//===-- EarlyIfConversion.cpp - If-conversion on SSA form machine code ----===// // // 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 // //===----------------------------------------------------------------------===// // // Early if-conversion is for out-of-order CPUs that don't have a lot of // predicable instructions. The goal is to eliminate conditional branches that // may mispredict. // // Instructions from both sides of the branch are executed specutatively, and a // cmov instruction selects the result. // //===----------------------------------------------------------------------===// #include "llvm/ADT/BitVector.h" #include "llvm/ADT/PostOrderIterator.h" #include "llvm/ADT/SetVector.h" #include "llvm/ADT/SmallPtrSet.h" #include "llvm/ADT/SparseSet.h" #include "llvm/ADT/Statistic.h" #include "llvm/CodeGen/MachineBranchProbabilityInfo.h" #include "llvm/CodeGen/MachineDominators.h" #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/MachineFunctionPass.h" #include "llvm/CodeGen/MachineInstr.h" #include "llvm/CodeGen/MachineLoopInfo.h" #include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h" #include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/CodeGen/MachineTraceMetrics.h" #include "llvm/CodeGen/Passes.h" #include "llvm/CodeGen/TargetInstrInfo.h" #include "llvm/CodeGen/TargetRegisterInfo.h" #include "llvm/CodeGen/TargetSubtargetInfo.h" #include "llvm/InitializePasses.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" using namespace llvm; #define DEBUG_TYPE "early-ifcvt" // Absolute maximum number of instructions allowed per speculated block. // This bypasses all other heuristics, so it should be set fairly high. static cl::opt BlockInstrLimit("early-ifcvt-limit", cl::init(30), cl::Hidden, cl::desc("Maximum number of instructions per speculated block.")); // Stress testing mode - disable heuristics. static cl::opt Stress("stress-early-ifcvt", cl::Hidden, cl::desc("Turn all knobs to 11")); STATISTIC(NumDiamondsSeen, "Number of diamonds"); STATISTIC(NumDiamondsConv, "Number of diamonds converted"); STATISTIC(NumTrianglesSeen, "Number of triangles"); STATISTIC(NumTrianglesConv, "Number of triangles converted"); //===----------------------------------------------------------------------===// // SSAIfConv //===----------------------------------------------------------------------===// // // The SSAIfConv class performs if-conversion on SSA form machine code after // determining if it is possible. The class contains no heuristics; external // code should be used to determine when if-conversion is a good idea. // // SSAIfConv can convert both triangles and diamonds: // // Triangle: Head Diamond: Head // | \ / \_ // | \ / | // | [TF]BB FBB TBB // | / \ / // | / \ / // Tail Tail // // Instructions in the conditional blocks TBB and/or FBB are spliced into the // Head block, and phis in the Tail block are converted to select instructions. // namespace { class SSAIfConv { const TargetInstrInfo *TII; const TargetRegisterInfo *TRI; MachineRegisterInfo *MRI; public: /// The block containing the conditional branch. MachineBasicBlock *Head; /// The block containing phis after the if-then-else. MachineBasicBlock *Tail; /// The 'true' conditional block as determined by analyzeBranch. MachineBasicBlock *TBB; /// The 'false' conditional block as determined by analyzeBranch. MachineBasicBlock *FBB; /// isTriangle - When there is no 'else' block, either TBB or FBB will be /// equal to Tail. bool isTriangle() const { return TBB == Tail || FBB == Tail; } /// Returns the Tail predecessor for the True side. MachineBasicBlock *getTPred() const { return TBB == Tail ? Head : TBB; } /// Returns the Tail predecessor for the False side. MachineBasicBlock *getFPred() const { return FBB == Tail ? Head : FBB; } /// Information about each phi in the Tail block. struct PHIInfo { MachineInstr *PHI; unsigned TReg, FReg; // Latencies from Cond+Branch, TReg, and FReg to DstReg. int CondCycles, TCycles, FCycles; PHIInfo(MachineInstr *phi) : PHI(phi), TReg(0), FReg(0), CondCycles(0), TCycles(0), FCycles(0) {} }; SmallVector PHIs; private: /// The branch condition determined by analyzeBranch. SmallVector Cond; /// Instructions in Head that define values used by the conditional blocks. /// The hoisted instructions must be inserted after these instructions. SmallPtrSet InsertAfter; /// Register units clobbered by the conditional blocks. BitVector ClobberedRegUnits; // Scratch pad for findInsertionPoint. SparseSet LiveRegUnits; /// Insertion point in Head for speculatively executed instructions form TBB /// and FBB. MachineBasicBlock::iterator InsertionPoint; /// Return true if all non-terminator instructions in MBB can be safely /// speculated. bool canSpeculateInstrs(MachineBasicBlock *MBB); /// Return true if all non-terminator instructions in MBB can be safely /// predicated. bool canPredicateInstrs(MachineBasicBlock *MBB); /// Scan through instruction dependencies and update InsertAfter array. /// Return false if any dependency is incompatible with if conversion. bool InstrDependenciesAllowIfConv(MachineInstr *I); /// Predicate all instructions of the basic block with current condition /// except for terminators. Reverse the condition if ReversePredicate is set. void PredicateBlock(MachineBasicBlock *MBB, bool ReversePredicate); /// Find a valid insertion point in Head. bool findInsertionPoint(); /// Replace PHI instructions in Tail with selects. void replacePHIInstrs(); /// Insert selects and rewrite PHI operands to use them. void rewritePHIOperands(); public: /// runOnMachineFunction - Initialize per-function data structures. void runOnMachineFunction(MachineFunction &MF) { TII = MF.getSubtarget().getInstrInfo(); TRI = MF.getSubtarget().getRegisterInfo(); MRI = &MF.getRegInfo(); LiveRegUnits.clear(); LiveRegUnits.setUniverse(TRI->getNumRegUnits()); ClobberedRegUnits.clear(); ClobberedRegUnits.resize(TRI->getNumRegUnits()); } /// canConvertIf - If the sub-CFG headed by MBB can be if-converted, /// initialize the internal state, and return true. /// If predicate is set try to predicate the block otherwise try to /// speculatively execute it. bool canConvertIf(MachineBasicBlock *MBB, bool Predicate = false); /// convertIf - If-convert the last block passed to canConvertIf(), assuming /// it is possible. Add any erased blocks to RemovedBlocks. void convertIf(SmallVectorImpl &RemovedBlocks, bool Predicate = false); }; } // end anonymous namespace /// canSpeculateInstrs - Returns true if all the instructions in MBB can safely /// be speculated. The terminators are not considered. /// /// If instructions use any values that are defined in the head basic block, /// the defining instructions are added to InsertAfter. /// /// Any clobbered regunits are added to ClobberedRegUnits. /// bool SSAIfConv::canSpeculateInstrs(MachineBasicBlock *MBB) { // Reject any live-in physregs. It's probably CPSR/EFLAGS, and very hard to // get right. if (!MBB->livein_empty()) { LLVM_DEBUG(dbgs() << printMBBReference(*MBB) << " has live-ins.\n"); return false; } unsigned InstrCount = 0; // Check all instructions, except the terminators. It is assumed that // terminators never have side effects or define any used register values. for (MachineInstr &MI : llvm::make_range(MBB->begin(), MBB->getFirstTerminator())) { if (MI.isDebugInstr()) continue; if (++InstrCount > BlockInstrLimit && !Stress) { LLVM_DEBUG(dbgs() << printMBBReference(*MBB) << " has more than " << BlockInstrLimit << " instructions.\n"); return false; } // There shouldn't normally be any phis in a single-predecessor block. if (MI.isPHI()) { LLVM_DEBUG(dbgs() << "Can't hoist: " << MI); return false; } // Don't speculate loads. Note that it may be possible and desirable to // speculate GOT or constant pool loads that are guaranteed not to trap, // but we don't support that for now. if (MI.mayLoad()) { LLVM_DEBUG(dbgs() << "Won't speculate load: " << MI); return false; } // We never speculate stores, so an AA pointer isn't necessary. bool DontMoveAcrossStore = true; if (!MI.isSafeToMove(nullptr, DontMoveAcrossStore)) { LLVM_DEBUG(dbgs() << "Can't speculate: " << MI); return false; } // Check for any dependencies on Head instructions. if (!InstrDependenciesAllowIfConv(&MI)) return false; } return true; } /// Check that there is no dependencies preventing if conversion. /// /// If instruction uses any values that are defined in the head basic block, /// the defining instructions are added to InsertAfter. bool SSAIfConv::InstrDependenciesAllowIfConv(MachineInstr *I) { for (const MachineOperand &MO : I->operands()) { if (MO.isRegMask()) { LLVM_DEBUG(dbgs() << "Won't speculate regmask: " << *I); return false; } if (!MO.isReg()) continue; Register Reg = MO.getReg(); // Remember clobbered regunits. if (MO.isDef() && Register::isPhysicalRegister(Reg)) for (MCRegUnitIterator Units(Reg.asMCReg(), TRI); Units.isValid(); ++Units) ClobberedRegUnits.set(*Units); if (!MO.readsReg() || !Register::isVirtualRegister(Reg)) continue; MachineInstr *DefMI = MRI->getVRegDef(Reg); if (!DefMI || DefMI->getParent() != Head) continue; if (InsertAfter.insert(DefMI).second) LLVM_DEBUG(dbgs() << printMBBReference(*I->getParent()) << " depends on " << *DefMI); if (DefMI->isTerminator()) { LLVM_DEBUG(dbgs() << "Can't insert instructions below terminator.\n"); return false; } } return true; } /// canPredicateInstrs - Returns true if all the instructions in MBB can safely /// be predicates. The terminators are not considered. /// /// If instructions use any values that are defined in the head basic block, /// the defining instructions are added to InsertAfter. /// /// Any clobbered regunits are added to ClobberedRegUnits. /// bool SSAIfConv::canPredicateInstrs(MachineBasicBlock *MBB) { // Reject any live-in physregs. It's probably CPSR/EFLAGS, and very hard to // get right. if (!MBB->livein_empty()) { LLVM_DEBUG(dbgs() << printMBBReference(*MBB) << " has live-ins.