//===- Reassociate.cpp - Reassociate binary expressions -------------------===// // // 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 reassociates commutative expressions in an order that is designed // to promote better constant propagation, GCSE, LICM, PRE, etc. // // For example: 4 + (x + 5) -> x + (4 + 5) // // In the implementation of this algorithm, constants are assigned rank = 0, // function arguments are rank = 1, and other values are assigned ranks // corresponding to the reverse post order traversal of current function // (starting at 2), which effectively gives values in deep loops higher rank // than values not in loops. // //===----------------------------------------------------------------------===// #include "llvm/Transforms/Scalar/Reassociate.h" #include "llvm/ADT/APFloat.h" #include "llvm/ADT/APInt.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/PostOrderIterator.h" #include "llvm/ADT/SetVector.h" #include "llvm/ADT/SmallPtrSet.h" #include "llvm/ADT/SmallSet.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/Statistic.h" #include "llvm/Analysis/GlobalsModRef.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/Analysis/ValueTracking.h" #include "llvm/IR/Argument.h" #include "llvm/IR/BasicBlock.h" #include "llvm/IR/CFG.h" #include "llvm/IR/Constant.h" #include "llvm/IR/Constants.h" #include "llvm/IR/Function.h" #include "llvm/IR/IRBuilder.h" #include "llvm/IR/InstrTypes.h" #include "llvm/IR/Instruction.h" #include "llvm/IR/Instructions.h" #include "llvm/IR/IntrinsicInst.h" #include "llvm/IR/Operator.h" #include "llvm/IR/PassManager.h" #include "llvm/IR/PatternMatch.h" #include "llvm/IR/Type.h" #include "llvm/IR/User.h" #include "llvm/IR/Value.h" #include "llvm/IR/ValueHandle.h" #include "llvm/Pass.h" #include "llvm/Support/Casting.h" #include "llvm/Support/Debug.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/raw_ostream.h" #include "llvm/Transforms/Scalar.h" #include #include #include using namespace llvm; using namespace reassociate; using namespace PatternMatch; #define DEBUG_TYPE "reassociate" STATISTIC(NumChanged, "Number of insts reassociated"); STATISTIC(NumAnnihil, "Number of expr tree annihilated"); STATISTIC(NumFactor , "Number of multiplies factored"); #ifndef NDEBUG /// Print out the expression identified in the Ops list. static void PrintOps(Instruction *I, const SmallVectorImpl &Ops) { Module *M = I->getModule(); dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " " << *Ops[0].Op->getType() << '\t'; for (unsigned i = 0, e = Ops.size(); i != e; ++i) { dbgs() << "[ "; Ops[i].Op->printAsOperand(dbgs(), false, M); dbgs() << ", #" << Ops[i].Rank << "] "; } } #endif /// Utility class representing a non-constant Xor-operand. We classify /// non-constant Xor-Operands into two categories: /// C1) The operand is in the form "X & C", where C is a constant and C != ~0 /// C2) /// C2.1) The operand is in the form of "X | C", where C is a non-zero /// constant. /// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this /// operand as "E | 0" class llvm::reassociate::XorOpnd { public: XorOpnd(Value *V); bool isInvalid() const { return SymbolicPart == nullptr; } bool isOrExpr() const { return isOr; } Value *getValue() const { return OrigVal; } Value *getSymbolicPart() const { return SymbolicPart; } unsigned getSymbolicRank() const { return SymbolicRank; } const APInt &getConstPart() const { return ConstPart; } void Invalidate() { SymbolicPart = OrigVal = nullptr; } void setSymbolicRank(unsigned R) { SymbolicRank = R; } private: Value *OrigVal; Value *SymbolicPart; APInt ConstPart; unsigned SymbolicRank; bool isOr; }; XorOpnd::XorOpnd(Value *V) { assert(!isa(V) && "No ConstantInt"); OrigVal = V; Instruction *I = dyn_cast(V); SymbolicRank = 0; if (I && (I->getOpcode() == Instruction::Or || I->getOpcode() == Instruction::And)) { Value *V0 = I->getOperand(0); Value *V1 = I->getOperand(1); const APInt *C; if (match(V0, m_APInt(C))) std::swap(V0, V1); if (match(V1, m_APInt(C))) { ConstPart = *C; SymbolicPart = V0; isOr = (I->getOpcode() == Instruction::Or); return; } } // view the operand as "V | 0" SymbolicPart = V; ConstPart = APInt::getNullValue(V->getType()->getScalarSizeInBits()); isOr = true; } /// Return true if V is an instruction of the specified opcode and if it /// only has one use. static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { auto *I = dyn_cast(V); if (I && I->hasOneUse() && I->getOpcode() == Opcode) if (!isa(I) || I->isFast()) return cast(I); return nullptr; } static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1, unsigned Opcode2) { auto *I = dyn_cast(V); if (I && I->hasOneUse() && (I->getOpcode() == Opcode1 || I->getOpcode() == Opcode2)) if (!isa(I) || I->isFast()) return cast(I); return nullptr; } void ReassociatePass::BuildRankMap(Function &F, ReversePostOrderTraversal &RPOT) { unsigned Rank = 2; // Assign distinct ranks to function arguments. for (auto &Arg : F.args()) { ValueRankMap[&Arg] = ++Rank; LLVM_DEBUG(dbgs() << "Calculated Rank[" << Arg.getName() << "] = " << Rank << "\n"); } // Traverse basic blocks in ReversePostOrder for (BasicBlock *BB : RPOT) { unsigned BBRank = RankMap[BB] = ++Rank << 16; // Walk the basic block, adding precomputed ranks for any instructions that // we cannot move. This ensures that the ranks for these instructions are // all different in the block. for (Instruction &I : *BB) if (mayBeMemoryDependent(I)) ValueRankMap[&I] = ++BBRank; } } unsigned ReassociatePass::getRank(Value *V) { Instruction *I = dyn_cast(V); if (!I) { if (isa(V)) return ValueRankMap[V]; // Function argument. return 0; // Otherwise it's a global or constant, rank 0. } if (unsigned Rank = ValueRankMap[I]) return Rank; // Rank already known? // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that // we can reassociate expressions for code motion! Since we do not recurse // for PHI nodes, we cannot have infinite recursion here, because there // cannot be loops in the value graph that do not go through PHI nodes. unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; for (unsigned i = 0, e = I->getNumOperands(); i != e && Rank != MaxRank; ++i) Rank = std::max(Rank, getRank(I->getOperand(i))); // If this is a 'not' or 'neg' instruction, do not count it for rank. This // assures us that X and ~X will have the same rank. if (!match(I, m_Not(m_Value())) && !match(I, m_Neg(m_Value())) && !match(I, m_FNeg(m_Value()))) ++Rank; LLVM_DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n"); return ValueRankMap[I] = Rank; } // Canonicalize constants to RHS. Otherwise, sort the operands by rank. void ReassociatePass::canonicalizeOperands(Instruction *I) { assert(isa(I) && "Expected binary operator."); assert(I->isCommutative() && "Expected commutative operator."); Value *LHS = I->getOperand(0); Value *RHS = I->getOperand(1); if (LHS == RHS || isa(RHS)) return; if (isa(LHS) || getRank(RHS) < getRank(LHS)) cast(I)->swapOperands(); } static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name, Instruction *InsertBefore, Value *FlagsOp) { if (S1->getType()->isIntOrIntVectorTy()) return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore); else { BinaryOperator *Res = BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore); Res->setFastMathFlags(cast(FlagsOp)->getFastMathFlags()); return Res; } } static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name, Instruction *InsertBefore, Value *FlagsOp) { if (S1->getType()->isIntOrIntVectorTy()) return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore); else { BinaryOperator *Res = BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore); Res->setFastMathFlags(cast(FlagsOp)->getFastMathFlags()); return Res; } } static BinaryOperator *CreateNeg(Value *S1, const Twine &Name, Instruction *InsertBefore, Value *FlagsOp) { if (S1->getType()->isIntOrIntVectorTy()) return BinaryOperator::CreateNeg(S1, Name, InsertBefore); else { BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore); Res->setFastMathFlags(cast(FlagsOp)->getFastMathFlags()); return Res; } } /// Replace 0-X with X*-1. static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) { assert((isa(Neg) || isa(Neg)) && "Expected a Negate!"); // FIXME: It's not safe to lower a unary FNeg into a FMul by -1.0. unsigned OpNo = isa(Neg) ? 1 : 0; Type *Ty = Neg->getType(); Constant *NegOne = Ty->isIntOrIntVectorTy() ? ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, -1.0); BinaryOperator *Res = CreateMul(Neg->getOperand(OpNo), NegOne, "", Neg, Neg); Neg->setOperand(OpNo, Constant::getNullValue(Ty)); // Drop use of op. Res->takeName(Neg); Neg->replaceAllUsesWith(Res); Res->setDebugLoc(Neg->getDebugLoc()); return Res; } /// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael /// function. This means that x^(2^k) === 1 mod 2^Bitwidth for /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic. /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every /// even x in Bitwidth-bit arithmetic. static unsigned CarmichaelShift(unsigned Bitwidth) { if (Bitwidth < 3) return Bitwidth - 1; return Bitwidth - 2; } /// Add the extra weight 'RHS' to the existing weight 'LHS', /// reducing the combined weight using any special properties of the operation. /// The existing weight LHS represents the computation X op X op ... op X where /// X occurs LHS times. The combined weight represents X op X op ... op X with /// X occurring LHS + RHS times. If op is "Xor" for example then the combined /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even; /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second. static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) { // If we were working with infinite precision arithmetic then the combined // weight would be LHS + RHS. But we are using finite precision arithmetic, // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct // for nilpotent operations and addition, but not for idempotent operations // and multiplication), so it is important to correctly reduce the combined // weight back into range if wrapping would be wrong. // If RHS is zero then the weight didn't change. if (RHS.isMinValue()) return; // If LHS is zero then the combined weight is RHS. if (LHS.isMinValue()) { LHS = RHS; return; } // From this point on we know that neither LHS nor RHS is zero. if (Instruction::isIdempotent(Opcode)) { // Idempotent means X op X === X, so any non-zero weight is equivalent to a // weight of 1. Keeping