X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FIfConversion.cpp;h=8264d6dbab81873bdc3f34472ba27d2c1d1a4668;hb=c496875f0c205ffadcec8060e1170e1c58e4eb55;hp=16d324735c8672174a6edc55e05c751e876a52ff;hpb=86050dc8cc0aaea8c9dfeb89de02cafbd7f48d92;p=oota-llvm.git diff --git a/lib/CodeGen/IfConversion.cpp b/lib/CodeGen/IfConversion.cpp index 16d324735c8..8264d6dbab8 100644 --- a/lib/CodeGen/IfConversion.cpp +++ b/lib/CodeGen/IfConversion.cpp @@ -12,22 +12,25 @@ //===----------------------------------------------------------------------===// #define DEBUG_TYPE "ifcvt" -#include "BranchFolding.h" -#include "llvm/Function.h" #include "llvm/CodeGen/Passes.h" -#include "llvm/CodeGen/MachineModuleInfo.h" +#include "BranchFolding.h" +#include "llvm/ADT/STLExtras.h" +#include "llvm/ADT/SmallSet.h" +#include "llvm/ADT/Statistic.h" +#include "llvm/CodeGen/MachineBranchProbabilityInfo.h" #include "llvm/CodeGen/MachineFunctionPass.h" -#include "llvm/Target/TargetInstrInfo.h" -#include "llvm/Target/TargetLowering.h" -#include "llvm/Target/TargetMachine.h" -#include "llvm/Target/TargetRegisterInfo.h" +#include "llvm/CodeGen/MachineInstrBuilder.h" +#include "llvm/CodeGen/MachineModuleInfo.h" +#include "llvm/CodeGen/MachineRegisterInfo.h" +#include "llvm/MC/MCInstrItineraries.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/ADT/DepthFirstIterator.h" -#include "llvm/ADT/Statistic.h" -#include "llvm/ADT/STLExtras.h" +#include "llvm/Target/TargetInstrInfo.h" +#include "llvm/Target/TargetLowering.h" +#include "llvm/Target/TargetMachine.h" +#include "llvm/Target/TargetRegisterInfo.h" using namespace llvm; // Hidden options for help debugging. @@ -60,6 +63,7 @@ STATISTIC(NumTriangleFRev, "Number of triangle (F/R) if-conversions performed"); STATISTIC(NumDiamonds, "Number of diamond if-conversions performed"); STATISTIC(NumIfConvBBs, "Number of if-converted blocks"); STATISTIC(NumDupBBs, "Number of duplicated blocks"); +STATISTIC(NumUnpred, "Number of true blocks of diamonds unpredicated"); namespace { class IfConverter : public MachineFunctionPass { @@ -91,6 +95,8 @@ namespace { /// ClobbersPred - True if BB could modify predicates (e.g. has /// cmp, call, etc.) /// NonPredSize - Number of non-predicated instructions. + /// ExtraCost - Extra cost for multi-cycle instructions. + /// ExtraCost2 - Some instructions are slower when predicated /// BB - Corresponding MachineBasicBlock. /// TrueBB / FalseBB- See AnalyzeBranch(). /// BrCond - Conditions for end of block conditional branches. @@ -106,6 +112,8 @@ namespace { bool CannotBeCopied : 1; bool ClobbersPred : 1; unsigned NonPredSize; + unsigned ExtraCost; + unsigned ExtraCost2; MachineBasicBlock *BB; MachineBasicBlock *TrueBB; MachineBasicBlock *FalseBB; @@ -115,7 +123,7 @@ namespace { IsAnalyzed(false), IsEnqueued(false), IsBrAnalyzable(false), HasFallThrough(false), IsUnpredicable(false), CannotBeCopied(false), ClobbersPred(false), NonPredSize(0), - BB(0), TrueBB(0), FalseBB(0) {} + ExtraCost(0), ExtraCost2(0), BB(0), TrueBB(0), FalseBB(0) {} }; /// IfcvtToken - Record information about pending if-conversions to attempt: @@ -139,31 +147,40 @@ namespace { : BBI(b), Kind(k), NeedSubsumption(s), NumDups(d), NumDups2(d2) {} }; - /// Roots - Basic blocks that do not have successors. These are the starting - /// points of Graph traversal. - std::vector Roots; - /// BBAnalysis - Results of if-conversion feasibility analysis indexed by /// basic block number. std::vector BBAnalysis; - const TargetLowering *TLI; + const TargetLoweringBase *TLI; const TargetInstrInfo *TII; const TargetRegisterInfo *TRI; + const InstrItineraryData *InstrItins; + const MachineBranchProbabilityInfo *MBPI; + MachineRegisterInfo *MRI; + + bool