\n"); return false; } unsigned InstrCount = 0; // Check all instructions, except the terminators. It is assumed that // terminators never have side effects or define any used register values. for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->getFirstTerminator(); I != E; ++I) { if (I->isDebugInstr()) continue; if (++InstrCount > BlockInstrLimit && !Stress) { LLVM_DEBUG(dbgs() << printMBBReference(*MBB) << " has more than " << BlockInstrLimit << " instructions.\n"); return false; } // There shouldn't normally be any phis in a single-predecessor block. if (I->isPHI()) { LLVM_DEBUG(dbgs() << "Can't predicate: " << *I); return false; } // Check that instruction is predicable and that it is not already // predicated. if (!TII->isPredicable(*I) || TII->isPredicated(*I)) { return false; } // Check for any dependencies on Head instructions. if (!InstrDependenciesAllowIfConv(&(*I))) return false; } return true; } // Apply predicate to all instructions in the machine block. void SSAIfConv::PredicateBlock(MachineBasicBlock *MBB, bool ReversePredicate) { auto Condition = Cond; if (ReversePredicate) TII->reverseBranchCondition(Condition); // Terminators don't need to be predicated as they will be removed. for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->getFirstTerminator(); I != E; ++I) { if (I->isDebugInstr()) continue; TII->PredicateInstruction(*I, Condition); } } /// Find an insertion point in Head for the speculated instructions. The /// insertion point must be: /// /// 1. Before any terminators. /// 2. After any instructions in InsertAfter. /// 3. Not have any clobbered regunits live. /// /// This function sets InsertionPoint and returns true when successful, it /// returns false if no valid insertion point could be found. /// bool SSAIfConv::findInsertionPoint() { // Keep track of live regunits before the current position. // Only track RegUnits that are also in ClobberedRegUnits. LiveRegUnits.clear(); SmallVector Reads; MachineBasicBlock::iterator FirstTerm = Head->getFirstTerminator(); MachineBasicBlock::iterator I = Head->end(); MachineBasicBlock::iterator B = Head->begin(); while (I != B) { --I; // Some of the conditional code depends in I. if (InsertAfter.count(&*I)) { LLVM_DEBUG(dbgs() << "Can't insert code after " << *I); return false; } // Update live regunits. for (const MachineOperand &MO : I->operands()) { // We're ignoring regmask operands. That is conservatively correct. if (!MO.isReg()) continue; Register Reg = MO.getReg(); if (!Register::isPhysicalRegister(Reg)) continue; // I clobbers Reg, so it isn't live before I. if (MO.isDef()) for (MCRegUnitIterator Units(Reg.asMCReg(), TRI); Units.isValid(); ++Units) LiveRegUnits.erase(*Units); // Unless I reads Reg. if (MO.readsReg()) Reads.push_back(Reg.asMCReg()); } // Anything read by I is live before I. while (!Reads.empty()) for (MCRegUnitIterator Units(Reads.pop_back_val(), TRI); Units.isValid(); ++Units) if (ClobberedRegUnits.test(*Units)) LiveRegUnits.insert(*Units); // We can't insert before a terminator. if (I != FirstTerm && I->isTerminator()) continue; // Some of the clobbered registers are live before I, not a valid insertion // point. if (!LiveRegUnits.empty()) { LLVM_DEBUG({ dbgs() << "Would clobber"; for (unsigned LRU : LiveRegUnits) dbgs() << ' ' << printRegUnit(LRU, TRI); dbgs() << " live before " << *I; }); continue; } // This is a valid insertion point. InsertionPoint = I; LLVM_DEBUG(dbgs() << "Can insert before " << *I); return true; } LLVM_DEBUG(dbgs() << "No legal insertion point found.\n"); return false; } /// canConvertIf - analyze the sub-cfg rooted in MBB, and return true if it is /// a potential candidate for if-conversion. Fill out the internal state. /// bool SSAIfConv::canConvertIf(MachineBasicBlock *MBB, bool Predicate) { Head = MBB; TBB = FBB = Tail = nullptr; if (Head->succ_size() != 2) return false; MachineBasicBlock *Succ0 = Head->succ_begin()[0]; MachineBasicBlock *Succ1 = Head->succ_begin()[1]; // Canonicalize so Succ0 has MBB as its single predecessor. if (Succ0->pred_size() != 1) std::swap(Succ0, Succ1); if (Succ0->pred_size() != 1 || Succ0->succ_size() != 1) return false; Tail = Succ0->succ_begin()[0]; // This is not a triangle. if (Tail != Succ1) { // Check for a diamond. We won't deal with any critical edges. if (Succ1->pred_size() != 1 || Succ1->succ_size() != 1 || Succ1->succ_begin()[0] != Tail) return false; LLVM_DEBUG(dbgs() << "\nDiamond: " << printMBBReference(*Head) << " -> " << printMBBReference(*Succ0) << "/" << printMBBReference(*Succ1) << " -> " << printMBBReference(*Tail) << '\n'); // Live-in physregs are tricky to get right when speculating code. if (!Tail->livein_empty()) { LLVM_DEBUG(dbgs() << "Tail has live-ins.