weights at zero or one also means that wrapping is // not a problem. assert(LHS == 1 && RHS == 1 && "Weights not reduced!"); return; // Return a weight of 1. } if (Instruction::isNilpotent(Opcode)) { // Nilpotent means X op X === 0, so reduce weights modulo 2. assert(LHS == 1 && RHS == 1 && "Weights not reduced!"); LHS = 0; // 1 + 1 === 0 modulo 2. return; } if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) { // TODO: Reduce the weight by exploiting nsw/nuw? LHS += RHS; return; } assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) && "Unknown associative operation!"); unsigned Bitwidth = LHS.getBitWidth(); // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth // bit number x, since either x is odd in which case x^CM = 1, or x is even in // which case both x^W and x^(W - CM) are zero. By subtracting off multiples // of CM like this weights can always be reduced to the range [0, CM+Bitwidth) // which by a happy accident means that they can always be represented using // Bitwidth bits. // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than // the Carmichael number). if (Bitwidth > 3) { /// CM - The value of Carmichael's lambda function. APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth)); // Any weight W >= Threshold can be replaced with W - CM. APInt Threshold = CM + Bitwidth; assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!"); // For Bitwidth 4 or more the following sum does not overflow. LHS += RHS; while (LHS.uge(Threshold)) LHS -= CM; } else { // To avoid problems with overflow do everything the same as above but using // a larger type. unsigned CM = 1U << CarmichaelShift(Bitwidth); unsigned Threshold = CM + Bitwidth; assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold && "Weights not reduced!"); unsigned Total = LHS.getZExtValue() + RHS.getZExtValue(); while (Total >= Threshold) Total -= CM; LHS = Total; } } using RepeatedValue = std::pair; /// Given an associative binary expression, return the leaf /// nodes in Ops along with their weights (how many times the leaf occurs). The /// original expression is the same as /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times /// op /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times /// op /// ... /// op /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times /// /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct. /// /// This routine may modify the function, in which case it returns 'true'. The /// changes it makes may well be destructive, changing the value computed by 'I' /// to something completely different. Thus if the routine returns 'true' then /// you MUST either replace I with a new expression computed from the Ops array, /// or use RewriteExprTree to put the values back in. /// /// A leaf node is either not a binary operation of the same kind as the root /// node 'I' (i.e. is not a binary operator at all, or is, but with a different /// opcode), or is the same kind of binary operator but has a use which either /// does not belong to the expression, or does belong to the expression but is /// a leaf node. Every leaf node has at least one use that is a non-leaf node /// of the expression, while for non-leaf nodes (except for the root 'I') every /// use is a non-leaf node of the expression. /// /// For example: /// expression graph node names /// /// + | I /// / \ | /// + + | A, B /// / \ / \ | /// * + * | C, D, E /// / \ / \ / \ | /// + * | F, G /// /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in /// that order) (C, 1), (E, 1), (F, 2), (G, 2). /// /// The expression is maximal: if some instruction is a binary operator of the /// same kind as 'I', and all of its uses are non-leaf nodes of the expression, /// then the instruction also belongs to the expression, is not a leaf node of /// it, and its operands also belong to the expression (but may be leaf nodes). /// /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in /// order to ensure that every non-root node in the expression has *exactly one* /// use by a non-leaf node of the expression. This destruction means that the /// caller MUST either replace 'I' with a new expression or use something like /// RewriteExprTree to put the values back in if the routine indicates that it /// made a change by returning 'true'. /// /// In the above example either the right operand of A or the left operand of B /// will be replaced by undef. If it is B's operand then this gives: /// /// + | I /// / \ | /// + + | A, B - operand of B replaced with undef /// / \ \ | /// * + * | C, D, E /// / \ / \ / \ | /// + * | F, G /// /// Note that such undef operands can only be reached by passing through 'I'. /// For example, if you visit operands recursively starting from a leaf node /// then you will never see such an undef operand unless you get back to 'I', /// which requires passing through a phi node. /// /// Note that this routine may also mutate binary operators of the wrong type /// that have all uses inside the expression (i.e. only used by non-leaf nodes /// of the expression) if it can turn them into binary operators of the right /// type and thus make the expression bigger. static bool LinearizeExprTree(Instruction *I, SmallVectorImpl &Ops) { assert((isa(I) || isa(I)) && "Expected a UnaryOperator or BinaryOperator!"); LLVM_DEBUG(dbgs() << "LINEARIZE: " << *I << '\n'); unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits(); unsigned Opcode = I->getOpcode(); assert(I->isAssociative() && I->isCommutative() && "Expected an associative and commutative operation!"); // Visit all operands of the expression, keeping track of their weight (the // number of paths from the expression root to the operand, or if you like // the number of times that operand occurs in the linearized expression). // For example, if I = X + A, where X = A + B, then I, X and B have weight 1 // while A has weight two. // Worklist of non-leaf nodes (their operands are in the expression too) along // with their weights, representing a certain number of paths to the operator. // If an operator occurs in the worklist multiple times then we found multiple // ways to get to it. SmallVector, 8> Worklist; // (Op, Weight) Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1))); bool Changed = false; // Leaves of the expression are values that either aren't the right kind of // operation (eg: a constant, or a multiply in an add tree), or are, but have // some uses that are not inside the expression. For example, in I = X + X, // X = A + B, the value X has two uses (by I) that are in the expression. If // X has any other uses, for example in a return instruction, then we consider // X to be a leaf, and won't analyze it further. When we first visit a value, // if it has more than one use then at first we conservatively consider it to // be a leaf. Later, as the expression is explored, we may discover some more // uses of the value from inside the expression. If all uses turn out to be // from within the expression (and the value is a binary operator of the right // kind) then the value is no longer considered to be a leaf, and its operands // are explored. // Leaves - Keeps track of the set of putative leaves as well as the number of // paths to each leaf seen so far. using LeafMap = DenseMap; LeafMap Leaves; // Leaf -> Total weight so far. SmallVector LeafOrder; // Ensure deterministic leaf output order. #ifndef NDEBUG SmallPtrSet Visited; // For sanity checking the iteration scheme. #endif while (!Worklist.empty()) { std::pair P = Worklist.pop_back_val(); I = P.first; // We examine the operands of this binary operator. for (unsigned OpIdx = 0; OpIdx < I->getNumOperands(); ++OpIdx) { // Visit operands. Value *Op = I->getOperand(OpIdx); APInt Weight = P.second; // Number of paths to this operand. LLVM_DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n"); assert(!Op->use_empty() && "No uses, so how did we get to it?!"); // If this is a binary operation of the right kind with only one use then // add its operands to the expression. if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) { assert(Visited.insert(Op).second && "Not first visit!"); LLVM_DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n"); Worklist.push_back(std::make_pair(BO, Weight)); continue; } // Appears to be a leaf. Is the operand already in the set of leaves? LeafMap::iterator It = Leaves.find(Op); if (It == Leaves.end()) { // Not in the leaf map. Must be the first time we saw this operand. assert(Visited.insert(Op).second && "Not first visit!"); if (!Op->hasOneUse()) { // This value has uses not accounted for by the expression, so it is // not safe to modify. Mark it as being a leaf. LLVM_DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n"); LeafOrder.push_back(Op); Leaves[Op] = Weight; continue; } // No uses outside the expression, try morphing it. } else { // Already in the leaf map. assert(It != Leaves.end() && Visited.count(Op) && "In leaf map but not visited!"); // Update the number of paths to the leaf. IncorporateWeight(It->second, Weight, Opcode); #if 0 // TODO: Re-enable once PR13021 is fixed. // The leaf already has one use from inside the expression. As we want // exactly one such use, drop this new use of the leaf. assert(!Op->hasOneUse() && "Only one use, but we got here twice!"); I->setOperand(OpIdx, UndefValue::get(I->getType())); Changed = true; // If the leaf is a binary operation of the right kind and we now see // that its multiple original uses were in fact all by nodes belonging // to the expression, then no longer consider it to be a leaf and add // its operands to the expression. if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) { LLVM_DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n"); Worklist.push_back(std::make_pair(BO, It->second)); Leaves.erase(It); continue; } #endif // If we still have uses that are not accounted for by the expression // then it is not safe to modify the value. if (!Op->hasOneUse()) continue; // No uses outside the expression, try morphing it. Weight = It->second; Leaves.erase(It); // Since the value may be morphed below. } // At this point we have a value which, first of all, is not a binary // expression of the right kind, and secondly, is only used inside the // expression. This means that it can safely be modified. See if we // can usefully morph it into an expression of the right kind. assert((!isa(Op) || cast(Op)->getOpcode() != Opcode || (isa(Op) && !cast(Op)->isFast())) && "Should have been handled above!"); assert(Op->hasOneUse() && "Has uses outside the expression tree!"); // If this is a multiply expression, turn any internal negations into // multiplies by -1 so they can be reassociated. if (Instruction *Tmp = dyn_cast(Op)) if ((Opcode == Instruction::Mul && match(Tmp, m_Neg(m_Value()))) || (Opcode == Instruction::FMul && match(Tmp, m_FNeg(m_Value())))) { LLVM_DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO "); Tmp = LowerNegateToMultiply(Tmp); LLVM_DEBUG(dbgs() << *Tmp << '\n'); Worklist.push_back(std::make_pair(Tmp, Weight)); Changed = true; continue; } // Failed to morph into an expression of the right type. This really is // a leaf. LLVM_DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n"); assert(!isReassociableOp(Op, Opcode) && "Value was morphed?"); LeafOrder.push_back(Op); Leaves[Op] = Weight; } } // The leaves, repeated according to their weights, represent the linearized // form of the expression. for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) { Value *V = LeafOrder[i]; LeafMap::iterator It = Leaves.find(V); if (It == Leaves.end()) // Node initially thought to be a leaf wasn't. continue; assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!"); APInt Weight = It->second; if (Weight.isMinValue()) // Leaf already output or weight reduction eliminated it. continue; // Ensure the leaf is only output once. It->second = 0; Ops.push_back(std::make_pair(V, Weight)); } // For nilpotent operations or addition there may be no operands, for example // because the expression was "X xor X" or consisted of 2^Bitwidth additions: // in both cases the weight reduces to 0 causing the value to be skipped. if (Ops.empty()) { Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType()); assert(Identity && "Associative operation without identity!"); Ops.emplace_back(Identity, APInt(Bitwidth, 1)); } return Changed; } /// Now that the operands for this expression tree are /// linearized and optimized, emit them in-order. void ReassociatePass::RewriteExprTree(BinaryOperator *I, SmallVectorImpl &Ops) { assert(Ops.size() > 1 && "Single values should be used directly!"); // Since our optimizations should never increase the number of operations, the // new expression can usually be written reusing the existing binary operators // from the original expression tree, without creating any new instructions, // though the rewritten expression may have a completely different topology. // We take care to not change anything if the new expression will be the same // as the original. If more than trivial changes (like commuting operands) // were made then we are obliged to clear out any optional subclass data like // nsw flags. /// NodesToRewrite - Nodes from the original expression available for writing /// the new expression into. SmallVector NodesToRewrite; unsigned Opcode = I->getOpcode(); BinaryOperator *Op = I; /// NotRewritable - The operands being written will be the leaves of the new /// expression and must not be used as inner nodes (via NodesToRewrite) by /// mistake. Inner nodes are always reassociable, and usually leaves are not /// (if they were they would have been incorporated into the expression and so /// would not be leaves), so most of the time there is no danger of this. But /// in rare cases a leaf may become reassociable if an optimization kills uses /// of it, or it may momentarily become reassociable during rewriting (below) /// due it being removed as an operand of one of its uses. Ensure that misuse /// of leaf nodes as inner nodes cannot occur by remembering all of the future /// leaves and refusing to reuse any of them as inner nodes. SmallPtrSet NotRewritable; for (unsigned i = 0, e = Ops.size(); i != e; ++i) NotRewritable.insert(Ops[i].Op); // ExpressionChanged - Non-null if the rewritten expression differs from the // original in some non-trivial way, requiring the clearing of optional flags. // Flags are cleared from the operator in ExpressionChanged up to I inclusive. BinaryOperator *ExpressionChanged = nullptr; for (unsigned i = 0; ; ++i) { // The last operation (which comes earliest in the IR) is special as both // operands will come from Ops, rather than just one with the other being // a subexpression. if (i+2 == Ops.size()) { Value *NewLHS = Ops[i].Op; Value *NewRHS = Ops[i+1].Op; Value *OldLHS = Op->getOperand(0); Value *OldRHS = Op->getOperand(1); if (NewLHS == OldLHS && NewRHS == OldRHS) // Nothing changed, leave it alone. break; if (NewLHS == OldRHS && NewRHS == OldLHS) { // The order of the operands was reversed. Swap them. LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n'); Op->swapOperands(); LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n'); MadeChange = true; ++NumChanged; break; } // The new operation differs non-trivially from the original. Overwrite // the old operands with the new ones. LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n'); if (NewLHS != OldLHS) { BinaryOperator *BO = isReassociableOp(OldLHS, Opcode); if (BO && !NotRewritable.count(BO)) NodesToRewrite.push_back(BO); Op->setOperand(0, NewLHS); } if (NewRHS != OldRHS) { BinaryOperator *BO = isReassociableOp(OldRHS, Opcode); if (BO && !NotRewritable.count(BO)) NodesToRewrite.push_back(BO); Op->setOperand(1, NewRHS); } LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n'); ExpressionChanged = Op; MadeChange = true; ++NumChanged; break; } // Not the last operation. The left-hand side will be a sub-expression // while the right-hand side will be the current element of Ops. Value *NewRHS = Ops[i].Op; if (NewRHS != Op->getOperand(1)) { LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n'); if (NewRHS == Op->getOperand(0)) { // The new right-hand side was already present as the left operand. If // we are lucky then swapping the operands will sort out both of them. Op->swapOperands(); } else { // Overwrite with the new right-hand side. BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode); if (BO && !NotRewritable.count(BO)) NodesToRewrite.push_back(BO); Op->setOperand(1, NewRHS); ExpressionChanged = Op; } LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n'); MadeChange = true; ++NumChanged; } // Now deal with the left-hand side. If this is already an operation node // from the original expression then just rewrite the rest of the expression // into it. BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode); if (BO && !NotRewritable.count(BO)) { Op = BO; continue; } // Otherwise, grab a spare node from the original expression and use that as // the left-hand side. If there are no nodes left then the optimizers made // an expression with more nodes than the original! This usually means that // they did something stupid but it might mean that the problem was just too // hard (finding the mimimal number of multiplications needed to realize a // multiplication expression is NP-complete). Whatever the reason, smart or // stupid, create a new node if there are none left. BinaryOperator *NewOp; if (NodesToRewrite.empty()) { Constant *Undef = UndefValue::get(I->getType()); NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode), Undef, Undef, "", I); if (NewOp->getType()->isFPOrFPVectorTy()) NewOp->setFastMathFlags(I->getFastMathFlags()); } else { NewOp = NodesToRewrite.pop_back_val(); } LLVM_DEBUG(dbgs() << "RA: " << *Op << '\n'); Op->setOperand(0, NewOp); LLVM_DEBUG(dbgs() << "TO: " << *Op << '\n'); ExpressionChanged = Op; MadeChange = true; ++NumChanged; Op = NewOp; } // If the expression changed non-trivially then clear out all subclass data // starting from the operator specified in ExpressionChanged, and compactify // the operators to just before the expression root to guarantee that the // expression tree is dominated by all of Ops. if (ExpressionChanged) do { // Preserve FastMathFlags. if (isa(I)) { FastMathFlags Flags = I->getFastMathFlags(); ExpressionChanged->clearSubclassOptionalData(); ExpressionChanged->setFastMathFlags(Flags); } else ExpressionChanged->clearSubclassOptionalData(); if (ExpressionChanged == I) break; // Discard any debug info related to the expressions that has changed (we // can leave debug infor related to the root, since the result of the // expression tree should be the same even after reassociation). replaceDbgUsesWithUndef(ExpressionChanged); ExpressionChanged->moveBefore(I); ExpressionChanged = cast(*ExpressionChanged->user_begin()); } while (true); // Throw away any left over nodes from the original expression. for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i) RedoInsts.insert(NodesToRewrite[i]); } /// Insert instructions before the instruction pointed to by BI, /// that computes the negative version of the value specified. The negative /// version of the value is returned, and BI is left pointing at the instruction /// that should be processed next by the reassociation pass. /// Also add intermediate instructions to the redo list that are modified while /// pushing the negates through adds. These will be revisited to see if /// additional opportunities have been exposed. static Value *NegateValue(Value *V, Instruction *BI, ReassociatePass::OrderedSet &ToRedo) { if (auto *C = dyn_cast(V)) return C->getType()->isFPOrFPVectorTy() ? ConstantExpr::getFNeg(C) : ConstantExpr::getNeg(C); // We are trying to expose opportunity for reassociation. One of the things // that we want to do to achieve this is to push a negation as deep into an // expression chain as possible, to expose the add instructions. In practice, // this means that we turn this: // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate // the constants. We assume that instcombine will clean up the mess later if // we introduce tons of unnecessary negation instructions. // if (BinaryOperator *I = isReassociableOp(V, Instruction::Add, Instruction::FAdd)) { // Push the negates through the add. I->setOperand(0, NegateValue(I->getOperand(0), BI, ToRedo)); I->setOperand(1, NegateValue(I->getOperand(1), BI, ToRedo)); if (I->getOpcode() == Instruction::Add) { I->setHasNoUnsignedWrap(false); I->setHasNoSignedWrap(false); } // We must move the add instruction here, because the neg instructions do // not dominate the old add instruction in general. By moving it, we are // assured that the neg instructions we just inserted dominate the // instruction we are about to insert after