pow2) if (C1Val.isPowerOf2() && C2Val.isZero()) { if (VT != MVT::i1) Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Cond); SDValue ShAmtC = DAG.getShiftAmountConstant(C1Val.exactLogBase2(), VT, DL); return DAG.getNode(ISD::SHL, DL, VT, Cond, ShAmtC); } if (SDValue V = foldSelectOfConstantsUsingSra(N, DAG)) return V; } return SDValue(); } // fold (select Cond, 0, 1) -> (xor Cond, 1) // We can't do this reliably if integer based booleans have different contents // to floating point based booleans. This is because we can't tell whether we // have an integer-based boolean or a floating-point-based boolean unless we // can find the SETCC that produced it and inspect its operands. This is // fairly easy if C is the SETCC node, but it can potentially be // undiscoverable (or not reasonably discoverable). For example, it could be // in another basic block or it could require searching a complicated // expression. if (CondVT.isInteger() && TLI.getBooleanContents(/*isVec*/false, /*isFloat*/true) == TargetLowering::ZeroOrOneBooleanContent && TLI.getBooleanContents(/*isVec*/false, /*isFloat*/false) == TargetLowering::ZeroOrOneBooleanContent && C1->isZero() && C2->isOne()) { SDValue NotCond = DAG.getNode(ISD::XOR, DL, CondVT, Cond, DAG.getConstant(1, DL, CondVT)); if (VT.bitsEq(CondVT)) return NotCond; return DAG.getZExtOrTrunc(NotCond, DL, VT); } return SDValue(); } static SDValue foldBoolSelectToLogic(SDNode *N, SelectionDAG &DAG) { assert((N->getOpcode() == ISD::SELECT || N->getOpcode() == ISD::VSELECT) && "Expected a (v)select"); SDValue Cond = N->getOperand(0); SDValue T = N->getOperand(1), F = N->getOperand(2); EVT VT = N->getValueType(0); if (VT != Cond.getValueType() || VT.getScalarSizeInBits() != 1) return SDValue(); // select Cond, Cond, F --> or Cond, F // select Cond, 1, F --> or Cond, F if (Cond == T || isOneOrOneSplat(T, /* AllowUndefs */ true)) return DAG.getNode(ISD::OR, SDLoc(N), VT, Cond, F); // select Cond, T, Cond --> and Cond, T // select Cond, T, 0 --> and Cond, T if (Cond == F || isNullOrNullSplat(F, /* AllowUndefs */ true)) return DAG.getNode(ISD::AND, SDLoc(N), VT, Cond, T); // select Cond, T, 1 --> or (not Cond), T if (isOneOrOneSplat(F, /* AllowUndefs */ true)) { SDValue NotCond = DAG.getNOT(SDLoc(N), Cond, VT); return DAG.getNode(ISD::OR, SDLoc(N), VT, NotCond, T); } // select Cond, 0, F --> and (not Cond), F if (isNullOrNullSplat(T, /* AllowUndefs */ true)) { SDValue NotCond = DAG.getNOT(SDLoc(N), Cond, VT); return DAG.getNode(ISD::AND, SDLoc(N), VT, NotCond, F); } return SDValue(); } static SDValue foldVSelectToSignBitSplatMask(SDNode *N, SelectionDAG &DAG) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); SDValue N2 = N->getOperand(2); EVT VT = N->getValueType(0); if (N0.getOpcode() != ISD::SETCC || !N0.hasOneUse()) return SDValue(); SDValue Cond0 = N0.getOperand(0); SDValue Cond1 = N0.getOperand(1); ISD::CondCode CC = cast(N0.getOperand(2))->get(); if (VT != Cond0.getValueType()) return SDValue(); // Match a signbit check of Cond0 as "Cond0 s<0". Swap select operands if the // compare is inverted from that pattern ("Cond0 s> -1"). if (CC == ISD::SETLT && isNullOrNullSplat(Cond1)) ; // This is the pattern we are looking for. else if (CC == ISD::SETGT && isAllOnesOrAllOnesSplat(Cond1)) std::swap(N1, N2); else return SDValue(); // (Cond0 s< 0) ? N1 : 0 --> (Cond0 s>> BW-1) & N1 if (isNullOrNullSplat(N2)) { SDLoc DL(N); SDValue ShiftAmt = DAG.getConstant(VT.getScalarSizeInBits() - 1, DL, VT); SDValue Sra = DAG.getNode(ISD::SRA, DL, VT, Cond0, ShiftAmt); return DAG.getNode(ISD::AND, DL, VT, Sra, N1); } // (Cond0 s< 0) ? -1 : N2 --> (Cond0 s>> BW-1) | N2 if (isAllOnesOrAllOnesSplat(N1)) { SDLoc DL(N); SDValue ShiftAmt = DAG.getConstant(VT.getScalarSizeInBits() - 1, DL, VT); SDValue Sra = DAG.getNode(ISD::SRA, DL, VT, Cond0, ShiftAmt); return DAG.getNode(ISD::OR, DL, VT, Sra, N2); } // If we have to invert the sign bit mask, only do that transform if the // target has a bitwise 'and not' instruction (the invert is free). // (Cond0 s< -0) ? 0 : N2 --> ~(Cond0 s>> BW-1) & N2 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); if (isNullOrNullSplat(N1) && TLI.hasAndNot(N1)) { SDLoc DL(N); SDValue ShiftAmt = DAG.getConstant(VT.getScalarSizeInBits() - 1, DL, VT); SDValue Sra = DAG.getNode(ISD::SRA, DL, VT, Cond0, ShiftAmt); SDValue Not = DAG.getNOT(DL, Sra, VT); return DAG.getNode(ISD::AND, DL, VT, Not, N2); } // TODO: There's another pattern in this family, but it may require // implementing hasOrNot() to check for profitability: // (Cond0 s> -1) ? -1 : N2 --> ~(Cond0 s>> BW-1) | N2 return SDValue(); } SDValue DAGCombiner::visitSELECT(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); SDValue N2 = N->getOperand(2); EVT VT = N->getValueType(0); EVT VT0 = N0.getValueType(); SDLoc DL(N); SDNodeFlags Flags = N->getFlags(); if (SDValue V = DAG.simplifySelect(N0, N1, N2)) return V; if (SDValue V = foldSelectOfConstants(N)) return V; if (SDValue V = foldBoolSelectToLogic(N, DAG)) return V; // If we can fold this based on the true/false value, do so. if (SimplifySelectOps(N, N1, N2)) return SDValue(N, 0); // Don't revisit N. if (VT0 == MVT::i1) { // The code in this block deals with the following 2 equivalences: // select(C0|C1, x, y) <=> select(C0, x, select(C1, x, y)) // select(C0&C1, x, y) <=> select(C0, select(C1, x, y), y) // The target can specify its preferred form with the // shouldNormalizeToSelectSequence() callback. However we always transform // to the right anyway if we find the inner select exists in the DAG anyway // and we always transform to the left side if we know that we can further // optimize the combination of the conditions. bool normalizeToSequence = TLI.shouldNormalizeToSelectSequence(*DAG.getContext(), VT); // select (and Cond0, Cond1), X, Y // -> select Cond0, (select Cond1, X, Y), Y if (N0->getOpcode() == ISD::AND && N0->hasOneUse()) { SDValue Cond0 = N0->getOperand(0); SDValue Cond1 = N0->getOperand(1); SDValue InnerSelect = DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Cond1, N1, N2, Flags); if (normalizeToSequence || !InnerSelect.use_empty()) return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Cond0, InnerSelect, N2, Flags); // Cleanup on failure. if (InnerSelect.use_empty()) recursivelyDeleteUnusedNodes(InnerSelect.getNode()); } // select (or Cond0, Cond1), X, Y -> select Cond0, X, (select Cond1, X, Y) if (N0->getOpcode() == ISD::OR && N0->hasOneUse()) { SDValue Cond0 = N0->getOperand(0); SDValue Cond1 = N0->getOperand(1); SDValue InnerSelect = DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Cond1, N1, N2, Flags); if (normalizeToSequence || !InnerSelect.use_empty()) return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Cond0, N1, InnerSelect, Flags); // Cleanup on failure. if (InnerSelect.use_empty()) recursivelyDeleteUnusedNodes(InnerSelect.getNode()); } // select Cond0, (select Cond1, X, Y), Y -> select (and Cond0, Cond1), X, Y if (N1->getOpcode() == ISD::SELECT && N1->hasOneUse()) { SDValue N1_0 = N1->getOperand(0); SDValue N1_1 = N1->getOperand(1); SDValue N1_2 = N1->getOperand(2); if (N1_2 == N2 && N0.getValueType() == N1_0.getValueType()) { // Create the actual and node if we can generate good code for it. if (!normalizeToSequence) { SDValue And = DAG.getNode(ISD::AND, DL, N0.getValueType(), N0, N1_0); return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), And, N1_1, N2, Flags); } // Otherwise see if we can optimize the "and" to a better pattern. if (SDValue Combined = visitANDLike(N0, N1_0, N)) { return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Combined, N1_1, N2, Flags); } } } // select Cond0, X, (select Cond1, X, Y) -> select (or Cond0, Cond1), X, Y if (N2->getOpcode() == ISD::SELECT && N2->hasOneUse()) { SDValue N2_0 = N2->getOperand(0); SDValue N2_1 = N2->getOperand(1); SDValue N2_2 = N2->getOperand(2); if (N2_1 == N1 && N0.getValueType() == N2_0.getValueType()) { // Create the actual or node if we can generate good code for it. if (!normalizeToSequence) { SDValue Or = DAG.getNode(ISD::OR, DL, N0.getValueType(), N0, N2_0); return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Or, N1, N2_2, Flags); } // Otherwise see if we can optimize to a better pattern. if (SDValue Combined = visitORLike(N0, N2_0, N)) return DAG.getNode(ISD::SELECT, DL, N1.getValueType(), Combined, N1, N2_2, Flags); } } } // select (not Cond), N1, N2 -> select Cond, N2, N1 if (SDValue F = extractBooleanFlip(N0, DAG, TLI, false)) { SDValue SelectOp = DAG.getSelect(DL, VT, F, N2, N1); SelectOp->setFlags(Flags); return SelectOp; } // Fold selects based on a setcc into other things, such as min/max/abs. if (N0.getOpcode() == ISD::SETCC) { SDValue Cond0 = N0.getOperand(0), Cond1 = N0.getOperand(1); ISD::CondCode CC = cast(N0.getOperand(2))->get(); // select (fcmp lt x, y), x, y -> fminnum x, y // select (fcmp gt x, y), x, y -> fmaxnum x, y // // This is OK if we don't care what happens if either operand is a NaN. if (N0.hasOneUse() && isLegalToCombineMinNumMaxNum(DAG, N1, N2, TLI)) if (SDValue FMinMax = combineMinNumMaxNum(DL, VT, Cond0, Cond1, N1, N2, CC, TLI, DAG)) return FMinMax; // Use 'unsigned add with overflow' to optimize an unsigned saturating add. // This is conservatively limited to pre-legal-operations to give targets // a chance to reverse the transform if they want to do that. Also, it is // unlikely that the pattern would be formed late, so it's probably not // worth going through the other checks. if (!LegalOperations && TLI.isOperationLegalOrCustom(ISD::UADDO, VT) && CC == ISD::SETUGT && N0.hasOneUse() && isAllOnesConstant(N1) && N2.getOpcode() == ISD::ADD && Cond0 == N2.getOperand(0)) { auto *C = dyn_cast(N2.getOperand(1)); auto *NotC = dyn_cast(Cond1); if (C && NotC && C->getAPIntValue() == ~NotC->getAPIntValue()) { // select (setcc Cond0, ~C, ugt), -1, (add Cond0, C) --> // uaddo Cond0, C; select uaddo.1, -1, uaddo.0 // // The IR equivalent of this transform would have this form: // %a = add %x, C // %c = icmp ugt %x, ~C // %r = select %c, -1, %a // => // %u = call {iN,i1} llvm.uadd.with.overflow(%x, C) // %u0 = extractvalue %u, 0 // %u1 = extractvalue %u, 1 // %r = select %u1, -1, %u0 SDVTList VTs = DAG.getVTList(VT, VT0); SDValue UAO = DAG.getNode(ISD::UADDO, DL, VTs, Cond0, N2.getOperand(1)); return DAG.getSelect(DL, VT, UAO.getValue(1), N1, UAO.getValue(0)); } } if (TLI.isOperationLegal(ISD::SELECT_CC, VT) || (!LegalOperations && TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT))) { // Any flags available in a select/setcc fold will be on the setcc as they // migrated from fcmp Flags = N0->getFlags(); SDValue SelectNode = DAG.getNode(ISD::SELECT_CC, DL, VT, Cond0, Cond1, N1, N2, N0.getOperand(2)); SelectNode->setFlags(Flags); return SelectNode; } if (SDValue NewSel = SimplifySelect(DL, N0, N1, N2)) return NewSel; } if (!VT.isVector()) if (SDValue BinOp = foldSelectOfBinops(N)) return BinOp; return SDValue(); } // This function assumes all the vselect's arguments are CONCAT_VECTOR // nodes and that the condition is a BV of ConstantSDNodes (or undefs). static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) { SDLoc DL(N); SDValue Cond = N->getOperand(0); SDValue LHS = N->getOperand(1); SDValue RHS = N->getOperand(2); EVT VT = N->getValueType(0); int NumElems = VT.getVectorNumElements(); assert(LHS.getOpcode() == ISD::CONCAT_VECTORS && RHS.getOpcode() == ISD::CONCAT_VECTORS && Cond.getOpcode() == ISD::BUILD_VECTOR); // CONCAT_VECTOR can take an arbitrary number of arguments. We only care about // binary ones here. if (LHS->getNumOperands() != 2 || RHS->getNumOperands() != 2) return SDValue(); // We're sure we have an even number of elements due to the // concat_vectors we have as arguments to vselect. // Skip BV elements until we find one that's not an UNDEF // After we find an UNDEF element, keep looping until we get to half the // length of the BV and see if all the non-undef nodes are the same. ConstantSDNode *BottomHalf = nullptr; for (int i = 0; i < NumElems / 2; ++i) { if (Cond->getOperand(i)->isUndef()) continue; if (BottomHalf == nullptr) BottomHalf = cast(Cond.getOperand(i)); else if (Cond->getOperand(i).getNode() != BottomHalf) return SDValue(); } // Do the same for the second half of the BuildVector ConstantSDNode *TopHalf = nullptr; for (int i = NumElems / 2; i < NumElems; ++i) { if (Cond->getOperand(i)->isUndef()) continue; if (TopHalf == nullptr) TopHalf = cast(Cond.getOperand(i)); else if (Cond->getOperand(i).getNode() != TopHalf) return SDValue(); } assert(TopHalf && BottomHalf && "One half of the selector was all UNDEFs and the other was all the " "same value. This should have been addressed before this function."); return DAG.getNode( ISD::CONCAT_VECTORS, DL, VT, BottomHalf->isZero() ? RHS->getOperand(0) : LHS->getOperand(0), TopHalf->isZero() ? RHS->getOperand(1) : LHS->getOperand(1)); } bool refineUniformBase(SDValue &BasePtr, SDValue &Index, bool IndexIsScaled, SelectionDAG &DAG) { if (!isNullConstant(BasePtr) || Index.getOpcode() != ISD::ADD) return false; // Only perform the transformation when existing operands can be reused. if (IndexIsScaled) return false; // For now we check only the LHS of the add. SDValue LHS = Index.getOperand(0); SDValue SplatVal = DAG.getSplatValue(LHS); if (!SplatVal || SplatVal.getValueType() != BasePtr.getValueType()) return false; BasePtr = SplatVal; Index = Index.getOperand(1); return true; } // Fold sext/zext of index into index type. bool refineIndexType(SDValue &Index, ISD::MemIndexType &IndexType, EVT DataVT, SelectionDAG &DAG) { const TargetLowering &TLI = DAG.getTargetLoweringInfo(); // It's always safe to look through zero extends. if (Index.getOpcode() == ISD::ZERO_EXTEND) { SDValue Op = Index.getOperand(0); if (TLI.shouldRemoveExtendFromGSIndex(Op.getValueType(), DataVT)) { IndexType = ISD::UNSIGNED_SCALED; Index = Op; return true; } if (ISD::isIndexTypeSigned(IndexType)) { IndexType = ISD::UNSIGNED_SCALED; return true; } } // It's only safe to look through sign extends when Index is signed. if (Index.getOpcode() == ISD::SIGN_EXTEND && ISD::isIndexTypeSigned(IndexType)) { SDValue Op = Index.getOperand(0); if (TLI.shouldRemoveExtendFromGSIndex(Op.getValueType(), DataVT)) { Index = Op; return true; } } return false; } SDValue DAGCombiner::visitMSCATTER(SDNode *N) { MaskedScatterSDNode *MSC = cast(N); SDValue Mask = MSC->getMask(); SDValue Chain = MSC->getChain(); SDValue Index = MSC->getIndex(); SDValue Scale = MSC->getScale(); SDValue StoreVal = MSC->getValue(); SDValue BasePtr = MSC->getBasePtr(); ISD::MemIndexType IndexType = MSC->getIndexType(); SDLoc DL(N); // Zap scatters with a zero mask. if (ISD::isConstantSplatVectorAllZeros(Mask.getNode())) return Chain; if (refineUniformBase(BasePtr, Index, MSC->isIndexScaled(), DAG)) { SDValue Ops[] = {Chain, StoreVal, Mask, BasePtr, Index, Scale}; return DAG.getMaskedScatter(DAG.getVTList(MVT::Other), MSC->getMemoryVT(), DL, Ops, MSC->getMemOperand(), IndexType, MSC->isTruncatingStore()); } if (refineIndexType(Index, IndexType, StoreVal.getValueType(), DAG)) { SDValue Ops[] = {Chain, StoreVal, Mask, BasePtr, Index, Scale}; return DAG.getMaskedScatter(DAG.getVTList(MVT::Other), MSC->getMemoryVT(), DL, Ops, MSC->getMemOperand(), IndexType, MSC->isTruncatingStore()); } return SDValue(); } SDValue DAGCombiner::visitMSTORE(SDNode *N) { MaskedStoreSDNode *MST = cast(N); SDValue Mask = MST->getMask(); SDValue Chain = MST->getChain(); SDValue Value = MST->getValue(); SDValue Ptr = MST->getBasePtr(); SDLoc DL(N); // Zap masked stores with a zero mask. if (ISD::isConstantSplatVectorAllZeros(Mask.getNode())) return Chain; // If this is a masked load with an all ones mask, we can use a unmasked load. // FIXME: Can we do this for indexed, compressing, or truncating stores? if (ISD::isConstantSplatVectorAllOnes(Mask.getNode()) && MST->isUnindexed() && !MST->isCompressingStore() && !MST->isTruncatingStore()) return DAG.getStore(MST->getChain(), SDLoc(N), MST->getValue(), MST->getBasePtr(), MST->getPointerInfo(), MST->getOriginalAlign(), MachineMemOperand::MOStore, MST->getAAInfo()); // Try transforming N to an indexed store. if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N)) return SDValue(N, 0); if (MST->isTruncatingStore() && MST->isUnindexed() && Value.getValueType().isInteger() && (!isa(Value) || !cast(Value)->isOpaque())) { APInt TruncDemandedBits = APInt::getLowBitsSet(Value.getScalarValueSizeInBits(), MST->getMemoryVT().getScalarSizeInBits()); // See if we can simplify the operation with // SimplifyDemandedBits, which only works if the value has a single use. if (SimplifyDemandedBits(Value, TruncDemandedBits)) { // Re-visit the store if anything changed and the store hasn't been merged // with another node (N is deleted) SimplifyDemandedBits will add Value's // node back to the worklist if necessary, but we also need to re-visit // the Store node itself. if (N->getOpcode() != ISD::DELETED_NODE) AddToWorklist(N); return SDValue(N, 0); } } // If this is a TRUNC followed by a masked store, fold this into a masked // truncating store. We can do this even if this is already a masked // truncstore. if ((Value.getOpcode() == ISD::TRUNCATE) && Value->hasOneUse() && MST->isUnindexed() && TLI.canCombineTruncStore(Value.getOperand(0).getValueType(), MST->getMemoryVT(), LegalOperations)) { auto Mask = TLI.promoteTargetBoolean(DAG, MST->getMask(), Value.getOperand(0).getValueType()); return DAG.getMaskedStore(Chain, SDLoc(N), Value.getOperand(0), Ptr, MST->getOffset(), Mask, MST->getMemoryVT(), MST->getMemOperand(), MST->getAddressingMode(), /*IsTruncating=*/true); } return SDValue(); } SDValue DAGCombiner::visitMGATHER(SDNode *N) { MaskedGatherSDNode *MGT = cast(N); SDValue Mask = MGT->getMask(); SDValue Chain = MGT->getChain(); SDValue Index = MGT->getIndex(); SDValue Scale = MGT->getScale(); SDValue PassThru = MGT->getPassThru(); SDValue BasePtr = MGT->getBasePtr(); ISD::MemIndexType IndexType = MGT->getIndexType(); SDLoc DL(N); // Zap gathers with a zero mask. if (ISD::isConstantSplatVectorAllZeros(Mask.getNode())) return CombineTo(N, PassThru, MGT->getChain()); if (refineUniformBase(BasePtr, Index, MGT->isIndexScaled(), DAG)) { SDValue Ops[] = {Chain, PassThru, Mask, BasePtr, Index, Scale}; return DAG.getMaskedGather( DAG.getVTList(N->getValueType(0), MVT::Other), MGT->getMemoryVT(), DL, Ops, MGT->getMemOperand(), IndexType, MGT->getExtensionType()); } if (refineIndexType(Index, IndexType, N->getValueType(0), DAG)) { SDValue Ops[] = {Chain, PassThru, Mask, BasePtr, Index, Scale}; return DAG.getMaskedGather( DAG.getVTList(N->getValueType(0), MVT::Other), MGT->getMemoryVT(), DL, Ops, MGT->getMemOperand(), IndexType, MGT->getExtensionType()); } return SDValue(); } SDValue DAGCombiner::visitMLOAD(SDNode *N) { MaskedLoadSDNode *MLD = cast(N); SDValue Mask = MLD->getMask(); SDLoc DL(N); // Zap masked loads with a zero mask. if (ISD::isConstantSplatVectorAllZeros(Mask.getNode())) return CombineTo(N, MLD->getPassThru(), MLD->getChain()); // If this is a masked load with an all ones mask, we can use a unmasked load. // FIXME: Can we do this for indexed, expanding, or extending loads? if (ISD::isConstantSplatVectorAllOnes(Mask.getNode()) && MLD->isUnindexed() && !MLD->isExpandingLoad() && MLD->getExtensionType() == ISD::NON_EXTLOAD) { SDValue NewLd = DAG.getLoad( N->getValueType(0), SDLoc(N), MLD->getChain(), MLD->getBasePtr(), MLD->getPointerInfo(), MLD->getOriginalAlign(), MachineMemOperand::MOLoad, MLD->getAAInfo(), MLD->getRanges()); return CombineTo(N, NewLd, NewLd.getValue(1)); } // Try transforming N to an indexed load. if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N)) return SDValue(N, 0); return SDValue(); } /// A vector select of 2 constant vectors can be simplified to math/logic to /// avoid a variable select instruction and possibly avoid constant loads. SDValue DAGCombiner::foldVSelectOfConstants(SDNode *N) { SDValue Cond = N->getOperand(0); SDValue N1 = N->getOperand(1); SDValue N2 = N->getOperand(2); EVT VT = N->getValueType(0); if (!Cond.hasOneUse() || Cond.getScalarValueSizeInBits() != 1 || !TLI.convertSelectOfConstantsToMath(VT) || !ISD::isBuildVectorOfConstantSDNodes(N1.getNode()) || !ISD::isBuildVectorOfConstantSDNodes(N2.getNode())) return SDValue(); // Check if we can use the condition value to increment/decrement a single // constant value. This simplifies a select to an add and removes a constant // load/materialization from the general case. bool AllAddOne = true; bool AllSubOne = true; unsigned Elts = VT.getVectorNumElements(); for (unsigned i = 0; i != Elts; ++i) { SDValue N1Elt = N1.getOperand(i); SDValue N2Elt = N2.getOperand(i); if (N1Elt.isUndef() || N2Elt.isUndef()) continue; if (N1Elt.getValueType() != N2Elt.getValueType()) continue; const APInt &C1 = cast(N1Elt)->getAPIntValue(); const APInt &C2 = cast(N2Elt)->getAPIntValue(); if (C1 != C2 + 1) AllAddOne = false; if (C1 != C2 - 1) AllSubOne = false; } // Further simplifications for the extra-special cases where the constants are // all 0 or all -1 should be implemented as folds of these patterns. SDLoc DL(N); if (AllAddOne || AllSubOne) { // vselect Cond, C+1, C --> add (zext Cond), C // vselect Cond, C-1, C --> add (sext Cond), C auto ExtendOpcode = AllAddOne ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND; SDValue ExtendedCond = DAG.getNode(ExtendOpcode, DL, VT, Cond); return DAG.getNode(ISD::ADD, DL, VT, ExtendedCond, N2); } // select Cond, Pow2C, 0 --> (zext Cond) << log2(Pow2C) APInt Pow2C; if (ISD::isConstantSplatVector(N1.getNode(), Pow2C) && Pow2C.isPowerOf2() && isNullOrNullSplat(N2)) { SDValue ZextCond = DAG.getZExtOrTrunc(Cond, DL, VT); SDValue ShAmtC = DAG.getConstant(Pow2C.exactLogBase2(), DL, VT); return DAG.getNode(ISD::SHL, DL, VT, ZextCond, ShAmtC); } if (SDValue V = foldSelectOfConstantsUsingSra(N, DAG)) return V; // The general case for select-of-constants: // vselect Cond, C1, C2 --> xor (and (sext Cond), (C1^C2)), C2 // ...but that only makes sense if a vselect is slower than 2 logic ops, so // leave that to a machine-specific pass. return SDValue(); } SDValue DAGCombiner::visitVSELECT(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); SDValue N2 = N->getOperand(2); EVT VT = N->getValueType(0); SDLoc DL(N); if (SDValue V = DAG.simplifySelect(N0, N1, N2)) return V; if (SDValue V = foldBoolSelectToLogic(N, DAG)) return V; // vselect (not Cond), N1, N2 -> vselect Cond, N2, N1 if (SDValue F = extractBooleanFlip(N0, DAG, TLI, false)) return DAG.getSelect(DL, VT, F, N2, N1); // Canonicalize integer abs. // vselect (setg[te] X, 0), X, -X -> // vselect (setgt X, -1), X, -X -> // vselect (setl[te] X, 0), -X, X -> // Y = sra (X, size(X)-1); xor (add (X, Y), Y) if (N0.getOpcode() == ISD::SETCC) { SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1); ISD::CondCode CC = cast(N0.getOperand(2))->get(); bool isAbs = false; bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode()); if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) || (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) && N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1)) isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode()); else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) && N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1)) isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode()); if (isAbs) { if (TLI.isOperationLegalOrCustom(ISD::ABS, VT)) return DAG.getNode(ISD::ABS, DL, VT, LHS); SDValue Shift = DAG.getNode(ISD::SRA, DL, VT, LHS, DAG.getConstant(VT.getScalarSizeInBits() - 1, DL, getShiftAmountTy(VT))); SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift); AddToWorklist(Shift.getNode()); AddToWorklist(Add.getNode()); return DAG.getNode(ISD::XOR, DL, VT, Add, Shift); } // vselect x, y (fcmp lt x, y) -> fminnum x, y // vselect x, y (fcmp gt x, y) -> fmaxnum x, y // // This is OK if we don't care about what happens if either operand is a // NaN. // if (N0.hasOneUse() && isLegalToCombineMinNumMaxNum(DAG, LHS, RHS, TLI)) { if (SDValue FMinMax = combineMinNumMaxNum(DL, VT, LHS, RHS, N1, N2, CC, TLI, DAG)) return FMinMax; } if (SDValue S = PerformMinMaxFpToSatCombine(LHS, RHS, N1, N2, CC, DAG)) return S; if (SDValue S = PerformUMinFpToSatCombine(LHS, RHS, N1, N2, CC, DAG)) return S; // If this select has a condition (setcc) with narrower operands than the // select, try to widen the compare to match the select width. // TODO: This should be extended to handle any constant. // TODO: This could be extended to handle non-loading patterns, but that // requires thorough testing to avoid regressions. if (isNullOrNullSplat(RHS)) { EVT NarrowVT = LHS.getValueType(); EVT WideVT = N1.getValueType().changeVectorElementTypeToInteger(); EVT SetCCVT = getSetCCResultType(LHS.getValueType()); unsigned SetCCWidth = SetCCVT.getScalarSizeInBits(); unsigned WideWidth = WideVT.getScalarSizeInBits(); bool IsSigned = isSignedIntSetCC(CC); auto LoadExtOpcode = IsSigned ? ISD::SEXTLOAD : ISD::ZEXTLOAD; if (LHS.getOpcode() == ISD::LOAD && LHS.hasOneUse() && SetCCWidth != 1 && SetCCWidth < WideWidth && TLI.isLoadExtLegalOrCustom(LoadExtOpcode, WideVT, NarrowVT) && TLI.isOperationLegalOrCustom(ISD::SETCC, WideVT)) { // Both compare operands can be widened for free. The LHS can use an // extended load, and the RHS is a constant: // vselect (ext (setcc load(X), C)), N1, N2 --> // vselect (setcc extload(X), C'), N1, N2 auto ExtOpcode = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; SDValue WideLHS = DAG.getNode(ExtOpcode, DL, WideVT, LHS); SDValue WideRHS = DAG.getNode(ExtOpcode, DL, WideVT, RHS); EVT WideSetCCVT = getSetCCResultType(WideVT); SDValue WideSetCC = DAG.getSetCC(DL, WideSetCCVT, WideLHS, WideRHS, CC); return DAG.getSelect(DL, N1.getValueType(), WideSetCC, N1, N2); } } // Match VSELECTs into add with unsigned saturation. if (hasOperation(ISD::UADDSAT, VT)) { // Check if one of the arms of the VSELECT is vector with all bits set. // If it's on the left side invert the predicate to simplify logic below. SDValue Other; ISD::CondCode SatCC = CC; if (ISD::isConstantSplatVectorAllOnes(N1.getNode())) { Other = N2; SatCC = ISD::getSetCCInverse(SatCC, VT.getScalarType()); } else if (ISD::isConstantSplatVectorAllOnes(N2.getNode())) { Other = N1; } if (Other && Other.getOpcode() == ISD::ADD) { SDValue CondLHS = LHS, CondRHS = RHS; SDValue OpLHS = Other.getOperand(0), OpRHS = Other.getOperand(1); // Canonicalize condition operands. if (SatCC == ISD::SETUGE) { std::swap(CondLHS, CondRHS); SatCC = ISD::SETULE; } // We can test against either of the addition operands. // x <= x+y ? x+y : ~0 --> uaddsat x, y // x+y >= x ? x+y : ~0 --> uaddsat x, y if (SatCC == ISD::SETULE && Other == CondRHS && (OpLHS == CondLHS || OpRHS == CondLHS)) return DAG.getNode(ISD::UADDSAT, DL, VT, OpLHS, OpRHS); if (OpRHS.getOpcode() == CondRHS.getOpcode() && (OpRHS.getOpcode() == ISD::BUILD_VECTOR || OpRHS.getOpcode() == ISD::SPLAT_VECTOR) && CondLHS == OpLHS) { // If the RHS is a constant we have to reverse the const // canonicalization. // x >= ~C ? x+C : ~0 --> uaddsat x, C auto MatchUADDSAT = [](ConstantSDNode *Op, ConstantSDNode *Cond) { return Cond->getAPIntValue() == ~Op->getAPIntValue(); }; if (SatCC == ISD::SETULE && ISD::matchBinaryPredicate(OpRHS, CondRHS, MatchUADDSAT)) return DAG.getNode(ISD::UADDSAT, DL, VT, OpLHS, OpRHS); } } } // Match VSELECTs into sub with unsigned saturation. if (hasOperation(ISD::USUBSAT, VT)) { // Check if one of the arms of the VSELECT is a zero vector. If it's on // the left side invert the predicate to simplify logic below. SDValue Other; ISD::CondCode SatCC = CC; if (ISD::isConstantSplatVectorAllZeros(N1.getNode())) { Other = N2; SatCC = ISD::getSetCCInverse(SatCC, VT.getScalarType()); } else if (ISD::isConstantSplatVectorAllZeros(N2.getNode())) { Other = N1; } // zext(x) >= y ? trunc(zext(x) - y) : 0 // --> usubsat(trunc(zext(x)),trunc(umin(y,SatLimit))) // zext(x) > y ? trunc(zext(x) - y) : 0 // --> usubsat(trunc(zext(x)),trunc(umin(y,SatLimit))) if (Other && Other.getOpcode() == ISD::TRUNCATE && Other.getOperand(0).getOpcode() == ISD::SUB && (SatCC == ISD::SETUGE || SatCC == ISD::SETUGT)) { SDValue OpLHS = Other.getOperand(0).getOperand(0); SDValue OpRHS = Other.getOperand(0).getOperand(1); if (LHS == OpLHS && RHS == OpRHS && LHS.getOpcode() == ISD::ZERO_EXTEND) if (SDValue R = getTruncatedUSUBSAT(VT, LHS.getValueType(), LHS, RHS, DAG, DL)) return R; } if (Other && Other.getNumOperands() == 2) { SDValue CondRHS = RHS; SDValue OpLHS = Other.getOperand(0), OpRHS = Other.getOperand(1); if (OpLHS == LHS) { // Look for a general sub with unsigned saturation first. // x >= y ? x-y : 0 --> usubsat x, y // x > y ? x-y : 0 --> usubsat x, y if ((SatCC == ISD::SETUGE || SatCC == ISD::SETUGT) && Other.getOpcode() == ISD::SUB && OpRHS == CondRHS) return DAG.getNode(ISD::USUBSAT, DL, VT, OpLHS, OpRHS); if (OpRHS.getOpcode() == ISD::BUILD_VECTOR || OpRHS.getOpcode() == ISD::SPLAT_VECTOR) { if (CondRHS.getOpcode() == ISD::BUILD_VECTOR || CondRHS.getOpcode() == ISD::SPLAT_VECTOR) { // If the RHS is a constant we have to reverse the const // canonicalization. // x > C-1 ? x+-C : 0 --> usubsat x, C auto MatchUSUBSAT = [](ConstantSDNode *Op, ConstantSDNode *Cond) { return (!Op && !Cond) || (Op && Cond && Cond->getAPIntValue() == (-Op->getAPIntValue() - 1)); }; if (SatCC == ISD::SETUGT && Other.getOpcode() == ISD::ADD && ISD::matchBinaryPredicate(OpRHS, CondRHS, MatchUSUBSAT, /*AllowUndefs*/ true)) { OpRHS = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), OpRHS); return DAG.getNode(ISD::USUBSAT, DL, VT, OpLHS, OpRHS); } // Another special case: If C was a sign bit, the sub has been // canonicalized into a xor. // FIXME: Would it be better to use computeKnownBits to // determine whether it's safe to decanonicalize the xor? // x s< 0 ? x^C : 0 --> usubsat x, C APInt SplatValue; if (SatCC == ISD::SETLT && Other.getOpcode() == ISD::XOR && ISD::isConstantSplatVector(OpRHS.getNode(), SplatValue) && ISD::isConstantSplatVectorAllZeros(CondRHS.getNode()) && SplatValue.isSignMask()) { // Note that we have to rebuild the RHS constant here to // ensure we don't rely on particular values of undef lanes. OpRHS = DAG.getConstant(SplatValue, DL, VT); return DAG.getNode(ISD::USUBSAT, DL, VT, OpLHS, OpRHS); } } } } } } } if (SimplifySelectOps(N, N1, N2)) return SDValue(N, 0); // Don't revisit N. // Fold (vselect all_ones, N1, N2) -> N1 if (ISD::isConstantSplatVectorAllOnes(N0.getNode())) return N1; // Fold (vselect all_zeros, N1, N2) -> N2 if (ISD::isConstantSplatVectorAllZeros(N0.getNode())) return N2; // The ConvertSelectToConcatVector function is assuming both the above // checks for (vselect (build_vector all{ones,zeros) ...) have been made // and addressed. if (N1.getOpcode() == ISD::CONCAT_VECTORS && N2.getOpcode() == ISD::CONCAT_VECTORS && ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) { if (SDValue CV = ConvertSelectToConcatVector(N, DAG)) return CV; } if (SDValue V = foldVSelectOfConstants(N)) return V; if (hasOperation(ISD::SRA, VT)) if (SDValue V = foldVSelectToSignBitSplatMask(N, DAG)) return V; if (SimplifyDemandedVectorElts(SDValue(N, 0))) return SDValue(N, 0); return SDValue(); } SDValue DAGCombiner::visitSELECT_CC(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); SDValue N2 = N->getOperand(2); SDValue N3 = N->getOperand(3); SDValue N4 = N->getOperand(4); ISD::CondCode CC = cast(N4)->get(); // fold select_cc lhs, rhs, x, x, cc -> x if (N2 == N3) return N2; // Determine if the condition we're dealing with is constant if (SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()), N0, N1, CC, SDLoc(N), false)) { AddToWorklist(SCC.getNode()); // cond always true -> true val // cond always false -> false val if (auto *SCCC = dyn_cast(SCC.getNode())) return SCCC->isZero() ? N3 : N2; // When the condition is UNDEF, just return the first operand. This is // coherent the DAG creation, no setcc node is created in this case if (SCC->isUndef()) return N2; // Fold to a simpler select_cc if (SCC.getOpcode() == ISD::SETCC) { SDValue SelectOp = DAG.getNode( ISD::SELECT_CC, SDLoc(N), N2.getValueType(), SCC.getOperand(0), SCC.getOperand(1), N2, N3, SCC.getOperand(2)); SelectOp->setFlags(SCC->getFlags()); return SelectOp; } } // If we can fold this based on the true/false value, do so. if (SimplifySelectOps(N, N2, N3)) return SDValue(N, 0); // Don't revisit N. // fold select_cc into other things, such as min/max/abs return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC); } SDValue DAGCombiner::visitSETCC(SDNode *N) { // setcc is very commonly used as an argument to brcond. This pattern // also lend itself to numerous combines and, as a result, it is desired // we keep the argument to a brcond as a setcc as much as possible. bool PreferSetCC = N->hasOneUse() && N->use_begin()->getOpcode() == ISD::BRCOND; ISD::CondCode Cond = cast(N->getOperand(2))->get(); EVT VT = N->getValueType(0); // SETCC(FREEZE(X), CONST, Cond) // => // FREEZE(SETCC(X, CONST, Cond)) // This is correct if FREEZE(X) has one use and SETCC(FREEZE(X), CONST, Cond) // isn't equivalent to true or false. // For example, SETCC(FREEZE(X), -128, SETULT) cannot be folded to // FREEZE(SETCC(X, -128, SETULT)) because X can be poison. // // This transformation is beneficial because visitBRCOND can fold // BRCOND(FREEZE(X)) to BRCOND(X). // Conservatively optimize integer comparisons only. if (PreferSetCC) { // Do this only when SETCC is going to be used by BRCOND. SDValue N0 = N->getOperand(0), N1 = N->getOperand(1); ConstantSDNode *N0C = dyn_cast(N0); ConstantSDNode *N1C = dyn_cast(N1); bool Updated = false; // Is 'X Cond C' always true or false? auto IsAlwaysTrueOrFalse = [](ISD::CondCode Cond, ConstantSDNode *C) { bool False = (Cond == ISD::SETULT && C->isZero()) || (Cond == ISD::SETLT && C->isMinSignedValue()) || (Cond == ISD::SETUGT && C->isAllOnes()) || (Cond == ISD::SETGT && C->isMaxSignedValue()); bool True = (Cond == ISD::SETULE && C->isAllOnes()) || (Cond == ISD::SETLE && C->isMaxSignedValue()) || (Cond == ISD::SETUGE && C->isZero()) || (Cond == ISD::SETGE && C->isMinSignedValue()); return True || False; }; if (N0->getOpcode() == ISD::FREEZE && N0.hasOneUse() && N1C) { if (!IsAlwaysTrueOrFalse(Cond, N1C)) { N0 = N0->getOperand(0); Updated = true; } } if (N1->getOpcode() == ISD::FREEZE && N1.hasOneUse() && N0C) { if (!IsAlwaysTrueOrFalse(ISD::getSetCCSwappedOperands(Cond), N0C)) { N1 = N1->getOperand(0); Updated = true; } } if (Updated) return DAG.getFreeze(DAG.getSetCC(SDLoc(N), VT, N0, N1, Cond)); } SDValue Combined = SimplifySetCC(VT, N->getOperand(0), N->getOperand(1), Cond, SDLoc(N), !PreferSetCC); if (!Combined) return SDValue(); // If we prefer to have a setcc, and we don't, we'll try our best to // recreate one using rebuildSetCC. if (PreferSetCC && Combined.getOpcode() != ISD::SETCC) { SDValue NewSetCC = rebuildSetCC(Combined); // We don't have anything interesting to combine to. if (NewSetCC.getNode() == N) return SDValue(); if (NewSetCC) return NewSetCC; } return Combined; } SDValue DAGCombiner::visitSETCCCARRY(SDNode *N) { SDValue LHS = N->getOperand(0); SDValue RHS = N->getOperand(1); SDValue Carry = N->getOperand(2); SDValue Cond = N->getOperand(3); // If Carry is false, fold to a regular SETCC. if (isNullConstant(Carry)) return DAG.getNode(ISD::SETCC, SDLoc(N), N->getVTList(), LHS, RHS, Cond); return SDValue(); } /// Check if N satisfies: /// N is used once. /// N is a Load. /// The load is compatible with ExtOpcode. It means /// If load has explicit zero/sign extension, ExpOpcode must have the same /// extension. /// Otherwise returns true. static bool isCompatibleLoad(SDValue N, unsigned ExtOpcode) { if (!N.hasOneUse()) return false; if (!isa(N)) return false; LoadSDNode *Load = cast(N); ISD::LoadExtType LoadExt = Load->getExtensionType(); if (LoadExt == ISD::NON_EXTLOAD || LoadExt == ISD::EXTLOAD) return true; // Now LoadExt is either SEXTLOAD or ZEXTLOAD, ExtOpcode must have the same // extension. if ((LoadExt == ISD::SEXTLOAD && ExtOpcode != ISD::SIGN_EXTEND) || (LoadExt == ISD::ZEXTLOAD && ExtOpcode != ISD::ZERO_EXTEND)) return false; return true; } /// Fold /// (sext (select c, load x, load y)) -> (select c, sextload x, sextload y) /// (zext (select c, load x, load y)) -> (select c, zextload x, zextload y) /// (aext (select c, load x, load y)) -> (select c, extload x, extload y) /// This function is called by the DAGCombiner when visiting sext/zext/aext /// dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND). static SDValue tryToFoldExtendSelectLoad(SDNode *N, const TargetLowering &TLI, SelectionDAG &DAG) { unsigned Opcode = N->getOpcode(); SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); SDLoc DL(N); assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND || Opcode == ISD::ANY_EXTEND) && "Expected EXTEND dag node in input!"); if (!(N0->getOpcode() == ISD::SELECT || N0->getOpcode() == ISD::VSELECT) || !N0.hasOneUse()) return SDValue(); SDValue Op1 = N0->getOperand(1); SDValue Op2 = N0->getOperand(2); if (!isCompatibleLoad(Op1, Opcode) || !isCompatibleLoad(Op2, Opcode)) return SDValue(); auto ExtLoadOpcode = ISD::EXTLOAD; if (Opcode == ISD::SIGN_EXTEND) ExtLoadOpcode = ISD::SEXTLOAD; else if (Opcode == ISD::ZERO_EXTEND) ExtLoadOpcode = ISD::ZEXTLOAD; LoadSDNode *Load1 = cast(Op1); LoadSDNode *Load2 = cast(Op2); if (!TLI.isLoadExtLegal(ExtLoadOpcode, VT, Load1->getMemoryVT()) || !TLI.isLoadExtLegal(ExtLoadOpcode, VT, Load2->getMemoryVT())) return SDValue(); SDValue Ext1 = DAG.getNode(Opcode, DL, VT, Op1); SDValue Ext2 = DAG.getNode(Opcode, DL, VT, Op2); return DAG.getSelect(DL, VT, N0->getOperand(0), Ext1, Ext2); } /// Try to fold a sext/zext/aext dag node into a ConstantSDNode or /// a build_vector of constants. /// This function is called by the DAGCombiner when visiting sext/zext/aext /// dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND). /// Vector extends are not folded if operations are legal; this is to /// avoid introducing illegal build_vector dag nodes. static SDValue tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI, SelectionDAG &DAG, bool LegalTypes) { unsigned Opcode = N->getOpcode(); SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); SDLoc DL(N); assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND || Opcode == ISD::ANY_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG || Opcode == ISD::ZERO_EXTEND_VECTOR_INREG) && "Expected EXTEND dag node in input!"); // fold (sext c1) -> c1 // fold (zext c1) -> c1 // fold (aext c1) -> c1 if (isa(N0)) return DAG.getNode(Opcode, DL, VT, N0); // fold (sext (select cond, c1, c2)) -> (select cond, sext c1, sext c2) // fold (zext (select cond, c1, c2)) -> (select cond, zext c1, zext c2) // fold (aext (select cond, c1, c2)) -> (select cond, sext c1, sext c2) if (N0->getOpcode() == ISD::SELECT) { SDValue Op1 = N0->getOperand(1); SDValue Op2 = N0->getOperand(2); if (isa(Op1) && isa(Op2) && (Opcode != ISD::ZERO_EXTEND || !TLI.isZExtFree(N0.getValueType(), VT))) { // For any_extend, choose sign extension of the constants to allow a // possible further transform to sign_extend_inreg.i.e. // // t1: i8 = select t0, Constant:i8<-1>, Constant:i8<0> // t2: i64 = any_extend t1 // --> // t3: i64 = select t0, Constant:i64<-1>, Constant:i64<0> // --> // t4: i64 = sign_extend_inreg t3 unsigned FoldOpc = Opcode; if (FoldOpc == ISD::ANY_EXTEND) FoldOpc = ISD::SIGN_EXTEND; return DAG.getSelect(DL, VT, N0->getOperand(0), DAG.getNode(FoldOpc, DL, VT, Op1), DAG.getNode(FoldOpc, DL, VT, Op2)); } } // fold (sext (build_vector AllConstants) -> (build_vector AllConstants) // fold (zext (build_vector AllConstants) -> (build_vector AllConstants) // fold (aext (build_vector AllConstants) -> (build_vector AllConstants) EVT SVT = VT.getScalarType(); if (!(VT.isVector() && (!LegalTypes || TLI.isTypeLegal(SVT)) && ISD::isBuildVectorOfConstantSDNodes(N0.getNode()))) return SDValue(); // We can fold this node into a build_vector. unsigned VTBits = SVT.getSizeInBits(); unsigned EVTBits = N0->getValueType(0).getScalarSizeInBits(); SmallVector Elts; unsigned NumElts = VT.getVectorNumElements(); // For zero-extensions, UNDEF elements still guarantee to have the upper // bits set to zero. bool IsZext = Opcode == ISD::ZERO_EXTEND || Opcode == ISD::ZERO_EXTEND_VECTOR_INREG; for (unsigned i = 0; i != NumElts; ++i) { SDValue Op = N0.getOperand(i); if (Op.isUndef()) { Elts.push_back(IsZext ? DAG.getConstant(0, DL, SVT) : DAG.getUNDEF(SVT)); continue; } SDLoc DL(Op); // Get the constant value and if needed trunc it to the size of the type. // Nodes like build_vector might have constants wider than the scalar type. APInt C = cast(Op)->getAPIntValue().zextOrTrunc(EVTBits); if (Opcode == ISD::SIGN_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG) Elts.push_back(DAG.getConstant(C.sext(VTBits), DL, SVT)); else Elts.push_back(DAG.getConstant(C.zext(VTBits), DL, SVT)); } return DAG.getBuildVector(VT, DL, Elts); } // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this: // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))" // transformation. Returns true if extension are possible and the above // mentioned transformation is profitable. static bool ExtendUsesToFormExtLoad(EVT VT, SDNode *N, SDValue N0, unsigned ExtOpc, SmallVectorImpl &ExtendNodes, const TargetLowering &TLI) { bool HasCopyToRegUses = false; bool isTruncFree = TLI.isTruncateFree(VT, N0.getValueType()); for (SDNode::use_iterator UI = N0->use_begin(), UE = N0->use_end(); UI != UE; ++UI) { SDNode *User = *UI; if (User == N) continue; if (UI.getUse().getResNo() != N0.getResNo()) continue; // FIXME: Only extend SETCC N, N and SETCC N, c for now. if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) { ISD::CondCode CC = cast(User->getOperand(2))->get(); if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC)) // Sign bits will be lost after a zext. return false; bool Add = false; for (unsigned i = 0; i != 2; ++i) { SDValue UseOp = User->getOperand(i); if (UseOp == N0) continue; if (!isa(UseOp)) return false; Add = true; } if (Add) ExtendNodes.push_back(User); continue; } // If truncates aren't free and there are users we can't // extend, it isn't worthwhile. if (!isTruncFree) return false; // Remember if this value is live-out. if (User->getOpcode() == ISD::CopyToReg) HasCopyToRegUses = true; } if (HasCopyToRegUses) { bool BothLiveOut = false; for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); UI != UE; ++UI) { SDUse &Use = UI.getUse(); if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) { BothLiveOut = true; break; } } if (BothLiveOut) // Both unextended and extended values are live out. There had better be // a good reason for the transformation. return ExtendNodes.size(); } return true; } void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl &SetCCs, SDValue OrigLoad, SDValue ExtLoad, ISD::NodeType ExtType) { // Extend SetCC uses if necessary. SDLoc DL(ExtLoad); for (SDNode *SetCC : SetCCs) { SmallVector Ops; for (unsigned j = 0; j != 2; ++j) { SDValue SOp = SetCC->getOperand(j); if (SOp == OrigLoad) Ops.push_back(ExtLoad); else Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp)); } Ops.push_back(SetCC->getOperand(2)); CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops)); } } // FIXME: Bring more similar combines here, common to sext/zext (maybe aext?). SDValue DAGCombiner::CombineExtLoad(SDNode *N) { SDValue N0 = N->getOperand(0); EVT DstVT = N->getValueType(0); EVT SrcVT = N0.getValueType(); assert((N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND) && "Unexpected node type (not an extend)!"); // fold (sext (load x)) to multiple smaller sextloads; same for zext. // For example, on a target with legal v4i32, but illegal v8i32, turn: // (v8i32 (sext (v8i16 (load x)))) // into: // (v8i32 (concat_vectors (v4i32 (sextload x)), // (v4i32 (sextload (x + 16))))) // Where uses of the original load, i.e.: // (v8i16 (load x)) // are replaced with: // (v8i16 (truncate // (v8i32 (concat_vectors (v4i32 (sextload x)), // (v4i32 (sextload (x + 16))))))) // // This combine is only applicable to illegal, but splittable, vectors. // All legal types, and illegal non-vector types, are handled elsewhere. // This combine is controlled by TargetLowering::isVectorLoadExtDesirable. // if (N0->getOpcode() != ISD::LOAD) return SDValue(); LoadSDNode *LN0 = cast(N0); if (!ISD::isNON_EXTLoad(LN0) || !ISD::isUNINDEXEDLoad(LN0) || !N0.hasOneUse() || !LN0->isSimple() || !DstVT.isVector() || !DstVT.isPow2VectorType() || !TLI.isVectorLoadExtDesirable(SDValue(N, 0))) return SDValue(); SmallVector SetCCs; if (!ExtendUsesToFormExtLoad(DstVT, N, N0, N->getOpcode(), SetCCs, TLI)) return SDValue(); ISD::LoadExtType ExtType = N->getOpcode() == ISD::SIGN_EXTEND ? ISD::SEXTLOAD : ISD::ZEXTLOAD; // Try to split the vector types to get down to legal types. EVT SplitSrcVT = SrcVT; EVT SplitDstVT = DstVT; while (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT) && SplitSrcVT.getVectorNumElements() > 1) { SplitDstVT = DAG.GetSplitDestVTs(SplitDstVT).first; SplitSrcVT = DAG.GetSplitDestVTs(SplitSrcVT).first; } if (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT)) return SDValue(); assert(!DstVT.isScalableVector() && "Unexpected scalable vector type"); SDLoc DL(N); const unsigned NumSplits = DstVT.getVectorNumElements() / SplitDstVT.getVectorNumElements(); const unsigned Stride = SplitSrcVT.getStoreSize(); SmallVector Loads; SmallVector Chains; SDValue BasePtr = LN0->getBasePtr(); for (unsigned Idx = 0; Idx < NumSplits; Idx++) { const unsigned Offset = Idx * Stride; const Align Align = commonAlignment(LN0->getAlign(), Offset); SDValue SplitLoad = DAG.getExtLoad( ExtType, SDLoc(LN0), SplitDstVT, LN0->getChain(), BasePtr, LN0->getPointerInfo().getWithOffset(Offset), SplitSrcVT, Align, LN0->getMemOperand()->getFlags(), LN0->getAAInfo()); BasePtr = DAG.getMemBasePlusOffset(BasePtr, TypeSize::Fixed(Stride), DL); Loads.push_back(SplitLoad.getValue(0)); Chains.push_back(SplitLoad.getValue(1)); } SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains); SDValue NewValue = DAG.getNode(ISD::CONCAT_VECTORS, DL, DstVT, Loads); // Simplify TF. AddToWorklist(NewChain.getNode()); CombineTo(N, NewValue); // Replace uses of the original load (before extension) // with a truncate of the concatenated sextloaded vectors. SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(), NewValue); ExtendSetCCUses(SetCCs, N0, NewValue, (ISD::NodeType)N->getOpcode()); CombineTo(N0.getNode(), Trunc, NewChain); return SDValue(N, 0); // Return N so it doesn't get rechecked! } // fold (zext (and/or/xor (shl/shr (load x), cst), cst)) -> // (and/or/xor (shl/shr (zextload x), (zext cst)), (zext cst)) SDValue DAGCombiner::CombineZExtLogicopShiftLoad(SDNode *N) { assert(N->getOpcode() == ISD::ZERO_EXTEND); EVT VT = N->getValueType(0); EVT OrigVT = N->getOperand(0).getValueType(); if (TLI.isZExtFree(OrigVT, VT)) return SDValue(); // and/or/xor SDValue N0 = N->getOperand(0); if (!(N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::XOR) || N0.getOperand(1).getOpcode() != ISD::Constant || (LegalOperations && !TLI.isOperationLegal(N0.getOpcode(), VT))) return SDValue(); // shl/shr SDValue N1 = N0->getOperand(0); if (!(N1.getOpcode() == ISD::SHL || N1.getOpcode() == ISD::SRL) || N1.getOperand(1).getOpcode() != ISD::Constant || (LegalOperations && !TLI.isOperationLegal(N1.getOpcode(), VT))) return SDValue(); // load if (!isa(N1.getOperand(0))) return SDValue(); LoadSDNode *Load = cast(N1.getOperand(0)); EVT MemVT = Load->getMemoryVT(); if (!TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT) || Load->getExtensionType() == ISD::SEXTLOAD || Load->isIndexed()) return SDValue(); // If the shift op is SHL, the logic op must be AND, otherwise the result // will be wrong. if (N1.getOpcode() == ISD::SHL && N0.getOpcode() != ISD::AND) return SDValue(); if (!N0.hasOneUse() || !N1.hasOneUse()) return SDValue(); SmallVector SetCCs; if (!ExtendUsesToFormExtLoad(VT, N1.getNode(), N1.getOperand(0), ISD::ZERO_EXTEND, SetCCs, TLI)) return SDValue(); // Actually do the transformation. SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(Load), VT, Load->getChain(), Load->getBasePtr(), Load->getMemoryVT(), Load->getMemOperand()); SDLoc DL1(N1); SDValue Shift = DAG.getNode(N1.getOpcode(), DL1, VT, ExtLoad, N1.getOperand(1)); APInt Mask = N0.getConstantOperandAPInt(1).zext(VT.getSizeInBits()); SDLoc DL0(N0); SDValue And = DAG.getNode(N0.getOpcode(), DL0, VT, Shift, DAG.getConstant(Mask, DL0, VT)); ExtendSetCCUses(SetCCs, N1.getOperand(0), ExtLoad, ISD::ZERO_EXTEND); CombineTo(N, And); if (SDValue(Load, 0).hasOneUse()) { DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), ExtLoad.getValue(1)); } else { SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(Load), Load->getValueType(0), ExtLoad); CombineTo(Load, Trunc, ExtLoad.getValue(1)); } // N0 is dead at this point. recursivelyDeleteUnusedNodes(N0.getNode()); return SDValue(N,0); // Return N so it doesn't get rechecked! } /// If we're narrowing or widening the result of a vector select and the final /// size is the same size as a setcc (compare) feeding the select, then try to /// apply the cast operation to the select's operands because matching vector /// sizes for a select condition and other operands should be more efficient. SDValue DAGCombiner::matchVSelectOpSizesWithSetCC(SDNode *Cast) { unsigned CastOpcode = Cast->getOpcode(); assert((CastOpcode == ISD::SIGN_EXTEND || CastOpcode == ISD::ZERO_EXTEND || CastOpcode == ISD::TRUNCATE || CastOpcode == ISD::FP_EXTEND || CastOpcode == ISD::FP_ROUND) && "Unexpected opcode for vector select narrowing/widening"); // We only do this transform before legal ops because the pattern may be // obfuscated by target-specific operations after legalization. Do not create // an illegal select op, however, because that may be difficult to lower. EVT VT = Cast->getValueType(0); if (LegalOperations || !TLI.isOperationLegalOrCustom(ISD::VSELECT, VT)) return SDValue(); SDValue VSel = Cast->getOperand(0); if (VSel.getOpcode() != ISD::VSELECT || !VSel.hasOneUse() || VSel.getOperand(0).getOpcode() != ISD::SETCC) return SDValue(); // Does the setcc have the same vector size as the casted select? SDValue SetCC = VSel.getOperand(0); EVT SetCCVT = getSetCCResultType(SetCC.getOperand(0).getValueType()); if (SetCCVT.getSizeInBits() != VT.getSizeInBits()) return SDValue(); // cast (vsel (setcc X), A, B) --> vsel (setcc X), (cast A), (cast B) SDValue A = VSel.getOperand(1); SDValue B = VSel.getOperand(2); SDValue CastA, CastB; SDLoc DL(Cast); if (CastOpcode == ISD::FP_ROUND) { // FP_ROUND (fptrunc) has an extra flag operand to pass along. CastA = DAG.getNode(CastOpcode, DL, VT, A, Cast->getOperand(1)); CastB = DAG.getNode(CastOpcode, DL, VT, B, Cast->getOperand(1)); } else { CastA = DAG.getNode(CastOpcode, DL, VT, A); CastB = DAG.getNode(CastOpcode, DL, VT, B); } return DAG.getNode(ISD::VSELECT, DL, VT, SetCC, CastA, CastB); } // fold ([s|z]ext ([s|z]extload x)) -> ([s|z]ext (truncate ([s|z]extload x))) // fold ([s|z]ext ( extload x)) -> ([s|z]ext (truncate ([s|z]extload x))) static SDValue tryToFoldExtOfExtload(SelectionDAG &DAG, DAGCombiner &Combiner, const TargetLowering &TLI, EVT VT, bool LegalOperations, SDNode *N, SDValue N0, ISD::LoadExtType ExtLoadType) { SDNode *N0Node = N0.getNode(); bool isAExtLoad = (ExtLoadType == ISD::SEXTLOAD) ? ISD::isSEXTLoad(N0Node) : ISD::isZEXTLoad(N0Node); if ((!isAExtLoad && !ISD::isEXTLoad(N0Node)) || !ISD::isUNINDEXEDLoad(N0Node) || !N0.hasOneUse()) return SDValue(); LoadSDNode *LN0 = cast(N0); EVT MemVT = LN0->getMemoryVT(); if ((LegalOperations || !LN0->isSimple() || VT.isVector()) && !TLI.isLoadExtLegal(ExtLoadType, VT, MemVT)) return SDValue(); SDValue ExtLoad = DAG.getExtLoad(ExtLoadType, SDLoc(LN0), VT, LN0->getChain(), LN0->getBasePtr(), MemVT, LN0->getMemOperand()); Combiner.CombineTo(N, ExtLoad); DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), ExtLoad.getValue(1)); if (LN0->use_empty()) Combiner.recursivelyDeleteUnusedNodes(LN0); return SDValue(N, 0); // Return N so it doesn't get rechecked! } // fold ([s|z]ext (load x)) -> ([s|z]ext (truncate ([s|z]extload x))) // Only generate vector extloads when 1) they're legal, and 2) they are // deemed desirable by the target. static SDValue tryToFoldExtOfLoad(SelectionDAG &DAG, DAGCombiner &Combiner, const TargetLowering &TLI, EVT VT, bool LegalOperations, SDNode *N, SDValue N0, ISD::LoadExtType ExtLoadType, ISD::NodeType ExtOpc) { // TODO: isFixedLengthVector() should be removed and any negative effects on // code generation being the result of that target's implementation of // isVectorLoadExtDesirable(). if (!ISD::isNON_EXTLoad(N0.getNode()) || !ISD::isUNINDEXEDLoad(N0.getNode()) || ((LegalOperations || VT.isFixedLengthVector() || !cast(N0)->isSimple()) && !TLI.isLoadExtLegal(ExtLoadType, VT, N0.getValueType()))) return {}; bool DoXform = true; SmallVector SetCCs; if (!N0.hasOneUse()) DoXform = ExtendUsesToFormExtLoad(VT, N, N0, ExtOpc, SetCCs, TLI); if (VT.isVector()) DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0)); if (!DoXform) return {}; LoadSDNode *LN0 = cast(N0); SDValue ExtLoad = DAG.getExtLoad(ExtLoadType, SDLoc(LN0), VT, LN0->getChain(), LN0->getBasePtr(), N0.getValueType(), LN0->getMemOperand()); Combiner.ExtendSetCCUses(SetCCs, N0, ExtLoad, ExtOpc); // If the load value is used only by N, replace it via CombineTo N. bool NoReplaceTrunc = SDValue(LN0, 0).hasOneUse(); Combiner.CombineTo(N, ExtLoad); if (NoReplaceTrunc) { DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), ExtLoad.getValue(1)); Combiner.recursivelyDeleteUnusedNodes(LN0); } else { SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(), ExtLoad); Combiner.CombineTo(LN0, Trunc, ExtLoad.getValue(1)); } return SDValue(N, 0); // Return N so it doesn't get rechecked! } static SDValue tryToFoldExtOfMaskedLoad(SelectionDAG &DAG, const TargetLowering &TLI, EVT VT, SDNode *N, SDValue N0, ISD::LoadExtType ExtLoadType, ISD::NodeType ExtOpc) { if (!N0.hasOneUse()) return SDValue(); MaskedLoadSDNode *Ld = dyn_cast(N0); if (!Ld || Ld->getExtensionType() != ISD::NON_EXTLOAD) return SDValue(); if (!TLI.isLoadExtLegalOrCustom(ExtLoadType, VT, Ld->getValueType(0))) return SDValue(); if (!TLI.isVectorLoadExtDesirable(SDValue(N, 0))) return SDValue(); SDLoc dl(Ld); SDValue PassThru = DAG.getNode(ExtOpc, dl, VT, Ld->getPassThru()); SDValue NewLoad = DAG.getMaskedLoad( VT, dl, Ld->getChain(), Ld->getBasePtr(), Ld->getOffset(), Ld->getMask(), PassThru, Ld->getMemoryVT(), Ld->getMemOperand(), Ld->getAddressingMode(), ExtLoadType, Ld->isExpandingLoad()); DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), SDValue(NewLoad.getNode(), 1)); return NewLoad; } static SDValue foldExtendedSignBitTest(SDNode *N, SelectionDAG &DAG, bool LegalOperations) { assert((N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND) && "Expected sext or zext"); SDValue SetCC = N->getOperand(0); if (LegalOperations || SetCC.getOpcode() != ISD::SETCC || !SetCC.hasOneUse() || SetCC.getValueType() != MVT::i1) return SDValue(); SDValue X = SetCC.getOperand(0); SDValue Ones = SetCC.getOperand(1); ISD::CondCode CC = cast(SetCC.getOperand(2))->get(); EVT VT = N->getValueType(0); EVT XVT = X.getValueType(); // setge X, C is canonicalized to setgt, so we do not need to match that // pattern. The setlt sibling is folded in SimplifySelectCC() because it does // not require the 'not' op. if (CC == ISD::SETGT && isAllOnesConstant(Ones) && VT == XVT) { // Invert and smear/shift the sign bit: // sext i1 (setgt iN X, -1) --> sra (not X), (N - 1) // zext i1 (setgt iN X, -1) --> srl (not X), (N - 1) SDLoc DL(N); unsigned ShCt = VT.getSizeInBits() - 1; const TargetLowering &TLI = DAG.getTargetLoweringInfo(); if (!TLI.shouldAvoidTransformToShift(VT, ShCt)) { SDValue NotX = DAG.getNOT(DL, X, VT); SDValue ShiftAmount = DAG.getConstant(ShCt, DL, VT); auto ShiftOpcode = N->getOpcode() == ISD::SIGN_EXTEND ? ISD::SRA : ISD::SRL; return DAG.getNode(ShiftOpcode, DL, VT, NotX, ShiftAmount); } } return SDValue(); } SDValue DAGCombiner::foldSextSetcc(SDNode *N) { SDValue N0 = N->getOperand(0); if (N0.getOpcode() != ISD::SETCC) return SDValue(); SDValue N00 = N0.getOperand(0); SDValue N01 = N0.getOperand(1); ISD::CondCode CC = cast(N0.getOperand(2))->get(); EVT VT = N->getValueType(0); EVT N00VT = N00.getValueType(); SDLoc DL(N); // Propagate fast-math-flags. SelectionDAG::FlagInserter FlagsInserter(DAG, N0->getFlags()); // On some architectures (such as SSE/NEON/etc) the SETCC result type is // the same size as the compared operands. Try to optimize sext(setcc()) // if this is the case. if (VT.isVector() && !LegalOperations && TLI.getBooleanContents(N00VT) == TargetLowering::ZeroOrNegativeOneBooleanContent) { EVT SVT = getSetCCResultType(N00VT); // If we already have the desired type, don't change it. if (SVT != N0.getValueType()) { // We know that the # elements of the results is the same as the // # elements of the compare (and the # elements of the compare result // for that matter). Check to see that they are the same size. If so, // we know that the element size of the sext'd result matches the // element size of the compare operands. if (VT.getSizeInBits() == SVT.getSizeInBits()) return DAG.getSetCC(DL, VT, N00, N01, CC); // If the desired elements are smaller or larger than the source // elements, we can use a matching integer vector type and then // truncate/sign extend. EVT MatchingVecType = N00VT.changeVectorElementTypeToInteger(); if (SVT == MatchingVecType) { SDValue VsetCC = DAG.getSetCC(DL, MatchingVecType, N00, N01, CC); return DAG.getSExtOrTrunc(VsetCC, DL, VT); } } // Try to eliminate the sext of a setcc by zexting the compare operands. if (N0.hasOneUse() && TLI.isOperationLegalOrCustom(ISD::SETCC, VT) && !TLI.isOperationLegalOrCustom(ISD::SETCC, SVT)) { bool IsSignedCmp = ISD::isSignedIntSetCC(CC); unsigned LoadOpcode = IsSignedCmp ? ISD::SEXTLOAD : ISD::ZEXTLOAD; unsigned ExtOpcode = IsSignedCmp ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; // We have an unsupported narrow vector compare op that would be legal // if extended to the destination type. See if the compare operands // can be freely extended to the destination type. auto IsFreeToExtend = [&](SDValue V) { if (isConstantOrConstantVector(V, /*NoOpaques*/ true)) return true; // Match a simple, non-extended load that can be converted to a // legal {z/s}ext-load. // TODO: Allow widening of an existing {z/s}ext-load? if (!