PreRegAlloc; bool MadeChange; int FnNum; public: static char ID; - IfConverter() : MachineFunctionPass(&ID), FnNum(-1) {} + IfConverter() : MachineFunctionPass(ID), FnNum(-1) { + initializeIfConverterPass(*PassRegistry::getPassRegistry()); + } + + virtual void getAnalysisUsage(AnalysisUsage &AU) const { + AU.addRequired(); + MachineFunctionPass::getAnalysisUsage(AU); + } virtual bool runOnMachineFunction(MachineFunction &MF); - virtual const char *getPassName() const { return "If Converter"; } private: bool ReverseBranchCondition(BBInfo &BBI); - bool ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const; + bool ValidSimple(BBInfo &TrueBBI, unsigned &Dups, + const BranchProbability &Prediction) const; bool ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, - bool FalseBranch, unsigned &Dups) const; + bool FalseBranch, unsigned &Dups, + const BranchProbability &Prediction) const; bool ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI, unsigned &Dups1, unsigned &Dups2) const; void ScanInstructions(BBInfo &BBI); @@ -181,15 +198,29 @@ namespace { void PredicateBlock(BBInfo &BBI, MachineBasicBlock::iterator E, SmallVectorImpl &Cond, - SmallSet &Redefs); + SmallSet &Redefs, + SmallSet *LaterRedefs = 0); void CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, SmallVectorImpl &Cond, SmallSet &Redefs, bool IgnoreBr = false); - void MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI); + void MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges = true); - bool MeetIfcvtSizeLimit(unsigned Size) const { - return Size > 0 && Size <= TLI->getIfCvtBlockSizeLimit(); + bool MeetIfcvtSizeLimit(MachineBasicBlock &BB, + unsigned Cycle, unsigned Extra, + const BranchProbability &Prediction) const { + return Cycle > 0 && TII->isProfitableToIfCvt(BB, Cycle, Extra, + Prediction); + } + + bool MeetIfcvtSizeLimit(MachineBasicBlock &TBB, + unsigned TCycle, unsigned TExtra, + MachineBasicBlock &FBB, + unsigned FCycle, unsigned FExtra, + const BranchProbability &Prediction) const { + return TCycle > 0 && FCycle > 0 && + TII->isProfitableToIfCvt(TBB, TCycle, TExtra, FBB, FCycle, FExtra, + Prediction); } // blockAlwaysFallThrough - Block ends without a terminator. @@ -224,25 +255,34 @@ namespace { char IfConverter::ID = 0; } -static RegisterPass -X("if-converter", "If Converter"); +char &llvm::IfConverterID = IfConverter::ID; -FunctionPass *llvm::createIfConverterPass() { return new IfConverter(); } +INITIALIZE_PASS_BEGIN(IfConverter, "if-converter", "If Converter", false, false) +INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo) +INITIALIZE_PASS_END(IfConverter, "if-converter", "If Converter", false, false) bool IfConverter::runOnMachineFunction(MachineFunction &MF) { TLI = MF.getTarget().getTargetLowering(); TII = MF.getTarget().getInstrInfo(); TRI = MF.getTarget().getRegisterInfo(); + MBPI = &getAnalysis(); + MRI = &MF.getRegInfo(); + InstrItins = MF.getTarget().getInstrItineraryData(); if (!TII) return false; - // Tail merge tend to expose more if-conversion opportunities. - BranchFolder BF(true); - bool BFChange = BF.OptimizeFunction(MF, TII, + PreRegAlloc = MRI->isSSA(); + + bool BFChange = false; + if (!PreRegAlloc) { + // Tail merge tend to expose more if-conversion opportunities. + BranchFolder BF(true, false); + BFChange = BF.OptimizeFunction(MF, TII, MF.getTarget().getRegisterInfo(), getAnalysisIfAvailable()); + } DEBUG(dbgs() << "\nIfcvt: function (" << ++FnNum << ") \'" - << MF.getFunction()->getName() << "\'"); + << MF.getName() << "\'"); if (FnNum < IfCvtFnStart || (IfCvtFnStop != -1 && FnNum > IfCvtFnStop)) { DEBUG(dbgs() << " skipped\n"); @@ -253,11 +293,6 @@ bool IfConverter::runOnMachineFunction(MachineFunction &MF) { MF.RenumberBlocks(); BBAnalysis.resize(MF.getNumBlockIDs()); - // Look for root nodes, i.e. blocks without successors. - for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) - if (I->succ_empty()) - Roots.push_back(I); - std::vector Tokens; MadeChange = false; unsigned NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + @@ -288,8 +323,7 @@ bool IfConverter::runOnMachineFunction(MachineFunction &MF) { bool RetVal = false; switch (Kind) { - default: assert(false && "Unexpected!"); - break; + default: llvm_unreachable("Unexpected!"); case ICSimple: case ICSimpleFalse: { bool isFalse = Kind == ICSimpleFalse; @@ -303,8 +337,8 @@ bool IfConverter::runOnMachineFunction(MachineFunction &MF) { RetVal = IfConvertSimple(BBI, Kind); DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n"); if (RetVal) { - if (isFalse) NumSimpleFalse++; - else NumSimple++; + if (isFalse) ++NumSimpleFalse; + else ++NumSimple; } break; } @@ -330,11 +364,11 @@ bool IfConverter::runOnMachineFunction(MachineFunction &MF) { DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n"); if (RetVal) { if (isFalse) { - if (isRev) NumTriangleFRev++; - else NumTriangleFalse++; + if (isRev) ++NumTriangleFRev; + else ++NumTriangleFalse; } else { - if (isRev) NumTriangleRev++; - else NumTriangle++; + if (isRev) ++NumTriangleRev; + else ++NumTriangle; } } break; @@ -346,7 +380,7 @@ bool IfConverter::runOnMachineFunction(MachineFunction &MF) { << BBI.FalseBB->getNumber() << ") "); RetVal = IfConvertDiamond(BBI, Kind, NumDups, NumDups2); DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n"); - if (RetVal) NumDiamonds++; + if (RetVal) ++NumDiamonds; break; } } @@ -372,11 +406,10 @@ bool IfConverter::runOnMachineFunction(MachineFunction &MF) { } Tokens.clear(); - Roots.clear(); BBAnalysis.clear(); if (MadeChange && IfCvtBranchFold) { - BranchFolder BF(false); + BranchFolder BF(false, false); BF.OptimizeFunction(MF, TII, MF.getTarget().getRegisterInfo(), getAnalysisIfAvailable()); @@ -426,7 +459,8 @@ static inline MachineBasicBlock *getNextBlock(MachineBasicBlock *BB) { /// predecessor) forms a valid simple shape for ifcvt. It also returns the /// number of instructions that the ifcvt would need to duplicate if performed /// in Dups. -bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const { +bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups, + const BranchProbability &Prediction) const { Dups = 0; if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) return false; @@ -436,7 +470,8 @@ bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const { if (TrueBBI.BB->pred_size() > 1) { if (TrueBBI.CannotBeCopied || - TrueBBI.NonPredSize > TLI->getIfCvtDupBlockSizeLimit()) + !TII->isProfitableToDupForIfCvt(*TrueBBI.BB, TrueBBI.NonPredSize, + Prediction)) return false; Dups = TrueBBI.NonPredSize; } @@ -451,7 +486,8 @@ bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const { /// returns the number of instructions that the ifcvt would need to duplicate /// if performed in 'Dups'. bool IfConverter::ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, - bool FalseBranch, unsigned &Dups) const { + bool FalseBranch, unsigned &Dups, + const BranchProbability &Prediction) const { Dups = 0; if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) return false; @@ -473,7 +509,7 @@ bool IfConverter::ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, ++Size; } } - if (Size > TLI->getIfCvtDupBlockSizeLimit()) + if (!TII->isProfitableToDupForIfCvt(*TrueBBI.BB, Size, Prediction)) return false; Dups = Size; } @@ -488,18 +524,6 @@ bool IfConverter::ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, return TExit && TExit == FalseBBI.BB; } -static -MachineBasicBlock::iterator