them. // I->moveBefore(BI); I->setName(I->getName()+".neg"); // Add the intermediate negates to the redo list as processing them later // could expose more reassociating opportunities. ToRedo.insert(I); return I; } // Okay, we need to materialize a negated version of V with an instruction. // Scan the use lists of V to see if we have one already. for (User *U : V->users()) { if (!match(U, m_Neg(m_Value())) && !match(U, m_FNeg(m_Value()))) continue; // We found one! Now we have to make sure that the definition dominates // this use. We do this by moving it to the entry block (if it is a // non-instruction value) or right after the definition. These negates will // be zapped by reassociate later, so we don't need much finesse here. Instruction *TheNeg = cast(U); // Verify that the negate is in this function, V might be a constant expr. if (TheNeg->getParent()->getParent() != BI->getParent()->getParent()) continue; bool FoundCatchSwitch = false; BasicBlock::iterator InsertPt; if (Instruction *InstInput = dyn_cast(V)) { if (InvokeInst *II = dyn_cast(InstInput)) { InsertPt = II->getNormalDest()->begin(); } else { InsertPt = ++InstInput->getIterator(); } const BasicBlock *BB = InsertPt->getParent(); // Make sure we don't move anything before PHIs or exception // handling pads. while (InsertPt != BB->end() && (isa(InsertPt) || InsertPt->isEHPad())) { if (isa(InsertPt)) // A catchswitch cannot have anything in the block except // itself and PHIs. We'll bail out below. FoundCatchSwitch = true; ++InsertPt; } } else { InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin(); } // We found a catchswitch in the block where we want to move the // neg. We cannot move anything into that block. Bail and just // create the neg before BI, as if we hadn't found an existing // neg. if (FoundCatchSwitch) break; TheNeg->moveBefore(&*InsertPt); if (TheNeg->getOpcode() == Instruction::Sub) { TheNeg->setHasNoUnsignedWrap(false); TheNeg->setHasNoSignedWrap(false); } else { TheNeg->andIRFlags(BI); } ToRedo.insert(TheNeg); return TheNeg; } // Insert a 'neg' instruction that subtracts the value from zero to get the // negation. BinaryOperator *NewNeg = CreateNeg(V, V->getName() + ".neg", BI, BI); ToRedo.insert(NewNeg); return NewNeg; } /// Return true if we should break up this subtract of X-Y into (X + -Y). static bool ShouldBreakUpSubtract(Instruction *Sub) { // If this is a negation, we can't split it up! if (match(Sub, m_Neg(m_Value())) || match(Sub, m_FNeg(m_Value()))) return false; // Don't breakup X - undef. if (isa(Sub->getOperand(1))) return false; // Don't bother to break this up unless either the LHS is an associable add or // subtract or if this is only used by one. Value *V0 = Sub->getOperand(0); if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) || isReassociableOp(V0, Instruction::Sub, Instruction::FSub)) return true; Value *V1 = Sub->getOperand(1); if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) || isReassociableOp(V1, Instruction::Sub, Instruction::FSub)) return true; Value *VB = Sub->user_back(); if (Sub->hasOneUse() && (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) || isReassociableOp(VB, Instruction::Sub, Instruction::FSub))) return true; return false; } /// If we have (X-Y), and if either X is an add, or if this is only used by an /// add, transform this into (X+(0-Y)) to promote better reassociation. static BinaryOperator *BreakUpSubtract(Instruction *Sub, ReassociatePass::OrderedSet &ToRedo) { // Convert a subtract into an add and a neg instruction. This allows sub // instructions to be commuted with other add instructions. // // Calculate the negative value of Operand 1 of the sub instruction, // and set it as the RHS of the add instruction we just made. Value *NegVal = NegateValue(Sub->getOperand(1), Sub, ToRedo); BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub); Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op. Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op. New->takeName(Sub); // Everyone now refers to the add instruction. Sub->replaceAllUsesWith(New); New->setDebugLoc(Sub->getDebugLoc()); LLVM_DEBUG(dbgs() << "Negated: " << *New << '\n'); return New; } /// If this is a shift of a reassociable multiply or is used by one, change /// this into a multiply by a constant to assist with further reassociation. static BinaryOperator *ConvertShiftToMul(Instruction *Shl) { Constant *MulCst = ConstantInt::get(Shl->getType(), 1); MulCst = ConstantExpr::getShl(MulCst, cast(Shl->getOperand(1))); BinaryOperator *Mul = BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl); Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op. Mul->takeName(Shl); // Everyone now refers to the mul instruction. Shl->replaceAllUsesWith(Mul); Mul->setDebugLoc(Shl->getDebugLoc()); // We can safely preserve the nuw flag in all cases. It's also safe to turn a // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special // handling. bool NSW = cast(Shl)->hasNoSignedWrap(); bool NUW = cast(Shl)->hasNoUnsignedWrap(); if (NSW && NUW) Mul->setHasNoSignedWrap(true); Mul->setHasNoUnsignedWrap(NUW); return Mul; } /// Scan backwards and forwards among values with the same rank as element i /// to see if X exists. If X does not exist, return i. This is useful when /// scanning for 'x' when we see '-x' because they both get the same rank. static unsigned FindInOperandList(const SmallVectorImpl &Ops, unsigned i, Value *X) { unsigned XRank = Ops[i].Rank; unsigned e = Ops.size(); for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) { if (Ops[j].Op == X) return j; if (Instruction *I1 = dyn_cast(Ops[j].Op)) if (Instruction *I2 = dyn_cast(X)) if (I1->isIdenticalTo(I2)) return j; } // Scan backwards. for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) { if (Ops[j].Op == X) return j; if (Instruction *I1 = dyn_cast(Ops[j].Op)) if (Instruction *I2 = dyn_cast(X)) if (I1->isIdenticalTo(I2)) return j; } return i; } /// Emit a tree of add instructions, summing Ops together /// and returning the result. Insert the tree before I. static Value *EmitAddTreeOfValues(Instruction *I, SmallVectorImpl &Ops) { if (Ops.size() == 1) return Ops.back(); Value *V1 = Ops.back(); Ops.pop_back(); Value *V2 = EmitAddTreeOfValues(I, Ops); return CreateAdd(V2, V1, "reass.add", I, I); } /// If V is an expression tree that is a multiplication sequence, /// and if this sequence contains a multiply by Factor, /// remove Factor from the tree and return the new tree. Value *ReassociatePass::RemoveFactorFromExpression(Value *V, Value *Factor) { BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul); if (!BO) return nullptr; SmallVector Tree; MadeChange |= LinearizeExprTree(BO, Tree); SmallVector Factors; Factors.reserve(Tree.size()); for (unsigned i = 0, e = Tree.size(); i != e; ++i) { RepeatedValue E = Tree[i]; Factors.append(E.second.getZExtValue(), ValueEntry(getRank(E.first), E.first)); } bool FoundFactor = false; bool NeedsNegate = false; for (unsigned i = 0, e = Factors.size(); i != e; ++i) { if (Factors[i].Op == Factor) { FoundFactor = true; Factors.erase(Factors.begin()+i); break; } // If this is a negative version of this factor, remove it. if (ConstantInt *FC1 = dyn_cast(Factor)) { if (ConstantInt *FC2 = dyn_cast(Factors[i].Op)) if (FC1->getValue() == -FC2->getValue()) { FoundFactor = NeedsNegate = true; Factors.erase(Factors.begin()+i); break; } } else if (ConstantFP *FC1 = dyn_cast(Factor)) { if (ConstantFP *FC2 = dyn_cast(Factors[i].Op)) { const APFloat &F1 = FC1->getValueAPF(); APFloat F2(FC2->getValueAPF()); F2.changeSign(); if (F1.compare(F2) == APFloat::cmpEqual) { FoundFactor = NeedsNegate = true; Factors.erase(Factors.begin() + i); break; } } } } if (!FoundFactor) { // Make sure to restore the operands to the expression tree. RewriteExprTree(BO, Factors); return nullptr; } BasicBlock::iterator InsertPt = ++BO->getIterator(); // If this was just a single multiply, remove the multiply and return the only // remaining operand. if (Factors.size() == 1) { RedoInsts.insert(BO); V = Factors[0].Op; } else { RewriteExprTree(BO, Factors); V = BO; } if (NeedsNegate) V = CreateNeg(V, "neg", &*InsertPt, BO); return V; } /// If V is a single-use multiply, recursively add its operands as factors, /// otherwise add V to the list of factors. /// /// Ops is the top-level list of add operands we're trying to factor. static void FindSingleUseMultiplyFactors(Value *V, SmallVectorImpl &Factors) { BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul); if (!BO) { Factors.push_back(V); return; } // Otherwise, add the LHS and RHS to the list of factors. FindSingleUseMultiplyFactors(BO->getOperand(1), Factors); FindSingleUseMultiplyFactors(BO->getOperand(0), Factors); } /// Optimize a series of operands to an 'and', 'or', or 'xor' instruction. /// This optimizes based on identities. If it can be reduced to a single Value, /// it is returned, otherwise the Ops list is mutated as necessary. static Value *OptimizeAndOrXor(unsigned Opcode, SmallVectorImpl &Ops) { // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. for (unsigned i = 0, e = Ops.size(); i != e; ++i) { // First, check for X and ~X in the operand list. assert(i < Ops.size()); Value *X; if (match(Ops[i].Op, m_Not(m_Value(X)))) { // Cannot occur for ^. unsigned FoundX = FindInOperandList(Ops, i, X); if (FoundX != i) { if (Opcode == Instruction::And) // ...&X&~X = 0 return Constant::getNullValue(X->getType()); if (Opcode == Instruction::Or) // ...