\n"); return false; } } else { LLVM_DEBUG(dbgs() << "\nTriangle: " << printMBBReference(*Head) << " -> " << printMBBReference(*Succ0) << " -> " << printMBBReference(*Tail) << '\n'); } // This is a triangle or a diamond. // Skip if we cannot predicate and there are no phis skip as there must be // side effects that can only be handled with predication. if (!Predicate && (Tail->empty() || !Tail->front().isPHI())) { LLVM_DEBUG(dbgs() << "No phis in tail.\n"); return false; } // The branch we're looking to eliminate must be analyzable. Cond.clear(); if (TII->analyzeBranch(*Head, TBB, FBB, Cond)) { LLVM_DEBUG(dbgs() << "Branch not analyzable.\n"); return false; } // This is weird, probably some sort of degenerate CFG. if (!TBB) { LLVM_DEBUG(dbgs() << "analyzeBranch didn't find conditional branch.\n"); return false; } // Make sure the analyzed branch is conditional; one of the successors // could be a landing pad. (Empty landing pads can be generated on Windows.) if (Cond.empty()) { LLVM_DEBUG(dbgs() << "analyzeBranch found an unconditional branch.\n"); return false; } // analyzeBranch doesn't set FBB on a fall-through branch. // Make sure it is always set. FBB = TBB == Succ0 ? Succ1 : Succ0; // Any phis in the tail block must be convertible to selects. PHIs.clear(); MachineBasicBlock *TPred = getTPred(); MachineBasicBlock *FPred = getFPred(); for (MachineBasicBlock::iterator I = Tail->begin(), E = Tail->end(); I != E && I->isPHI(); ++I) { PHIs.push_back(&*I); PHIInfo &PI = PHIs.back(); // Find PHI operands corresponding to TPred and FPred. for (unsigned i = 1; i != PI.PHI->getNumOperands(); i += 2) { if (PI.PHI->getOperand(i+1).getMBB() == TPred) PI.TReg = PI.PHI->getOperand(i).getReg(); if (PI.PHI->getOperand(i+1).getMBB() == FPred) PI.FReg = PI.PHI->getOperand(i).getReg(); } assert(Register::isVirtualRegister(PI.TReg) && "Bad PHI"); assert(Register::isVirtualRegister(PI.FReg) && "Bad PHI"); // Get target information. if (!TII->canInsertSelect(*Head, Cond, PI.PHI->getOperand(0).getReg(), PI.TReg, PI.FReg, PI.CondCycles, PI.TCycles, PI.FCycles)) { LLVM_DEBUG(dbgs() << "Can't convert: " << *PI.PHI); return false; } } // Check that the conditional instructions can be speculated. InsertAfter.clear(); ClobberedRegUnits.reset(); if (Predicate) { if (TBB != Tail && !canPredicateInstrs(TBB)) return false; if (FBB != Tail && !canPredicateInstrs(FBB)) return false; } else { if (TBB != Tail && !canSpeculateInstrs(TBB)) return false; if (FBB != Tail && !canSpeculateInstrs(FBB)) return false; } // Try to find a valid insertion point for the speculated instructions in the // head basic block. if (!findInsertionPoint()) return false; if (isTriangle()) ++NumTrianglesSeen; else ++NumDiamondsSeen; return true; } /// \return true iff the two registers are known to have the same value. static bool hasSameValue(const MachineRegisterInfo &MRI, const TargetInstrInfo *TII, Register TReg, Register FReg) { if (TReg == FReg) return true; if (!TReg.isVirtual() || !FReg.isVirtual()) return false; const MachineInstr *TDef = MRI.getUniqueVRegDef(TReg); const MachineInstr *FDef = MRI.getUniqueVRegDef(FReg); if (!TDef || !FDef) return false; // If there are side-effects, all bets are off. if (TDef->hasUnmodeledSideEffects()) return false; // If the instruction could modify memory, or there may be some intervening // store between the two, we can't consider them to be equal. if (TDef->mayLoadOrStore() && !TDef->isDereferenceableInvariantLoad(nullptr)) return false; // We also can't guarantee that they are the same if, for example, the // instructions are both a copy from a physical reg, because some other // instruction may have modified the value in that reg between the two // defining insts. if (any_of(TDef->uses(), [](const MachineOperand &MO) { return MO.isReg() && MO.getReg().isPhysical(); })) return false; // Check whether the two defining instructions produce the same value(s). if (!TII->produceSameValue(*TDef, *FDef, &MRI)) return false; // Further, check that the two defs come from corresponding operands. int TIdx = TDef->findRegisterDefOperandIdx(TReg); int FIdx = FDef->findRegisterDefOperandIdx(FReg); if (TIdx == -1 || FIdx == -1) return false; return TIdx == FIdx; } /// replacePHIInstrs - Completely replace PHI instructions with selects. /// This is possible when the only Tail predecessors are the if-converted /// blocks. void SSAIfConv::replacePHIInstrs() { assert(Tail->pred_size() == 2 && "Cannot replace PHIs"); MachineBasicBlock::iterator FirstTerm = Head->getFirstTerminator(); assert(FirstTerm != Head->end() && "No terminators"); DebugLoc HeadDL = FirstTerm->getDebugLoc(); // Convert all PHIs to select instructions inserted