(ISD::isNON_EXTLoad(V.getNode()) && ISD::isUNINDEXEDLoad(V.getNode()) && cast(V)->isSimple() && TLI.isLoadExtLegal(LoadOpcode, VT, V.getValueType()))) return false; // Non-chain users of this value must either be the setcc in this // sequence or extends that can be folded into the new {z/s}ext-load. for (SDNode::use_iterator UI = V->use_begin(), UE = V->use_end(); UI != UE; ++UI) { // Skip uses of the chain and the setcc. SDNode *User = *UI; if (UI.getUse().getResNo() != 0 || User == N0.getNode()) continue; // Extra users must have exactly the same cast we are about to create. // TODO: This restriction could be eased if ExtendUsesToFormExtLoad() // is enhanced similarly. if (User->getOpcode() != ExtOpcode || User->getValueType(0) != VT) return false; } return true; }; if (IsFreeToExtend(N00) && IsFreeToExtend(N01)) { SDValue Ext0 = DAG.getNode(ExtOpcode, DL, VT, N00); SDValue Ext1 = DAG.getNode(ExtOpcode, DL, VT, N01); return DAG.getSetCC(DL, VT, Ext0, Ext1, CC); } } } // sext(setcc x, y, cc) -> (select (setcc x, y, cc), T, 0) // Here, T can be 1 or -1, depending on the type of the setcc and // getBooleanContents(). unsigned SetCCWidth = N0.getScalarValueSizeInBits(); // To determine the "true" side of the select, we need to know the high bit // of the value returned by the setcc if it evaluates to true. // If the type of the setcc is i1, then the true case of the select is just // sext(i1 1), that is, -1. // If the type of the setcc is larger (say, i8) then the value of the high // bit depends on getBooleanContents(), so ask TLI for a real "true" value // of the appropriate width. SDValue ExtTrueVal = (SetCCWidth == 1) ? DAG.getAllOnesConstant(DL, VT) : DAG.getBoolConstant(true, DL, VT, N00VT); SDValue Zero = DAG.getConstant(0, DL, VT); if (SDValue SCC = SimplifySelectCC(DL, N00, N01, ExtTrueVal, Zero, CC, true)) return SCC; if (!VT.isVector() && !TLI.convertSelectOfConstantsToMath(VT)) { EVT SetCCVT = getSetCCResultType(N00VT); // Don't do this transform for i1 because there's a select transform // that would reverse it. // TODO: We should not do this transform at all without a target hook // because a sext is likely cheaper than a select? if (SetCCVT.getScalarSizeInBits() != 1 && (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, N00VT))) { SDValue SetCC = DAG.getSetCC(DL, SetCCVT, N00, N01, CC); return DAG.getSelect(DL, VT, SetCC, ExtTrueVal, Zero); } } return SDValue(); } SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); SDLoc DL(N); // sext(undef) = 0 because the top bit will all be the same. if (N0.isUndef()) return DAG.getConstant(0, DL, VT); if (SDValue Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes)) return Res; // fold (sext (sext x)) -> (sext x) // fold (sext (aext x)) -> (sext x) if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) return DAG.getNode(ISD::SIGN_EXTEND, DL, VT, N0.getOperand(0)); if (N0.getOpcode() == ISD::TRUNCATE) { // fold (sext (truncate (load x))) -> (sext (smaller load x)) // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n))) if (SDValue NarrowLoad = reduceLoadWidth(N0.getNode())) { SDNode *oye = N0.getOperand(0).getNode(); if (NarrowLoad.getNode() != N0.getNode()) { CombineTo(N0.getNode(), NarrowLoad); // CombineTo deleted the truncate, if needed, but not what's under it. AddToWorklist(oye); } return SDValue(N, 0); // Return N so it doesn't get rechecked! } // See if the value being truncated is already sign extended. If so, just // eliminate the trunc/sext pair. SDValue Op = N0.getOperand(0); unsigned OpBits = Op.getScalarValueSizeInBits(); unsigned MidBits = N0.getScalarValueSizeInBits(); unsigned DestBits = VT.getScalarSizeInBits(); unsigned NumSignBits = DAG.ComputeNumSignBits(Op); if (OpBits == DestBits) { // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign // bits, it is already ready. if (NumSignBits > DestBits-MidBits) return Op; } else if (OpBits < DestBits) { // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign // bits, just sext from i32. if (NumSignBits > OpBits-MidBits) return DAG.getNode(ISD::SIGN_EXTEND, DL, VT, Op); } else { // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign // bits, just truncate to i32. if (NumSignBits > OpBits-MidBits) return DAG.getNode(ISD::TRUNCATE, DL, VT, Op); } // fold (sext (truncate x)) -> (sextinreg x). if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, N0.getValueType())) { if (OpBits < DestBits) Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op); else if (OpBits > DestBits) Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op); return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, Op, DAG.getValueType(N0.getValueType())); } } // Try to simplify (sext (load x)). if (SDValue foldedExt = tryToFoldExtOfLoad(DAG, *this, TLI, VT, LegalOperations, N, N0, ISD::SEXTLOAD, ISD::SIGN_EXTEND)) return foldedExt; if (SDValue foldedExt = tryToFoldExtOfMaskedLoad(DAG, TLI, VT, N, N0, ISD::SEXTLOAD, ISD::SIGN_EXTEND)) return foldedExt; // fold (sext (load x)) to multiple smaller sextloads. // Only on illegal but splittable vectors. if (SDValue ExtLoad = CombineExtLoad(N)) return ExtLoad; // Try to simplify (sext (sextload x)). if (SDValue foldedExt = tryToFoldExtOfExtload( DAG, *this, TLI, VT, LegalOperations, N, N0, ISD::SEXTLOAD)) return foldedExt; // fold (sext (and/or/xor (load x), cst)) -> // (and/or/xor (sextload x), (sext cst)) if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::XOR) && isa(N0.getOperand(0)) && N0.getOperand(1).getOpcode() == ISD::Constant && (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) { LoadSDNode *LN00 = cast(N0.getOperand(0)); EVT MemVT = LN00->getMemoryVT(); if (TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, MemVT) && LN00->getExtensionType() != ISD::ZEXTLOAD && LN00->isUnindexed()) { SmallVector SetCCs; bool DoXform = ExtendUsesToFormExtLoad(VT, N0.getNode(), N0.getOperand(0), ISD::SIGN_EXTEND, SetCCs, TLI); if (DoXform) { SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN00), VT, LN00->getChain(), LN00->getBasePtr(), LN00->getMemoryVT(), LN00->getMemOperand()); APInt Mask = N0.getConstantOperandAPInt(1).sext(VT.getSizeInBits()); SDValue And = DAG.getNode(N0.getOpcode(), DL, VT, ExtLoad, DAG.getConstant(Mask, DL, VT)); ExtendSetCCUses(SetCCs, N0.getOperand(0), ExtLoad, ISD::SIGN_EXTEND); bool NoReplaceTruncAnd = !N0.hasOneUse(); bool NoReplaceTrunc = SDValue(LN00, 0).hasOneUse(); CombineTo(N, And); // If N0 has multiple uses, change other uses as well. if (NoReplaceTruncAnd) { SDValue TruncAnd = DAG.getNode(ISD::TRUNCATE, DL, N0.getValueType(), And); CombineTo(N0.getNode(), TruncAnd); } if (NoReplaceTrunc) { DAG.ReplaceAllUsesOfValueWith(SDValue(LN00, 1), ExtLoad.getValue(1)); } else { SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(LN00), LN00->getValueType(0), ExtLoad); CombineTo(LN00, Trunc, ExtLoad.getValue(1)); } return SDValue(N,0); // Return N so it doesn't get rechecked! } } } if (SDValue V = foldExtendedSignBitTest(N, DAG, LegalOperations)) return V; if (SDValue V = foldSextSetcc(N)) return V; // fold (sext x) -> (zext x) if the sign bit is known zero. if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) && DAG.SignBitIsZero(N0)) return DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0); if (SDValue NewVSel = matchVSelectOpSizesWithSetCC(N)) return NewVSel; // Eliminate this sign extend by doing a negation in the destination type: // sext i32 (0 - (zext i8 X to i32)) to i64 --> 0 - (zext i8 X to i64) if (N0.getOpcode() == ISD::SUB && N0.hasOneUse() && isNullOrNullSplat(N0.getOperand(0)) && N0.getOperand(1).getOpcode() == ISD::ZERO_EXTEND && TLI.isOperationLegalOrCustom(ISD::SUB, VT)) { SDValue Zext = DAG.getZExtOrTrunc(N0.getOperand(1).getOperand(0), DL, VT); return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Zext); } // Eliminate this sign extend by doing a decrement in the destination type: // sext i32 ((zext i8 X to i32) + (-1)) to i64 --> (zext i8 X to i64) + (-1) if (N0.getOpcode() == ISD::ADD && N0.hasOneUse() && isAllOnesOrAllOnesSplat(N0.getOperand(1)) && N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND && TLI.isOperationLegalOrCustom(ISD::ADD, VT)) { SDValue Zext = DAG.getZExtOrTrunc(N0.getOperand(0).getOperand(0), DL, VT); return DAG.getNode(ISD::ADD, DL, VT, Zext, DAG.getAllOnesConstant(DL, VT)); } // fold sext (not i1 X) -> add (zext i1 X), -1 // TODO: This could be extended to handle bool vectors. if (N0.getValueType() == MVT::i1 && isBitwiseNot(N0) && N0.hasOneUse() && (!LegalOperations || (TLI.isOperationLegal(ISD::ZERO_EXTEND, VT) && TLI.isOperationLegal(ISD::ADD, VT)))) { // If we can eliminate the 'not', the sext form should be better if (SDValue NewXor = visitXOR(N0.getNode())) { // Returning N0 is a form of in-visit replacement that may have // invalidated N0. if (NewXor.getNode() == N0.getNode()) { // Return SDValue here as the xor should have already been replaced in // this sext. return SDValue(); } // Return a new sext with the new xor. return DAG.getNode(ISD::SIGN_EXTEND, DL, VT, NewXor); } SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)); return DAG.getNode(ISD::ADD, DL, VT, Zext, DAG.getAllOnesConstant(DL, VT)); } if (SDValue Res = tryToFoldExtendSelectLoad(N, TLI, DAG)) return Res; return SDValue(); } // isTruncateOf - If N is a truncate of some other value, return true, record // the value being truncated in Op and which of Op's bits are zero/one in Known. // This function computes KnownBits to avoid a duplicated call to // computeKnownBits in the caller. static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op, KnownBits &Known) { if (N->getOpcode() == ISD::TRUNCATE) { Op = N->getOperand(0); Known = DAG.computeKnownBits(Op); return true; } if (N.getOpcode() != ISD::SETCC || N.getValueType().getScalarType() != MVT::i1 || cast(N.getOperand(2))->get() != ISD::SETNE) return false; SDValue Op0 = N->getOperand(0); SDValue Op1 = N->getOperand(1); assert(Op0.getValueType() == Op1.getValueType()); if (isNullOrNullSplat(Op0)) Op = Op1; else if (isNullOrNullSplat(Op1)) Op = Op0; else return false; Known = DAG.computeKnownBits(Op); return (Known.Zero | 1).isAllOnes(); } /// Given an extending node with a pop-count operand, if the target does not /// support a pop-count in the narrow source type but does support it in the /// destination type, widen the pop-count to the destination type. static SDValue widenCtPop(SDNode *Extend, SelectionDAG &DAG) { assert((Extend->getOpcode() == ISD::ZERO_EXTEND || Extend->getOpcode() == ISD::ANY_EXTEND) && "Expected extend op"); SDValue CtPop = Extend->getOperand(0); if (CtPop.getOpcode() != ISD::CTPOP || !CtPop.hasOneUse()) return SDValue(); EVT VT = Extend->getValueType(0); const TargetLowering &TLI = DAG.getTargetLoweringInfo(); if (TLI.isOperationLegalOrCustom(ISD::CTPOP, CtPop.getValueType()) || !TLI.isOperationLegalOrCustom(ISD::CTPOP, VT)) return SDValue(); // zext (ctpop X) --> ctpop (zext X) SDLoc DL(Extend); SDValue NewZext = DAG.getZExtOrTrunc(CtPop.getOperand(0), DL, VT); return DAG.getNode(ISD::CTPOP, DL, VT, NewZext); } SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); // zext(undef) = 0 if (N0.isUndef()) return DAG.getConstant(0, SDLoc(N), VT); if (SDValue Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes)) return Res; // fold (zext (zext x)) -> (zext x) // fold (zext (aext x)) -> (zext x) if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0.getOperand(0)); // fold (zext (truncate x)) -> (zext x) or // (zext (truncate x)) -> (truncate x) // This is valid when the truncated bits of x are already zero. SDValue Op; KnownBits Known; if (isTruncateOf(DAG, N0, Op, Known)) { APInt TruncatedBits = (Op.getScalarValueSizeInBits() == N0.getScalarValueSizeInBits()) ? APInt(Op.getScalarValueSizeInBits(), 0) : APInt::getBitsSet(Op.getScalarValueSizeInBits(), N0.getScalarValueSizeInBits(), std::min(Op.getScalarValueSizeInBits(), VT.getScalarSizeInBits())); if (TruncatedBits.isSubsetOf(Known.Zero)) return DAG.getZExtOrTrunc(Op, SDLoc(N), VT); } // fold (zext (truncate x)) -> (and x, mask) if (N0.getOpcode() == ISD::TRUNCATE) { // fold (zext (truncate (load x))) -> (zext (smaller load x)) // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n))) if (SDValue NarrowLoad = reduceLoadWidth(N0.getNode())) { SDNode *oye = N0.getOperand(0).getNode(); if (NarrowLoad.getNode() != N0.getNode()) { CombineTo(N0.getNode(), NarrowLoad); // CombineTo deleted the truncate, if needed, but not what's under it. AddToWorklist(oye); } return SDValue(N, 0); // Return N so it doesn't get rechecked! } EVT SrcVT = N0.getOperand(0).getValueType(); EVT MinVT = N0.getValueType(); // Try to mask before the extension to avoid having to generate a larger mask, // possibly over several sub-vectors. if (SrcVT.bitsLT(VT) && VT.isVector()) { if (!LegalOperations || (TLI.isOperationLegal(ISD::AND, SrcVT) && TLI.isOperationLegal(ISD::ZERO_EXTEND, VT))) { SDValue Op = N0.getOperand(0); Op = DAG.getZeroExtendInReg(Op, SDLoc(N), MinVT); AddToWorklist(Op.getNode()); SDValue ZExtOrTrunc = DAG.getZExtOrTrunc(Op, SDLoc(N), VT); // Transfer the debug info; the new node is equivalent to N0. DAG.transferDbgValues(N0, ZExtOrTrunc); return ZExtOrTrunc; } } if (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT)) { SDValue Op = DAG.getAnyExtOrTrunc(N0.getOperand(0), SDLoc(N), VT); AddToWorklist(Op.getNode()); SDValue And = DAG.getZeroExtendInReg(Op, SDLoc(N), MinVT); // We may safely transfer the debug info describing the truncate node over // to the equivalent and operation. DAG.transferDbgValues(N0, And); return And; } } // Fold (zext (and (trunc x), cst)) -> (and x, cst), // if either of the casts is not free. if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::TRUNCATE && N0.getOperand(1).getOpcode() == ISD::Constant && (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(), N0.getValueType()) || !TLI.isZExtFree(N0.getValueType(), VT))) { SDValue X = N0.getOperand(0).getOperand(0); X = DAG.getAnyExtOrTrunc(X, SDLoc(X), VT); APInt Mask = N0.getConstantOperandAPInt(1).zext(VT.getSizeInBits()); SDLoc DL(N); return DAG.getNode(ISD::AND, DL, VT, X, DAG.getConstant(Mask, DL, VT)); } // Try to simplify (zext (load x)). if (SDValue foldedExt = tryToFoldExtOfLoad(DAG, *this, TLI, VT, LegalOperations, N, N0, ISD::ZEXTLOAD, ISD::ZERO_EXTEND)) return foldedExt; if (SDValue foldedExt = tryToFoldExtOfMaskedLoad(DAG, TLI, VT, N, N0, ISD::ZEXTLOAD, ISD::ZERO_EXTEND)) return foldedExt; // fold (zext (load x)) to multiple smaller zextloads. // Only on illegal but splittable vectors. if (SDValue ExtLoad = CombineExtLoad(N)) return ExtLoad; // fold (zext (and/or/xor (load x), cst)) -> // (and/or/xor (zextload x), (zext cst)) // Unless (and (load x) cst) will match as a zextload already and has // additional users. if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::XOR) && isa(N0.getOperand(0)) && N0.getOperand(1).getOpcode() == ISD::Constant && (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) { LoadSDNode *LN00 = cast(N0.getOperand(0)); EVT MemVT = LN00->getMemoryVT(); if (TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT) && LN00->getExtensionType() != ISD::SEXTLOAD && LN00->isUnindexed()) { bool DoXform = true; SmallVector SetCCs; if (!N0.hasOneUse()) { if (N0.getOpcode() == ISD::AND) { auto *AndC = cast(N0.getOperand(1)); EVT LoadResultTy = AndC->getValueType(0); EVT ExtVT; if (isAndLoadExtLoad(AndC, LN00, LoadResultTy, ExtVT)) DoXform = false; } } if (DoXform) DoXform = ExtendUsesToFormExtLoad(VT, N0.getNode(), N0.getOperand(0), ISD::ZERO_EXTEND, SetCCs, TLI); if (DoXform) { SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN00), VT, LN00->getChain(), LN00->getBasePtr(), LN00->getMemoryVT(), LN00->getMemOperand()); APInt Mask = N0.getConstantOperandAPInt(1).zext(VT.getSizeInBits()); SDLoc DL(N); SDValue And = DAG.getNode(N0.getOpcode(), DL, VT, ExtLoad, DAG.getConstant(Mask, DL, VT)); ExtendSetCCUses(SetCCs, N0.getOperand(0), ExtLoad, ISD::ZERO_EXTEND); bool NoReplaceTruncAnd = !N0.hasOneUse(); bool NoReplaceTrunc = SDValue(LN00, 0).hasOneUse(); CombineTo(N, And); // If N0 has multiple uses, change other uses as well. if (NoReplaceTruncAnd) { SDValue TruncAnd = DAG.getNode(ISD::TRUNCATE, DL, N0.getValueType(), And); CombineTo(N0.getNode(), TruncAnd); } if (NoReplaceTrunc) { DAG.ReplaceAllUsesOfValueWith(SDValue(LN00, 1), ExtLoad.getValue(1)); } else { SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(LN00), LN00->getValueType(0), ExtLoad); CombineTo(LN00, Trunc, ExtLoad.getValue(1)); } return SDValue(N,0); // Return N so it doesn't get rechecked! } } } // fold (zext (and/or/xor (shl/shr (load x), cst), cst)) -> // (and/or/xor (shl/shr (zextload x), (zext cst)), (zext cst)) if (SDValue ZExtLoad = CombineZExtLogicopShiftLoad(N)) return ZExtLoad; // Try to simplify (zext (zextload x)). if (SDValue foldedExt = tryToFoldExtOfExtload( DAG, *this, TLI, VT, LegalOperations, N, N0, ISD::ZEXTLOAD)) return foldedExt; if (SDValue V = foldExtendedSignBitTest(N, DAG, LegalOperations)) return V; if (N0.getOpcode() == ISD::SETCC) { // Propagate fast-math-flags. SelectionDAG::FlagInserter FlagsInserter(DAG, N0->getFlags()); // Only do this before legalize for now. if (!LegalOperations && VT.isVector() && N0.getValueType().getVectorElementType() == MVT::i1) { EVT N00VT = N0.getOperand(0).getValueType(); if (getSetCCResultType(N00VT) == N0.getValueType()) return SDValue(); // We know that the # elements of the results is the same as the # // elements of the compare (and the # elements of the compare result for // that matter). Check to see that they are the same size. If so, we know // that the element size of the sext'd result matches the element size of // the compare operands. SDLoc DL(N); if (VT.getSizeInBits() == N00VT.getSizeInBits()) { // zext(setcc) -> zext_in_reg(vsetcc) for vectors. SDValue VSetCC = DAG.getNode(ISD::SETCC, DL, VT, N0.getOperand(0), N0.getOperand(1), N0.getOperand(2)); return DAG.getZeroExtendInReg(VSetCC, DL, N0.getValueType()); } // If the desired elements are smaller or larger than the source // elements we can use a matching integer vector type and then // truncate/any extend followed by zext_in_reg. EVT MatchingVectorType = N00VT.changeVectorElementTypeToInteger(); SDValue VsetCC = DAG.getNode(ISD::SETCC, DL, MatchingVectorType, N0.getOperand(0), N0.getOperand(1), N0.getOperand(2)); return DAG.getZeroExtendInReg(DAG.getAnyExtOrTrunc(VsetCC, DL, VT), DL, N0.getValueType()); } // zext(setcc x,y,cc) -> zext(select x, y, true, false, cc) SDLoc DL(N); EVT N0VT = N0.getValueType(); EVT N00VT = N0.getOperand(0).getValueType(); if (SDValue SCC = SimplifySelectCC( DL, N0.getOperand(0), N0.getOperand(1), DAG.getBoolConstant(true, DL, N0VT, N00VT), DAG.getBoolConstant(false, DL, N0VT, N00VT), cast(N0.getOperand(2))->get(), true)) return DAG.getNode(ISD::ZERO_EXTEND, DL, VT, SCC); } // (zext (shl (zext x), cst)) -> (shl (zext x), cst) if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) && isa(N0.getOperand(1)) && N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse()) { SDValue ShAmt = N0.getOperand(1); if (N0.getOpcode() == ISD::SHL) { SDValue InnerZExt = N0.getOperand(0); // If the original shl may be shifting out bits, do not perform this // transformation. unsigned KnownZeroBits = InnerZExt.getValueSizeInBits() - InnerZExt.getOperand(0).getValueSizeInBits(); if (cast(ShAmt)->getAPIntValue().ugt(KnownZeroBits)) return SDValue(); } SDLoc DL(N); // Ensure that the shift amount is wide enough for the shifted value. if (Log2_32_Ceil(VT.getSizeInBits()) > ShAmt.getValueSizeInBits()) ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt); return DAG.getNode(N0.getOpcode(), DL, VT, DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)), ShAmt); } if (SDValue NewVSel = matchVSelectOpSizesWithSetCC(N)) return NewVSel; if (SDValue NewCtPop = widenCtPop(N, DAG)) return NewCtPop; if (SDValue Res = tryToFoldExtendSelectLoad(N, TLI, DAG)) return Res; return SDValue(); } SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); // aext(undef) = undef if (N0.isUndef()) return DAG.getUNDEF(VT); if (SDValue Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes)) return Res; // fold (aext (aext x)) -> (aext x) // fold (aext (zext x)) -> (zext x) // fold (aext (sext x)) -> (sext x) if (N0.getOpcode() == ISD::ANY_EXTEND || N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::SIGN_EXTEND) return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0)); // fold (aext (truncate (load x))) -> (aext (smaller load x)) // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n))) if (N0.getOpcode() == ISD::TRUNCATE) { if (SDValue NarrowLoad = reduceLoadWidth(N0.getNode())) { SDNode *oye = N0.getOperand(0).getNode(); if (NarrowLoad.getNode() != N0.getNode()) { CombineTo(N0.getNode(), NarrowLoad); // CombineTo deleted the truncate, if needed, but not what's under it. AddToWorklist(oye); } return SDValue(N, 0); // Return N so it doesn't get rechecked! } } // fold (aext (truncate x)) if (N0.getOpcode() == ISD::TRUNCATE) return DAG.getAnyExtOrTrunc(N0.getOperand(0), SDLoc(N), VT); // Fold (aext (and (trunc x), cst)) -> (and x, cst) // if the trunc is not free. if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::TRUNCATE && N0.getOperand(1).getOpcode() == ISD::Constant && !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(), N0.getValueType())) { SDLoc DL(N); SDValue X = DAG.getAnyExtOrTrunc(N0.getOperand(0).getOperand(0), DL, VT); SDValue Y = DAG.getNode(ISD::ANY_EXTEND, DL, VT, N0.getOperand(1)); assert(isa(Y) && "Expected constant to be folded!"); return DAG.getNode(ISD::AND, DL, VT, X, Y); } // fold (aext (load x)) -> (aext (truncate (extload x))) // None of the supported targets knows how to perform load and any_ext // on vectors in one instruction, so attempt to fold to zext instead. if (VT.isVector()) { // Try to simplify (zext (load x)). if (SDValue foldedExt = tryToFoldExtOfLoad(DAG, *this, TLI, VT, LegalOperations, N, N0, ISD::ZEXTLOAD, ISD::ZERO_EXTEND)) return foldedExt; } else if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) && TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) { bool DoXform = true; SmallVector SetCCs; if (!N0.hasOneUse()) DoXform = ExtendUsesToFormExtLoad(VT, N, N0, ISD::ANY_EXTEND, SetCCs, TLI); if (DoXform) { LoadSDNode *LN0 = cast(N0); SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT, LN0->getChain(), LN0->getBasePtr(), N0.getValueType(), LN0->getMemOperand()); ExtendSetCCUses(SetCCs, N0, ExtLoad, ISD::ANY_EXTEND); // If the load value is used only by N, replace it via CombineTo N. bool NoReplaceTrunc = N0.hasOneUse(); CombineTo(N, ExtLoad); if (NoReplaceTrunc) { DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), ExtLoad.getValue(1)); recursivelyDeleteUnusedNodes(LN0); } else { SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(), ExtLoad); CombineTo(LN0, Trunc, ExtLoad.getValue(1)); } return SDValue(N, 0); // Return N so it doesn't get rechecked! } } // fold (aext (zextload x)) -> (aext (truncate (zextload x))) // fold (aext (sextload x)) -> (aext (truncate (sextload x))) // fold (aext ( extload x)) -> (aext (truncate (extload x))) if (N0.getOpcode() == ISD::LOAD && !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) { LoadSDNode *LN0 = cast(N0); ISD::LoadExtType ExtType = LN0->getExtensionType(); EVT MemVT = LN0->getMemoryVT(); if (!LegalOperations || TLI.isLoadExtLegal(ExtType, VT, MemVT)) { SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N), VT, LN0->getChain(), LN0->getBasePtr(), MemVT, LN0->getMemOperand()); CombineTo(N, ExtLoad); DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), ExtLoad.getValue(1)); recursivelyDeleteUnusedNodes(LN0); return SDValue(N, 0); // Return N so it doesn't get rechecked! } } if (N0.getOpcode() == ISD::SETCC) { // Propagate fast-math-flags. SelectionDAG::FlagInserter FlagsInserter(DAG, N0->getFlags()); // For vectors: // aext(setcc) -> vsetcc // aext(setcc) -> truncate(vsetcc) // aext(setcc) -> aext(vsetcc) // Only do this before legalize for now. if (VT.isVector() && !LegalOperations) { EVT N00VT = N0.getOperand(0).getValueType(); if (getSetCCResultType(N00VT) == N0.getValueType()) return SDValue(); // We know that the # elements of the results is the same as the // # elements of the compare (and the # elements of the compare result // for that matter). Check to see that they are the same size. If so, // we know that the element size of the sext'd result matches the // element size of the compare operands. if (VT.getSizeInBits() == N00VT.getSizeInBits()) return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0), N0.getOperand(1), cast(N0.getOperand(2))->get()); // If the desired elements are smaller or larger than the source // elements we can use a matching integer vector type and then // truncate/any extend EVT MatchingVectorType = N00VT.changeVectorElementTypeToInteger(); SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0), N0.getOperand(1), cast(N0.getOperand(2))->get()); return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT); } // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc SDLoc DL(N); if (SDValue SCC = SimplifySelectCC( DL, N0.getOperand(0), N0.getOperand(1), DAG.getConstant(1, DL, VT), DAG.getConstant(0, DL, VT), cast(N0.getOperand(2))->get(), true)) return SCC; } if (SDValue NewCtPop = widenCtPop(N, DAG)) return NewCtPop; if (SDValue Res = tryToFoldExtendSelectLoad(N, TLI, DAG)) return Res; return SDValue(); } SDValue DAGCombiner::visitAssertExt(SDNode *N) { unsigned Opcode = N->getOpcode(); SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT AssertVT = cast(N1)->getVT(); // fold (assert?ext (assert?ext x, vt), vt) -> (assert?ext x, vt) if (N0.getOpcode() == Opcode && AssertVT == cast(N0.getOperand(1))->getVT()) return N0; if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() && N0.getOperand(0).getOpcode() == Opcode) { // We have an assert, truncate, assert sandwich. Make one stronger assert // by asserting on the smallest asserted type to the larger source type. // This eliminates the later assert: // assert (trunc (assert X, i8) to iN), i1 --> trunc (assert X, i1) to iN // assert (trunc (assert X, i1) to iN), i8 --> trunc (assert X, i1) to iN SDLoc DL(N); SDValue BigA = N0.getOperand(0); EVT BigA_AssertVT = cast(BigA.getOperand(1))->getVT(); EVT MinAssertVT = AssertVT.bitsLT(BigA_AssertVT) ? AssertVT : BigA_AssertVT; SDValue MinAssertVTVal = DAG.getValueType(MinAssertVT); SDValue NewAssert = DAG.getNode(Opcode, DL, BigA.getValueType(), BigA.getOperand(0), MinAssertVTVal); return DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), NewAssert); } // If we have (AssertZext (truncate (AssertSext X, iX)), iY) and Y is smaller // than X. Just move the AssertZext in front of the truncate and drop the // AssertSExt. if (N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() && N0.getOperand(0).getOpcode() == ISD::AssertSext && Opcode == ISD::AssertZext) { SDValue BigA = N0.getOperand(0); EVT BigA_AssertVT = cast(BigA.getOperand(1))->getVT(); if (AssertVT.bitsLT(BigA_AssertVT)) { SDLoc DL(N); SDValue NewAssert = DAG.getNode(Opcode, DL, BigA.getValueType(), BigA.getOperand(0), N1); return DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), NewAssert); } } return SDValue(); } SDValue DAGCombiner::visitAssertAlign(SDNode *N) { SDLoc DL(N); Align AL = cast(N)->getAlign(); SDValue N0 = N->getOperand(0); // Fold (assertalign (assertalign x, AL0), AL1) -> // (assertalign x, max(AL0, AL1)) if (auto *AAN = dyn_cast(N0)) return DAG.getAssertAlign(DL, N0.getOperand(0), std::max(AL, AAN->getAlign())); // In rare cases, there are trivial arithmetic ops in source operands. Sink // this assert down to source operands so that those arithmetic ops could be // exposed to the DAG combining. switch (N0.getOpcode()) { default: break; case ISD::ADD: case ISD::SUB: { unsigned AlignShift = Log2(AL); SDValue LHS = N0.getOperand(0); SDValue RHS = N0.getOperand(1); unsigned LHSAlignShift = DAG.computeKnownBits(LHS).countMinTrailingZeros(); unsigned RHSAlignShift = DAG.computeKnownBits(RHS).countMinTrailingZeros(); if (LHSAlignShift >= AlignShift || RHSAlignShift >= AlignShift) { if (LHSAlignShift < AlignShift) LHS = DAG.getAssertAlign(DL, LHS, AL); if (RHSAlignShift < AlignShift) RHS = DAG.getAssertAlign(DL, RHS, AL); return DAG.getNode(N0.getOpcode(), DL, N0.getValueType(), LHS, RHS); } break; } } return SDValue(); } /// If the result of a load is shifted/masked/truncated to an effectively /// narrower type, try to transform the load to a narrower type and/or /// use an extending load. SDValue DAGCombiner::reduceLoadWidth(SDNode *N) { unsigned Opc = N->getOpcode(); ISD::LoadExtType ExtType = ISD::NON_EXTLOAD; SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); EVT ExtVT = VT; // This transformation isn't valid for vector loads. if (VT.isVector()) return SDValue(); // The ShAmt variable is used to indicate that we've consumed a right // shift. I.e. we want to narrow the width of the load by skipping to load the // ShAmt least significant bits. unsigned ShAmt = 0; // A special case is when the least significant bits from the load are masked // away, but using an AND rather than a right shift. HasShiftedOffset is used // to indicate that the narrowed load should be left-shifted ShAmt bits to get // the result. bool HasShiftedOffset = false; // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then // extended to VT. if (Opc == ISD::SIGN_EXTEND_INREG) { ExtType = ISD::SEXTLOAD; ExtVT = cast(N->getOperand(1))->getVT(); } else if (Opc == ISD::SRL || Opc == ISD::SRA) { // Another special-case: SRL/SRA is basically zero/sign-extending a narrower // value, or it may be shifting a higher subword, half or byte into the // lowest bits. // Only handle shift with constant shift amount, and the shiftee must be a // load. auto *LN = dyn_cast(N0); auto *N1C = dyn_cast(N->getOperand(1)); if (!N1C || !LN) return SDValue(); // If the shift amount is larger than the memory type then we're not // accessing any of the loaded bytes. ShAmt = N1C->getZExtValue(); uint64_t MemoryWidth = LN->getMemoryVT().getScalarSizeInBits(); if (MemoryWidth <= ShAmt) return SDValue(); // Attempt to fold away the SRL by using ZEXTLOAD and SRA by using SEXTLOAD. ExtType = Opc == ISD::SRL ? ISD::ZEXTLOAD : ISD::SEXTLOAD; ExtVT = EVT::getIntegerVT(*DAG.getContext(), MemoryWidth - ShAmt); // If original load is a SEXTLOAD then we can't simply replace it by a // ZEXTLOAD (we could potentially replace it by a more narrow SEXTLOAD // followed by a ZEXT, but that is not handled at the moment). Similarly if // the original load is a ZEXTLOAD and we want to use a SEXTLOAD. if ((LN->getExtensionType() == ISD::SEXTLOAD || LN->getExtensionType() == ISD::ZEXTLOAD) && LN->getExtensionType() != ExtType) return SDValue(); } else if (Opc == ISD::AND) { // An AND with a constant mask is the same as a truncate + zero-extend. auto AndC = dyn_cast(N->getOperand(1)); if (!AndC) return SDValue(); const APInt &Mask = AndC->getAPIntValue(); unsigned ActiveBits = 0; if (Mask.isMask()) { ActiveBits = Mask.countTrailingOnes(); } else if (Mask.isShiftedMask(ShAmt, ActiveBits)) { HasShiftedOffset = true; } else { return SDValue(); } ExtType = ISD::ZEXTLOAD; ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits); } // In case Opc==SRL we've already prepared ExtVT/ExtType/ShAmt based on doing // a right shift. Here we redo some of those checks, to possibly adjust the // ExtVT even further based on "a masking AND". We could also end up here for // other reasons (e.g. based on Opc==TRUNCATE) and that is why some checks // need to be done here as well. if (Opc == ISD::SRL || N0.getOpcode() == ISD::SRL) { SDValue SRL = Opc == ISD::SRL ? SDValue(N, 0) : N0; // Bail out when the SRL has more than one use. This is done for historical // (undocumented) reasons. Maybe intent was to guard the AND-masking below // check below? And maybe it could be non-profitable to do the transform in // case the SRL has multiple uses and we get here with Opc!=ISD::SRL? // FIXME: Can't we just skip this check for the Opc==ISD::SRL case. if (!SRL.hasOneUse()) return SDValue(); // Only handle shift with constant shift amount, and the shiftee must be a // load. auto *LN = dyn_cast(SRL.getOperand(0)); auto *SRL1C = dyn_cast(SRL.getOperand(1)); if (!SRL1C || !LN) return SDValue(); // If the shift amount is larger than the input type then we're not // accessing any of the loaded bytes. If the load was a zextload/extload // then the result of the shift+trunc is zero/undef (handled elsewhere). ShAmt = SRL1C->getZExtValue(); uint64_t MemoryWidth = LN->getMemoryVT().getSizeInBits(); if (ShAmt >= MemoryWidth) return SDValue(); // Because a SRL must be assumed to *need* to zero-extend the high bits // (as opposed to anyext the high bits), we can't combine the zextload // lowering of SRL and an sextload. if (LN->getExtensionType() == ISD::SEXTLOAD) return SDValue(); // Avoid reading outside the memory accessed by the original load (could // happened if we only adjust the load base pointer by ShAmt). Instead we // try to narrow the load even further. The typical scenario here is: // (i64 (truncate (i96 (srl (load x), 64)))) -> // (i64 (truncate (i96 (zextload (load i32 + offset) from i32)))) if (ExtVT.getScalarSizeInBits() > MemoryWidth - ShAmt) { // Don't replace sextload by zextload. if (ExtType == ISD::SEXTLOAD) return SDValue(); // Narrow the load. ExtType = ISD::ZEXTLOAD; ExtVT = EVT::getIntegerVT(*DAG.getContext(), MemoryWidth - ShAmt); } // If the SRL is only used by a masking AND, we may be able to adjust // the ExtVT to make the AND redundant. SDNode *Mask = *(SRL->use_begin()); if (SRL.hasOneUse() && Mask->getOpcode() == ISD::AND && isa(Mask->getOperand(1))) { const APInt& ShiftMask = Mask->getConstantOperandAPInt(1); if (ShiftMask.isMask()) { EVT MaskedVT = EVT::getIntegerVT(*DAG.getContext(), ShiftMask.countTrailingOnes()); // If the mask is smaller, recompute the type. if ((ExtVT.getScalarSizeInBits() > MaskedVT.getScalarSizeInBits()) && TLI.isLoadExtLegal(ExtType, SRL.getValueType(), MaskedVT)) ExtVT = MaskedVT; } } N0 = SRL.getOperand(0); } // If the load is shifted left (and the result isn't shifted back right), we // can fold a truncate through the shift. The typical scenario is that N // points at a TRUNCATE here so the attempted fold is: // (truncate (shl (load x), c))) -> (shl (narrow load x), c) // ShLeftAmt will indicate how much a narrowed load should be shifted left. unsigned ShLeftAmt = 0; if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() && ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) { if (ConstantSDNode *N01 = dyn_cast(N0.getOperand(1))) { ShLeftAmt = N01->getZExtValue(); N0 = N0.getOperand(0); } } // If we haven't found a load, we can't narrow it. if (!isa(N0)) return SDValue(); LoadSDNode *LN0 = cast(N0); // Reducing the width of a volatile load is illegal. For atomics, we may be // able to reduce the width provided we never widen again. (see D66309) if (!LN0->isSimple() || !isLegalNarrowLdSt(LN0, ExtType, ExtVT, ShAmt)) return SDValue(); auto AdjustBigEndianShift = [&](unsigned ShAmt) { unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits().getFixedSize(); unsigned EVTStoreBits = ExtVT.getStoreSizeInBits().getFixedSize(); return LVTStoreBits - EVTStoreBits - ShAmt; }; // We need to adjust the pointer to the load by ShAmt bits in order to load // the correct bytes. unsigned PtrAdjustmentInBits = DAG.getDataLayout().isBigEndian() ? AdjustBigEndianShift(ShAmt) : ShAmt; uint64_t PtrOff = PtrAdjustmentInBits / 8; Align NewAlign = commonAlignment(LN0->getAlign(), PtrOff); SDLoc DL(LN0); // The original load itself didn't wrap, so an offset within it doesn't. SDNodeFlags Flags; Flags.setNoUnsignedWrap(true); SDValue NewPtr = DAG.getMemBasePlusOffset(LN0->getBasePtr(), TypeSize::Fixed(PtrOff), DL, Flags); AddToWorklist(NewPtr.getNode()); SDValue Load; if (ExtType == ISD::NON_EXTLOAD) Load = DAG.getLoad(VT, DL, LN0->getChain(), NewPtr, LN0->getPointerInfo().getWithOffset(PtrOff), NewAlign, LN0->getMemOperand()->getFlags(), LN0->getAAInfo()); else Load = DAG.getExtLoad(ExtType, DL, VT, LN0->getChain(), NewPtr, LN0->getPointerInfo().getWithOffset(PtrOff), ExtVT, NewAlign, LN0->getMemOperand()->getFlags(), LN0->getAAInfo()); // Replace the old load's chain with the new load's chain. WorklistRemover DeadNodes(*this); DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1)); // Shift the result left, if we've swallowed a left shift. SDValue Result = Load; if (ShLeftAmt != 0) { EVT ShImmTy = getShiftAmountTy(Result.getValueType()); if (!isUIntN(ShImmTy.getScalarSizeInBits(), ShLeftAmt)) ShImmTy = VT; // If the shift amount is as large as the result size (but, presumably, // no larger than the source) then the useful bits of the result are // zero; we can't simply return the shortened shift, because the result // of that operation is undefined. if (ShLeftAmt >= VT.getScalarSizeInBits()) Result = DAG.getConstant(0, DL, VT); else Result = DAG.getNode(ISD::SHL, DL, VT, Result, DAG.getConstant(ShLeftAmt, DL, ShImmTy)); } if (HasShiftedOffset) { // We're using a shifted mask, so the load now has an offset. This means // that data has been loaded into the lower bytes than it would have been // before, so we need to shl the loaded data into the correct position in the // register. SDValue ShiftC = DAG.getConstant(ShAmt, DL, VT); Result = DAG.getNode(ISD::SHL, DL, VT, Result, ShiftC); DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result); } // Return the new loaded value. return Result; } SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); EVT ExtVT = cast(N1)->getVT(); unsigned VTBits = VT.getScalarSizeInBits(); unsigned ExtVTBits = ExtVT.getScalarSizeInBits(); // sext_vector_inreg(undef) = 0 because the top bit will all be the same. if (N0.isUndef()) return DAG.getConstant(0, SDLoc(N), VT); // fold (sext_in_reg c1) -> c1 if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1); // If the input is already sign extended, just drop the extension. if (ExtVTBits >= DAG.ComputeMaxSignificantBits(N0)) return N0; // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && ExtVT.bitsLT(cast(N0.getOperand(1))->getVT())) return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0.getOperand(0), N1); // fold (sext_in_reg (sext x)) -> (sext x) // fold (sext_in_reg (aext x)) -> (sext x) // if x is small enough or if we know that x has more than 1 sign bit and the // sign_extend_inreg is extending from one of them. if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) { SDValue N00 = N0.getOperand(0); unsigned N00Bits = N00.getScalarValueSizeInBits(); if ((N00Bits <= ExtVTBits || DAG.ComputeMaxSignificantBits(N00) <= ExtVTBits) && (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT))) return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00); } // fold (sext_in_reg (*_extend_vector_inreg x)) -> (sext_vector_inreg x) // if x is small enough or if we know that x has more than 1 sign bit and the // sign_extend_inreg is extending from one of them. if (N0.getOpcode() == ISD::ANY_EXTEND_VECTOR_INREG || N0.getOpcode() == ISD::SIGN_EXTEND_VECTOR_INREG || N0.getOpcode() == ISD::ZERO_EXTEND_VECTOR_INREG) { SDValue N00 = N0.getOperand(0); unsigned N00Bits = N00.getScalarValueSizeInBits(); unsigned DstElts = N0.getValueType().getVectorMinNumElements(); unsigned SrcElts = N00.getValueType().getVectorMinNumElements(); bool IsZext = N0.getOpcode() == ISD::ZERO_EXTEND_VECTOR_INREG; APInt DemandedSrcElts = APInt::getLowBitsSet(SrcElts, DstElts); if ((N00Bits == ExtVTBits || (!IsZext && (N00Bits < ExtVTBits || DAG.ComputeMaxSignificantBits(N00) <= ExtVTBits))) && (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_VECTOR_INREG, VT))) return DAG.getNode(ISD::SIGN_EXTEND_VECTOR_INREG, SDLoc(N), VT, N00); } // fold (sext_in_reg (zext x)) -> (sext x) // iff we are extending the source sign bit. if (N0.getOpcode() == ISD::ZERO_EXTEND) { SDValue N00 = N0.getOperand(0); if (N00.getScalarValueSizeInBits() == ExtVTBits && (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT))) return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1); } // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero. if (DAG.MaskedValueIsZero(N0, APInt::getOneBitSet(VTBits, ExtVTBits - 1))) return DAG.getZeroExtendInReg(N0, SDLoc(N), ExtVT); // fold operands of sext_in_reg based on knowledge that the top bits are not // demanded. if (SimplifyDemandedBits(SDValue(N, 0))) return SDValue(N, 0); // fold (sext_in_reg (load x)) -> (smaller sextload x) // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits)) if (SDValue NarrowLoad = reduceLoadWidth(N)) return NarrowLoad; // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24) // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible. // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above. if (N0.getOpcode() == ISD::SRL) { if (auto *ShAmt = dyn_cast(N0.getOperand(1))) if (ShAmt->getAPIntValue().ule(VTBits - ExtVTBits)) { // We can turn this into an SRA iff the input to the SRL is already sign // extended enough. unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0)); if (((VTBits - ExtVTBits) - ShAmt->getZExtValue()) < InSignBits) return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0), N0.getOperand(1)); } } // fold (sext_inreg (extload x)) -> (sextload x) // If sextload is not supported by target, we can only do the combine when // load has one use. Doing otherwise can block folding the extload with other // extends that the target does support. if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) && ExtVT == cast(N0)->getMemoryVT() && ((!LegalOperations && cast(N0)->isSimple() && N0.hasOneUse()) || TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, ExtVT))) { LoadSDNode *LN0 = cast(N0); SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT, LN0->getChain(), LN0->getBasePtr(), ExtVT, LN0->getMemOperand()); CombineTo(N, ExtLoad); CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); AddToWorklist(ExtLoad.getNode()); return SDValue(N, 0); // Return N so it doesn't get rechecked! } // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse() && ExtVT == cast(N0)->getMemoryVT() && ((!LegalOperations && cast(N0)->isSimple()) && TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, ExtVT))) { LoadSDNode *LN0 = cast(N0); SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT, LN0->getChain(), LN0->getBasePtr(), ExtVT, LN0->getMemOperand()); CombineTo(N, ExtLoad); CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); return SDValue(N, 0); // Return N so it doesn't get rechecked! } // fold (sext_inreg (masked_load x)) -> (sext_masked_load x) // ignore it if the masked load is already sign extended if (MaskedLoadSDNode *Ld = dyn_cast(N0)) { if (ExtVT == Ld->getMemoryVT() && N0.hasOneUse() && Ld->getExtensionType() != ISD::LoadExtType::NON_EXTLOAD && TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, ExtVT)) { SDValue ExtMaskedLoad = DAG.getMaskedLoad( VT, SDLoc(N), Ld->getChain(), Ld->getBasePtr(), Ld->getOffset(), Ld->getMask(), Ld->getPassThru(), ExtVT, Ld->getMemOperand(), Ld->getAddressingMode(), ISD::SEXTLOAD, Ld->isExpandingLoad()); CombineTo(N, ExtMaskedLoad); CombineTo(N0.getNode(), ExtMaskedLoad, ExtMaskedLoad.getValue(1)); return SDValue(N, 0); // Return N so it doesn't get rechecked! } } // fold (sext_inreg (masked_gather x)) -> (sext_masked_gather x) if (auto *GN0 = dyn_cast(N0)) { if (SDValue(GN0, 0).hasOneUse() && ExtVT == GN0->getMemoryVT() && TLI.isVectorLoadExtDesirable(SDValue(SDValue(GN0, 0)))) { SDValue Ops[] = {GN0->getChain(), GN0->getPassThru(), GN0->getMask(), GN0->getBasePtr(), GN0->getIndex(), GN0->getScale()}; SDValue ExtLoad = DAG.getMaskedGather( DAG.getVTList(VT, MVT::Other), ExtVT, SDLoc(N), Ops, GN0->getMemOperand(), GN0->getIndexType(), ISD::SEXTLOAD); CombineTo(N, ExtLoad); CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); AddToWorklist(ExtLoad.getNode()); return SDValue(N, 0); // Return N so it doesn't get rechecked! } } // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16)) if (ExtVTBits <= 16 && N0.getOpcode() == ISD::OR) { if (SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0), N0.getOperand(1), false)) return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, BSwap, N1); } return SDValue(); } SDValue DAGCombiner::visitEXTEND_VECTOR_INREG(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); // {s/z}ext_vector_inreg(undef) = 0 because the top bits must be the same. if (N0.isUndef()) return DAG.getConstant(0, SDLoc(N), VT); if (SDValue Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes)) return Res; if (SimplifyDemandedVectorElts(SDValue(N, 0))) return SDValue(N, 0); return SDValue(); } SDValue DAGCombiner::visitTRUNCATE(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); EVT SrcVT = N0.getValueType(); bool isLE = DAG.getDataLayout().isLittleEndian(); // noop truncate if (SrcVT == VT) return N0; // fold (truncate (truncate x)) -> (truncate x) if (N0.getOpcode() == ISD::TRUNCATE) return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0)); // fold (truncate c1) -> c1 if (DAG.isConstantIntBuildVectorOrConstantInt(N0)) { SDValue C = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0); if (C.getNode() != N) return C; } // fold (truncate (ext x)) -> (ext x) or (truncate x) or x if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) { // if the source is smaller than the dest, we still need an extend. if (N0.getOperand(0).getValueType().bitsLT(VT)) return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0)); // if the source is larger than the dest, than we just need the truncate. if (N0.getOperand(0).getValueType().bitsGT(VT)) return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0)); // if the source and dest are the same type, we can drop both the extend // and the truncate. return N0.getOperand(0); } // Try to narrow a truncate-of-sext_in_reg to the destination type: // trunc (sign_ext_inreg X, iM) to iN --> sign_ext_inreg (trunc X to iN), iM if (!LegalTypes && N0.getOpcode() == ISD::SIGN_EXTEND_INREG && N0.hasOneUse()) { SDValue X = N0.getOperand(0); SDValue ExtVal = N0.getOperand(1); EVT ExtVT = cast(ExtVal)->getVT(); if (ExtVT.bitsLT(VT)) { SDValue TrX = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X); return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, TrX, ExtVal); } } // If this is anyext(trunc), don't fold it, allow ourselves to be folded. if (N->hasOneUse() && (N->use_begin()->getOpcode() == ISD::ANY_EXTEND)) return SDValue(); // Fold extract-and-trunc into a narrow extract. For example: // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1) // i32 y = TRUNCATE(i64 x) // -- becomes -- // v16i8 b = BITCAST (v2i64 val) // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8) // // Note: We only run this optimization after type legalization (which often // creates this pattern) and before operation legalization after which // we need to be more careful about the vector instructions that we generate. if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT && LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) { EVT VecTy = N0.getOperand(0).getValueType(); EVT ExTy = N0.getValueType(); EVT TrTy = N->getValueType(0); auto EltCnt = VecTy.getVectorElementCount(); unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits(); auto NewEltCnt = EltCnt * SizeRatio; EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, NewEltCnt); assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size"); SDValue EltNo = N0->getOperand(1); if (isa(EltNo) && isTypeLegal(NVT)) { int Elt = cast(EltNo)->getZExtValue(); int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1)); SDLoc DL(N); return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, TrTy, DAG.getBitcast(NVT, N0.getOperand(0)), DAG.getVectorIdxConstant(Index, DL)); } } // trunc (select c, a, b) -> select c, (trunc a), (trunc b) if (N0.getOpcode() == ISD::SELECT && N0.hasOneUse()) { if ((!LegalOperations || TLI.isOperationLegal(ISD::SELECT, SrcVT)) && TLI.isTruncateFree(SrcVT, VT)) { SDLoc SL(N0); SDValue Cond = N0.getOperand(0); SDValue TruncOp0 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1)); SDValue TruncOp1 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(2)); return DAG.getNode(ISD::SELECT, SDLoc(N), VT, Cond, TruncOp0, TruncOp1); } } // trunc (shl x, K) -> shl (trunc x), K => K < VT.getScalarSizeInBits() if (N0.getOpcode() == ISD::SHL && N0.hasOneUse() && (!LegalOperations || TLI.isOperationLegal(ISD::SHL, VT)) && TLI.isTypeDesirableForOp(ISD::SHL, VT)) { SDValue Amt = N0.getOperand(1); KnownBits Known = DAG.computeKnownBits(Amt); unsigned Size = VT.getScalarSizeInBits(); if (Known.countMaxActiveBits() <= Log2_32(Size)) { SDLoc SL(N); EVT AmtVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout()); SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(0)); if (AmtVT != Amt.getValueType()) { Amt = DAG.getZExtOrTrunc(Amt, SL, AmtVT); AddToWorklist(Amt.getNode()); } return DAG.getNode(ISD::SHL, SL, VT, Trunc, Amt); } } if (SDValue V = foldSubToUSubSat(VT, N0.getNode())) return V; // Attempt to pre-truncate BUILD_VECTOR sources. if (N0.getOpcode() == ISD::BUILD_VECTOR && !LegalOperations && TLI.isTruncateFree(SrcVT.getScalarType(), VT.getScalarType()) && // Avoid creating illegal types if running after type legalizer. (!LegalTypes || TLI.isTypeLegal(VT.getScalarType()))) { SDLoc DL(N); EVT SVT = VT.getScalarType(); SmallVector TruncOps; for (const SDValue &Op : N0->op_values()) { SDValue TruncOp = DAG.getNode(ISD::TRUNCATE, DL, SVT, Op); TruncOps.push_back(TruncOp); } return DAG.getBuildVector(VT, DL, TruncOps); } // Fold a series of buildvector, bitcast, and truncate if possible. // For example fold // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to // (2xi32 (buildvector x, y)). if (Level == AfterLegalizeVectorOps && VT.isVector() && N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() && N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR && N0.getOperand(0).hasOneUse()) { SDValue BuildVect = N0.getOperand(0); EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType(); EVT TruncVecEltTy = VT.getVectorElementType(); // Check that the element types match. if (BuildVectEltTy == TruncVecEltTy) { // Now we only need to compute the offset of the truncated elements. unsigned BuildVecNumElts = BuildVect.getNumOperands(); unsigned TruncVecNumElts = VT.getVectorNumElements(); unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts; assert((BuildVecNumElts % TruncVecNumElts) == 0 && "Invalid number of elements"); SmallVector Opnds; for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset) Opnds.push_back(BuildVect.getOperand(i)); return DAG.getBuildVector(VT, SDLoc(N), Opnds); } } // fold (truncate (load x)) -> (smaller load x) // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits)) if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) { if (SDValue Reduced = reduceLoadWidth(N)) return Reduced; // Handle the case where the load remains an extending load even // after truncation. if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) { LoadSDNode *LN0 = cast(N0); if (LN0->isSimple() && LN0->getMemoryVT().bitsLT(VT)) { SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0), VT, LN0->getChain(), LN0->getBasePtr(), LN0->getMemoryVT(), LN0->getMemOperand()); DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1)); return NewLoad; } } } // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)), // where ... are all 'undef'. if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) { SmallVector VTs; SDValue V; unsigned Idx = 0; unsigned NumDefs = 0; for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) { SDValue X = N0.getOperand(i); if (!X.isUndef()) { V = X; Idx = i; NumDefs++; } // Stop if more than one members are non-undef. if (NumDefs > 1) break; VTs.push_back(EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(), X.getValueType().getVectorElementCount())); } if (NumDefs == 0) return DAG.getUNDEF(VT); if (NumDefs == 1) { assert(V.getNode() && "The single defined operand is empty!"); SmallVector Opnds; for (unsigned i = 0, e = VTs.size(); i != e; ++i) { if (i != Idx) { Opnds.push_back(DAG.getUNDEF(VTs[i])); continue; } SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V); AddToWorklist(NV.getNode()); Opnds.push_back(NV); } return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds); } } // Fold truncate of a bitcast of a vector to an extract of the low vector // element. // // e.g. trunc (i64 (bitcast v2i32:x)) -> extract_vector_elt v2i32:x, idx if (N0.getOpcode() == ISD::BITCAST && !VT.isVector()) { SDValue VecSrc = N0.getOperand(0); EVT VecSrcVT = VecSrc.getValueType(); if (VecSrcVT.isVector() && VecSrcVT.getScalarType() == VT && (!LegalOperations || TLI.isOperationLegal(ISD::EXTRACT_VECTOR_ELT, VecSrcVT))) { SDLoc SL(N); unsigned Idx = isLE ? 0 : VecSrcVT.getVectorNumElements() - 1; return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, VT, VecSrc, DAG.getVectorIdxConstant(Idx, SL)); } } // Simplify the operands using demanded-bits information. if (SimplifyDemandedBits(SDValue(N, 0))) return SDValue(N, 0); // See if we can simplify the input to this truncate through knowledge that // only the low bits are being used. // For example "trunc (or (shl x, 8), y)" // -> trunc y // Currently we only perform this optimization on scalars because vectors // may have different active low bits. if (!VT.isVector()) { APInt Mask = APInt::getLowBitsSet(N0.getValueSizeInBits(), VT.getSizeInBits()); if (SDValue Shorter = DAG.GetDemandedBits(N0, Mask)) return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter); } // fold (truncate (extract_subvector(ext x))) -> // (extract_subvector x) // TODO: This can be generalized to cover cases where the truncate and extract // do not fully cancel each other out. if (!LegalTypes && N0.getOpcode() == ISD::EXTRACT_SUBVECTOR) { SDValue N00 = N0.getOperand(0); if (N00.getOpcode() == ISD::SIGN_EXTEND || N00.getOpcode() == ISD::ZERO_EXTEND || N00.getOpcode() == ISD::ANY_EXTEND) { if (N00.getOperand(0)->getValueType(0).getVectorElementType() == VT.getVectorElementType()) return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N0->getOperand(0)), VT, N00.getOperand(0), N0.getOperand(1)); } } if (SDValue NewVSel = matchVSelectOpSizesWithSetCC(N)) return NewVSel; // Narrow a suitable binary operation with a non-opaque constant operand by // moving it ahead of the truncate. This is limited to pre-legalization // because targets may prefer a wider type during later combines and invert // this transform. switch (N0.getOpcode()) { case ISD::ADD: case ISD::SUB: case ISD::MUL: case ISD::AND: case ISD::OR: case ISD::XOR: if (!LegalOperations && N0.hasOneUse() && (isConstantOrConstantVector(N0.getOperand(0), true) || isConstantOrConstantVector(N0.getOperand(1), true))) { // TODO: We already restricted this to pre-legalization, but for vectors // we are extra cautious to not create an unsupported operation. // Target-specific changes are likely needed to avoid regressions here. if (VT.isScalarInteger() || TLI.isOperationLegal(N0.getOpcode(), VT)) { SDLoc DL(N); SDValue NarrowL = DAG.getNode(ISD::TRUNCATE, DL, VT, N0.getOperand(0)); SDValue NarrowR = DAG.getNode(ISD::TRUNCATE, DL, VT, N0.getOperand(1)); return DAG.getNode(N0.getOpcode(), DL, VT, NarrowL, NarrowR); } } break; case ISD::ADDE: case ISD::ADDCARRY: // (trunc adde(X, Y, Carry)) -> (adde trunc(X), trunc(Y), Carry) // (trunc addcarry(X, Y, Carry)) -> (addcarry trunc(X), trunc(Y), Carry) // When the adde's carry is not used. // We only do for addcarry before legalize operation if (((!LegalOperations && N0.getOpcode() == ISD::ADDCARRY) || TLI.isOperationLegal(N0.getOpcode(), VT)) && N0.hasOneUse() && !N0->hasAnyUseOfValue(1)) { SDLoc DL(N); SDValue X = DAG.getNode(ISD::TRUNCATE, DL, VT, N0.getOperand(0)); SDValue Y = DAG.getNode(ISD::TRUNCATE, DL, VT, N0.getOperand(1)); SDVTList VTs = DAG.getVTList(VT, N0->getValueType(1)); return DAG.getNode(N0.getOpcode(), DL, VTs, X, Y, N0.getOperand(2)); } break; case ISD::USUBSAT: // Truncate the USUBSAT only if LHS is a known zero-extension, its not // enough to know that the upper bits are zero we must ensure that we don't // introduce an extra truncate. if (!LegalOperations && N0.hasOneUse() && N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND && N0.getOperand(0).getOperand(0).getScalarValueSizeInBits() <= VT.getScalarSizeInBits() && hasOperation(N0.getOpcode(), VT)) { return getTruncatedUSUBSAT(VT, SrcVT, N0.getOperand(0), N0.getOperand(1), DAG, SDLoc(N)); } break; } return SDValue(); } static SDNode *getBuildPairElt(SDNode *N, unsigned i) { SDValue Elt = N->getOperand(i); if (Elt.getOpcode() != ISD::MERGE_VALUES) return Elt.getNode(); return Elt.getOperand(Elt.getResNo()).getNode(); } /// build_pair (load, load) -> load /// if load locations are consecutive. SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) { assert(N->getOpcode() == ISD::BUILD_PAIR); auto *LD1 = dyn_cast(getBuildPairElt(N, 0)); auto *LD2 = dyn_cast(getBuildPairElt(N, 1)); // A BUILD_PAIR is always having the least significant part in elt 0 and the // most significant part in elt 1. So when combining into one large load, we // need to consider the endianness. if (DAG.getDataLayout().isBigEndian()) std::swap(LD1, LD2); if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !ISD::isNON_EXTLoad(LD2) || !LD1->hasOneUse() || !LD2->hasOneUse() || LD1->getAddressSpace() != LD2->getAddressSpace()) return SDValue(); bool LD1Fast = false; EVT LD1VT = LD1->getValueType(0); unsigned LD1Bytes = LD1VT.getStoreSize(); if ((!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) && DAG.areNonVolatileConsecutiveLoads(LD2, LD1, LD1Bytes, 1) && TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT, *LD1->getMemOperand(), &LD1Fast) && LD1Fast) return DAG.getLoad(VT, SDLoc(N), LD1->getChain(), LD1->getBasePtr(), LD1->getPointerInfo(), LD1->getAlign()); return SDValue(); } static unsigned getPPCf128HiElementSelector(const SelectionDAG &DAG) { // On little-endian machines, bitcasting from ppcf128 to i128 does swap the Hi // and Lo parts; on big-endian machines it doesn't. return DAG.getDataLayout().isBigEndian() ? 1 : 0; } static SDValue foldBitcastedFPLogic(SDNode *N, SelectionDAG &DAG, const TargetLowering &TLI) { // If this is not a bitcast to an FP type or if the target doesn't have // IEEE754-compliant FP logic, we're done. EVT VT = N->getValueType(0); if (!VT.isFloatingPoint() || !TLI.hasBitPreservingFPLogic(VT)) return SDValue(); // TODO: Handle cases where the integer constant is a different scalar // bitwidth to the FP. SDValue N0 = N->getOperand(0); EVT SourceVT = N0.getValueType(); if (VT.getScalarSizeInBits() != SourceVT.getScalarSizeInBits()) return SDValue(); unsigned FPOpcode; APInt SignMask; switch (N0.getOpcode()) { case ISD::AND: FPOpcode = ISD::FABS; SignMask = ~APInt::getSignMask(SourceVT.getScalarSizeInBits()); break; case ISD::XOR: FPOpcode = ISD::FNEG; SignMask = APInt::getSignMask(SourceVT.getScalarSizeInBits()); break; case ISD::OR: FPOpcode = ISD::FABS; SignMask = APInt::getSignMask(SourceVT.getScalarSizeInBits()); break; default: return SDValue(); } // Fold (bitcast int (and (bitcast fp X to int), 0x7fff...) to fp) -> fabs X // Fold (bitcast int (xor (bitcast fp X to int), 0x8000...) to fp) -> fneg X // Fold (bitcast int (or (bitcast fp X to int), 0x8000...) to fp) -> // fneg (fabs X) SDValue LogicOp0 = N0.getOperand(0); ConstantSDNode *LogicOp1 = isConstOrConstSplat(N0.getOperand(1), true); if (LogicOp1 && LogicOp1->getAPIntValue() == SignMask && LogicOp0.getOpcode() == ISD::BITCAST && LogicOp0.getOperand(0).getValueType() == VT) { SDValue FPOp = DAG.getNode(FPOpcode, SDLoc(N), VT, LogicOp0.getOperand(0)); NumFPLogicOpsConv++; if (N0.getOpcode() == ISD::OR) return DAG.getNode(ISD::FNEG, SDLoc(N), VT, FPOp); return FPOp; } return SDValue(); } SDValue DAGCombiner::visitBITCAST(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); if (N0.isUndef()) return DAG.getUNDEF(VT); // If the input is a BUILD_VECTOR with all constant elements, fold this now. // Only do this before legalize types, unless both types are integer and the // scalar type is legal. Only do this before legalize ops, since the target // maybe depending on the bitcast. // First check to see if this is all constant. // TODO: Support FP bitcasts after legalize types. if (VT.isVector() && (!LegalTypes || (!LegalOperations && VT.isInteger() && N0.getValueType().isInteger() && TLI.isTypeLegal(VT.getVectorElementType()))) && N0.getOpcode() == ISD::BUILD_VECTOR && N0->hasOneUse() && cast(N0)->isConstant()) return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), VT.getVectorElementType()); // If the input is a constant, let getNode fold it. if (isIntOrFPConstant(N0)) { // If we can't allow illegal operations, we need to check that this is just // a fp -> int or int -> conversion and that the resulting operation will // be legal. if (!LegalOperations || (isa(N0) && VT.isFloatingPoint() && !VT.isVector() && TLI.isOperationLegal(ISD::ConstantFP, VT)) || (isa(N0) && VT.isInteger() && !VT.isVector() && TLI.isOperationLegal(ISD::Constant, VT))) { SDValue C = DAG.getBitcast(VT, N0); if (C.getNode() != N) return C; } } // (conv (conv x, t1), t2) -> (conv x, t2) if (N0.getOpcode() == ISD::BITCAST) return DAG.getBitcast(VT, N0.getOperand(0)); // fold (conv (load x)) -> (load (conv*)x) // If the resultant load doesn't need a higher alignment than the original! if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() && // Do not remove the cast if the types differ in endian layout. TLI.hasBigEndianPartOrdering(N0.getValueType(), DAG.getDataLayout()) == TLI.hasBigEndianPartOrdering(VT, DAG.getDataLayout()) && // If the load is volatile, we only want to change the load type if the // resulting load is legal. Otherwise we might increase the number of // memory accesses. We don't care if the original type was legal or not // as we assume software couldn't rely on the number of accesses of an // illegal type. ((!LegalOperations && cast(N0)->isSimple()) || TLI.isOperationLegal(ISD::LOAD, VT))) { LoadSDNode *LN0 = cast(N0); if (TLI.isLoadBitCastBeneficial(N0.getValueType(), VT, DAG, *LN0->getMemOperand())) { SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(), LN0->getPointerInfo(), LN0->getAlign(), LN0->getMemOperand()->getFlags(), LN0->getAAInfo()); DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1)); return Load; } } if (SDValue V = foldBitcastedFPLogic(N, DAG, TLI)) return V; // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit) // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit)) // // For ppc_fp128: // fold (bitcast (fneg x)) -> // flipbit = signbit // (xor (bitcast x) (build_pair flipbit, flipbit)) // // fold (bitcast (fabs x)) -> // flipbit = (and (extract_element (bitcast x), 0), signbit) // (xor (bitcast x) (build_pair flipbit, flipbit)) // This often reduces constant pool loads. if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) || (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) && N0->hasOneUse() && VT.isInteger() && !VT.isVector() && !N0.getValueType().isVector()) { SDValue NewConv = DAG.getBitcast(VT, N0.getOperand(0)); AddToWorklist(NewConv.getNode()); SDLoc DL(N); if (N0.getValueType() == MVT::ppcf128 && !LegalTypes) { assert(VT.getSizeInBits() == 128); SDValue SignBit = DAG.getConstant( APInt::getSignMask(VT.getSizeInBits() / 2), SDLoc(N0), MVT::i64); SDValue FlipBit; if (N0.getOpcode() == ISD::FNEG) { FlipBit = SignBit; AddToWorklist(FlipBit.getNode()); } else { assert(N0.getOpcode() == ISD::FABS); SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, SDLoc(NewConv), MVT::i64, NewConv, DAG.getIntPtrConstant(getPPCf128HiElementSelector(DAG), SDLoc(NewConv))); AddToWorklist(Hi.getNode()); FlipBit = DAG.getNode(ISD::AND, SDLoc(N0), MVT::i64, Hi, SignBit); AddToWorklist(FlipBit.getNode()); } SDValue FlipBits = DAG.getNode(ISD::BUILD_PAIR, SDLoc(N0), VT, FlipBit, FlipBit); AddToWorklist(FlipBits.getNode()); return DAG.getNode(ISD::XOR, DL, VT, NewConv, FlipBits); } APInt SignBit = APInt::getSignMask(VT.getSizeInBits()); if (N0.getOpcode() == ISD::FNEG) return DAG.getNode(ISD::XOR, DL, VT, NewConv, DAG.getConstant(SignBit, DL, VT)); assert(N0.getOpcode() == ISD::FABS); return DAG.getNode(ISD::AND, DL, VT, NewConv, DAG.getConstant(~SignBit, DL, VT)); } // fold (bitconvert (fcopysign cst, x)) -> // (or (and (bitconvert x), sign), (and cst, (not sign))) // Note that we don't handle (copysign x, cst) because this can always be // folded to an fneg or fabs. // // For ppc_fp128: // fold (bitcast (fcopysign cst, x)) -> // flipbit = (and (extract_element // (xor (bitcast cst), (bitcast x)), 0), // signbit) // (xor (bitcast cst) (build_pair flipbit, flipbit)) if (N0.getOpcode() == ISD::FCOPYSIGN && N0->hasOneUse() && isa(N0.getOperand(0)) && VT.isInteger() && !VT.isVector()) { unsigned OrigXWidth = N0.getOperand(1).getValueSizeInBits(); EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth); if (isTypeLegal(IntXVT)) { SDValue X = DAG.getBitcast(IntXVT, N0.getOperand(1)); AddToWorklist(X.getNode()); // If X has a different width than the result/lhs, sext it or truncate it. unsigned VTWidth = VT.getSizeInBits(); if (OrigXWidth < VTWidth) { X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X); AddToWorklist(X.getNode()); } else if (OrigXWidth > VTWidth) { // To get the sign bit in the right place, we have to shift it right // before truncating. SDLoc DL(X); X = DAG.getNode(ISD::SRL, DL, X.getValueType(), X, DAG.getConstant(OrigXWidth-VTWidth, DL, X.getValueType())); AddToWorklist(X.getNode()); X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X); AddToWorklist(X.getNode()); } if (N0.getValueType() == MVT::ppcf128 && !LegalTypes) { APInt SignBit = APInt::getSignMask(VT.getSizeInBits() / 2); SDValue Cst = DAG.getBitcast(VT, N0.getOperand(0)); AddToWorklist(Cst.getNode()); SDValue X = DAG.getBitcast(VT, N0.getOperand(1)); AddToWorklist(X.getNode()); SDValue XorResult = DAG.getNode(ISD::XOR, SDLoc(N0), VT, Cst, X); AddToWorklist(XorResult.getNode()); SDValue XorResult64 = DAG.getNode( ISD::EXTRACT_ELEMENT, SDLoc(XorResult), MVT::i64, XorResult, DAG.getIntPtrConstant(getPPCf128HiElementSelector(DAG), SDLoc(XorResult))); AddToWorklist(XorResult64.getNode()); SDValue FlipBit = DAG.getNode(ISD::AND, SDLoc(XorResult64), MVT::i64, XorResult64, DAG.getConstant(SignBit, SDLoc(XorResult64), MVT::i64)); AddToWorklist(FlipBit.getNode()); SDValue FlipBits = DAG.getNode(ISD::BUILD_PAIR, SDLoc(N0), VT, FlipBit, FlipBit); AddToWorklist(FlipBits.getNode()); return DAG.getNode(ISD::XOR, SDLoc(N), VT, Cst, FlipBits); } APInt SignBit = APInt::getSignMask(VT.getSizeInBits()); X = DAG.getNode(ISD::AND, SDLoc(X), VT, X, DAG.getConstant(SignBit, SDLoc(X), VT)); AddToWorklist(X.getNode()); SDValue Cst = DAG.getBitcast(VT, N0.getOperand(0)); Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT, Cst, DAG.getConstant(~SignBit, SDLoc(Cst), VT)); AddToWorklist(Cst.getNode()); return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst); } } // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive. if (N0.getOpcode() == ISD::BUILD_PAIR) if (SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT)) return CombineLD; // Remove double bitcasts from shuffles - this is often a legacy of // XformToShuffleWithZero being used to combine bitmaskings (of // float vectors bitcast to integer vectors) into shuffles. // bitcast(shuffle(bitcast(s0),bitcast(s1))) -> shuffle(s0,s1) if (Level < AfterLegalizeDAG && TLI.isTypeLegal(VT) && VT.isVector() && N0->getOpcode() == ISD::VECTOR_SHUFFLE && N0.hasOneUse() && VT.getVectorNumElements() >= N0.getValueType().getVectorNumElements() && !(VT.getVectorNumElements() % N0.getValueType().getVectorNumElements())) { ShuffleVectorSDNode *SVN = cast(N0); // If operands are a bitcast, peek through if it casts the original VT. // If operands are a constant, just bitcast back to original VT. auto PeekThroughBitcast = [&](SDValue Op) { if (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).getValueType() == VT) return SDValue(Op.getOperand(0)); if (Op.isUndef() || isAnyConstantBuildVector(Op)) return DAG.getBitcast(VT, Op); return SDValue(); }; // FIXME: If either input vector is bitcast, try to convert the shuffle to // the result type of this bitcast. This would eliminate at least one // bitcast. See the transform in InstCombine. SDValue SV0 = PeekThroughBitcast(N0->getOperand(0)); SDValue SV1 = PeekThroughBitcast(N0->getOperand(1)); if (!