|X|~X = -1 return Constant::getAllOnesValue(X->getType()); } } // Next, check for duplicate pairs of values, which we assume are next to // each other, due to our sorting criteria. assert(i < Ops.size()); if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { if (Opcode == Instruction::And || Opcode == Instruction::Or) { // Drop duplicate values for And and Or. Ops.erase(Ops.begin()+i); --i; --e; ++NumAnnihil; continue; } // Drop pairs of values for Xor. assert(Opcode == Instruction::Xor); if (e == 2) return Constant::getNullValue(Ops[0].Op->getType()); // Y ^ X^X -> Y Ops.erase(Ops.begin()+i, Ops.begin()+i+2); i -= 1; e -= 2; ++NumAnnihil; } } return nullptr; } /// Helper function of CombineXorOpnd(). It creates a bitwise-and /// instruction with the given two operands, and return the resulting /// instruction. There are two special cases: 1) if the constant operand is 0, /// it will return NULL. 2) if the constant is ~0, the symbolic operand will /// be returned. static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd, const APInt &ConstOpnd) { if (ConstOpnd.isNullValue()) return nullptr; if (ConstOpnd.isAllOnesValue()) return Opnd; Instruction *I = BinaryOperator::CreateAnd( Opnd, ConstantInt::get(Opnd->getType(), ConstOpnd), "and.ra", InsertBefore); I->setDebugLoc(InsertBefore->getDebugLoc()); return I; } // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd" // into "R ^ C", where C would be 0, and R is a symbolic value. // // If it was successful, true is returned, and the "R" and "C" is returned // via "Res" and "ConstOpnd", respectively; otherwise, false is returned, // and both "Res" and "ConstOpnd" remain unchanged. bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, APInt &ConstOpnd, Value *&Res) { // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2 // = ((x | c1) ^ c1) ^ (c1 ^ c2) // = (x & ~c1) ^ (c1 ^ c2) // It is useful only when c1 == c2. if (!Opnd1->isOrExpr() || Opnd1->getConstPart().isNullValue()) return false; if (!Opnd1->getValue()->hasOneUse()) return false; const APInt &C1 = Opnd1->getConstPart(); if (C1 != ConstOpnd) return false; Value *X = Opnd1->getSymbolicPart(); Res = createAndInstr(I, X, ~C1); // ConstOpnd was C2, now C1 ^ C2. ConstOpnd ^= C1; if (Instruction *T = dyn_cast(Opnd1->getValue())) RedoInsts.insert(T); return true; } // Helper function of OptimizeXor(). It tries to simplify // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a // symbolic value. // // If it was successful, true is returned, and the "R" and "C" is returned // via "Res" and "ConstOpnd", respectively (If the entire expression is // evaluated to a constant, the Res is set to NULL); otherwise, false is // returned, and both "Res" and "ConstOpnd" remain unchanged. bool ReassociatePass::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2, APInt &ConstOpnd, Value *&Res) { Value *X = Opnd1->getSymbolicPart(); if (X != Opnd2->getSymbolicPart()) return false; // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.) int DeadInstNum = 1; if (Opnd1->getValue()->hasOneUse()) DeadInstNum++; if (Opnd2->getValue()->hasOneUse()) DeadInstNum++; // Xor-Rule 2: // (x | c1) ^ (x & c2) // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3 // if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) { if (Opnd2->isOrExpr()) std::swap(Opnd1, Opnd2); const APInt &C1 = Opnd1->getConstPart(); const APInt &C2 = Opnd2->getConstPart(); APInt C3((~C1) ^ C2); // Do not increase code size! if (!C3.isNullValue() && !C3.isAllOnesValue()) { int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2; if (NewInstNum > DeadInstNum) return false; } Res = createAndInstr(I, X, C3); ConstOpnd ^= C1; } else if (Opnd1->isOrExpr()) { // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2 // const APInt &C1 = Opnd1->getConstPart(); const APInt &C2 = Opnd2->getConstPart(); APInt C3 = C1 ^ C2; // Do not increase code size if (!C3.isNullValue() && !C3.isAllOnesValue()) { int NewInstNum = ConstOpnd.getBoolValue() ? 1 : 2; if (NewInstNum > DeadInstNum) return false; } Res = createAndInstr(I, X, C3); ConstOpnd ^= C3; } else { // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2)) // const APInt &C1 = Opnd1->getConstPart(); const APInt &C2 = Opnd2->getConstPart(); APInt C3 = C1 ^ C2; Res = createAndInstr(I, X, C3); } // Put the original operands in the Redo list; hope they will be deleted // as dead code. if (Instruction *T = dyn_cast(Opnd1->getValue())) RedoInsts.insert(T); if (Instruction *T = dyn_cast(Opnd2->getValue())) RedoInsts.insert(T); return true; } /// Optimize a series of operands to an 'xor' instruction. If it can be reduced /// to a single Value, it is returned, otherwise the Ops list is mutated as /// necessary. Value *ReassociatePass::OptimizeXor(Instruction *I, SmallVectorImpl &Ops) { if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops)) return V; if (Ops.size() == 1) return nullptr; SmallVector Opnds; SmallVector OpndPtrs; Type *Ty = Ops[0].Op->getType(); APInt ConstOpnd(Ty->getScalarSizeInBits(), 0); // Step 1: Convert ValueEntry to XorOpnd for (unsigned i = 0, e = Ops.size(); i != e; ++i) { Value *V = Ops[i].Op; const APInt *C; // TODO: Support non-splat vectors. if (match(V, m_APInt(C))) { ConstOpnd ^= *C; } else { XorOpnd O(V); O.setSymbolicRank(getRank(O.getSymbolicPart())); Opnds.push_back(O); } } // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds". // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate // the "OpndPtrs" as well. For the similar reason, do not fuse this loop // with the previous loop --- the iterator of the "Opnds" may be invalidated // when new elements are added to the vector. for (unsigned i = 0, e = Opnds.size(); i != e; ++i) OpndPtrs.push_back(&Opnds[i]); // Step 2: Sort the Xor-Operands in a way such that the operands containing // the same symbolic value cluster together. For instance, the input operand // sequence ("x | 123", "y & 456", "x & 789") will be sorted into: // ("x | 123", "x & 789", "y & 456"). // // The purpose is twofold: // 1) Cluster together the operands sharing the same symbolic-value. // 2) Operand having smaller symbolic-value-rank is permuted earlier, which // could potentially shorten crital path, and expose more loop-invariants. // Note that values' rank are basically defined in RPO order (FIXME). // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2", // "z" in the order of X-Y-Z is better than any other orders. llvm::stable_sort(OpndPtrs, [](XorOpnd *LHS, XorOpnd *RHS) { return LHS->getSymbolicRank() < RHS->getSymbolicRank(); }); // Step 3: Combine adjacent operands XorOpnd *PrevOpnd = nullptr; bool Changed = false; for (unsigned i = 0, e = Opnds.size(); i < e; i++) { XorOpnd *CurrOpnd = OpndPtrs[i]; // The combined value Value *CV; // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd" if (!ConstOpnd.isNullValue() && CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) { Changed = true; if (CV) *CurrOpnd = XorOpnd(CV); else { CurrOpnd->Invalidate(); continue; } } if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) { PrevOpnd = CurrOpnd; continue; } // step 3.2: When previous and current operands share the same symbolic // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd" if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) { // Remove previous operand PrevOpnd->Invalidate(); if (CV) { *CurrOpnd = XorOpnd(CV); PrevOpnd = CurrOpnd; } else { CurrOpnd->Invalidate(); PrevOpnd = nullptr; } Changed = true; } } // Step 4: Reassemble the Ops if (Changed) { Ops.clear(); for (unsigned int i = 0, e = Opnds.size(); i < e; i++) { XorOpnd &O = Opnds[i]; if (O.isInvalid()) continue; ValueEntry VE(getRank(O.getValue()), O.getValue()); Ops.push_back(VE); } if (!ConstOpnd.isNullValue()) { Value *C = ConstantInt::get(Ty, ConstOpnd); ValueEntry VE(getRank(C), C); Ops.push_back(VE); } unsigned Sz = Ops.size(); if (Sz == 1) return Ops.back().Op; if (Sz == 0) { assert(ConstOpnd.isNullValue()); return ConstantInt::get(Ty, ConstOpnd); } } return nullptr; } /// Optimize a series of operands to an 'add' instruction. This /// optimizes based on identities. If it can be reduced to a single Value, it /// is returned, otherwise the Ops list is mutated as necessary. Value *ReassociatePass::OptimizeAdd(Instruction *I, SmallVectorImpl &Ops) { // Scan the operand lists looking for X and -X pairs. If we find any, we // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it, // scan for any // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z. for (unsigned i = 0, e = Ops.size(); i != e; ++i) { Value *TheOp = Ops[i].Op; // Check to see if we've seen this operand before. If so, we factor all // instances of the operand together. Due to our sorting criteria, we know // that these need to be next to each other in the vector. if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) { // Rescan the list, remove all instances of this operand from the expr. unsigned NumFound = 0; do { Ops.erase(Ops.begin()+i); ++NumFound; } while (i != Ops.size() && Ops[i].Op == TheOp); LLVM_DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n'); ++NumFactor; // Insert a new multiply. Type *Ty = TheOp->getType(); Constant *C = Ty->isIntOrIntVectorTy() ? ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound); Instruction *Mul = CreateMul(TheOp, C, "factor", I, I); // Now that we have inserted a multiply, optimize it. This allows us to // handle cases that require multiple factoring steps, such as this: // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6 RedoInsts.insert(Mul); // If every add operand was a duplicate, return the multiply. if (Ops.empty()) return Mul; // Otherwise, we had some input that didn't have the dupe, such as // "A + A + B" -> "A*2 + B". Add the new multiply to the list of // things being added by this operation. Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul)); --i; e = Ops.size(); continue; } // Check for X and -X or X and ~X in the operand list. Value *X; if (!match(TheOp, m_Neg(m_Value(X))) && !match(TheOp, m_Not(m_Value(X))) && !match(TheOp, m_FNeg(m_Value(X)))) continue; unsigned FoundX = FindInOperandList(Ops, i, X); if (FoundX == i) continue; // Remove X and -X from the operand list. if (Ops.size() == 2 && (match(TheOp, m_Neg(m_Value())) || match(TheOp, m_FNeg(m_Value())))) return Constant::getNullValue(X->getType()); // Remove X and ~X from the operand list. if (Ops.size() == 2 && match(TheOp, m_Not(m_Value()))) return Constant::getAllOnesValue(X->getType()); Ops.erase(Ops.begin()+i); if (i < FoundX) --FoundX; else --i; // Need to back up an extra one. Ops.erase(Ops.begin()+FoundX); ++NumAnnihil; --i; // Revisit element. e -= 2; // Removed two elements. // if X and ~X we append -1 to the operand list. if (match(TheOp, m_Not(m_Value()))) { Value *V = Constant::getAllOnesValue(X->getType()); Ops.insert(Ops.end(), ValueEntry(getRank(V), V)); e += 1; } } // Scan the operand list, checking to see if there are any common factors // between operands. Consider something like A*A+A*B*C+D. We would like to // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. // To efficiently find this, we count the number of times a factor occurs // for any ADD operands that are MULs. DenseMap FactorOccurrences; // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4) // where they are actually the same multiply. unsigned MaxOcc = 0; Value *MaxOccVal = nullptr; for (unsigned i = 0, e = Ops.size(); i != e; ++i) { BinaryOperator *BOp = isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul); if (!BOp) continue; // Compute all of the factors of this added value. SmallVector Factors; FindSingleUseMultiplyFactors(BOp, Factors); assert(Factors.size() > 1 && "Bad linearize!"); // Add one to FactorOccurrences for each unique factor in this op. SmallPtrSet Duplicates; for (unsigned i = 0, e = Factors.size(); i != e; ++i) { Value *Factor = Factors[i]; if (!Duplicates.insert(Factor).second) continue; unsigned Occ = ++FactorOccurrences[Factor]; if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; } // If Factor is a negative constant, add the negated value as a factor // because we can percolate the negate out. Watch for minint, which // cannot be positivified. if (ConstantInt *CI = dyn_cast(Factor)) { if (CI->isNegative() && !CI->isMinValue(true)) { Factor = ConstantInt::get(CI->getContext(), -CI->getValue()); if (!Duplicates.insert(Factor).second) continue; unsigned Occ = ++FactorOccurrences[Factor]; if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; } } } else if (ConstantFP *CF = dyn_cast(Factor)) { if (CF->isNegative()) { APFloat F(CF->getValueAPF()); F.changeSign(); Factor = ConstantFP::get(CF->getContext(), F); if (!Duplicates.insert(Factor).second) continue; unsigned Occ = ++FactorOccurrences[Factor]; if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; } } } } } // If any factor occurred more than one time, we can pull it out. if (MaxOcc > 1) { LLVM_DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n'); ++NumFactor; // Create a new instruction that uses the MaxOccVal twice. If we don't do // this, we could otherwise run into situations where removing a factor // from an expression will drop a use of maxocc, and this can cause // RemoveFactorFromExpression on successive values to behave differently. Instruction *DummyInst = I->getType()->isIntOrIntVectorTy() ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal) : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal); SmallVector NewMulOps; for (unsigned i = 0; i != Ops.size(); ++i) { // Only try to remove factors from expressions we're allowed to. BinaryOperator *BOp = isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul); if (!BOp) continue; if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { // The factorized operand may occur several times. Convert them all in // one fell swoop. for (unsigned j = Ops.size(); j != i;) { --j; if (Ops[j].Op == Ops[i].Op) { NewMulOps.push_back(V); Ops.erase(Ops.begin()+j); } } --i; } } // No need for extra uses anymore. DummyInst->deleteValue(); unsigned NumAddedValues = NewMulOps.size(); Value *V = EmitAddTreeOfValues(I, NewMulOps); // Now that we have inserted the add tree, optimize it. This allows us to // handle cases that require multiple factoring steps, such as this: // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) assert(NumAddedValues > 1 && "Each occurrence should contribute a value"); (void)NumAddedValues; if (Instruction *VI = dyn_cast(V)) RedoInsts.insert(VI); // Create the multiply. Instruction *V2 = CreateMul(V, MaxOccVal, "reass.mul", I, I); // Rerun associate on the multiply in case the inner expression turned into // a multiply. We want to make sure that we keep things in canonical form. RedoInsts.insert(V2); // If every add operand included the factor (e.g. "A*B + A*C"), then the // entire result expression is just the multiply "A*(B+C)". if (Ops.empty()) return V2; // Otherwise, we had some input that didn't have the factor, such as // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of // things being added by this operation. Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); } return nullptr; } /// Build up a vector of value/power pairs factoring a product. /// /// Given a series of multiplication operands, build a vector of factors and /// the powers each is raised to when forming the final product. Sort them in /// the order of descending power. /// /// (x*x) -> [(x, 2)] /// ((x*x)*x) -> [(x, 3)] /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)] /// /// \returns Whether any factors have a power greater than one. static bool collectMultiplyFactors(SmallVectorImpl &Ops, SmallVectorImpl &Factors) { // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this. // Compute the sum of powers of simplifiable factors. unsigned FactorPowerSum = 0; for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) { Value *Op = Ops[Idx-1].Op; // Count the number of occurrences of this value. unsigned Count = 1; for (; Idx < Size && Ops[Idx].Op == Op; ++Idx) ++Count; // Track for simplification all factors which occur 2 or more times. if (Count > 1) FactorPowerSum += Count; } // We can only simplify factors if the sum of the powers of our simplifiable // factors is 4 or higher. When that is the case, we will *always* have // a simplification. This is an important invariant to prevent cyclicly // trying to simplify already minimal formations. if (FactorPowerSum < 4) return false; // Now gather the simplifiable factors, removing them from Ops. FactorPowerSum = 0; for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) { Value *Op = Ops[Idx-1].Op; // Count the number of occurrences of this value. unsigned Count = 1; for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx) ++Count; if (Count == 1) continue; // Move an even number of occurrences to Factors. Count &= ~1U; Idx -= Count; FactorPowerSum += Count; Factors.push_back(Factor(Op, Count)); Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count); } // None of the adjustments above should have reduced the sum of factor powers // below our mininum of '4'. assert(FactorPowerSum >= 4); llvm::stable_sort(Factors, [](const Factor &LHS, const Factor &RHS) { return LHS.Power > RHS.Power; }); return true; } /// Build a tree of multiplies, computing the product of Ops. static Value *buildMultiplyTree(IRBuilder<> &Builder, SmallVectorImpl &Ops) { if (Ops.size() == 1) return Ops.back(); Value *LHS = Ops.pop_back_val(); do { if (LHS->getType()->isIntOrIntVectorTy()) LHS = Builder.CreateMul(LHS, Ops.pop_back_val()); else LHS = Builder.CreateFMul(LHS, Ops.pop_back_val()); } while (!Ops.empty()); return LHS; } /// Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*... /// /// Given a vector of values raised to various powers, where no two values are /// equal and the powers are sorted in decreasing order, compute the minimal /// DAG of multiplies to compute the final product, and return that product /// value. Value * ReassociatePass::buildMinimalMultiplyDAG(IRBuilder<> &Builder, SmallVectorImpl &Factors) { assert(Factors[0].Power); SmallVector OuterProduct; for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size(); Idx < Size && Factors[Idx].Power > 0; ++Idx) { if (Factors[Idx].Power != Factors[LastIdx].Power) { LastIdx = Idx; continue; } // We want to multiply across all the factors with the same power so that // we can raise them to that power as a single entity. Build a mini tree // for that. SmallVector InnerProduct; InnerProduct.push_back(Factors[LastIdx].Base); do { InnerProduct.push_back(Factors[Idx].Base); ++Idx; } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power); // Reset the base value of the first factor to the new expression tree. // We'll remove all the factors with the same power in a second pass. Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct); if (Instruction *MI = dyn_cast(M)) RedoInsts.insert(MI); LastIdx = Idx; } // Unique factors with equal powers -- we've folded them into the first one's // base. Factors.erase(std::unique(Factors.begin(), Factors.end(), [](const Factor &LHS, const Factor &RHS) { return LHS.Power == RHS.Power; }), Factors.end()); // Iteratively collect the base of each factor with an add power into the // outer product, and halve each power in preparation for squaring the // expression. for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) { if (Factors[Idx].Power & 1) OuterProduct.push_back(Factors[Idx].Base); Factors[Idx].Power >>= 1; } if (Factors[0].Power) { Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors); OuterProduct.push_back(SquareRoot); OuterProduct.push_back(SquareRoot); } if (OuterProduct.size() == 1) return OuterProduct.front(); Value *V = buildMultiplyTree(Builder, OuterProduct); return V; } Value *ReassociatePass::OptimizeMul(BinaryOperator *I, SmallVectorImpl &Ops) { // We can only optimize the multiplies when there is a chain of more than // three, such that a balanced tree might require fewer total multiplies. if (Ops.size() < 4) return nullptr; // Try to turn linear trees of multiplies without other uses of the // intermediate stages into minimal multiply DAGs with perfect sub-expression // re-use. SmallVector Factors; if (!collectMultiplyFactors(Ops, Factors)) return nullptr; // All distinct factors, so nothing left for us to do. IRBuilder<> Builder(I); // The reassociate transformation for FP operations is performed only // if unsafe algebra is permitted by FastMathFlags. Propagate those flags // to the newly generated operations. if (auto FPI = dyn_cast(I)) Builder.setFastMathFlags(FPI->getFastMathFlags()); Value *V = buildMinimalMultiplyDAG(Builder, Factors); if (Ops.empty()) return V; ValueEntry NewEntry = ValueEntry(getRank(V), V); Ops.insert(llvm::lower_bound(Ops, NewEntry), NewEntry); return nullptr; } Value *ReassociatePass::OptimizeExpression(BinaryOperator *I, SmallVectorImpl &Ops) { // Now that we have the linearized expression tree, try to optimize it. // Start by folding any constants that we found. Constant *Cst = nullptr; unsigned Opcode = I->getOpcode(); while (!Ops.empty() && isa(Ops.back().Op)) { Constant *C = cast(Ops.pop_back_val().Op); Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C; } // If there was nothing but constants then we are done. if (Ops.empty()) return Cst; // Put the combined constant back at the end of the operand list, except if // there is no point. For example, an add of 0 gets dropped here, while a // multiplication by zero turns the whole expression into zero. if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) { if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType())) return Cst; Ops.push_back(ValueEntry(0, Cst)); } if (Ops.size() == 1) return Ops[0].Op; // Handle destructive annihilation due to identities between elements in the // argument list here. unsigned NumOps = Ops.size(); switch (Opcode) { default: break; case