before FirstTerm. for (unsigned i = 0, e = PHIs.size(); i != e; ++i) { PHIInfo &PI = PHIs[i]; LLVM_DEBUG(dbgs() << "If-converting " << *PI.PHI); Register DstReg = PI.PHI->getOperand(0).getReg(); if (hasSameValue(*MRI, TII, PI.TReg, PI.FReg)) { // We do not need the select instruction if both incoming values are // equal, but we do need a COPY. BuildMI(*Head, FirstTerm, HeadDL, TII->get(TargetOpcode::COPY), DstReg) .addReg(PI.TReg); } else { TII->insertSelect(*Head, FirstTerm, HeadDL, DstReg, Cond, PI.TReg, PI.FReg); } LLVM_DEBUG(dbgs() << " --> " << *std::prev(FirstTerm)); PI.PHI->eraseFromParent(); PI.PHI = nullptr; } } /// rewritePHIOperands - When there are additional Tail predecessors, insert /// select instructions in Head and rewrite PHI operands to use the selects. /// Keep the PHI instructions in Tail to handle the other predecessors. void SSAIfConv::rewritePHIOperands() { MachineBasicBlock::iterator FirstTerm = Head->getFirstTerminator(); assert(FirstTerm != Head->end() && "No terminators"); DebugLoc HeadDL = FirstTerm->getDebugLoc(); // Convert all PHIs to select instructions inserted before FirstTerm. for (unsigned i = 0, e = PHIs.size(); i != e; ++i) { PHIInfo &PI = PHIs[i]; unsigned DstReg = 0; LLVM_DEBUG(dbgs() << "If-converting " << *PI.PHI); if (hasSameValue(*MRI, TII, PI.TReg, PI.FReg)) { // We do not need the select instruction if both incoming values are // equal. DstReg = PI.TReg; } else { Register PHIDst = PI.PHI->getOperand(0).getReg(); DstReg = MRI->createVirtualRegister(MRI->getRegClass(PHIDst)); TII->insertSelect(*Head, FirstTerm, HeadDL, DstReg, Cond, PI.TReg, PI.FReg); LLVM_DEBUG(dbgs() << " --> " << *std::prev(FirstTerm)); } // Rewrite PHI operands TPred -> (DstReg, Head), remove FPred. for (unsigned i = PI.PHI->getNumOperands(); i != 1; i -= 2) { MachineBasicBlock *MBB = PI.PHI->getOperand(i-1).getMBB(); if (MBB == getTPred()) { PI.PHI->getOperand(i-1).setMBB(Head); PI.PHI->getOperand(i-2).setReg(DstReg); } else if (MBB == getFPred()) { PI.PHI->RemoveOperand(i-1); PI.PHI->RemoveOperand(i-2); } } LLVM_DEBUG(dbgs() << " --> " << *PI.PHI); } } /// convertIf - Execute the if conversion after canConvertIf has determined the /// feasibility. /// /// Any basic blocks erased will be added to RemovedBlocks. /// void SSAIfConv::convertIf(SmallVectorImpl &RemovedBlocks, bool Predicate) { assert(Head && Tail && TBB && FBB && "Call canConvertIf first."); // Update statistics. if (isTriangle()) ++NumTrianglesConv; else ++NumDiamondsConv; // Move all instructions into Head, except for the terminators. if (TBB != Tail) { if (Predicate) PredicateBlock(TBB, /*ReversePredicate=*/false); Head->splice(InsertionPoint, TBB, TBB->begin(), TBB->getFirstTerminator()); } if (FBB != Tail) { if (Predicate) PredicateBlock(FBB, /*ReversePredicate=*/true); Head->splice(InsertionPoint, FBB, FBB->begin(), FBB->getFirstTerminator()); } // Are there extra Tail predecessors? bool ExtraPreds = Tail->pred_size() != 2; if (ExtraPreds) rewritePHIOperands(); else replacePHIInstrs(); // Fix up the CFG, temporarily leave Head without any successors. Head->removeSuccessor(TBB); Head->removeSuccessor(FBB, true); if (TBB != Tail) TBB->removeSuccessor(Tail, true); if (FBB != Tail) FBB->removeSuccessor(Tail, true); // Fix up Head's terminators. // It should become a single branch or a fallthrough. DebugLoc HeadDL = Head->getFirstTerminator()->getDebugLoc(); TII->removeBranch(*Head); // Erase the now empty conditional blocks. It is likely that Head can fall // through to Tail, and we can join the two blocks. if (TBB != Tail) { RemovedBlocks.push_back(TBB); TBB->eraseFromParent(); } if (FBB != Tail) { RemovedBlocks.push_back(FBB); FBB->eraseFromParent(); } assert(Head->succ_empty() && "Additional head successors?"); if (!ExtraPreds && Head->isLayoutSuccessor(Tail)) { // Splice Tail onto the end of Head. LLVM_DEBUG(dbgs() << "Joining tail " << printMBBReference(*Tail) << " into head " << printMBBReference(*Head) << '\n'); Head->splice(Head->end(), Tail, Tail->begin(), Tail->end()); Head->transferSuccessorsAndUpdatePHIs(Tail); RemovedBlocks.push_back(Tail); Tail->eraseFromParent(); } else { // We need a branch to Tail, let code placement work it out later. LLVM_DEBUG(dbgs() << "Converting to unconditional branch.