firstNonBranchInst(MachineBasicBlock *BB, - const TargetInstrInfo *TII) { - MachineBasicBlock::iterator I = BB->end(); - while (I != BB->begin()) { - --I; - if (!I->getDesc().isBranch()) - break; - } - return I; -} - /// ValidDiamond - Returns true if the 'true' and 'false' blocks (along /// with their common predecessor) forms a valid diamond shape for ifcvt. bool IfConverter::ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI, @@ -528,64 +552,70 @@ bool IfConverter::ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI, (TrueBBI.ClobbersPred && FalseBBI.ClobbersPred)) return false; - MachineBasicBlock::iterator TI = TrueBBI.BB->begin(); - MachineBasicBlock::iterator FI = FalseBBI.BB->begin(); + // Count duplicate instructions at the beginning of the true and false blocks. + MachineBasicBlock::iterator TIB = TrueBBI.BB->begin(); + MachineBasicBlock::iterator FIB = FalseBBI.BB->begin(); MachineBasicBlock::iterator TIE = TrueBBI.BB->end(); MachineBasicBlock::iterator FIE = FalseBBI.BB->end(); - // Skip dbg_value instructions - while (TI != TIE && TI->isDebugValue()) - ++TI; - while (FI != FIE && FI->isDebugValue()) - ++FI; - while (TI != TIE && FI != FIE) { + while (TIB != TIE && FIB != FIE) { // Skip dbg_value instructions. These do not count. - if (TI->isDebugValue()) { - while (TI != TIE && TI->isDebugValue()) - ++TI; - if (TI == TIE) + if (TIB->isDebugValue()) { + while (TIB != TIE && TIB->isDebugValue()) + ++TIB; + if (TIB == TIE) break; } - if (FI->isDebugValue()) { - while (FI != FIE && FI->isDebugValue()) - ++FI; - if (FI == FIE) + if (FIB->isDebugValue()) { + while (FIB != FIE && FIB->isDebugValue()) + ++FIB; + if (FIB == FIE) break; } - if (!TI->isIdenticalTo(FI)) + if (!TIB->isIdenticalTo(FIB)) break; ++Dups1; - ++TI; - ++FI; + ++TIB; + ++FIB; } - TI = firstNonBranchInst(TrueBBI.BB, TII); - FI = firstNonBranchInst(FalseBBI.BB, TII); - MachineBasicBlock::iterator TIB = TrueBBI.BB->begin(); - MachineBasicBlock::iterator FIB = FalseBBI.BB->begin(); - // Skip dbg_value instructions at end of the bb's. - while (TI != TIB && TI->isDebugValue()) - --TI; - while (FI != FIB && FI->isDebugValue()) - --FI; - while (TI != TIB && FI != FIB) { + // Now, in preparation for counting duplicate instructions at the ends of the + // blocks, move the end iterators up past any branch instructions. + while (TIE != TIB) { + --TIE; + if (!TIE->isBranch()) + break; + } + while (FIE != FIB) { + --FIE; + if (!FIE->isBranch()) + break; + } + + // If Dups1 includes all of a block, then don't count duplicate + // instructions at the end of the blocks. + if (TIB == TIE || FIB == FIE) + return true; + + // Count duplicate instructions at the ends of the blocks. + while (TIE != TIB && FIE != FIB) { // Skip dbg_value instructions. These do not count. - if (TI->isDebugValue()) { - while (TI != TIB && TI->isDebugValue()) - --TI; - if (TI == TIB) + if (TIE->isDebugValue()) { + while (TIE != TIB && TIE->isDebugValue()) + --TIE; + if (TIE == TIB) break; } - if (FI->isDebugValue()) { - while (FI != FIB && FI->isDebugValue()) - --FI; - if (FI == FIB) + if (FIE->isDebugValue()) { + while (FIE != FIB && FIE->isDebugValue()) + --FIE; + if (FIE == FIB) break; } - if (!TI->isIdenticalTo(FI)) + if (!TIE->isIdenticalTo(FIE)) break; ++Dups2; - --TI; - --FI; + --TIE; + --FIE; } return true; @@ -600,7 +630,7 @@ void IfConverter::ScanInstructions(BBInfo &BBI) { if (BBI.IsDone) return; - bool AlreadyPredicated = BBI.Predicate.size() > 0; + bool AlreadyPredicated = !BBI.Predicate.empty(); // First analyze the end of BB branches. BBI.TrueBB = BBI.FalseBB = NULL; BBI.BrCond.clear(); @@ -622,23 +652,30 @@ void IfConverter::ScanInstructions(BBInfo &BBI) { // Then scan all the instructions. BBI.NonPredSize = 