Instruction::And: case Instruction::Or: if (Value *Result = OptimizeAndOrXor(Opcode, Ops)) return Result; break; case Instruction::Xor: if (Value *Result = OptimizeXor(I, Ops)) return Result; break; case Instruction::Add: case Instruction::FAdd: if (Value *Result = OptimizeAdd(I, Ops)) return Result; break; case Instruction::Mul: case Instruction::FMul: if (Value *Result = OptimizeMul(I, Ops)) return Result; break; } if (Ops.size() != NumOps) return OptimizeExpression(I, Ops); return nullptr; } // Remove dead instructions and if any operands are trivially dead add them to // Insts so they will be removed as well. void ReassociatePass::RecursivelyEraseDeadInsts(Instruction *I, OrderedSet &Insts) { assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!"); SmallVector Ops(I->op_begin(), I->op_end()); ValueRankMap.erase(I); Insts.remove(I); RedoInsts.remove(I); I->eraseFromParent(); for (auto Op : Ops) if (Instruction *OpInst = dyn_cast(Op)) if (OpInst->use_empty()) Insts.insert(OpInst); } /// Zap the given instruction, adding interesting operands to the work list. void ReassociatePass::EraseInst(Instruction *I) { assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!"); LLVM_DEBUG(dbgs() << "Erasing dead inst: "; I->dump()); SmallVector Ops(I->op_begin(), I->op_end()); // Erase the dead instruction. ValueRankMap.erase(I); RedoInsts.remove(I); I->eraseFromParent(); // Optimize its operands. SmallPtrSet Visited; // Detect self-referential nodes. for (unsigned i = 0, e = Ops.size(); i != e; ++i) if (Instruction *Op = dyn_cast(Ops[i])) { // If this is a node in an expression tree, climb to the expression root // and add that since that's where optimization actually happens. unsigned Opcode = Op->getOpcode(); while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode && Visited.insert(Op).second) Op = Op->user_back(); // The instruction we're going to push may be coming from a // dead block, and Reassociate skips the processing of unreachable // blocks because it's a waste of time and also because it can // lead to infinite loop due to LLVM's non-standard definition // of dominance. if (ValueRankMap.find(Op) != ValueRankMap.end()) RedoInsts.insert(Op); } MadeChange = true; } /// Recursively analyze an expression to build a list of instructions that have /// negative floating-point constant operands. The caller can then transform /// the list to create positive constants for better reassociation and CSE. static void getNegatibleInsts(Value *V, SmallVectorImpl &Candidates) { // Handle only one-use instructions. Combining negations does not justify // replicating instructions. Instruction *I; if (!match(V, m_OneUse(m_Instruction(I)))) return; // Handle expressions of multiplications and divisions. // TODO: This could look through floating-point casts. const APFloat *C; switch (I->getOpcode()) { case Instruction::FMul: // Not expecting non-canonical code here. Bail out and wait. if (match(I->getOperand(0), m_Constant())) break; if (match(I->getOperand(1), m_APFloat(C)) && C->isNegative()) { Candidates.push_back(I); LLVM_DEBUG(dbgs() << "FMul with negative constant: " << *I << '\n'); } getNegatibleInsts(I->getOperand(0), Candidates); getNegatibleInsts(I->getOperand(1), Candidates); break; case Instruction::FDiv: // Not expecting non-canonical code here. Bail out and wait. if (match(I->getOperand(0), m_Constant()) && match(I->getOperand(1), m_Constant())) break; if ((match(I->getOperand(0), m_APFloat(C)) && C->isNegative()) || (match(I->getOperand(1), m_APFloat(C)) && C->isNegative())) { Candidates.push_back(I); LLVM_DEBUG(dbgs() << "FDiv with negative constant: " << *I << '\n'); } getNegatibleInsts(I->getOperand(0), Candidates); getNegatibleInsts(I->getOperand(1), Candidates); break; default: break; } } /// Given an fadd/fsub with an operand that is a one-use instruction /// (the fadd/fsub), try to change negative floating-point constants into /// positive constants to increase potential for reassociation and CSE. Instruction *ReassociatePass::canonicalizeNegFPConstantsForOp(Instruction *I, Instruction *Op, Value *OtherOp) { assert((I->getOpcode() == Instruction::FAdd || I->getOpcode() == Instruction::FSub) && "Expected fadd/fsub"); // Collect instructions with negative FP constants from the subtree that ends // in Op. SmallVector Candidates; getNegatibleInsts(Op, Candidates); if (Candidates.empty()) return nullptr; // Don't canonicalize x + (-Constant * y) -> x - (Constant * y), if the // resulting subtract will be broken up later. This can get us into an // infinite loop during reassociation. bool IsFSub = I->getOpcode() == Instruction::FSub; bool NeedsSubtract = !IsFSub && Candidates.size() % 2 == 1; if (NeedsSubtract && ShouldBreakUpSubtract(I)) return nullptr; for (Instruction *Negatible : Candidates) { const APFloat *C; if (match(Negatible->getOperand(0), m_APFloat(C))) { assert(!match(Negatible->getOperand(1), m_Constant()) && "Expecting only 1 constant operand"); assert(C->isNegative() && "Expected negative FP constant"); Negatible->setOperand(0, ConstantFP::get(Negatible->getType(), abs(*C))); MadeChange = true; } if (match(Negatible->getOperand(1), m_APFloat(C))) { assert(!match(Negatible->getOperand(0), m_Constant()) && "Expecting only 1 constant operand"); assert(C->isNegative() && "Expected negative FP constant"); Negatible->setOperand(1, ConstantFP::get(Negatible->getType(), abs(*C))); MadeChange = true; } } assert(MadeChange == true && "Negative constant candidate was not changed"); // Negations cancelled out. if (Candidates.size() % 2 == 0) return I; // Negate the final operand in the expression by flipping the opcode of this // fadd/fsub. assert(Candidates.size() % 2 == 1 && "Expected odd number"); IRBuilder<> Builder(I); Value *NewInst = IsFSub ? Builder.CreateFAddFMF(OtherOp, Op, I) : Builder.CreateFSubFMF(OtherOp, Op, I); I->replaceAllUsesWith(NewInst); RedoInsts.insert(I); return dyn_cast(NewInst); } /// Canonicalize expressions that contain a negative floating-point constant /// of the following form: /// OtherOp + (subtree) -> OtherOp {+/-} (canonical subtree) /// (subtree) + OtherOp -> OtherOp {+/-} (canonical subtree) /// OtherOp - (subtree) -> OtherOp {+/-} (canonical subtree) /// /// The fadd/fsub opcode may be switched to allow folding a negation into the /// input instruction. Instruction *ReassociatePass::canonicalizeNegFPConstants(Instruction *I) { LLVM_DEBUG(dbgs() << "Combine negations for: " << *I << '\n'); Value *X; Instruction *Op; if (match(I, m_FAdd(m_Value(X), m_OneUse(m_Instruction(Op))))) if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, X)) I = R; if (match(I, m_FAdd(m_OneUse(m_Instruction(Op)), m_Value(X)))) if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, X)) I = R; if (match(I, m_FSub(m_Value(X), m_OneUse(m_Instruction(Op))))) if (Instruction *R = canonicalizeNegFPConstantsForOp(I, Op, X)) I = R; return I; } /// Inspect and optimize the given instruction. Note that erasing /// instructions is not allowed. void ReassociatePass::OptimizeInst(Instruction *I) { // Only consider operations that we understand. if (!isa(I) && !isa(I)) return; if (I->getOpcode() == Instruction::Shl && isa(I->getOperand(1))) // If an operand of this shift is a reassociable multiply, or if the shift // is used by a reassociable multiply or add, turn into a multiply. if (isReassociableOp(I->getOperand(0), Instruction::Mul) || (I->hasOneUse() && (isReassociableOp(I->user_back(), Instruction::Mul) || isReassociableOp(I->user_back(), Instruction::Add)))) { Instruction *NI = ConvertShiftToMul(I); RedoInsts.insert(I); MadeChange = true; I = NI; } // Commute binary operators, to canonicalize the order of their operands. // This can potentially expose more CSE opportunities, and makes writing other // transformations simpler. if (I->isCommutative()) canonicalizeOperands(I); // Canonicalize negative constants out of expressions. if (Instruction *Res = canonicalizeNegFPConstants(I)) I = Res; // Don't optimize floating-point instructions unless they are 'fast'. if (I->getType()->isFPOrFPVectorTy() && !I->isFast()) return; // Do not reassociate boolean (i1) expressions. We want to preserve the // original order of evaluation for short-circuited comparisons that // SimplifyCFG has folded to AND/OR expressions. If the expression // is not further optimized, it is likely to be transformed back to a // short-circuited form for code gen, and the source order may have been // optimized for the most likely conditions. if (I->getType()->isIntegerTy(1)) return; // If this is a subtract instruction which is not already in negate form, // see if we can convert it to X+-Y. if (I->getOpcode() == Instruction::Sub) { if (ShouldBreakUpSubtract(I)) { Instruction *NI = BreakUpSubtract(I, RedoInsts); RedoInsts.insert(I); MadeChange = true; I = NI; } else if (match(I, m_Neg(m_Value()))) { // Otherwise, this is a negation. See if the operand is a multiply tree // and if this is not an inner node of a multiply tree. if (isReassociableOp(I->getOperand(1), Instruction::Mul) && (!I->hasOneUse() || !isReassociableOp(I->user_back(), Instruction::Mul))) { Instruction *NI = LowerNegateToMultiply(I); // If the negate was simplified, revisit the users to see if we can // reassociate further. for (User *U : NI->users()) { if (BinaryOperator *Tmp = dyn_cast(U)) RedoInsts.insert(Tmp); } RedoInsts.insert(I); MadeChange = true; I = NI; } } } else if (I->getOpcode() == Instruction::FNeg || I->getOpcode() == Instruction::FSub) { if (ShouldBreakUpSubtract(I)) { Instruction *NI = BreakUpSubtract(I, RedoInsts); RedoInsts.insert(I); MadeChange = true; I = NI; } else if (match(I, m_FNeg(m_Value()))) { // Otherwise, this is a negation. See if the operand is a multiply tree // and if this is not an inner node of a multiply tree. Value *Op = isa(I) ? I->getOperand(1) : I->getOperand(0); if (isReassociableOp(Op, Instruction::FMul) && (!I->hasOneUse() || !isReassociableOp(I->user_back(), Instruction::FMul))) { // If the negate was simplified, revisit the users to see if we can // reassociate further. Instruction *NI = LowerNegateToMultiply(I); for (User *U : NI->users()) { if (BinaryOperator *Tmp = dyn_cast(U)) RedoInsts.insert(Tmp); } RedoInsts.insert(I); MadeChange = true; I = NI; } } } // If this instruction is an associative binary operator, process it. if (!I->isAssociative()) return; BinaryOperator *BO = cast(I); // If this is an interior node