\n"); SmallVector EmptyCond; TII->insertBranch(*Head, Tail, nullptr, EmptyCond, HeadDL); Head->addSuccessor(Tail); } LLVM_DEBUG(dbgs() << *Head); } //===----------------------------------------------------------------------===// // EarlyIfConverter Pass //===----------------------------------------------------------------------===// namespace { class EarlyIfConverter : public MachineFunctionPass { const TargetInstrInfo *TII; const TargetRegisterInfo *TRI; MCSchedModel SchedModel; MachineRegisterInfo *MRI; MachineDominatorTree *DomTree; MachineLoopInfo *Loops; MachineTraceMetrics *Traces; MachineTraceMetrics::Ensemble *MinInstr; SSAIfConv IfConv; public: static char ID; EarlyIfConverter() : MachineFunctionPass(ID) {} void getAnalysisUsage(AnalysisUsage &AU) const override; bool runOnMachineFunction(MachineFunction &MF) override; StringRef getPassName() const override { return "Early If-Conversion"; } private: bool tryConvertIf(MachineBasicBlock*); void invalidateTraces(); bool shouldConvertIf(); }; } // end anonymous namespace char EarlyIfConverter::ID = 0; char &llvm::EarlyIfConverterID = EarlyIfConverter::ID; INITIALIZE_PASS_BEGIN(EarlyIfConverter, DEBUG_TYPE, "Early If Converter", false, false) INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo) INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) INITIALIZE_PASS_DEPENDENCY(MachineTraceMetrics) INITIALIZE_PASS_END(EarlyIfConverter, DEBUG_TYPE, "Early If Converter", false, false) void EarlyIfConverter::getAnalysisUsage(AnalysisUsage &AU) const { AU.addRequired(); AU.addRequired(); AU.addPreserved(); AU.addRequired(); AU.addPreserved(); AU.addRequired(); AU.addPreserved(); MachineFunctionPass::getAnalysisUsage(AU); } namespace { /// Update the dominator tree after if-conversion erased some blocks. void updateDomTree(MachineDominatorTree *DomTree, const SSAIfConv &IfConv, ArrayRef Removed) { // convertIf can remove TBB, FBB, and Tail can be merged into Head. // TBB and FBB should not dominate any blocks. // Tail children should be transferred to Head. MachineDomTreeNode *HeadNode = DomTree->getNode(IfConv.Head); for (auto B : Removed) { MachineDomTreeNode *Node = DomTree->getNode(B); assert(Node != HeadNode && "Cannot erase the head node"); while (Node->getNumChildren()) { assert(Node->getBlock() == IfConv.Tail && "Unexpected children"); DomTree->changeImmediateDominator(Node->back(), HeadNode); } DomTree->eraseNode(B); } } /// Update LoopInfo after if-conversion. void updateLoops(MachineLoopInfo *Loops, ArrayRef Removed) { if (!Loops) return; // If-conversion doesn't change loop structure, and it doesn't mess with back // edges, so updating LoopInfo is simply removing the dead blocks. for (auto B : Removed) Loops->removeBlock(B); } } // namespace /// Invalidate MachineTraceMetrics before if-conversion. void EarlyIfConverter::invalidateTraces() { Traces->verifyAnalysis(); Traces->invalidate(IfConv.Head); Traces->invalidate(IfConv.Tail); Traces->invalidate(IfConv.TBB); Traces->invalidate(IfConv.FBB); Traces->verifyAnalysis(); } // Adjust cycles with downward saturation. static unsigned adjCycles(unsigned Cyc, int Delta) { if (Delta < 0 && Cyc + Delta > Cyc) return 0; return Cyc + Delta; } namespace { /// Helper class to simplify emission of cycle counts into optimization remarks. struct Cycles { const char *Key; unsigned Value; }; template Remark &operator<<(Remark &R, Cycles C) { return R << ore::NV(C.Key, C.Value) << (C.Value == 1 ? " cycle" : " cycles"); } } // anonymous namespace /// Apply cost model and heuristics to the if-conversion in IfConv. /// Return true if the conversion is a good idea. /// bool EarlyIfConverter::shouldConvertIf() { // Stress testing mode disables all cost considerations. if (Stress) return true; if (!MinInstr) MinInstr = Traces->getEnsemble(MachineTraceMetrics::TS_MinInstrCount); MachineTraceMetrics::Trace TBBTrace = MinInstr->getTrace(IfConv.getTPred()); MachineTraceMetrics::Trace FBBTrace = MinInstr->getTrace(IfConv.getFPred()); LLVM_DEBUG(dbgs() << "TBB: " << TBBTrace << "FBB: " << FBBTrace); unsigned MinCrit = std::min(TBBTrace.getCriticalPath(), FBBTrace.getCriticalPath()); // Set a somewhat arbitrary limit on the critical path extension we accept. unsigned CritLimit = SchedModel.MispredictPenalty/2; MachineBasicBlock &MBB = *IfConv.Head; MachineOptimizationRemarkEmitter MORE(*MBB.getParent(), nullptr); // If-conversion only makes sense when there is unexploited ILP. Compute the // maximum-ILP resource length of the trace after if-conversion. Compare it // to the shortest critical path. SmallVector ExtraBlocks; if (IfConv.TBB != IfConv.Tail) ExtraBlocks.push_back(IfConv.TBB); unsigned ResLength = FBBTrace.getResourceLength(ExtraBlocks); LLVM_DEBUG(dbgs() << "Resource length " << ResLength << ", minimal critical path " << MinCrit << '\n'); if (ResLength > MinCrit + CritLimit) { LLVM_DEBUG(dbgs() << "Not enough available ILP.