(SV0 && SV1)) return SDValue(); int MaskScale = VT.getVectorNumElements() / N0.getValueType().getVectorNumElements(); SmallVector NewMask; for (int M : SVN->getMask()) for (int i = 0; i != MaskScale; ++i) NewMask.push_back(M < 0 ? -1 : M * MaskScale + i); SDValue LegalShuffle = TLI.buildLegalVectorShuffle(VT, SDLoc(N), SV0, SV1, NewMask, DAG); if (LegalShuffle) return LegalShuffle; } return SDValue(); } SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) { EVT VT = N->getValueType(0); return CombineConsecutiveLoads(N, VT); } SDValue DAGCombiner::visitFREEZE(SDNode *N) { SDValue N0 = N->getOperand(0); if (DAG.isGuaranteedNotToBeUndefOrPoison(N0, /*PoisonOnly*/ false)) return N0; // Fold freeze(bitcast(x)) -> bitcast(freeze(x)). // TODO: Replace with pushFreezeToPreventPoisonFromPropagating fold. if (N0.getOpcode() == ISD::BITCAST) return DAG.getBitcast(N->getValueType(0), DAG.getNode(ISD::FREEZE, SDLoc(N0), N0.getOperand(0).getValueType(), N0.getOperand(0))); return SDValue(); } /// We know that BV is a build_vector node with Constant, ConstantFP or Undef /// operands. DstEltVT indicates the destination element value type. SDValue DAGCombiner:: ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) { EVT SrcEltVT = BV->getValueType(0).getVectorElementType(); // If this is already the right type, we're done. if (SrcEltVT == DstEltVT) return SDValue(BV, 0); unsigned SrcBitSize = SrcEltVT.getSizeInBits(); unsigned DstBitSize = DstEltVT.getSizeInBits(); // If this is a conversion of N elements of one type to N elements of another // type, convert each element. This handles FP<->INT cases. if (SrcBitSize == DstBitSize) { SmallVector Ops; for (SDValue Op : BV->op_values()) { // If the vector element type is not legal, the BUILD_VECTOR operands // are promoted and implicitly truncated. Make that explicit here. if (Op.getValueType() != SrcEltVT) Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op); Ops.push_back(DAG.getBitcast(DstEltVT, Op)); AddToWorklist(Ops.back().getNode()); } EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, BV->getValueType(0).getVectorNumElements()); return DAG.getBuildVector(VT, SDLoc(BV), Ops); } // Otherwise, we're growing or shrinking the elements. To avoid having to // handle annoying details of growing/shrinking FP values, we convert them to // int first. if (SrcEltVT.isFloatingPoint()) { // Convert the input float vector to a int vector where the elements are the // same sizes. EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits()); BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode(); SrcEltVT = IntVT; } // Now we know the input is an integer vector. If the output is a FP type, // convert to integer first, then to FP of the right size. if (DstEltVT.isFloatingPoint()) { EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits()); SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode(); // Next, convert to FP elements of the same size. return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT); } // Okay, we know the src/dst types are both integers of differing types. assert(SrcEltVT.isInteger() && DstEltVT.isInteger()); // TODO: Should ConstantFoldBITCASTofBUILD_VECTOR always take a // BuildVectorSDNode? auto *BVN = cast(BV); // Extract the constant raw bit data. BitVector UndefElements; SmallVector RawBits; bool IsLE = DAG.getDataLayout().isLittleEndian(); if (!BVN->getConstantRawBits(IsLE, DstBitSize, RawBits, UndefElements)) return SDValue(); SDLoc DL(BV); SmallVector Ops; for (unsigned I = 0, E = RawBits.size(); I != E; ++I) { if (UndefElements[I]) Ops.push_back(DAG.getUNDEF(DstEltVT)); else Ops.push_back(DAG.getConstant(RawBits[I], DL, DstEltVT)); } EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size()); return DAG.getBuildVector(VT, DL, Ops); } // Returns true if floating point contraction is allowed on the FMUL-SDValue // `N` static bool isContractableFMUL(const TargetOptions &Options, SDValue N) { assert(N.getOpcode() == ISD::FMUL); return Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath || N->getFlags().hasAllowContract(); } // Returns true if `N` can assume no infinities involved in its computation. static bool hasNoInfs(const TargetOptions &Options, SDValue N) { return Options.NoInfsFPMath || N->getFlags().hasNoInfs(); } /// Try to perform FMA combining on a given FADD node. SDValue DAGCombiner::visitFADDForFMACombine(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); SDLoc SL(N); const TargetOptions &Options = DAG.getTarget().Options; // Floating-point multiply-add with intermediate rounding. bool HasFMAD = (LegalOperations && TLI.isFMADLegal(DAG, N)); // Floating-point multiply-add without intermediate rounding. bool HasFMA = TLI.isFMAFasterThanFMulAndFAdd(DAG.getMachineFunction(), VT) && (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT)); // No valid opcode, do not combine. if (!HasFMAD && !HasFMA) return SDValue(); bool CanReassociate = Options.UnsafeFPMath || N->getFlags().hasAllowReassociation(); bool AllowFusionGlobally = (Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath || HasFMAD); // If the addition is not contractable, do not combine. if (!AllowFusionGlobally && !N->getFlags().hasAllowContract()) return SDValue(); if (TLI.generateFMAsInMachineCombiner(VT, OptLevel)) return SDValue(); // Always prefer FMAD to FMA for precision. unsigned PreferredFusedOpcode = HasFMAD ? ISD::FMAD : ISD::FMA; bool Aggressive = TLI.enableAggressiveFMAFusion(VT); auto isFusedOp = [&](SDValue N) { unsigned Opcode = N.getOpcode(); return Opcode == ISD::FMA || Opcode == ISD::FMAD; }; // Is the node an FMUL and contractable either due to global flags or // SDNodeFlags. auto isContractableFMUL = [AllowFusionGlobally](SDValue N) { if (N.getOpcode() != ISD::FMUL) return false; return AllowFusionGlobally || N->getFlags().hasAllowContract(); }; // If we have two choices trying to fold (fadd (fmul u, v), (fmul x, y)), // prefer to fold the multiply with fewer uses. if (Aggressive && isContractableFMUL(N0) && isContractableFMUL(N1)) { if (N0->use_size() > N1->use_size()) std::swap(N0, N1); } // fold (fadd (fmul x, y), z) -> (fma x, y, z) if (isContractableFMUL(N0) && (Aggressive || N0->hasOneUse())) { return DAG.getNode(PreferredFusedOpcode, SL, VT, N0.getOperand(0), N0.getOperand(1), N1); } // fold (fadd x, (fmul y, z)) -> (fma y, z, x) // Note: Commutes FADD operands. if (isContractableFMUL(N1) && (Aggressive || N1->hasOneUse())) { return DAG.getNode(PreferredFusedOpcode, SL, VT, N1.getOperand(0), N1.getOperand(1), N0); } // fadd (fma A, B, (fmul C, D)), E --> fma A, B, (fma C, D, E) // fadd E, (fma A, B, (fmul C, D)) --> fma A, B, (fma C, D, E) // This requires reassociation because it changes the order of operations. SDValue FMA, E; if (CanReassociate && isFusedOp(N0) && N0.getOperand(2).getOpcode() == ISD::FMUL && N0.hasOneUse() && N0.getOperand(2).hasOneUse()) { FMA = N0; E = N1; } else if (CanReassociate && isFusedOp(N1) && N1.getOperand(2).getOpcode() == ISD::FMUL && N1.hasOneUse() && N1.getOperand(2).hasOneUse()) { FMA = N1; E = N0; } if (FMA && E) { SDValue A = FMA.getOperand(0); SDValue B = FMA.getOperand(1); SDValue C = FMA.getOperand(2).getOperand(0); SDValue D = FMA.getOperand(2).getOperand(1); SDValue CDE = DAG.getNode(PreferredFusedOpcode, SL, VT, C, D, E); return DAG.getNode(PreferredFusedOpcode, SL, VT, A, B, CDE); } // Look through FP_EXTEND nodes to do more combining. // fold (fadd (fpext (fmul x, y)), z) -> (fma (fpext x), (fpext y), z) if (N0.getOpcode() == ISD::FP_EXTEND) { SDValue N00 = N0.getOperand(0); if (isContractableFMUL(N00) && TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT, N00.getValueType())) { return DAG.getNode(PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FP_EXTEND, SL, VT, N00.getOperand(0)), DAG.getNode(ISD::FP_EXTEND, SL, VT, N00.getOperand(1)), N1); } } // fold (fadd x, (fpext (fmul y, z))) -> (fma (fpext y), (fpext z), x) // Note: Commutes FADD operands. if (N1.getOpcode() == ISD::FP_EXTEND) { SDValue N10 = N1.getOperand(0); if (isContractableFMUL(N10) && TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT, N10.getValueType())) { return DAG.getNode(PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FP_EXTEND, SL, VT, N10.getOperand(0)), DAG.getNode(ISD::FP_EXTEND, SL, VT, N10.getOperand(1)), N0); } } // More folding opportunities when target permits. if (Aggressive) { // fold (fadd (fma x, y, (fpext (fmul u, v))), z) // -> (fma x, y, (fma (fpext u), (fpext v), z)) auto FoldFAddFMAFPExtFMul = [&](SDValue X, SDValue Y, SDValue U, SDValue V, SDValue Z) { return DAG.getNode(PreferredFusedOpcode, SL, VT, X, Y, DAG.getNode(PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FP_EXTEND, SL, VT, U), DAG.getNode(ISD::FP_EXTEND, SL, VT, V), Z)); }; if (isFusedOp(N0)) { SDValue N02 = N0.getOperand(2); if (N02.getOpcode() == ISD::FP_EXTEND) { SDValue N020 = N02.getOperand(0); if (isContractableFMUL(N020) && TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT, N020.getValueType())) { return FoldFAddFMAFPExtFMul(N0.getOperand(0), N0.getOperand(1), N020.getOperand(0), N020.getOperand(1), N1); } } } // fold (fadd (fpext (fma x, y, (fmul u, v))), z) // -> (fma (fpext x), (fpext y), (fma (fpext u), (fpext v), z)) // FIXME: This turns two single-precision and one double-precision // operation into two double-precision operations, which might not be // interesting for all targets, especially GPUs. auto FoldFAddFPExtFMAFMul = [&](SDValue X, SDValue Y, SDValue U, SDValue V, SDValue Z) { return DAG.getNode( PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FP_EXTEND, SL, VT, X), DAG.getNode(ISD::FP_EXTEND, SL, VT, Y), DAG.getNode(PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FP_EXTEND, SL, VT, U), DAG.getNode(ISD::FP_EXTEND, SL, VT, V), Z)); }; if (N0.getOpcode() == ISD::FP_EXTEND) { SDValue N00 = N0.getOperand(0); if (isFusedOp(N00)) { SDValue N002 = N00.getOperand(2); if (isContractableFMUL(N002) && TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT, N00.getValueType())) { return FoldFAddFPExtFMAFMul(N00.getOperand(0), N00.getOperand(1), N002.getOperand(0), N002.getOperand(1), N1); } } } // fold (fadd x, (fma y, z, (fpext (fmul u, v))) // -> (fma y, z, (fma (fpext u), (fpext v), x)) if (isFusedOp(N1)) { SDValue N12 = N1.getOperand(2); if (N12.getOpcode() == ISD::FP_EXTEND) { SDValue N120 = N12.getOperand(0); if (isContractableFMUL(N120) && TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT, N120.getValueType())) { return FoldFAddFMAFPExtFMul(N1.getOperand(0), N1.getOperand(1), N120.getOperand(0), N120.getOperand(1), N0); } } } // fold (fadd x, (fpext (fma y, z, (fmul u, v))) // -> (fma (fpext y), (fpext z), (fma (fpext u), (fpext v), x)) // FIXME: This turns two single-precision and one double-precision // operation into two double-precision operations, which might not be // interesting for all targets, especially GPUs. if (N1.getOpcode() == ISD::FP_EXTEND) { SDValue N10 = N1.getOperand(0); if (isFusedOp(N10)) { SDValue N102 = N10.getOperand(2); if (isContractableFMUL(N102) && TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT, N10.getValueType())) { return FoldFAddFPExtFMAFMul(N10.getOperand(0), N10.getOperand(1), N102.getOperand(0), N102.getOperand(1), N0); } } } } return SDValue(); } /// Try to perform FMA combining on a given FSUB node. SDValue DAGCombiner::visitFSUBForFMACombine(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); SDLoc SL(N); const TargetOptions &Options = DAG.getTarget().Options; // Floating-point multiply-add with intermediate rounding. bool HasFMAD = (LegalOperations && TLI.isFMADLegal(DAG, N)); // Floating-point multiply-add without intermediate rounding. bool HasFMA = TLI.isFMAFasterThanFMulAndFAdd(DAG.getMachineFunction(), VT) && (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT)); // No valid opcode, do not combine. if (!HasFMAD && !HasFMA) return SDValue(); const SDNodeFlags Flags = N->getFlags(); bool AllowFusionGlobally = (Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath || HasFMAD); // If the subtraction is not contractable, do not combine. if (!AllowFusionGlobally && !N->getFlags().hasAllowContract()) return SDValue(); if (TLI.generateFMAsInMachineCombiner(VT, OptLevel)) return SDValue(); // Always prefer FMAD to FMA for precision. unsigned PreferredFusedOpcode = HasFMAD ? ISD::FMAD : ISD::FMA; bool Aggressive = TLI.enableAggressiveFMAFusion(VT); bool NoSignedZero = Options.NoSignedZerosFPMath || Flags.hasNoSignedZeros(); // Is the node an FMUL and contractable either due to global flags or // SDNodeFlags. auto isContractableFMUL = [AllowFusionGlobally](SDValue N) { if (N.getOpcode() != ISD::FMUL) return false; return AllowFusionGlobally || N->getFlags().hasAllowContract(); }; // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z)) auto tryToFoldXYSubZ = [&](SDValue XY, SDValue Z) { if (isContractableFMUL(XY) && (Aggressive || XY->hasOneUse())) { return DAG.getNode(PreferredFusedOpcode, SL, VT, XY.getOperand(0), XY.getOperand(1), DAG.getNode(ISD::FNEG, SL, VT, Z)); } return SDValue(); }; // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x) // Note: Commutes FSUB operands. auto tryToFoldXSubYZ = [&](SDValue X, SDValue YZ) { if (isContractableFMUL(YZ) && (Aggressive || YZ->hasOneUse())) { return DAG.getNode(PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FNEG, SL, VT, YZ.getOperand(0)), YZ.getOperand(1), X); } return SDValue(); }; // If we have two choices trying to fold (fsub (fmul u, v), (fmul x, y)), // prefer to fold the multiply with fewer uses. if (isContractableFMUL(N0) && isContractableFMUL(N1) && (N0->use_size() > N1->use_size())) { // fold (fsub (fmul a, b), (fmul c, d)) -> (fma (fneg c), d, (fmul a, b)) if (SDValue V = tryToFoldXSubYZ(N0, N1)) return V; // fold (fsub (fmul a, b), (fmul c, d)) -> (fma a, b, (fneg (fmul c, d))) if (SDValue V = tryToFoldXYSubZ(N0, N1)) return V; } else { // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z)) if (SDValue V = tryToFoldXYSubZ(N0, N1)) return V; // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x) if (SDValue V = tryToFoldXSubYZ(N0, N1)) return V; } // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z)) if (N0.getOpcode() == ISD::FNEG && isContractableFMUL(N0.getOperand(0)) && (Aggressive || (N0->hasOneUse() && N0.getOperand(0).hasOneUse()))) { SDValue N00 = N0.getOperand(0).getOperand(0); SDValue N01 = N0.getOperand(0).getOperand(1); return DAG.getNode(PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FNEG, SL, VT, N00), N01, DAG.getNode(ISD::FNEG, SL, VT, N1)); } // Look through FP_EXTEND nodes to do more combining. // fold (fsub (fpext (fmul x, y)), z) // -> (fma (fpext x), (fpext y), (fneg z)) if (N0.getOpcode() == ISD::FP_EXTEND) { SDValue N00 = N0.getOperand(0); if (isContractableFMUL(N00) && TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT, N00.getValueType())) { return DAG.getNode(PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FP_EXTEND, SL, VT, N00.getOperand(0)), DAG.getNode(ISD::FP_EXTEND, SL, VT, N00.getOperand(1)), DAG.getNode(ISD::FNEG, SL, VT, N1)); } } // fold (fsub x, (fpext (fmul y, z))) // -> (fma (fneg (fpext y)), (fpext z), x) // Note: Commutes FSUB operands. if (N1.getOpcode() == ISD::FP_EXTEND) { SDValue N10 = N1.getOperand(0); if (isContractableFMUL(N10) && TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT, N10.getValueType())) { return DAG.getNode( PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FNEG, SL, VT, DAG.getNode(ISD::FP_EXTEND, SL, VT, N10.getOperand(0))), DAG.getNode(ISD::FP_EXTEND, SL, VT, N10.getOperand(1)), N0); } } // fold (fsub (fpext (fneg (fmul, x, y))), z) // -> (fneg (fma (fpext x), (fpext y), z)) // Note: This could be removed with appropriate canonicalization of the // input expression into (fneg (fadd (fpext (fmul, x, y)), z). However, the // orthogonal flags -fp-contract=fast and -enable-unsafe-fp-math prevent // from implementing the canonicalization in visitFSUB. if (N0.getOpcode() == ISD::FP_EXTEND) { SDValue N00 = N0.getOperand(0); if (N00.getOpcode() == ISD::FNEG) { SDValue N000 = N00.getOperand(0); if (isContractableFMUL(N000) && TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT, N00.getValueType())) { return DAG.getNode( ISD::FNEG, SL, VT, DAG.getNode(PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FP_EXTEND, SL, VT, N000.getOperand(0)), DAG.getNode(ISD::FP_EXTEND, SL, VT, N000.getOperand(1)), N1)); } } } // fold (fsub (fneg (fpext (fmul, x, y))), z) // -> (fneg (fma (fpext x)), (fpext y), z) // Note: This could be removed with appropriate canonicalization of the // input expression into (fneg (fadd (fpext (fmul, x, y)), z). However, the // orthogonal flags -fp-contract=fast and -enable-unsafe-fp-math prevent // from implementing the canonicalization in visitFSUB. if (N0.getOpcode() == ISD::FNEG) { SDValue N00 = N0.getOperand(0); if (N00.getOpcode() == ISD::FP_EXTEND) { SDValue N000 = N00.getOperand(0); if (isContractableFMUL(N000) && TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT, N000.getValueType())) { return DAG.getNode( ISD::FNEG, SL, VT, DAG.getNode(PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FP_EXTEND, SL, VT, N000.getOperand(0)), DAG.getNode(ISD::FP_EXTEND, SL, VT, N000.getOperand(1)), N1)); } } } auto isReassociable = [Options](SDNode *N) { return Options.UnsafeFPMath || N->getFlags().hasAllowReassociation(); }; auto isContractableAndReassociableFMUL = [isContractableFMUL, isReassociable](SDValue N) { return isContractableFMUL(N) && isReassociable(N.getNode()); }; auto isFusedOp = [&](SDValue N) { unsigned Opcode = N.getOpcode(); return Opcode == ISD::FMA || Opcode == ISD::FMAD; }; // More folding opportunities when target permits. if (Aggressive && isReassociable(N)) { bool CanFuse = Options.UnsafeFPMath || N->getFlags().hasAllowContract(); // fold (fsub (fma x, y, (fmul u, v)), z) // -> (fma x, y (fma u, v, (fneg z))) if (CanFuse && isFusedOp(N0) && isContractableAndReassociableFMUL(N0.getOperand(2)) && N0->hasOneUse() && N0.getOperand(2)->hasOneUse()) { return DAG.getNode(PreferredFusedOpcode, SL, VT, N0.getOperand(0), N0.getOperand(1), DAG.getNode(PreferredFusedOpcode, SL, VT, N0.getOperand(2).getOperand(0), N0.getOperand(2).getOperand(1), DAG.getNode(ISD::FNEG, SL, VT, N1))); } // fold (fsub x, (fma y, z, (fmul u, v))) // -> (fma (fneg y), z, (fma (fneg u), v, x)) if (CanFuse && isFusedOp(N1) && isContractableAndReassociableFMUL(N1.getOperand(2)) && N1->hasOneUse() && NoSignedZero) { SDValue N20 = N1.getOperand(2).getOperand(0); SDValue N21 = N1.getOperand(2).getOperand(1); return DAG.getNode( PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FNEG, SL, VT, N1.getOperand(0)), N1.getOperand(1), DAG.getNode(PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FNEG, SL, VT, N20), N21, N0)); } // fold (fsub (fma x, y, (fpext (fmul u, v))), z) // -> (fma x, y (fma (fpext u), (fpext v), (fneg z))) if (isFusedOp(N0) && N0->hasOneUse()) { SDValue N02 = N0.getOperand(2); if (N02.getOpcode() == ISD::FP_EXTEND) { SDValue N020 = N02.getOperand(0); if (isContractableAndReassociableFMUL(N020) && TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT, N020.getValueType())) { return DAG.getNode( PreferredFusedOpcode, SL, VT, N0.getOperand(0), N0.getOperand(1), DAG.getNode( PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FP_EXTEND, SL, VT, N020.getOperand(0)), DAG.getNode(ISD::FP_EXTEND, SL, VT, N020.getOperand(1)), DAG.getNode(ISD::FNEG, SL, VT, N1))); } } } // fold (fsub (fpext (fma x, y, (fmul u, v))), z) // -> (fma (fpext x), (fpext y), // (fma (fpext u), (fpext v), (fneg z))) // FIXME: This turns two single-precision and one double-precision // operation into two double-precision operations, which might not be // interesting for all targets, especially GPUs. if (N0.getOpcode() == ISD::FP_EXTEND) { SDValue N00 = N0.getOperand(0); if (isFusedOp(N00)) { SDValue N002 = N00.getOperand(2); if (isContractableAndReassociableFMUL(N002) && TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT, N00.getValueType())) { return DAG.getNode( PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FP_EXTEND, SL, VT, N00.getOperand(0)), DAG.getNode(ISD::FP_EXTEND, SL, VT, N00.getOperand(1)), DAG.getNode( PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FP_EXTEND, SL, VT, N002.getOperand(0)), DAG.getNode(ISD::FP_EXTEND, SL, VT, N002.getOperand(1)), DAG.getNode(ISD::FNEG, SL, VT, N1))); } } } // fold (fsub x, (fma y, z, (fpext (fmul u, v)))) // -> (fma (fneg y), z, (fma (fneg (fpext u)), (fpext v), x)) if (isFusedOp(N1) && N1.getOperand(2).getOpcode() == ISD::FP_EXTEND && N1->hasOneUse()) { SDValue N120 = N1.getOperand(2).getOperand(0); if (isContractableAndReassociableFMUL(N120) && TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT, N120.getValueType())) { SDValue N1200 = N120.getOperand(0); SDValue N1201 = N120.getOperand(1); return DAG.getNode( PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FNEG, SL, VT, N1.getOperand(0)), N1.getOperand(1), DAG.getNode(PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FNEG, SL, VT, DAG.getNode(ISD::FP_EXTEND, SL, VT, N1200)), DAG.getNode(ISD::FP_EXTEND, SL, VT, N1201), N0)); } } // fold (fsub x, (fpext (fma y, z, (fmul u, v)))) // -> (fma (fneg (fpext y)), (fpext z), // (fma (fneg (fpext u)), (fpext v), x)) // FIXME: This turns two single-precision and one double-precision // operation into two double-precision operations, which might not be // interesting for all targets, especially GPUs. if (N1.getOpcode() == ISD::FP_EXTEND && isFusedOp(N1.getOperand(0))) { SDValue CvtSrc = N1.getOperand(0); SDValue N100 = CvtSrc.getOperand(0); SDValue N101 = CvtSrc.getOperand(1); SDValue N102 = CvtSrc.getOperand(2); if (isContractableAndReassociableFMUL(N102) && TLI.isFPExtFoldable(DAG, PreferredFusedOpcode, VT, CvtSrc.getValueType())) { SDValue N1020 = N102.getOperand(0); SDValue N1021 = N102.getOperand(1); return DAG.getNode( PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FNEG, SL, VT, DAG.getNode(ISD::FP_EXTEND, SL, VT, N100)), DAG.getNode(ISD::FP_EXTEND, SL, VT, N101), DAG.getNode(PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FNEG, SL, VT, DAG.getNode(ISD::FP_EXTEND, SL, VT, N1020)), DAG.getNode(ISD::FP_EXTEND, SL, VT, N1021), N0)); } } } return SDValue(); } /// Try to perform FMA combining on a given FMUL node based on the distributive /// law x * (y + 1) = x * y + x and variants thereof (commuted versions, /// subtraction instead of addition). SDValue DAGCombiner::visitFMULForFMADistributiveCombine(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); SDLoc SL(N); assert(N->getOpcode() == ISD::FMUL && "Expected FMUL Operation"); const TargetOptions &Options = DAG.getTarget().Options; // The transforms below are incorrect when x == 0 and y == inf, because the // intermediate multiplication produces a nan. SDValue FAdd = N0.getOpcode() == ISD::FADD ? N0 : N1; if (!hasNoInfs(Options, FAdd)) return SDValue(); // Floating-point multiply-add without intermediate rounding. bool HasFMA = isContractableFMUL(Options, SDValue(N, 0)) && TLI.isFMAFasterThanFMulAndFAdd(DAG.getMachineFunction(), VT) && (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT)); // Floating-point multiply-add with intermediate rounding. This can result // in a less precise result due to the changed rounding order. bool HasFMAD = Options.UnsafeFPMath && (LegalOperations && TLI.isFMADLegal(DAG, N)); // No valid opcode, do not combine. if (!HasFMAD && !HasFMA) return SDValue(); // Always prefer FMAD to FMA for precision. unsigned PreferredFusedOpcode = HasFMAD ? ISD::FMAD : ISD::FMA; bool Aggressive = TLI.enableAggressiveFMAFusion(VT); // fold (fmul (fadd x0, +1.0), y) -> (fma x0, y, y) // fold (fmul (fadd x0, -1.0), y) -> (fma x0, y, (fneg y)) auto FuseFADD = [&](SDValue X, SDValue Y) { if (X.getOpcode() == ISD::FADD && (Aggressive || X->hasOneUse())) { if (auto *C = isConstOrConstSplatFP(X.getOperand(1), true)) { if (C->isExactlyValue(+1.0)) return DAG.getNode(PreferredFusedOpcode, SL, VT, X.getOperand(0), Y, Y); if (C->isExactlyValue(-1.0)) return DAG.getNode(PreferredFusedOpcode, SL, VT, X.getOperand(0), Y, DAG.getNode(ISD::FNEG, SL, VT, Y)); } } return SDValue(); }; if (SDValue FMA = FuseFADD(N0, N1)) return FMA; if (SDValue FMA = FuseFADD(N1, N0)) return FMA; // fold (fmul (fsub +1.0, x1), y) -> (fma (fneg x1), y, y) // fold (fmul (fsub -1.0, x1), y) -> (fma (fneg x1), y, (fneg y)) // fold (fmul (fsub x0, +1.0), y) -> (fma x0, y, (fneg y)) // fold (fmul (fsub x0, -1.0), y) -> (fma x0, y, y) auto FuseFSUB = [&](SDValue X, SDValue Y) { if (X.getOpcode() == ISD::FSUB && (Aggressive || X->hasOneUse())) { if (auto *C0 = isConstOrConstSplatFP(X.getOperand(0), true)) { if (C0->isExactlyValue(+1.0)) return DAG.getNode(PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FNEG, SL, VT, X.getOperand(1)), Y, Y); if (C0->isExactlyValue(-1.0)) return DAG.getNode(PreferredFusedOpcode, SL, VT, DAG.getNode(ISD::FNEG, SL, VT, X.getOperand(1)), Y, DAG.getNode(ISD::FNEG, SL, VT, Y)); } if (auto *C1 = isConstOrConstSplatFP(X.getOperand(1), true)) { if (C1->isExactlyValue(+1.0)) return DAG.getNode(PreferredFusedOpcode, SL, VT, X.getOperand(0), Y, DAG.getNode(ISD::FNEG, SL, VT, Y)); if (C1->isExactlyValue(-1.0)) return DAG.getNode(PreferredFusedOpcode, SL, VT, X.getOperand(0), Y, Y); } } return SDValue(); }; if (SDValue FMA = FuseFSUB(N0, N1)) return FMA; if (SDValue FMA = FuseFSUB(N1, N0)) return FMA; return SDValue(); } SDValue DAGCombiner::visitFADD(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); bool N0CFP = DAG.isConstantFPBuildVectorOrConstantFP(N0); bool N1CFP = DAG.isConstantFPBuildVectorOrConstantFP(N1); EVT VT = N->getValueType(0); SDLoc DL(N); const TargetOptions &Options = DAG.getTarget().Options; SDNodeFlags Flags = N->getFlags(); SelectionDAG::FlagInserter FlagsInserter(DAG, N); if (SDValue R = DAG.simplifyFPBinop(N->getOpcode(), N0, N1, Flags)) return R; // fold (fadd c1, c2) -> c1 + c2 if (SDValue C = DAG.FoldConstantArithmetic(ISD::FADD, DL, VT, {N0, N1})) return C; // canonicalize constant to RHS if (N0CFP && !N1CFP) return DAG.getNode(ISD::FADD, DL, VT, N1, N0); // fold vector ops if (VT.isVector()) if (SDValue FoldedVOp = SimplifyVBinOp(N, DL)) return FoldedVOp; // N0 + -0.0 --> N0 (also allowed with +0.0 and fast-math) ConstantFPSDNode *N1C = isConstOrConstSplatFP(N1, true); if (N1C && N1C->isZero()) if (N1C->isNegative() || Options.NoSignedZerosFPMath || Flags.hasNoSignedZeros()) return N0; if (SDValue NewSel = foldBinOpIntoSelect(N)) return NewSel; // fold (fadd A, (fneg B)) -> (fsub A, B) if (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) if (SDValue NegN1 = TLI.getCheaperNegatedExpression( N1, DAG, LegalOperations, ForCodeSize)) return DAG.getNode(ISD::FSUB, DL, VT, N0, NegN1); // fold (fadd (fneg A), B) -> (fsub B, A) if (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) if (SDValue NegN0 = TLI.getCheaperNegatedExpression( N0, DAG, LegalOperations, ForCodeSize)) return DAG.getNode(ISD::FSUB, DL, VT, N1, NegN0); auto isFMulNegTwo = [](SDValue FMul) { if (!FMul.hasOneUse() || FMul.getOpcode() != ISD::FMUL) return false; auto *C = isConstOrConstSplatFP(FMul.getOperand(1), true); return C && C->isExactlyValue(-2.0); }; // fadd (fmul B, -2.0), A --> fsub A, (fadd B, B) if (isFMulNegTwo(N0)) { SDValue B = N0.getOperand(0); SDValue Add = DAG.getNode(ISD::FADD, DL, VT, B, B); return DAG.getNode(ISD::FSUB, DL, VT, N1, Add); } // fadd A, (fmul B, -2.0) --> fsub A, (fadd B, B) if (isFMulNegTwo(N1)) { SDValue B = N1.getOperand(0); SDValue Add = DAG.getNode(ISD::FADD, DL, VT, B, B); return DAG.getNode(ISD::FSUB, DL, VT, N0, Add); } // No FP constant should be created after legalization as Instruction // Selection pass has a hard time dealing with FP constants. bool AllowNewConst = (Level < AfterLegalizeDAG); // If nnan is enabled, fold lots of things. if ((Options.NoNaNsFPMath || Flags.hasNoNaNs()) && AllowNewConst) { // If allowed, fold (fadd (fneg x), x) -> 0.0 if (N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1) return DAG.getConstantFP(0.0, DL, VT); // If allowed, fold (fadd x, (fneg x)) -> 0.0 if (N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0) return DAG.getConstantFP(0.0, DL, VT); } // If 'unsafe math' or reassoc and nsz, fold lots of things. // TODO: break out portions of the transformations below for which Unsafe is // considered and which do not require both nsz and reassoc if (((Options.UnsafeFPMath && Options.NoSignedZerosFPMath) || (Flags.hasAllowReassociation() && Flags.hasNoSignedZeros())) && AllowNewConst) { // fadd (fadd x, c1), c2 -> fadd x, c1 + c2 if (N1CFP && N0.getOpcode() == ISD::FADD && DAG.isConstantFPBuildVectorOrConstantFP(N0.getOperand(1))) { SDValue NewC = DAG.getNode(ISD::FADD, DL, VT, N0.getOperand(1), N1); return DAG.getNode(ISD::FADD, DL, VT, N0.getOperand(0), NewC); } // We can fold chains of FADD's of the same value into multiplications. // This transform is not safe in general because we are reducing the number // of rounding steps. if (TLI.isOperationLegalOrCustom(ISD::FMUL, VT) && !N0CFP && !N1CFP) { if (N0.getOpcode() == ISD::FMUL) { bool CFP00 = DAG.isConstantFPBuildVectorOrConstantFP(N0.getOperand(0)); bool CFP01 = DAG.isConstantFPBuildVectorOrConstantFP(N0.getOperand(1)); // (fadd (fmul x, c), x) -> (fmul x, c+1) if (CFP01 && !CFP00 && N0.getOperand(0) == N1) { SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, N0.getOperand(1), DAG.getConstantFP(1.0, DL, VT)); return DAG.getNode(ISD::FMUL, DL, VT, N1, NewCFP); } // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2) if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD && N1.getOperand(0) == N1.getOperand(1) && N0.getOperand(0) == N1.getOperand(0)) { SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, N0.getOperand(1), DAG.getConstantFP(2.0, DL, VT)); return DAG.getNode(ISD::FMUL, DL, VT, N0.getOperand(0), NewCFP); } } if (N1.getOpcode() == ISD::FMUL) { bool CFP10 = DAG.isConstantFPBuildVectorOrConstantFP(N1.getOperand(0)); bool CFP11 = DAG.isConstantFPBuildVectorOrConstantFP(N1.getOperand(1)); // (fadd x, (fmul x, c)) -> (fmul x, c+1) if (CFP11 && !CFP10 && N1.getOperand(0) == N0) { SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, N1.getOperand(1), DAG.getConstantFP(1.0, DL, VT)); return DAG.getNode(ISD::FMUL, DL, VT, N0, NewCFP); } // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2) if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD && N0.getOperand(0) == N0.getOperand(1) && N1.getOperand(0) == N0.getOperand(0)) { SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, N1.getOperand(1), DAG.getConstantFP(2.0, DL, VT)); return DAG.getNode(ISD::FMUL, DL, VT, N1.getOperand(0), NewCFP); } } if (N0.getOpcode() == ISD::FADD) { bool CFP00 = DAG.isConstantFPBuildVectorOrConstantFP(N0.getOperand(0)); // (fadd (fadd x, x), x) -> (fmul x, 3.0) if (!CFP00 && N0.getOperand(0) == N0.getOperand(1) && (N0.getOperand(0) == N1)) { return DAG.getNode(ISD::FMUL, DL, VT, N1, DAG.getConstantFP(3.0, DL, VT)); } } if (N1.getOpcode() == ISD::FADD) { bool CFP10 = DAG.isConstantFPBuildVectorOrConstantFP(N1.getOperand(0)); // (fadd x, (fadd x, x)) -> (fmul x, 3.0) if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) && N1.getOperand(0) == N0) { return DAG.getNode(ISD::FMUL, DL, VT, N0, DAG.getConstantFP(3.0, DL, VT)); } } // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0) if (N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD && N0.getOperand(0) == N0.getOperand(1) && N1.getOperand(0) == N1.getOperand(1) && N0.getOperand(0) == N1.getOperand(0)) { return DAG.getNode(ISD::FMUL, DL, VT, N0.getOperand(0), DAG.getConstantFP(4.0, DL, VT)); } } } // enable-unsafe-fp-math // FADD -> FMA combines: if (SDValue Fused = visitFADDForFMACombine(N)) { AddToWorklist(Fused.getNode()); return Fused; } return SDValue(); } SDValue DAGCombiner::visitSTRICT_FADD(SDNode *N) { SDValue Chain = N->getOperand(0); SDValue N0 = N->getOperand(1); SDValue N1 = N->getOperand(2); EVT VT = N->getValueType(0); EVT ChainVT = N->getValueType(1); SDLoc DL(N); SelectionDAG::FlagInserter FlagsInserter(DAG, N); // fold (strict_fadd A, (fneg B)) -> (strict_fsub A, B) if (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::STRICT_FSUB, VT)) if (SDValue NegN1 = TLI.getCheaperNegatedExpression( N1, DAG, LegalOperations, ForCodeSize)) { return DAG.getNode(ISD::STRICT_FSUB, DL, DAG.getVTList(VT, ChainVT), {Chain, N0, NegN1}); } // fold (strict_fadd (fneg A), B) -> (strict_fsub B, A) if (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::STRICT_FSUB, VT)) if (SDValue NegN0 = TLI.getCheaperNegatedExpression( N0, DAG, LegalOperations, ForCodeSize)) { return DAG.getNode(ISD::STRICT_FSUB, DL, DAG.getVTList(VT, ChainVT), {Chain, N1, NegN0}); } return SDValue(); } SDValue DAGCombiner::visitFSUB(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0, true); ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1, true); EVT VT = N->getValueType(0); SDLoc DL(N); const TargetOptions &Options = DAG.getTarget().Options; const SDNodeFlags Flags = N->getFlags(); SelectionDAG::FlagInserter FlagsInserter(DAG, N); if (SDValue R = DAG.simplifyFPBinop(N->getOpcode(), N0, N1, Flags)) return R; // fold (fsub c1, c2) -> c1-c2 if (SDValue C = DAG.FoldConstantArithmetic(ISD::FSUB, DL, VT, {N0, N1})) return C; // fold vector ops if (VT.isVector()) if (SDValue FoldedVOp = SimplifyVBinOp(N, DL)) return FoldedVOp; if (SDValue NewSel = foldBinOpIntoSelect(N)) return NewSel; // (fsub A, 0) -> A if (N1CFP && N1CFP->isZero()) { if (!N1CFP->isNegative() || Options.NoSignedZerosFPMath || Flags.hasNoSignedZeros()) { return N0; } } if (N0 == N1) { // (fsub x, x) -> 0.0 if (Options.NoNaNsFPMath || Flags.hasNoNaNs()) return DAG.getConstantFP(0.0f, DL, VT); } // (fsub -0.0, N1) -> -N1 if (N0CFP && N0CFP->isZero()) { if (N0CFP->isNegative() || (Options.NoSignedZerosFPMath || Flags.hasNoSignedZeros())) { // We cannot replace an FSUB(+-0.0,X) with FNEG(X) when denormals are // flushed to zero, unless all users treat denorms as zero (DAZ). // FIXME: This transform will change the sign of a NaN and the behavior // of a signaling NaN. It is only valid when a NoNaN flag is present. DenormalMode DenormMode = DAG.getDenormalMode(VT); if (DenormMode == DenormalMode::getIEEE()) { if (SDValue NegN1 = TLI.getNegatedExpression(N1, DAG, LegalOperations, ForCodeSize)) return NegN1; if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT)) return DAG.getNode(ISD::FNEG, DL, VT, N1); } } } if (((Options.UnsafeFPMath && Options.NoSignedZerosFPMath) || (Flags.hasAllowReassociation() && Flags.hasNoSignedZeros())) && N1.getOpcode() == ISD::FADD) { // X - (X + Y) -> -Y if (N0 == N1->getOperand(0)) return DAG.getNode(ISD::FNEG, DL, VT, N1->getOperand(1)); // X - (Y + X) -> -Y if (N0 == N1->getOperand(1)) return DAG.getNode(ISD::FNEG, DL, VT, N1->getOperand(0)); } // fold (fsub A, (fneg B)) -> (fadd A, B) if (SDValue NegN1 = TLI.getNegatedExpression(N1, DAG, LegalOperations, ForCodeSize)) return DAG.getNode(ISD::FADD, DL, VT, N0, NegN1); // FSUB -> FMA combines: if (SDValue Fused = visitFSUBForFMACombine(N)) { AddToWorklist(Fused.getNode()); return Fused; } return SDValue(); } SDValue DAGCombiner::visitFMUL(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1, true); EVT VT = N->getValueType(0); SDLoc DL(N); const TargetOptions &Options = DAG.getTarget().Options; const SDNodeFlags Flags = N->getFlags(); SelectionDAG::FlagInserter FlagsInserter(DAG, N); if (SDValue R = DAG.simplifyFPBinop(N->getOpcode(), N0, N1, Flags)) return R; // fold (fmul c1, c2) -> c1*c2 if (SDValue C = DAG.FoldConstantArithmetic(ISD::FMUL, DL, VT, {N0, N1})) return C; // canonicalize constant to RHS if (DAG.isConstantFPBuildVectorOrConstantFP(N0) && !DAG.isConstantFPBuildVectorOrConstantFP(N1)) return DAG.getNode(ISD::FMUL, DL, VT, N1, N0); // fold vector ops if (VT.isVector()) if (SDValue FoldedVOp = SimplifyVBinOp(N, DL)) return FoldedVOp; if (SDValue NewSel = foldBinOpIntoSelect(N)) return NewSel; if (Options.UnsafeFPMath || Flags.hasAllowReassociation()) { // fmul (fmul X, C1), C2 -> fmul X, C1 * C2 if (DAG.isConstantFPBuildVectorOrConstantFP(N1) && N0.getOpcode() == ISD::FMUL) { SDValue N00 = N0.getOperand(0); SDValue N01 = N0.getOperand(1); // Avoid an infinite loop by making sure that N00 is not a constant // (the inner multiply has not been constant folded yet). if (DAG.isConstantFPBuildVectorOrConstantFP(N01) && !DAG.isConstantFPBuildVectorOrConstantFP(N00)) { SDValue MulConsts = DAG.getNode(ISD::FMUL, DL, VT, N01, N1); return DAG.getNode(ISD::FMUL, DL, VT, N00, MulConsts); } } // Match a special-case: we convert X * 2.0 into fadd. // fmul (fadd X, X), C -> fmul X, 2.0 * C if (N0.getOpcode() == ISD::FADD && N0.hasOneUse() && N0.getOperand(0) == N0.getOperand(1)) { const SDValue Two = DAG.getConstantFP(2.0, DL, VT); SDValue MulConsts = DAG.getNode(ISD::FMUL, DL, VT, Two, N1); return DAG.getNode(ISD::FMUL, DL, VT, N0.getOperand(0), MulConsts); } } // fold (fmul X, 2.0) -> (fadd X, X) if (N1CFP && N1CFP->isExactlyValue(+2.0)) return DAG.getNode(ISD::FADD, DL, VT, N0, N0); // fold (fmul X, -1.0) -> (fsub -0.0, X) if (N1CFP && N1CFP->isExactlyValue(-1.0)) { if (!LegalOperations || TLI.isOperationLegal(ISD::FSUB, VT)) { return DAG.getNode(ISD::FSUB, DL, VT, DAG.getConstantFP(-0.0, DL, VT), N0, Flags); } } // -N0 * -N1 --> N0 * N1 TargetLowering::NegatibleCost CostN0 = TargetLowering::NegatibleCost::Expensive; TargetLowering::NegatibleCost CostN1 = TargetLowering::NegatibleCost::Expensive; SDValue NegN0 = TLI.getNegatedExpression(N0, DAG, LegalOperations, ForCodeSize, CostN0); SDValue NegN1 = TLI.getNegatedExpression(N1, DAG, LegalOperations, ForCodeSize, CostN1); if (NegN0 && NegN1 && (CostN0 == TargetLowering::NegatibleCost::Cheaper || CostN1 == TargetLowering::NegatibleCost::Cheaper)) return DAG.getNode(ISD::FMUL, DL, VT, NegN0, NegN1); // fold (fmul X, (select (fcmp X > 0.0), -1.0, 1.0)) -> (fneg (fabs X)) // fold (fmul X, (select (fcmp X > 0.0), 1.0, -1.0)) -> (fabs X) if (Flags.hasNoNaNs() && Flags.hasNoSignedZeros() && (N0.getOpcode() == ISD::SELECT || N1.getOpcode() == ISD::SELECT) && TLI.isOperationLegal(ISD::FABS, VT)) { SDValue Select = N0, X = N1; if (Select.getOpcode() != ISD::SELECT) std::swap(Select, X); SDValue Cond = Select.getOperand(0); auto TrueOpnd = dyn_cast(Select.getOperand(1)); auto FalseOpnd = dyn_cast(Select.getOperand(2)); if (TrueOpnd && FalseOpnd && Cond.getOpcode() == ISD::SETCC && Cond.getOperand(0) == X && isa(Cond.getOperand(1)) && cast(Cond.getOperand(1))->isExactlyValue(0.0)) { ISD::CondCode CC = cast(Cond.getOperand(2))->get(); switch (CC) { default: break; case ISD::SETOLT: case ISD::SETULT: case ISD::SETOLE: case ISD::SETULE: case ISD::SETLT: case ISD::SETLE: std::swap(TrueOpnd, FalseOpnd); LLVM_FALLTHROUGH; case ISD::SETOGT: case ISD::SETUGT: case ISD::SETOGE: case ISD::SETUGE: case ISD::SETGT: case ISD::SETGE: if (TrueOpnd->isExactlyValue(-1.0) && FalseOpnd->isExactlyValue(1.0) && TLI.isOperationLegal(ISD::FNEG, VT)) return DAG.getNode(ISD::FNEG, DL, VT, DAG.getNode(ISD::FABS, DL, VT, X)); if (TrueOpnd->isExactlyValue(1.0) && FalseOpnd->isExactlyValue(-1.0)) return DAG.getNode(ISD::FABS, DL, VT, X); break; } } } // FMUL -> FMA combines: if (SDValue Fused = visitFMULForFMADistributiveCombine(N)) { AddToWorklist(Fused.getNode()); return Fused; } return SDValue(); } SDValue DAGCombiner::visitFMA(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); SDValue N2 = N->getOperand(2); ConstantFPSDNode *N0CFP = dyn_cast(N0); ConstantFPSDNode *N1CFP = dyn_cast(N1); EVT VT = N->getValueType(0); SDLoc DL(N); const TargetOptions &Options = DAG.getTarget().Options; // FMA nodes have flags that propagate to the created nodes. SelectionDAG::FlagInserter FlagsInserter(DAG, N); bool CanReassociate = Options.UnsafeFPMath || N->getFlags().hasAllowReassociation(); // Constant fold FMA. if (isa(N0) && isa(N1) && isa(N2)) { return DAG.getNode(ISD::FMA, DL, VT, N0, N1, N2); } // (-N0 * -N1) + N2 --> (N0 * N1) + N2 TargetLowering::NegatibleCost CostN0 = TargetLowering::NegatibleCost::Expensive; TargetLowering::NegatibleCost CostN1 = TargetLowering::NegatibleCost::Expensive; SDValue NegN0 = TLI.getNegatedExpression(N0, DAG, LegalOperations, ForCodeSize, CostN0); SDValue NegN1 = TLI.getNegatedExpression(N1, DAG, LegalOperations, ForCodeSize, CostN1); if (NegN0 && NegN1 && (CostN0 == TargetLowering::NegatibleCost::Cheaper || CostN1 == TargetLowering::NegatibleCost::Cheaper)) return DAG.getNode(ISD::FMA, DL, VT, NegN0, NegN1, N2); // FIXME: use fast math flags instead of Options.UnsafeFPMath if (Options.UnsafeFPMath) { if (N0CFP && N0CFP->isZero()) return N2; if (N1CFP && N1CFP->isZero()) return N2; } if (N0CFP && N0CFP->isExactlyValue(1.0)) return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2); if (N1CFP && N1CFP->isExactlyValue(1.0)) return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2); // Canonicalize (fma c, x, y) -> (fma x, c, y) if (DAG.isConstantFPBuildVectorOrConstantFP(N0) && !DAG.isConstantFPBuildVectorOrConstantFP(N1)) return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2); if (CanReassociate) { // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2) if (N2.getOpcode() == ISD::FMUL && N0 == N2.getOperand(0) && DAG.isConstantFPBuildVectorOrConstantFP(N1) && DAG.isConstantFPBuildVectorOrConstantFP(N2.getOperand(1))) { return DAG.getNode(ISD::FMUL, DL, VT, N0, DAG.getNode(ISD::FADD, DL, VT, N1, N2.getOperand(1))); } // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y) if (N0.getOpcode() == ISD::FMUL && DAG.isConstantFPBuildVectorOrConstantFP(N1) && DAG.isConstantFPBuildVectorOrConstantFP(N0.getOperand(1))) { return DAG.getNode(ISD::FMA, DL, VT, N0.getOperand(0), DAG.getNode(ISD::FMUL, DL, VT, N1, N0.getOperand(1)), N2); } } // (fma x, -1, y) -> (fadd (fneg x), y) if (N1CFP) { if (N1CFP->isExactlyValue(1.0)) return DAG.getNode(ISD::FADD, DL, VT, N0, N2); if (N1CFP->isExactlyValue(-1.0) && (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) { SDValue RHSNeg = DAG.getNode(ISD::FNEG, DL, VT, N0); AddToWorklist(RHSNeg.getNode()); return DAG.getNode(ISD::FADD, DL, VT, N2, RHSNeg); } // fma (fneg x), K, y -> fma x -K, y if (N0.getOpcode() == ISD::FNEG && (TLI.isOperationLegal(ISD::ConstantFP, VT) || (N1.hasOneUse() && !TLI.isFPImmLegal(N1CFP->getValueAPF(), VT, ForCodeSize)))) { return DAG.getNode(ISD::FMA, DL, VT, N0.getOperand(0), DAG.getNode(ISD::FNEG, DL, VT, N1), N2); } } if (CanReassociate) { // (fma x, c, x) -> (fmul x, (c+1)) if (N1CFP && N0 == N2) { return DAG.getNode( ISD::FMUL, DL, VT, N0, DAG.getNode(ISD::FADD, DL, VT, N1, DAG.getConstantFP(1.0, DL, VT))); } // (fma x, c, (fneg x)) -> (fmul x, (c-1)) if (N1CFP && N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0) { return DAG.getNode( ISD::FMUL, DL, VT, N0, DAG.getNode(ISD::FADD, DL, VT, N1, DAG.getConstantFP(-1.0, DL, VT))); } } // fold ((fma (fneg X), Y, (fneg Z)) -> fneg (fma X, Y, Z)) // fold ((fma X, (fneg Y), (fneg Z)) -> fneg (fma X, Y, Z)) if (!TLI.isFNegFree(VT)) if (SDValue Neg = TLI.getCheaperNegatedExpression( SDValue(N, 0), DAG, LegalOperations, ForCodeSize)) return DAG.getNode(ISD::FNEG, DL, VT, Neg); return SDValue(); } // Combine multiple FDIVs with the same divisor into multiple FMULs by the // reciprocal. // E.g., (a / D; b / D;) -> (recip = 1.0 / D; a * recip; b * recip) // Notice that this is not always beneficial. One reason is different targets // may have different costs for FDIV and FMUL, so sometimes the cost of two // FDIVs may be lower than the cost of one FDIV and two FMULs. Another reason // is the critical path is increased from "one FDIV" to "one FDIV + one FMUL". SDValue DAGCombiner::combineRepeatedFPDivisors(SDNode *N) { // TODO: Limit this transform based on optsize/minsize - it always creates at // least 1 extra instruction. But the perf win may be substantial enough // that only minsize should restrict this. bool UnsafeMath = DAG.getTarget().Options.UnsafeFPMath; const SDNodeFlags Flags = N->getFlags(); if (LegalDAG || (!UnsafeMath && !Flags.hasAllowReciprocal())) return SDValue(); // Skip if current node is a reciprocal/fneg-reciprocal. SDValue N0 = N->getOperand(0), N1 = N->getOperand(1); ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0, /* AllowUndefs */ true); if (N0CFP && (N0CFP->isExactlyValue(1.0) || N0CFP->isExactlyValue(-1.0))) return SDValue(); // Exit early if the target does not want this transform or if there can't // possibly be enough uses of the divisor to make the transform worthwhile. unsigned MinUses = TLI.combineRepeatedFPDivisors(); // For splat vectors, scale the number of uses by the splat factor. If we can // convert the division into a scalar op, that will likely be much faster. unsigned NumElts = 1; EVT VT = N->getValueType(0); if (VT.isVector() && DAG.isSplatValue(N1)) NumElts = VT.getVectorMinNumElements(); if (!MinUses || (N1->use_size() * NumElts) < MinUses) return SDValue(); // Find all FDIV users of the same divisor. // Use a set because duplicates may be present in the user list. SetVector Users; for (auto *U : N1->uses()) { if (U->getOpcode() == ISD::FDIV && U->getOperand(1) == N1) { // Skip X/sqrt(X) that has not been simplified to sqrt(X) yet. if (U->getOperand(1).getOpcode() == ISD::FSQRT && U->getOperand(0) == U->getOperand(1).getOperand(0) && U->getFlags().hasAllowReassociation() && U->getFlags().hasNoSignedZeros()) continue; // This division is eligible for optimization only if global unsafe math // is enabled or if this division allows reciprocal formation. if (UnsafeMath || U->getFlags().hasAllowReciprocal()) Users.insert(U); } } // Now that we have the actual number of divisor uses, make sure it meets // the minimum threshold specified by the target. if ((Users.size() * NumElts) < MinUses) return SDValue(); SDLoc DL(N); SDValue FPOne = DAG.getConstantFP(1.0, DL, VT); SDValue Reciprocal = DAG.getNode(ISD::FDIV, DL, VT, FPOne, N1, Flags); // Dividend / Divisor -> Dividend * Reciprocal for (auto *U : Users) { SDValue Dividend = U->getOperand(0); if (Dividend != FPOne) { SDValue NewNode = DAG.getNode(ISD::FMUL, SDLoc(U), VT, Dividend, Reciprocal, Flags); CombineTo(U, NewNode); } else if (U != Reciprocal.getNode()) { // In the absence of fast-math-flags, this user node is always the // same node as Reciprocal, but with FMF they may be different nodes. CombineTo(U, Reciprocal); } } return SDValue(N, 0); // N was replaced. } SDValue DAGCombiner::visitFDIV(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); SDLoc DL(N); const TargetOptions &Options = DAG.getTarget().Options; SDNodeFlags Flags = N->getFlags(); SelectionDAG::FlagInserter FlagsInserter(DAG, N); if (SDValue R = DAG.simplifyFPBinop(N->getOpcode(), N0, N1, Flags)) return R; // fold (fdiv c1, c2) -> c1/c2 if (SDValue C = DAG.FoldConstantArithmetic(ISD::FDIV, DL, VT, {N0, N1})) return C; // fold vector ops if (VT.isVector()) if (SDValue FoldedVOp = SimplifyVBinOp(N, DL)) return FoldedVOp; if (SDValue NewSel = foldBinOpIntoSelect(N)) return NewSel; if (SDValue V = combineRepeatedFPDivisors(N)) return V; if (Options.UnsafeFPMath || Flags.hasAllowReciprocal()) { // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable. if (auto *N1CFP = dyn_cast(N1)) { // Compute the reciprocal 1.0 / c2. const APFloat &N1APF = N1CFP->getValueAPF(); APFloat Recip(N1APF.getSemantics(), 1); // 1.0 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven); // Only do the transform if the reciprocal is a legal fp immediate that // isn't too nasty (eg NaN, denormal, ...). if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty (!LegalOperations || // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM // backend)... we should handle this gracefully after Legalize. // TLI.isOperationLegalOrCustom(ISD::ConstantFP, VT) || TLI.isOperationLegal(ISD::ConstantFP, VT) || TLI.isFPImmLegal(Recip, VT, ForCodeSize))) return DAG.getNode(ISD::FMUL, DL, VT, N0, DAG.getConstantFP(Recip, DL, VT)); } // If this FDIV is part of a reciprocal square root, it may be folded // into a target-specific square root estimate instruction. if (N1.getOpcode() == ISD::FSQRT) { if (SDValue RV = buildRsqrtEstimate(N1.getOperand(0), Flags)) return DAG.getNode(ISD::FMUL, DL, VT, N0, RV); } else if (N1.getOpcode() == ISD::FP_EXTEND && N1.getOperand(0).getOpcode() == ISD::FSQRT) { if (SDValue RV = buildRsqrtEstimate(N1.getOperand(0).getOperand(0), Flags)) { RV = DAG.getNode(ISD::FP_EXTEND, SDLoc(N1), VT, RV); AddToWorklist(RV.getNode()); return DAG.getNode(ISD::FMUL, DL, VT, N0, RV); } } else if (N1.getOpcode() == ISD::FP_ROUND && N1.getOperand(0).getOpcode() == ISD::FSQRT) { if (SDValue RV = buildRsqrtEstimate(N1.getOperand(0).getOperand(0), Flags)) { RV = DAG.getNode(ISD::FP_ROUND, SDLoc(N1), VT, RV, N1.getOperand(1)); AddToWorklist(RV.getNode()); return DAG.getNode(ISD::FMUL, DL, VT, N0, RV); } } else if (N1.getOpcode() == ISD::FMUL) { // Look through an FMUL. Even though this won't remove the FDIV directly, // it's still worthwhile to get rid of the FSQRT if possible. SDValue Sqrt, Y; if (N1.getOperand(0).getOpcode() == ISD::FSQRT) { Sqrt = N1.getOperand(0); Y = N1.getOperand(1); } else if (N1.getOperand(1).getOpcode() == ISD::FSQRT) { Sqrt = N1.getOperand(1); Y = N1.getOperand(0); } if (Sqrt.getNode()) { // If the other multiply operand is known positive, pull it into the // sqrt. That will eliminate the division if we convert to an estimate. if (Flags.hasAllowReassociation() && N1.hasOneUse() && N1->getFlags().hasAllowReassociation() && Sqrt.hasOneUse()) { SDValue A; if (Y.getOpcode() == ISD::FABS && Y.hasOneUse()) A = Y.getOperand(0); else if (Y == Sqrt.getOperand(0)) A = Y; if (A) { // X / (fabs(A) * sqrt(Z)) --> X / sqrt(A*A*Z) --> X * rsqrt(A*A*Z) // X / (A * sqrt(A)) --> X / sqrt(A*A*A) --> X * rsqrt(A*A*A) SDValue AA = DAG.getNode(ISD::FMUL, DL, VT, A, A); SDValue AAZ = DAG.getNode(ISD::FMUL, DL, VT, AA, Sqrt.getOperand(0)); if (SDValue Rsqrt = buildRsqrtEstimate(AAZ, Flags)) return DAG.getNode(ISD::FMUL, DL, VT, N0, Rsqrt); // Estimate creation failed. Clean up speculatively created nodes. recursivelyDeleteUnusedNodes(AAZ.getNode()); } } // We found a FSQRT, so try to make this fold: // X / (Y * sqrt(Z)) -> X * (rsqrt(Z) / Y) if (SDValue Rsqrt = buildRsqrtEstimate(Sqrt.getOperand(0), Flags)) { SDValue Div = DAG.getNode(ISD::FDIV, SDLoc(N1), VT, Rsqrt, Y); AddToWorklist(Div.getNode()); return DAG.getNode(ISD::FMUL, DL, VT, N0, Div); } } } // Fold into a reciprocal estimate and multiply instead of a real divide. if (Options.NoInfsFPMath || Flags.hasNoInfs()) if (SDValue RV = BuildDivEstimate(N0, N1, Flags)) return RV; } // Fold X/Sqrt(X) -> Sqrt(X) if ((Options.NoSignedZerosFPMath || Flags.hasNoSignedZeros()) && (Options.UnsafeFPMath || Flags.hasAllowReassociation())) if (N1.getOpcode() == ISD::FSQRT && N0 == N1.getOperand(0)) return N1; // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y) TargetLowering::NegatibleCost CostN0 = TargetLowering::NegatibleCost::Expensive; TargetLowering::NegatibleCost CostN1 = TargetLowering::NegatibleCost::Expensive; SDValue NegN0 = TLI.getNegatedExpression(N0, DAG, LegalOperations, ForCodeSize, CostN0); SDValue NegN1 = TLI.getNegatedExpression(N1, DAG, LegalOperations, ForCodeSize, CostN1); if (NegN0 && NegN1 && (CostN0 == TargetLowering::NegatibleCost::Cheaper || CostN1 == TargetLowering::NegatibleCost::Cheaper)) return DAG.getNode(ISD::FDIV, SDLoc(N), VT, NegN0, NegN1); return SDValue(); } SDValue DAGCombiner::visitFREM(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); SDNodeFlags Flags = N->getFlags(); SelectionDAG::FlagInserter FlagsInserter(DAG, N); if (SDValue R = DAG.simplifyFPBinop(N->getOpcode(), N0, N1, Flags)) return R; // fold (frem c1, c2) -> fmod(c1,c2) if (SDValue C = DAG.FoldConstantArithmetic(ISD::FREM, SDLoc(N), VT, {N0, N1})) return C; if (SDValue NewSel = foldBinOpIntoSelect(N)) return NewSel; return SDValue(); } SDValue DAGCombiner::visitFSQRT(SDNode *N) { SDNodeFlags Flags = N->getFlags(); const TargetOptions &Options = DAG.getTarget().Options; // Require 'ninf' flag since sqrt(+Inf) = +Inf, but the estimation goes as: // sqrt(+Inf) == rsqrt(+Inf) * +Inf = 0 * +Inf = NaN if (!Flags.hasApproximateFuncs() || (!Options.NoInfsFPMath && !Flags.hasNoInfs())) return SDValue(); SDValue N0 = N->getOperand(0); if (TLI.isFsqrtCheap(N0, DAG)) return SDValue(); // FSQRT nodes have flags that propagate to the created nodes. // TODO: If this is N0/sqrt(N0), and we reach this node before trying to // transform the fdiv, we may produce a sub-optimal estimate sequence // because the reciprocal calculation may not have to filter out a // 0.0 input. return buildSqrtEstimate(N0, Flags); } /// copysign(x, fp_extend(y)) -> copysign(x, y) /// copysign(x, fp_round(y)) -> copysign(x, y) static inline bool CanCombineFCOPYSIGN_EXTEND_ROUND(SDNode *N) { SDValue N1 = N->getOperand(1); if ((N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)) { EVT N1VT = N1->getValueType(0); EVT N1Op0VT = N1->getOperand(0).getValueType(); // Always fold no-op FP casts. if (N1VT == N1Op0VT) return true; // Do not optimize out type conversion of f128 type yet. // For some targets like x86_64, configuration is changed to keep one f128 // value in one SSE register, but instruction selection cannot handle // FCOPYSIGN on SSE registers yet. if (N1Op0VT == MVT::f128) return false; // Avoid mismatched vector operand types, for better instruction selection. if (N1Op0VT.isVector()) return false; return true; } return false; } SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); // fold (fcopysign c1, c2) -> fcopysign(c1,c2) if (SDValue C = DAG.FoldConstantArithmetic(ISD::FCOPYSIGN, SDLoc(N), VT, {N0, N1})) return C; if (ConstantFPSDNode *N1C = isConstOrConstSplatFP(N->getOperand(1))) { const APFloat &V = N1C->getValueAPF(); // copysign(x, c1) -> fabs(x) iff ispos(c1) // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1) if (!V.isNegative()) { if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT)) return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0); } else { if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT)) return DAG.getNode(ISD::FNEG, SDLoc(N), VT, DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0)); } } // copysign(fabs(x), y) -> copysign(x, y) // copysign(fneg(x), y) -> copysign(x, y) // copysign(copysign(x,z), y) -> copysign(x, y) if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN) return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0.getOperand(0), N1); // copysign(x, abs(y)) -> abs(x) if (N1.getOpcode() == ISD::FABS) return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0); // copysign(x, copysign(y,z)) -> copysign(x, z) if (N1.getOpcode() == ISD::FCOPYSIGN) return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1.getOperand(1)); // copysign(x, fp_extend(y)) -> copysign(x, y) // copysign(x, fp_round(y)) -> copysign(x, y) if (CanCombineFCOPYSIGN_EXTEND_ROUND(N)) return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1.getOperand(0)); return SDValue(); } SDValue DAGCombiner::visitFPOW(SDNode *N) { ConstantFPSDNode *ExponentC = isConstOrConstSplatFP(N->getOperand(1)); if (!ExponentC) return SDValue(); SelectionDAG::FlagInserter FlagsInserter(DAG, N); // Try to convert x ** (1/3) into cube root. // TODO: Handle the various flavors of long double. // TODO: Since we're approximating, we don't need an exact 1/3 exponent. // Some range near 1/3 should be fine. EVT VT = N->getValueType(0); if ((VT == MVT::f32 && ExponentC->getValueAPF().isExactlyValue(1.0f/3.0f)) || (VT == MVT::f64 && ExponentC->getValueAPF().isExactlyValue(1.0/3.0))) { // pow(-0.0, 1/3) = +0.0; cbrt(-0.0) = -0.0. // pow(-inf, 1/3) = +inf; cbrt(-inf) = -inf. // pow(-val, 1/3) = nan; cbrt(-val) = -num. // For regular numbers, rounding may cause the results to differ. // Therefore, we require { nsz ninf nnan afn } for this transform. // TODO: We could select out the special cases if we don't have nsz/ninf. SDNodeFlags Flags = N->getFlags(); if (!Flags.hasNoSignedZeros() || !Flags.hasNoInfs() || !Flags.hasNoNaNs() || !Flags.hasApproximateFuncs()) return SDValue(); // Do not create a cbrt() libcall if the target does not have it, and do not // turn a pow that has lowering support into a cbrt() libcall. if (!DAG.getLibInfo().has(LibFunc_cbrt) || (!DAG.getTargetLoweringInfo().isOperationExpand(ISD::FPOW, VT) && DAG.getTargetLoweringInfo().isOperationExpand(ISD::FCBRT, VT))) return SDValue(); return DAG.getNode(ISD::FCBRT, SDLoc(N), VT, N->getOperand(0)); } // Try to convert x ** (1/4) and x ** (3/4) into square roots. // x ** (1/2) is canonicalized to sqrt, so we do not bother with that case. // TODO: This could be extended (using a target hook) to handle smaller // power-of-2 fractional exponents. bool ExponentIs025 = ExponentC->getValueAPF().isExactlyValue(0.25); bool ExponentIs075 = ExponentC->getValueAPF().isExactlyValue(0.75); if (ExponentIs025 || ExponentIs075) { // pow(-0.0, 0.25) = +0.0; sqrt(sqrt(-0.0)) = -0.0. // pow(-inf, 0.25) = +inf; sqrt(sqrt(-inf)) = NaN. // pow(-0.0, 0.75) = +0.0; sqrt(-0.0) * sqrt(sqrt(-0.0)) = +0.0. // pow(-inf, 0.75) = +inf; sqrt(-inf) * sqrt(sqrt(-inf)) = NaN. // For regular numbers, rounding may cause the results to differ. // Therefore, we require { nsz ninf afn } for this transform. // TODO: We could select out the special cases if we don't have nsz/ninf. SDNodeFlags Flags = N->getFlags(); // We only need no signed zeros for the 0.25 case. if ((!Flags.hasNoSignedZeros() && ExponentIs025) || !Flags.hasNoInfs() || !Flags.hasApproximateFuncs()) return SDValue(); // Don't double the number of libcalls. We are trying to inline fast code. if (!DAG.getTargetLoweringInfo().isOperationLegalOrCustom(ISD::FSQRT, VT)) return SDValue(); // Assume that libcalls are the smallest code. // TODO: This restriction should probably be lifted for vectors. if (ForCodeSize) return SDValue(); // pow(X, 0.25) --> sqrt(sqrt(X)) SDLoc DL(N); SDValue Sqrt = DAG.getNode(ISD::FSQRT, DL, VT, N->getOperand(0)); SDValue SqrtSqrt = DAG.getNode(ISD::FSQRT, DL, VT, Sqrt); if (ExponentIs025) return SqrtSqrt; // pow(X, 0.75) --> sqrt(X) * sqrt(sqrt(X)) return DAG.getNode(ISD::FMUL, DL, VT, Sqrt, SqrtSqrt); } return SDValue(); } static SDValue foldFPToIntToFP(SDNode *N, SelectionDAG &DAG, const TargetLowering &TLI) { // We only do this if the target has legal ftrunc. Otherwise, we'd likely be // replacing casts with a libcall. We also must be allowed to ignore -0.0 // because FTRUNC will return -0.0 for (-1.0, -0.0), but using integer // conversions would return +0.0. // FIXME: We should be able to use node-level FMF here. // TODO: If strict math, should we use FABS (+ range check for signed cast)? EVT VT = N->getValueType(0); if (!TLI.isOperationLegal(ISD::FTRUNC, VT) || !DAG.getTarget().Options.NoSignedZerosFPMath) return SDValue(); // fptosi/fptoui round towards zero, so converting from FP to integer and // back is the same as an 'ftrunc': [us]itofp (fpto[us]i X) --> ftrunc X SDValue N0 = N->getOperand(0); if (N->getOpcode() == ISD::SINT_TO_FP && N0.getOpcode() == ISD::FP_TO_SINT && N0.getOperand(0).getValueType() == VT) return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0.getOperand(0)); if (N->getOpcode() == ISD::UINT_TO_FP && N0.getOpcode() == ISD::FP_TO_UINT && N0.getOperand(0).getValueType() == VT) return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0.getOperand(0)); return SDValue(); } SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); EVT OpVT = N0.getValueType(); // [us]itofp(undef) = 0, because the result value is bounded. if (N0.isUndef()) return DAG.getConstantFP(0.0, SDLoc(N), VT); // fold (sint_to_fp c1) -> c1fp if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && // ...but only if the target supports immediate floating-point values (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::ConstantFP, VT))) return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0); // If the input is a legal type, and SINT_TO_FP is not legal on this target, // but UINT_TO_FP is legal on this target, try to convert. if (!hasOperation(ISD::SINT_TO_FP, OpVT) && hasOperation(ISD::UINT_TO_FP, OpVT)) { // If the sign bit is known to be zero, we can change this to UINT_TO_FP. if (DAG.SignBitIsZero(N0)) return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0); } // The next optimizations are desirable only if SELECT_CC can be lowered. // fold (sint_to_fp (setcc x, y, cc)) -> (select (setcc x, y, cc), -1.0, 0.0) if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 && !VT.isVector() && (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::ConstantFP, VT))) { SDLoc DL(N); return DAG.getSelect(DL, VT, N0, DAG.getConstantFP(-1.0, DL, VT), DAG.getConstantFP(0.0, DL, VT)); } // fold (sint_to_fp (zext (setcc x, y, cc))) -> // (select (setcc x, y, cc), 1.0, 0.0) if (N0.getOpcode() == ISD::ZERO_EXTEND && N0.getOperand(0).getOpcode() == ISD::SETCC && !VT.isVector() && (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::ConstantFP, VT))) { SDLoc DL(N); return DAG.getSelect(DL, VT, N0.getOperand(0), DAG.getConstantFP(1.0, DL, VT), DAG.getConstantFP(0.0, DL, VT)); } if (SDValue FTrunc = foldFPToIntToFP(N, DAG, TLI)) return FTrunc; return SDValue(); } SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); EVT OpVT = N0.getValueType(); // [us]itofp(undef) = 0, because the result value is bounded. if (N0.isUndef()) return DAG.getConstantFP(0.0, SDLoc(N), VT); // fold (uint_to_fp c1) -> c1fp if (DAG.isConstantIntBuildVectorOrConstantInt(N0) && // ...but only if the target supports immediate floating-point values (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::ConstantFP, VT))) return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0); // If the input is a legal type, and UINT_TO_FP is not legal on this target, // but SINT_TO_FP is legal on this target, try to convert. if (!hasOperation(ISD::UINT_TO_FP, OpVT) && hasOperation(ISD::SINT_TO_FP, OpVT)) { // If the sign bit is known to be zero, we can change this to SINT_TO_FP. if (DAG.SignBitIsZero(N0)) return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0); } // fold (uint_to_fp (setcc x, y, cc)) -> (select (setcc x, y, cc), 1.0, 0.0) if (N0.getOpcode() == ISD::SETCC && !VT.isVector() && (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::ConstantFP, VT))) { SDLoc DL(N); return DAG.getSelect(DL, VT, N0, DAG.getConstantFP(1.0, DL, VT), DAG.getConstantFP(0.0, DL, VT)); } if (SDValue FTrunc = foldFPToIntToFP(N, DAG, TLI)) return FTrunc; return SDValue(); } // Fold (fp_to_{s/u}int ({s/u}int_to_fpx)) -> zext x, sext x, trunc x, or x static SDValue FoldIntToFPToInt(SDNode *N, SelectionDAG &DAG) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); if (N0.getOpcode() != ISD::UINT_TO_FP && N0.getOpcode() != ISD::SINT_TO_FP) return SDValue(); SDValue Src = N0.getOperand(0); EVT SrcVT = Src.getValueType(); bool IsInputSigned = N0.getOpcode() == ISD::SINT_TO_FP; bool IsOutputSigned = N->getOpcode() == ISD::FP_TO_SINT; // We can safely assume the conversion won't overflow the output range, // because (for example) (uint8_t)18293.f is undefined behavior. // Since we can assume the conversion won't overflow, our decision as to // whether the input will fit in the float should depend on the minimum // of the input range and output range. // This means this is also safe for a signed input and unsigned output, since // a negative input would lead to undefined behavior. unsigned InputSize = (int)SrcVT.getScalarSizeInBits() - IsInputSigned; unsigned OutputSize = (int)VT.getScalarSizeInBits(); unsigned ActualSize = std::min(InputSize, OutputSize); const fltSemantics &sem = DAG.EVTToAPFloatSemantics(N0.getValueType()); // We can only fold away the float conversion if the input range can be // represented exactly in the float range. if (APFloat::semanticsPrecision(sem) >= ActualSize) { if (VT.getScalarSizeInBits() > SrcVT.getScalarSizeInBits()) { unsigned ExtOp = IsInputSigned && IsOutputSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; return DAG.getNode(ExtOp, SDLoc(N), VT, Src); } if (VT.getScalarSizeInBits() < SrcVT.getScalarSizeInBits()) return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Src); return DAG.getBitcast(VT, Src); } return SDValue(); } SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); // fold (fp_to_sint undef) -> undef if (N0.isUndef()) return DAG.getUNDEF(VT); // fold (fp_to_sint c1fp) -> c1 if (DAG.isConstantFPBuildVectorOrConstantFP(N0)) return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0); return FoldIntToFPToInt(N, DAG); } SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); // fold (fp_to_uint undef) -> undef if (N0.isUndef()) return DAG.getUNDEF(VT); // fold (fp_to_uint c1fp) -> c1 if (DAG.isConstantFPBuildVectorOrConstantFP(N0)) return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0); return FoldIntToFPToInt(N, DAG); } SDValue DAGCombiner::visitFP_ROUND(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); ConstantFPSDNode *N0CFP = dyn_cast(N0); EVT VT = N->getValueType(0); // fold (fp_round c1fp) -> c1fp if (N0CFP) return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1); // fold (fp_round (fp_extend x)) -> x if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType()) return N0.getOperand(0); // fold (fp_round (fp_round x)) -> (fp_round x) if (N0.getOpcode() == ISD::FP_ROUND) { const bool NIsTrunc = N->getConstantOperandVal(1) == 1; const bool N0IsTrunc = N0.getConstantOperandVal(1) == 1; // Skip this folding if it results in an fp_round from f80 to f16. // // f80 to f16 always generates an expensive (and as yet, unimplemented) // libcall to __truncxfhf2 instead of selecting native f16 conversion // instructions from f32 or f64. Moreover, the first (value-preserving) // fp_round from f80 to either f32 or f64 may become a NOP in platforms like // x86. if (N0.getOperand(0).getValueType() == MVT::f80 && VT == MVT::f16) return SDValue(); // If the first fp_round isn't a value preserving truncation, it might // introduce a tie in the second fp_round, that wouldn't occur in the // single-step fp_round we want to fold to. // In other words, double rounding isn't the same as rounding. // Also, this is a value preserving truncation iff both fp_round's are. if (DAG.getTarget().Options.UnsafeFPMath || N0IsTrunc) { SDLoc DL(N); return DAG.getNode(ISD::FP_ROUND, DL, VT, N0.getOperand(0), DAG.getIntPtrConstant(NIsTrunc && N0IsTrunc, DL)); } } // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y) if (N0.getOpcode() == ISD::FCOPYSIGN && N0->hasOneUse()) { SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT, N0.getOperand(0), N1); AddToWorklist(Tmp.getNode()); return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, Tmp, N0.getOperand(1)); } if (SDValue NewVSel = matchVSelectOpSizesWithSetCC(N)) return NewVSel; return SDValue(); } SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded. if (N->hasOneUse() && N->use_begin()->getOpcode() == ISD::FP_ROUND) return SDValue(); // fold (fp_extend c1fp) -> c1fp if (DAG.isConstantFPBuildVectorOrConstantFP(N0)) return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0); // fold (fp_extend (fp16_to_fp op)) -> (fp16_to_fp op) if (N0.getOpcode() == ISD::FP16_TO_FP && TLI.getOperationAction(ISD::FP16_TO_FP, VT) == TargetLowering::Legal) return DAG.getNode(ISD::FP16_TO_FP, SDLoc(N), VT, N0.getOperand(0)); // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the // value of X. if (N0.getOpcode() == ISD::FP_ROUND && N0.getConstantOperandVal(1) == 1) { SDValue In = N0.getOperand(0); if (In.getValueType() == VT) return In; if (VT.bitsLT(In.getValueType())) return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, In, N0.getOperand(1)); return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In); } // fold (fpext (load x)) -> (fpext (fptrunc (extload x))) if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() && TLI.isLoadExtLegalOrCustom(ISD::EXTLOAD, VT, N0.getValueType())) { LoadSDNode *LN0 = cast(N0); SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT, LN0->getChain(), LN0->getBasePtr(), N0.getValueType(), LN0->getMemOperand()); CombineTo(N, ExtLoad); CombineTo(N0.getNode(), DAG.getNode(ISD::FP_ROUND, SDLoc(N0), N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1, SDLoc(N0))), ExtLoad.getValue(1)); return SDValue(N, 0); // Return N so it doesn't get rechecked! } if (SDValue NewVSel = matchVSelectOpSizesWithSetCC(N)) return NewVSel; return SDValue(); } SDValue DAGCombiner::visitFCEIL(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); // fold (fceil c1) -> fceil(c1) if (DAG.isConstantFPBuildVectorOrConstantFP(N0)) return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0); return SDValue(); } SDValue DAGCombiner::visitFTRUNC(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); // fold (ftrunc c1) -> ftrunc(c1) if (DAG.isConstantFPBuildVectorOrConstantFP(N0)) return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0); // fold ftrunc (known rounded int x) -> x // ftrunc is a part of fptosi/fptoui expansion on some targets, so this is // likely to be generated to extract integer from a rounded floating value. switch (N0.getOpcode()) { default: break; case ISD::FRINT: case ISD::FTRUNC: case ISD::FNEARBYINT: case ISD::FFLOOR: case ISD::FCEIL: return N0; } return SDValue(); } SDValue DAGCombiner::visitFFLOOR(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); // fold (ffloor c1) -> ffloor(c1) if (DAG.isConstantFPBuildVectorOrConstantFP(N0)) return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0); return SDValue(); } SDValue DAGCombiner::visitFNEG(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); SelectionDAG::FlagInserter FlagsInserter(DAG, N); // Constant fold FNEG. if (DAG.isConstantFPBuildVectorOrConstantFP(N0)) return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0); if (SDValue NegN0 = TLI.getNegatedExpression(N0, DAG, LegalOperations, ForCodeSize)) return NegN0; // -(X-Y) -> (Y-X) is unsafe because when X==Y, -0.0 != +0.0 // FIXME: This is duplicated in getNegatibleCost, but getNegatibleCost doesn't // know it was called from a context with a nsz flag if the input fsub does // not. if (N0.getOpcode() == ISD::FSUB && (DAG.getTarget().Options.NoSignedZerosFPMath || N->getFlags().hasNoSignedZeros()) && N0.hasOneUse()) { return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0.getOperand(1), N0.getOperand(0)); } if (SDValue Cast = foldSignChangeInBitcast(N)) return Cast; return SDValue(); } SDValue DAGCombiner::visitFMinMax(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); const SDNodeFlags Flags = N->getFlags(); unsigned Opc = N->getOpcode(); bool PropagatesNaN = Opc == ISD::FMINIMUM || Opc == ISD::FMAXIMUM; bool IsMin = Opc == ISD::FMINNUM || Opc == ISD::FMINIMUM; SelectionDAG::FlagInserter FlagsInserter(DAG, N); // Constant fold. if (SDValue C = DAG.FoldConstantArithmetic(Opc, SDLoc(N), VT, {N0, N1})) return C; // Canonicalize to constant on RHS. if (DAG.isConstantFPBuildVectorOrConstantFP(N0) && !DAG.isConstantFPBuildVectorOrConstantFP(N1)) return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N1, N0); if (const ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1)) { const APFloat &AF = N1CFP->getValueAPF(); // minnum(X, nan) -> X // maxnum(X, nan) -> X // minimum(X, nan) -> nan // maximum(X, nan) -> nan if (AF.isNaN()) return PropagatesNaN ? N->getOperand(1) : N->getOperand(0); // In the following folds, inf can be replaced with the largest finite // float, if the ninf flag is set. if (AF.isInfinity() || (Flags.hasNoInfs() && AF.isLargest())) { // minnum(X, -inf) -> -inf // maxnum(X, +inf) -> +inf // minimum(X, -inf) -> -inf if nnan // maximum(X, +inf) -> +inf if nnan if (IsMin == AF.isNegative() && (!PropagatesNaN || Flags.hasNoNaNs())) return N->getOperand(1); // minnum(X, +inf) -> X if nnan // maxnum(X, -inf) -> X if nnan // minimum(X, +inf) -> X // maximum(X, -inf) -> X if (IsMin != AF.isNegative() && (PropagatesNaN || Flags.hasNoNaNs())) return N->getOperand(0); } } return SDValue(); } SDValue DAGCombiner::visitFABS(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); // fold (fabs c1) -> fabs(c1) if (DAG.isConstantFPBuildVectorOrConstantFP(N0)) return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0); // fold (fabs (fabs x)) -> (fabs x) if (N0.getOpcode() == ISD::FABS) return N->getOperand(0); // fold (fabs (fneg x)) -> (fabs x) // fold (fabs (fcopysign x, y)) -> (fabs x) if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN) return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0)); if (SDValue Cast = foldSignChangeInBitcast(N)) return Cast; return SDValue(); } SDValue DAGCombiner::visitBRCOND(SDNode *N) { SDValue Chain = N->getOperand(0); SDValue N1 = N->getOperand(1); SDValue N2 = N->getOperand(2); // BRCOND(FREEZE(cond)) is equivalent to BRCOND(cond) (both are // nondeterministic jumps). if (N1->getOpcode() == ISD::FREEZE && N1.hasOneUse()) { return DAG.getNode(ISD::BRCOND, SDLoc(N), MVT::Other, Chain, N1->getOperand(0), N2); } // If N is a constant we could fold this into a fallthrough or unconditional // branch. However that doesn't happen very often in normal code, because // Instcombine/SimplifyCFG should have handled the available opportunities. // If we did this folding here, it would be necessary to update the // MachineBasicBlock CFG, which is awkward. // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal // on the target. if (N1.getOpcode() == ISD::SETCC && TLI.isOperationLegalOrCustom(ISD::BR_CC, N1.getOperand(0).getValueType())) { return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other, Chain, N1.getOperand(2), N1.getOperand(0), N1.getOperand(1), N2); } if (N1.hasOneUse()) { // rebuildSetCC calls visitXor which may change the Chain when there is a // STRICT_FSETCC/STRICT_FSETCCS involved. Use a handle to track changes. HandleSDNode ChainHandle(Chain); if (SDValue NewN1 = rebuildSetCC(N1)) return DAG.getNode(ISD::BRCOND, SDLoc(N), MVT::Other, ChainHandle.getValue(), NewN1, N2); } return SDValue(); } SDValue DAGCombiner::rebuildSetCC(SDValue N) { if (N.getOpcode() == ISD::SRL || (N.getOpcode() == ISD::TRUNCATE && (N.getOperand(0).hasOneUse() && N.getOperand(0).getOpcode() == ISD::SRL))) { // Look pass the truncate. if (N.getOpcode() == ISD::TRUNCATE) N = N.getOperand(0); // Match this pattern so that we can generate simpler code: // // %a = ... // %b = and i32 %a, 2 // %c = srl i32 %b, 1 // brcond i32 %c ... // // into // // %a = ... // %b = and i32 %a, 2 // %c = setcc eq %b, 0 // brcond %c ... // // This applies only when the AND constant value has one bit set and the // SRL constant is equal to the log2 of the AND constant. The back-end is // smart enough to convert the result into a TEST/JMP sequence. SDValue Op0 = N.getOperand(0); SDValue Op1 = N.getOperand(1); if (Op0.getOpcode() == ISD::AND && Op1.getOpcode() == ISD::Constant) { SDValue AndOp1 = Op0.getOperand(1); if (AndOp1.getOpcode() == ISD::Constant) { const APInt &AndConst = cast(AndOp1)->getAPIntValue(); if (AndConst.isPowerOf2() && cast(Op1)->getAPIntValue() == AndConst.logBase2()) { SDLoc DL(N); return DAG.getSetCC(DL, getSetCCResultType(Op0.getValueType()), Op0, DAG.getConstant(0, DL, Op0.getValueType()), ISD::SETNE); } } } } // Transform (brcond (xor x, y)) -> (brcond (setcc, x, y, ne)) // Transform (brcond (xor (xor x, y), -1)) -> (brcond (setcc, x, y, eq)) if (N.getOpcode() == ISD::XOR) { // Because we may call this on a speculatively constructed // SimplifiedSetCC Node, we need to simplify this node first. // Ideally this should be folded into SimplifySetCC and not // here. For now, grab a handle to N so we don't lose it from // replacements interal to the visit. HandleSDNode XORHandle(N); while (N.getOpcode() == ISD::XOR) { SDValue Tmp = visitXOR(N.getNode()); // No simplification done. if (!Tmp.getNode()) break; // Returning N is form in-visit replacement that may invalidated // N. Grab value from Handle. if (Tmp.getNode() == N.getNode()) N = XORHandle.getValue(); else // Node simplified. Try simplifying again. N = Tmp; } if (N.getOpcode() != ISD::XOR) return N; SDValue Op0 = N->getOperand(0); SDValue Op1 = N->getOperand(1); if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) { bool Equal = false; // (brcond (xor (xor x, y), -1)) -> (brcond (setcc x, y, eq)) if (isBitwiseNot(N) && Op0.hasOneUse() && Op0.getOpcode() == ISD::XOR && Op0.getValueType() == MVT::i1) { N = Op0; Op0 = N->getOperand(0); Op1 = N->getOperand(1); Equal = true; } EVT SetCCVT = N.getValueType(); if (LegalTypes) SetCCVT = getSetCCResultType(SetCCVT); // Replace the uses of XOR with SETCC return DAG.getSetCC(SDLoc(N), SetCCVT, Op0, Op1, Equal ? ISD::SETEQ : ISD::SETNE); } } return SDValue(); } // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB. // SDValue DAGCombiner::visitBR_CC(SDNode *N) { CondCodeSDNode *CC = cast(N->getOperand(1)); SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3); // If N is a constant we could fold this into a fallthrough or unconditional // branch. However that doesn't happen very often in normal code, because // Instcombine/SimplifyCFG should have handled the available opportunities. // If we did this folding here, it would be necessary to update the // MachineBasicBlock CFG, which is awkward. // Use SimplifySetCC to simplify SETCC's. SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()), CondLHS, CondRHS, CC->get(), SDLoc(N), false); if (Simp.getNode()) AddToWorklist(Simp.getNode()); // fold to a simpler setcc if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC) return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other, N->getOperand(0), Simp.getOperand(2), Simp.getOperand(0), Simp.getOperand(1), N->getOperand(4)); return SDValue(); } static bool getCombineLoadStoreParts(SDNode *N, unsigned Inc, unsigned Dec, bool &IsLoad, bool &IsMasked, SDValue &Ptr, const TargetLowering &TLI) { if (LoadSDNode *LD = dyn_cast(N)) { if (LD->isIndexed()) return false; EVT VT = LD->getMemoryVT(); if (!TLI.isIndexedLoadLegal(Inc, VT) && !TLI.isIndexedLoadLegal(Dec, VT)) return false; Ptr = LD->getBasePtr(); } else if (StoreSDNode *ST = dyn_cast(N)) { if (ST->isIndexed()) return false; EVT VT = ST->getMemoryVT(); if (!TLI.isIndexedStoreLegal(Inc, VT) && !TLI.isIndexedStoreLegal(Dec, VT)) return false; Ptr = ST->getBasePtr(); IsLoad = false; } else if (MaskedLoadSDNode *LD = dyn_cast(N)) { if (LD->isIndexed()) return false; EVT VT = LD->getMemoryVT(); if (!TLI.isIndexedMaskedLoadLegal(Inc, VT) && !TLI.isIndexedMaskedLoadLegal(Dec, VT)) return false; Ptr = LD->getBasePtr(); IsMasked = true; } else if (MaskedStoreSDNode *ST = dyn_cast(N)) { if (ST->isIndexed()) return false; EVT VT = ST->getMemoryVT(); if (!TLI.isIndexedMaskedStoreLegal(Inc, VT) && !TLI.isIndexedMaskedStoreLegal(Dec, VT)) return false; Ptr = ST->getBasePtr(); IsLoad = false; IsMasked = true; } else { return false; } return true; } /// Try turning a load/store into a pre-indexed load/store when the base /// pointer is an add or subtract and it has other uses besides the load/store. /// After the transformation, the new indexed load/store has effectively folded /// the add/subtract in and all of its other uses are redirected to the /// new load/store. bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) { if (Level < AfterLegalizeDAG) return false; bool IsLoad = true; bool IsMasked = false; SDValue Ptr; if (!getCombineLoadStoreParts(N, ISD::PRE_INC, ISD::PRE_DEC, IsLoad, IsMasked, Ptr, TLI)) return false; // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail // out. There is no reason to make this a preinc/predec. if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) || Ptr->hasOneUse()) return false; // Ask the target to do addressing mode selection. SDValue BasePtr; SDValue Offset; ISD::MemIndexedMode AM = ISD::UNINDEXED; if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG)) return false; // Backends without true r+i pre-indexed forms may need to pass a // constant base with a variable offset so that constant coercion // will work with the patterns in canonical form. bool Swapped = false; if (isa(BasePtr)) { std::swap(BasePtr, Offset); Swapped = true; } // Don't create a indexed load / store with zero offset. if (isNullConstant(Offset)) return false; // Try turning it into a pre-indexed load / store except when: // 1) The new base ptr is a frame index. // 2) If N is a store and the new base ptr is either the same as or is a // predecessor of the value being stored. // 3) Another use of old base ptr is a predecessor of N. If ptr is folded // that would create a cycle. // 4) All uses are load / store ops that use it as old base ptr. // Check #1. Preinc'ing a frame index would require copying the stack pointer // (plus the implicit offset) to a register to preinc anyway. if (isa(BasePtr) || isa(BasePtr)) return false; // Check #2. if (!IsLoad) { SDValue Val = IsMasked ? cast(N)->getValue() : cast(N)->getValue(); // Would require a copy. if (Val == BasePtr) return false; // Would create a cycle. if (Val == Ptr || Ptr->isPredecessorOf(Val.getNode())) return false; } // Caches for hasPredecessorHelper. SmallPtrSet Visited; SmallVector Worklist; Worklist.push_back(N); // If the offset is a constant, there may be other adds of constants that // can be folded with this one. We should do this to avoid having to keep // a copy of the original base pointer. SmallVector OtherUses; if (isa(Offset)) for (SDNode::use_iterator UI = BasePtr->use_begin(), UE = BasePtr->use_end(); UI != UE; ++UI) { SDUse &Use = UI.getUse(); // Skip the use that is Ptr and uses of other results from BasePtr's // node (important for nodes that return multiple results). if (Use.getUser() == Ptr.getNode() || Use != BasePtr) continue; if (SDNode::hasPredecessorHelper(Use.getUser(), Visited, Worklist)) continue; if (Use.getUser()->getOpcode() != ISD::ADD && Use.getUser()->getOpcode() != ISD::SUB) { OtherUses.clear(); break; } SDValue Op1 = Use.getUser()->getOperand((UI.getOperandNo() + 1) & 1); if (!isa(Op1)) { OtherUses.clear(); break; } // FIXME: In some cases, we can be smarter about this. if (Op1.getValueType() != Offset.getValueType()) { OtherUses.clear(); break; } OtherUses.push_back(Use.getUser()); } if (Swapped) std::swap(BasePtr, Offset); // Now check for #3 and #4. bool RealUse = false; for (SDNode *Use : Ptr->uses()) { if (Use == N) continue; if (SDNode::hasPredecessorHelper(Use, Visited, Worklist)) return false; // If Ptr may be folded in addressing mode of other use, then it's // not profitable to do this transformation. if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI)) RealUse = true; } if (!RealUse) return false; SDValue Result; if (!IsMasked) { if (IsLoad) Result = DAG.getIndexedLoad(SDValue(N, 0), SDLoc(N), BasePtr, Offset, AM); else Result = DAG.getIndexedStore(SDValue(N, 0), SDLoc(N), BasePtr, Offset, AM); } else { if (IsLoad) Result = DAG.getIndexedMaskedLoad(SDValue(N, 0), SDLoc(N), BasePtr, Offset, AM); else Result = DAG.getIndexedMaskedStore(SDValue(N, 0), SDLoc(N), BasePtr, Offset, AM); } ++PreIndexedNodes; ++NodesCombined; LLVM_DEBUG(dbgs() << "\nReplacing.4 "; N->dump(&DAG); dbgs() << "\nWith: "; Result.dump(&DAG); dbgs() << '\n'); WorklistRemover DeadNodes(*this); if (IsLoad) { DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0)); DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2)); } else { DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1)); } // Finally, since the node is now dead, remove it from the graph. deleteAndRecombine(N); if (Swapped) std::swap(BasePtr, Offset); // Replace other uses of BasePtr that can be updated to use Ptr for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) { unsigned OffsetIdx = 1; if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode()) OffsetIdx = 0; assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() == BasePtr.getNode() && "Expected BasePtr operand"); // We need to replace ptr0 in the following expression: // x0 * offset0 + y0 * ptr0 = t0 // knowing that // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store) // // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the // indexed load/store and the expression that needs to be re-written. // // Therefore, we have: // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1 auto *CN = cast(OtherUses[i]->getOperand(OffsetIdx)); const APInt &Offset0 = CN->getAPIntValue(); const APInt &Offset1 = cast(Offset)->getAPIntValue(); int X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1; int Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1; int X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1; int Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1; unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD; APInt CNV = Offset0; if (X0 < 0) CNV = -CNV; if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1; else CNV = CNV - Offset1; SDLoc DL(OtherUses[i]); // We can now generate the new expression. SDValue NewOp1 = DAG.getConstant(CNV, DL, CN->getValueType(0)); SDValue NewOp2 = Result.getValue(IsLoad ? 1 : 0); SDValue NewUse = DAG.getNode(Opcode, DL, OtherUses[i]->getValueType(0), NewOp1, NewOp2); DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse); deleteAndRecombine(OtherUses[i]); } // Replace the uses of Ptr with uses of the updated base value. DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(IsLoad ? 1 : 0)); deleteAndRecombine(Ptr.getNode()); AddToWorklist(Result.getNode()); return true; } static bool shouldCombineToPostInc(SDNode *N, SDValue Ptr, SDNode *PtrUse, SDValue &BasePtr, SDValue &Offset, ISD::MemIndexedMode &AM, SelectionDAG &DAG, const TargetLowering &TLI) { if (PtrUse == N || (PtrUse->getOpcode() != ISD::ADD && PtrUse->getOpcode() != ISD::SUB)) return false; if (!TLI.getPostIndexedAddressParts(N, PtrUse, BasePtr, Offset, AM, DAG)) return false; // Don't create a indexed load / store with zero offset. if (isNullConstant(Offset)) return false; if (isa(BasePtr) || isa(BasePtr)) return false; SmallPtrSet Visited; for (SDNode *Use : BasePtr->uses()) { if (Use == Ptr.getNode()) continue; // No if there's a later user which could perform the index instead. if (isa(Use)) { bool IsLoad = true; bool IsMasked = false; SDValue OtherPtr; if (getCombineLoadStoreParts(Use, ISD::POST_INC, ISD::POST_DEC, IsLoad, IsMasked, OtherPtr, TLI)) { SmallVector Worklist; Worklist.push_back(Use); if (SDNode::hasPredecessorHelper(N, Visited, Worklist)) return false; } } // If all the uses are load / store addresses, then don't do the // transformation. if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB) { for (SDNode *UseUse : Use->uses()) if (canFoldInAddressingMode(Use, UseUse, DAG, TLI)) return false; } } return true; } static SDNode *getPostIndexedLoadStoreOp(SDNode *N, bool &IsLoad, bool &IsMasked, SDValue &Ptr, SDValue &BasePtr, SDValue &Offset, ISD::MemIndexedMode &AM, SelectionDAG &DAG, const TargetLowering &TLI) { if (!getCombineLoadStoreParts(N, ISD::POST_INC, ISD::POST_DEC, IsLoad, IsMasked, Ptr, TLI) || Ptr->hasOneUse()) return nullptr; // Try turning it into a post-indexed load / store except when // 1) All uses are load / store ops that use it as base ptr (and // it may be folded as addressing mmode). // 2) Op must be independent of N, i.e. Op is neither a predecessor // nor a successor of N. Otherwise, if Op is folded that would // create a cycle. for (SDNode *Op : Ptr->uses()) { // Check for #1. if (!shouldCombineToPostInc(N, Ptr, Op, BasePtr, Offset, AM, DAG, TLI)) continue; // Check for #2. SmallPtrSet Visited; SmallVector Worklist; // Ptr is predecessor to both N and Op. Visited.insert(Ptr.getNode()); Worklist.push_back(N); Worklist.push_back(Op); if (!SDNode::hasPredecessorHelper(N, Visited, Worklist) && !SDNode::hasPredecessorHelper(Op, Visited, Worklist)) return Op; } return nullptr; } /// Try to combine a load/store with a add/sub of the base pointer node into a /// post-indexed load/store. The transformation folded the add/subtract into the /// new indexed load/store effectively and all of its uses are redirected to the /// new load/store. bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) { if (Level < AfterLegalizeDAG) return false; bool IsLoad = true; bool IsMasked = false; SDValue Ptr; SDValue BasePtr; SDValue Offset; ISD::MemIndexedMode AM = ISD::UNINDEXED; SDNode *Op = getPostIndexedLoadStoreOp(N, IsLoad, IsMasked, Ptr, BasePtr, Offset, AM, DAG, TLI); if (!Op) return false; SDValue Result; if (!IsMasked) Result = IsLoad ? DAG.getIndexedLoad(SDValue(N, 0), SDLoc(N), BasePtr, Offset, AM) : DAG.getIndexedStore(SDValue(N, 0), SDLoc(N), BasePtr, Offset, AM); else Result = IsLoad ? DAG.getIndexedMaskedLoad(SDValue(N, 0), SDLoc(N), BasePtr, Offset, AM) : DAG.getIndexedMaskedStore(SDValue(N, 0), SDLoc(N), BasePtr, Offset, AM); ++PostIndexedNodes; ++NodesCombined; LLVM_DEBUG(dbgs() << "\nReplacing.5 "; N->dump(&DAG); dbgs() << "\nWith: "; Result.dump(&DAG); dbgs() << '\n'); WorklistRemover DeadNodes(*this); if (IsLoad) { DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0)); DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2)); } else { DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1)); } // Finally, since the node is now dead, remove it from the graph. deleteAndRecombine(N); // Replace the uses of Use with uses of the updated base value. DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0), Result.getValue(IsLoad ? 1 : 0)); deleteAndRecombine(Op); return true; } /// Return the base-pointer arithmetic from an indexed \p LD. SDValue DAGCombiner::SplitIndexingFromLoad(LoadSDNode *LD) { ISD::MemIndexedMode AM = LD->getAddressingMode(); assert(AM != ISD::UNINDEXED); SDValue BP = LD->getOperand(1); SDValue Inc = LD->getOperand(2); // Some backends use TargetConstants for load offsets, but don't expect // TargetConstants in general ADD nodes. We can convert these constants into // regular Constants (if the constant is not opaque). assert((Inc.getOpcode() != ISD::TargetConstant || !cast(Inc)->isOpaque()) && "Cannot split out indexing using opaque target constants"); if (Inc.getOpcode() == ISD::TargetConstant) { ConstantSDNode *ConstInc = cast(Inc); Inc = DAG.getConstant(*ConstInc->getConstantIntValue(), SDLoc(Inc), ConstInc->getValueType(0)); } unsigned Opc = (AM == ISD::PRE_INC || AM == ISD::POST_INC ? ISD::ADD : ISD::SUB); return DAG.getNode(Opc, SDLoc(LD), BP.getSimpleValueType(), BP, Inc); } static inline ElementCount numVectorEltsOrZero(EVT T) { return T.isVector() ? T.getVectorElementCount() : ElementCount::getFixed(0); } bool DAGCombiner::getTruncatedStoreValue(StoreSDNode *ST, SDValue &Val) { Val = ST->getValue(); EVT STType = Val.getValueType(); EVT STMemType = ST->getMemoryVT(); if (STType == STMemType) return true; if (isTypeLegal(STMemType)) return false; // fail. if (STType.isFloatingPoint() && STMemType.isFloatingPoint() && TLI.isOperationLegal(ISD::FTRUNC, STMemType)) { Val = DAG.getNode(ISD::FTRUNC, SDLoc(ST), STMemType, Val); return true; } if (numVectorEltsOrZero(STType) == numVectorEltsOrZero(STMemType) && STType.isInteger() && STMemType.isInteger()) { Val = DAG.getNode(ISD::TRUNCATE, SDLoc(ST), STMemType, Val); return true; } if (STType.getSizeInBits() == STMemType.getSizeInBits()) { Val = DAG.getBitcast(STMemType, Val); return true; } return false; // fail. } bool DAGCombiner::extendLoadedValueToExtension(LoadSDNode *LD, SDValue &Val) { EVT LDMemType = LD->getMemoryVT(); EVT LDType = LD->getValueType(0); assert(Val.getValueType() == LDMemType && "Attempting to extend value of non-matching type"); if (LDType == LDMemType) return true; if (LDMemType.isInteger() && LDType.isInteger()) { switch (LD->getExtensionType()) { case ISD::NON_EXTLOAD: Val = DAG.getBitcast(LDType, Val); return true; case ISD::EXTLOAD: Val = DAG.getNode(ISD::ANY_EXTEND, SDLoc(LD), LDType, Val); return true; case ISD::SEXTLOAD: Val = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(LD), LDType, Val); return true; case ISD::ZEXTLOAD: Val = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(LD), LDType, Val); return true; } } return false; } SDValue DAGCombiner::ForwardStoreValueToDirectLoad(LoadSDNode *LD) { if (OptLevel == CodeGenOpt::None || !LD->isSimple()) return SDValue(); SDValue Chain = LD->getOperand(0); StoreSDNode *ST = dyn_cast(Chain.getNode()); // TODO: Relax this restriction for unordered atomics (see D66309) if (!ST || !ST->isSimple()) return SDValue(); EVT LDType = LD->getValueType(0); EVT LDMemType = LD->getMemoryVT(); EVT STMemType = ST->getMemoryVT(); EVT STType = ST->getValue().getValueType(); // There are two cases to consider here: // 1. The store is fixed width and the load is scalable. In this case we // don't know at compile time if the store completely envelops the load // so we abandon the optimisation. // 2. The store is scalable and the load is fixed width. We could // potentially support a limited number of cases here, but there has been // no cost-benefit analysis to prove it's worth it. bool LdStScalable = LDMemType.isScalableVector(); if (LdStScalable != STMemType.isScalableVector()) return SDValue(); // If we are dealing with scalable vectors on a big endian platform the // calculation of offsets below becomes trickier, since we do not know at // compile time the absolute size of the vector. Until we've done more // analysis on big-endian platforms it seems better to bail out for now. if (LdStScalable && DAG.getDataLayout().isBigEndian()) return SDValue(); BaseIndexOffset BasePtrLD = BaseIndexOffset::match(LD, DAG); BaseIndexOffset BasePtrST = BaseIndexOffset::match(ST, DAG); int64_t Offset; if (!BasePtrST.equalBaseIndex(BasePtrLD, DAG, Offset)) return SDValue(); // Normalize for Endianness. After this Offset=0 will denote that the least // significant bit in the loaded value maps to the least significant bit in // the stored value). With Offset=n (for n > 0) the loaded value starts at the // n:th least significant byte of the stored value. if (DAG.getDataLayout().isBigEndian()) Offset = ((int64_t)STMemType.getStoreSizeInBits().getFixedSize() - (int64_t)LDMemType.getStoreSizeInBits().getFixedSize()) / 8 - Offset; // Check that the stored value cover all bits that are loaded. bool STCoversLD; TypeSize LdMemSize = LDMemType.getSizeInBits(); TypeSize StMemSize = STMemType.getSizeInBits(); if (LdStScalable) STCoversLD = (Offset == 0) && LdMemSize == StMemSize; else STCoversLD = (Offset >= 0) && (Offset * 8 + LdMemSize.getFixedSize() <= StMemSize.getFixedSize()); auto ReplaceLd = [&](LoadSDNode *LD, SDValue Val, SDValue Chain) -> SDValue { if (LD->isIndexed()) { // Cannot handle opaque target constants and we must respect the user's // request not to split indexes from loads. if (!canSplitIdx(LD)) return SDValue(); SDValue Idx = SplitIndexingFromLoad(LD); SDValue Ops[] = {Val, Idx, Chain}; return CombineTo(LD, Ops, 3); } return CombineTo(LD, Val, Chain); }; if (!STCoversLD) return SDValue(); // Memory as copy space (potentially masked). if (Offset == 0 && LDType == STType && STMemType == LDMemType) { // Simple case: Direct non-truncating forwarding if (LDType.getSizeInBits() == LdMemSize) return ReplaceLd(LD, ST->getValue(), Chain); // Can we model the truncate and extension with an and mask? if (STType.isInteger() && LDMemType.isInteger() && !STType.isVector() && !LDMemType.isVector() && LD->getExtensionType() != ISD::SEXTLOAD) { // Mask to size of LDMemType auto Mask = DAG.getConstant(APInt::getLowBitsSet(STType.getFixedSizeInBits(), StMemSize.getFixedSize()), SDLoc(ST), STType); auto Val = DAG.getNode(ISD::AND, SDLoc(LD), LDType, ST->getValue(), Mask); return ReplaceLd(LD, Val, Chain); } } // TODO: Deal with nonzero offset. if (LD->getBasePtr().isUndef() || Offset != 0) return SDValue(); // Model necessary truncations / extenstions. SDValue Val; // Truncate Value To Stored Memory Size. do { if (!getTruncatedStoreValue(ST, Val)) continue; if (!isTypeLegal(LDMemType)) continue; if (STMemType != LDMemType) { // TODO: Support vectors? This requires extract_subvector/bitcast. if (!STMemType.isVector() && !LDMemType.isVector() && STMemType.isInteger() && LDMemType.isInteger()) Val = DAG.getNode(ISD::TRUNCATE, SDLoc(LD), LDMemType, Val); else continue; } if (!extendLoadedValueToExtension(LD, Val)) continue; return ReplaceLd(LD, Val, Chain); } while (false); // On failure, cleanup dead nodes we may have created. if (Val->use_empty()) deleteAndRecombine(Val.getNode()); return SDValue(); } SDValue DAGCombiner::visitLOAD(SDNode *N) { LoadSDNode *LD = cast(N); SDValue Chain = LD->getChain(); SDValue Ptr = LD->getBasePtr(); // If load is not volatile and there are no uses of the loaded value (and // the updated indexed value in case of indexed loads), change uses of the // chain value into uses of the chain input (i.e. delete the dead load). // TODO: Allow this for unordered atomics (see D66309) if (LD->isSimple()) { if (N->getValueType(1) == MVT::Other) { // Unindexed loads. if (!N->hasAnyUseOfValue(0)) { // It's not safe to use the two value CombineTo variant here. e.g. // v1, chain2 = load chain1, loc // v2, chain3 = load chain2, loc // v3 = add v2, c // Now we replace use of chain2 with chain1. This makes the second load // isomorphic to the one we are deleting, and thus makes this load live. LLVM_DEBUG(dbgs() << "\nReplacing.6 "; N->dump(&DAG); dbgs() << "\nWith chain: "; Chain.dump(&DAG); dbgs() << "\n"); WorklistRemover DeadNodes(*this); DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain); AddUsersToWorklist(Chain.getNode()); if (N->use_empty()) deleteAndRecombine(N); return SDValue(N, 0); // Return N so it doesn't get rechecked! } } else { // Indexed loads. assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?"); // If this load has an opaque TargetConstant offset, then we cannot split // the indexing into an add/sub directly (that TargetConstant may not be // valid for a different type of node, and we cannot convert an opaque // target constant into a regular constant). bool CanSplitIdx = canSplitIdx(LD); if (!N->hasAnyUseOfValue(0) && (CanSplitIdx || !N->hasAnyUseOfValue(1))) { SDValue Undef = DAG.getUNDEF(N->getValueType(0)); SDValue Index; if (N->hasAnyUseOfValue(1) && CanSplitIdx) { Index = SplitIndexingFromLoad(LD); // Try to fold the base pointer arithmetic into subsequent loads and // stores. AddUsersToWorklist(N); } else Index = DAG.getUNDEF(N->getValueType(1)); LLVM_DEBUG(dbgs() << "\nReplacing.7 "; N->dump(&DAG); dbgs() << "\nWith: "; Undef.dump(&DAG); dbgs() << " and 2 other values\n"); WorklistRemover DeadNodes(*this); DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef); DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Index); DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain); deleteAndRecombine(N); return SDValue(N, 0); // Return N so it doesn't get rechecked! } } } // If this load is directly stored, replace the load value with the stored // value. if (auto V = ForwardStoreValueToDirectLoad(LD)) return V; // Try to infer better alignment information than the load already has. if (OptLevel != CodeGenOpt::None && LD->isUnindexed() && !LD->isAtomic()) { if (MaybeAlign Alignment = DAG.InferPtrAlign(Ptr)) { if (*Alignment > LD->getAlign() && isAligned(*Alignment, LD->getSrcValueOffset())) { SDValue NewLoad = DAG.getExtLoad( LD->getExtensionType(), SDLoc(N), LD->getValueType(0), Chain, Ptr, LD->getPointerInfo(), LD->getMemoryVT(), *Alignment, LD->getMemOperand()->getFlags(), LD->getAAInfo()); // NewLoad will always be N as we are only refining the alignment assert(NewLoad.getNode() == N); (void)NewLoad; } } } if (LD->isUnindexed()) { // Walk up chain skipping non-aliasing memory nodes. SDValue BetterChain = FindBetterChain(LD, Chain); // If there is a better chain. if (Chain != BetterChain) { SDValue ReplLoad; // Replace the chain to void dependency. if (LD->getExtensionType() == ISD::NON_EXTLOAD) { ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD), BetterChain, Ptr, LD->getMemOperand()); } else { ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), LD->getValueType(0), BetterChain, Ptr, LD->getMemoryVT(), LD->getMemOperand()); } // Create token factor to keep old chain connected. SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Chain, ReplLoad.getValue(1)); // Replace uses with load result and token factor return CombineTo(N, ReplLoad.getValue(0), Token); } } // Try transforming N to an indexed load. if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N)) return SDValue(N, 0); // Try to slice up N to more direct loads if the slices are mapped to // different register banks or pairing can take place. if (SliceUpLoad(N)) return SDValue(N, 0); return SDValue(); } namespace { /// Helper structure used to slice a load in smaller loads. /// Basically a slice is obtained from the following sequence: /// Origin = load Ty1, Base /// Shift = srl Ty1 Origin, CstTy Amount /// Inst = trunc Shift to Ty2 /// /// Then, it will be rewritten into: /// Slice = load SliceTy, Base + SliceOffset /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2 /// /// SliceTy is deduced from the number of bits that are actually used to /// build Inst. struct LoadedSlice { /// Helper structure used to compute the cost of a slice. struct Cost { /// Are we optimizing for code size. bool ForCodeSize = false; /// Various cost. unsigned Loads = 0; unsigned Truncates = 0; unsigned CrossRegisterBanksCopies = 0; unsigned ZExts = 0; unsigned Shift = 0; explicit Cost(bool ForCodeSize) : ForCodeSize(ForCodeSize) {} /// Get the cost of one isolated slice. Cost(const LoadedSlice &LS, bool ForCodeSize) : ForCodeSize(ForCodeSize), Loads(1) { EVT TruncType = LS.Inst->getValueType(0); EVT LoadedType = LS.getLoadedType(); if (TruncType != LoadedType && !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType)) ZExts = 1; } /// Account for slicing gain in the current cost. /// Slicing provide a few gains like removing a shift or a /// truncate. This method allows to grow the cost of the original /// load with the gain from this slice. void addSliceGain(const LoadedSlice &LS) { // Each slice saves a truncate. const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo(); if (!TLI.isTruncateFree(LS.Inst->getOperand(0).getValueType(), LS.Inst->getValueType(0))) ++Truncates; // If there is a shift amount, this slice gets rid of it. if (LS.Shift) ++Shift; // If this slice can merge a cross register bank copy, account for it. if (LS.canMergeExpensiveCrossRegisterBankCopy()) ++CrossRegisterBanksCopies; } Cost &operator+=(const Cost &RHS) { Loads += RHS.Loads; Truncates += RHS.Truncates; CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies; ZExts += RHS.ZExts; Shift += RHS.Shift; return *this; } bool operator==(const Cost &RHS) const { return Loads == RHS.Loads && Truncates == RHS.Truncates && CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies && ZExts == RHS.ZExts && Shift == RHS.Shift; } bool operator!=(const Cost &RHS) const { return !(*this == RHS); } bool operator<(const Cost &RHS) const { // Assume cross register banks copies are as expensive as loads. // FIXME: Do we want some more target hooks? unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies; unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies; // Unless we are optimizing for code size, consider the // expensive operation first. if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS) return ExpensiveOpsLHS < ExpensiveOpsRHS; return (Truncates + ZExts + Shift + ExpensiveOpsLHS) < (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS); } bool operator>(const Cost &RHS) const { return RHS < *this; } bool operator<=(const Cost &RHS) const { return !(RHS < *this); } bool operator>=(const Cost &RHS) const { return !(*this < RHS); } }; // The last instruction that represent the slice. This should be a // truncate instruction. SDNode *Inst; // The original load instruction. LoadSDNode *Origin; // The right shift amount in bits from the original load. unsigned Shift; // The DAG from which Origin came from. // This is used to get some contextual information about legal types, etc. SelectionDAG *DAG; LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr, unsigned Shift = 0, SelectionDAG *DAG = nullptr) : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {} /// Get the bits used in a chunk of bits \p BitWidth large. /// \return Result is \p BitWidth and has used bits set to 1 and /// not used bits set to 0. APInt getUsedBits() const { // Reproduce the trunc(lshr) sequence: // - Start from the truncated value. // - Zero extend to the desired bit width. // - Shift left. assert(Origin && "No original load to compare against."); unsigned BitWidth = Origin->getValueSizeInBits(0); assert(Inst && "This slice is not bound to an instruction"); assert(Inst->getValueSizeInBits(0) <= BitWidth && "Extracted slice is bigger than the whole type!"); APInt UsedBits(Inst->getValueSizeInBits(0), 0); UsedBits.setAllBits(); UsedBits = UsedBits.zext(BitWidth); UsedBits <<= Shift; return UsedBits; } /// Get the size of the slice to be loaded in bytes. unsigned getLoadedSize() const { unsigned SliceSize = getUsedBits().countPopulation(); assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte."); return SliceSize / 8; } /// Get the type that will be loaded for this slice. /// Note: This may not be the final type for the slice. EVT getLoadedType() const { assert(DAG && "Missing context"); LLVMContext &Ctxt = *DAG->getContext(); return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8); } /// Get the alignment of the load used for this slice. Align getAlign() const { Align Alignment = Origin->getAlign(); uint64_t Offset = getOffsetFromBase(); if (Offset != 0) Alignment = commonAlignment(Alignment, Alignment.value() + Offset); return Alignment; } /// Check if this slice can be rewritten with legal operations. bool isLegal() const { // An invalid slice is not legal. if (!Origin || !Inst || !DAG) return false; // Offsets are for indexed load only, we do not handle that. if (!Origin->getOffset().isUndef()) return false; const TargetLowering &TLI = DAG->getTargetLoweringInfo(); // Check that the type is legal. EVT SliceType = getLoadedType(); if (!TLI.isTypeLegal(SliceType)) return false; // Check that the load is legal for this type. if (!TLI.isOperationLegal(ISD::LOAD, SliceType)) return false; // Check that the offset can be computed. // 1. Check its type. EVT PtrType = Origin->getBasePtr().getValueType(); if (PtrType == MVT::Untyped || PtrType.isExtended()) return false; // 2. Check that it fits in the immediate. if (!TLI.isLegalAddImmediate(getOffsetFromBase())) return false; // 3. Check that the computation is legal. if (!TLI.isOperationLegal(ISD::ADD, PtrType)) return false; // Check that the zext is legal if it needs one. EVT TruncateType = Inst->getValueType(0); if (TruncateType != SliceType && !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType)) return false; return true; } /// Get the offset in bytes of this slice in the original chunk of /// bits. /// \pre DAG != nullptr. uint64_t getOffsetFromBase() const { assert(DAG && "Missing context."); bool IsBigEndian = DAG->getDataLayout().isBigEndian(); assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported."); uint64_t Offset = Shift / 8; unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8; assert(!(Origin->getValueSizeInBits(0) & 0x7) && "The size of the original loaded type is not a multiple of a" " byte."); // If Offset is bigger than TySizeInBytes, it means we are loading all // zeros. This should have been optimized before in the process. assert(TySizeInBytes > Offset && "Invalid shift amount for given loaded size"); if (IsBigEndian) Offset = TySizeInBytes - Offset - getLoadedSize(); return Offset; } /// Generate the sequence of instructions to load the slice /// represented by this object and redirect the uses of this slice to /// this new sequence of instructions. /// \pre this->Inst && this->Origin are valid Instructions and this /// object passed the legal check: LoadedSlice::isLegal returned true. /// \return The last instruction of the sequence used to load the slice. SDValue loadSlice() const { assert(Inst && Origin && "Unable to replace a non-existing slice."); const SDValue &OldBaseAddr = Origin->getBasePtr(); SDValue BaseAddr = OldBaseAddr; // Get the offset in that chunk of bytes w.r.t. the endianness. int64_t Offset = static_cast(getOffsetFromBase()); assert(Offset >= 0 && "Offset too big to fit in int64_t!"); if (Offset) { // BaseAddr = BaseAddr + Offset. EVT ArithType = BaseAddr.getValueType(); SDLoc DL(Origin); BaseAddr = DAG->getNode(ISD::ADD, DL, ArithType, BaseAddr, DAG->getConstant(Offset, DL, ArithType)); } // Create the type of the loaded slice according to its size. EVT SliceType = getLoadedType(); // Create the load for the slice. SDValue LastInst = DAG->getLoad(SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr, Origin->getPointerInfo().getWithOffset(Offset), getAlign(), Origin->getMemOperand()->getFlags()); // If the final type is not the same as the loaded type, this means that // we have to pad with zero. Create a zero extend for that. EVT FinalType = Inst->getValueType(0); if (SliceType != FinalType) LastInst = DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst); return LastInst; } /// Check if this slice can be merged with an expensive cross register /// bank copy. E.g., /// i = load i32 /// f = bitcast i32 i to float bool canMergeExpensiveCrossRegisterBankCopy() const { if (!Inst || !Inst->hasOneUse()) return false; SDNode *Use = *Inst->use_begin(); if (Use->getOpcode() != ISD::BITCAST) return false; assert(DAG && "Missing context"); const TargetLowering &TLI = DAG->getTargetLoweringInfo(); EVT ResVT = Use->getValueType(0); const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT(), Use->isDivergent()); const TargetRegisterClass *ArgRC = TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT(), Use->getOperand(0)->isDivergent()); if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT)) return false; // At this point, we know that we perform a cross-register-bank copy. // Check if it is expensive. const TargetRegisterInfo *TRI = DAG->getSubtarget().getRegisterInfo(); // Assume bitcasts are cheap, unless both register classes do not // explicitly share a common sub class. if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC)) return false; // Check if it will be merged with the load. // 1. Check the alignment / fast memory access constraint. bool IsFast = false; if (!TLI.allowsMemoryAccess(*DAG->getContext(), DAG->getDataLayout(), ResVT, Origin->getAddressSpace(), getAlign(), Origin->getMemOperand()->getFlags(), &IsFast) || !IsFast) return false; // 2. Check that the load is a legal operation for that type. if (!TLI.isOperationLegal(ISD::LOAD, ResVT)) return false; // 3. Check that we do not have a zext in the way. if (Inst->getValueType(0) != getLoadedType()) return false; return true; } }; } // end anonymous namespace /// Check that all bits set in \p UsedBits form a dense region, i.e., /// \p UsedBits looks like 0..0 1..1 0..0. static bool areUsedBitsDense(const APInt &UsedBits) { // If all the bits are one, this is dense! if (UsedBits.isAllOnes()) return true; // Get rid of the unused bits on the right. APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros()); // Get rid of the unused bits on the left. if (NarrowedUsedBits.countLeadingZeros()) NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits()); // Check that the chunk of bits is completely used. return NarrowedUsedBits.isAllOnes(); } /// Check whether or not \p First and \p Second are next to each other /// in memory. This means that there is no hole between the bits loaded /// by \p First and the bits loaded by \p Second. static bool areSlicesNextToEachOther(const LoadedSlice &First, const LoadedSlice &Second) { assert(First.Origin == Second.Origin && First.Origin && "Unable to match different memory origins."); APInt UsedBits = First.getUsedBits(); assert((UsedBits & Second.getUsedBits()) == 0 && "Slices are not supposed to overlap."); UsedBits |= Second.getUsedBits(); return areUsedBitsDense(UsedBits); } /// Adjust the \p GlobalLSCost according to the target /// paring capabilities and the layout of the slices. /// \pre \p GlobalLSCost should account for at least as many loads as /// there is in the slices in \p LoadedSlices. static void adjustCostForPairing(SmallVectorImpl &LoadedSlices, LoadedSlice::Cost &GlobalLSCost) { unsigned NumberOfSlices = LoadedSlices.size(); // If there is less than 2 elements, no pairing is possible. if (NumberOfSlices < 2) return; // Sort the slices so that elements that are likely to be next to each // other in memory are next to each other in the list. llvm::sort(LoadedSlices, [](const LoadedSlice &LHS, const LoadedSlice &RHS) { assert(LHS.Origin == RHS.Origin && "Different bases not implemented."); return LHS.getOffsetFromBase() < RHS.getOffsetFromBase(); }); const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo(); // First (resp. Second) is the first (resp. Second) potentially candidate // to be placed in a paired load. const LoadedSlice *First = nullptr; const LoadedSlice *Second = nullptr; for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice, // Set the beginning of the pair. First = Second) { Second = &LoadedSlices[CurrSlice]; // If First is NULL, it means we start a new pair. // Get to the next slice. if (!First) continue; EVT LoadedType = First->getLoadedType(); // If the types of the slices are different, we cannot pair them. if (LoadedType != Second->getLoadedType()) continue; // Check if the target supplies paired loads for this type. Align RequiredAlignment; if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) { // move to the next pair, this type is hopeless. Second = nullptr; continue; } // Check if we meet the alignment requirement. if (First->getAlign() < RequiredAlignment) continue; // Check that both loads are next to each other in memory. if (!areSlicesNextToEachOther(*First, *Second)) continue; assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!"); --GlobalLSCost.Loads; // Move to the next pair. Second = nullptr; } } /// Check the profitability of all involved LoadedSlice. /// Currently, it is considered profitable if there is exactly two /// involved slices (1) which are (2) next to each other in memory, and /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3). /// /// Note: The order of the elements in \p LoadedSlices may be modified, but not /// the elements themselves. /// /// FIXME: When the cost model will be mature enough, we can relax /// constraints (1) and (2). static bool isSlicingProfitable(SmallVectorImpl &LoadedSlices, const APInt &UsedBits, bool ForCodeSize) { unsigned NumberOfSlices = LoadedSlices.size(); if (StressLoadSlicing) return NumberOfSlices > 1; // Check (1). if (NumberOfSlices != 2) return false; // Check (2). if (!areUsedBitsDense(UsedBits)) return false; // Check (3). LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize); // The original code has one big load. OrigCost.Loads = 1; for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) { const LoadedSlice &LS = LoadedSlices[CurrSlice]; // Accumulate the cost of all the slices. LoadedSlice::Cost SliceCost(LS, ForCodeSize); GlobalSlicingCost += SliceCost; // Account as cost in the original configuration the gain obtained // with the current slices. OrigCost.addSliceGain(LS); } // If the target supports paired load, adjust the cost accordingly. adjustCostForPairing(LoadedSlices, GlobalSlicingCost); return OrigCost > GlobalSlicingCost; } /// If the given load, \p LI, is used only by trunc or trunc(lshr) /// operations, split it in the various pieces being extracted. /// /// This sort of thing is introduced by SROA. /// This slicing takes care not to insert overlapping loads. /// \pre LI is a simple load (i.e., not an atomic or volatile load). bool DAGCombiner::SliceUpLoad(SDNode *N) { if (Level < AfterLegalizeDAG) return false; LoadSDNode *LD = cast(N); if (!LD->isSimple() || !ISD::isNormalLoad(LD) || !LD->getValueType(0).isInteger()) return false; // The algorithm to split up a load of a scalable vector into individual // elements currently requires knowing the length of the loaded type, // so will need adjusting to work on scalable vectors. if (LD->getValueType(0).isScalableVector()) return false; // Keep track of already used bits to detect overlapping values. // In that case, we will just abort the transformation. APInt UsedBits(LD->getValueSizeInBits(0), 0); SmallVector LoadedSlices; // Check if this load is used as several smaller chunks of bits. // Basically, look for uses in trunc or trunc(lshr) and record a new chain // of computation for each trunc. for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end(); UI != UIEnd; ++UI) { // Skip the uses of the chain. if (UI.getUse().getResNo() != 0) continue; SDNode *User = *UI; unsigned Shift = 0; // Check if this is a trunc(lshr). if (User->getOpcode() == ISD::SRL && User->hasOneUse() && isa(User->getOperand(1))) { Shift = User->getConstantOperandVal(1); User = *User->use_begin(); } // At this point, User is a Truncate, iff we encountered, trunc or // trunc(lshr). if (User->getOpcode() != ISD::TRUNCATE) return false; // The width of the type must be a power of 2 and greater than 8-bits. // Otherwise the load cannot be represented in LLVM IR. // Moreover, if we shifted with a non-8-bits multiple, the slice // will be across several bytes. We do not support that. unsigned Width = User->getValueSizeInBits(0); if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7)) return false; // Build the slice for this chain of computations. LoadedSlice LS(User, LD, Shift, &DAG); APInt CurrentUsedBits = LS.getUsedBits(); // Check if this slice overlaps with another. if ((CurrentUsedBits & UsedBits) != 0) return false; // Update the bits used globally. UsedBits |= CurrentUsedBits; // Check if the new slice would be legal. if (!LS.isLegal()) return false; // Record the slice. LoadedSlices.push_back(LS); } // Abort slicing if it does not seem to be profitable. if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize)) return false; ++SlicedLoads; // Rewrite each chain to use an independent load. // By construction, each chain can be represented by a unique load. // Prepare the argument for the new token factor for all the slices. SmallVector ArgChains; for (const LoadedSlice &LS : LoadedSlices) { SDValue SliceInst = LS.loadSlice(); CombineTo(LS.Inst, SliceInst, true); if (SliceInst.getOpcode() != ISD::LOAD) SliceInst = SliceInst.getOperand(0); assert(SliceInst->getOpcode() == ISD::LOAD && "It takes more than a zext to get to the loaded slice!!"); ArgChains.push_back(SliceInst.getValue(1)); } SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other, ArgChains); DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain); AddToWorklist(Chain.getNode()); return true; } /// Check to see if V is (and load (ptr), imm), where the load is having /// specific bytes cleared out. If so, return the byte size being masked out /// and the shift amount. static std::pair CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) { std::pair Result(0, 0); // Check for the structure we're looking for. if (V->getOpcode() != ISD::AND || !isa(V->getOperand(1)) || !ISD::isNormalLoad(V->getOperand(0).getNode())) return Result; // Check the chain and pointer. LoadSDNode *LD = cast(V->getOperand(0)); if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer. // This only handles simple types. if (V.getValueType() != MVT::i16 && V.getValueType() != MVT::i32 && V.getValueType() != MVT::i64) return Result; // Check the constant mask. Invert it so that the bits being masked out are // 0 and the bits being kept are 1. Use getSExtValue so that leading bits // follow the sign bit for uniformity. uint64_t NotMask = ~cast(V->getOperand(1))->getSExtValue(); unsigned NotMaskLZ = countLeadingZeros(NotMask); if (NotMaskLZ & 7) return Result; // Must be multiple of a byte. unsigned NotMaskTZ = countTrailingZeros(NotMask); if (NotMaskTZ & 7) return Result; // Must be multiple of a byte. if (NotMaskLZ == 64) return Result; // All zero mask. // See if we have a continuous run of bits. If so, we have 0*1+0* if (countTrailingOnes(NotMask >> NotMaskTZ) + NotMaskTZ + NotMaskLZ != 64) return Result; // Adjust NotMaskLZ down to be from the actual size of the int instead of i64. if (V.getValueType() != MVT::i64 && NotMaskLZ) NotMaskLZ -= 64-V.getValueSizeInBits(); unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8; switch (MaskedBytes) { case 1: case 2: case 4: break; default: return Result; // All one mask, or 5-byte mask. } // Verify that the first bit starts at a multiple of mask so that the access // is aligned the same as the access width. if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result; // For narrowing to be valid, it must be the case that the load the // immediately preceding memory operation before the store. if (LD == Chain.getNode()) ; // ok. else if (Chain->getOpcode() == ISD::TokenFactor && SDValue(LD, 1).hasOneUse()) { // LD has only 1 chain use so they are no indirect dependencies. if (!LD->isOperandOf(Chain.getNode())) return Result; } else return Result; // Fail. Result.first = MaskedBytes; Result.second = NotMaskTZ/8; return Result; } /// Check to see if IVal is something that provides a value as specified by /// MaskInfo. If so, replace the specified store with a narrower store of /// truncated IVal. static SDValue ShrinkLoadReplaceStoreWithStore(const std::pair &MaskInfo, SDValue IVal, StoreSDNode *St, DAGCombiner *DC) { unsigned NumBytes = MaskInfo.first; unsigned ByteShift = MaskInfo.second; SelectionDAG &DAG = DC->getDAG(); // Check to see if IVal is all zeros in the part being masked in by the 'or' // that uses this. If not, this is not a replacement. APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(), ByteShift*8, (ByteShift+NumBytes)*8); if (!DAG.MaskedValueIsZero(IVal, Mask)) return SDValue(); // Check that it is legal on the target to do this. It is legal if the new // VT we're shrinking to (i8/i16/i32) is legal or we're still before type // legalization. If the source type is legal, but the store type isn't, see // if we can use a truncating store. MVT VT = MVT::getIntegerVT(NumBytes * 8); const TargetLowering &TLI = DAG.getTargetLoweringInfo(); bool UseTruncStore; if (DC->isTypeLegal(VT)) UseTruncStore = false; else if (TLI.isTypeLegal(IVal.getValueType()) && TLI.isTruncStoreLegal(IVal.getValueType(), VT)) UseTruncStore = true; else return SDValue(); // Check that the target doesn't think this is a bad idea. if (St->getMemOperand() && !TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT, *St->getMemOperand())) return SDValue(); // Okay, we can do this! Replace the 'St' store with a store of IVal that is // shifted by ByteShift and truncated down to NumBytes. if (ByteShift) { SDLoc DL(IVal); IVal = DAG.getNode(ISD::SRL, DL, IVal.getValueType(), IVal, DAG.getConstant(ByteShift*8, DL, DC->getShiftAmountTy(IVal.getValueType()))); } // Figure out the offset for the store and the alignment of the access. unsigned StOffset; if (DAG.getDataLayout().isLittleEndian()) StOffset = ByteShift; else StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes; SDValue Ptr = St->getBasePtr(); if (StOffset) { SDLoc DL(IVal); Ptr = DAG.getMemBasePlusOffset(Ptr, TypeSize::Fixed(StOffset), DL); } ++OpsNarrowed; if (UseTruncStore) return DAG.getTruncStore(St->getChain(), SDLoc(St), IVal, Ptr, St->getPointerInfo().getWithOffset(StOffset), VT, St->getOriginalAlign()); // Truncate down to the new size. IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal); return DAG .getStore(St->getChain(), SDLoc(St), IVal, Ptr, St->getPointerInfo().getWithOffset(StOffset), St->getOriginalAlign()); } /// Look for sequence of load / op / store where op is one of 'or', 'xor', and /// 'and' of immediates. If 'op' is only touching some of the loaded bits, try /// narrowing the load and store if it would end up being a win for performance /// or code size. SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) { StoreSDNode *ST = cast(N); if (!ST->isSimple()) return SDValue(); SDValue Chain = ST->getChain(); SDValue Value = ST->getValue(); SDValue Ptr = ST->getBasePtr(); EVT VT = Value.getValueType(); if (ST->isTruncatingStore() || VT.isVector()) return SDValue(); unsigned Opc = Value.getOpcode(); if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) || !Value.hasOneUse()) return SDValue(); // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst // is a byte mask indicating a consecutive number of bytes, check to see if // Y is known to provide just those bytes. If so, we try to replace the // load + replace + store sequence with a single (narrower) store, which makes // the load dead. if (Opc == ISD::OR && EnableShrinkLoadReplaceStoreWithStore) { std::pair MaskedLoad; MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain); if (MaskedLoad.first) if (SDValue NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad, Value.getOperand(1), ST,this)) return NewST; // Or is commutative, so try swapping X and Y. MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain); if (MaskedLoad.first) if (SDValue NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad, Value.getOperand(0), ST,this)) return NewST; } if (!EnableReduceLoadOpStoreWidth) return SDValue(); if (Value.getOperand(1).getOpcode() != ISD::Constant) return SDValue(); SDValue N0 = Value.getOperand(0); if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() && Chain == SDValue(N0.getNode(), 1)) { LoadSDNode *LD = cast(N0); if (LD->getBasePtr() != Ptr || LD->getPointerInfo().getAddrSpace() != ST->getPointerInfo().getAddrSpace()) return SDValue(); // Find the type to narrow it the load / op / store to. SDValue N1 = Value.getOperand(1); unsigned BitWidth = N1.getValueSizeInBits(); APInt Imm = cast(N1)->getAPIntValue(); if (Opc == ISD::AND) Imm ^= APInt::getAllOnes(BitWidth); if (Imm == 0 || Imm.isAllOnes()) return SDValue(); unsigned ShAmt = Imm.countTrailingZeros(); unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1; unsigned NewBW = NextPowerOf2(MSB - ShAmt); EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW); // The narrowing should be profitable, the load/store operation should be // legal (or custom) and the store size should be equal to the NewVT width. while (NewBW < BitWidth && (NewVT.getStoreSizeInBits() != NewBW || !TLI.isOperationLegalOrCustom(Opc, NewVT) || !TLI.isNarrowingProfitable(VT, NewVT))) { NewBW = NextPowerOf2(NewBW); NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW); } if (NewBW >= BitWidth) return SDValue(); // If the lsb changed does not start at the type bitwidth boundary, // start at the previous one. if (ShAmt % NewBW) ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW; APInt Mask = APInt::getBitsSet(BitWidth, ShAmt, std::min(BitWidth, ShAmt + NewBW)); if ((Imm & Mask) == Imm) { APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW); if (Opc == ISD::AND) NewImm ^= APInt::getAllOnes(NewBW); uint64_t PtrOff = ShAmt / 8; // For big endian targets, we need to adjust the offset to the pointer to // load the correct bytes. if (DAG.getDataLayout().isBigEndian()) PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff; bool IsFast = false; Align NewAlign = commonAlignment(LD->getAlign(), PtrOff); if (!TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), NewVT, LD->getAddressSpace(), NewAlign, LD->getMemOperand()->getFlags(), &IsFast) || !IsFast) return SDValue(); SDValue NewPtr = DAG.getMemBasePlusOffset(Ptr, TypeSize::Fixed(PtrOff), SDLoc(LD)); SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0), LD->getChain(), NewPtr, LD->getPointerInfo().getWithOffset(PtrOff), NewAlign, LD->getMemOperand()->getFlags(), LD->getAAInfo()); SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD, DAG.getConstant(NewImm, SDLoc(Value), NewVT)); SDValue NewST = DAG.getStore(Chain, SDLoc(N), NewVal, NewPtr, ST->getPointerInfo().getWithOffset(PtrOff), NewAlign); AddToWorklist(NewPtr.getNode()); AddToWorklist(NewLD.getNode()); AddToWorklist(NewVal.getNode()); WorklistRemover DeadNodes(*this); DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1)); ++OpsNarrowed; return NewST; } } return SDValue(); } /// For a given floating point load / store pair, if the load value isn't used /// by any other operations, then consider transforming the pair to integer /// load / store operations if the target deems the transformation profitable. SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) { StoreSDNode *ST = cast(N); SDValue Value = ST->getValue(); if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) && Value.hasOneUse()) { LoadSDNode *LD = cast(Value); EVT VT = LD->getMemoryVT(); if (!VT.isFloatingPoint() || VT != ST->getMemoryVT() || LD->isNonTemporal() || ST->isNonTemporal() || LD->getPointerInfo().getAddrSpace() != 0 || ST->getPointerInfo().getAddrSpace() != 0) return SDValue(); TypeSize VTSize = VT.getSizeInBits(); // We don't know the size of scalable types at compile time so we cannot // create an integer of the equivalent size. if (VTSize.isScalable()) return SDValue(); bool FastLD = false, FastST = false; EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VTSize.getFixedSize()); if (!TLI.isOperationLegal(ISD::LOAD, IntVT) || !TLI.isOperationLegal(ISD::STORE, IntVT) || !