of a reassociable tree, ignore it until we // get to the root of the tree, to avoid N^2 analysis. unsigned Opcode = BO->getOpcode(); if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) { // During the initial run we will get to the root of the tree. // But if we get here while we are redoing instructions, there is no // guarantee that the root will be visited. So Redo later if (BO->user_back() != BO && BO->getParent() == BO->user_back()->getParent()) RedoInsts.insert(BO->user_back()); return; } // If this is an add tree that is used by a sub instruction, ignore it // until we process the subtract. if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add && cast(BO->user_back())->getOpcode() == Instruction::Sub) return; if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd && cast(BO->user_back())->getOpcode() == Instruction::FSub) return; ReassociateExpression(BO); } void ReassociatePass::ReassociateExpression(BinaryOperator *I) { // First, walk the expression tree, linearizing the tree, collecting the // operand information. SmallVector Tree; MadeChange |= LinearizeExprTree(I, Tree); SmallVector Ops; Ops.reserve(Tree.size()); for (unsigned i = 0, e = Tree.size(); i != e; ++i) { RepeatedValue E = Tree[i]; Ops.append(E.second.getZExtValue(), ValueEntry(getRank(E.first), E.first)); } LLVM_DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n'); // Now that we have linearized the tree to a list and have gathered all of // the operands and their ranks, sort the operands by their rank. Use a // stable_sort so that values with equal ranks will have their relative // positions maintained (and so the compiler is deterministic). Note that // this sorts so that the highest ranking values end up at the beginning of // the vector. llvm::stable_sort(Ops); // Now that we have the expression tree in a convenient // sorted form, optimize it globally if possible. if (Value *V = OptimizeExpression(I, Ops)) { if (V == I) // Self-referential expression in unreachable code. return; // This expression tree simplified to something that isn't a tree, // eliminate it. LLVM_DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n'); I->replaceAllUsesWith(V); if (Instruction *VI = dyn_cast(V)) if (I->getDebugLoc()) VI->setDebugLoc(I->getDebugLoc()); RedoInsts.insert(I); ++NumAnnihil; return; } // We want to sink immediates as deeply as possible except in the case where // this is a multiply tree used only by an add, and the immediate is a -1. // In this case we reassociate to put the negation on the outside so that we // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y if (I->hasOneUse()) { if (I->getOpcode() == Instruction::Mul && cast(I->user_back())->getOpcode() == Instruction::Add && isa(Ops.back().Op) && cast(Ops.back().Op)->isMinusOne()) { ValueEntry Tmp = Ops.pop_back_val(); Ops.insert(Ops.begin(), Tmp); } else if (I->getOpcode() == Instruction::FMul && cast(I->user_back())->getOpcode() == Instruction::FAdd && isa(Ops.back().Op) && cast(Ops.back().Op)->isExactlyValue(-1.0)) { ValueEntry Tmp = Ops.pop_back_val(); Ops.insert(Ops.begin(), Tmp); } } LLVM_DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n'); if (Ops.size() == 1) { if (Ops[0].Op == I) // Self-referential expression in unreachable code. return; // This expression tree simplified to something that isn't a tree, // eliminate it. I->replaceAllUsesWith(Ops[0].Op); if (Instruction *OI = dyn_cast(Ops[0].Op)) OI->setDebugLoc(I->getDebugLoc()); RedoInsts.insert(I); return; } if (Ops.size() > 2 && Ops.size() <= GlobalReassociateLimit) { // Find the pair with the highest count in the pairmap and move it to the // back of the list so that it can later be CSE'd. // example: // a*b*c*d*e // if c*e is the most "popular" pair, we can express this as // (((c*e)*d)*b)*a unsigned Max = 1; unsigned BestRank = 0; std::pair BestPair; unsigned Idx = I->getOpcode() - Instruction::BinaryOpsBegin; for (unsigned i = 0; i < Ops.size() - 1; ++i) for (unsigned j = i + 1; j < Ops.size(); ++j) { unsigned Score = 0; Value *Op0 = Ops[i].Op; Value *Op1 = Ops[j].Op; if (std::less()(Op1, Op0)) std::swap(Op0, Op1); auto it = PairMap[Idx].find({Op0, Op1}); if (it != PairMap[Idx].end()) { // Functions like BreakUpSubtract() can erase the Values we're using // as keys and create new Values after we built the PairMap. There's a // small chance that the new nodes can have the same address as // something already in the table. We shouldn't accumulate the stored // score in that case as it refers to the wrong Value. if (it->second.isValid()) Score += it->second.Score; } unsigned MaxRank = std::max(Ops[i].Rank, Ops[j].Rank); if (Score > Max || (Score == Max && MaxRank < BestRank)) { BestPair = {i, j}; Max = Score; BestRank = MaxRank; } } if (Max > 1) { auto Op0 = Ops[BestPair.first]; auto Op1 = Ops[BestPair.second]; Ops.erase(&Ops[BestPair.second]); Ops.erase(&Ops[BestPair.first]); Ops.push_back(Op0); Ops.push_back(Op1); } } // Now that we ordered and optimized the expressions, splat them back into // the expression tree, removing any unneeded nodes. RewriteExprTree(I, Ops); } void ReassociatePass::BuildPairMap(ReversePostOrderTraversal &RPOT) { // Make a "pairmap" of how often each operand pair occurs. for (BasicBlock *BI : RPOT) { for (Instruction &I : *BI) { if (!I.isAssociative()) continue; // Ignore nodes that aren't at the root of trees. if (I.hasOneUse() && I.user_back()->getOpcode() == I.getOpcode()) continue; // Collect all operands in a single reassociable expression. // Since Reassociate has already been run once, we can assume things // are already canonical according to Reassociation's regime. SmallVector Worklist = { I.getOperand(0), I.getOperand(1) }; SmallVector Ops; while (!Worklist.empty() && Ops.size() <= GlobalReassociateLimit) { Value *Op = Worklist.pop_back_val(); Instruction *OpI = dyn_cast(Op); if (!OpI || OpI->getOpcode() != I.getOpcode() || !OpI->hasOneUse()) { Ops.push_back(Op); continue; } // Be paranoid about self-referencing expressions in unreachable code. if (OpI->getOperand(0) != OpI) Worklist.push_back(OpI->getOperand(0)); if (OpI->getOperand(1) != OpI) Worklist.push_back(OpI->getOperand(1)); } // Skip extremely long expressions. if (Ops.size() > GlobalReassociateLimit) continue; // Add all pairwise combinations of operands to the pair map. unsigned BinaryIdx = I.getOpcode() - Instruction::BinaryOpsBegin; SmallSet, 32> Visited; for (unsigned i = 0; i < Ops.size() - 1; ++i) { for (unsigned j = i + 1; j < Ops.size(); ++j) { // Canonicalize operand orderings. Value *Op0 = Ops[i]; Value *Op1 = Ops[j]; if (std::less()(Op1, Op0)) std::swap(Op0, Op1); if (!Visited.insert({Op0, Op1}).second) continue; auto res = PairMap[BinaryIdx].insert({{Op0, Op1}, {Op0, Op1, 1}}); if (!res.second) { // If either key value has been erased then we've got the same // address by coincidence. That can't happen here because nothing is // erasing values but it can happen by the time we're querying the // map. assert(res.first->second.isValid() && "WeakVH invalidated"); ++res.first->second.Score; } } } } } } PreservedAnalyses ReassociatePass::run(Function &F, FunctionAnalysisManager &) { // Get the functions basic blocks in Reverse Post Order. This order is used by // BuildRankMap to pre calculate ranks correctly. It also excludes dead basic // blocks (it has been seen that the analysis in this pass could hang when // analysing dead basic blocks). ReversePostOrderTraversal RPOT(&F); // Calculate the rank map for F. BuildRankMap(F, RPOT); // Build the pair map before running reassociate. // Technically this would be more accurate if we did it after one round // of reassociation, but in practice it doesn't seem to help much on // real-world code, so don't waste the compile time running reassociate // twice. // If a user wants, they could expicitly run reassociate twice in their // pass pipeline for further potential gains. // It might also be possible to update the pair map during runtime, but the // overhead of that may be large if there's many reassociable chains. BuildPairMap(RPOT); MadeChange = false; // Traverse the same blocks that were analysed by BuildRankMap. for (BasicBlock *BI : RPOT) { assert(RankMap.count(&*BI) && "BB should be ranked."); // Optimize every instruction in the basic block. for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;) if (isInstructionTriviallyDead(&*II)) { EraseInst(&*II++); } else { OptimizeInst(&*II); assert(II->getParent() == &*BI && "Moved to a different block!"); ++II; } // Make a copy of all the instructions to be redone so we can remove dead // instructions. OrderedSet ToRedo(RedoInsts); // Iterate over all instructions to be reevaluated and remove trivially dead // instructions. If any operand of the trivially dead instruction becomes // dead mark it for deletion as well. Continue this process until all // trivially dead instructions have been removed. while (!ToRedo.empty()) { Instruction *I = ToRedo.pop_back_val(); if (isInstructionTriviallyDead(I)) { RecursivelyEraseDeadInsts(I, ToRedo); MadeChange = true; } } // Now that we have removed dead instructions, we can reoptimize the // remaining instructions. while (!RedoInsts.empty()) { Instruction *I = RedoInsts.front(); RedoInsts.erase(RedoInsts.begin()); if (isInstructionTriviallyDead(I)) EraseInst(I); else OptimizeInst(I); } } // We are done with the rank map and pair map. RankMap.clear(); ValueRankMap.clear(); for (auto &Entry : PairMap) Entry.clear(); if (MadeChange) { PreservedAnalyses PA; PA.preserveSet(); PA.preserve(); return PA; } return PreservedAnalyses::all(); } namespace { class ReassociateLegacyPass : public FunctionPass { ReassociatePass Impl; public: static char ID; // Pass identification, replacement for typeid ReassociateLegacyPass() : FunctionPass(ID) { initializeReassociateLegacyPassPass(*PassRegistry::getPassRegistry()); } bool runOnFunction(Function &F) override { if (skipFunction(F)) return false; FunctionAnalysisManager DummyFAM; auto PA = Impl.run(F, DummyFAM); return !PA.areAllPreserved(); } void getAnalysisUsage(AnalysisUsage &AU) const override { AU.setPreservesCFG(); AU.addPreserved(); } }; } // end anonymous namespace char ReassociateLegacyPass::ID = 0; INITIALIZE_PASS(ReassociateLegacyPass, "reassociate", "Reassociate expressions", false, false) // Public interface to the Reassociate pass FunctionPass *llvm::createReassociatePass() { return new ReassociateLegacyPass(); }