\n"); MORE.emit([&]() { MachineOptimizationRemarkMissed R(DEBUG_TYPE, "IfConversion", MBB.findDebugLoc(MBB.back()), &MBB); R << "did not if-convert branch: the resulting critical path (" << Cycles{"ResLength", ResLength} << ") would extend the shorter leg's critical path (" << Cycles{"MinCrit", MinCrit} << ") by more than the threshold of " << Cycles{"CritLimit", CritLimit} << ", which cannot be hidden by available ILP."; return R; }); return false; } // Assume that the depth of the first head terminator will also be the depth // of the select instruction inserted, as determined by the flag dependency. // TBB / FBB data dependencies may delay the select even more. MachineTraceMetrics::Trace HeadTrace = MinInstr->getTrace(IfConv.Head); unsigned BranchDepth = HeadTrace.getInstrCycles(*IfConv.Head->getFirstTerminator()).Depth; LLVM_DEBUG(dbgs() << "Branch depth: " << BranchDepth << '\n'); // Look at all the tail phis, and compute the critical path extension caused // by inserting select instructions. MachineTraceMetrics::Trace TailTrace = MinInstr->getTrace(IfConv.Tail); struct CriticalPathInfo { unsigned Extra; // Count of extra cycles that the component adds. unsigned Depth; // Absolute depth of the component in cycles. }; CriticalPathInfo Cond{}; CriticalPathInfo TBlock{}; CriticalPathInfo FBlock{}; bool ShouldConvert = true; for (unsigned i = 0, e = IfConv.PHIs.size(); i != e; ++i) { SSAIfConv::PHIInfo &PI = IfConv.PHIs[i]; unsigned Slack = TailTrace.getInstrSlack(*PI.PHI); unsigned MaxDepth = Slack + TailTrace.getInstrCycles(*PI.PHI).Depth; LLVM_DEBUG(dbgs() << "Slack " << Slack << ":\t" << *PI.PHI); // The condition is pulled into the critical path. unsigned CondDepth = adjCycles(BranchDepth, PI.CondCycles); if (CondDepth > MaxDepth) { unsigned Extra = CondDepth - MaxDepth; LLVM_DEBUG(dbgs() << "Condition adds " << Extra << " cycles.\n"); if (Extra > Cond.Extra) Cond = {Extra, CondDepth}; if (Extra > CritLimit) { LLVM_DEBUG(dbgs() << "Exceeds limit of " << CritLimit << '\n'); ShouldConvert = false; } } // The TBB value is pulled into the critical path. unsigned TDepth = adjCycles(TBBTrace.getPHIDepth(*PI.PHI), PI.TCycles); if (TDepth > MaxDepth) { unsigned Extra = TDepth - MaxDepth; LLVM_DEBUG(dbgs() << "TBB data adds " << Extra << " cycles.\n"); if (Extra > TBlock.Extra) TBlock = {Extra, TDepth}; if (Extra > CritLimit) { LLVM_DEBUG(dbgs() << "Exceeds limit of " << CritLimit << '\n'); ShouldConvert = false; } } // The FBB value is pulled into the critical path. unsigned FDepth = adjCycles(FBBTrace.getPHIDepth(*PI.PHI), PI.FCycles); if (FDepth > MaxDepth) { unsigned Extra = FDepth - MaxDepth; LLVM_DEBUG(dbgs() << "FBB data adds " << Extra << " cycles.\n"); if (Extra > FBlock.Extra) FBlock = {Extra, FDepth}; if (Extra > CritLimit) { LLVM_DEBUG(dbgs() << "Exceeds limit of " << CritLimit << '\n'); ShouldConvert = false; } } } // Organize by "short" and "long" legs, since the diagnostics get confusing // when referring to the "true" and "false" sides of the branch, given that // those don't always correlate with what the user wrote in source-terms. const CriticalPathInfo Short = TBlock.Extra > FBlock.Extra ? FBlock : TBlock; const CriticalPathInfo Long = TBlock.Extra > FBlock.Extra ? TBlock : FBlock; if (ShouldConvert) { MORE.emit([&]() { MachineOptimizationRemark R(DEBUG_TYPE, "IfConversion", MBB.back().getDebugLoc(), &MBB); R << "performing if-conversion on branch: the condition adds " << Cycles{"CondCycles", Cond.Extra} << " to the critical path"; if (Short.Extra > 0) R << ", and the short leg adds another " << Cycles{"ShortCycles", Short.Extra}; if (Long.Extra > 0) R << ", and the long leg adds another " << Cycles{"LongCycles", Long.Extra}; R << ", each staying under the threshold of " << Cycles{"CritLimit", CritLimit} << "."; return R; }); } else { MORE.emit([&]() { MachineOptimizationRemarkMissed R(DEBUG_TYPE, "IfConversion", MBB.back().getDebugLoc(), &MBB); R << "did not if-convert branch: the condition would add " << Cycles{"CondCycles", Cond.Extra} << " to the critical path"; if (Cond.Extra > CritLimit) R << " exceeding the limit of " << Cycles{"CritLimit", CritLimit}; if (Short.Extra > 0) { R << ", and the short leg would add another " << Cycles{"ShortCycles", Short.Extra}; if (Short.Extra > CritLimit) R << " exceeding the limit of " << Cycles{"CritLimit", CritLimit}; } if (Long.Extra > 0) { R << ", and the long leg would add another " << Cycles{"LongCycles", Long.Extra}; if (Long.Extra > CritLimit) R << " exceeding the limit of " << Cycles{"CritLimit", CritLimit}; } R << "."; return R; }); } return ShouldConvert; } /// Attempt repeated if-conversion on MBB, return true if successful. /// bool EarlyIfConverter::tryConvertIf(MachineBasicBlock *MBB) { bool Changed = false; while (IfConv.canConvertIf(MBB) && shouldConvertIf()) { // If-convert MBB and update analyses. invalidateTraces(); SmallVector RemovedBlocks; IfConv.convertIf(RemovedBlocks); Changed = true; updateDomTree(DomTree, IfConv, RemovedBlocks); updateLoops(Loops, RemovedBlocks); } return Changed; } bool EarlyIfConverter::runOnMachineFunction(MachineFunction &MF) { LLVM_DEBUG(dbgs() << "********** EARLY IF-CONVERSION **********\n" << "********** Function: " << MF.getName() << '\n'); if (skipFunction(MF.getFunction())) return false; // Only run if conversion if the target wants