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) || !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT) || !TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), IntVT, *LD->getMemOperand(), &FastLD) || !TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), IntVT, *ST->getMemOperand(), &FastST) || !FastLD || !FastST) return SDValue(); SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value), LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(), LD->getAlign()); SDValue NewST = DAG.getStore(ST->getChain(), SDLoc(N), NewLD, ST->getBasePtr(), ST->getPointerInfo(), ST->getAlign()); AddToWorklist(NewLD.getNode()); AddToWorklist(NewST.getNode()); WorklistRemover DeadNodes(*this); DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1)); ++LdStFP2Int; return NewST; } return SDValue(); } // This is a helper function for visitMUL to check the profitability // of folding (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2). // MulNode is the original multiply, AddNode is (add x, c1), // and ConstNode is c2. // // If the (add x, c1) has multiple uses, we could increase // the number of adds if we make this transformation. // It would only be worth doing this if we can remove a // multiply in the process. Check for that here. // To illustrate: // (A + c1) * c3 // (A + c2) * c3 // We're checking for cases where we have common "c3 * A" expressions. bool DAGCombiner::isMulAddWithConstProfitable(SDNode *MulNode, SDValue AddNode, SDValue ConstNode) { APInt Val; // If the add only has one use, and the target thinks the folding is // profitable or does not lead to worse code, this would be OK to do. if (AddNode->hasOneUse() && TLI.isMulAddWithConstProfitable(AddNode, ConstNode)) return true; // Walk all the users of the constant with which we're multiplying. for (SDNode *Use : ConstNode->uses()) { if (Use == MulNode) // This use is the one we're on right now. Skip it. continue; if (Use->getOpcode() == ISD::MUL) { // We have another multiply use. SDNode *OtherOp; SDNode *MulVar = AddNode.getOperand(0).getNode(); // OtherOp is what we're multiplying against the constant. if (Use->getOperand(0) == ConstNode) OtherOp = Use->getOperand(1).getNode(); else OtherOp = Use->getOperand(0).getNode(); // Check to see if multiply is with the same operand of our "add". // // ConstNode = CONST // Use = ConstNode * A <-- visiting Use. OtherOp is A. // ... // AddNode = (A + c1) <-- MulVar is A. // = AddNode * ConstNode <-- current visiting instruction. // // If we make this transformation, we will have a common // multiply (ConstNode * A) that we can save. if (OtherOp == MulVar) return true; // Now check to see if a future expansion will give us a common // multiply. // // ConstNode = CONST // AddNode = (A + c1) // ... = AddNode * ConstNode <-- current visiting instruction. // ... // OtherOp = (A + c2) // Use = OtherOp * ConstNode <-- visiting Use. // // If we make this transformation, we will have a common // multiply (CONST * A) after we also do the same transformation // to the "t2" instruction. if (OtherOp->getOpcode() == ISD::ADD && DAG.isConstantIntBuildVectorOrConstantInt(OtherOp->getOperand(1)) && OtherOp->getOperand(0).getNode() == MulVar) return true; } } // Didn't find a case where this would be profitable. return false; } SDValue DAGCombiner::getMergeStoreChains(SmallVectorImpl &StoreNodes, unsigned NumStores) { SmallVector Chains; SmallPtrSet Visited; SDLoc StoreDL(StoreNodes[0].MemNode); for (unsigned i = 0; i < NumStores; ++i) { Visited.insert(StoreNodes[i].MemNode); } // don't include nodes that are children or repeated nodes. for (unsigned i = 0; i < NumStores; ++i) { if (Visited.insert(StoreNodes[i].MemNode->getChain().getNode()).second) Chains.push_back(StoreNodes[i].MemNode->getChain()); } assert(Chains.size() > 0 && "Chain should have generated a chain"); return DAG.getTokenFactor(StoreDL, Chains); } bool DAGCombiner::mergeStoresOfConstantsOrVecElts( SmallVectorImpl &StoreNodes, EVT MemVT, unsigned NumStores, bool IsConstantSrc, bool UseVector, bool UseTrunc) { // Make sure we have something to merge. if (NumStores < 2) return false; assert((!UseTrunc || !UseVector) && "This optimization cannot emit a vector truncating store"); // The latest Node in the DAG. SDLoc DL(StoreNodes[0].MemNode); TypeSize ElementSizeBits = MemVT.getStoreSizeInBits(); unsigned SizeInBits = NumStores * ElementSizeBits; unsigned NumMemElts = MemVT.isVector() ? MemVT.getVectorNumElements() : 1; Optional Flags; AAMDNodes AAInfo; for (unsigned I = 0; I != NumStores; ++I) { StoreSDNode *St = cast(StoreNodes[I].MemNode); if (!Flags) { Flags = St->getMemOperand()->getFlags(); AAInfo = St->getAAInfo(); continue; } // Skip merging if there's an inconsistent flag. if (Flags != St->getMemOperand()->getFlags()) return false; // Concatenate AA metadata. AAInfo = AAInfo.concat(St->getAAInfo()); } EVT StoreTy; if (UseVector) { unsigned Elts = NumStores * NumMemElts; // Get the type for the merged vector store. StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(), Elts); } else StoreTy = EVT::getIntegerVT(*DAG.getContext(), SizeInBits); SDValue StoredVal; if (UseVector) { if (IsConstantSrc) { SmallVector BuildVector; for (unsigned I = 0; I != NumStores; ++I) { StoreSDNode *St = cast(StoreNodes[I].MemNode); SDValue Val = St->getValue(); // If constant is of the wrong type, convert it now. if (MemVT != Val.getValueType()) { Val = peekThroughBitcasts(Val); // Deal with constants of wrong size. if (ElementSizeBits != Val.getValueSizeInBits()) { EVT IntMemVT = EVT::getIntegerVT(*DAG.getContext(), MemVT.getSizeInBits()); if (isa(Val)) { // Not clear how to truncate FP values. return false; } if (auto *C = dyn_cast(Val)) Val = DAG.getConstant(C->getAPIntValue() .zextOrTrunc(Val.getValueSizeInBits()) .zextOrTrunc(ElementSizeBits), SDLoc(C), IntMemVT); } // Make sure correctly size type is the correct type. Val = DAG.getBitcast(MemVT, Val); } BuildVector.push_back(Val); } StoredVal = DAG.getNode(MemVT.isVector() ? ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, DL, StoreTy, BuildVector); } else { SmallVector Ops; for (unsigned i = 0; i < NumStores; ++i) { StoreSDNode *St = cast(StoreNodes[i].MemNode); SDValue Val = peekThroughBitcasts(St->getValue()); // All operands of BUILD_VECTOR / CONCAT_VECTOR must be of // type MemVT. If the underlying value is not the correct // type, but it is an extraction of an appropriate vector we // can recast Val to be of the correct type. This may require // converting between EXTRACT_VECTOR_ELT and // EXTRACT_SUBVECTOR. if ((MemVT != Val.getValueType()) && (Val.getOpcode() == ISD::EXTRACT_VECTOR_ELT || Val.getOpcode() == ISD::EXTRACT_SUBVECTOR)) { EVT MemVTScalarTy = MemVT.getScalarType(); // We may need to add a bitcast here to get types to line up. if (MemVTScalarTy != Val.getValueType().getScalarType()) { Val = DAG.getBitcast(MemVT, Val); } else { unsigned OpC = MemVT.isVector() ? ISD::EXTRACT_SUBVECTOR : ISD::EXTRACT_VECTOR_ELT; SDValue Vec = Val.getOperand(0); SDValue Idx = Val.getOperand(1); Val = DAG.getNode(OpC, SDLoc(Val), MemVT, Vec, Idx); } } Ops.push_back(Val); } // Build the extracted vector elements back into a vector. StoredVal = DAG.getNode(MemVT.isVector() ? ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, DL, StoreTy, Ops); } } else { // We should always use a vector store when merging extracted vector // elements, so this path implies a store of constants. assert(IsConstantSrc && "Merged vector elements should use vector store"); APInt StoreInt(SizeInBits, 0); // Construct a single integer constant which is made of the smaller // constant inputs. bool IsLE = DAG.getDataLayout().isLittleEndian(); for (unsigned i = 0; i < NumStores; ++i) { unsigned Idx = IsLE ? (NumStores - 1 - i) : i; StoreSDNode *St = cast(StoreNodes[Idx].MemNode); SDValue Val = St->getValue(); Val = peekThroughBitcasts(Val); StoreInt <<= ElementSizeBits; if (ConstantSDNode *C = dyn_cast(Val)) { StoreInt |= C->getAPIntValue() .zextOrTrunc(ElementSizeBits) .zextOrTrunc(SizeInBits); } else if (ConstantFPSDNode *C = dyn_cast(Val)) { StoreInt |= C->getValueAPF() .bitcastToAPInt() .zextOrTrunc(ElementSizeBits) .zextOrTrunc(SizeInBits); // If fp truncation is necessary give up for now. if (MemVT.getSizeInBits() != ElementSizeBits) return false; } else { llvm_unreachable("Invalid constant element type"); } } // Create the new Load and Store operations. StoredVal = DAG.getConstant(StoreInt, DL, StoreTy); } LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode; SDValue NewChain = getMergeStoreChains(StoreNodes, NumStores); // make sure we use trunc store if it's necessary to be legal. SDValue NewStore; if (!UseTrunc) { NewStore = DAG.getStore(NewChain, DL, StoredVal, FirstInChain->getBasePtr(), FirstInChain->getPointerInfo(), FirstInChain->getAlign(), *Flags, AAInfo); } else { // Must be realized as a trunc store EVT LegalizedStoredValTy = TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType()); unsigned LegalizedStoreSize = LegalizedStoredValTy.getSizeInBits(); ConstantSDNode *C = cast(StoredVal); SDValue ExtendedStoreVal = DAG.getConstant(C->getAPIntValue().zextOrTrunc(LegalizedStoreSize), DL, LegalizedStoredValTy); NewStore = DAG.getTruncStore( NewChain, DL, ExtendedStoreVal, FirstInChain->getBasePtr(), FirstInChain->getPointerInfo(), StoredVal.getValueType() /*TVT*/, FirstInChain->getAlign(), *Flags, AAInfo); } // Replace all merged stores with the new store. for (unsigned i = 0; i < NumStores; ++i) CombineTo(StoreNodes[i].MemNode, NewStore); AddToWorklist(NewChain.getNode()); return true; } void DAGCombiner::getStoreMergeCandidates( StoreSDNode *St, SmallVectorImpl &StoreNodes, SDNode *&RootNode) { // This holds the base pointer, index, and the offset in bytes from the base // pointer. We must have a base and an offset. Do not handle stores to undef // base pointers. BaseIndexOffset BasePtr = BaseIndexOffset::match(St, DAG); if (!BasePtr.getBase().getNode() || BasePtr.getBase().isUndef()) return; SDValue Val = peekThroughBitcasts(St->getValue()); StoreSource StoreSrc = getStoreSource(Val); assert(StoreSrc != StoreSource::Unknown && "Expected known source for store"); // Match on loadbaseptr if relevant. EVT MemVT = St->getMemoryVT(); BaseIndexOffset LBasePtr; EVT LoadVT; if (StoreSrc == StoreSource::Load) { auto *Ld = cast(Val); LBasePtr = BaseIndexOffset::match(Ld, DAG); LoadVT = Ld->getMemoryVT(); // Load and store should be the same type. if (MemVT != LoadVT) return; // Loads must only have one use. if (!Ld->hasNUsesOfValue(1, 0)) return; // The memory operands must not be volatile/indexed/atomic. // TODO: May be able to relax for unordered atomics (see D66309) if (!Ld->isSimple() || Ld->isIndexed()) return; } auto CandidateMatch = [&](StoreSDNode *Other, BaseIndexOffset &Ptr, int64_t &Offset) -> bool { // The memory operands must not be volatile/indexed/atomic. // TODO: May be able to relax for unordered atomics (see D66309) if (!Other->isSimple() || Other->isIndexed()) return false; // Don't mix temporal stores with non-temporal stores. if (St->isNonTemporal() != Other->isNonTemporal()) return false; SDValue OtherBC = peekThroughBitcasts(Other->getValue()); // Allow merging constants of different types as integers. bool NoTypeMatch = (MemVT.isInteger()) ? !MemVT.bitsEq(Other->getMemoryVT()) : Other->getMemoryVT() != MemVT; switch (StoreSrc) { case StoreSource::Load: { if (NoTypeMatch) return false; // The Load's Base Ptr must also match. auto *OtherLd = dyn_cast(OtherBC); if (!OtherLd) return false; BaseIndexOffset LPtr = BaseIndexOffset::match(OtherLd, DAG); if (LoadVT != OtherLd->getMemoryVT()) return false; // Loads must only have one use. if (!OtherLd->hasNUsesOfValue(1, 0)) return false; // The memory operands must not be volatile/indexed/atomic. // TODO: May be able to relax for unordered atomics (see D66309) if (!OtherLd->isSimple() || OtherLd->isIndexed()) return false; // Don't mix temporal loads with non-temporal loads. if (cast(Val)->isNonTemporal() != OtherLd->isNonTemporal()) return false; if (!(LBasePtr.equalBaseIndex(LPtr, DAG))) return false; break; } case StoreSource::Constant: if (NoTypeMatch) return false; if (!isIntOrFPConstant(OtherBC)) return false; break; case StoreSource::Extract: // Do not merge truncated stores here. if (Other->isTruncatingStore()) return false; if (!MemVT.bitsEq(OtherBC.getValueType())) return false; if (OtherBC.getOpcode() != ISD::EXTRACT_VECTOR_ELT && OtherBC.getOpcode() != ISD::EXTRACT_SUBVECTOR) return false; break; default: llvm_unreachable("Unhandled store source for merging"); } Ptr = BaseIndexOffset::match(Other, DAG); return (BasePtr.equalBaseIndex(Ptr, DAG, Offset)); }; // Check if the pair of StoreNode and the RootNode already bail out many // times which is over the limit in dependence check. auto OverLimitInDependenceCheck = [&](SDNode *StoreNode, SDNode *RootNode) -> bool { auto RootCount = StoreRootCountMap.find(StoreNode); return RootCount != StoreRootCountMap.end() && RootCount->second.first == RootNode && RootCount->second.second > StoreMergeDependenceLimit; }; auto TryToAddCandidate = [&](SDNode::use_iterator UseIter) { // This must be a chain use. if (UseIter.getOperandNo() != 0) return; if (auto *OtherStore = dyn_cast(*UseIter)) { BaseIndexOffset Ptr; int64_t PtrDiff; if (CandidateMatch(OtherStore, Ptr, PtrDiff) && !OverLimitInDependenceCheck(OtherStore, RootNode)) StoreNodes.push_back(MemOpLink(OtherStore, PtrDiff)); } }; // We looking for a root node which is an ancestor to all mergable // stores. We search up through a load, to our root and then down // through all children. For instance we will find Store{1,2,3} if // St is Store1, Store2. or Store3 where the root is not a load // which always true for nonvolatile ops. TODO: Expand // the search to find all valid candidates through multiple layers of loads. // // Root // |-------|-------| // Load Load Store3 // | | // Store1 Store2 // // FIXME: We should be able to climb and // descend TokenFactors to find candidates as well. RootNode = St->getChain().getNode(); unsigned NumNodesExplored = 0; const unsigned MaxSearchNodes = 1024; if (auto *Ldn = dyn_cast(RootNode)) { RootNode = Ldn->getChain().getNode(); for (auto I = RootNode->use_begin(), E = RootNode->use_end(); I != E && NumNodesExplored < MaxSearchNodes; ++I, ++NumNodesExplored) { if (I.getOperandNo() == 0 && isa(*I)) { // walk down chain for (auto I2 = (*I)->use_begin(), E2 = (*I)->use_end(); I2 != E2; ++I2) TryToAddCandidate(I2); } // Check stores that depend on the root (e.g. Store 3 in the chart above). if (I.getOperandNo() == 0 && isa(*I)) { TryToAddCandidate(I); } } } else { for (auto I = RootNode->use_begin(), E = RootNode->use_end(); I != E && NumNodesExplored < MaxSearchNodes; ++I, ++NumNodesExplored) TryToAddCandidate(I); } } // We need to check that merging these stores does not cause a loop in the // DAG. Any store candidate may depend on another candidate indirectly through // its operands. Check in parallel by searching up from operands of candidates. bool DAGCombiner::checkMergeStoreCandidatesForDependencies( SmallVectorImpl &StoreNodes, unsigned NumStores, SDNode *RootNode) { // FIXME: We should be able to truncate a full search of // predecessors by doing a BFS and keeping tabs the originating // stores from which worklist nodes come from in a similar way to // TokenFactor simplfication. SmallPtrSet Visited; SmallVector Worklist; // RootNode is a predecessor to all candidates so we need not search // past it. Add RootNode (peeking through TokenFactors). Do not count // these towards size check. Worklist.push_back(RootNode); while (!Worklist.empty()) { auto N = Worklist.pop_back_val(); if (!Visited.insert(N).second) continue; // Already present in Visited. if (N->getOpcode() == ISD::TokenFactor) { for (SDValue Op : N->ops()) Worklist.push_back(Op.getNode()); } } // Don't count pruning nodes towards max. unsigned int Max = 1024 + Visited.size(); // Search Ops of store candidates. for (unsigned i = 0; i < NumStores; ++i) { SDNode *N = StoreNodes[i].MemNode; // Of the 4 Store Operands: // * Chain (Op 0) -> We have already considered these // in candidate selection, but only by following the // chain dependencies. We could still have a chain // dependency to a load, that has a non-chain dep to // another load, that depends on a store, etc. So it is // possible to have dependencies that consist of a mix // of chain and non-chain deps, and we need to include // chain operands in the analysis here.. // * Value (Op 1) -> Cycles may happen (e.g. through load chains) // * Address (Op 2) -> Merged addresses may only vary by a fixed constant, // but aren't necessarily fromt the same base node, so // cycles possible (e.g. via indexed store). // * (Op 3) -> Represents the pre or post-indexing offset (or undef for // non-indexed stores). Not constant on all targets (e.g. ARM) // and so can participate in a cycle. for (unsigned j = 0; j < N->getNumOperands(); ++j) Worklist.push_back(N->getOperand(j).getNode()); } // Search through DAG. We can stop early if we find a store node. for (unsigned i = 0; i < NumStores; ++i) if (SDNode::hasPredecessorHelper(StoreNodes[i].MemNode, Visited, Worklist, Max)) { // If the searching bail out, record the StoreNode and RootNode in the // StoreRootCountMap. If we have seen the pair many times over a limit, // we won't add the StoreNode into StoreNodes set again. if (Visited.size() >= Max) { auto &RootCount = StoreRootCountMap[StoreNodes[i].MemNode]; if (RootCount.first == RootNode) RootCount.second++; else RootCount = {RootNode, 1}; } return false; } return true; } unsigned DAGCombiner::getConsecutiveStores(SmallVectorImpl &StoreNodes, int64_t ElementSizeBytes) const { while (true) { // Find a store past the width of the first store. size_t StartIdx = 0; while ((StartIdx + 1 < StoreNodes.size()) && StoreNodes[StartIdx].OffsetFromBase + ElementSizeBytes != StoreNodes[StartIdx + 1].OffsetFromBase) ++StartIdx; // Bail if we don't have enough candidates to merge. if (StartIdx + 1 >= StoreNodes.size()) return 0; // Trim stores that overlapped with the first store. if (StartIdx) StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + StartIdx); // Scan the memory operations on the chain and find the first // non-consecutive store memory address. unsigned NumConsecutiveStores = 1; int64_t StartAddress = StoreNodes[0].OffsetFromBase; // Check that the addresses are consecutive starting from the second // element in the list of stores. for (unsigned i = 1, e = StoreNodes.size(); i < e; ++i) { int64_t CurrAddress = StoreNodes[i].OffsetFromBase; if (CurrAddress - StartAddress != (ElementSizeBytes * i)) break; NumConsecutiveStores = i + 1; } if (NumConsecutiveStores > 1) return NumConsecutiveStores; // There are no consecutive stores at the start of the list. // Remove the first store and try again. StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + 1); } } bool DAGCombiner::tryStoreMergeOfConstants( SmallVectorImpl &StoreNodes, unsigned NumConsecutiveStores, EVT MemVT, SDNode *RootNode, bool AllowVectors) { LLVMContext &Context = *DAG.getContext(); const DataLayout &DL = DAG.getDataLayout(); int64_t ElementSizeBytes = MemVT.getStoreSize(); unsigned NumMemElts = MemVT.isVector() ? MemVT.getVectorNumElements() : 1; bool MadeChange = false; // Store the constants into memory as one consecutive store. while (NumConsecutiveStores >= 2) { LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode; unsigned FirstStoreAS = FirstInChain->getAddressSpace(); Align FirstStoreAlign = FirstInChain->getAlign(); unsigned LastLegalType = 1; unsigned LastLegalVectorType = 1; bool LastIntegerTrunc = false; bool NonZero = false; unsigned FirstZeroAfterNonZero = NumConsecutiveStores; for (unsigned i = 0; i < NumConsecutiveStores; ++i) { StoreSDNode *ST = cast(StoreNodes[i].MemNode); SDValue StoredVal = ST->getValue(); bool IsElementZero = false; if (ConstantSDNode *C = dyn_cast(StoredVal)) IsElementZero = C->isZero(); else if (ConstantFPSDNode *C = dyn_cast(StoredVal)) IsElementZero = C->getConstantFPValue()->isNullValue(); if (IsElementZero) { if (NonZero && FirstZeroAfterNonZero == NumConsecutiveStores) FirstZeroAfterNonZero = i; } NonZero |= !IsElementZero; // Find a legal type for the constant store. unsigned SizeInBits = (i + 1) * ElementSizeBytes * 8; EVT StoreTy = EVT::getIntegerVT(Context, SizeInBits); bool IsFast = false; // Break early when size is too large to be legal. if (StoreTy.getSizeInBits() > MaximumLegalStoreInBits) break; if (TLI.isTypeLegal(StoreTy) && TLI.canMergeStoresTo(FirstStoreAS, StoreTy, DAG.getMachineFunction()) && TLI.allowsMemoryAccess(Context, DL, StoreTy, *FirstInChain->getMemOperand(), &IsFast) && IsFast) { LastIntegerTrunc = false; LastLegalType = i + 1; // Or check whether a truncstore is legal. } else if (TLI.getTypeAction(Context, StoreTy) == TargetLowering::TypePromoteInteger) { EVT LegalizedStoredValTy = TLI.getTypeToTransformTo(Context, StoredVal.getValueType()); if (TLI.isTruncStoreLegal(LegalizedStoredValTy, StoreTy) && TLI.canMergeStoresTo(FirstStoreAS, LegalizedStoredValTy, DAG.getMachineFunction()) && TLI.allowsMemoryAccess(Context, DL, StoreTy, *FirstInChain->getMemOperand(), &IsFast) && IsFast) { LastIntegerTrunc = true; LastLegalType = i + 1; } } // We only use vectors if the constant is known to be zero or the // target allows it and the function is not marked with the // noimplicitfloat attribute. if ((!NonZero || TLI.storeOfVectorConstantIsCheap(MemVT, i + 1, FirstStoreAS)) && AllowVectors) { // Find a legal type for the vector store. unsigned Elts = (i + 1) * NumMemElts; EVT Ty = EVT::getVectorVT(Context, MemVT.getScalarType(), Elts); if (TLI.isTypeLegal(Ty) && TLI.isTypeLegal(MemVT) && TLI.canMergeStoresTo(FirstStoreAS, Ty, DAG.getMachineFunction()) && TLI.allowsMemoryAccess(Context, DL, Ty, *FirstInChain->getMemOperand(), &IsFast) && IsFast) LastLegalVectorType = i + 1; } } bool UseVector = (LastLegalVectorType > LastLegalType) && AllowVectors; unsigned NumElem = (UseVector) ? LastLegalVectorType : LastLegalType; bool UseTrunc = LastIntegerTrunc && !UseVector; // Check if we found a legal integer type that creates a meaningful // merge. if (NumElem < 2) { // We know that candidate stores are in order and of correct // shape. While there is no mergeable sequence from the // beginning one may start later in the sequence. The only // reason a merge of size N could have failed where another of // the same size would not have, is if the alignment has // improved or we've dropped a non-zero value. Drop as many // candidates as we can here. unsigned NumSkip = 1; while ((NumSkip < NumConsecutiveStores) && (NumSkip < FirstZeroAfterNonZero) && (StoreNodes[NumSkip].MemNode->getAlign() <= FirstStoreAlign)) NumSkip++; StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + NumSkip); NumConsecutiveStores -= NumSkip; continue; } // Check that we can merge these candidates without causing a cycle. if (!checkMergeStoreCandidatesForDependencies(StoreNodes, NumElem, RootNode)) { StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + NumElem); NumConsecutiveStores -= NumElem; continue; } MadeChange |= mergeStoresOfConstantsOrVecElts(StoreNodes, MemVT, NumElem, /*IsConstantSrc*/ true, UseVector, UseTrunc); // Remove merged stores for next iteration. StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + NumElem); NumConsecutiveStores -= NumElem; } return MadeChange; } bool DAGCombiner::tryStoreMergeOfExtracts( SmallVectorImpl &StoreNodes, unsigned NumConsecutiveStores, EVT MemVT, SDNode *RootNode) { LLVMContext &Context = *DAG.getContext(); const DataLayout &DL = DAG.getDataLayout(); unsigned NumMemElts = MemVT.isVector() ? MemVT.getVectorNumElements() : 1; bool MadeChange = false; // Loop on Consecutive Stores on success. while (NumConsecutiveStores >= 2) { LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode; unsigned FirstStoreAS = FirstInChain->getAddressSpace(); Align FirstStoreAlign = FirstInChain->getAlign(); unsigned NumStoresToMerge = 1; for (unsigned i = 0; i < NumConsecutiveStores; ++i) { // Find a legal type for the vector store. unsigned Elts = (i + 1) * NumMemElts; EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(), Elts); bool IsFast = false; // Break early when size is too large to be legal. if (Ty.getSizeInBits() > MaximumLegalStoreInBits) break; if (TLI.isTypeLegal(Ty) && TLI.canMergeStoresTo(FirstStoreAS, Ty, DAG.getMachineFunction()) && TLI.allowsMemoryAccess(Context, DL, Ty, *FirstInChain->getMemOperand(), &IsFast) && IsFast) NumStoresToMerge = i + 1; } // Check if we found a legal integer type creating a meaningful // merge. if (NumStoresToMerge < 2) { // We know that candidate stores are in order and of correct // shape. While there is no mergeable sequence from the // beginning one may start later in the sequence. The only // reason a merge of size N could have failed where another of // the same size would not have, is if the alignment has // improved. Drop as many candidates as we can here. unsigned NumSkip = 1; while ((NumSkip < NumConsecutiveStores) && (StoreNodes[NumSkip].MemNode->getAlign() <= FirstStoreAlign)) NumSkip++; StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + NumSkip); NumConsecutiveStores -= NumSkip; continue; } // Check that we can merge these candidates without causing a cycle. if (!checkMergeStoreCandidatesForDependencies(StoreNodes, NumStoresToMerge, RootNode)) { StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + NumStoresToMerge); NumConsecutiveStores -= NumStoresToMerge; continue; } MadeChange |= mergeStoresOfConstantsOrVecElts( StoreNodes, MemVT, NumStoresToMerge, /*IsConstantSrc*/ false, /*UseVector*/ true, /*UseTrunc*/ false); StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + NumStoresToMerge); NumConsecutiveStores -= NumStoresToMerge; } return MadeChange; } bool DAGCombiner::tryStoreMergeOfLoads(SmallVectorImpl &StoreNodes, unsigned NumConsecutiveStores, EVT MemVT, SDNode *RootNode, bool AllowVectors, bool IsNonTemporalStore, bool IsNonTemporalLoad) { LLVMContext &Context = *DAG.getContext(); const DataLayout &DL = DAG.getDataLayout(); int64_t ElementSizeBytes = MemVT.getStoreSize(); unsigned NumMemElts = MemVT.isVector() ? MemVT.getVectorNumElements() : 1; bool MadeChange = false; // Look for load nodes which are used by the stored values. SmallVector LoadNodes; // Find acceptable loads. Loads need to have the same chain (token factor), // must not be zext, volatile, indexed, and they must be consecutive. BaseIndexOffset LdBasePtr; for (unsigned i = 0; i < NumConsecutiveStores; ++i) { StoreSDNode *St = cast(StoreNodes[i].MemNode); SDValue Val = peekThroughBitcasts(St->getValue()); LoadSDNode *Ld = cast(Val); BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld, DAG); // If this is not the first ptr that we check. int64_t LdOffset = 0; if (LdBasePtr.getBase().getNode()) { // The base ptr must be the same. if (!LdBasePtr.equalBaseIndex(LdPtr, DAG, LdOffset)) break; } else { // Check that all other base pointers are the same as this one. LdBasePtr = LdPtr; } // We found a potential memory operand to merge. LoadNodes.push_back(MemOpLink(Ld, LdOffset)); } while (NumConsecutiveStores >= 2 && LoadNodes.size() >= 2) { Align RequiredAlignment; bool NeedRotate = false; if (LoadNodes.size() == 2) { // If we have load/store pair instructions and we only have two values, // don't bother merging. if (TLI.hasPairedLoad(MemVT, RequiredAlignment) && StoreNodes[0].MemNode->getAlign() >= RequiredAlignment) { StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + 2); LoadNodes.erase(LoadNodes.begin(), LoadNodes.begin() + 2); break; } // If the loads are reversed, see if we can rotate the halves into place. int64_t Offset0 = LoadNodes[0].OffsetFromBase; int64_t Offset1 = LoadNodes[1].OffsetFromBase; EVT PairVT = EVT::getIntegerVT(Context, ElementSizeBytes * 8 * 2); if (Offset0 - Offset1 == ElementSizeBytes && (hasOperation(ISD::ROTL, PairVT) || hasOperation(ISD::ROTR, PairVT))) { std::swap(LoadNodes[0], LoadNodes[1]); NeedRotate = true; } } LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode; unsigned FirstStoreAS = FirstInChain->getAddressSpace(); Align FirstStoreAlign = FirstInChain->getAlign(); LoadSDNode *FirstLoad = cast(LoadNodes[0].MemNode); // Scan the memory operations on the chain and find the first // non-consecutive load memory address. These variables hold the index in // the store node array. unsigned LastConsecutiveLoad = 1; // This variable refers to the size and not index in the array. unsigned LastLegalVectorType = 1; unsigned LastLegalIntegerType = 1; bool isDereferenceable = true; bool DoIntegerTruncate = false; int64_t StartAddress = LoadNodes[0].OffsetFromBase; SDValue LoadChain = FirstLoad->getChain(); for (unsigned i = 1; i < LoadNodes.size(); ++i) { // All loads must share the same chain. if (LoadNodes[i].MemNode->getChain() != LoadChain) break; int64_t CurrAddress = LoadNodes[i].OffsetFromBase; if (CurrAddress - StartAddress != (ElementSizeBytes * i)) break; LastConsecutiveLoad = i; if (isDereferenceable && !LoadNodes[i].MemNode->isDereferenceable()) isDereferenceable = false; // Find a legal type for the vector store. unsigned Elts = (i + 1) * NumMemElts; EVT StoreTy = EVT::getVectorVT(Context, MemVT.getScalarType(), Elts); // Break early when size is too large to be legal. if (StoreTy.getSizeInBits() > MaximumLegalStoreInBits) break; bool IsFastSt = false; bool IsFastLd = false; // Don't try vector types if we need a rotate. We may still fail the // legality checks for the integer type, but we can't handle the rotate // case with vectors. // FIXME: We could use a shuffle in place of the rotate. if (!NeedRotate && TLI.isTypeLegal(StoreTy) && TLI.canMergeStoresTo(FirstStoreAS, StoreTy, DAG.getMachineFunction()) && TLI.allowsMemoryAccess(Context, DL, StoreTy, *FirstInChain->getMemOperand(), &IsFastSt) && IsFastSt && TLI.allowsMemoryAccess(Context, DL, StoreTy, *FirstLoad->getMemOperand(), &IsFastLd) && IsFastLd) { LastLegalVectorType = i + 1; } // Find a legal type for the integer store. unsigned SizeInBits = (i + 1) * ElementSizeBytes * 8; StoreTy = EVT::getIntegerVT(Context, SizeInBits); if (TLI.isTypeLegal(StoreTy) && TLI.canMergeStoresTo(FirstStoreAS, StoreTy, DAG.getMachineFunction()) && TLI.allowsMemoryAccess(Context, DL, StoreTy, *FirstInChain->getMemOperand(), &IsFastSt) && IsFastSt && TLI.allowsMemoryAccess(Context, DL, StoreTy, *FirstLoad->getMemOperand(), &IsFastLd) && IsFastLd) { LastLegalIntegerType = i + 1; DoIntegerTruncate = false; // Or check whether a truncstore and extload is legal. } else if (TLI.getTypeAction(Context, StoreTy) == TargetLowering::TypePromoteInteger) { EVT LegalizedStoredValTy = TLI.getTypeToTransformTo(Context, StoreTy); if (TLI.isTruncStoreLegal(LegalizedStoredValTy, StoreTy) && TLI.canMergeStoresTo(FirstStoreAS, LegalizedStoredValTy, DAG.getMachineFunction()) && TLI.isLoadExtLegal(ISD::ZEXTLOAD, LegalizedStoredValTy, StoreTy) && TLI.isLoadExtLegal(ISD::SEXTLOAD, LegalizedStoredValTy, StoreTy) && TLI.isLoadExtLegal(ISD::EXTLOAD, LegalizedStoredValTy, StoreTy) && TLI.allowsMemoryAccess(Context, DL, StoreTy, *FirstInChain->getMemOperand(), &IsFastSt) && IsFastSt && TLI.allowsMemoryAccess(Context, DL, StoreTy, *FirstLoad->getMemOperand(), &IsFastLd) && IsFastLd) { LastLegalIntegerType = i + 1; DoIntegerTruncate = true; } } } // Only use vector types if the vector type is larger than the integer // type. If they are the same, use integers. bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && AllowVectors; unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType); // We add +1 here because the LastXXX variables refer to location while // the NumElem refers to array/index size. unsigned NumElem = std::min(NumConsecutiveStores, LastConsecutiveLoad + 1); NumElem = std::min(LastLegalType, NumElem); Align FirstLoadAlign = FirstLoad->getAlign(); if (NumElem < 2) { // We know that candidate stores are in order and of correct // shape. While there is no mergeable sequence from the // beginning one may start later in the sequence. The only // reason a merge of size N could have failed where another of // the same size would not have is if the alignment or either // the load or store has improved. Drop as many candidates as we // can here. unsigned NumSkip = 1; while ((NumSkip < LoadNodes.size()) && (LoadNodes[NumSkip].MemNode->getAlign() <= FirstLoadAlign) && (StoreNodes[NumSkip].MemNode->getAlign() <= FirstStoreAlign)) NumSkip++; StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + NumSkip); LoadNodes.erase(LoadNodes.begin(), LoadNodes.begin() + NumSkip); NumConsecutiveStores -= NumSkip; continue; } // Check that we can merge these candidates without causing a cycle. if (!checkMergeStoreCandidatesForDependencies(StoreNodes, NumElem, RootNode)) { StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + NumElem); LoadNodes.erase(LoadNodes.begin(), LoadNodes.begin() + NumElem); NumConsecutiveStores -= NumElem; continue; } // Find if it is better to use vectors or integers to load and store // to memory. EVT JointMemOpVT; if (UseVectorTy) { // Find a legal type for the vector store. unsigned Elts = NumElem * NumMemElts; JointMemOpVT = EVT::getVectorVT(Context, MemVT.getScalarType(), Elts); } else { unsigned SizeInBits = NumElem * ElementSizeBytes * 8; JointMemOpVT = EVT::getIntegerVT(Context, SizeInBits); } SDLoc LoadDL(LoadNodes[0].MemNode); SDLoc StoreDL(StoreNodes[0].MemNode); // The merged loads are required to have the same incoming chain, so // using the first's chain is acceptable. SDValue NewStoreChain = getMergeStoreChains(StoreNodes, NumElem); AddToWorklist(NewStoreChain.getNode()); MachineMemOperand::Flags LdMMOFlags = isDereferenceable ? MachineMemOperand::MODereferenceable : MachineMemOperand::MONone; if (IsNonTemporalLoad) LdMMOFlags |= MachineMemOperand::MONonTemporal; MachineMemOperand::Flags StMMOFlags = IsNonTemporalStore ? MachineMemOperand::MONonTemporal : MachineMemOperand::MONone; SDValue NewLoad, NewStore; if (UseVectorTy || !DoIntegerTruncate) { NewLoad = DAG.getLoad( JointMemOpVT, LoadDL, FirstLoad->getChain(), FirstLoad->getBasePtr(), FirstLoad->getPointerInfo(), FirstLoadAlign, LdMMOFlags); SDValue StoreOp = NewLoad; if (NeedRotate) { unsigned LoadWidth = ElementSizeBytes * 8 * 2; assert(JointMemOpVT == EVT::getIntegerVT(Context, LoadWidth) && "Unexpected type for rotate-able load pair"); SDValue RotAmt = DAG.getShiftAmountConstant(LoadWidth / 2, JointMemOpVT, LoadDL); // Target can convert to the identical ROTR if it does not have ROTL. StoreOp = DAG.getNode(ISD::ROTL, LoadDL, JointMemOpVT, NewLoad, RotAmt); } NewStore = DAG.getStore( NewStoreChain, StoreDL, StoreOp, FirstInChain->getBasePtr(), FirstInChain->getPointerInfo(), FirstStoreAlign, StMMOFlags); } else { // This must be the truncstore/extload case EVT ExtendedTy = TLI.getTypeToTransformTo(*DAG.getContext(), JointMemOpVT); NewLoad = DAG.getExtLoad(ISD::EXTLOAD, LoadDL, ExtendedTy, FirstLoad->getChain(), FirstLoad->getBasePtr(), FirstLoad->getPointerInfo(), JointMemOpVT, FirstLoadAlign, LdMMOFlags); NewStore = DAG.getTruncStore( NewStoreChain, StoreDL, NewLoad, FirstInChain->getBasePtr(), FirstInChain->getPointerInfo(), JointMemOpVT, FirstInChain->getAlign(), FirstInChain->getMemOperand()->getFlags()); } // Transfer chain users from old loads to the new load. for (unsigned i = 0; i < NumElem; ++i) { LoadSDNode *Ld = cast(LoadNodes[i].MemNode); DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), SDValue(NewLoad.getNode(), 1)); } // Replace all stores with the new store. Recursively remove corresponding // values if they are no longer used. for (unsigned i = 0; i < NumElem; ++i) { SDValue Val = StoreNodes[i].MemNode->getOperand(1); CombineTo(StoreNodes[i].MemNode, NewStore); if (Val->use_empty()) recursivelyDeleteUnusedNodes(Val.getNode()); } MadeChange = true; StoreNodes.erase(StoreNodes.begin(), StoreNodes.begin() + NumElem); LoadNodes.erase(LoadNodes.begin(), LoadNodes.begin() + NumElem); NumConsecutiveStores -= NumElem; } return MadeChange; } bool DAGCombiner::mergeConsecutiveStores(StoreSDNode *St) { if (OptLevel == CodeGenOpt::None || !EnableStoreMerging) return false; // TODO: Extend this function to merge stores of scalable vectors. // (i.e. two stores can be merged to one // store since we know is exactly twice as large as // ). Until then, bail out for scalable vectors. EVT MemVT = St->getMemoryVT(); if (MemVT.isScalableVector()) return false; if (!MemVT.isSimple() || MemVT.getSizeInBits() * 2 > MaximumLegalStoreInBits) return false; // This function cannot currently deal with non-byte-sized memory sizes. int64_t ElementSizeBytes = MemVT.getStoreSize(); if (ElementSizeBytes * 8 != (int64_t)MemVT.getSizeInBits()) return false; // Do not bother looking at stored values that are not constants, loads, or // extracted vector elements. SDValue StoredVal = peekThroughBitcasts(St->getValue()); const StoreSource StoreSrc = getStoreSource(StoredVal); if (StoreSrc == StoreSource::Unknown) return false; SmallVector StoreNodes; SDNode *RootNode; // Find potential store merge candidates by searching through chain sub-DAG getStoreMergeCandidates(St, StoreNodes, RootNode); // Check if there is anything to merge. if (StoreNodes.size() < 2) return false; // Sort the memory operands according to their distance from the // base pointer. llvm::sort(StoreNodes, [](MemOpLink LHS, MemOpLink RHS) { return LHS.OffsetFromBase < RHS.OffsetFromBase; }); bool AllowVectors = !DAG.getMachineFunction().getFunction().hasFnAttribute( Attribute::NoImplicitFloat); bool IsNonTemporalStore = St->isNonTemporal(); bool IsNonTemporalLoad = StoreSrc == StoreSource::Load && cast(StoredVal)->isNonTemporal(); // Store Merge attempts to merge the lowest stores. This generally // works out as if successful, as the remaining stores are checked // after the first collection of stores is merged. However, in the // case that a non-mergeable store is found first, e.g., {p[-2], // p[0], p[1], p[2], p[3]}, we would fail and miss the subsequent // mergeable cases. To prevent this, we prune such stores from the // front of StoreNodes here. bool MadeChange = false; while (StoreNodes.size() > 1) { unsigned NumConsecutiveStores = getConsecutiveStores(StoreNodes, ElementSizeBytes); // There are no more stores in the list to examine. if (NumConsecutiveStores == 0) return MadeChange; // We have at least 2 consecutive stores. Try to merge them. assert(NumConsecutiveStores >= 2 && "Expected at least 2 stores"); switch (StoreSrc) { case StoreSource::Constant: MadeChange |= tryStoreMergeOfConstants(StoreNodes, NumConsecutiveStores, MemVT, RootNode, AllowVectors); break; case StoreSource::Extract: MadeChange |= tryStoreMergeOfExtracts(StoreNodes, NumConsecutiveStores, MemVT, RootNode); break; case StoreSource::Load: MadeChange |= tryStoreMergeOfLoads(StoreNodes, NumConsecutiveStores, MemVT, RootNode, AllowVectors, IsNonTemporalStore, IsNonTemporalLoad); break; default: llvm_unreachable("Unhandled store source type"); } } return MadeChange; } SDValue DAGCombiner::replaceStoreChain(StoreSDNode *ST, SDValue BetterChain) { SDLoc SL(ST); SDValue ReplStore; // Replace the chain to avoid dependency. if (ST->isTruncatingStore()) { ReplStore = DAG.getTruncStore(BetterChain, SL, ST->getValue(), ST->getBasePtr(), ST->getMemoryVT(), ST->getMemOperand()); } else { ReplStore = DAG.getStore(BetterChain, SL, ST->getValue(), ST->getBasePtr(), ST->getMemOperand()); } // Create token to keep both nodes around. SDValue Token = DAG.getNode(ISD::TokenFactor, SL, MVT::Other, ST->getChain(), ReplStore); // Make sure the new and old chains are cleaned up. AddToWorklist(Token.getNode()); // Don't add users to work list. return CombineTo(ST, Token, false); } SDValue DAGCombiner::replaceStoreOfFPConstant(StoreSDNode *ST) { SDValue Value = ST->getValue(); if (Value.getOpcode() == ISD::TargetConstantFP) return SDValue(); if (!ISD::isNormalStore(ST)) return SDValue(); SDLoc DL(ST); SDValue Chain = ST->getChain(); SDValue Ptr = ST->getBasePtr(); const ConstantFPSDNode *CFP = cast(Value); // NOTE: If the original store is volatile, this transform must not increase // the number of stores. For example, on x86-32 an f64 can be stored in one // processor operation but an i64 (which is not legal) requires two. So the // transform should not be done in this case. SDValue Tmp; switch (CFP->getSimpleValueType(0).SimpleTy) { default: llvm_unreachable("Unknown FP type"); case MVT::f16: // We don't do this for these yet. case MVT::bf16: case MVT::f80: case MVT::f128: case MVT::ppcf128: return SDValue(); case MVT::f32: if ((isTypeLegal(MVT::i32) && !LegalOperations && ST->isSimple()) || TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) { Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF(). bitcastToAPInt().getZExtValue(), SDLoc(CFP), MVT::i32); return DAG.getStore(Chain, DL, Tmp, Ptr, ST->getMemOperand()); } return SDValue(); case MVT::f64: if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations && ST->isSimple()) || TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) { Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt(). getZExtValue(), SDLoc(CFP), MVT::i64); return DAG.getStore(Chain, DL, Tmp, Ptr, ST->getMemOperand()); } if (ST->isSimple() && TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) { // Many FP stores are not made apparent until after legalize, e.g. for // argument passing. Since this is so common, custom legalize the // 64-bit integer store into two 32-bit stores. uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue(); SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, SDLoc(CFP), MVT::i32); SDValue Hi = DAG.getConstant(Val >> 32, SDLoc(CFP), MVT::i32); if (DAG.getDataLayout().isBigEndian()) std::swap(Lo, Hi); MachineMemOperand::Flags MMOFlags = ST->getMemOperand()->getFlags(); AAMDNodes AAInfo = ST->getAAInfo(); SDValue St0 = DAG.getStore(Chain, DL, Lo, Ptr, ST->getPointerInfo(), ST->getOriginalAlign(), MMOFlags, AAInfo); Ptr = DAG.getMemBasePlusOffset(Ptr, TypeSize::Fixed(4), DL); SDValue St1 = DAG.getStore(Chain, DL, Hi, Ptr, ST->getPointerInfo().getWithOffset(4), ST->getOriginalAlign(), MMOFlags, AAInfo); return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, St0, St1); } return SDValue(); } } SDValue DAGCombiner::visitSTORE(SDNode *N) { StoreSDNode *ST = cast(N); SDValue Chain = ST->getChain(); SDValue Value = ST->getValue(); SDValue Ptr = ST->getBasePtr(); // If this is a store of a bit convert, store the input value if the // resultant store does not need a higher alignment than the original. if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() && ST->isUnindexed()) { EVT SVT = Value.getOperand(0).getValueType(); // If the store is volatile, we only want to change the store type if the // resulting store is legal. Otherwise we might increase the number of // memory accesses. We don't care if the original type was legal or not // as we assume software couldn't rely on the number of accesses of an // illegal type. // TODO: May be able to relax for unordered atomics (see D66309) if (((!LegalOperations && ST->isSimple()) || TLI.isOperationLegal(ISD::STORE, SVT)) && TLI.isStoreBitCastBeneficial(Value.getValueType(), SVT, DAG, *ST->getMemOperand())) { return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0), Ptr, ST->getMemOperand()); } } // Turn 'store undef, Ptr' -> nothing. if (Value.isUndef() && ST->isUnindexed()) return Chain; // Try to infer better alignment information than the store already has. if (OptLevel != CodeGenOpt::None && ST->isUnindexed() && !ST->isAtomic()) { if (MaybeAlign Alignment = DAG.InferPtrAlign(Ptr)) { if (*Alignment > ST->getAlign() && isAligned(*Alignment, ST->getSrcValueOffset())) { SDValue NewStore = DAG.getTruncStore(Chain, SDLoc(N), Value, Ptr, ST->getPointerInfo(), ST->getMemoryVT(), *Alignment, ST->getMemOperand()->getFlags(), ST->getAAInfo()); // NewStore will always be N as we are only refining the alignment assert(NewStore.getNode() == N); (void)NewStore; } } } // Try transforming a pair floating point load / store ops to integer // load / store ops. if (SDValue NewST = TransformFPLoadStorePair(N)) return NewST; // Try transforming several stores into STORE (BSWAP). if (SDValue Store = mergeTruncStores(ST)) return Store; if (ST->isUnindexed()) { // Walk up chain skipping non-aliasing memory nodes, on this store and any // adjacent stores. if (findBetterNeighborChains(ST)) { // replaceStoreChain uses CombineTo, which handled all of the worklist // manipulation. Return the original node to not do anything else. return SDValue(ST, 0); } Chain = ST->getChain(); } // FIXME: is there such a thing as a truncating indexed store? if (ST->isTruncatingStore() && ST->isUnindexed() && Value.getValueType().isInteger() && (!isa(Value) || !cast(Value)->isOpaque())) { // Convert a truncating store of a extension into a standard store. if ((Value.getOpcode() == ISD::ZERO_EXTEND || Value.getOpcode() == ISD::SIGN_EXTEND || Value.getOpcode() == ISD::ANY_EXTEND) && Value.getOperand(0).getValueType() == ST->getMemoryVT() && TLI.isOperationLegalOrCustom(ISD::STORE, ST->getMemoryVT())) return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0), Ptr, ST->getMemOperand()); APInt TruncDemandedBits = APInt::getLowBitsSet(Value.getScalarValueSizeInBits(), ST->getMemoryVT().getScalarSizeInBits()); // See if we can simplify the input to this truncstore with knowledge that // only the low bits are being used. For example: // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8" AddToWorklist(Value.getNode()); if (SDValue Shorter = DAG.GetDemandedBits(Value, TruncDemandedBits)) return DAG.getTruncStore(Chain, SDLoc(N), Shorter, Ptr, ST->getMemoryVT(), ST->getMemOperand()); // Otherwise, see if we can simplify the operation with // SimplifyDemandedBits, which only works if the value has a single use. if (SimplifyDemandedBits(Value, TruncDemandedBits)) { // Re-visit the store if anything changed and the store hasn't been merged // with another node (N is deleted) SimplifyDemandedBits will add Value's // node back to the worklist if necessary, but we also need to re-visit // the Store node itself. if (N->getOpcode() != ISD::DELETED_NODE) AddToWorklist(N); return SDValue(N, 0); } } // If this is a load followed by a store to the same location, then the store // is dead/noop. // TODO: Can relax for unordered atomics (see D66309) if (LoadSDNode *Ld = dyn_cast(Value)) { if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() && ST->isUnindexed() && ST->isSimple() && Ld->getAddressSpace() == ST->getAddressSpace() && // There can't be any side effects between the load and store, such as // a call or store. Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) { // The store is dead, remove it. return Chain; } } // TODO: Can relax for unordered atomics (see D66309) if (StoreSDNode *ST1 = dyn_cast(Chain)) { if (ST->isUnindexed() && ST->isSimple() && ST1->isUnindexed() && ST1->isSimple()) { if (OptLevel != CodeGenOpt::None && ST1->getBasePtr() == Ptr && ST1->getValue() == Value && ST->getMemoryVT() == ST1->getMemoryVT() && ST->getAddressSpace() == ST1->getAddressSpace()) { // If this is a store followed by a store with the same value to the // same location, then the store is dead/noop. return Chain; } if (OptLevel != CodeGenOpt::None && ST1->hasOneUse() && !ST1->getBasePtr().isUndef() && // BaseIndexOffset and the code below requires knowing the size // of a vector, so bail out if MemoryVT is scalable. !ST->getMemoryVT().isScalableVector() && !ST1->getMemoryVT().isScalableVector() && ST->getAddressSpace() == ST1->getAddressSpace()) { const BaseIndexOffset STBase = BaseIndexOffset::match(ST, DAG); const BaseIndexOffset ChainBase = BaseIndexOffset::match(ST1, DAG); unsigned STBitSize = ST->getMemoryVT().getFixedSizeInBits(); unsigned ChainBitSize = ST1->getMemoryVT().getFixedSizeInBits(); // If this is a store who's preceding store to a subset of the current // location and no one other node is chained to that store we can // effectively drop the store. Do not remove stores to undef as they may // be used as data sinks. if (STBase.contains(DAG, STBitSize, ChainBase, ChainBitSize)) { CombineTo(ST1, ST1->getChain()); return SDValue(); } } } } // If this is an FP_ROUND or TRUNC followed by a store, fold this into a // truncating store. We can do this even if this is already a truncstore. if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE) && Value->hasOneUse() && ST->isUnindexed() && TLI.canCombineTruncStore(Value.getOperand(0).getValueType(), ST->getMemoryVT(), LegalOperations)) { return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0), Ptr, ST->getMemoryVT(), ST->getMemOperand()); } // Always perform this optimization before types are legal. If the target // prefers, also try this after legalization to catch stores that were created // by intrinsics or other nodes. if (!LegalTypes || (TLI.mergeStoresAfterLegalization(ST->getMemoryVT()))) { while (true) { // There can be multiple store sequences on the same chain. // Keep trying to merge store sequences until we are unable to do so // or until we merge the last store on the chain. bool Changed = mergeConsecutiveStores(ST); if (!Changed) break; // Return N as merge only uses CombineTo and no worklist clean // up is necessary. if (N->getOpcode() == ISD::DELETED_NODE || !isa(N)) return SDValue(N, 0); } } // Try transforming N to an indexed store. if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N)) return SDValue(N, 0); // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr' // // Make sure to do this only after attempting to merge stores in order to // avoid changing the types of some subset of stores due to visit order, // preventing their merging. if (isa(ST->getValue())) { if (SDValue NewSt = replaceStoreOfFPConstant(ST)) return NewSt; } if (SDValue NewSt = splitMergedValStore(ST)) return NewSt; return ReduceLoadOpStoreWidth(N); } SDValue DAGCombiner::visitLIFETIME_END(SDNode *N) { const auto *LifetimeEnd = cast(N); if (!LifetimeEnd->hasOffset()) return SDValue(); const BaseIndexOffset LifetimeEndBase(N->getOperand(1), SDValue(), LifetimeEnd->getOffset(), false); // We walk up the chains to find stores. SmallVector Chains = {N->getOperand(0)}; while (!Chains.empty()) { SDValue Chain = Chains.pop_back_val(); if (!Chain.hasOneUse()) continue; switch (Chain.getOpcode()) { case ISD::TokenFactor: for (unsigned Nops = Chain.getNumOperands(); Nops;) Chains.push_back(Chain.getOperand(--Nops)); break; case ISD::LIFETIME_START: case ISD::LIFETIME_END: // We can forward past any lifetime start/end that can be proven not to // alias the node. if (!mayAlias(Chain.getNode(), N)) Chains.push_back(Chain.getOperand(0)); break; case ISD::STORE: { StoreSDNode *ST = dyn_cast(Chain); // TODO: Can relax for unordered atomics (see D66309) if (!ST->isSimple() || ST->isIndexed()) continue; const TypeSize StoreSize = ST->getMemoryVT().getStoreSize(); // The bounds of a scalable store are not known until runtime, so this // store cannot be elided. if (StoreSize.isScalable()) continue; const BaseIndexOffset StoreBase = BaseIndexOffset::match(ST, DAG); // If we store purely within object bounds just before its lifetime ends, // we can remove the store. if (LifetimeEndBase.contains(DAG, LifetimeEnd->getSize() * 8, StoreBase, StoreSize.getFixedSize() * 8)) { LLVM_DEBUG(dbgs() << "\nRemoving store:"; StoreBase.dump(); dbgs() << "\nwithin LIFETIME_END of : "; LifetimeEndBase.dump(); dbgs() << "\n"); CombineTo(ST, ST->getChain()); return SDValue(N, 0); } } } } return SDValue(); } /// For the instruction sequence of store below, F and I values /// are bundled together as an i64 value before being stored into memory. /// Sometimes it is more efficent to generate separate stores for F and I, /// which can remove the bitwise instructions or sink them to colder places. /// /// (store (or (zext (bitcast F to i32) to i64), /// (shl (zext I to i64), 32)), addr) --> /// (store F, addr) and (store I, addr+4) /// /// Similarly, splitting for other merged store can also be beneficial, like: /// For pair of {i32, i32}, i64 store --> two i32 stores. /// For pair of {i32, i16}, i64 store --> two i32 stores. /// For pair of {i16, i16}, i32 store --> two i16 stores. /// For pair of {i16, i8}, i32 store --> two i16 stores. /// For pair of {i8, i8}, i16 store --> two i8 stores. /// /// We allow each target to determine specifically which kind of splitting is /// supported. /// /// The store patterns are commonly seen from the simple code snippet below /// if only std::make_pair(...) is sroa transformed before inlined into hoo. /// void goo(const std::pair &); /// hoo() { /// ... /// goo(std::make_pair(tmp, ftmp)); /// ... /// } /// SDValue DAGCombiner::splitMergedValStore(StoreSDNode *ST) { if (OptLevel == CodeGenOpt::None) return SDValue(); // Can't change the number of memory accesses for a volatile store or break // atomicity for an atomic one. if (!ST->isSimple()) return SDValue(); SDValue Val = ST->getValue(); SDLoc DL(ST); // Match OR operand. if (!Val.getValueType().isScalarInteger() || Val.getOpcode() != ISD::OR) return SDValue(); // Match SHL operand and get Lower and Higher parts of Val. SDValue Op1 = Val.getOperand(0); SDValue Op2 = Val.getOperand(1); SDValue Lo, Hi; if (Op1.getOpcode() != ISD::SHL) { std::swap(Op1, Op2); if (Op1.getOpcode() != ISD::SHL) return SDValue(); } Lo = Op2; Hi = Op1.getOperand(0); if (!Op1.hasOneUse()) return SDValue(); // Match shift amount to HalfValBitSize. unsigned HalfValBitSize = Val.getValueSizeInBits() / 2; ConstantSDNode *ShAmt = dyn_cast(Op1.getOperand(1)); if (!ShAmt || ShAmt->getAPIntValue() != HalfValBitSize) return SDValue(); // Lo and Hi are zero-extended from int with size less equal than 32 // to i64. if (Lo.getOpcode() != ISD::ZERO_EXTEND || !Lo.hasOneUse() || !Lo.getOperand(0).getValueType().isScalarInteger() || Lo.getOperand(0).getValueSizeInBits() > HalfValBitSize || Hi.getOpcode() != ISD::ZERO_EXTEND || !Hi.hasOneUse() || !Hi.getOperand(0).getValueType().isScalarInteger() || Hi.getOperand(0).getValueSizeInBits() > HalfValBitSize) return SDValue(); // Use the EVT of low and high parts before bitcast as the input // of target query. EVT LowTy = (Lo.getOperand(0).getOpcode() == ISD::BITCAST) ? Lo.getOperand(0).getValueType() : Lo.getValueType(); EVT HighTy = (Hi.getOperand(0).getOpcode() == ISD::BITCAST) ? Hi.getOperand(0).getValueType() : Hi.getValueType(); if (!TLI.isMultiStoresCheaperThanBitsMerge(LowTy, HighTy)) return SDValue(); // Start to split store. MachineMemOperand::Flags MMOFlags = ST->getMemOperand()->getFlags(); AAMDNodes AAInfo = ST->getAAInfo(); // Change the sizes of Lo and Hi's value types to HalfValBitSize. EVT VT = EVT::getIntegerVT(*DAG.getContext(), HalfValBitSize); Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Lo.getOperand(0)); Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Hi.getOperand(0)); SDValue Chain = ST->getChain(); SDValue Ptr = ST->getBasePtr(); // Lower value store. SDValue St0 = DAG.getStore(Chain, DL, Lo, Ptr, ST->getPointerInfo(), ST->getOriginalAlign(), MMOFlags, AAInfo); Ptr = DAG.getMemBasePlusOffset(Ptr, TypeSize::Fixed(HalfValBitSize / 8), DL); // Higher value store. SDValue St1 = DAG.getStore( St0, DL, Hi, Ptr, ST->getPointerInfo().getWithOffset(HalfValBitSize / 8), ST->getOriginalAlign(), MMOFlags, AAInfo); return St1; } /// Convert a disguised subvector insertion into a shuffle: SDValue DAGCombiner::combineInsertEltToShuffle(SDNode *N, unsigned InsIndex) { assert(N->getOpcode() == ISD::INSERT_VECTOR_ELT && "Expected extract_vector_elt"); SDValue InsertVal = N->getOperand(1); SDValue Vec = N->getOperand(0); // (insert_vector_elt (vector_shuffle X, Y), (extract_vector_elt X, N), // InsIndex) // --> (vector_shuffle X, Y) and variations where shuffle operands may be // CONCAT_VECTORS. if (Vec.getOpcode() == ISD::VECTOR_SHUFFLE && Vec.hasOneUse() && InsertVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT && isa(InsertVal.getOperand(1))) { ShuffleVectorSDNode *SVN = cast(Vec.getNode()); ArrayRef Mask = SVN->getMask(); SDValue X = Vec.getOperand(0); SDValue Y = Vec.getOperand(1); // Vec's operand 0 is using indices from 0 to N-1 and // operand 1 from N to 2N - 1, where N is the number of // elements in the vectors. SDValue InsertVal0 = InsertVal.getOperand(0); int ElementOffset = -1; // We explore the inputs of the shuffle in order to see if we find the // source of the extract_vector_elt. If so, we can use it to modify the // shuffle rather than perform an insert_vector_elt. SmallVector, 8> ArgWorkList; ArgWorkList.emplace_back(Mask.size(), Y); ArgWorkList.emplace_back(0, X); while (!ArgWorkList.empty()) { int ArgOffset; SDValue ArgVal; std::tie(ArgOffset, ArgVal) = ArgWorkList.pop_back_val(); if (ArgVal == InsertVal0) { ElementOffset = ArgOffset; break; } // Peek through concat_vector. if (ArgVal.getOpcode() == ISD::CONCAT_VECTORS) { int CurrentArgOffset = ArgOffset + ArgVal.getValueType().getVectorNumElements(); int Step = ArgVal.getOperand(0).getValueType().getVectorNumElements(); for (SDValue Op : reverse(ArgVal->ops())) { CurrentArgOffset -= Step; ArgWorkList.emplace_back(CurrentArgOffset, Op); } // Make sure we went through all the elements and did not screw up index // computation. assert(CurrentArgOffset == ArgOffset); } } // If we failed to find a match, see if we can replace an UNDEF shuffle // operand. if (ElementOffset == -1 && Y.isUndef() && InsertVal0.getValueType() == Y.getValueType()) { ElementOffset = Mask.size(); Y = InsertVal0; } if (ElementOffset != -1) { SmallVector NewMask(Mask.begin(), Mask.end()); auto *ExtrIndex = cast(InsertVal.getOperand(1)); NewMask[InsIndex] = ElementOffset + ExtrIndex->getZExtValue(); assert(NewMask[InsIndex] < (int)(2 * Vec.getValueType().getVectorNumElements()) && NewMask[InsIndex] >= 0 && "NewMask[InsIndex] is out of bound"); SDValue LegalShuffle = TLI.buildLegalVectorShuffle(Vec.getValueType(), SDLoc(N), X, Y, NewMask, DAG); if (LegalShuffle) return LegalShuffle; } } // insert_vector_elt V, (bitcast X from vector type), IdxC --> // bitcast(shuffle (bitcast V), (extended X), Mask) // Note: We do not use an insert_subvector node because that requires a // legal subvector type. if (InsertVal.getOpcode() != ISD::BITCAST || !InsertVal.hasOneUse() || !InsertVal.getOperand(0).getValueType().isVector()) return SDValue(); SDValue SubVec = InsertVal.getOperand(0); SDValue DestVec = N->getOperand(0); EVT SubVecVT = SubVec.getValueType(); EVT VT = DestVec.getValueType(); unsigned NumSrcElts = SubVecVT.getVectorNumElements(); // If the source only has a single vector element, the cost of creating adding // it to a vector is likely to exceed the cost of a insert_vector_elt. if (NumSrcElts == 1) return SDValue(); unsigned ExtendRatio = VT.getSizeInBits() / SubVecVT.getSizeInBits(); unsigned NumMaskVals = ExtendRatio * NumSrcElts; // Step 1: Create a shuffle mask that implements this insert operation. The // vector that we are inserting into will be operand 0 of the shuffle, so // those elements are just 'i'. The inserted subvector is in the first // positions of operand 1 of the shuffle. Example: // insert v4i32 V, (v2i16 X), 2 --> shuffle v8i16 V', X', {0,1,2,3,8,9,6,7} SmallVector Mask(NumMaskVals); for (unsigned i = 0; i != NumMaskVals; ++i) { if (i / NumSrcElts == InsIndex) Mask[i] = (i % NumSrcElts) + NumMaskVals; else Mask[i] = i; } // Bail out if the target can not handle the shuffle we want to create. EVT SubVecEltVT = SubVecVT.getVectorElementType(); EVT ShufVT = EVT::getVectorVT(*DAG.getContext(), SubVecEltVT, NumMaskVals); if (!TLI.isShuffleMaskLegal(Mask, ShufVT)) return SDValue(); // Step 2: Create a wide vector from the inserted source vector by appending // undefined elements. This is the same size as our destination vector. SDLoc DL(N); SmallVector ConcatOps(ExtendRatio, DAG.getUNDEF(SubVecVT)); ConcatOps[0] = SubVec; SDValue PaddedSubV = DAG.getNode(ISD::CONCAT_VECTORS, DL, ShufVT, ConcatOps); // Step 3: Shuffle in the padded subvector. SDValue DestVecBC = DAG.getBitcast(ShufVT, DestVec); SDValue Shuf = DAG.getVectorShuffle(ShufVT, DL, DestVecBC, PaddedSubV, Mask); AddToWorklist(PaddedSubV.getNode()); AddToWorklist(DestVecBC.getNode()); AddToWorklist(Shuf.getNode()); return DAG.getBitcast(VT, Shuf); } SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) { SDValue InVec = N->getOperand(0); SDValue InVal = N->getOperand(1); SDValue EltNo = N->getOperand(2); SDLoc DL(N); EVT VT = InVec.getValueType(); auto *IndexC = dyn_cast(EltNo); // Insert into out-of-bounds element is undefined. if (IndexC && VT.isFixedLengthVector() && IndexC->getZExtValue() >= VT.getVectorNumElements()) return DAG.getUNDEF(VT); // Remove redundant insertions: // (insert_vector_elt x (extract_vector_elt x idx) idx) -> x if (InVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT && InVec == InVal.getOperand(0) && EltNo == InVal.getOperand(1)) return InVec; if (!IndexC) { // If this is variable insert to undef vector, it might be better to splat: // inselt undef, InVal, EltNo --> build_vector < InVal, InVal, ... > if (InVec.isUndef() && TLI.shouldSplatInsEltVarIndex(VT)) { if (VT.isScalableVector()) return DAG.getSplatVector(VT, DL, InVal); SmallVector Ops(VT.getVectorNumElements(), InVal); return DAG.getBuildVector(VT, DL, Ops); } return SDValue(); } if (VT.isScalableVector()) return SDValue(); unsigned NumElts = VT.getVectorNumElements(); // We must know which element is being inserted for folds below here. unsigned Elt = IndexC->getZExtValue(); if (SDValue Shuf = combineInsertEltToShuffle(N, Elt)) return Shuf; // Handle <1 x ???> vector insertion special cases. if (NumElts == 1) { // insert_vector_elt(x, extract_vector_elt(y, 0), 0) -> y if (InVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT && InVal.getOperand(0).getValueType() == VT && isNullConstant(InVal.getOperand(1))) return InVal.getOperand(0); } // Canonicalize insert_vector_elt dag nodes. // Example: // (insert_vector_elt (insert_vector_elt A, Idx0), Idx1) // -> (insert_vector_elt (insert_vector_elt A, Idx1), Idx0) // // Do this only if the child insert_vector node has one use; also // do this only if indices are both constants and Idx1 < Idx0. if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT && InVec.hasOneUse() && isa(InVec.getOperand(2))) { unsigned OtherElt = InVec.getConstantOperandVal(2); if (Elt < OtherElt) { // Swap nodes. SDValue NewOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, InVec.getOperand(0), InVal, EltNo); AddToWorklist(NewOp.getNode()); return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(InVec.getNode()), VT, NewOp, InVec.getOperand(1), InVec.getOperand(2)); } } // Attempt to convert an insert_vector_elt chain into a legal build_vector. if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) { // vXi1 vector - we don't need to recurse. if (NumElts == 1) return DAG.getBuildVector(VT, DL, {InVal}); // If we haven't already collected the element, insert into the op list. EVT MaxEltVT = InVal.getValueType(); auto AddBuildVectorOp = [&](SmallVectorImpl &Ops, SDValue Elt, unsigned Idx) { if (!Ops[Idx]) { Ops[Idx] = Elt; if (VT.isInteger()) { EVT EltVT = Elt.getValueType(); MaxEltVT = MaxEltVT.bitsGE(EltVT) ? MaxEltVT : EltVT; } } }; // Ensure all the operands are the same value type, fill any missing // operands with UNDEF and create the BUILD_VECTOR. auto CanonicalizeBuildVector = [&](SmallVectorImpl &Ops) { assert(Ops.size() == NumElts && "Unexpected vector size"); for (SDValue &Op : Ops) { if (Op) Op = VT.isInteger() ? DAG.getAnyExtOrTrunc(Op, DL, MaxEltVT) : Op; else Op = DAG.getUNDEF(MaxEltVT); } return DAG.getBuildVector(VT, DL, Ops); }; SmallVector Ops(NumElts, SDValue()); Ops[Elt] = InVal; // Recurse up a INSERT_VECTOR_ELT chain to build a BUILD_VECTOR. for (SDValue CurVec = InVec; CurVec;) { // UNDEF - build new BUILD_VECTOR from already inserted operands. if (CurVec.isUndef()) return CanonicalizeBuildVector(Ops); // BUILD_VECTOR - insert unused operands and build new BUILD_VECTOR. if (CurVec.getOpcode() == ISD::BUILD_VECTOR && CurVec.hasOneUse()) { for (unsigned I = 0; I != NumElts; ++I) AddBuildVectorOp(Ops, CurVec.getOperand(I), I); return CanonicalizeBuildVector(Ops); } // SCALAR_TO_VECTOR - insert unused scalar and build new BUILD_VECTOR. if (CurVec.getOpcode() == ISD::SCALAR_TO_VECTOR && CurVec.hasOneUse()) { AddBuildVectorOp(Ops, CurVec.getOperand(0), 0); return CanonicalizeBuildVector(Ops); } // INSERT_VECTOR_ELT - insert operand and continue up the chain. if (CurVec.getOpcode() == ISD::INSERT_VECTOR_ELT && CurVec.hasOneUse()) if (auto *CurIdx = dyn_cast(CurVec.getOperand(2))) if (CurIdx->getAPIntValue().ult(NumElts)) { unsigned Idx = CurIdx->getZExtValue(); AddBuildVectorOp(Ops, CurVec.getOperand(1), Idx); // Found entire BUILD_VECTOR. if (all_of(Ops, [](SDValue Op) { return !!Op; })) return CanonicalizeBuildVector(Ops); CurVec = CurVec->getOperand(0); continue; } // Failed to find a match in the chain - bail. break; } } return SDValue(); } SDValue DAGCombiner::scalarizeExtractedVectorLoad(SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad) { assert(OriginalLoad->isSimple()); EVT ResultVT = EVE->getValueType(0); EVT VecEltVT = InVecVT.getVectorElementType(); // If the vector element type is not a multiple of a byte then we are unable // to correctly compute an address to load only the extracted element as a // scalar. if (!VecEltVT.isByteSized()) return SDValue(); ISD::LoadExtType ExtTy = ResultVT.bitsGT(VecEltVT) ? ISD::NON_EXTLOAD : ISD::EXTLOAD; if (!TLI.isOperationLegalOrCustom(ISD::LOAD, VecEltVT) || !TLI.shouldReduceLoadWidth(OriginalLoad, ExtTy, VecEltVT)) return SDValue(); Align Alignment = OriginalLoad->getAlign(); MachinePointerInfo MPI; SDLoc DL(EVE); if (auto *ConstEltNo = dyn_cast(EltNo)) { int Elt = ConstEltNo->getZExtValue(); unsigned PtrOff = VecEltVT.getSizeInBits() * Elt / 8; MPI = OriginalLoad->getPointerInfo().getWithOffset(PtrOff); Alignment = commonAlignment(Alignment, PtrOff); } else { // Discard the pointer info except the address space because the memory // operand can't represent this new access since the offset is variable. MPI = MachinePointerInfo(OriginalLoad->getPointerInfo().getAddrSpace()); Alignment = commonAlignment(Alignment, VecEltVT.getSizeInBits() / 8); } bool IsFast = false; if (!TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VecEltVT, OriginalLoad->getAddressSpace(), Alignment, OriginalLoad->getMemOperand()->getFlags(), &IsFast) || !IsFast) return SDValue(); SDValue NewPtr = TLI.getVectorElementPointer(DAG, OriginalLoad->getBasePtr(), InVecVT, EltNo); // We are replacing a vector load with a scalar load. The new load must have // identical memory op ordering to the original. SDValue Load; if (ResultVT.bitsGT(VecEltVT)) { // If the result type of vextract is wider than the load, then issue an // extending load instead. ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, ResultVT, VecEltVT) ? ISD::ZEXTLOAD : ISD::EXTLOAD; Load = DAG.getExtLoad(ExtType, DL, ResultVT, OriginalLoad->getChain(), NewPtr, MPI, VecEltVT, Alignment, OriginalLoad->getMemOperand()->getFlags(), OriginalLoad->getAAInfo()); DAG.makeEquivalentMemoryOrdering(OriginalLoad, Load); } else { // The result type is narrower or the same width as the vector element Load = DAG.getLoad(VecEltVT, DL, OriginalLoad->getChain(), NewPtr, MPI, Alignment, OriginalLoad->getMemOperand()->getFlags(), OriginalLoad->getAAInfo()); DAG.makeEquivalentMemoryOrdering(OriginalLoad, Load); if (ResultVT.bitsLT(VecEltVT)) Load = DAG.getNode(ISD::TRUNCATE, DL, ResultVT, Load); else Load = DAG.getBitcast(ResultVT, Load); } ++OpsNarrowed; return Load; } /// Transform a vector binary operation into a scalar binary operation by moving /// the math/logic after an extract element of a vector. static SDValue scalarizeExtractedBinop(SDNode *ExtElt, SelectionDAG &DAG, bool LegalOperations) { const TargetLowering &TLI = DAG.getTargetLoweringInfo(); SDValue Vec = ExtElt->getOperand(0); SDValue Index = ExtElt->getOperand(1); auto *IndexC = dyn_cast(Index); if (!IndexC || !TLI.isBinOp(Vec.getOpcode()) || !Vec.hasOneUse() || Vec->getNumValues() != 1) return SDValue(); // Targets may want to avoid this to prevent an expensive register transfer. if (!TLI.shouldScalarizeBinop(Vec)) return SDValue(); // Extracting an element of a vector constant is constant-folded, so this // transform is just replacing a vector op with a scalar op while moving the // extract. SDValue Op0 = Vec.getOperand(0); SDValue Op1 = Vec.getOperand(1); APInt SplatVal; if (isAnyConstantBuildVector(Op0, true) || ISD::isConstantSplatVector(Op0.getNode(), SplatVal) || isAnyConstantBuildVector(Op1, true) || ISD::isConstantSplatVector(Op1.getNode(), SplatVal)) { // extractelt (binop X, C), IndexC --> binop (extractelt X, IndexC), C' // extractelt (binop C, X), IndexC --> binop C', (extractelt X, IndexC) SDLoc DL(ExtElt); EVT VT = ExtElt->getValueType(0); SDValue Ext0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Op0, Index); SDValue Ext1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Op1, Index); return DAG.getNode(Vec.getOpcode(), DL, VT, Ext0, Ext1); } return SDValue(); } SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) { SDValue VecOp = N->getOperand(0); SDValue Index = N->getOperand(1); EVT ScalarVT = N->getValueType(0); EVT VecVT = VecOp.getValueType(); if (VecOp.isUndef()) return DAG.getUNDEF(ScalarVT); // extract_vector_elt (insert_vector_elt vec, val, idx), idx) -> val // // This only really matters if the index is non-constant since other combines // on the constant elements already work. SDLoc DL(N); if (VecOp.getOpcode() == ISD::INSERT_VECTOR_ELT && Index == VecOp.getOperand(2)) { SDValue Elt = VecOp.getOperand(1); return VecVT.isInteger() ? DAG.getAnyExtOrTrunc(Elt, DL, ScalarVT) : Elt; } // (vextract (scalar_to_vector val, 0) -> val if (VecOp.getOpcode() == ISD::SCALAR_TO_VECTOR) { // Only 0'th element of SCALAR_TO_VECTOR is defined. if (DAG.isKnownNeverZero(Index)) return DAG.getUNDEF(ScalarVT); // Check if the result type doesn't match the inserted element type. A // SCALAR_TO_VECTOR may truncate the inserted element and the // EXTRACT_VECTOR_ELT may widen the extracted vector. SDValue InOp = VecOp.getOperand(0); if (InOp.getValueType() != ScalarVT) { assert(InOp.getValueType().isInteger() && ScalarVT.isInteger() && InOp.getValueType().bitsGT(ScalarVT)); return DAG.getNode(ISD::TRUNCATE, DL, ScalarVT, InOp); } return InOp; } // extract_vector_elt of out-of-bounds element -> UNDEF auto *IndexC = dyn_cast(Index); if (IndexC && VecVT.isFixedLengthVector() && IndexC->getAPIntValue().uge(VecVT.getVectorNumElements())) return DAG.getUNDEF(ScalarVT); // extract_vector_elt (build_vector x, y), 1 -> y if (((IndexC && VecOp.getOpcode() == ISD::BUILD_VECTOR) || VecOp.getOpcode() == ISD::SPLAT_VECTOR) && TLI.isTypeLegal(VecVT) && (VecOp.hasOneUse() || TLI.aggressivelyPreferBuildVectorSources(VecVT))) { assert((VecOp.getOpcode() != ISD::BUILD_VECTOR || VecVT.isFixedLengthVector()) && "BUILD_VECTOR used for scalable vectors"); unsigned IndexVal = VecOp.getOpcode() == ISD::BUILD_VECTOR ? IndexC->getZExtValue() : 0; SDValue Elt = VecOp.getOperand(IndexVal); EVT InEltVT = Elt.getValueType(); // Sometimes build_vector's scalar input types do not match result type. if (ScalarVT == InEltVT) return Elt; // TODO: It may be useful to truncate if free if the build_vector implicitly // converts. } if (SDValue BO = scalarizeExtractedBinop(N, DAG, LegalOperations)) return BO; if (VecVT.isScalableVector()) return SDValue(); // All the code from this point onwards assumes fixed width vectors, but it's // possible that some of the combinations could be made to work for scalable // vectors too. unsigned NumElts = VecVT.getVectorNumElements(); unsigned VecEltBitWidth = VecVT.getScalarSizeInBits(); // TODO: These transforms should not require the 'hasOneUse' restriction, but // there are regressions on multiple targets without it. We can end up with a // mess of scalar and vector code if we reduce only part of the DAG to scalar. if (IndexC && VecOp.getOpcode() == ISD::BITCAST && VecVT.isInteger() && VecOp.hasOneUse()) { // The vector index of the LSBs of the source depend on the endian-ness. bool IsLE = DAG.getDataLayout().isLittleEndian(); unsigned ExtractIndex = IndexC->getZExtValue(); // extract_elt (v2i32 (bitcast i64:x)), BCTruncElt -> i32 (trunc i64:x) unsigned BCTruncElt = IsLE ? 0 : NumElts - 1; SDValue BCSrc = VecOp.getOperand(0); if (ExtractIndex == BCTruncElt && BCSrc.getValueType().isScalarInteger()) return DAG.getNode(ISD::TRUNCATE, DL, ScalarVT, BCSrc); if (LegalTypes && BCSrc.getValueType().isInteger() && BCSrc.getOpcode() == ISD::SCALAR_TO_VECTOR) { // ext_elt (bitcast (scalar_to_vec i64 X to v2i64) to v4i32), TruncElt --> // trunc i64 X to i32 SDValue X = BCSrc.getOperand(0); assert(X.getValueType().isScalarInteger() && ScalarVT.isScalarInteger() && "Extract element and scalar to vector can't change element type " "from FP to integer."); unsigned XBitWidth = X.getValueSizeInBits(); BCTruncElt = IsLE ? 0 : XBitWidth / VecEltBitWidth - 1; // An extract element return value type can be wider than its vector // operand element type. In that case, the high bits are undefined, so // it's possible that we may need to extend rather than truncate. if (ExtractIndex == BCTruncElt && XBitWidth > VecEltBitWidth) { assert(XBitWidth % VecEltBitWidth == 0 && "Scalar bitwidth must be a multiple of vector element bitwidth"); return DAG.getAnyExtOrTrunc(X, DL, ScalarVT); } } } // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT. // We only perform this optimization before the op legalization phase because // we may introduce new vector instructions which are not backed by TD // patterns. For example on AVX, extracting elements from a wide vector // without using extract_subvector. However, if we can find an underlying // scalar value, then we can always use that. if (IndexC && VecOp.getOpcode() == ISD::VECTOR_SHUFFLE) { auto *Shuf = cast(VecOp); // Find the new index to extract from. int OrigElt = Shuf->getMaskElt(IndexC->getZExtValue()); // Extracting an undef index is undef. if (OrigElt == -1) return DAG.getUNDEF(ScalarVT); // Select the right vector half to extract from. SDValue SVInVec; if (OrigElt < (int)NumElts) { SVInVec = VecOp.getOperand(0); } else { SVInVec = VecOp.getOperand(1); OrigElt -= NumElts; } if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) { SDValue InOp = SVInVec.getOperand(OrigElt); if (InOp.getValueType() != ScalarVT) { assert(InOp.getValueType().isInteger() && ScalarVT.isInteger()); InOp = DAG.getSExtOrTrunc(InOp, DL, ScalarVT); } return InOp; } // FIXME: We should handle recursing on other vector shuffles and // scalar_to_vector here as well. if (!LegalOperations || // FIXME: Should really be just isOperationLegalOrCustom. TLI.isOperationLegal(ISD::EXTRACT_VECTOR_ELT, VecVT) || TLI.isOperationExpand(ISD::VECTOR_SHUFFLE, VecVT)) { return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ScalarVT, SVInVec, DAG.getVectorIdxConstant(OrigElt, DL)); } } // If only EXTRACT_VECTOR_ELT nodes use the source vector we can // simplify it based on the (valid) extraction indices. if (llvm::all_of(VecOp->uses(), [&](SDNode *Use) { return Use->getOpcode() == ISD::EXTRACT_VECTOR_ELT && Use->getOperand(0) == VecOp && isa(Use->getOperand(1)); })) { APInt DemandedElts = APInt::getZero(NumElts); for (SDNode *Use : VecOp->uses()) { auto *CstElt = cast(Use->getOperand(1)); if (CstElt->getAPIntValue().ult(NumElts)) DemandedElts.setBit(CstElt->getZExtValue()); } if (SimplifyDemandedVectorElts(VecOp, DemandedElts, true)) { // We simplified the vector operand of this extract element. If this // extract is not dead, visit it again so it is folded properly. if (N->getOpcode() != ISD::DELETED_NODE) AddToWorklist(N); return SDValue(N, 0); } APInt DemandedBits = APInt::getAllOnes(VecEltBitWidth); if (SimplifyDemandedBits(VecOp, DemandedBits, DemandedElts, true)) { // We simplified the vector operand of this extract element. If this // extract is not dead, visit it again so it is folded properly. if (N->getOpcode() != ISD::DELETED_NODE) AddToWorklist(N); return SDValue(N, 0); } } // Everything under here is trying to match an extract of a loaded value. // If the result of load has to be truncated, then it's not necessarily // profitable. bool BCNumEltsChanged = false; EVT ExtVT = VecVT.getVectorElementType(); EVT LVT = ExtVT; if (ScalarVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, ScalarVT)) return SDValue(); if (VecOp.getOpcode() == ISD::BITCAST) { // Don't duplicate a load with other uses. if (!VecOp.hasOneUse()) return SDValue(); EVT BCVT = VecOp.getOperand(0).getValueType(); if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType())) return SDValue(); if (NumElts != BCVT.getVectorNumElements()) BCNumEltsChanged = true; VecOp = VecOp.getOperand(0); ExtVT = BCVT.getVectorElementType(); } // extract (vector load $addr), i --> load $addr + i * size if (!LegalOperations && !IndexC && VecOp.hasOneUse() && ISD::isNormalLoad(VecOp.getNode()) && !Index->hasPredecessor(VecOp.getNode())) { auto *VecLoad = dyn_cast(VecOp); if (VecLoad && VecLoad->isSimple()) return scalarizeExtractedVectorLoad(N, VecVT, Index, VecLoad); } // Perform only after legalization to ensure build_vector / vector_shuffle // optimizations have already been done. if (!LegalOperations || !IndexC) return SDValue(); // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size) // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size) // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr) int Elt = IndexC->getZExtValue(); LoadSDNode *LN0 = nullptr; if (ISD::isNormalLoad(VecOp.getNode())) { LN0 = cast(VecOp); } else if (VecOp.getOpcode() == ISD::SCALAR_TO_VECTOR && VecOp.getOperand(0).getValueType() == ExtVT && ISD::isNormalLoad(VecOp.getOperand(0).getNode())) { // Don't duplicate a load with other uses. if (!VecOp.hasOneUse()) return SDValue(); LN0 = cast(VecOp.getOperand(0)); } if (auto *Shuf = dyn_cast(VecOp)) { // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1) // => // (load $addr+1*size) // Don't duplicate a load with other uses. if (!VecOp.hasOneUse()) return SDValue(); // If the bit convert changed the number of elements, it is unsafe // to examine the mask. if (BCNumEltsChanged) return SDValue(); // Select the input vector, guarding against out of range extract vector. int Idx = (Elt > (int)NumElts) ? -1 : Shuf->getMaskElt(Elt); VecOp = (Idx < (int)NumElts) ? VecOp.getOperand(0) : VecOp.getOperand(1); if (VecOp.getOpcode() == ISD::BITCAST) { // Don't duplicate a load with other uses. if (!VecOp.hasOneUse()) return SDValue(); VecOp = VecOp.getOperand(0); } if (ISD::isNormalLoad(VecOp.getNode())) { LN0 = cast(VecOp); Elt = (Idx < (int)NumElts) ? Idx : Idx - (int)NumElts; Index = DAG.getConstant(Elt, DL, Index.getValueType()); } } else if (VecOp.getOpcode() == ISD::CONCAT_VECTORS && !BCNumEltsChanged && VecVT.getVectorElementType() == ScalarVT && (!LegalTypes || TLI.isTypeLegal( VecOp.getOperand(0).getValueType().getVectorElementType()))) { // extract_vector_elt (concat_vectors v2i16:a, v2i16:b), 0 // -> extract_vector_elt a, 0 // extract_vector_elt (concat_vectors v2i16:a, v2i16:b), 1 // -> extract_vector_elt a, 1 // extract_vector_elt (concat_vectors v2i16:a, v2i16:b), 2 // -> extract_vector_elt b, 0 // extract_vector_elt (concat_vectors v2i16:a, v2i16:b), 3 // -> extract_vector_elt b, 1 SDLoc SL(N); EVT ConcatVT = VecOp.getOperand(0).getValueType(); unsigned ConcatNumElts = ConcatVT.getVectorNumElements(); SDValue NewIdx = DAG.getConstant(Elt % ConcatNumElts, SL, Index.getValueType()); SDValue ConcatOp = VecOp.getOperand(Elt / ConcatNumElts); SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, ConcatVT.getVectorElementType(), ConcatOp, NewIdx); return DAG.getNode(ISD::BITCAST, SL, ScalarVT, Elt); } // Make sure we found a non-volatile load and the extractelement is // the only use. if (!LN0 || !LN0->hasNUsesOfValue(1,0) || !LN0->isSimple()) return SDValue(); // If Idx was -1 above, Elt is going to be -1, so just return undef. if (Elt == -1) return DAG.getUNDEF(LVT); return scalarizeExtractedVectorLoad(N, VecVT, Index, LN0); } // Simplify (build_vec (ext )) to (bitcast (build_vec )) SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) { // We perform this optimization post type-legalization because // the type-legalizer often scalarizes integer-promoted vectors. // Performing this optimization before may create bit-casts which // will be type-legalized to complex code sequences. // We perform this optimization only before the operation legalizer because we // may introduce illegal operations. if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes) return SDValue(); unsigned NumInScalars = N->getNumOperands(); SDLoc DL(N); EVT VT = N->getValueType(0); // Check to see if this is a BUILD_VECTOR of a bunch of values // which come from any_extend or zero_extend nodes. If so, we can create // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR // optimizations. We do not handle sign-extend because we can't fill the sign // using shuffles. EVT SourceType = MVT::Other; bool AllAnyExt = true; for (unsigned i = 0; i != NumInScalars; ++i) { SDValue In = N->getOperand(i); // Ignore undef inputs. if (In.isUndef()) continue; bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND; bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND; // Abort if the element is not an extension. if (!ZeroExt && !AnyExt) { SourceType = MVT::Other; break; } // The input is a ZeroExt or AnyExt. Check the original type. EVT InTy = In.getOperand(0).getValueType(); // Check that all of the widened source types are the same. if (SourceType == MVT::Other) // First time. SourceType = InTy; else if (InTy != SourceType) { // Multiple income types. Abort. SourceType = MVT::Other; break; } // Check if all of the extends are ANY_EXTENDs. AllAnyExt &= AnyExt; } // In order to have valid types, all of the inputs must be extended from the // same source type and all of the inputs must be any or zero extend. // Scalar sizes must be a power of two. EVT OutScalarTy = VT.getScalarType(); bool ValidTypes = SourceType != MVT::Other && isPowerOf2_32(OutScalarTy.getSizeInBits()) && isPowerOf2_32(SourceType.getSizeInBits()); // Create a new simpler BUILD_VECTOR sequence which other optimizations can // turn into a single shuffle instruction. if (!ValidTypes) return SDValue(); // If we already have a splat buildvector, then don't fold it if it means // introducing zeros. if (!AllAnyExt && DAG.isSplatValue(SDValue(N, 0), /*AllowUndefs*/ true)) return SDValue(); bool isLE = DAG.getDataLayout().isLittleEndian(); unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits(); assert(ElemRatio > 1 && "Invalid element size ratio"); SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType): DAG.getConstant(0, DL, SourceType); unsigned NewBVElems = ElemRatio * VT.getVectorNumElements(); SmallVector Ops(NewBVElems, Filler); // Populate the new build_vector for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { SDValue Cast = N->getOperand(i); assert((Cast.getOpcode() == ISD::ANY_EXTEND || Cast.getOpcode() == ISD::ZERO_EXTEND || Cast.isUndef()) && "Invalid cast opcode"); SDValue In; if (Cast.isUndef()) In = DAG.getUNDEF(SourceType); else In = Cast->getOperand(0); unsigned Index = isLE ? (i * ElemRatio) : (i * ElemRatio + (ElemRatio - 1)); assert(Index < Ops.size() && "Invalid index"); Ops[Index] = In; } // The type of the new BUILD_VECTOR node. EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems); assert(VecVT.getSizeInBits() == VT.getSizeInBits() && "Invalid vector size"); // Check if the new vector type is legal. if (!isTypeLegal(VecVT) || (!TLI.isOperationLegal(ISD::BUILD_VECTOR, VecVT) && TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))) return SDValue(); // Make the new BUILD_VECTOR. SDValue BV = DAG.getBuildVector(VecVT, DL, Ops); // The new BUILD_VECTOR node has the potential to be further optimized. AddToWorklist(BV.getNode()); // Bitcast to the desired type. return DAG.getBitcast(VT, BV); } // Simplify (build_vec (trunc $1) // (trunc (srl $1 half-width)) // (trunc (srl $1 (2 * half-width))) …) // to (bitcast $1) SDValue DAGCombiner::reduceBuildVecTruncToBitCast(SDNode *N) { assert(N->getOpcode() == ISD::BUILD_VECTOR && "Expected build vector"); // Only for little endian if (!DAG.getDataLayout().isLittleEndian()) return SDValue(); SDLoc DL(N); EVT VT = N->getValueType(0); EVT OutScalarTy = VT.getScalarType(); uint64_t ScalarTypeBitsize = OutScalarTy.getSizeInBits(); // Only for power of two types to be sure that bitcast works well if (!isPowerOf2_64(ScalarTypeBitsize)) return SDValue(); unsigned NumInScalars = N->getNumOperands(); // Look through bitcasts auto PeekThroughBitcast = [](SDValue Op) { if (Op.getOpcode() == ISD::BITCAST) return Op.getOperand(0); return Op; }; // The source value where all the parts are extracted. SDValue Src; for (unsigned i = 0; i != NumInScalars; ++i) { SDValue In = PeekThroughBitcast(N->getOperand(i)); // Ignore undef inputs. if (In.isUndef()) continue; if (In.getOpcode() != ISD::TRUNCATE) return SDValue(); In = PeekThroughBitcast(In.getOperand(0)); if (In.getOpcode() != ISD::SRL) { // For now only build_vec without shuffling, handle shifts here in the // future. if (i != 0) return SDValue(); Src = In; } else { // In is SRL SDValue part = PeekThroughBitcast(In.getOperand(0)); if (!Src) { Src = part; } else if (Src != part) { // Vector parts do not stem from the same variable return SDValue(); } SDValue ShiftAmtVal = In.getOperand(1); if (!isa(ShiftAmtVal)) return SDValue(); uint64_t ShiftAmt = In.getConstantOperandVal(1); // The extracted value is not extracted at the right position if (ShiftAmt != i * ScalarTypeBitsize) return SDValue(); } } // Only cast if the size is the same if (Src.getValueType().getSizeInBits() != VT.getSizeInBits()) return SDValue(); return DAG.getBitcast(VT, Src); } SDValue DAGCombiner::createBuildVecShuffle(const SDLoc &DL, SDNode *N, ArrayRef VectorMask, SDValue VecIn1, SDValue VecIn2, unsigned LeftIdx, bool DidSplitVec) { SDValue ZeroIdx = DAG.getVectorIdxConstant(0, DL); EVT VT = N->getValueType(0); EVT InVT1 = VecIn1.getValueType(); EVT InVT2 = VecIn2.getNode() ? VecIn2.getValueType() : InVT1; unsigned NumElems = VT.getVectorNumElements(); unsigned ShuffleNumElems = NumElems; // If we artificially split a vector in two already, then the offsets in the // operands will all be based off of VecIn1, even those in VecIn2. unsigned Vec2Offset = DidSplitVec ? 0 : InVT1.getVectorNumElements(); uint64_t VTSize = VT.getFixedSizeInBits(); uint64_t InVT1Size = InVT1.getFixedSizeInBits(); uint64_t InVT2Size = InVT2.getFixedSizeInBits(); assert(InVT2Size <= InVT1Size && "Inputs must be sorted to be in non-increasing vector size order."); // We can't generate a shuffle node with mismatched input and output types. // Try to make the types match the type of the output. if (InVT1 != VT || InVT2 != VT) { if ((VTSize % InVT1Size == 0) && InVT1 == InVT2) { // If the output vector length is a multiple of both input lengths, // we can concatenate them and pad the rest with undefs. unsigned NumConcats = VTSize / InVT1Size; assert(NumConcats >= 2 && "Concat needs at least two inputs!"); SmallVector ConcatOps(NumConcats, DAG.getUNDEF(InVT1)); ConcatOps[0] = VecIn1; ConcatOps[1] = VecIn2 ? VecIn2 : DAG.getUNDEF(InVT1); VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, ConcatOps); VecIn2 = SDValue(); } else if (InVT1Size == VTSize * 2) { if (!TLI.isExtractSubvectorCheap(VT, InVT1, NumElems)) return SDValue(); if (!VecIn2.getNode()) { // If we only have one input vector, and it's twice the size of the // output, split it in two. VecIn2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, VecIn1, DAG.getVectorIdxConstant(NumElems, DL)); VecIn1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, VecIn1, ZeroIdx); // Since we now have shorter input vectors, adjust the offset of the // second vector's start. Vec2Offset = NumElems; } else { assert(InVT2Size <= InVT1Size && "Second input is not going to be larger than the first one."); // VecIn1 is wider than the output, and we have another, possibly // smaller input. Pad the smaller input with undefs, shuffle at the // input vector width, and extract the output. // The shuffle type is different than VT, so check legality again. if (LegalOperations && !TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, InVT1)) return SDValue(); // Legalizing INSERT_SUBVECTOR is tricky - you basically have to // lower it back into a BUILD_VECTOR. So if the inserted type is // illegal, don't even try. if (InVT1 != InVT2) { if (!TLI.isTypeLegal(InVT2)) return SDValue(); VecIn2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, InVT1, DAG.getUNDEF(InVT1), VecIn2, ZeroIdx); } ShuffleNumElems = NumElems * 2; } } else if (InVT2Size * 2 == VTSize && InVT1Size == VTSize) { SmallVector ConcatOps(2, DAG.getUNDEF(InVT2)); ConcatOps[0] = VecIn2; VecIn2 = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, ConcatOps); } else { // TODO: Support cases where the length mismatch isn't exactly by a // factor of 2. // TODO: Move this check upwards, so that if we have bad type // mismatches, we don't create any DAG nodes. return SDValue(); } } // Initialize mask to undef. SmallVector Mask(ShuffleNumElems, -1); // Only need to run up to the number of elements actually used, not the // total number of elements in the shuffle - if we are shuffling a wider // vector, the high lanes should be set to undef. for (unsigned i = 0; i != NumElems; ++i) { if (VectorMask[i] <= 0) continue; unsigned ExtIndex = N->getOperand(i).getConstantOperandVal(1); if (VectorMask[i] == (int)LeftIdx) { Mask[i] = ExtIndex; } else if (VectorMask[i] == (int)LeftIdx + 1) { Mask[i] = Vec2Offset + ExtIndex; } } // The type the input vectors may have changed above. InVT1 = VecIn1.getValueType(); // If we already have a VecIn2, it should have the same type as VecIn1. // If we don't, get an undef/zero vector of the appropriate type. VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(InVT1); assert(InVT1 == VecIn2.getValueType() && "Unexpected second input type."); SDValue Shuffle = DAG.getVectorShuffle(InVT1, DL, VecIn1, VecIn2, Mask); if (ShuffleNumElems > NumElems) Shuffle = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Shuffle, ZeroIdx); return Shuffle; } static SDValue reduceBuildVecToShuffleWithZero(SDNode *BV, SelectionDAG &DAG) { assert(BV->getOpcode() == ISD::BUILD_VECTOR && "Expected build vector"); // First, determine where the build vector is not undef. // TODO: We could extend this to handle zero elements as well as undefs. int NumBVOps = BV->getNumOperands(); int ZextElt = -1; for (int i = 0; i != NumBVOps; ++i) { SDValue Op = BV->getOperand(i); if (Op.isUndef()) continue; if (ZextElt == -1) ZextElt = i; else return SDValue(); } // Bail out if there's no non-undef element. if (ZextElt == -1) return SDValue(); // The build vector contains some number of undef elements and exactly // one other element. That other element must be a zero-extended scalar // extracted from a vector at a constant index to turn this into a shuffle. // Also, require that the build vector does not implicitly truncate/extend // its elements. // TODO: This could be enhanced to allow ANY_EXTEND as well as ZERO_EXTEND. EVT VT = BV->getValueType(0); SDValue Zext = BV->getOperand(ZextElt); if (Zext.getOpcode() != ISD::ZERO_EXTEND || !Zext.hasOneUse() || Zext.getOperand(0).getOpcode() != ISD::EXTRACT_VECTOR_ELT || !isa(Zext.getOperand(0).getOperand(1)) || Zext.getValueSizeInBits() != VT.getScalarSizeInBits()) return SDValue(); // The zero-extend must be a multiple of the source size, and we must be // building a vector of the same size as the source of the extract element. SDValue Extract = Zext.getOperand(0); unsigned DestSize = Zext.getValueSizeInBits(); unsigned SrcSize = Extract.getValueSizeInBits(); if (DestSize % SrcSize != 0 || Extract.getOperand(0).getValueSizeInBits() != VT.getSizeInBits()) return SDValue(); // Create a shuffle mask that will combine the extracted element with zeros // and undefs. int ZextRatio = DestSize / SrcSize; int NumMaskElts = NumBVOps * ZextRatio; SmallVector ShufMask(NumMaskElts, -1); for (int i = 0; i != NumMaskElts; ++i) { if (i / ZextRatio == ZextElt) { // The low bits of the (potentially translated) extracted element map to // the source vector. The high bits map to zero. We will use a zero vector // as the 2nd source operand of the shuffle, so use the 1st element of // that vector (mask value is number-of-elements) for the high bits. if (i % ZextRatio == 0) ShufMask[i] = Extract.getConstantOperandVal(1); else ShufMask[i] = NumMaskElts; } // Undef elements of the build vector remain undef because we initialize // the shuffle mask with -1. } // buildvec undef, ..., (zext (extractelt V, IndexC)), undef... --> // bitcast (shuffle V, ZeroVec, VectorMask) SDLoc DL(BV); EVT VecVT = Extract.getOperand(0).getValueType(); SDValue ZeroVec = DAG.getConstant(0, DL, VecVT); const TargetLowering &TLI = DAG.getTargetLoweringInfo(); SDValue Shuf = TLI.buildLegalVectorShuffle(VecVT, DL, Extract.getOperand(0), ZeroVec, ShufMask, DAG); if (!Shuf) return SDValue(); return DAG.getBitcast(VT, Shuf); } // FIXME: promote to STLExtras. template static auto getFirstIndexOf(R &&Range, const T &Val) { auto I = find(Range, Val); if (I == Range.end()) return static_cast(-1); return std::distance(Range.begin(), I); } // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT // operations. If the types of the vectors we're extracting from allow it, // turn this into a vector_shuffle node. SDValue DAGCombiner::reduceBuildVecToShuffle(SDNode *N) { SDLoc DL(N); EVT VT = N->getValueType(0); // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes. if (!isTypeLegal(VT)) return SDValue(); if (SDValue V = reduceBuildVecToShuffleWithZero(N, DAG)) return V; // May only combine to shuffle after legalize if shuffle is legal. if (LegalOperations && !TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, VT)) return SDValue(); bool UsesZeroVector = false; unsigned NumElems = N->getNumOperands(); // Record, for each element of the newly built vector, which input vector // that element comes from. -1 stands for undef, 0 for the zero vector, // and positive values for the input vectors. // VectorMask maps each element to its vector number, and VecIn maps vector // numbers to their initial SDValues. SmallVector VectorMask(NumElems, -1); SmallVector VecIn; VecIn.push_back(SDValue()); for (unsigned i = 0; i != NumElems; ++i) { SDValue Op = N->getOperand(i); if (Op.isUndef()) continue; // See if we can use a blend with a zero vector. // TODO: Should we generalize this to a blend with an arbitrary constant // vector? if (isNullConstant(Op) || isNullFPConstant(Op)) { UsesZeroVector = true; VectorMask[i] = 0; continue; } // Not an undef or zero. If the input is something other than an // EXTRACT_VECTOR_ELT with an in-range constant index, bail out. if (Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT || !isa(Op.getOperand(1))) return SDValue(); SDValue ExtractedFromVec = Op.getOperand(0); if (ExtractedFromVec.getValueType().isScalableVector()) return SDValue(); const APInt &ExtractIdx = Op.getConstantOperandAPInt(1); if (ExtractIdx.uge(ExtractedFromVec.getValueType().getVectorNumElements())) return SDValue(); // All inputs must have the same element type as the output. if (VT.getVectorElementType() != ExtractedFromVec.getValueType().getVectorElementType()) return SDValue(); // Have we seen this input vector before? // The vectors are expected to be tiny (usually 1 or 2 elements), so using // a map back from SDValues to numbers isn't worth it. int Idx = getFirstIndexOf(VecIn, ExtractedFromVec); if (Idx == -1) { // A new source vector? Idx = VecIn.size(); VecIn.push_back(ExtractedFromVec); } VectorMask[i] = Idx; } // If we didn't find at least one input vector, bail out. if (VecIn.size() < 2) return SDValue(); // If all the Operands of BUILD_VECTOR extract from same // vector, then split the vector efficiently based on the maximum // vector access index and adjust the VectorMask and // VecIn accordingly. bool DidSplitVec = false; if (VecIn.size() == 2) { unsigned MaxIndex = 0; unsigned NearestPow2 = 0; SDValue Vec = VecIn.back(); EVT InVT = Vec.getValueType(); SmallVector IndexVec(NumElems, 0); for (unsigned i = 0; i < NumElems; i++) { if (VectorMask[i] <= 0) continue; unsigned Index = N->getOperand(i).getConstantOperandVal(1); IndexVec[i] = Index; MaxIndex = std::max(MaxIndex, Index); } NearestPow2 = PowerOf2Ceil(MaxIndex); if (InVT.isSimple() && NearestPow2 > 2 && MaxIndex < NearestPow2 && NumElems * 2 < NearestPow2) { unsigned SplitSize = NearestPow2 / 2; EVT SplitVT = EVT::getVectorVT(*DAG.getContext(), InVT.getVectorElementType(), SplitSize); if (TLI.isTypeLegal(SplitVT) && SplitSize + SplitVT.getVectorNumElements() <= InVT.getVectorNumElements()) { SDValue VecIn2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, Vec, DAG.getVectorIdxConstant(SplitSize, DL)); SDValue VecIn1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, Vec, DAG.getVectorIdxConstant(0, DL)); VecIn.pop_back(); VecIn.push_back(VecIn1); VecIn.push_back(VecIn2); DidSplitVec = true; for (unsigned i = 0; i < NumElems; i++) { if (VectorMask[i] <= 0) continue; VectorMask[i] = (IndexVec[i] < SplitSize) ? 1 : 2; } } } } // Sort input vectors by decreasing vector element count, // while preserving the relative order of equally-sized vectors. // Note that we keep the first "implicit zero vector as-is. SmallVector SortedVecIn(VecIn); llvm::stable_sort(MutableArrayRef(SortedVecIn).drop_front(), [](const SDValue &a, const SDValue &b) { return a.getValueType().getVectorNumElements() > b.getValueType().getVectorNumElements(); }); // We now also need to rebuild the VectorMask, because it referenced element // order in VecIn, and we just sorted them. for (int &SourceVectorIndex : VectorMask) { if (SourceVectorIndex <= 0) continue; unsigned Idx = getFirstIndexOf(SortedVecIn, VecIn[SourceVectorIndex]); assert(Idx > 0 && Idx < SortedVecIn.size() && VecIn[SourceVectorIndex] == SortedVecIn[Idx] && "Remapping failure"); SourceVectorIndex = Idx; } VecIn = std::move(SortedVecIn); // TODO: Should this fire if some of the input vectors has illegal type (like // it does now), or should we let legalization run its course first? // Shuffle phase: // Take pairs of vectors, and shuffle them so that the result has elements // from these vectors in the correct places. // For example, given: // t10: i32 = extract_vector_elt t1, Constant:i64<0> // t11: i32 = extract_vector_elt t2, Constant:i64<0> // t12: i32 = extract_vector_elt t3, Constant:i64<0> // t13: i32 = extract_vector_elt t1, Constant:i64<1> // t14: v4i32 = BUILD_VECTOR t10, t11, t12, t13 // We will generate: // t20: v4i32 = vector_shuffle<0,4,u,1> t1, t2 // t21: v4i32 = vector_shuffle t3, undef SmallVector Shuffles; for (unsigned In = 0, Len = (VecIn.size() / 2); In < Len; ++In) { unsigned LeftIdx = 2 * In + 1; SDValue VecLeft = VecIn[LeftIdx]; SDValue VecRight = (LeftIdx + 1) < VecIn.size() ? VecIn[LeftIdx + 1] : SDValue(); if (SDValue Shuffle = createBuildVecShuffle(DL, N, VectorMask, VecLeft, VecRight, LeftIdx, DidSplitVec)) Shuffles.push_back(Shuffle); else return SDValue(); } // If we need the zero vector as an "ingredient" in the blend tree, add it // to the list of shuffles. if (UsesZeroVector) Shuffles.push_back(VT.isInteger() ? DAG.getConstant(0, DL, VT) : DAG.getConstantFP(0.0, DL, VT)); // If we only have one shuffle, we're done. if (Shuffles.size() == 1) return Shuffles[0]; // Update the vector mask to point to the post-shuffle vectors. for (int &Vec : VectorMask) if (Vec == 0) Vec = Shuffles.size() - 1; else Vec = (Vec - 1) / 2; // More than one shuffle. Generate a binary tree of blends, e.g. if from // the previous step we got the set of shuffles t10, t11, t12, t13, we will // generate: // t10: v8i32 = vector_shuffle<0,8,u,u,u,u,u,u> t1, t2 // t11: v8i32 = vector_shuffle t3, t4 // t12: v8i32 = vector_shuffle t5, t6 // t13: v8i32 = vector_shuffle t7, t8 // t20: v8i32 = vector_shuffle<0,1,10,11,u,u,u,u> t10, t11 // t21: v8i32 = vector_shuffle t12, t13 // t30: v8i32 = vector_shuffle<0,1,2,3,12,13,14,15> t20, t21 // Make sure the initial size of the shuffle list is even. if (Shuffles.size() % 2) Shuffles.push_back(DAG.getUNDEF(VT)); for (unsigned CurSize = Shuffles.size(); CurSize > 1; CurSize /= 2) { if (CurSize % 2) { Shuffles[CurSize] = DAG.getUNDEF(VT); CurSize++; } for (unsigned In = 0, Len = CurSize / 2; In < Len; ++In) { int Left = 2 * In; int Right = 2 * In + 1; SmallVector Mask(NumElems, -1); SDValue L = Shuffles[Left]; ArrayRef LMask; bool IsLeftShuffle = L.getOpcode() == ISD::VECTOR_SHUFFLE && L.use_empty() && L.getOperand(1).isUndef() && L.getOperand(0).getValueType() == L.getValueType(); if (IsLeftShuffle) { LMask = cast(L.getNode())->getMask(); L = L.getOperand(0); } SDValue R = Shuffles[Right]; ArrayRef RMask; bool IsRightShuffle = R.getOpcode() == ISD::VECTOR_SHUFFLE && R.use_empty() && R.getOperand(1).isUndef() && R.getOperand(0).getValueType() == R.getValueType(); if (IsRightShuffle) { RMask = cast(R.getNode())->getMask(); R = R.getOperand(0); } for (unsigned I = 0; I != NumElems; ++I) { if (VectorMask[I] == Left) { Mask[I] = I; if (IsLeftShuffle) Mask[I] = LMask[I]; VectorMask[I] = In; } else if (VectorMask[I] == Right) { Mask[I] = I + NumElems; if (IsRightShuffle) Mask[I] = RMask[I] + NumElems; VectorMask[I] = In; } } Shuffles[In] = DAG.getVectorShuffle(VT, DL, L, R, Mask); } } return Shuffles[0]; } // Try to turn a build vector of zero extends of extract vector elts into a // a vector zero extend and possibly an extract subvector. // TODO: Support sign extend? // TODO: Allow undef elements? SDValue DAGCombiner::convertBuildVecZextToZext(SDNode *N) { if (LegalOperations) return SDValue(); EVT VT = N->getValueType(0); bool FoundZeroExtend = false; SDValue Op0 = N->getOperand(0); auto checkElem = [&](SDValue Op) -> int64_t { unsigned Opc = Op.getOpcode(); FoundZeroExtend |= (Opc == ISD::ZERO_EXTEND); if ((Opc == ISD::ZERO_EXTEND || Opc == ISD::ANY_EXTEND) && Op.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT && Op0.getOperand(0).getOperand(0) == Op.getOperand(0).getOperand(0)) if (auto *C = dyn_cast(Op.getOperand(0).getOperand(1))) return C->getZExtValue(); return -1; }; // Make sure the first element matches // (zext (extract_vector_elt X, C)) // Offset must be a constant multiple of the // known-minimum vector length of the result type. int64_t Offset = checkElem(Op0); if (Offset < 0 || (Offset % VT.getVectorNumElements()) != 0) return SDValue(); unsigned NumElems = N->getNumOperands(); SDValue In = Op0.getOperand(0).getOperand(0); EVT InSVT = In.getValueType().getScalarType(); EVT InVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumElems); // Don't create an illegal input type after type legalization. if (LegalTypes && !TLI.isTypeLegal(InVT)) return SDValue(); // Ensure all the elements come from the same vector and are adjacent. for (unsigned i = 1; i != NumElems; ++i) { if ((Offset + i) != checkElem(N->getOperand(i))) return SDValue(); } SDLoc DL(N); In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InVT, In, Op0.getOperand(0).getOperand(1)); return DAG.getNode(FoundZeroExtend ? ISD::ZERO_EXTEND : ISD::ANY_EXTEND, DL, VT, In); } SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) { EVT VT = N->getValueType(0); // A vector built entirely of undefs is undef. if (ISD::allOperandsUndef(N)) return DAG.getUNDEF(VT); // If this is a splat of a bitcast from another vector, change to a // concat_vector. // For example: // (build_vector (i64 (bitcast (v2i32 X))), (i64 (bitcast (v2i32 X)))) -> // (v2i64 (bitcast (concat_vectors (v2i32 X), (v2i32 X)))) // // If X is a build_vector itself, the concat can become a larger build_vector. // TODO: Maybe this is useful for non-splat too? if (!LegalOperations) { if (SDValue Splat = cast(N)->getSplatValue()) { Splat = peekThroughBitcasts(Splat); EVT SrcVT = Splat.getValueType(); if (SrcVT.isVector()) { unsigned NumElts = N->getNumOperands() * SrcVT.getVectorNumElements(); EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(), NumElts); if (!LegalTypes || TLI.isTypeLegal(NewVT)) { SmallVector Ops(N->getNumOperands(), Splat); SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), NewVT, Ops); return DAG.getBitcast(VT, Concat); } } } } // Check if we can express BUILD VECTOR via subvector extract. if (!LegalTypes && (N->getNumOperands() > 1)) { SDValue Op0 = N->getOperand(0); auto checkElem = [&](SDValue Op) -> uint64_t { if ((Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT) && (Op0.getOperand(0) == Op.getOperand(0))) if (auto CNode = dyn_cast(Op.getOperand(1))) return CNode->getZExtValue(); return -1; }; int Offset = checkElem(Op0); for (unsigned i = 0; i < N->getNumOperands(); ++i) { if (Offset + i != checkElem(N->getOperand(i))) { Offset = -1; break; } } if ((Offset == 0) && (Op0.getOperand(0).getValueType() == N->getValueType(0))) return Op0.getOperand(0); if ((Offset != -1) && ((Offset % N->getValueType(0).getVectorNumElements()) == 0)) // IDX must be multiple of output size. return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N), N->getValueType(0), Op0.getOperand(0), Op0.getOperand(1)); } if (SDValue V = convertBuildVecZextToZext(N)) return V; if (SDValue V = reduceBuildVecExtToExtBuildVec(N)) return V; if (SDValue V = reduceBuildVecTruncToBitCast(N)) return V; if (SDValue V = reduceBuildVecToShuffle(N)) return V; // A splat of a single element is a SPLAT_VECTOR if supported on the target. // Do this late as some of the above may replace the splat. if (TLI.getOperationAction(ISD::SPLAT_VECTOR, VT) != TargetLowering::Expand) if (SDValue V = cast(N)->getSplatValue()) { assert(!V.isUndef() && "Splat of undef should have been handled earlier"); return DAG.getNode(ISD::SPLAT_VECTOR, SDLoc(N), VT, V); } return SDValue(); } static SDValue combineConcatVectorOfScalars(SDNode *N, SelectionDAG &DAG) { const TargetLowering &TLI = DAG.getTargetLoweringInfo(); EVT OpVT = N->getOperand(0).getValueType(); // If the operands are legal vectors, leave them alone. if (TLI.isTypeLegal(OpVT)) return SDValue(); SDLoc DL(N); EVT VT = N->getValueType(0); SmallVector Ops; EVT SVT = EVT::getIntegerVT(*DAG.getContext(), OpVT.getSizeInBits()); SDValue ScalarUndef = DAG.getNode(ISD::UNDEF, DL, SVT); // Keep track of what we encounter. bool AnyInteger = false; bool AnyFP = false; for (const SDValue &Op : N->ops()) { if (ISD::BITCAST == Op.getOpcode() && !Op.getOperand(0).getValueType().isVector()) Ops.push_back(Op.getOperand(0)); else if (ISD::UNDEF == Op.getOpcode()) Ops.push_back(ScalarUndef); else return SDValue(); // Note whether we encounter an integer or floating point scalar. // If it's neither, bail out, it could be something weird like x86mmx. EVT LastOpVT = Ops.back().getValueType(); if (LastOpVT.isFloatingPoint()) AnyFP = true; else if (LastOpVT.isInteger()) AnyInteger = true; else return SDValue(); } // If any of the operands is a floating point scalar bitcast to a vector, // use floating point types throughout, and bitcast everything. // Replace UNDEFs by another scalar UNDEF node, of the final desired type. if (AnyFP) { SVT = EVT::getFloatingPointVT(OpVT.getSizeInBits()); ScalarUndef = DAG.getNode(ISD::UNDEF, DL, SVT); if (AnyInteger) { for (SDValue &Op : Ops) { if (Op.getValueType() == SVT) continue; if (Op.isUndef()) Op = ScalarUndef; else Op = DAG.getBitcast(SVT, Op); } } } EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SVT, VT.getSizeInBits() / SVT.getSizeInBits()); return DAG.getBitcast(VT, DAG.getBuildVector(VecVT, DL, Ops)); } // Attempt to merge nested concat_vectors/undefs. // Fold concat_vectors(concat_vectors(x,y,z,w),u,u,concat_vectors(a,b,c,d)) // --> concat_vectors(x,y,z,w,u,u,u,u,u,u,u,u,a,b,c,d) static SDValue combineConcatVectorOfConcatVectors(SDNode *N, SelectionDAG &DAG) { EVT VT = N->getValueType(0); // Ensure we're concatenating UNDEF and CONCAT_VECTORS nodes of similar types. EVT SubVT; SDValue FirstConcat; for (const SDValue &Op : N->ops()) { if (Op.isUndef()) continue; if (Op.getOpcode() != ISD::CONCAT_VECTORS) return SDValue(); if (!FirstConcat) { SubVT = Op.getOperand(0).getValueType(); if (!DAG.getTargetLoweringInfo().isTypeLegal(SubVT)) return SDValue(); FirstConcat = Op; continue; } if (SubVT != Op.getOperand(0).getValueType()) return SDValue(); } assert(FirstConcat && "Concat of all-undefs found"); SmallVector ConcatOps; for (const SDValue &Op : N->ops()) { if (Op.isUndef()) { ConcatOps.append(FirstConcat->getNumOperands(), DAG.getUNDEF(SubVT)); continue; } ConcatOps.append(Op->op_begin(), Op->op_end()); } return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, ConcatOps); } // Check to see if this is a CONCAT_VECTORS of a bunch of EXTRACT_SUBVECTOR // operations. If so, and if the EXTRACT_SUBVECTOR vector inputs come from at // most two distinct vectors the same size as the result, attempt to turn this // into a legal shuffle. static SDValue combineConcatVectorOfExtracts(SDNode *N, SelectionDAG &DAG) { EVT VT = N->getValueType(0); EVT OpVT = N->getOperand(0).getValueType(); // We currently can't generate an appropriate shuffle for a scalable vector. if (VT.isScalableVector()) return SDValue(); int NumElts = VT.getVectorNumElements(); int NumOpElts = OpVT.getVectorNumElements(); SDValue SV0 = DAG.getUNDEF(VT), SV1 = DAG.getUNDEF(VT); SmallVector Mask; for (SDValue Op : N->ops()) { Op = peekThroughBitcasts(Op); // UNDEF nodes convert to UNDEF shuffle mask values. if (Op.isUndef()) { Mask.append((unsigned)NumOpElts, -1); continue; } if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR) return SDValue(); // What vector are we extracting the subvector from and at what index? SDValue ExtVec = Op.getOperand(0); int ExtIdx = Op.getConstantOperandVal(1); // We want the EVT of the original extraction to correctly scale the // extraction index. EVT ExtVT = ExtVec.getValueType(); ExtVec = peekThroughBitcasts(ExtVec); // UNDEF nodes convert to UNDEF shuffle mask values. if (ExtVec.isUndef()) { Mask.append((unsigned)NumOpElts, -1); continue; } // Ensure that we are extracting a subvector from a vector the same // size as the result. if (ExtVT.getSizeInBits() != VT.getSizeInBits()) return SDValue(); // Scale the subvector index to account for any bitcast. int NumExtElts = ExtVT.getVectorNumElements(); if (0 == (NumExtElts % NumElts)) ExtIdx /= (NumExtElts / NumElts); else if (0 == (NumElts % NumExtElts)) ExtIdx *= (NumElts / NumExtElts); else return SDValue(); // At most we can reference 2 inputs in the final shuffle. if (SV0.isUndef() || SV0 == ExtVec) { SV0 = ExtVec; for (int i = 0; i != NumOpElts; ++i) Mask.push_back(i + ExtIdx); } else if (SV1.isUndef() || SV1 == ExtVec) { SV1 = ExtVec; for (int i = 0; i != NumOpElts; ++i) Mask.push_back(i + ExtIdx + NumElts); } else { return SDValue(); } } const TargetLowering &TLI = DAG.getTargetLoweringInfo(); return TLI.buildLegalVectorShuffle(VT, SDLoc(N), DAG.getBitcast(VT, SV0), DAG.getBitcast(VT, SV1), Mask, DAG); } static SDValue combineConcatVectorOfCasts(SDNode *N, SelectionDAG &DAG) { unsigned CastOpcode = N->getOperand(0).getOpcode(); switch (CastOpcode) { case ISD::SINT_TO_FP: case ISD::UINT_TO_FP: case ISD::FP_TO_SINT: case ISD::FP_TO_UINT: // TODO: Allow more opcodes? // case ISD::BITCAST: // case ISD::TRUNCATE: // case ISD::ZERO_EXTEND: // case ISD::SIGN_EXTEND: // case ISD::FP_EXTEND: break; default: return SDValue(); } EVT SrcVT = N->getOperand(0).getOperand(0).getValueType(); if (!SrcVT.isVector()) return SDValue(); // All operands of the concat must be the same kind of cast from the same // source type. SmallVector SrcOps; for (SDValue Op : N->ops()) { if (Op.getOpcode() != CastOpcode || !Op.hasOneUse() || Op.getOperand(0).getValueType() != SrcVT) return SDValue(); SrcOps.push_back(Op.getOperand(0)); } // The wider cast must be supported by the target. This is unusual because // the operation support type parameter depends on the opcode. In addition, // check the other type in the cast to make sure this is really legal. EVT VT = N->getValueType(0); EVT SrcEltVT = SrcVT.getVectorElementType(); ElementCount NumElts = SrcVT.getVectorElementCount() * N->getNumOperands(); EVT ConcatSrcVT = EVT::getVectorVT(*DAG.getContext(), SrcEltVT, NumElts); const TargetLowering &TLI = DAG.getTargetLoweringInfo(); switch (CastOpcode) { case ISD::SINT_TO_FP: case ISD::UINT_TO_FP: if (!TLI.isOperationLegalOrCustom(CastOpcode, ConcatSrcVT) || !TLI.isTypeLegal(VT)) return SDValue(); break; case ISD::FP_TO_SINT: case ISD::FP_TO_UINT: if (!TLI.isOperationLegalOrCustom(CastOpcode, VT) || !TLI.isTypeLegal(ConcatSrcVT)) return SDValue(); break; default: llvm_unreachable("Unexpected cast opcode"); } // concat (cast X), (cast Y)... -> cast (concat X, Y...) SDLoc DL(N); SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, DL, ConcatSrcVT, SrcOps); return DAG.getNode(CastOpcode, DL, VT, NewConcat); } SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) { // If we only have one input vector, we don't need to do any concatenation. if (N->getNumOperands() == 1) return N->getOperand(0); // Check if all of the operands are undefs. EVT VT = N->getValueType(0); if (ISD::allOperandsUndef(N)) return DAG.getUNDEF(VT); // Optimize concat_vectors where all but the first of the vectors are undef. if (all_of(drop_begin(N->ops()), [](const SDValue &Op) { return Op.isUndef(); })) { SDValue In = N->getOperand(0); assert(In.getValueType().isVector() && "Must concat vectors"); // If the input is a concat_vectors, just make a larger concat by padding // with smaller undefs. if (In.getOpcode() == ISD::CONCAT_VECTORS && In.hasOneUse()) { unsigned NumOps = N->getNumOperands() * In.getNumOperands(); SmallVector Ops(In->op_begin(), In->op_end()); Ops.resize(NumOps, DAG.getUNDEF(Ops[0].getValueType())); return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops); } SDValue Scalar = peekThroughOneUseBitcasts(In); // concat_vectors(scalar_to_vector(scalar), undef) -> // scalar_to_vector(scalar) if (!LegalOperations && Scalar.getOpcode() == ISD::SCALAR_TO_VECTOR && Scalar.hasOneUse()) { EVT SVT = Scalar.getValueType().getVectorElementType(); if (SVT == Scalar.getOperand(0).getValueType()) Scalar = Scalar.getOperand(0); } // concat_vectors(scalar, undef) -> scalar_to_vector(scalar) if (!Scalar.getValueType().isVector()) { // If the bitcast type isn't legal, it might be a trunc of a legal type; // look through the trunc so we can still do the transform: // concat_vectors(trunc(scalar), undef) -> scalar_to_vector(scalar) if (Scalar->getOpcode() == ISD::TRUNCATE && !TLI.isTypeLegal(Scalar.getValueType()) && TLI.isTypeLegal(Scalar->getOperand(0).getValueType())) Scalar = Scalar->getOperand(0); EVT SclTy = Scalar.getValueType(); if (!SclTy.isFloatingPoint() && !SclTy.isInteger()) return SDValue(); // Bail out if the vector size is not a multiple of the scalar size. if (VT.getSizeInBits() % SclTy.getSizeInBits()) return SDValue(); unsigned VNTNumElms = VT.getSizeInBits() / SclTy.getSizeInBits(); if (VNTNumElms < 2) return SDValue(); EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy, VNTNumElms); if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType())) return SDValue(); SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N), NVT, Scalar); return DAG.getBitcast(VT, Res); } } // Fold any combination of BUILD_VECTOR or UNDEF nodes into one BUILD_VECTOR. // We have already tested above for an UNDEF only concatenation. // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...)) // -> (BUILD_VECTOR A, B, ..., C, D, ...) auto IsBuildVectorOrUndef = [](const SDValue &Op) { return ISD::UNDEF == Op.getOpcode() || ISD::BUILD_VECTOR == Op.getOpcode(); }; if (llvm::all_of(N->ops(), IsBuildVectorOrUndef)) { SmallVector Opnds; EVT SVT = VT.getScalarType(); EVT MinVT = SVT; if (!SVT.isFloatingPoint()) { // If BUILD_VECTOR are from built from integer, they may have different // operand types. Get the smallest type and truncate all operands to it. bool FoundMinVT = false; for (const SDValue &Op : N->ops()) if (ISD::BUILD_VECTOR == Op.getOpcode()) { EVT OpSVT = Op.getOperand(0).getValueType(); MinVT = (!FoundMinVT || OpSVT.bitsLE(MinVT)) ? OpSVT : MinVT; FoundMinVT = true; } assert(FoundMinVT && "Concat vector type mismatch"); } for (const SDValue &Op : N->ops()) { EVT OpVT = Op.getValueType(); unsigned NumElts = OpVT.getVectorNumElements(); if (ISD::UNDEF == Op.getOpcode()) Opnds.append(NumElts, DAG.getUNDEF(MinVT)); if (ISD::BUILD_VECTOR == Op.getOpcode()) { if (SVT.isFloatingPoint()) { assert(SVT == OpVT.getScalarType() && "Concat vector type mismatch"); Opnds.append(Op->op_begin(), Op->op_begin() + NumElts); } else { for (unsigned i = 0; i != NumElts; ++i) Opnds.push_back( DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinVT, Op.getOperand(i))); } } } assert(VT.getVectorNumElements() == Opnds.size() && "Concat vector type mismatch"); return DAG.getBuildVector(VT, SDLoc(N), Opnds); } // Fold CONCAT_VECTORS of only bitcast scalars (or undef) to BUILD_VECTOR. // FIXME: Add support for concat_vectors(bitcast(vec0),bitcast(vec1),...). if (SDValue V = combineConcatVectorOfScalars(N, DAG)) return V; if (Level < AfterLegalizeVectorOps && TLI.isTypeLegal(VT)) { // Fold CONCAT_VECTORS of CONCAT_VECTORS (or undef) to VECTOR_SHUFFLE. if (SDValue V = combineConcatVectorOfConcatVectors(N, DAG)) return V; // Fold CONCAT_VECTORS of EXTRACT_SUBVECTOR (or undef) to VECTOR_SHUFFLE. if (SDValue V = combineConcatVectorOfExtracts(N, DAG)) return V; } if (SDValue V = combineConcatVectorOfCasts(N, DAG)) return V; // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR // nodes often generate nop CONCAT_VECTOR nodes. Scan the CONCAT_VECTOR // operands and look for a CONCAT operations that place the incoming vectors // at the exact same location. // // For scalable vectors, EXTRACT_SUBVECTOR indexes are implicitly scaled. SDValue SingleSource = SDValue(); unsigned PartNumElem = N->getOperand(0).getValueType().getVectorMinNumElements(); for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { SDValue Op = N->getOperand(i); if (Op.isUndef()) continue; // Check if this is the identity extract: if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR) return SDValue(); // Find the single incoming vector for the extract_subvector. if (SingleSource.getNode()) { if (Op.getOperand(0) != SingleSource) return SDValue(); } else { SingleSource = Op.getOperand(0); // Check the source type is the same as the type of the result. // If not, this concat may extend the vector, so we can not // optimize it away. if (SingleSource.getValueType() != N->getValueType(0)) return SDValue(); } // Check that we are reading from the identity index. unsigned IdentityIndex = i * PartNumElem; if (Op.getConstantOperandAPInt(1) != IdentityIndex) return SDValue(); } if (SingleSource.getNode()) return SingleSource; return SDValue(); } // Helper that peeks through INSERT_SUBVECTOR/CONCAT_VECTORS to find // if the subvector can be sourced for free. static SDValue getSubVectorSrc(SDValue V, SDValue Index, EVT SubVT) { if (V.getOpcode() == ISD::INSERT_SUBVECTOR && V.getOperand(1).getValueType() == SubVT && V.getOperand(2) == Index) { return V.getOperand(1); } auto *IndexC = dyn_cast(Index); if (IndexC && V.getOpcode() == ISD::CONCAT_VECTORS && V.getOperand(0).getValueType() == SubVT && (IndexC->getZExtValue() % SubVT.getVectorMinNumElements()) == 0) { uint64_t SubIdx = IndexC->getZExtValue() / SubVT.getVectorMinNumElements(); return V.getOperand(SubIdx); } return SDValue(); } static SDValue narrowInsertExtractVectorBinOp(SDNode *Extract, SelectionDAG &DAG, bool LegalOperations) { const TargetLowering &TLI = DAG.getTargetLoweringInfo(); SDValue BinOp = Extract->getOperand(0); unsigned BinOpcode = BinOp.getOpcode(); if (!TLI.isBinOp(BinOpcode) || BinOp->getNumValues() != 1) return SDValue(); EVT VecVT = BinOp.getValueType(); SDValue Bop0 = BinOp.getOperand(0), Bop1 = BinOp.getOperand(1); if (VecVT != Bop0.getValueType() || VecVT != Bop1.getValueType()) return SDValue(); SDValue Index = Extract->getOperand(1); EVT SubVT = Extract->getValueType(0); if (!TLI.isOperationLegalOrCustom(BinOpcode, SubVT, LegalOperations)) return SDValue(); SDValue Sub0 = getSubVectorSrc(Bop0, Index, SubVT); SDValue Sub1 = getSubVectorSrc(Bop1, Index, SubVT); // TODO: We could handle the case where only 1 operand is being inserted by // creating an extract of the other operand, but that requires checking // number of uses and/or costs. if (!Sub0 || !Sub1) return SDValue(); // We are inserting both operands of the wide binop only to extract back // to the narrow vector size. Eliminate all of the insert/extract: // ext (binop (ins ?, X, Index), (ins ?, Y, Index)), Index --> binop X, Y return DAG.getNode(BinOpcode, SDLoc(Extract), SubVT, Sub0, Sub1, BinOp->getFlags()); } /// If we are extracting a subvector produced by a wide binary operator try /// to use a narrow binary operator and/or avoid concatenation and extraction. static SDValue narrowExtractedVectorBinOp(SDNode *Extract, SelectionDAG &DAG, bool LegalOperations) { // TODO: Refactor with the caller (visitEXTRACT_SUBVECTOR), so we can share // some of these bailouts with other transforms. if (SDValue V = narrowInsertExtractVectorBinOp(Extract, DAG, LegalOperations)) return V; // The extract index must be a constant, so we can map it to a concat operand. auto *ExtractIndexC = dyn_cast(Extract->getOperand(1)); if (!ExtractIndexC) return SDValue(); // We are looking for an optionally bitcasted wide vector binary operator // feeding an extract subvector. const TargetLowering &TLI = DAG.getTargetLoweringInfo(); SDValue BinOp = peekThroughBitcasts(Extract->getOperand(0)); unsigned BOpcode = BinOp.getOpcode(); if (!TLI.isBinOp(BOpcode) || BinOp->getNumValues() != 1) return SDValue(); // Exclude the fake form of fneg (fsub -0.0, x) because that is likely to be // reduced to the unary fneg when it is visited, and we probably want to deal // with fneg in a target-specific way. if (BOpcode == ISD::FSUB) { auto *C = isConstOrConstSplatFP(BinOp.getOperand(0), /*AllowUndefs*/ true); if (C && C->getValueAPF().isNegZero()) return SDValue(); } // The binop must be a vector type, so we can extract some fraction of it. EVT WideBVT = BinOp.getValueType(); // The optimisations below currently assume we are dealing with fixed length // vectors. It is possible to add support for scalable vectors, but at the // moment we've done no analysis to prove whether they are profitable or not. if (!WideBVT.isFixedLengthVector()) return SDValue(); EVT VT = Extract->getValueType(0); unsigned ExtractIndex = ExtractIndexC->getZExtValue(); assert(ExtractIndex % VT.getVectorNumElements() == 0 && "Extract index is not a multiple of the vector length."); // Bail out if this is not a proper multiple width extraction. unsigned WideWidth = WideBVT.getSizeInBits(); unsigned NarrowWidth = VT.getSizeInBits(); if (WideWidth % NarrowWidth != 0) return SDValue(); // Bail out if we are extracting a fraction of a single operation. This can // occur because we potentially looked through a bitcast of the binop. unsigned NarrowingRatio = WideWidth / NarrowWidth; unsigned WideNumElts = WideBVT.getVectorNumElements(); if (WideNumElts % NarrowingRatio != 0) return SDValue(); // Bail out if the target does not support a narrower version of the binop. EVT NarrowBVT = EVT::getVectorVT(*DAG.getContext(), WideBVT.getScalarType(), WideNumElts / NarrowingRatio); if (!TLI.isOperationLegalOrCustomOrPromote(BOpcode, NarrowBVT)) return SDValue(); // If extraction is cheap, we don't need to look at the binop operands // for concat ops. The narrow binop alone makes this transform profitable. // We can't just reuse the original extract index operand because we may have // bitcasted. unsigned ConcatOpNum = ExtractIndex / VT.getVectorNumElements(); unsigned ExtBOIdx = ConcatOpNum * NarrowBVT.getVectorNumElements(); if (TLI.isExtractSubvectorCheap(NarrowBVT, WideBVT, ExtBOIdx) && BinOp.hasOneUse() && Extract->getOperand(0)->hasOneUse()) { // extract (binop B0, B1), N --> binop (extract B0, N), (extract B1, N) SDLoc DL(Extract); SDValue NewExtIndex = DAG.getVectorIdxConstant(ExtBOIdx, DL); SDValue X = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, NarrowBVT, BinOp.getOperand(0), NewExtIndex); SDValue Y = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, NarrowBVT, BinOp.getOperand(1), NewExtIndex); SDValue NarrowBinOp = DAG.getNode(BOpcode, DL, NarrowBVT, X, Y, BinOp->getFlags()); return DAG.getBitcast(VT, NarrowBinOp); } // Only handle the case where we are doubling and then halving. A larger ratio // may require more than two narrow binops to replace the wide binop. if (NarrowingRatio != 2) return SDValue(); // TODO: The motivating case for this transform is an x86 AVX1 target. That // target has temptingly almost legal versions of bitwise logic ops in 256-bit // flavors, but no other 256-bit integer support. This could be extended to // handle any binop, but that may require fixing/adding other folds to avoid // codegen regressions. if (BOpcode != ISD::AND && BOpcode != ISD::OR && BOpcode != ISD::XOR) return SDValue(); // We need at least one concatenation operation of a binop operand to make // this transform worthwhile. The concat must double the input vector sizes. auto GetSubVector = [ConcatOpNum](SDValue V) -> SDValue { if (V.getOpcode() == ISD::CONCAT_VECTORS && V.getNumOperands() == 2) return V.getOperand(ConcatOpNum); return SDValue(); }; SDValue SubVecL = GetSubVector(peekThroughBitcasts(BinOp.getOperand(0))); SDValue SubVecR = GetSubVector(peekThroughBitcasts(BinOp.getOperand(1))); if (SubVecL || SubVecR) { // If a binop operand was not the result of a concat, we must extract a // half-sized operand for our new narrow binop: // extract (binop (concat X1, X2), (concat Y1, Y2)), N --> binop XN, YN // extract (binop (concat X1, X2), Y), N --> binop XN, (extract Y, IndexC) // extract (binop X, (concat Y1, Y2)), N --> binop (extract X, IndexC), YN SDLoc DL(Extract); SDValue IndexC = DAG.getVectorIdxConstant(ExtBOIdx, DL); SDValue X = SubVecL ? DAG.getBitcast(NarrowBVT, SubVecL) : DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, NarrowBVT, BinOp.getOperand(0), IndexC); SDValue Y = SubVecR ? DAG.getBitcast(NarrowBVT, SubVecR) : DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, NarrowBVT, BinOp.getOperand(1), IndexC); SDValue NarrowBinOp = DAG.getNode(BOpcode, DL, NarrowBVT, X, Y); return DAG.getBitcast(VT, NarrowBinOp); } return SDValue(); } /// If we are extracting a subvector from a wide vector load, convert to a /// narrow load to eliminate the extraction: /// (extract_subvector (load wide vector)) --> (load narrow vector) static SDValue narrowExtractedVectorLoad(SDNode *Extract, SelectionDAG &DAG) { // TODO: Add support for big-endian. The offset calculation must be adjusted. if (DAG.getDataLayout().isBigEndian()) return SDValue(); auto *Ld = dyn_cast(Extract->getOperand(0)); if (!Ld || Ld->getExtensionType() || !Ld->isSimple()) return SDValue(); // Allow targets to opt-out. EVT VT = Extract->getValueType(0); // We can only create byte sized loads. if (!VT.isByteSized()) return SDValue(); unsigned Index = Extract->getConstantOperandVal(1); unsigned NumElts = VT.getVectorMinNumElements(); // The definition of EXTRACT_SUBVECTOR states that the index must be a // multiple of the minimum number of elements in the result type. assert(Index % NumElts == 0 && "The extract subvector index is not a " "multiple of the result's element count"); // It's fine to use TypeSize here as we know the offset will not be negative. TypeSize Offset = VT.getStoreSize() * (Index / NumElts); const TargetLowering &TLI = DAG.getTargetLoweringInfo(); if (!TLI.shouldReduceLoadWidth(Ld, Ld->getExtensionType(), VT)) return SDValue(); // The narrow load will be offset from the base address of the old load if // we are extracting from something besides index 0 (little-endian). SDLoc DL(Extract); // TODO: Use "BaseIndexOffset" to make this more effective. SDValue NewAddr = DAG.getMemBasePlusOffset(Ld->getBasePtr(), Offset, DL); uint64_t StoreSize = MemoryLocation::getSizeOrUnknown(VT.getStoreSize()); MachineFunction &MF = DAG.getMachineFunction(); MachineMemOperand *MMO; if (Offset.isScalable()) { MachinePointerInfo MPI = MachinePointerInfo(Ld->getPointerInfo().getAddrSpace()); MMO = MF.getMachineMemOperand(Ld->getMemOperand(), MPI, StoreSize); } else MMO = MF.getMachineMemOperand(Ld->getMemOperand(), Offset.getFixedSize(), StoreSize); SDValue NewLd = DAG.getLoad(VT, DL, Ld->getChain(), NewAddr, MMO); DAG.makeEquivalentMemoryOrdering(Ld, NewLd); return NewLd; } /// Given EXTRACT_SUBVECTOR(VECTOR_SHUFFLE(Op0, Op1, Mask)), /// try to produce VECTOR_SHUFFLE(EXTRACT_SUBVECTOR(Op?, ?), /// EXTRACT_SUBVECTOR(Op?, ?), /// Mask')) /// iff it is legal and profitable to do so. Notably, the trimmed mask /// (containing only the elements that are extracted) /// must reference at most two subvectors. static SDValue foldExtractSubvectorFromShuffleVector(SDNode *N, SelectionDAG &DAG, const TargetLowering &TLI, bool LegalOperations) { assert(N->getOpcode() == ISD::EXTRACT_SUBVECTOR && "Must only be called on EXTRACT_SUBVECTOR's"); SDValue N0 = N->getOperand(0); // Only deal with non-scalable vectors. EVT NarrowVT = N->getValueType(0); EVT WideVT = N0.getValueType(); if (!NarrowVT.isFixedLengthVector() || !WideVT.isFixedLengthVector()) return SDValue(); // The operand must be a shufflevector. auto *WideShuffleVector = dyn_cast(N0); if (!WideShuffleVector) return SDValue(); // The old shuffleneeds to go away. if (!WideShuffleVector->hasOneUse()) return SDValue(); // And the narrow shufflevector that we'll form must be legal. if (LegalOperations && !TLI.isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, NarrowVT)) return SDValue(); uint64_t FirstExtractedEltIdx = N->getConstantOperandVal(1); int NumEltsExtracted = NarrowVT.getVectorNumElements(); assert((FirstExtractedEltIdx % NumEltsExtracted) == 0 && "Extract index is not a multiple of the output vector length."); int WideNumElts = WideVT.getVectorNumElements(); SmallVector NewMask; NewMask.reserve(NumEltsExtracted); SmallSetVector, 2> DemandedSubvectors; // Try to decode the wide mask into narrow mask from at most two subvectors. for (int M : WideShuffleVector->getMask().slice(FirstExtractedEltIdx, NumEltsExtracted)) { assert((M >= -1) && (M < (2 * WideNumElts)) && "Out-of-bounds shuffle mask?"); if (M < 0) { // Does not depend on operands, does not require adjustment. NewMask.emplace_back(M); continue; } // From which operand of the shuffle does this shuffle mask element pick? int WideShufOpIdx = M / WideNumElts; // Which element of that operand is picked? int OpEltIdx = M % WideNumElts; assert((OpEltIdx + WideShufOpIdx * WideNumElts) == M && "Shuffle mask vector decomposition failure."); // And which NumEltsExtracted-sized subvector of that operand is that? int OpSubvecIdx = OpEltIdx / NumEltsExtracted; // And which element within that subvector of that operand is that? int OpEltIdxInSubvec = OpEltIdx % NumEltsExtracted; assert((OpEltIdxInSubvec + OpSubvecIdx * NumEltsExtracted) == OpEltIdx && "Shuffle mask subvector decomposition failure."); assert((OpEltIdxInSubvec + OpSubvecIdx * NumEltsExtracted + WideShufOpIdx * WideNumElts) == M && "Shuffle mask full decomposition failure."); SDValue Op = WideShuffleVector->getOperand(WideShufOpIdx); if (Op.isUndef()) { // Picking from an undef operand. Let's adjust mask instead. NewMask.emplace_back(-1); continue; } // Profitability check: only deal with extractions from the first subvector. if (OpSubvecIdx != 0) return SDValue(); const std::pair DemandedSubvector = std::make_pair(Op, OpSubvecIdx); if (DemandedSubvectors.insert(DemandedSubvector)) { if (DemandedSubvectors.size() > 2) return SDValue(); // We can't handle more than two subvectors. // How many elements into the WideVT does this subvector start? int Index = NumEltsExtracted * OpSubvecIdx; // Bail out if the extraction isn't going to be cheap. if (!TLI.isExtractSubvectorCheap(NarrowVT, WideVT, Index)) return SDValue(); } // Ok, but from which operand of the new shuffle will this element pick? int NewOpIdx = getFirstIndexOf(DemandedSubvectors.getArrayRef(), DemandedSubvector); assert((NewOpIdx == 0 || NewOpIdx == 1) && "Unexpected operand index."); int AdjM = OpEltIdxInSubvec + NewOpIdx * NumEltsExtracted; NewMask.emplace_back(AdjM); } assert(NewMask.size() == (unsigned)NumEltsExtracted && "Produced bad mask."); assert(DemandedSubvectors.size() <= 2 && "Should have ended up demanding at most two subvectors."); // Did we discover that the shuffle does not actually depend on operands? if (DemandedSubvectors.empty()) return DAG.getUNDEF(NarrowVT); // We still perform the exact same EXTRACT_SUBVECTOR, just on different // operand[s]/index[es], so there is no point in checking for it's legality. // Do not turn a legal shuffle into an illegal one. if (TLI.isShuffleMaskLegal(WideShuffleVector->getMask(), WideVT) && !TLI.isShuffleMaskLegal(NewMask, NarrowVT)) return SDValue(); SDLoc DL(N); SmallVector NewOps; for (const std::pair &DemandedSubvector : DemandedSubvectors) { // How many elements into the WideVT does this subvector start? int Index = NumEltsExtracted * DemandedSubvector.second; SDValue IndexC = DAG.getVectorIdxConstant(Index, DL); NewOps.emplace_back(DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, NarrowVT, DemandedSubvector.first, IndexC)); } assert((NewOps.size() == 1 || NewOps.size() == 2) && "Should end up with either one or two ops"); // If we ended up with only one operand, pad with an undef. if (NewOps.size() == 1) NewOps.emplace_back(DAG.getUNDEF(NarrowVT)); return DAG.getVectorShuffle(NarrowVT, DL, NewOps[0], NewOps[1], NewMask); } SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode *N) { EVT NVT = N->getValueType(0); SDValue V = N->getOperand(0); uint64_t ExtIdx = N->getConstantOperandVal(1); // Extract from UNDEF is UNDEF. if (V.isUndef()) return DAG.getUNDEF(NVT); if (TLI.isOperationLegalOrCustomOrPromote(ISD::LOAD, NVT)) if (SDValue NarrowLoad = narrowExtractedVectorLoad(N, DAG)) return NarrowLoad; // Combine an extract of an extract into a single extract_subvector. // ext (ext X, C), 0 --> ext X, C if (ExtIdx == 0 && V.getOpcode() == ISD::EXTRACT_SUBVECTOR && V.hasOneUse()) { if (TLI.isExtractSubvectorCheap(NVT, V.getOperand(0).getValueType(), V.getConstantOperandVal(1)) && TLI.isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, NVT)) { return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N), NVT, V.getOperand(0), V.getOperand(1)); } } // ty1 extract_vector(ty2 splat(V))) -> ty1 splat(V) if (V.getOpcode() == ISD::SPLAT_VECTOR) if (DAG.isConstantValueOfAnyType(V.getOperand(0)) || V.hasOneUse()) if (!LegalOperations || TLI.isOperationLegal(ISD::SPLAT_VECTOR, NVT)) return DAG.getSplatVector(NVT, SDLoc(N), V.getOperand(0)); // Try to move vector bitcast after extract_subv by scaling extraction index: // extract_subv (bitcast X), Index --> bitcast (extract_subv X, Index') if (V.getOpcode() == ISD::BITCAST && V.getOperand(0).getValueType().isVector() && (!LegalOperations || TLI.isOperationLegal(ISD::BITCAST, NVT))) { SDValue SrcOp = V.getOperand(0); EVT SrcVT = SrcOp.getValueType(); unsigned SrcNumElts = SrcVT.getVectorMinNumElements(); unsigned DestNumElts = V.getValueType().getVectorMinNumElements(); if ((SrcNumElts % DestNumElts) == 0) { unsigned SrcDestRatio = SrcNumElts / DestNumElts; ElementCount NewExtEC = NVT.getVectorElementCount() * SrcDestRatio; EVT NewExtVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getScalarType(), NewExtEC); if (TLI.isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, NewExtVT)) { SDLoc DL(N); SDValue NewIndex = DAG.getVectorIdxConstant(ExtIdx * SrcDestRatio, DL); SDValue NewExtract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, NewExtVT, V.getOperand(0), NewIndex); return DAG.getBitcast(NVT, NewExtract); } } if ((DestNumElts % SrcNumElts) == 0) { unsigned DestSrcRatio = DestNumElts / SrcNumElts; if (NVT.getVectorElementCount().isKnownMultipleOf(DestSrcRatio)) { ElementCount NewExtEC = NVT.getVectorElementCount().divideCoefficientBy(DestSrcRatio); EVT ScalarVT = SrcVT.getScalarType(); if ((ExtIdx % DestSrcRatio) == 0) { SDLoc DL(N); unsigned IndexValScaled = ExtIdx / DestSrcRatio; EVT NewExtVT = EVT::getVectorVT(*DAG.getContext(), ScalarVT, NewExtEC); if (TLI.isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, NewExtVT)) { SDValue NewIndex = DAG.getVectorIdxConstant(IndexValScaled, DL); SDValue NewExtract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, NewExtVT, V.getOperand(0), NewIndex); return DAG.getBitcast(NVT, NewExtract); } if (NewExtEC.isScalar() && TLI.isOperationLegalOrCustom(ISD::EXTRACT_VECTOR_ELT, ScalarVT)) { SDValue NewIndex = DAG.getVectorIdxConstant(IndexValScaled, DL); SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ScalarVT, V.getOperand(0), NewIndex); return DAG.getBitcast(NVT, NewExtract); } } } } } if (V.getOpcode() == ISD::CONCAT_VECTORS) { unsigned ExtNumElts = NVT.getVectorMinNumElements(); EVT ConcatSrcVT = V.getOperand(0).getValueType(); assert(ConcatSrcVT.getVectorElementType() == NVT.getVectorElementType() && "Concat and extract subvector do not change element type"); assert((ExtIdx % ExtNumElts) == 0 && "Extract index is not a multiple of the input vector length."); unsigned ConcatSrcNumElts = ConcatSrcVT.getVectorMinNumElements(); unsigned ConcatOpIdx = ExtIdx / ConcatSrcNumElts; // If the concatenated source types match this extract, it's a direct // simplification: // extract_subvec (concat V1, V2, ...), i --> Vi if (NVT.getVectorElementCount() == ConcatSrcVT.getVectorElementCount()) return V.getOperand(ConcatOpIdx); // If the concatenated source vectors are a multiple length of this extract, // then extract a fraction of one of those source vectors directly from a // concat operand. Example: // v2i8 extract_subvec (v16i8 concat (v8i8 X), (v8i8 Y), 14 --> // v2i8 extract_subvec v8i8 Y, 6 if (NVT.isFixedLengthVector() && ConcatSrcVT.isFixedLengthVector() && ConcatSrcNumElts % ExtNumElts == 0) { SDLoc DL(N); unsigned NewExtIdx = ExtIdx - ConcatOpIdx * ConcatSrcNumElts; assert(NewExtIdx + ExtNumElts <= ConcatSrcNumElts && "Trying to extract from >1 concat operand?"); assert(NewExtIdx % ExtNumElts == 0 && "Extract index is not a multiple of the input vector length."); SDValue NewIndexC = DAG.getVectorIdxConstant(NewExtIdx, DL); return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, NVT, V.getOperand(ConcatOpIdx), NewIndexC); } } if (SDValue V = foldExtractSubvectorFromShuffleVector(N, DAG, TLI, LegalOperations)) return V; V = peekThroughBitcasts(V); // If the input is a build vector. Try to make a smaller build vector. if (V.getOpcode() == ISD::BUILD_VECTOR) { EVT InVT = V.getValueType(); unsigned ExtractSize = NVT.getSizeInBits(); unsigned EltSize = InVT.getScalarSizeInBits(); // Only do this if we won't split any elements. if (ExtractSize % EltSize == 0) { unsigned NumElems = ExtractSize / EltSize; EVT EltVT = InVT.getVectorElementType(); EVT ExtractVT = NumElems == 1 ? EltVT : EVT::getVectorVT(*DAG.getContext(), EltVT, NumElems); if ((Level < AfterLegalizeDAG || (NumElems == 1 || TLI.isOperationLegal(ISD::BUILD_VECTOR, ExtractVT))) && (!LegalTypes || TLI.isTypeLegal(ExtractVT))) { unsigned IdxVal = (ExtIdx * NVT.getScalarSizeInBits()) / EltSize; if (NumElems == 1) { SDValue Src = V->getOperand(IdxVal); if (EltVT != Src.getValueType()) Src = DAG.getNode(ISD::TRUNCATE, SDLoc(N), InVT, Src); return DAG.getBitcast(NVT, Src); } // Extract the pieces from the original build_vector. SDValue BuildVec = DAG.getBuildVector(ExtractVT, SDLoc(N), V->ops().slice(IdxVal, NumElems)); return DAG.getBitcast(NVT, BuildVec); } } } if (V.getOpcode() == ISD::INSERT_SUBVECTOR) { // Handle only simple case where vector being inserted and vector // being extracted are of same size. EVT SmallVT = V.getOperand(1).getValueType(); if (!NVT.bitsEq(SmallVT)) return SDValue(); // Combine: // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx) // Into: // indices are equal or bit offsets are equal => V1 // otherwise => (extract_subvec V1, ExtIdx) uint64_t InsIdx = V.getConstantOperandVal(2); if (InsIdx * SmallVT.getScalarSizeInBits() == ExtIdx * NVT.getScalarSizeInBits()) { if (LegalOperations && !TLI.isOperationLegal(ISD::BITCAST, NVT)) return SDValue(); return DAG.getBitcast(NVT, V.getOperand(1)); } return DAG.getNode( ISD::EXTRACT_SUBVECTOR, SDLoc(N), NVT, DAG.getBitcast(N->getOperand(0).getValueType(), V.getOperand(0)), N->getOperand(1)); } if (SDValue NarrowBOp = narrowExtractedVectorBinOp(N, DAG, LegalOperations)) return NarrowBOp; if (SimplifyDemandedVectorElts(SDValue(N, 0))) return SDValue(N, 0); return SDValue(); } /// Try to convert a wide shuffle of concatenated vectors into 2 narrow shuffles /// followed by concatenation. Narrow vector ops may have better performance /// than wide ops, and this can unlock further narrowing of other vector ops. /// Targets can invert this transform later if it is not profitable. static SDValue foldShuffleOfConcatUndefs(ShuffleVectorSDNode *Shuf, SelectionDAG &DAG) { SDValue N0 = Shuf->getOperand(0), N1 = Shuf->getOperand(1); if (N0.getOpcode() != ISD::CONCAT_VECTORS || N0.getNumOperands() != 2 || N1.getOpcode() != ISD::CONCAT_VECTORS || N1.getNumOperands() != 2 || !N0.getOperand(1).isUndef() || !N1.getOperand(1).isUndef()) return SDValue(); // Split the wide shuffle mask into halves. Any mask element that is accessing // operand 1 is offset down to account for narrowing of the vectors. ArrayRef Mask = Shuf->getMask(); EVT VT = Shuf->getValueType(0); unsigned NumElts = VT.getVectorNumElements(); unsigned HalfNumElts = NumElts / 2; SmallVector Mask0(HalfNumElts, -1); SmallVector Mask1(HalfNumElts, -1); for (unsigned i = 0; i != NumElts; ++i) { if (Mask[i] == -1) continue; // If we reference the upper (undef) subvector then the element is undef. if ((Mask[i] % NumElts) >= HalfNumElts) continue; int M = Mask[i] < (int)NumElts ? Mask[i] : Mask[i] - (int)HalfNumElts; if (i < HalfNumElts) Mask0[i] = M; else Mask1[i - HalfNumElts] = M; } // Ask the target if this is a valid transform. const TargetLowering &TLI = DAG.getTargetLoweringInfo(); EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), VT.getScalarType(), HalfNumElts); if (!TLI.isShuffleMaskLegal(Mask0, HalfVT) || !TLI.isShuffleMaskLegal(Mask1, HalfVT)) return SDValue(); // shuffle (concat X, undef), (concat Y, undef), Mask --> // concat (shuffle X, Y, Mask0), (shuffle X, Y, Mask1) SDValue X = N0.getOperand(0), Y = N1.getOperand(0); SDLoc DL(Shuf); SDValue Shuf0 = DAG.getVectorShuffle(HalfVT, DL, X, Y, Mask0); SDValue Shuf1 = DAG.getVectorShuffle(HalfVT, DL, X, Y, Mask1); return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Shuf0, Shuf1); } // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat, // or turn a shuffle of a single concat into simpler shuffle then concat. static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) { EVT VT = N->getValueType(0); unsigned NumElts = VT.getVectorNumElements(); SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); ShuffleVectorSDNode *SVN = cast(N); ArrayRef Mask = SVN->getMask(); SmallVector Ops; EVT ConcatVT = N0.getOperand(0).getValueType(); unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements(); unsigned NumConcats = NumElts / NumElemsPerConcat; auto IsUndefMaskElt = [](int i) { return i == -1; }; // Special case: shuffle(concat(A,B)) can be more efficiently represented // as concat(shuffle(A,B),UNDEF) if the shuffle doesn't set any of the high // half vector elements. if (NumElemsPerConcat * 2 == NumElts && N1.isUndef() && llvm::all_of(Mask.slice(NumElemsPerConcat, NumElemsPerConcat), IsUndefMaskElt)) { N0 = DAG.getVectorShuffle(ConcatVT, SDLoc(N), N0.getOperand(0), N0.getOperand(1), Mask.slice(0, NumElemsPerConcat)); N1 = DAG.getUNDEF(ConcatVT); return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, N0, N1); } // Look at every vector that's inserted. We're looking for exact // subvector-sized copies from a concatenated vector for (unsigned I = 0; I != NumConcats; ++I) { unsigned Begin = I * NumElemsPerConcat; ArrayRef SubMask = Mask.slice(Begin, NumElemsPerConcat); // Make sure we're dealing with a copy. if (llvm::all_of(SubMask, IsUndefMaskElt)) { Ops.push_back(DAG.getUNDEF(ConcatVT)); continue; } int OpIdx = -1; for (int i = 0; i != (int)NumElemsPerConcat; ++i) { if (IsUndefMaskElt(SubMask[i])) continue; if ((SubMask[i] % (int)NumElemsPerConcat) != i) return SDValue(); int EltOpIdx = SubMask[i] / NumElemsPerConcat; if (0 <= OpIdx && EltOpIdx != OpIdx) return SDValue(); OpIdx = EltOpIdx; } assert(0 <= OpIdx && "Unknown concat_vectors op"); if (OpIdx < (int)N0.getNumOperands()) Ops.push_back(N0.getOperand(OpIdx)); else Ops.push_back(N1.getOperand(OpIdx - N0.getNumOperands())); } return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops); } // Attempt to combine a shuffle of 2 inputs of 'scalar sources' - // BUILD_VECTOR or SCALAR_TO_VECTOR into a single BUILD_VECTOR. // // SHUFFLE(BUILD_VECTOR(), BUILD_VECTOR()) -> BUILD_VECTOR() is always // a simplification in some sense, but it isn't appropriate in general: some // BUILD_VECTORs are substantially cheaper than others. The general case // of a BUILD_VECTOR requires inserting each element individually (or // performing the equivalent in a temporary stack variable). A BUILD_VECTOR of // all constants is a single constant pool load. A BUILD_VECTOR where each // element is identical is a splat. A BUILD_VECTOR where most of the operands // are undef lowers to a small number of element insertions. // // To deal with this, we currently use a bunch of mostly arbitrary heuristics. // We don't fold shuffles where one side is a non-zero constant, and we don't // fold shuffles if the resulting (non-splat) BUILD_VECTOR would have duplicate // non-constant operands. This seems to work out reasonably well in practice. static SDValue combineShuffleOfScalars(ShuffleVectorSDNode *SVN, SelectionDAG &DAG, const TargetLowering &TLI) { EVT VT = SVN->getValueType(0); unsigned NumElts = VT.getVectorNumElements(); SDValue N0 = SVN->getOperand(0); SDValue N1 = SVN->getOperand(1); if (!N0->hasOneUse()) return SDValue(); // If only one of N1,N2 is constant, bail out if it is not ALL_ZEROS as // discussed above. if (!N1.isUndef()) { if (!N1->hasOneUse()) return SDValue(); bool N0AnyConst = isAnyConstantBuildVector(N0); bool N1AnyConst = isAnyConstantBuildVector(N1); if (N0AnyConst && !N1AnyConst && !ISD::isBuildVectorAllZeros(N0.getNode())) return SDValue(); if (!N0AnyConst && N1AnyConst && !ISD::isBuildVectorAllZeros(N1.getNode())) return SDValue(); } // If both inputs are splats of the same value then we can safely merge this // to a single BUILD_VECTOR with undef elements based on the shuffle mask. bool IsSplat = false; auto *BV0 = dyn_cast(N0); auto *BV1 = dyn_cast(N1); if (BV0 && BV1) if (SDValue Splat0 = BV0->getSplatValue()) IsSplat = (Splat0 == BV1->getSplatValue()); SmallVector Ops; SmallSet DuplicateOps; for (int M : SVN->getMask()) { SDValue Op = DAG.getUNDEF(VT.getScalarType()); if (M >= 0) { int Idx = M < (int)NumElts ? M : M - NumElts; SDValue &S = (M < (int)NumElts ? N0 : N1); if (S.getOpcode() == ISD::BUILD_VECTOR) { Op = S.getOperand(Idx); } else if (S.getOpcode() == ISD::SCALAR_TO_VECTOR) { SDValue Op0 = S.getOperand(0); Op = Idx == 0 ? Op0 : DAG.getUNDEF(Op0.getValueType()); } else { // Operand can't be combined - bail out. return SDValue(); } } // Don't duplicate a non-constant BUILD_VECTOR operand unless we're // generating a splat; semantically, this is fine, but it's likely to // generate low-quality code if the target can't reconstruct an appropriate // shuffle. if (!Op.isUndef() && !isIntOrFPConstant(Op)) if (!IsSplat && !DuplicateOps.insert(Op).second) return SDValue(); Ops.push_back(Op); } // BUILD_VECTOR requires all inputs to be of the same type, find the // maximum type and extend them all. EVT SVT = VT.getScalarType(); if (SVT.isInteger()) for (SDValue &Op : Ops) SVT = (SVT.bitsLT(Op.getValueType()) ? Op.getValueType() : SVT); if (SVT != VT.getScalarType()) for (SDValue &Op : Ops) Op = Op.isUndef() ? DAG.getUNDEF(SVT) : (TLI.isZExtFree(Op.getValueType(), SVT) ? DAG.getZExtOrTrunc(Op, SDLoc(SVN), SVT) : DAG.getSExtOrTrunc(Op, SDLoc(SVN), SVT)); return DAG.getBuildVector(VT, SDLoc(SVN), Ops); } // Match shuffles that can be converted to any_vector_extend_in_reg. // This is often generated during legalization. // e.g. v4i32 <0,u,1,u> -> (v2i64 any_vector_extend_in_reg(v4i32 src)) // TODO Add support for ZERO_EXTEND_VECTOR_INREG when we have a test case. static SDValue combineShuffleToVectorExtend(ShuffleVectorSDNode *SVN, SelectionDAG &DAG, const TargetLowering &TLI, bool LegalOperations) { EVT VT = SVN->getValueType(0); bool IsBigEndian = DAG.getDataLayout().isBigEndian(); // TODO Add support for big-endian when we have a test case. if (!VT.isInteger() || IsBigEndian) return SDValue(); unsigned NumElts = VT.getVectorNumElements(); unsigned EltSizeInBits = VT.getScalarSizeInBits(); ArrayRef Mask = SVN->getMask(); SDValue N0 = SVN->getOperand(0); // shuffle<0,-1,1,-1> == (v2i64 anyextend_vector_inreg(v4i32)) auto isAnyExtend = [&Mask, &NumElts](unsigned Scale) { for (unsigned i = 0; i != NumElts; ++i) { if (Mask[i] < 0) continue; if ((i % Scale) == 0 && Mask[i] == (int)(i / Scale)) continue; return false; } return true; }; // Attempt to match a '*_extend_vector_inreg' shuffle, we just search for // power-of-2 extensions as they are the most likely. for (unsigned Scale = 2; Scale < NumElts; Scale *= 2) { // Check for non power of 2 vector sizes if (NumElts % Scale != 0) continue; if (!isAnyExtend(Scale)) continue; EVT OutSVT = EVT::getIntegerVT(*DAG.getContext(), EltSizeInBits * Scale); EVT OutVT = EVT::getVectorVT(*DAG.getContext(), OutSVT, NumElts / Scale); // Never create an illegal type. Only create unsupported operations if we // are pre-legalization. if (TLI.isTypeLegal(OutVT)) if (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::ANY_EXTEND_VECTOR_INREG, OutVT)) return DAG.getBitcast(VT, DAG.getNode(ISD::ANY_EXTEND_VECTOR_INREG, SDLoc(SVN), OutVT, N0)); } return SDValue(); } // Detect 'truncate_vector_inreg' style shuffles that pack the lower parts of // each source element of a large type into the lowest elements of a smaller // destination type. This is often generated during legalization. // If the source node itself was a '*_extend_vector_inreg' node then we should // then be able to remove it. static SDValue combineTruncationShuffle(ShuffleVectorSDNode *SVN, SelectionDAG &DAG) { EVT VT = SVN->getValueType(0); bool IsBigEndian = DAG.getDataLayout().isBigEndian(); // TODO Add support for big-endian when we have a test case. if (!VT.isInteger() || IsBigEndian) return SDValue(); SDValue N0 = peekThroughBitcasts(SVN->getOperand(0)); unsigned Opcode = N0.getOpcode(); if (Opcode != ISD::ANY_EXTEND_VECTOR_INREG && Opcode != ISD::SIGN_EXTEND_VECTOR_INREG && Opcode != ISD::ZERO_EXTEND_VECTOR_INREG) return SDValue(); SDValue N00 = N0.getOperand(0); ArrayRef Mask = SVN->getMask(); unsigned NumElts = VT.getVectorNumElements(); unsigned EltSizeInBits = VT.getScalarSizeInBits(); unsigned ExtSrcSizeInBits = N00.getScalarValueSizeInBits(); unsigned ExtDstSizeInBits = N0.getScalarValueSizeInBits(); if (ExtDstSizeInBits % ExtSrcSizeInBits != 0) return SDValue(); unsigned ExtScale = ExtDstSizeInBits / ExtSrcSizeInBits; // (v4i32 truncate_vector_inreg(v2i64)) == shuffle<0,2-1,-1> // (v8i16 truncate_vector_inreg(v4i32)) == shuffle<0,2,4,6,-1,-1,-1,-1> // (v8i16 truncate_vector_inreg(v2i64)) == shuffle<0,4,-1,-1,-1,-1,-1,-1> auto isTruncate = [&Mask, &NumElts](unsigned Scale) { for (unsigned i = 0; i != NumElts; ++i) { if (Mask[i] < 0) continue; if ((i * Scale) < NumElts && Mask[i] == (int)(i * Scale)) continue; return false; } return true; }; // At the moment we just handle the case where we've truncated back to the // same size as before the extension. // TODO: handle more extension/truncation cases as cases arise. if (EltSizeInBits != ExtSrcSizeInBits) return SDValue(); // We can remove *extend_vector_inreg only if the truncation happens at // the same scale as the extension. if (isTruncate(ExtScale)) return DAG.getBitcast(VT, N00); return SDValue(); } // Combine shuffles of splat-shuffles of the form: // shuffle (shuffle V, undef, splat-mask), undef, M // If splat-mask contains undef elements, we need to be careful about // introducing undef's in the folded mask which are not the result of composing // the masks of the shuffles. static SDValue combineShuffleOfSplatVal(ShuffleVectorSDNode *Shuf, SelectionDAG &DAG) { if (!Shuf->getOperand(1).isUndef()) return SDValue(); // If the inner operand is a known splat with no undefs, just return that directly. // TODO: Create DemandedElts mask from Shuf's mask. // TODO: Allow undef elements and merge with the shuffle code below. if (DAG.isSplatValue(Shuf->getOperand(0), /*AllowUndefs*/ false)) return Shuf->getOperand(0); auto *Splat = dyn_cast(Shuf->getOperand(0)); if (!Splat || !Splat->isSplat()) return SDValue(); ArrayRef ShufMask = Shuf->getMask(); ArrayRef SplatMask = Splat->getMask(); assert(ShufMask.size() == SplatMask.size() && "Mask length mismatch"); // Prefer simplifying to the splat-shuffle, if possible. This is legal if // every undef mask element in the splat-shuffle has a corresponding undef // element in the user-shuffle's mask or if the composition of mask elements // would result in undef. // Examples for (shuffle (shuffle v, undef, SplatMask), undef, UserMask): // * UserMask=[0,2,u,u], SplatMask=[2,u,2,u] -> [2,2,u,u] // In this case it is not legal to simplify to the splat-shuffle because we // may be exposing the users of the shuffle an undef element at index 1 // which was not there before the combine. // * UserMask=[0,u,2,u], SplatMask=[2,u,2,u] -> [2,u,2,u] // In this case the composition of masks yields SplatMask, so it's ok to // simplify to the splat-shuffle. // * UserMask=[3,u,2,u], SplatMask=[2,u,2,u] -> [u,u,2,u] // In this case the composed mask includes all undef elements of SplatMask // and in addition sets element zero to undef. It is safe to simplify to // the splat-shuffle. auto CanSimplifyToExistingSplat = [](ArrayRef UserMask, ArrayRef SplatMask) { for (unsigned i = 0, e = UserMask.size(); i != e; ++i) if (UserMask[i] != -1 && SplatMask[i] == -1 && SplatMask[UserMask[i]] != -1) return false; return true; }; if (CanSimplifyToExistingSplat(ShufMask, SplatMask)) return Shuf->getOperand(0); // Create a new shuffle with a mask that is composed of the two shuffles' // masks. SmallVector NewMask; for (int Idx : ShufMask) NewMask.push_back(Idx == -1 ? -1 : SplatMask[Idx]); return DAG.getVectorShuffle(Splat->getValueType(0), SDLoc(Splat), Splat->getOperand(0), Splat->getOperand(1), NewMask); } // Combine shuffles of bitcasts into a shuffle of the bitcast type, providing // the mask can be treated as a larger type. static SDValue combineShuffleOfBitcast(ShuffleVectorSDNode *SVN, SelectionDAG &DAG, const TargetLowering &TLI, bool LegalOperations) { SDValue Op0 = SVN->getOperand(0); SDValue Op1 = SVN->getOperand(1); EVT VT = SVN->getValueType(0); if (Op0.getOpcode() != ISD::BITCAST) return SDValue(); EVT InVT = Op0.getOperand(0).getValueType(); if (!InVT.isVector() || (!Op1.isUndef() && (Op1.getOpcode() != ISD::BITCAST || Op1.getOperand(0).getValueType() != InVT))) return SDValue(); if (isAnyConstantBuildVector(Op0.getOperand(0)) && (Op1.isUndef() || isAnyConstantBuildVector(Op1.getOperand(0)))) return SDValue(); int VTLanes = VT.getVectorNumElements(); int InLanes = InVT.getVectorNumElements(); if (VTLanes <= InLanes || VTLanes % InLanes != 0 || (LegalOperations && !TLI.isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, InVT))) return SDValue(); int Factor = VTLanes / InLanes; // Check that each group of lanes in the mask are either undef or make a valid // mask for the wider lane type. ArrayRef Mask = SVN->getMask(); SmallVector NewMask; if (!widenShuffleMaskElts(Factor, Mask, NewMask)) return SDValue(); if (!TLI.isShuffleMaskLegal(NewMask, InVT)) return SDValue(); // Create the new shuffle with the new mask and bitcast it back to the // original type. SDLoc DL(SVN); Op0 = Op0.getOperand(0); Op1 = Op1.isUndef() ? DAG.getUNDEF(InVT) : Op1.getOperand(0); SDValue NewShuf = DAG.getVectorShuffle(InVT, DL, Op0, Op1, NewMask); return DAG.getBitcast(VT, NewShuf); } /// Combine shuffle of shuffle of the form: /// shuf (shuf X, undef, InnerMask), undef, OuterMask --> splat X static SDValue formSplatFromShuffles(ShuffleVectorSDNode *OuterShuf, SelectionDAG &DAG) { if (!OuterShuf->getOperand(1).isUndef()) return SDValue(); auto *InnerShuf = dyn_cast(OuterShuf->getOperand(0)); if (!InnerShuf || !InnerShuf->getOperand(1).isUndef()) return SDValue(); ArrayRef OuterMask = OuterShuf->getMask(); ArrayRef InnerMask = InnerShuf->getMask(); unsigned NumElts = OuterMask.size(); assert(NumElts == InnerMask.size() && "Mask length mismatch"); SmallVector CombinedMask(NumElts, -1); int SplatIndex = -1; for (unsigned i = 0; i != NumElts; ++i) { // Undef lanes remain undef. int OuterMaskElt = OuterMask[i]; if (OuterMaskElt == -1) continue; // Peek through the shuffle masks to get the underlying source element. int InnerMaskElt = InnerMask[OuterMaskElt]; if (InnerMaskElt == -1) continue; // Initialize the splatted element. if (SplatIndex == -1) SplatIndex = InnerMaskElt; // Non-matching index - this is not a splat. if (SplatIndex != InnerMaskElt) return SDValue(); CombinedMask[i] = InnerMaskElt; } assert((all_of(CombinedMask, [](int M) { return M == -1; }) || getSplatIndex(CombinedMask) != -1) && "Expected a splat mask"); // TODO: The transform may be a win even if the mask is not legal. EVT VT = OuterShuf->getValueType(0); assert(VT == InnerShuf->getValueType(0) && "Expected matching shuffle types"); if (!DAG.getTargetLoweringInfo().isShuffleMaskLegal(CombinedMask, VT)) return SDValue(); return DAG.getVectorShuffle(VT, SDLoc(OuterShuf), InnerShuf->getOperand(0), InnerShuf->getOperand(1), CombinedMask); } /// If the shuffle mask is taking exactly one element from the first vector /// operand and passing through all other elements from the second vector /// operand, return the index of the mask element that is choosing an element /// from the first operand. Otherwise, return -1. static int getShuffleMaskIndexOfOneElementFromOp0IntoOp1(ArrayRef Mask) { int MaskSize = Mask.size(); int EltFromOp0 = -1; // TODO: This does not match if there are undef elements in the shuffle mask. // Should we ignore undefs in the shuffle mask instead? The trade-off is // removing an instruction (a shuffle), but losing the knowledge that some // vector lanes are not needed. for (int i = 0; i != MaskSize; ++i) { if (Mask[i] >= 0 && Mask[i] < MaskSize) { // We're looking for a shuffle of exactly one element from operand 0. if (EltFromOp0 != -1) return -1; EltFromOp0 = i; } else if (Mask[i] != i + MaskSize) { // Nothing from operand 1 can change lanes. return -1; } } return EltFromOp0; } /// If a shuffle inserts exactly one element from a source vector operand into /// another vector operand and we can access the specified element as a scalar, /// then we can eliminate the shuffle. static SDValue replaceShuffleOfInsert(ShuffleVectorSDNode *Shuf, SelectionDAG &DAG) { // First, check if we are taking one element of a vector and shuffling that // element into another vector. ArrayRef Mask = Shuf->getMask(); SmallVector CommutedMask(Mask.begin(), Mask.end()); SDValue Op0 = Shuf->getOperand(0); SDValue Op1 = Shuf->getOperand(1); int ShufOp0Index = getShuffleMaskIndexOfOneElementFromOp0IntoOp1(Mask); if (ShufOp0Index == -1) { // Commute mask and check again. ShuffleVectorSDNode::commuteMask(CommutedMask); ShufOp0Index = getShuffleMaskIndexOfOneElementFromOp0IntoOp1(CommutedMask); if (ShufOp0Index == -1) return SDValue(); // Commute operands to match the commuted shuffle mask. std::swap(Op0, Op1); Mask = CommutedMask; } // The shuffle inserts exactly one element from operand 0 into operand 1. // Now see if we can access that element as a scalar via a real insert element // instruction. // TODO: We can try harder to locate the element as a scalar. Examples: it // could be an operand of SCALAR_TO_VECTOR, BUILD_VECTOR, or a constant. assert(Mask[ShufOp0Index] >= 0 && Mask[ShufOp0Index] < (int)Mask.size() && "Shuffle mask value must be from operand 0"); if (Op0.getOpcode() != ISD::INSERT_VECTOR_ELT) return SDValue(); auto *InsIndexC = dyn_cast(Op0.getOperand(2)); if (!InsIndexC || InsIndexC->getSExtValue() != Mask[ShufOp0Index]) return SDValue(); // There's an existing insertelement with constant insertion index, so we // don't need to check the legality/profitability of a replacement operation // that differs at most in the constant value. The target should be able to // lower any of those in a similar way. If not, legalization will expand this // to a scalar-to-vector plus shuffle. // // Note that the shuffle may move the scalar from the position that the insert // element used. Therefore, our new insert element occurs at the shuffle's // mask index value, not the insert's index value. // shuffle (insertelt v1, x, C), v2, mask --> insertelt v2, x, C' SDValue NewInsIndex = DAG.getVectorIdxConstant(ShufOp0Index, SDLoc(Shuf)); return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(Shuf), Op0.getValueType(), Op1, Op0.getOperand(1), NewInsIndex); } /// If we have a unary shuffle of a shuffle, see if it can be folded away /// completely. This has the potential to lose undef knowledge because the first /// shuffle may not have an undef mask element where the second one does. So /// only call this after doing simplifications based on demanded elements. static SDValue simplifyShuffleOfShuffle(ShuffleVectorSDNode *Shuf) { // shuf (shuf0 X, Y, Mask0), undef, Mask auto *Shuf0 = dyn_cast(Shuf->getOperand(0)); if (!Shuf0 || !Shuf->getOperand(1).isUndef()) return SDValue(); ArrayRef Mask = Shuf->getMask(); ArrayRef Mask0 = Shuf0->getMask(); for (int i = 0, e = (int)Mask.size(); i != e; ++i) { // Ignore undef elements. if (Mask[i] == -1) continue; assert(Mask[i] >= 0 && Mask[i] < e && "Unexpected shuffle mask value"); // Is the element of the shuffle operand chosen by this shuffle the same as // the element chosen by the shuffle operand itself? if (Mask0[Mask[i]] != Mask0[i]) return SDValue(); } // Every element of this shuffle is identical to the result of the previous // shuffle, so we can replace this value. return Shuf->getOperand(0); } SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) { EVT VT = N->getValueType(0); unsigned NumElts = VT.getVectorNumElements(); SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG"); // Canonicalize shuffle undef, undef -> undef if (N0.isUndef() && N1.isUndef()) return DAG.getUNDEF(VT); ShuffleVectorSDNode *SVN = cast(N); // Canonicalize shuffle v, v -> v, undef if (N0 == N1) return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT), createUnaryMask(SVN->getMask(), NumElts)); // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask. if (N0.isUndef()) return DAG.getCommutedVectorShuffle(*SVN); // Remove references to rhs if it is undef if (N1.isUndef()) { bool Changed = false; SmallVector NewMask; for (unsigned i = 0; i != NumElts; ++i) { int Idx = SVN->getMaskElt(i); if (Idx >= (int)NumElts) { Idx = -1; Changed = true; } NewMask.push_back(Idx); } if (Changed) return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, NewMask); } if (SDValue InsElt = replaceShuffleOfInsert(SVN, DAG)) return InsElt; // A shuffle of a single vector that is a splatted value can always be folded. if (SDValue V = combineShuffleOfSplatVal(SVN, DAG)) return V; if (SDValue V = formSplatFromShuffles(SVN, DAG)) return V; // If it is a splat, check if the argument vector is another splat or a // build_vector. if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) { int SplatIndex = SVN->getSplatIndex(); if (N0.hasOneUse() && TLI.isExtractVecEltCheap(VT, SplatIndex) && TLI.isBinOp(N0.getOpcode()) && N0->getNumValues() == 1) { // splat (vector_bo L, R), Index --> // splat (scalar_bo (extelt L, Index), (extelt R, Index)) SDValue L = N0.getOperand(0), R = N0.getOperand(1); SDLoc DL(N); EVT EltVT = VT.getScalarType(); SDValue Index = DAG.getVectorIdxConstant(SplatIndex, DL); SDValue ExtL = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, L, Index); SDValue ExtR = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, R, Index); SDValue NewBO = DAG.getNode(N0.getOpcode(), DL, EltVT, ExtL, ExtR, N0->getFlags()); SDValue Insert = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VT, NewBO); SmallVector ZeroMask(VT.getVectorNumElements(), 0); return DAG.getVectorShuffle(VT, DL, Insert, DAG.getUNDEF(VT), ZeroMask); } // splat(scalar_to_vector(x), 0) -> build_vector(x,...,x) // splat(insert_vector_elt(v, x, c), c) -> build_vector(x,...,x) if ((!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) && N0.hasOneUse()) { if (N0.getOpcode() == ISD::SCALAR_TO_VECTOR && SplatIndex == 0) return DAG.getSplatBuildVector(VT, SDLoc(N), N0.getOperand(0)); if (N0.getOpcode() == ISD::INSERT_VECTOR_ELT) if (auto *Idx = dyn_cast(N0.getOperand(2))) if (Idx->getAPIntValue() == SplatIndex) return DAG.getSplatBuildVector(VT, SDLoc(N), N0.getOperand(1)); } // If this is a bit convert that changes the element type of the vector but // not the number of vector elements, look through it. Be careful not to // look though conversions that change things like v4f32 to v2f64. SDNode *V = N0.getNode(); if (V->getOpcode() == ISD::BITCAST) { SDValue ConvInput = V->getOperand(0); if (ConvInput.getValueType().isVector() && ConvInput.getValueType().getVectorNumElements() == NumElts) V = ConvInput.getNode(); } if (V->getOpcode() == ISD::BUILD_VECTOR) { assert(V->getNumOperands() == NumElts && "BUILD_VECTOR has wrong number of operands"); SDValue Base; bool AllSame = true; for (unsigned i = 0; i != NumElts; ++i) { if (!V->getOperand(i).isUndef()) { Base = V->getOperand(i); break; } } // Splat of , return if (!Base.getNode()) return N0; for (unsigned i = 0; i != NumElts; ++i) { if (V->getOperand(i) != Base) { AllSame = false; break; } } // Splat of , return if (AllSame) return N0; // Canonicalize any other splat as a build_vector. SDValue Splatted = V->getOperand(SplatIndex); SmallVector Ops(NumElts, Splatted); SDValue NewBV = DAG.getBuildVector(V->getValueType(0), SDLoc(N), Ops); // We may have jumped through bitcasts, so the type of the // BUILD_VECTOR may not match the type of the shuffle. if (V->getValueType(0) != VT) NewBV = DAG.getBitcast(VT, NewBV); return NewBV; } } // Simplify source operands based on shuffle mask. if (SimplifyDemandedVectorElts(SDValue(N, 0))) return SDValue(N, 0); // This is intentionally placed after demanded elements simplification because // it could eliminate knowledge of undef elements created by this shuffle. if (SDValue ShufOp = simplifyShuffleOfShuffle(SVN)) return ShufOp; // Match shuffles that can be converted to any_vector_extend_in_reg. if (SDValue V = combineShuffleToVectorExtend(SVN, DAG, TLI, LegalOperations)) return V; // Combine "truncate_vector_in_reg" style shuffles. if (SDValue V = combineTruncationShuffle(SVN, DAG)) return V; if (N0.getOpcode() == ISD::CONCAT_VECTORS && Level < AfterLegalizeVectorOps && (N1.isUndef() || (N1.getOpcode() == ISD::CONCAT_VECTORS && N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) { if (SDValue V = partitionShuffleOfConcats(N, DAG)) return V; } // A shuffle of a concat of the same narrow vector can be reduced to use // only low-half elements of a concat with undef: // shuf (concat X, X), undef, Mask --> shuf (concat X, undef), undef, Mask' if (N0.getOpcode() == ISD::CONCAT_VECTORS && N1.isUndef() && N0.getNumOperands() == 2 && N0.getOperand(0) == N0.getOperand(1)) { int HalfNumElts = (int)NumElts / 2; SmallVector NewMask; for (unsigned i = 0; i != NumElts; ++i) { int Idx = SVN->getMaskElt(i); if (Idx >= HalfNumElts) { assert(Idx < (int)NumElts && "Shuffle mask chooses undef op"); Idx -= HalfNumElts; } NewMask.push_back(Idx); } if (TLI.isShuffleMaskLegal(NewMask, VT)) { SDValue UndefVec = DAG.getUNDEF(N0.getOperand(0).getValueType()); SDValue NewCat = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, N0.getOperand(0), UndefVec); return DAG.getVectorShuffle(VT, SDLoc(N), NewCat, N1, NewMask); } } // See if we can replace a shuffle with an insert_subvector. // e.g. v2i32 into v8i32: // shuffle(lhs,concat(rhs0,rhs1,rhs2,rhs3),0,1,2,3,10,11,6,7). // --> insert_subvector(lhs,rhs1,4). if (Level < AfterLegalizeVectorOps && TLI.isTypeLegal(VT) && TLI.isOperationLegalOrCustom(ISD::INSERT_SUBVECTOR, VT)) { auto ShuffleToInsert = [&](SDValue LHS, SDValue RHS, ArrayRef Mask) { // Ensure RHS subvectors are legal. assert(RHS.getOpcode() == ISD::CONCAT_VECTORS && "Can't find subvectors"); EVT SubVT = RHS.getOperand(0).getValueType(); int NumSubVecs = RHS.getNumOperands(); int NumSubElts = SubVT.getVectorNumElements(); assert((NumElts % NumSubElts) == 0 && "Subvector mismatch"); if (!TLI.isTypeLegal(SubVT)) return SDValue(); // Don't bother if we have an unary shuffle (matches undef + LHS elts). if (all_of(Mask, [NumElts](int M) { return M < (int)NumElts; })) return SDValue(); // Search [NumSubElts] spans for RHS sequence. // TODO: Can we avoid nested loops to increase performance? SmallVector InsertionMask(NumElts); for (int SubVec = 0; SubVec != NumSubVecs; ++SubVec) { for (int SubIdx = 0; SubIdx != (int)NumElts; SubIdx += NumSubElts) { // Reset mask to identity. std::iota(InsertionMask.begin(), InsertionMask.end(), 0); // Add subvector insertion. std::iota(InsertionMask.begin() + SubIdx, InsertionMask.begin() + SubIdx + NumSubElts, NumElts + (SubVec * NumSubElts)); // See if the shuffle mask matches the reference insertion mask. bool MatchingShuffle = true; for (int i = 0; i != (int)NumElts; ++i) { int ExpectIdx = InsertionMask[i]; int ActualIdx = Mask[i]; if (0 <= ActualIdx && ExpectIdx != ActualIdx) { MatchingShuffle = false; break; } } if (MatchingShuffle) return DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N), VT, LHS, RHS.getOperand(SubVec), DAG.getVectorIdxConstant(SubIdx, SDLoc(N))); } } return SDValue(); }; ArrayRef Mask = SVN->getMask(); if (N1.getOpcode() == ISD::CONCAT_VECTORS) if (SDValue InsertN1 = ShuffleToInsert(N0, N1, Mask)) return InsertN1; if (N0.getOpcode() == ISD::CONCAT_VECTORS) { SmallVector CommuteMask(Mask.begin(), Mask.end()); ShuffleVectorSDNode::commuteMask(CommuteMask); if (SDValue InsertN0 = ShuffleToInsert(N1, N0, CommuteMask)) return InsertN0; } } // If we're not performing a select/blend shuffle, see if we can convert the // shuffle into a AND node, with all the out-of-lane elements are known zero. if (Level < AfterLegalizeDAG && TLI.isTypeLegal(VT)) { bool IsInLaneMask = true; ArrayRef Mask = SVN->getMask(); SmallVector ClearMask(NumElts, -1); APInt DemandedLHS = APInt::getNullValue(NumElts); APInt DemandedRHS = APInt::getNullValue(NumElts); for (int I = 0; I != (int)NumElts; ++I) { int M = Mask[I]; if (M < 0) continue; ClearMask[I] = M == I ? I : (I + NumElts); IsInLaneMask &= (M == I) || (M == (int)(I + NumElts)); if (M != I) { APInt &Demanded = M < (int)NumElts ? DemandedLHS : DemandedRHS; Demanded.setBit(M % NumElts); } } // TODO: Should we try to mask with N1 as well? if (!IsInLaneMask && (!DemandedLHS.isNullValue() || !DemandedRHS.isNullValue()) && (DemandedLHS.isNullValue() || DAG.MaskedVectorIsZero(N0, DemandedLHS)) && (DemandedRHS.isNullValue() || DAG.MaskedVectorIsZero(N1, DemandedRHS))) { SDLoc DL(N); EVT IntVT = VT.changeVectorElementTypeToInteger(); EVT IntSVT = VT.getVectorElementType().changeTypeToInteger(); // Transform the type to a legal type so that the buildvector constant // elements are not illegal. Make sure that the result is larger than the // original type, incase the value is split into two (eg i64->i32). if (!TLI.isTypeLegal(IntSVT) && LegalTypes) IntSVT = TLI.getTypeToTransformTo(*DAG.getContext(), IntSVT); if (IntSVT.getSizeInBits() >= IntVT.getScalarSizeInBits()) { SDValue ZeroElt = DAG.getConstant(0, DL, IntSVT); SDValue AllOnesElt = DAG.getAllOnesConstant(DL, IntSVT); SmallVector AndMask(NumElts, DAG.getUNDEF(IntSVT)); for (int I = 0; I != (int)NumElts; ++I) if (0 <= Mask[I]) AndMask[I] = Mask[I] == I ? AllOnesElt : ZeroElt; // See if a clear mask is legal instead of going via // XformToShuffleWithZero which loses UNDEF mask elements. if (TLI.isVectorClearMaskLegal(ClearMask, IntVT)) return DAG.getBitcast( VT, DAG.getVectorShuffle(IntVT, DL, DAG.getBitcast(IntVT, N0), DAG.getConstant(0, DL, IntVT), ClearMask)); if (TLI.isOperationLegalOrCustom(ISD::AND, IntVT)) return DAG.getBitcast( VT, DAG.getNode(ISD::AND, DL, IntVT, DAG.getBitcast(IntVT, N0), DAG.getBuildVector(IntVT, DL, AndMask))); } } } // Attempt to combine a shuffle of 2 inputs of 'scalar sources' - // BUILD_VECTOR or SCALAR_TO_VECTOR into a single BUILD_VECTOR. if (Level < AfterLegalizeDAG && TLI.isTypeLegal(VT)) if (SDValue Res = combineShuffleOfScalars(SVN, DAG, TLI)) return Res; // If this shuffle only has a single input that is a bitcasted shuffle, // attempt to merge the 2 shuffles and suitably bitcast the inputs/output // back to their original types. if (N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() && N1.isUndef() && Level < AfterLegalizeVectorOps && TLI.isTypeLegal(VT)) { SDValue BC0 = peekThroughOneUseBitcasts(N0); if (BC0.getOpcode() == ISD::VECTOR_SHUFFLE && BC0.hasOneUse()) { EVT SVT = VT.getScalarType(); EVT InnerVT = BC0->getValueType(0); EVT InnerSVT = InnerVT.getScalarType(); // Determine which shuffle works with the smaller scalar type. EVT ScaleVT = SVT.bitsLT(InnerSVT) ? VT : InnerVT; EVT ScaleSVT = ScaleVT.getScalarType(); if (TLI.isTypeLegal(ScaleVT) && 0 == (InnerSVT.getSizeInBits() % ScaleSVT.getSizeInBits()) && 0 == (SVT.getSizeInBits() % ScaleSVT.getSizeInBits())) { int InnerScale = InnerSVT.getSizeInBits() / ScaleSVT.getSizeInBits(); int OuterScale = SVT.getSizeInBits() / ScaleSVT.getSizeInBits(); // Scale the shuffle masks to the smaller scalar type. ShuffleVectorSDNode *InnerSVN = cast(BC0); SmallVector InnerMask; SmallVector OuterMask; narrowShuffleMaskElts(InnerScale, InnerSVN->getMask(), InnerMask); narrowShuffleMaskElts(OuterScale, SVN->getMask(), OuterMask); // Merge the shuffle masks. SmallVector NewMask; for (int M : OuterMask) NewMask.push_back(M < 0 ? -1 : InnerMask[M]); // Test for shuffle mask legality over both commutations. SDValue SV0 = BC0->getOperand(0); SDValue SV1 = BC0->getOperand(1); bool LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT); if (!LegalMask) { std::swap(SV0, SV1); ShuffleVectorSDNode::commuteMask(NewMask); LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT); } if (LegalMask) { SV0 = DAG.getBitcast(ScaleVT, SV0); SV1 = DAG.getBitcast(ScaleVT, SV1); return DAG.getBitcast( VT, DAG.getVectorShuffle(ScaleVT, SDLoc(N), SV0, SV1, NewMask)); } } } } // Match shuffles of bitcasts, so long as the mask can be treated as the // larger type. if (SDValue V = combineShuffleOfBitcast(SVN, DAG, TLI, LegalOperations)) return V; // Compute the combined shuffle mask for a shuffle with SV0 as the first // operand, and SV1 as the second operand. // i.e. Merge SVN(OtherSVN, N1) -> shuffle(SV0, SV1, Mask) iff Commute = false // Merge SVN(N1, OtherSVN) -> shuffle(SV0, SV1, Mask') iff Commute = true auto MergeInnerShuffle = [NumElts, &VT](bool Commute, ShuffleVectorSDNode *SVN, ShuffleVectorSDNode *OtherSVN, SDValue N1, const TargetLowering &TLI, SDValue &SV0, SDValue &SV1, SmallVectorImpl &Mask) -> bool { // Don't try to fold splats; they're likely to simplify somehow, or they // might be free. if (OtherSVN->isSplat()) return false; SV0 = SV1 = SDValue(); Mask.clear(); for (unsigned i = 0; i != NumElts; ++i) { int Idx = SVN->getMaskElt(i); if (Idx < 0) { // Propagate Undef. Mask.push_back(Idx); continue; } if (Commute) Idx = (Idx < (int)NumElts) ? (Idx + NumElts) : (Idx - NumElts); SDValue CurrentVec; if (Idx < (int)NumElts) { // This shuffle index refers to the inner shuffle N0. Lookup the inner // shuffle mask to identify which vector is actually referenced. Idx = OtherSVN->getMaskElt(Idx); if (Idx < 0) { // Propagate Undef. Mask.push_back(Idx); continue; } CurrentVec = (Idx < (int)NumElts) ? OtherSVN->getOperand(0) : OtherSVN->getOperand(1); } else { // This shuffle index references an element within N1. CurrentVec = N1; } // Simple case where 'CurrentVec' is UNDEF. if (CurrentVec.isUndef()) { Mask.push_back(-1); continue; } // Canonicalize the shuffle index. We don't know yet if CurrentVec // will be the first or second operand of the combined shuffle. Idx = Idx % NumElts; if (!SV0.getNode() || SV0 == CurrentVec) { // Ok. CurrentVec is the left hand side. // Update the mask accordingly. SV0 = CurrentVec; Mask.push_back(Idx); continue; } if (!SV1.getNode() || SV1 == CurrentVec) { // Ok. CurrentVec is the right hand side. // Update the mask accordingly. SV1 = CurrentVec; Mask.push_back(Idx + NumElts); continue; } // Last chance - see if the vector is another shuffle and if it // uses one of the existing candidate shuffle ops. if (auto *CurrentSVN = dyn_cast(CurrentVec)) { int InnerIdx = CurrentSVN->getMaskElt(Idx); if (InnerIdx < 0) { Mask.push_back(-1); continue; } SDValue InnerVec = (InnerIdx < (int)NumElts) ? CurrentSVN->getOperand(0) : CurrentSVN->getOperand(1); if (InnerVec.isUndef()) { Mask.push_back(-1); continue; } InnerIdx %= NumElts; if (InnerVec == SV0) { Mask.push_back(InnerIdx); continue; } if (InnerVec == SV1) { Mask.push_back(InnerIdx + NumElts); continue; } } // Bail out if we cannot convert the shuffle pair into a single shuffle. return false; } if (llvm::all_of(Mask, [](int M) { return M < 0; })) return true; // Avoid introducing shuffles with illegal mask. // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2) // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2) // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2) // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, A, M2) // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, A, M2) // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, B, M2) if (TLI.isShuffleMaskLegal(Mask, VT)) return true; std::swap(SV0, SV1); ShuffleVectorSDNode::commuteMask(Mask); return TLI.isShuffleMaskLegal(Mask, VT); }; if (Level < AfterLegalizeDAG && TLI.isTypeLegal(VT)) { // Canonicalize shuffles according to rules: // shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A) // shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B) // shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B) if (N1.getOpcode() == ISD::VECTOR_SHUFFLE && N0.getOpcode() != ISD::VECTOR_SHUFFLE) { // The incoming shuffle must be of the same type as the result of the // current shuffle. assert(N1->getOperand(0).getValueType() == VT && "Shuffle types don't match"); SDValue SV0 = N1->getOperand(0); SDValue SV1 = N1->getOperand(1); bool HasSameOp0 = N0 == SV0; bool IsSV1Undef = SV1.isUndef(); if (HasSameOp0 || IsSV1Undef || N0 == SV1) // Commute the operands of this shuffle so merging below will trigger. return DAG.getCommutedVectorShuffle(*SVN); } // Canonicalize splat shuffles to the RHS to improve merging below. // shuffle(splat(A,u), shuffle(C,D)) -> shuffle'(shuffle(C,D), splat(A,u)) if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && N1.getOpcode() == ISD::VECTOR_SHUFFLE && cast(N0)->isSplat() && !cast(N1)->isSplat()) { return DAG.getCommutedVectorShuffle(*SVN); } // Try to fold according to rules: // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2) // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2) // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2) // Don't try to fold shuffles with illegal type. // Only fold if this shuffle is the only user of the other shuffle. // Try matching shuffle(C,shuffle(A,B)) commutted patterns as well. for (int i = 0; i != 2; ++i) { if (N->getOperand(i).getOpcode() == ISD::VECTOR_SHUFFLE && N->isOnlyUserOf(N->getOperand(i).getNode())) { // The incoming shuffle must be of the same type as the result of the // current shuffle. auto *OtherSV = cast(N->getOperand(i)); assert(OtherSV->getOperand(0).getValueType() == VT && "Shuffle types don't match"); SDValue SV0, SV1; SmallVector Mask; if (MergeInnerShuffle(i != 0, SVN, OtherSV, N->getOperand(1 - i), TLI, SV0, SV1, Mask)) { // Check if all indices in Mask are Undef. In case, propagate Undef. if (llvm::all_of(Mask, [](int M) { return M < 0; })) return DAG.getUNDEF(VT); return DAG.getVectorShuffle(VT, SDLoc(N), SV0 ? SV0 : DAG.getUNDEF(VT), SV1 ? SV1 : DAG.getUNDEF(VT), Mask); } } } // Merge shuffles through binops if we are able to merge it with at least // one other shuffles. // shuffle(bop(shuffle(x,y),shuffle(z,w)),undef) // shuffle(bop(shuffle(x,y),shuffle(z,w)),bop(shuffle(a,b),shuffle(c,d))) unsigned SrcOpcode = N0.getOpcode(); if (TLI.isBinOp(SrcOpcode) && N->isOnlyUserOf(N0.getNode()) && (N1.isUndef() || (SrcOpcode == N1.getOpcode() && N->isOnlyUserOf(N1.getNode())))) { // Get binop source ops, or just pass on the undef. SDValue Op00 = N0.getOperand(0); SDValue Op01 = N0.getOperand(1); SDValue Op10 = N1.isUndef() ? N1 : N1.getOperand(0); SDValue Op11 = N1.isUndef() ? N1 : N1.getOperand(1); // TODO: We might be able to relax the VT check but we don't currently // have any isBinOp() that has different result/ops VTs so play safe until // we have test coverage. if (Op00.getValueType() == VT && Op10.getValueType() == VT && Op01.getValueType() == VT && Op11.getValueType() == VT && (Op00.getOpcode() == ISD::VECTOR_SHUFFLE || Op10.getOpcode() == ISD::VECTOR_SHUFFLE || Op01.getOpcode() == ISD::VECTOR_SHUFFLE || Op11.getOpcode() == ISD::VECTOR_SHUFFLE)) { auto CanMergeInnerShuffle = [&](SDValue &SV0, SDValue &SV1, SmallVectorImpl &Mask, bool LeftOp, bool Commute) { SDValue InnerN = Commute ? N1 : N0; SDValue Op0 = LeftOp ? Op00 : Op01; SDValue Op1 = LeftOp ? Op10 : Op11; if (Commute) std::swap(Op0, Op1); // Only accept the merged shuffle if we don't introduce undef elements, // or the inner shuffle already contained undef elements. auto *SVN0 = dyn_cast(Op0); return SVN0 && InnerN->isOnlyUserOf(SVN0) && MergeInnerShuffle(Commute, SVN, SVN0, Op1, TLI, SV0, SV1, Mask) && (llvm::any_of(SVN0->getMask(), [](int M) { return M < 0; }) || llvm::none_of(Mask, [](int M) { return M < 0; })); }; // Ensure we don't increase the number of shuffles - we must merge a // shuffle from at least one of the LHS and RHS ops. bool MergedLeft = false; SDValue LeftSV0, LeftSV1; SmallVector LeftMask; if (CanMergeInnerShuffle(LeftSV0, LeftSV1, LeftMask, true, false) || CanMergeInnerShuffle(LeftSV0, LeftSV1, LeftMask, true, true)) { MergedLeft = true; } else { LeftMask.assign(SVN->getMask().begin(), SVN->getMask().end()); LeftSV0 = Op00, LeftSV1 = Op10; } bool MergedRight = false; SDValue RightSV0, RightSV1; SmallVector RightMask; if (CanMergeInnerShuffle(RightSV0, RightSV1, RightMask, false, false) || CanMergeInnerShuffle(RightSV0, RightSV1, RightMask, false, true)) { MergedRight = true; } else { RightMask.assign(SVN->getMask().begin(), SVN->getMask().end()); RightSV0 = Op01, RightSV1 = Op11; } if (MergedLeft || MergedRight) { SDLoc DL(N); SDValue LHS = DAG.getVectorShuffle( VT, DL, LeftSV0 ? LeftSV0 : DAG.getUNDEF(VT), LeftSV1 ? LeftSV1 : DAG.getUNDEF(VT), LeftMask); SDValue RHS = DAG.getVectorShuffle( VT, DL, RightSV0 ? RightSV0 : DAG.getUNDEF(VT), RightSV1 ? RightSV1 : DAG.getUNDEF(VT), RightMask); return DAG.getNode(SrcOpcode, DL, VT, LHS, RHS); } } } } if (SDValue V = foldShuffleOfConcatUndefs(SVN, DAG)) return V; return SDValue(); } SDValue DAGCombiner::visitSCALAR_TO_VECTOR(SDNode *N) { SDValue InVal = N->getOperand(0); EVT VT = N->getValueType(0); // Replace a SCALAR_TO_VECTOR(EXTRACT_VECTOR_ELT(V,C0)) pattern // with a VECTOR_SHUFFLE and possible truncate. if (InVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT && VT.isFixedLengthVector() && InVal->getOperand(0).getValueType().isFixedLengthVector()) { SDValue InVec = InVal->getOperand(0); SDValue EltNo = InVal->getOperand(1); auto InVecT = InVec.getValueType(); if (ConstantSDNode *C0 = dyn_cast(EltNo)) { SmallVector NewMask(InVecT.getVectorNumElements(), -1); int Elt = C0->getZExtValue(); NewMask[0] = Elt; // If we have an implict truncate do truncate here as long as it's legal. // if it's not legal, this should if (VT.getScalarType() != InVal.getValueType() && InVal.getValueType().isScalarInteger() && isTypeLegal(VT.getScalarType())) { SDValue Val = DAG.getNode(ISD::TRUNCATE, SDLoc(InVal), VT.getScalarType(), InVal); return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N), VT, Val); } if (VT.getScalarType() == InVecT.getScalarType() && VT.getVectorNumElements() <= InVecT.getVectorNumElements()) { SDValue LegalShuffle = TLI.buildLegalVectorShuffle(InVecT, SDLoc(N), InVec, DAG.getUNDEF(InVecT), NewMask, DAG); if (LegalShuffle) { // If the initial vector is the correct size this shuffle is a // valid result. if (VT == InVecT) return LegalShuffle; // If not we must truncate the vector. if (VT.getVectorNumElements() != InVecT.getVectorNumElements()) { SDValue ZeroIdx = DAG.getVectorIdxConstant(0, SDLoc(N)); EVT SubVT = EVT::getVectorVT(*DAG.getContext(), InVecT.getVectorElementType(), VT.getVectorNumElements()); return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N), SubVT, LegalShuffle, ZeroIdx); } } } } } return SDValue(); } SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) { EVT VT = N->getValueType(0); SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); SDValue N2 = N->getOperand(2); uint64_t InsIdx = N->getConstantOperandVal(2); // If inserting an UNDEF, just return the original vector. if (N1.isUndef()) return N0; // If this is an insert of an extracted vector into an undef vector, we can // just use the input to the extract. if (N0.isUndef() && N1.getOpcode() == ISD::EXTRACT_SUBVECTOR && N1.getOperand(1) == N2 && N1.getOperand(0).getValueType() == VT) return N1.getOperand(0); // Simplify scalar inserts into an undef vector: // insert_subvector undef, (splat X), N2 -> splat X if (N0.isUndef() && N1.getOpcode() == ISD::SPLAT_VECTOR) return DAG.getNode(ISD::SPLAT_VECTOR, SDLoc(N), VT, N1.getOperand(0)); // If we are inserting a bitcast value into an undef, with the same // number of elements, just use the bitcast input of the extract. // i.e. INSERT_SUBVECTOR UNDEF (BITCAST N1) N2 -> // BITCAST (INSERT_SUBVECTOR UNDEF N1 N2) if (N0.isUndef() && N1.getOpcode() == ISD::BITCAST && N1.getOperand(0).getOpcode() == ISD::EXTRACT_SUBVECTOR && N1.getOperand(0).getOperand(1) == N2 && N1.getOperand(0).getOperand(0).getValueType().getVectorElementCount() == VT.getVectorElementCount() && N1.getOperand(0).getOperand(0).getValueType().getSizeInBits() == VT.getSizeInBits()) { return DAG.getBitcast(VT, N1.getOperand(0).getOperand(0)); } // If both N1 and N2 are bitcast values on which insert_subvector // would makes sense, pull the bitcast through. // i.e. INSERT_SUBVECTOR (BITCAST N0) (BITCAST N1) N2 -> // BITCAST (INSERT_SUBVECTOR N0 N1 N2) if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST) { SDValue CN0 = N0.getOperand(0); SDValue CN1 = N1.getOperand(0); EVT CN0VT = CN0.getValueType(); EVT CN1VT = CN1.getValueType(); if (CN0VT.isVector() && CN1VT.isVector() && CN0VT.getVectorElementType() == CN1VT.getVectorElementType() && CN0VT.getVectorElementCount() == VT.getVectorElementCount()) { SDValue NewINSERT = DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N), CN0.getValueType(), CN0, CN1, N2); return DAG.getBitcast(VT, NewINSERT); } } // Combine INSERT_SUBVECTORs where we are inserting to the same index. // INSERT_SUBVECTOR( INSERT_SUBVECTOR( Vec, SubOld, Idx ), SubNew, Idx ) // --> INSERT_SUBVECTOR( Vec, SubNew, Idx ) if (N0.getOpcode() == ISD::INSERT_SUBVECTOR && N0.getOperand(1).getValueType() == N1.getValueType() && N0.getOperand(2) == N2) return DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N), VT, N0.getOperand(0), N1, N2); // Eliminate an intermediate insert into an undef vector: // insert_subvector undef, (insert_subvector undef, X, 0), N2 --> // insert_subvector undef, X, N2 if (N0.isUndef() && N1.getOpcode() == ISD::INSERT_SUBVECTOR && N1.getOperand(0).isUndef() && isNullConstant(N1.getOperand(2))) return DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N), VT, N0, N1.getOperand(1), N2); // Push subvector bitcasts to the output, adjusting the index as we go. // insert_subvector(bitcast(v), bitcast(s), c1) // -> bitcast(insert_subvector(v, s, c2)) if ((N0.isUndef() || N0.getOpcode() == ISD::BITCAST) && N1.getOpcode() == ISD::BITCAST) { SDValue N0Src = peekThroughBitcasts(N0); SDValue N1Src = peekThroughBitcasts(N1); EVT N0SrcSVT = N0Src.getValueType().getScalarType(); EVT N1SrcSVT = N1Src.getValueType().getScalarType(); if ((N0.isUndef() || N0SrcSVT == N1SrcSVT) && N0Src.getValueType().isVector() && N1Src.getValueType().isVector()) { EVT NewVT; SDLoc DL(N); SDValue NewIdx; LLVMContext &Ctx = *DAG.getContext(); ElementCount NumElts = VT.getVectorElementCount(); unsigned EltSizeInBits = VT.getScalarSizeInBits(); if ((EltSizeInBits % N1SrcSVT.getSizeInBits()) == 0) { unsigned Scale = EltSizeInBits / N1SrcSVT.getSizeInBits(); NewVT = EVT::getVectorVT(Ctx, N1SrcSVT, NumElts * Scale); NewIdx = DAG.getVectorIdxConstant(InsIdx * Scale, DL); } else if ((N1SrcSVT.getSizeInBits() % EltSizeInBits) == 0) { unsigned Scale = N1SrcSVT.getSizeInBits() / EltSizeInBits; if (NumElts.isKnownMultipleOf(Scale) && (InsIdx % Scale) == 0) { NewVT = EVT::getVectorVT(Ctx, N1SrcSVT, NumElts.divideCoefficientBy(Scale)); NewIdx = DAG.getVectorIdxConstant(InsIdx / Scale, DL); } } if (NewIdx && hasOperation(ISD::INSERT_SUBVECTOR, NewVT)) { SDValue Res = DAG.getBitcast(NewVT, N0Src); Res = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, NewVT, Res, N1Src, NewIdx); return DAG.getBitcast(VT, Res); } } } // Canonicalize insert_subvector dag nodes. // Example: // (insert_subvector (insert_subvector A, Idx0), Idx1) // -> (insert_subvector (insert_subvector A, Idx1), Idx0) if (N0.getOpcode() == ISD::INSERT_SUBVECTOR && N0.hasOneUse() && N1.getValueType() == N0.getOperand(1).getValueType()) { unsigned OtherIdx = N0.getConstantOperandVal(2); if (InsIdx < OtherIdx) { // Swap nodes. SDValue NewOp = DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N), VT, N0.getOperand(0), N1, N2); AddToWorklist(NewOp.getNode()); return DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N0.getNode()), VT, NewOp, N0.getOperand(1), N0.getOperand(2)); } } // If the input vector is a concatenation, and the insert replaces // one of the pieces, we can optimize into a single concat_vectors. if (N0.getOpcode() == ISD::CONCAT_VECTORS && N0.hasOneUse() && N0.getOperand(0).getValueType() == N1.getValueType() && N0.getOperand(0).getValueType().isScalableVector() == N1.getValueType().isScalableVector()) { unsigned Factor = N1.getValueType().getVectorMinNumElements(); SmallVector Ops(N0->op_begin(), N0->op_end()); Ops[InsIdx / Factor] = N1; return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops); } // Simplify source operands based on insertion. if (SimplifyDemandedVectorElts(SDValue(N, 0))) return SDValue(N, 0); return SDValue(); } SDValue DAGCombiner::visitFP_TO_FP16(SDNode *N) { SDValue N0 = N->getOperand(0); // fold (fp_to_fp16 (fp16_to_fp op)) -> op if (N0->getOpcode() == ISD::FP16_TO_FP) return N0->getOperand(0); return SDValue(); } SDValue DAGCombiner::visitFP16_TO_FP(SDNode *N) { SDValue N0 = N->getOperand(0); // fold fp16_to_fp(op & 0xffff) -> fp16_to_fp(op) if (!TLI.shouldKeepZExtForFP16Conv() && N0->getOpcode() == ISD::AND) { ConstantSDNode *AndConst = getAsNonOpaqueConstant(N0.getOperand(1)); if (AndConst && AndConst->getAPIntValue() == 0xffff) { return DAG.getNode(ISD::FP16_TO_FP, SDLoc(N), N->getValueType(0), N0.getOperand(0)); } } return SDValue(); } SDValue DAGCombiner::visitFP_TO_BF16(SDNode *N) { SDValue N0 = N->getOperand(0); // fold (fp_to_bf16 (bf16_to_fp op)) -> op if (N0->getOpcode() == ISD::BF16_TO_FP) return N0->getOperand(0); return SDValue(); } SDValue DAGCombiner::visitVECREDUCE(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N0.getValueType(); unsigned Opcode = N->getOpcode(); // VECREDUCE over 1-element vector is just an extract. if (VT.getVectorElementCount().isScalar()) { SDLoc dl(N); SDValue Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT.getVectorElementType(), N0, DAG.getVectorIdxConstant(0, dl)); if (Res.getValueType() != N->getValueType(0)) Res = DAG.getNode(ISD::ANY_EXTEND, dl, N->getValueType(0), Res); return Res; } // On an boolean vector an and/or reduction is the same as a umin/umax // reduction. Convert them if the latter is legal while the former isn't. if (Opcode == ISD::VECREDUCE_AND || Opcode == ISD::VECREDUCE_OR) { unsigned NewOpcode = Opcode == ISD::VECREDUCE_AND ? ISD::VECREDUCE_UMIN : ISD::VECREDUCE_UMAX; if (!TLI.isOperationLegalOrCustom(Opcode, VT) && TLI.isOperationLegalOrCustom(NewOpcode, VT) && DAG.ComputeNumSignBits(N0) == VT.getScalarSizeInBits()) return DAG.getNode(NewOpcode, SDLoc(N), N->getValueType(0), N0); } // vecreduce_or(insert_subvector(zero or undef, val)) -> vecreduce_or(val) // vecreduce_and(insert_subvector(ones or undef, val)) -> vecreduce_and(val) if (N0.getOpcode() == ISD::INSERT_SUBVECTOR && TLI.isTypeLegal(N0.getOperand(1).getValueType())) { SDValue Vec = N0.getOperand(0); SDValue Subvec = N0.getOperand(1); if ((Opcode == ISD::VECREDUCE_OR && (N0.getOperand(0).isUndef() || isNullOrNullSplat(Vec))) || (Opcode == ISD::VECREDUCE_AND && (N0.getOperand(0).isUndef() || isAllOnesOrAllOnesSplat(Vec)))) return DAG.getNode(Opcode, SDLoc(N), N->getValueType(0), Subvec); } return SDValue(); } SDValue DAGCombiner::visitVPOp(SDNode *N) { // VP operations in which all vector elements are disabled - either by // determining that the mask is all false or that the EVL is 0 - can be // eliminated. bool AreAllEltsDisabled = false; if (auto EVLIdx = ISD::getVPExplicitVectorLengthIdx(N->getOpcode())) AreAllEltsDisabled |= isNullConstant(N->getOperand(*EVLIdx)); if (auto MaskIdx = ISD::getVPMaskIdx(N->getOpcode())) AreAllEltsDisabled |= ISD::isConstantSplatVectorAllZeros(N->getOperand(*MaskIdx).getNode()); // This is the only generic VP combine we support for now. if (!AreAllEltsDisabled) return SDValue(); // Binary operations can be replaced by UNDEF. if (ISD::isVPBinaryOp(N->getOpcode())) return DAG.getUNDEF(N->getValueType(0)); // VP Memory operations can be replaced by either the chain (stores) or the // chain + undef (loads). if (const auto *MemSD = dyn_cast(N)) { if (MemSD->writeMem()) return MemSD->getChain(); return CombineTo(N, DAG.getUNDEF(N->getValueType(0)), MemSD->getChain()); } // Reduction operations return the start operand when no elements are active. if (ISD::isVPReduction(N->getOpcode())) return N->getOperand(0); return SDValue(); } /// Returns a vector_shuffle if it able to transform an AND to a vector_shuffle /// with the destination vector and a zero vector. /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==> /// vector_shuffle V, Zero, <0, 4, 2, 4> SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) { assert(N->getOpcode() == ISD::AND && "Unexpected opcode!"); EVT VT = N->getValueType(0); SDValue LHS = N->getOperand(0); SDValue RHS = peekThroughBitcasts(N->getOperand(1)); SDLoc DL(N); // Make sure we're not running after operation legalization where it // may have custom lowered the vector shuffles. if (LegalOperations) return SDValue(); if (RHS.getOpcode() != ISD::BUILD_VECTOR) return SDValue(); EVT RVT = RHS.getValueType(); unsigned NumElts = RHS.getNumOperands(); // Attempt to create a valid clear mask, splitting the mask into // sub elements and checking to see if each is // all zeros or all ones - suitable for shuffle masking. auto BuildClearMask = [&](int Split) { int NumSubElts = NumElts * Split; int NumSubBits = RVT.getScalarSizeInBits() / Split; SmallVector Indices; for (int i = 0; i != NumSubElts; ++i) { int EltIdx = i / Split; int SubIdx = i % Split; SDValue Elt = RHS.getOperand(EltIdx); // X & undef --> 0 (not undef). So this lane must be converted to choose // from the zero constant vector (same as if the element had all 0-bits). if (Elt.isUndef()) { Indices.push_back(i + NumSubElts); continue; } APInt Bits; if (isa(Elt)) Bits = cast(Elt)->getAPIntValue(); else if (isa(Elt)) Bits = cast(Elt)->getValueAPF().bitcastToAPInt(); else return SDValue(); // Extract the sub element from the constant bit mask. if (DAG.getDataLayout().isBigEndian()) Bits = Bits.extractBits(NumSubBits, (Split - SubIdx - 1) * NumSubBits); else Bits = Bits.extractBits(NumSubBits, SubIdx * NumSubBits); if (Bits.isAllOnes()) Indices.push_back(i); else if (Bits == 0) Indices.push_back(i + NumSubElts); else return SDValue(); } // Let's see if the target supports this vector_shuffle. EVT ClearSVT = EVT::getIntegerVT(*DAG.getContext(), NumSubBits); EVT ClearVT = EVT::getVectorVT(*DAG.getContext(), ClearSVT, NumSubElts); if (!TLI.isVectorClearMaskLegal(Indices, ClearVT)) return SDValue(); SDValue Zero = DAG.getConstant(0, DL, ClearVT); return DAG.getBitcast(VT, DAG.getVectorShuffle(ClearVT, DL, DAG.getBitcast(ClearVT, LHS), Zero, Indices)); }; // Determine maximum split level (byte level masking). int MaxSplit = 1; if (RVT.getScalarSizeInBits() % 8 == 0) MaxSplit = RVT.getScalarSizeInBits() / 8; for (int Split = 1; Split <= MaxSplit; ++Split) if (RVT.getScalarSizeInBits() % Split == 0) if (SDValue S = BuildClearMask(Split)) return S; return SDValue(); } /// If a vector binop is performed on splat values, it may be profitable to /// extract, scalarize, and insert/splat. static SDValue scalarizeBinOpOfSplats(SDNode *N, SelectionDAG &DAG, const SDLoc &DL) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); unsigned Opcode = N->getOpcode(); EVT VT = N->getValueType(0); EVT EltVT = VT.getVectorElementType(); const TargetLowering &TLI = DAG.getTargetLoweringInfo(); // TODO: Remove/replace the extract cost check? If the elements are available // as scalars, then there may be no extract cost. Should we ask if // inserting a scalar back into a vector is cheap instead? int Index0, Index1; SDValue Src0 = DAG.getSplatSourceVector(N0, Index0); SDValue Src1 = DAG.getSplatSourceVector(N1, Index1); // Extract element from splat_vector should be free. // TODO: use DAG.isSplatValue instead? bool IsBothSplatVector = N0.getOpcode() == ISD::SPLAT_VECTOR && N1.getOpcode() == ISD::SPLAT_VECTOR; if (!Src0 || !Src1 || Index0 != Index1 || Src0.getValueType().getVectorElementType() != EltVT || Src1.getValueType().getVectorElementType() != EltVT || !(IsBothSplatVector || TLI.isExtractVecEltCheap(VT, Index0)) || !TLI.isOperationLegalOrCustom(Opcode, EltVT)) return SDValue(); SDValue IndexC = DAG.getVectorIdxConstant(Index0, DL); SDValue X = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, Src0, IndexC); SDValue Y = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, Src1, IndexC); SDValue ScalarBO = DAG.getNode(Opcode, DL, EltVT, X, Y, N->getFlags()); // If all lanes but 1 are undefined, no need to splat the scalar result. // TODO: Keep track of undefs and use that info in the general case. if (N0.getOpcode() == ISD::BUILD_VECTOR && N0.getOpcode() == N1.getOpcode() && count_if(N0->ops(), [](SDValue V) { return !V.isUndef(); }) == 1 && count_if(N1->ops(), [](SDValue V) { return !V.isUndef(); }) == 1) { // bo (build_vec ..undef, X, undef...), (build_vec ..undef, Y, undef...) --> // build_vec ..undef, (bo X, Y), undef... SmallVector Ops(VT.getVectorNumElements(), DAG.getUNDEF(EltVT)); Ops[Index0] = ScalarBO; return DAG.getBuildVector(VT, DL, Ops); } // bo (splat X, Index), (splat Y, Index) --> splat (bo X, Y), Index if (VT.isScalableVector()) return DAG.getSplatVector(VT, DL, ScalarBO); SmallVector Ops(VT.getVectorNumElements(), ScalarBO); return DAG.getBuildVector(VT, DL, Ops); } /// Visit a binary vector operation, like ADD. SDValue DAGCombiner::SimplifyVBinOp(SDNode *N, const SDLoc &DL) { EVT VT = N->getValueType(0); assert(VT.isVector() && "SimplifyVBinOp only works on vectors!"); SDValue LHS = N->getOperand(0); SDValue RHS = N->getOperand(1); unsigned Opcode = N->getOpcode(); SDNodeFlags Flags = N->getFlags(); // Move unary shuffles with identical masks after a vector binop: // VBinOp (shuffle A, Undef, Mask), (shuffle B, Undef, Mask)) // --> shuffle (VBinOp A, B), Undef, Mask // This does not require type legality checks because we are creating the // same types of operations that are in the original sequence. We do have to // restrict ops like integer div that have immediate UB (eg, div-by-zero) // though. This code is adapted from the identical transform in instcombine. if (Opcode != ISD::UDIV && Opcode != ISD::SDIV && Opcode != ISD::UREM && Opcode != ISD::SREM && Opcode != ISD::UDIVREM && Opcode != ISD::SDIVREM) { auto *Shuf0 = dyn_cast(LHS); auto *Shuf1 = dyn_cast(RHS); if (Shuf0 && Shuf1 && Shuf0->getMask().equals(Shuf1->getMask()) && LHS.getOperand(1).isUndef() && RHS.getOperand(1).isUndef() && (LHS.hasOneUse() || RHS.hasOneUse() || LHS == RHS)) { SDValue NewBinOp = DAG.getNode(Opcode, DL, VT, LHS.getOperand(0), RHS.getOperand(0), Flags); SDValue UndefV = LHS.getOperand(1); return DAG.getVectorShuffle(VT, DL, NewBinOp, UndefV, Shuf0->getMask()); } // Try to sink a splat shuffle after a binop with a uniform constant. // This is limited to cases where neither the shuffle nor the constant have // undefined elements because that could be poison-unsafe or inhibit // demanded elements analysis. It is further limited to not change a splat // of an inserted scalar because that may be optimized better by // load-folding or other target-specific behaviors. if (isConstOrConstSplat(RHS) && Shuf0 && is_splat(Shuf0->getMask()) && Shuf0->hasOneUse() && Shuf0->getOperand(1).isUndef() && Shuf0->getOperand(0).getOpcode() != ISD::INSERT_VECTOR_ELT) { // binop (splat X), (splat C) --> splat (binop X, C) SDValue X = Shuf0->getOperand(0); SDValue NewBinOp = DAG.getNode(Opcode, DL, VT, X, RHS, Flags); return DAG.getVectorShuffle(VT, DL, NewBinOp, DAG.getUNDEF(VT), Shuf0->getMask()); } if (isConstOrConstSplat(LHS) && Shuf1 && is_splat(Shuf1->getMask()) && Shuf1->hasOneUse() && Shuf1->getOperand(1).isUndef() && Shuf1->getOperand(0).getOpcode() != ISD::INSERT_VECTOR_ELT) { // binop (splat C), (splat X) --> splat (binop C, X) SDValue X = Shuf1->getOperand(0); SDValue NewBinOp = DAG.getNode(Opcode, DL, VT, LHS, X, Flags); return DAG.getVectorShuffle(VT, DL, NewBinOp, DAG.getUNDEF(VT), Shuf1->getMask()); } } // The following pattern is likely to emerge with vector reduction ops. Moving // the binary operation ahead of insertion may allow using a narrower vector // instruction that has better performance than the wide version of the op: // VBinOp (ins undef, X, Z), (ins undef, Y, Z) --> ins VecC, (VBinOp X, Y), Z if (LHS.getOpcode() == ISD::INSERT_SUBVECTOR && LHS.getOperand(0).isUndef() && RHS.getOpcode() == ISD::INSERT_SUBVECTOR && RHS.getOperand(0).isUndef() && LHS.getOperand(2) == RHS.getOperand(2) && (LHS.hasOneUse() || RHS.hasOneUse())) { SDValue X = LHS.getOperand(1); SDValue Y = RHS.getOperand(1); SDValue Z = LHS.getOperand(2); EVT NarrowVT = X.getValueType(); if (NarrowVT == Y.getValueType() && TLI.isOperationLegalOrCustomOrPromote(Opcode, NarrowVT, LegalOperations)) { // (binop undef, undef) may not return undef, so compute that result. SDValue VecC = DAG.getNode(Opcode, DL, VT, DAG.getUNDEF(VT), DAG.getUNDEF(VT)); SDValue NarrowBO = DAG.getNode(Opcode, DL, NarrowVT, X, Y); return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, VecC, NarrowBO, Z); } } // Make sure all but the first op are undef or constant. auto ConcatWithConstantOrUndef = [](SDValue Concat) { return Concat.getOpcode() == ISD::CONCAT_VECTORS && all_of(drop_begin(Concat->ops()), [](const SDValue &Op) { return Op.isUndef() || ISD::isBuildVectorOfConstantSDNodes(Op.getNode()); }); }; // The following pattern is likely to emerge with vector reduction ops. Moving // the binary operation ahead of the concat may allow using a narrower vector // instruction that has better performance than the wide version of the op: // VBinOp (concat X, undef/constant), (concat Y, undef/constant) --> // concat (VBinOp X, Y), VecC if (ConcatWithConstantOrUndef(LHS) && ConcatWithConstantOrUndef(RHS) && (LHS.hasOneUse() || RHS.hasOneUse())) { EVT NarrowVT = LHS.getOperand(0).getValueType(); if (NarrowVT == RHS.getOperand(0).getValueType() && TLI.isOperationLegalOrCustomOrPromote(Opcode, NarrowVT)) { unsigned NumOperands = LHS.getNumOperands(); SmallVector ConcatOps; for (unsigned i = 0; i != NumOperands; ++i) { // This constant fold for operands 1 and up. ConcatOps.push_back(DAG.getNode(Opcode, DL, NarrowVT, LHS.getOperand(i), RHS.getOperand(i))); } return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, ConcatOps); } } if (SDValue V = scalarizeBinOpOfSplats(N, DAG, DL)) return V; return SDValue(); } SDValue DAGCombiner::SimplifySelect(const SDLoc &DL, SDValue N0, SDValue N1, SDValue N2) { assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!"); SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2, cast(N0.getOperand(2))->get()); // If we got a simplified select_cc node back from SimplifySelectCC, then // break it down into a new SETCC node, and a new SELECT node, and then return // the SELECT node, since we were called with a SELECT node. if (SCC.getNode()) { // Check to see if we got a select_cc back (to turn into setcc/select). // Otherwise, just return whatever node we got back, like fabs. if (SCC.getOpcode() == ISD::SELECT_CC) { const SDNodeFlags Flags = N0->getFlags(); SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0), N0.getValueType(), SCC.getOperand(0), SCC.getOperand(1), SCC.getOperand(4), Flags); AddToWorklist(SETCC.getNode()); SDValue SelectNode = DAG.getSelect(SDLoc(SCC), SCC.getValueType(), SETCC, SCC.getOperand(2), SCC.getOperand(3)); SelectNode->setFlags(Flags); return SelectNode; } return SCC; } return SDValue(); } /// Given a SELECT or a SELECT_CC node, where LHS and RHS are the two values /// being selected between, see if we can simplify the select. Callers of this /// should assume that TheSelect is deleted if this returns true. As such, they /// should return the appropriate thing (e.g. the node) back to the top-level of /// the DAG combiner loop to avoid it being looked at. bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS, SDValue RHS) { // fold (select (setcc x, [+-]0.0, *lt), NaN, (fsqrt x)) // The select + setcc is redundant, because fsqrt returns NaN for X < 0. if (const ConstantFPSDNode *NaN = isConstOrConstSplatFP(LHS)) { if (NaN->isNaN() && RHS.getOpcode() == ISD::FSQRT) { // We have: (select (setcc ?, ?, ?), NaN, (fsqrt ?)) SDValue Sqrt = RHS; ISD::CondCode CC; SDValue CmpLHS; const ConstantFPSDNode *Zero = nullptr; if (TheSelect->getOpcode() == ISD::SELECT_CC) { CC = cast(TheSelect->getOperand(4))->get(); CmpLHS = TheSelect->getOperand(0); Zero = isConstOrConstSplatFP(TheSelect->getOperand(1)); } else { // SELECT or VSELECT SDValue Cmp = TheSelect->getOperand(0); if (Cmp.getOpcode() == ISD::SETCC) { CC = cast(Cmp.getOperand(2))->get(); CmpLHS = Cmp.getOperand(0); Zero = isConstOrConstSplatFP(Cmp.getOperand(1)); } } if (Zero && Zero->isZero() && Sqrt.getOperand(0) == CmpLHS && (CC == ISD::SETOLT || CC == ISD::SETULT || CC == ISD::SETLT)) { // We have: (select (setcc x, [+-]0.0, *lt), NaN, (fsqrt x)) CombineTo(TheSelect, Sqrt); return true; } } } // Cannot simplify select with vector condition if (TheSelect->getOperand(0).getValueType().isVector()) return false; // If this is a select from two identical things, try to pull the operation // through the select. if (LHS.getOpcode() != RHS.getOpcode() || !LHS.hasOneUse() || !RHS.hasOneUse()) return false; // If this is a load and the token chain is identical, replace the select // of two loads with a load through a select of the address to load from. // This triggers in things like "select bool X, 10.0, 123.0" after the FP // constants have been dropped into the constant pool. if (LHS.getOpcode() == ISD::LOAD) { LoadSDNode *LLD = cast(LHS); LoadSDNode *RLD = cast(RHS); // Token chains must be identical. if (LHS.getOperand(0) != RHS.getOperand(0) || // Do not let this transformation reduce the number of volatile loads. // Be conservative for atomics for the moment // TODO: This does appear to be legal for unordered atomics (see D66309) !LLD->isSimple() || !RLD->isSimple() || // FIXME: If either is a pre/post inc/dec load, // we'd need to split out the address adjustment. LLD->isIndexed() || RLD->isIndexed() || // If this is an EXTLOAD, the VT's must match. LLD->getMemoryVT() != RLD->getMemoryVT() || // If this is an EXTLOAD, the kind of extension must match. (LLD->getExtensionType() != RLD->getExtensionType() && // The only exception is if one of the extensions is anyext. LLD->getExtensionType() != ISD::EXTLOAD && RLD->getExtensionType() != ISD::EXTLOAD) || // FIXME: this discards src value information. This is // over-conservative. It would be beneficial to be able to remember // both potential memory locations. Since we are discarding // src value info, don't do the transformation if the memory // locations are not in the default address space. LLD->getPointerInfo().getAddrSpace() != 0 || RLD->getPointerInfo().getAddrSpace() != 0 || // We can't produce a CMOV of a TargetFrameIndex since we won't // generate the address generation required. LLD->getBasePtr().getOpcode() == ISD::TargetFrameIndex || RLD->getBasePtr().getOpcode() == ISD::TargetFrameIndex || !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(), LLD->getBasePtr().getValueType())) return false; // The loads must not depend on one another. if (LLD->isPredecessorOf(RLD) || RLD->isPredecessorOf(LLD)) return false; // Check that the select condition doesn't reach either load. If so, // folding this will induce a cycle into the DAG. If not, this is safe to // xform, so create a select of the addresses. SmallPtrSet Visited; SmallVector Worklist; // Always fail if LLD and RLD are not independent. TheSelect is a // predecessor to all Nodes in question so we need not search past it. Visited.insert(TheSelect); Worklist.push_back(LLD); Worklist.push_back(RLD); if (SDNode::hasPredecessorHelper(LLD, Visited, Worklist) || SDNode::hasPredecessorHelper(RLD, Visited, Worklist)) return false; SDValue Addr; if (TheSelect->getOpcode() == ISD::SELECT) { // We cannot do this optimization if any pair of {RLD, LLD} is a // predecessor to {RLD, LLD, CondNode}. As we've already compared the // Loads, we only need to check if CondNode is a successor to one of the // loads. We can further avoid this if there's no use of their chain // value. SDNode *CondNode = TheSelect->getOperand(0).getNode(); Worklist.push_back(CondNode); if ((LLD->hasAnyUseOfValue(1) && SDNode::hasPredecessorHelper(LLD, Visited, Worklist)) || (RLD->hasAnyUseOfValue(1) && SDNode::hasPredecessorHelper(RLD, Visited, Worklist))) return false; Addr = DAG.getSelect(SDLoc(TheSelect), LLD->getBasePtr().getValueType(), TheSelect->getOperand(0), LLD->getBasePtr(), RLD->getBasePtr()); } else { // Otherwise SELECT_CC // We cannot do this optimization if any pair of {RLD, LLD} is a // predecessor to {RLD, LLD, CondLHS, CondRHS}. As we've already compared // the Loads, we only need to check if CondLHS/CondRHS is a successor to // one of the loads. We can further avoid this if there's no use of their // chain value. SDNode *CondLHS = TheSelect->getOperand(0).getNode(); SDNode *CondRHS = TheSelect->getOperand(1).getNode(); Worklist.push_back(CondLHS); Worklist.push_back(CondRHS); if ((LLD->hasAnyUseOfValue(1) && SDNode::hasPredecessorHelper(LLD, Visited, Worklist)) || (RLD->hasAnyUseOfValue(1) && SDNode::hasPredecessorHelper(RLD, Visited, Worklist))) return false; Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect), LLD->getBasePtr().getValueType(), TheSelect->getOperand(0), TheSelect->getOperand(1), LLD->getBasePtr(), RLD->getBasePtr(), TheSelect->getOperand(4)); } SDValue Load; // It is safe to replace the two loads if they have different alignments, // but the new load must be the minimum (most restrictive) alignment of the // inputs. Align Alignment = std::min(LLD->getAlign(), RLD->getAlign()); MachineMemOperand::Flags MMOFlags = LLD->getMemOperand()->getFlags(); if (!RLD->isInvariant()) MMOFlags &= ~MachineMemOperand::MOInvariant; if (!RLD->isDereferenceable()) MMOFlags &= ~MachineMemOperand::MODereferenceable; if (LLD->getExtensionType() == ISD::NON_EXTLOAD) { // FIXME: Discards pointer and AA info. Load = DAG.getLoad(TheSelect->getValueType(0), SDLoc(TheSelect), LLD->getChain(), Addr, MachinePointerInfo(), Alignment, MMOFlags); } else { // FIXME: Discards pointer and AA info. Load = DAG.getExtLoad( LLD->getExtensionType() == ISD::EXTLOAD ? RLD->getExtensionType() : LLD->getExtensionType(), SDLoc(TheSelect), TheSelect->getValueType(0), LLD->getChain(), Addr, MachinePointerInfo(), LLD->getMemoryVT(), Alignment, MMOFlags); } // Users of the select now use the result of the load. CombineTo(TheSelect, Load); // Users of the old loads now use the new load's chain. We know the // old-load value is dead now. CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1)); CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1)); return true; } return false; } /// Try to fold an expression of the form (N0 cond N1) ? N2 : N3 to a shift and /// bitwise 'and'. SDValue DAGCombiner::foldSelectCCToShiftAnd(const SDLoc &DL, SDValue N0, SDValue N1, SDValue N2, SDValue N3, ISD::CondCode CC) { // If this is a select where the false operand is zero and the compare is a // check of the sign bit, see if we can perform the "gzip trick": // select_cc setlt X, 0, A, 0 -> and (sra X, size(X)-1), A // select_cc setgt X, 0, A, 0 -> and (not (sra X, size(X)-1)), A EVT XType = N0.getValueType(); EVT AType = N2.getValueType(); if (!isNullConstant(N3) || !XType.bitsGE(AType)) return SDValue(); // If the comparison is testing for a positive value, we have to invert // the sign bit mask, so only do that transform if the target has a bitwise // 'and not' instruction (the invert is free). if (CC == ISD::SETGT && TLI.hasAndNot(N2)) { // (X > -1) ? A : 0 // (X > 0) ? X : 0 <-- This is canonical signed max. if (!(isAllOnesConstant(N1) || (isNullConstant(N1) && N0 == N2))) return SDValue(); } else if (CC == ISD::SETLT) { // (X < 0) ? A : 0 // (X < 1) ? X : 0 <-- This is un-canonicalized signed min. if (!(isNullConstant(N1) || (isOneConstant(N1) && N0 == N2))) return SDValue(); } else { return SDValue(); } // and (sra X, size(X)-1), A -> "and (srl X, C2), A" iff A is a single-bit // constant. EVT ShiftAmtTy = getShiftAmountTy(N0.getValueType()); auto *N2C = dyn_cast(N2.getNode()); if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue() - 1)) == 0)) { unsigned ShCt = XType.getSizeInBits() - N2C->getAPIntValue().logBase2() - 1; if (!TLI.shouldAvoidTransformToShift(XType, ShCt)) { SDValue ShiftAmt = DAG.getConstant(ShCt, DL, ShiftAmtTy); SDValue Shift = DAG.getNode(ISD::SRL, DL, XType, N0, ShiftAmt); AddToWorklist(Shift.getNode()); if (XType.bitsGT(AType)) { Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift); AddToWorklist(Shift.getNode()); } if (CC == ISD::SETGT) Shift = DAG.getNOT(DL, Shift, AType); return DAG.getNode(ISD::AND, DL, AType, Shift, N2); } } unsigned ShCt = XType.getSizeInBits() - 1; if (TLI.shouldAvoidTransformToShift(XType, ShCt)) return SDValue(); SDValue ShiftAmt = DAG.getConstant(ShCt, DL, ShiftAmtTy); SDValue Shift = DAG.getNode(ISD::SRA, DL, XType, N0, ShiftAmt); AddToWorklist(Shift.getNode()); if (XType.bitsGT(AType)) { Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift); AddToWorklist(Shift.getNode()); } if (CC == ISD::SETGT) Shift = DAG.getNOT(DL, Shift, AType); return DAG.getNode(ISD::AND, DL, AType, Shift, N2); } // Fold select(cc, binop(), binop()) -> binop(select(), select()) etc. SDValue DAGCombiner::foldSelectOfBinops(SDNode *N) { SDValue N0 = N->getOperand(0); SDValue N1 = N->getOperand(1); SDValue N2 = N->getOperand(2); EVT VT = N->getValueType(0); SDLoc DL(N); unsigned BinOpc = N1.getOpcode(); if (!TLI.isBinOp(BinOpc) || (N2.getOpcode() != BinOpc)) return SDValue(); // The use checks are intentionally on SDNode because we may be dealing // with opcodes that produce more than one SDValue. // TODO: Do we really need to check N0 (the condition operand of the select)? // But removing that clause could cause an infinite loop... if (!N0->hasOneUse() || !N1->hasOneUse() || !N2->hasOneUse()) return SDValue(); // Binops may include opcodes that return multiple values, so all values // must be created/propagated from the newly created binops below. SDVTList OpVTs = N1->getVTList(); // Fold select(cond, binop(x, y), binop(z, y)) // --> binop(select(cond, x, z), y) if (N1.getOperand(1) == N2.getOperand(1)) { SDValue NewSel = DAG.getSelect(DL, VT, N0, N1.getOperand(0), N2.getOperand(0)); SDValue NewBinOp = DAG.getNode(BinOpc, DL, OpVTs, NewSel, N1.getOperand(1)); NewBinOp->setFlags(N1->getFlags()); NewBinOp->intersectFlagsWith(N2->getFlags()); return NewBinOp; } // Fold select(cond, binop(x, y), binop(x, z)) // --> binop(x, select(cond, y, z)) // Second op VT might be different (e.g. shift amount type) if (N1.getOperand(0) == N2.getOperand(0) && VT == N1.getOperand(1).getValueType() && VT == N2.getOperand(1).getValueType()) { SDValue NewSel = DAG.getSelect(DL, VT, N0, N1.getOperand(1), N2.getOperand(1)); SDValue NewBinOp = DAG.getNode(BinOpc, DL, OpVTs, N1.getOperand(0), NewSel); NewBinOp->setFlags(N1->getFlags()); NewBinOp->intersectFlagsWith(N2->getFlags()); return NewBinOp; } // TODO: Handle isCommutativeBinOp patterns as well? return SDValue(); } // Transform (fneg/fabs (bitconvert x)) to avoid loading constant pool values. SDValue DAGCombiner::foldSignChangeInBitcast(SDNode *N) { SDValue N0 = N->getOperand(0); EVT VT = N->getValueType(0); bool IsFabs = N->getOpcode() == ISD::FABS; bool IsFree = IsFabs ? TLI.isFAbsFree(VT) : TLI.isFNegFree(VT); if (IsFree || N0.getOpcode() != ISD::BITCAST || !N0.hasOneUse()) return SDValue(); SDValue Int = N0.getOperand(0); EVT IntVT = Int.getValueType(); // The operand to cast should be integer. if (!IntVT.isInteger() || IntVT.isVector()) return SDValue(); // (fneg (bitconvert x)) -> (bitconvert (xor x sign)) // (fabs (bitconvert x)) -> (bitconvert (and x ~sign)) APInt SignMask; if (N0.getValueType().isVector()) { // For vector, create a sign mask (0x80...) or its inverse (for fabs, // 0x7f...) per element and splat it. SignMask = APInt::getSignMask(N0.getScalarValueSizeInBits()); if (IsFabs) SignMask = ~SignMask; SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask); } else { // For scalar, just use the sign mask (0x80... or the inverse, 0x7f...) SignMask = APInt::getSignMask(IntVT.getSizeInBits()); if (IsFabs) SignMask = ~SignMask; } SDLoc DL(N0); Int = DAG.getNode(IsFabs ? ISD::AND : ISD::XOR, DL, IntVT, Int, DAG.getConstant(SignMask, DL, IntVT)); AddToWorklist(Int.getNode()); return DAG.getBitcast(VT, Int); } /// Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)" /// where "tmp" is a constant pool entry containing an array with 1.0 and 2.0 /// in it. This may be a win when the constant is not otherwise available /// because it replaces two constant pool loads with one. SDValue DAGCombiner::convertSelectOfFPConstantsToLoadOffset( const SDLoc &DL, SDValue N0, SDValue N1, SDValue N2, SDValue N3, ISD::CondCode CC) { if (!TLI.reduceSelectOfFPConstantLoads(N0.getValueType())) return SDValue(); // If we are before legalize types, we want the other legalization to happen // first (for example, to avoid messing with soft float). auto *TV = dyn_cast(N2); auto *FV = dyn_cast(N3); EVT VT = N2.getValueType(); if (!TV || !FV || !TLI.isTypeLegal(VT)) return SDValue(); // If a constant can be materialized without loads, this does not make sense. if (TLI.getOperationAction(ISD::ConstantFP, VT) == TargetLowering::Legal || TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0), ForCodeSize) || TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0), ForCodeSize)) return SDValue(); // If both constants have multiple uses, then we won't need to do an extra // load. The values are likely around in registers for other users. if (!TV->hasOneUse() && !FV->hasOneUse()) return SDValue(); Constant *Elts[] = { const_cast(FV->getConstantFPValue()), const_cast(TV->getConstantFPValue()) }; Type *FPTy = Elts[0]->getType(); const DataLayout &TD = DAG.getDataLayout(); // Create a ConstantArray of the two constants. Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts); SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(DAG.getDataLayout()), TD.getPrefTypeAlign(FPTy)); Align Alignment = cast(CPIdx)->getAlign(); // Get offsets to the 0 and 1 elements of the array, so we can select between // them. SDValue Zero = DAG.getIntPtrConstant(0, DL); unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType()); SDValue One = DAG.getIntPtrConstant(EltSize, SDLoc(FV)); SDValue Cond = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()), N0, N1, CC); AddToWorklist(Cond.getNode()); SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(), Cond, One, Zero); AddToWorklist(CstOffset.getNode()); CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx, CstOffset); AddToWorklist(CPIdx.getNode()); return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx, MachinePointerInfo::getConstantPool( DAG.getMachineFunction()), Alignment); } /// Simplify an expression of the form (N0 cond N1) ? N2 : N3 /// where 'cond' is the comparison specified by CC. SDValue DAGCombiner::SimplifySelectCC(const SDLoc &DL, SDValue N0, SDValue N1, SDValue N2, SDValue N3, ISD::CondCode CC, bool NotExtCompare) { // (x ? y : y) -> y. if (N2 == N3) return N2; EVT CmpOpVT = N0.getValueType(); EVT CmpResVT = getSetCCResultType(CmpOpVT); EVT VT = N2.getValueType(); auto *N1C = dyn_cast(N1.getNode()); auto *N2C = dyn_cast(N2.getNode()); auto *N3C = dyn_cast(N3.getNode()); // Determine if the condition we're dealing with is constant. if (SDValue SCC = DAG.FoldSetCC(CmpResVT, N0, N1, CC, DL)) { AddToWorklist(SCC.getNode()); if (auto *SCCC = dyn_cast(SCC)) { // fold select_cc true, x, y -> x // fold select_cc false, x, y -> y return !(SCCC->isZero()) ? N2 : N3; } } if (SDValue V = convertSelectOfFPConstantsToLoadOffset(DL, N0, N1, N2, N3, CC)) return V; if (SDValue V = foldSelectCCToShiftAnd(DL, N0, N1, N2, N3, CC)) return V; // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A) // where y is has a single bit set. // A plaintext description would be, we can turn the SELECT_CC into an AND // when the condition can be materialized as an all-ones register. Any // single bit-test can be materialized as an all-ones register with // shift-left and shift-right-arith. if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND && N0->getValueType(0) == VT && isNullConstant(N1) && isNullConstant(N2)) { SDValue AndLHS = N0->getOperand(0); auto *ConstAndRHS = dyn_cast(N0->getOperand(1)); if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) { // Shift the tested bit over the sign bit. const APInt &AndMask = ConstAndRHS->getAPIntValue(); unsigned ShCt = AndMask.getBitWidth() - 1; if (!TLI.shouldAvoidTransformToShift(VT, ShCt)) { SDValue ShlAmt = DAG.getConstant(AndMask.countLeadingZeros(), SDLoc(AndLHS), getShiftAmountTy(AndLHS.getValueType())); SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt); // Now arithmetic right shift it all the way over, so the result is // either all-ones, or zero. SDValue ShrAmt = DAG.getConstant(ShCt, SDLoc(Shl), getShiftAmountTy(Shl.getValueType())); SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt); return DAG.getNode(ISD::AND, DL, VT, Shr, N3); } } } // fold select C, 16, 0 -> shl C, 4 bool Fold = N2C && isNullConstant(N3) && N2C->getAPIntValue().isPowerOf2(); bool Swap = N3C && isNullConstant(N2) && N3C->getAPIntValue().isPowerOf2(); if ((Fold || Swap) && TLI.getBooleanContents(CmpOpVT) == TargetLowering::ZeroOrOneBooleanContent && (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, CmpOpVT))) { if (Swap) { CC = ISD::getSetCCInverse(CC, CmpOpVT); std::swap(N2C, N3C); } // If the caller doesn't want us to simplify this into a zext of a compare, // don't do it. if (NotExtCompare && N2C->isOne()) return SDValue(); SDValue Temp, SCC; // zext (setcc n0, n1) if (LegalTypes) { SCC = DAG.getSetCC(DL, CmpResVT, N0, N1, CC); if (VT.bitsLT(SCC.getValueType())) Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2), VT); else Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2), VT, SCC); } else { SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC); Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2), VT, SCC); } AddToWorklist(SCC.getNode()); AddToWorklist(Temp.getNode()); if (N2C->isOne()) return Temp; unsigned ShCt = N2C->getAPIntValue().logBase2(); if (TLI.shouldAvoidTransformToShift(VT, ShCt)) return SDValue(); // shl setcc result by log2 n2c return DAG.getNode(ISD::SHL, DL, N2.getValueType(), Temp, DAG.getConstant(ShCt, SDLoc(Temp), getShiftAmountTy(Temp.getValueType()))); } // select_cc seteq X, 0, sizeof(X), ctlz(X) -> ctlz(X) // select_cc seteq X, 0, sizeof(X), ctlz_zero_undef(X) -> ctlz(X) // select_cc seteq X, 0, sizeof(X), cttz(X) -> cttz(X) // select_cc seteq X, 0, sizeof(X), cttz_zero_undef(X) -> cttz(X) // select_cc setne X, 0, ctlz(X), sizeof(X) -> ctlz(X) // select_cc setne X, 0, ctlz_zero_undef(X), sizeof(X) -> ctlz(X) // select_cc setne X, 0, cttz(X), sizeof(X) -> cttz(X) // select_cc setne X, 0, cttz_zero_undef(X), sizeof(X) -> cttz(X) if (N1C && N1C->isZero() && (CC == ISD::SETEQ || CC == ISD::SETNE)) { SDValue ValueOnZero = N2; SDValue Count = N3; // If the condition is NE instead of E, swap the operands. if (CC == ISD::SETNE) std::swap(ValueOnZero, Count); // Check if the value on zero is a constant equal to the bits in the type. if (auto *ValueOnZeroC = dyn_cast(ValueOnZero)) { if (ValueOnZeroC->getAPIntValue() == VT.getSizeInBits()) { // If the other operand is cttz/cttz_zero_undef of N0, and cttz is // legal, combine to just cttz. if ((Count.getOpcode() == ISD::CTTZ || Count.getOpcode() == ISD::CTTZ_ZERO_UNDEF) && N0 == Count.getOperand(0) && (!LegalOperations || TLI.isOperationLegal(ISD::CTTZ, VT))) return DAG.getNode(ISD::CTTZ, DL, VT, N0); // If the other operand is ctlz/ctlz_zero_undef of N0, and ctlz is // legal, combine to just ctlz. if ((Count.getOpcode() == ISD::CTLZ || Count.getOpcode() == ISD::CTLZ_ZERO_UNDEF) && N0 == Count.getOperand(0) && (!LegalOperations || TLI.isOperationLegal(ISD::CTLZ, VT))) return DAG.getNode(ISD::CTLZ, DL, VT, N0); } } } // Fold select_cc setgt X, -1, C, ~C -> xor (ashr X, BW-1), C // Fold select_cc setlt X, 0, C, ~C -> xor (ashr X, BW-1), ~C if (!NotExtCompare && N1C && N2C && N3C && N2C->getAPIntValue() == ~N3C->getAPIntValue() && ((N1C->isAllOnes() && CC == ISD::SETGT) || (N1C->isZero() && CC == ISD::SETLT)) && !TLI.shouldAvoidTransformToShift(VT, CmpOpVT.getScalarSizeInBits() - 1)) { SDValue ASR = DAG.getNode( ISD::SRA, DL, CmpOpVT, N0, DAG.getConstant(CmpOpVT.getScalarSizeInBits() - 1, DL, CmpOpVT)); return DAG.getNode(ISD::XOR, DL, VT, DAG.getSExtOrTrunc(ASR, DL, VT), DAG.getSExtOrTrunc(CC == ISD::SETLT ? N3 : N2, DL, VT)); } if (SDValue S = PerformMinMaxFpToSatCombine(N0, N1, N2, N3, CC, DAG)) return S; if (SDValue S = PerformUMinFpToSatCombine(N0, N1, N2, N3, CC, DAG)) return S; return SDValue(); } /// This is a stub for TargetLowering::SimplifySetCC. SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond, const SDLoc &DL, bool foldBooleans) { TargetLowering::DAGCombinerInfo DagCombineInfo(DAG, Level, false, this); return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL); } /// Given an ISD::SDIV node expressing a divide by constant, return /// a DAG expression to select that will generate the same value by multiplying /// by a magic number. /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide". SDValue DAGCombiner::BuildSDIV(SDNode *N) { // when optimising for minimum size, we don't want to expand a div to a mul // and a shift. if (DAG.getMachineFunction().getFunction().hasMinSize()) return SDValue(); SmallVector Built; if (SDValue S = TLI.BuildSDIV(N, DAG, LegalOperations, Built)) { for (SDNode *N : Built) AddToWorklist(N); return S; } return SDValue(); } /// Given an ISD::SDIV node expressing a divide by constant power of 2, return a /// DAG expression that will generate the same value by right shifting. SDValue DAGCombiner::BuildSDIVPow2(SDNode *N) { ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1)); if (!C) return SDValue(); // Avoid division by zero. if (C->isZero()) return SDValue(); SmallVector Built; if (SDValue S = TLI.BuildSDIVPow2(N, C->getAPIntValue(), DAG, Built)) { for (SDNode *N : Built) AddToWorklist(N); return S; } return SDValue(); } /// Given an ISD::UDIV node expressing a divide by constant, return a DAG /// expression that will generate the same value by multiplying by a magic /// number. /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide". SDValue DAGCombiner::BuildUDIV(SDNode *N) { // when optimising for minimum size, we don't want to expand a div to a mul // and a shift. if (DAG.getMachineFunction().getFunction().hasMinSize()) return SDValue(); SmallVector Built; if (SDValue S = TLI.BuildUDIV(N, DAG, LegalOperations, Built)) { for (SDNode *N : Built) AddToWorklist(N); return S; } return SDValue(); } /// Given an ISD::SREM node expressing a remainder by constant power of 2, /// return a DAG expression that will generate the same value. SDValue DAGCombiner::BuildSREMPow2(SDNode *N) { ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1)); if (!C) return SDValue(); // Avoid division by zero. if (C->isZero()) return SDValue(); SmallVector Built; if (SDValue S = TLI.BuildSREMPow2(N, C->getAPIntValue(), DAG, Built)) { for (SDNode *N : Built) AddToWorklist(N); return S; } return SDValue(); } /// Determines the LogBase2 value for a non-null input value using the /// transform: LogBase2(V) = (EltBits - 1) - ctlz(V). SDValue DAGCombiner::BuildLogBase2(SDValue V, const SDLoc &DL) { EVT VT = V.getValueType(); SDValue Ctlz = DAG.getNode(ISD::CTLZ, DL, VT, V); SDValue Base = DAG.getConstant(VT.getScalarSizeInBits() - 1, DL, VT); SDValue LogBase2 = DAG.getNode(ISD::SUB, DL, VT, Base, Ctlz); return LogBase2; } /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i) /// For the reciprocal, we need to find the zero of the function: /// F(X) = 1/X - A [which has a zero at X = 1/A] /// => /// X_{i+1} = X_i (2 - A X_i) = X_i + X_i (1 - A X_i) [this second form /// does not require additional intermediate precision] /// For the last iteration, put numerator N into it to gain more precision: /// Result = N X_i + X_i (N - N A X_i) SDValue DAGCombiner::BuildDivEstimate(SDValue N, SDValue Op, SDNodeFlags Flags) { if (LegalDAG) return SDValue(); // TODO: Handle extended types? EVT VT = Op.getValueType(); if (VT.getScalarType() != MVT::f16 && VT.getScalarType() != MVT::f32 && VT.getScalarType() != MVT::f64) return SDValue(); // If estimates are explicitly disabled for this function, we're done. MachineFunction &MF = DAG.getMachineFunction(); int Enabled = TLI.getRecipEstimateDivEnabled(VT, MF); if (Enabled == TLI.ReciprocalEstimate::Disabled) return SDValue(); // Estimates may be explicitly enabled for this type with a custom number of // refinement steps. int Iterations = TLI.getDivRefinementSteps(VT, MF); if (SDValue Est = TLI.getRecipEstimate(Op, DAG, Enabled, Iterations)) { AddToWorklist(Est.getNode()); SDLoc DL(Op); if (Iterations) { SDValue FPOne = DAG.getConstantFP(1.0, DL, VT); // Newton iterations: Est = Est + Est (N - Arg * Est) // If this is the last iteration, also multiply by the numerator. for (int i = 0; i < Iterations; ++i) { SDValue MulEst = Est; if (i == Iterations - 1) { MulEst = DAG.getNode(ISD::FMUL, DL, VT, N, Est, Flags); AddToWorklist(MulEst.getNode()); } SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Op, MulEst, Flags); AddToWorklist(NewEst.getNode()); NewEst = DAG.getNode(ISD::FSUB, DL, VT, (i == Iterations - 1 ? N : FPOne), NewEst, Flags); AddToWorklist(NewEst.getNode()); NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst, Flags); AddToWorklist(NewEst.getNode()); Est = DAG.getNode(ISD::FADD, DL, VT, MulEst, NewEst, Flags); AddToWorklist(Est.getNode()); } } else { // If no iterations are available, multiply with N. Est = DAG.getNode(ISD::FMUL, DL, VT, Est, N, Flags); AddToWorklist(Est.getNode()); } return Est; } return SDValue(); } /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i) /// For the reciprocal sqrt, we need to find the zero of the function: /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)] /// => /// X_{i+1} = X_i (1.5 - A X_i^2 / 2) /// As a result, we precompute A/2 prior to the iteration loop. SDValue DAGCombiner::buildSqrtNROneConst(SDValue Arg, SDValue Est, unsigned Iterations, SDNodeFlags Flags, bool Reciprocal) { EVT VT = Arg.getValueType(); SDLoc DL(Arg); SDValue ThreeHalves = DAG.getConstantFP(1.5, DL, VT); // We now need 0.5 * Arg which we can write as (1.5 * Arg - Arg) so that // this entire sequence requires only one FP constant. SDValue HalfArg = DAG.getNode(ISD::FMUL, DL, VT, ThreeHalves, Arg, Flags); HalfArg = DAG.getNode(ISD::FSUB, DL, VT, HalfArg, Arg, Flags); // Newton iterations: Est = Est * (1.5 - HalfArg * Est * Est) for (unsigned i = 0; i < Iterations; ++i) { SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, Est, Flags); NewEst = DAG.getNode(ISD::FMUL, DL, VT, HalfArg, NewEst, Flags); NewEst = DAG.getNode(ISD::FSUB, DL, VT, ThreeHalves, NewEst, Flags); Est = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst, Flags); } // If non-reciprocal square root is requested, multiply the result by Arg. if (!Reciprocal) Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Arg, Flags); return Est; } /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i) /// For the reciprocal sqrt, we need to find the zero of the function: /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)] /// => /// X_{i+1} = (-0.5 * X_i) * (A * X_i * X_i + (-3.0)) SDValue DAGCombiner::buildSqrtNRTwoConst(SDValue Arg, SDValue Est, unsigned Iterations, SDNodeFlags Flags, bool Reciprocal) { EVT VT = Arg.getValueType(); SDLoc DL(Arg); SDValue MinusThree = DAG.getConstantFP(-3.0, DL, VT); SDValue MinusHalf = DAG.getConstantFP(-0.5, DL, VT); // This routine must enter the loop below to work correctly // when (Reciprocal == false). assert(Iterations > 0); // Newton iterations for reciprocal square root: // E = (E * -0.5) * ((A * E) * E + -3.0) for (unsigned i = 0; i < Iterations; ++i) { SDValue AE = DAG.getNode(ISD::FMUL, DL, VT, Arg, Est, Flags); SDValue AEE = DAG.getNode(ISD::FMUL, DL, VT, AE, Est, Flags); SDValue RHS = DAG.getNode(ISD::FADD, DL, VT, AEE, MinusThree, Flags); // When calculating a square root at the last iteration build: // S = ((A * E) * -0.5) * ((A * E) * E + -3.0) // (notice a common subexpression) SDValue LHS; if (Reciprocal || (i + 1) < Iterations) { // RSQRT: LHS = (E * -0.5) LHS = DAG.getNode(ISD::FMUL, DL, VT, Est, MinusHalf, Flags); } else { // SQRT: LHS = (A * E) * -0.5 LHS = DAG.getNode(ISD::FMUL, DL, VT, AE, MinusHalf, Flags); } Est = DAG.getNode(ISD::FMUL, DL, VT, LHS, RHS, Flags); } return Est; } /// Build code to calculate either rsqrt(Op) or sqrt(Op). In the latter case /// Op*rsqrt(Op) is actually computed, so additional postprocessing is needed if /// Op can be zero. SDValue DAGCombiner::buildSqrtEstimateImpl(SDValue Op, SDNodeFlags Flags, bool Reciprocal) { if (LegalDAG) return SDValue(); // TODO: Handle extended types? EVT VT = Op.getValueType(); if (VT.getScalarType() != MVT::f16 && VT.getScalarType() != MVT::f32 && VT.getScalarType() != MVT::f64) return SDValue(); // If estimates are explicitly disabled for this function, we're done. MachineFunction &MF = DAG.getMachineFunction(); int Enabled = TLI.getRecipEstimateSqrtEnabled(VT, MF); if (Enabled == TLI.ReciprocalEstimate::Disabled) return SDValue(); // Estimates may be explicitly enabled for this type with a custom number of // refinement steps. int Iterations = TLI.getSqrtRefinementSteps(VT, MF); bool UseOneConstNR = false; if (SDValue Est = TLI.getSqrtEstimate(Op, DAG, Enabled, Iterations, UseOneConstNR, Reciprocal)) { AddToWorklist(Est.getNode()); if (Iterations) Est = UseOneConstNR ? buildSqrtNROneConst(Op, Est, Iterations, Flags, Reciprocal) : buildSqrtNRTwoConst(Op, Est, Iterations, Flags, Reciprocal); if (!Reciprocal) { SDLoc DL(Op); // Try the target specific test first. SDValue Test = TLI.getSqrtInputTest(Op, DAG, DAG.getDenormalMode(VT)); // The estimate is now completely wrong if the input was exactly 0.0 or // possibly a denormal. Force the answer to 0.0 or value provided by // target for those cases. Est = DAG.getNode( Test.getValueType().isVector() ? ISD::VSELECT : ISD::SELECT, DL, VT, Test, TLI.getSqrtResultForDenormInput(Op, DAG), Est); } return Est; } return SDValue(); } SDValue DAGCombiner::buildRsqrtEstimate(SDValue Op, SDNodeFlags Flags) { return buildSqrtEstimateImpl(Op, Flags, true); } SDValue DAGCombiner::buildSqrtEstimate(SDValue Op, SDNodeFlags Flags) { return buildSqrtEstimateImpl(Op, Flags, false); } /// Return true if there is any possibility that the two addresses overlap. bool DAGCombiner::mayAlias(SDNode *Op0, SDNode *Op1) const { struct MemUseCharacteristics { bool IsVolatile; bool IsAtomic; SDValue BasePtr; int64_t Offset; Optional NumBytes; MachineMemOperand *MMO; }; auto getCharacteristics = [](SDNode *N) -> MemUseCharacteristics { if (const auto *LSN = dyn_cast(N)) { int64_t Offset = 0; if (auto *C = dyn_cast(LSN->getOffset())) Offset = (LSN->getAddressingMode() == ISD::PRE_INC) ? C->getSExtValue() : (LSN->getAddressingMode() == ISD::PRE_DEC) ? -1 * C->getSExtValue() : 0; uint64_t Size = MemoryLocation::getSizeOrUnknown(LSN->getMemoryVT().getStoreSize()); return {LSN->isVolatile(), LSN->isAtomic(), LSN->getBasePtr(), Offset /*base offset*/, Optional(Size), LSN->getMemOperand()}; } if (const auto *LN = cast(N)) return {false /*isVolatile*/, /*isAtomic*/ false, LN->getOperand(1), (LN->hasOffset()) ? LN->getOffset() : 0, (LN->hasOffset()) ? Optional(LN->getSize()) : Optional(), (MachineMemOperand *)nullptr}; // Default. return {false /*isvolatile*/, /*isAtomic*/ false, SDValue(), (int64_t)0 /*offset*/, Optional() /*size*/, (MachineMemOperand *)nullptr}; }; MemUseCharacteristics MUC0 = getCharacteristics(Op0), MUC1 = getCharacteristics(Op1); // If they are to the same address, then they must be aliases. if (MUC0.BasePtr.getNode() && MUC0.BasePtr == MUC1.BasePtr && MUC0.Offset == MUC1.Offset) return true; // If they are both volatile then they cannot be reordered. if (MUC0.IsVolatile && MUC1.IsVolatile) return true; // Be conservative about atomics for the moment // TODO: This is way overconservative for unordered atomics (see D66309) if (MUC0.IsAtomic && MUC1.IsAtomic) return true; if (MUC0.MMO && MUC1.MMO) { if ((MUC0.MMO->isInvariant() && MUC1.MMO->isStore()) || (MUC1.MMO->isInvariant() && MUC0.MMO->isStore())) return false; } // Try to prove that there is aliasing, or that there is no aliasing. Either // way, we can return now. If nothing can be proved, proceed with more tests. bool IsAlias; if (BaseIndexOffset::computeAliasing(Op0, MUC0.NumBytes, Op1, MUC1.NumBytes, DAG, IsAlias)) return IsAlias; // The following all rely on MMO0 and MMO1 being valid. Fail conservatively if // either are not known. if (!MUC0.MMO || !MUC1.MMO) return true; // If one operation reads from invariant memory, and the other may store, they // cannot alias. These should really be checking the equivalent of mayWrite, // but it only matters for memory nodes other than load /store. if ((MUC0.MMO->isInvariant() && MUC1.MMO->isStore()) || (MUC1.MMO->isInvariant() && MUC0.MMO->isStore())) return false; // If we know required SrcValue1 and SrcValue2 have relatively large // alignment compared to the size and offset of the access, we may be able // to prove they do not alias. This check is conservative for now to catch // cases created by splitting vector types, it only works when the offsets are // multiples of the size of the data. int64_t SrcValOffset0 = MUC0.MMO->getOffset(); int64_t SrcValOffset1 = MUC1.MMO->getOffset(); Align OrigAlignment0 = MUC0.MMO->getBaseAlign(); Align OrigAlignment1 = MUC1.MMO->getBaseAlign(); auto &Size0 = MUC0.NumBytes; auto &Size1 = MUC1.NumBytes; if (OrigAlignment0 == OrigAlignment1 && SrcValOffset0 != SrcValOffset1 && Size0.has_value() && Size1.has_value() && *Size0 == *Size1 && OrigAlignment0 > *Size0 && SrcValOffset0 % *Size0 == 0 && SrcValOffset1 % *Size1 == 0) { int64_t OffAlign0 = SrcValOffset0 % OrigAlignment0.value(); int64_t OffAlign1 = SrcValOffset1 % OrigAlignment1.value(); // There is no overlap between these relatively aligned accesses of // similar size. Return no alias. if ((OffAlign0 + *Size0) <= OffAlign1 || (OffAlign1 + *Size1) <= OffAlign0) return false; } bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0 ? CombinerGlobalAA : DAG.getSubtarget().useAA(); #ifndef NDEBUG if (CombinerAAOnlyFunc.getNumOccurrences() && CombinerAAOnlyFunc != DAG.getMachineFunction().getName()) UseAA = false; #endif if (UseAA && AA && MUC0.MMO->getValue() && MUC1.MMO->getValue() && Size0 && Size1) { // Use alias analysis information. int64_t MinOffset = std::min(SrcValOffset0, SrcValOffset1); int64_t Overlap0 = *Size0 + SrcValOffset0 - MinOffset; int64_t Overlap1 = *Size1 + SrcValOffset1 - MinOffset; if (AA->isNoAlias( MemoryLocation(MUC0.MMO->getValue(), Overlap0, UseTBAA ? MUC0.MMO->getAAInfo() : AAMDNodes()), MemoryLocation(MUC1.MMO->getValue(), Overlap1, UseTBAA ? MUC1.MMO->getAAInfo() : AAMDNodes()))) return false; } // Otherwise we have to assume they alias. return true; } /// Walk up chain skipping non-aliasing memory nodes, /// looking for aliasing nodes and adding them to the Aliases vector. void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain, SmallVectorImpl &Aliases) { SmallVector Chains; // List of chains to visit. SmallPtrSet Visited; // Visited node set. // Get alias information for node. // TODO: relax aliasing for unordered atomics (see D66309) const bool IsLoad = isa(N) && cast(N)->isSimple(); // Starting off. Chains.push_back(OriginalChain); unsigned Depth = 0; // Attempt to improve chain by a single step auto ImproveChain = [&](SDValue &C) -> bool { switch (C.getOpcode()) { case ISD::EntryToken: // No need to mark EntryToken. C = SDValue(); return true; case ISD::LOAD: case ISD::STORE: { // Get alias information for C. // TODO: Relax aliasing for unordered atomics (see D66309) bool IsOpLoad = isa(C.getNode()) && cast(C.getNode())->isSimple(); if ((IsLoad && IsOpLoad) || !mayAlias(N, C.getNode())) { // Look further up the chain. C = C.getOperand(0); return true; } // Alias, so stop here. return false; } case ISD::CopyFromReg: // Always forward past past CopyFromReg. C = C.getOperand(0); return true; case ISD::LIFETIME_START: case ISD::LIFETIME_END: { // We can forward past any lifetime start/end that can be proven not to // alias the memory access. if (!mayAlias(N, C.getNode())) { // Look further up the chain. C = C.getOperand(0); return true; } return false; } default: return false; } }; // Look at each chain and determine if it is an alias. If so, add it to the // aliases list. If not, then continue up the chain looking for the next // candidate. while (!Chains.empty()) { SDValue Chain = Chains.pop_back_val(); // Don't bother if we've seen Chain before. if (!Visited.insert(Chain.getNode()).second) continue; // For TokenFactor nodes, look at each operand and only continue up the // chain until we reach the depth limit. // // FIXME: The depth check could be made to return the last non-aliasing // chain we found before we hit a tokenfactor rather than the original // chain. if (Depth > TLI.getGatherAllAliasesMaxDepth()) { Aliases.clear(); Aliases.push_back(OriginalChain); return; } if (Chain.getOpcode() == ISD::TokenFactor) { // We have to check each of the operands of the token factor for "small" // token factors, so we queue them up. Adding the operands to the queue // (stack) in reverse order maintains the original order and increases the // likelihood that getNode will find a matching token factor (CSE.) if (Chain.getNumOperands() > 16) { Aliases.push_back(Chain); continue; } for (unsigned n = Chain.getNumOperands(); n;) Chains.push_back(Chain.getOperand(--n)); ++Depth; continue; } // Everything else if (ImproveChain(Chain)) { // Updated Chain Found, Consider new chain if one exists. if (Chain.getNode()) Chains.push_back(Chain); ++Depth; continue; } // No Improved Chain Possible, treat as Alias. Aliases.push_back(Chain); } } /// Walk up chain skipping non-aliasing memory nodes, looking for a better chain /// (aliasing node.) SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) { if (OptLevel == CodeGenOpt::None) return OldChain; // Ops for replacing token factor. SmallVector Aliases; // Accumulate all the aliases to this node. GatherAllAliases(N, OldChain, Aliases); // If no operands then chain to entry token. if (Aliases.size() == 0) return DAG.getEntryNode(); // If a single operand then chain to it. We don't need to revisit it. if (Aliases.size() == 1) return Aliases[0]; // Construct a custom tailored token factor. return DAG.getTokenFactor(SDLoc(N), Aliases); } namespace { // TODO: Replace with with std::monostate when we move to C++17. struct UnitT { } Unit; bool operator==(const UnitT &, const UnitT &) { return true; } bool operator!=(const UnitT &, const UnitT &) { return false; } } // namespace // This function tries to collect a bunch of potentially interesting // nodes to improve the chains of, all at once. This might seem // redundant, as this function gets called when visiting every store // node, so why not let the work be done on each store as it's visited? // // I believe this is mainly important because mergeConsecutiveStores // is unable to deal with merging stores of different sizes, so unless // we improve the chains of all the potential candidates up-front // before running mergeConsecutiveStores, it might only see some of // the nodes that will eventually be candidates, and then not be able // to go from a partially-merged state to the desired final // fully-merged state. bool DAGCombiner::parallelizeChainedStores(StoreSDNode *St) { SmallVector ChainedStores; StoreSDNode *STChain = St; // Intervals records which offsets from BaseIndex have been covered. In // the common case, every store writes to the immediately previous address // space and thus merged with the previous interval at insertion time. using IMap = llvm::IntervalMap>; IMap::Allocator A; IMap Intervals(A); // This holds the base pointer, index, and the offset in bytes from the base // pointer. const BaseIndexOffset BasePtr = BaseIndexOffset::match(St, DAG); // We must have a base and an offset. if (!BasePtr.getBase().getNode()) return false; // Do not handle stores to undef base pointers. if (BasePtr.getBase().isUndef()) return false; // Do not handle stores to opaque types if (St->getMemoryVT().isZeroSized()) return false; // BaseIndexOffset assumes that offsets are fixed-size, which // is not valid for scalable vectors where the offsets are // scaled by `vscale`, so bail out early. if (St->getMemoryVT().isScalableVector()) return false; // Add ST's interval. Intervals.insert(0, (St->getMemoryVT().getSizeInBits() + 7) / 8, Unit); while (StoreSDNode *Chain = dyn_cast(STChain->getChain())) { if (Chain->getMemoryVT().isScalableVector()) return false; // If the chain has more than one use, then we can't reorder the mem ops. if (!SDValue(Chain, 0)->hasOneUse()) break; // TODO: Relax for unordered atomics (see D66309) if (!Chain->isSimple() || Chain->isIndexed()) break; // Find the base pointer and offset for this memory node. const BaseIndexOffset Ptr = BaseIndexOffset::match(Chain, DAG); // Check that the base pointer is the same as the original one. int64_t Offset; if (!BasePtr.equalBaseIndex(Ptr, DAG, Offset)) break; int64_t Length = (Chain->getMemoryVT().getSizeInBits() + 7) / 8; // Make sure we don't overlap with other intervals by checking the ones to // the left or right before inserting. auto I = Intervals.find(Offset); // If there's a next interval, we should end before it. if (I != Intervals.end() && I.start() < (Offset + Length)) break; // If there's a previous interval, we should start after it. if (I != Intervals.begin() && (--I).stop() <= Offset) break; Intervals.insert(Offset, Offset + Length, Unit); ChainedStores.push_back(Chain); STChain = Chain; } // If we didn't find a chained store, exit. if (ChainedStores.size() == 0) return false; // Improve all chained stores (St and ChainedStores members) starting from // where the store chain ended and return single TokenFactor. SDValue NewChain = STChain->getChain(); SmallVector TFOps; for (unsigned I = ChainedStores.size(); I;) { StoreSDNode *S = ChainedStores[--I]; SDValue BetterChain = FindBetterChain(S, NewChain); S = cast(DAG.UpdateNodeOperands( S, BetterChain, S->getOperand(1), S->getOperand(2), S->getOperand(3))); TFOps.push_back(SDValue(S, 0)); ChainedStores[I] = S; } // Improve St's chain. Use a new node to avoid creating a loop from CombineTo. SDValue BetterChain = FindBetterChain(St, NewChain); SDValue NewST; if (St->isTruncatingStore()) NewST = DAG.getTruncStore(BetterChain, SDLoc(St), St->getValue(), St->getBasePtr(), St->getMemoryVT(), St->getMemOperand()); else NewST = DAG.getStore(BetterChain, SDLoc(St), St->getValue(), St->getBasePtr(), St->getMemOperand()); TFOps.push_back(NewST); // If we improved every element of TFOps, then we've lost the dependence on // NewChain to successors of St and we need to add it back to TFOps. Do so at // the beginning to keep relative order consistent with FindBetterChains. auto hasImprovedChain = [&](SDValue ST) -> bool { return ST->getOperand(0) != NewChain; }; bool AddNewChain = llvm::all_of(TFOps, hasImprovedChain); if (AddNewChain) TFOps.insert(TFOps.begin(), NewChain); SDValue TF = DAG.getTokenFactor(SDLoc(STChain), TFOps); CombineTo(St, TF); // Add TF and its operands to the worklist. AddToWorklist(TF.getNode()); for (const SDValue &Op : TF->ops()) AddToWorklist(Op.getNode()); AddToWorklist(STChain); return true; } bool DAGCombiner::findBetterNeighborChains(StoreSDNode *St) { if (OptLevel == CodeGenOpt::None) return false; const BaseIndexOffset BasePtr = BaseIndexOffset::match(St, DAG); // We must have a base and an offset. if (!BasePtr.getBase().getNode()) return false; // Do not handle stores to undef base pointers. if (BasePtr.getBase().isUndef()) return false; // Directly improve a chain of disjoint stores starting at St. if (parallelizeChainedStores(St)) return true; // Improve St's Chain.. SDValue BetterChain = FindBetterChain(St, St->getChain()); if (St->getChain() != BetterChain) { replaceStoreChain(St, BetterChain); return true; } return false; } /// This is the entry point for the file. void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis *AA, CodeGenOpt::Level OptLevel) { /// This is the main entry point to this class. DAGCombiner(*this, AA, OptLevel).Run(Level); }