it. const TargetSubtargetInfo &STI = MF.getSubtarget(); if (!STI.enableEarlyIfConversion()) return false; TII = STI.getInstrInfo(); TRI = STI.getRegisterInfo(); SchedModel = STI.getSchedModel(); MRI = &MF.getRegInfo(); DomTree = &getAnalysis(); Loops = getAnalysisIfAvailable(); Traces = &getAnalysis(); MinInstr = nullptr; bool Changed = false; IfConv.runOnMachineFunction(MF); // Visit blocks in dominator tree post-order. The post-order enables nested // if-conversion in a single pass. The tryConvertIf() function may erase // blocks, but only blocks dominated by the head block. This makes it safe to // update the dominator tree while the post-order iterator is still active. for (auto DomNode : post_order(DomTree)) if (tryConvertIf(DomNode->getBlock())) Changed = true; return Changed; } //===----------------------------------------------------------------------===// // EarlyIfPredicator Pass //===----------------------------------------------------------------------===// namespace { class EarlyIfPredicator : public MachineFunctionPass { const TargetInstrInfo *TII; const TargetRegisterInfo *TRI; TargetSchedModel SchedModel; MachineRegisterInfo *MRI; MachineDominatorTree *DomTree; MachineBranchProbabilityInfo *MBPI; MachineLoopInfo *Loops; SSAIfConv IfConv; public: static char ID; EarlyIfPredicator() : MachineFunctionPass(ID) {} void getAnalysisUsage(AnalysisUsage &AU) const override; bool runOnMachineFunction(MachineFunction &MF) override; StringRef getPassName() const override { return "Early If-predicator"; } protected: bool tryConvertIf(MachineBasicBlock *); bool shouldConvertIf(); }; } // end anonymous namespace #undef DEBUG_TYPE #define DEBUG_TYPE "early-if-predicator" char EarlyIfPredicator::ID = 0; char &llvm::EarlyIfPredicatorID = EarlyIfPredicator::ID; INITIALIZE_PASS_BEGIN(EarlyIfPredicator, DEBUG_TYPE, "Early If Predicator", false, false) INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo) INITIALIZE_PASS_END(EarlyIfPredicator, DEBUG_TYPE, "Early If Predicator", false, false) void EarlyIfPredicator::getAnalysisUsage(AnalysisUsage &AU) const { AU.addRequired(); AU.addRequired(); AU.addPreserved(); AU.addRequired(); AU.addPreserved(); MachineFunctionPass::getAnalysisUsage(AU); } /// Apply the target heuristic to decide if the transformation is profitable. bool EarlyIfPredicator::shouldConvertIf() { auto TrueProbability = MBPI->getEdgeProbability(IfConv.Head, IfConv.TBB); if (IfConv.isTriangle()) { MachineBasicBlock &IfBlock = (IfConv.TBB == IfConv.Tail) ? *IfConv.FBB : *IfConv.TBB; unsigned ExtraPredCost = 0; unsigned Cycles = 0; for (MachineInstr &I : IfBlock) { unsigned NumCycles = SchedModel.computeInstrLatency(&I, false); if (NumCycles > 1) Cycles += NumCycles - 1; ExtraPredCost += TII->getPredicationCost(I); } return TII->isProfitableToIfCvt(IfBlock, Cycles, ExtraPredCost, TrueProbability); } unsigned TExtra = 0; unsigned FExtra = 0; unsigned TCycle = 0; unsigned FCycle = 0; for (MachineInstr &I : *IfConv.TBB) { unsigned NumCycles = SchedModel.computeInstrLatency(&I, false); if (NumCycles > 1) TCycle += NumCycles - 1; TExtra += TII->getPredicationCost(I); } for (MachineInstr &I : *IfConv.FBB) { unsigned NumCycles = SchedModel.computeInstrLatency(&I, false); if (NumCycles > 1) FCycle += NumCycles - 1; FExtra += TII->getPredicationCost(I); } return TII->isProfitableToIfCvt(*IfConv.TBB, TCycle, TExtra, *IfConv.FBB, FCycle, FExtra, TrueProbability); } /// Attempt repeated if-conversion on MBB, return true if successful. /// bool EarlyIfPredicator::tryConvertIf(MachineBasicBlock *MBB) { bool Changed = false; while (IfConv.canConvertIf(MBB, /*Predicate*/ true) && shouldConvertIf()) { // If-convert MBB and update analyses. SmallVector RemovedBlocks; IfConv.convertIf(RemovedBlocks, /*Predicate*/ true); Changed = true; updateDomTree(DomTree, IfConv, RemovedBlocks); updateLoops(Loops, RemovedBlocks); } return Changed; } bool EarlyIfPredicator::runOnMachineFunction(MachineFunction &MF) { LLVM_DEBUG(dbgs() << "********** EARLY IF-PREDICATOR **********\n" << "********** Function: " << MF.getName() << '\n'); if (skipFunction(MF.getFunction())) return false; const TargetSubtargetInfo &STI = MF.getSubtarget(); TII = STI.getInstrInfo(); TRI = STI.getRegisterInfo(); MRI = &MF.getRegInfo(); SchedModel.init(&STI); DomTree = &getAnalysis(); Loops = getAnalysisIfAvailable(); MBPI = &getAnalysis(); bool Changed = false; IfConv.runOnMachineFunction(MF); // Visit blocks in dominator tree post-order. The post-order enables nested // if-conversion in a single pass. The tryConvertIf() function may erase // blocks, but only blocks dominated by the head block. This makes it safe to // update the dominator tree while the post-order iterator is still active. for (auto DomNode : post_order(DomTree)) if (tryConvertIf(DomNode->getBlock())) Changed = true; return Changed; }