0; + BBI.ExtraCost = 0; + BBI.ExtraCost2 = 0; BBI.ClobbersPred = false; for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end(); I != E; ++I) { if (I->isDebugValue()) continue; - const TargetInstrDesc &TID = I->getDesc(); - if (TID.isNotDuplicable()) + if (I->isNotDuplicable()) BBI.CannotBeCopied = true; bool isPredicated = TII->isPredicated(I); - bool isCondBr = BBI.IsBrAnalyzable && TID.isConditionalBranch(); + bool isCondBr = BBI.IsBrAnalyzable && I->isConditionalBranch(); if (!isCondBr) { - if (!isPredicated) + if (!isPredicated) { BBI.NonPredSize++; - else if (!AlreadyPredicated) { + unsigned ExtraPredCost = 0; + unsigned NumCycles = TII->getInstrLatency(InstrItins, &*I, + &ExtraPredCost); + if (NumCycles > 1) + BBI.ExtraCost += NumCycles-1; + BBI.ExtraCost2 += ExtraPredCost; + } else if (!AlreadyPredicated) { // FIXME: This instruction is already predicated before the // if-conversion pass. It's probably something like a conditional move. // Mark this block unpredicable for now. @@ -722,8 +759,9 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, ScanInstructions(BBI); - // Unanalyzable or ends with fallthrough or unconditional branch. - if (!BBI.IsBrAnalyzable || BBI.BrCond.empty()) { + // Unanalyzable or ends with fallthrough or unconditional branch, or if is not + // considered for ifcvt anymore. + if (!BBI.IsBrAnalyzable || BBI.BrCond.empty() || BBI.IsDone) { BBI.IsBeingAnalyzed = false; BBI.IsAnalyzed = true; return BBI; @@ -757,12 +795,18 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, unsigned Dups = 0; unsigned Dups2 = 0; - bool TNeedSub = TrueBBI.Predicate.size() > 0; - bool FNeedSub = FalseBBI.Predicate.size() > 0; + bool TNeedSub = !TrueBBI.Predicate.empty(); + bool FNeedSub = !FalseBBI.Predicate.empty(); bool Enqueued = false; + + BranchProbability Prediction = MBPI->getEdgeProbability(BB, TrueBBI.BB); + if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI, Dups, Dups2) && - MeetIfcvtSizeLimit(TrueBBI.NonPredSize - (Dups + Dups2)) && - MeetIfcvtSizeLimit(FalseBBI.NonPredSize - (Dups + Dups2)) && + MeetIfcvtSizeLimit(*TrueBBI.BB, (TrueBBI.NonPredSize - (Dups + Dups2) + + TrueBBI.ExtraCost), TrueBBI.ExtraCost2, + *FalseBBI.BB, (FalseBBI.NonPredSize - (Dups + Dups2) + + FalseBBI.ExtraCost),FalseBBI.ExtraCost2, + Prediction) && FeasibilityAnalysis(TrueBBI, BBI.BrCond) && FeasibilityAnalysis(FalseBBI, RevCond)) { // Diamond: @@ -778,8 +822,9 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, Enqueued = true; } - if (ValidTriangle(TrueBBI, FalseBBI, false, Dups) && - MeetIfcvtSizeLimit(TrueBBI.NonPredSize) && + if (ValidTriangle(TrueBBI, FalseBBI, false, Dups, Prediction) && + MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost, + TrueBBI.ExtraCost2, Prediction) && FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) { // Triangle: // EBB @@ -792,15 +837,17 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, Enqueued = true; } - if (ValidTriangle(TrueBBI, FalseBBI, true, Dups) && - MeetIfcvtSizeLimit(TrueBBI.NonPredSize) && + if (ValidTriangle(TrueBBI, FalseBBI, true, Dups, Prediction) && + MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost, + TrueBBI.ExtraCost2, Prediction) && FeasibilityAnalysis(TrueBBI, BBI.BrCond, true, true)) { Tokens.push_back(new IfcvtToken(BBI, ICTriangleRev, TNeedSub, Dups)); Enqueued = true; } - if (ValidSimple(TrueBBI, Dups) && - MeetIfcvtSizeLimit(TrueBBI.NonPredSize) && + if (ValidSimple(TrueBBI, Dups, Prediction) && + MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost, + TrueBBI.ExtraCost2, Prediction) && FeasibilityAnalysis(TrueBBI, BBI.BrCond)) { // Simple (split, no rejoin): // EBB @@ -815,22 +862,30 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, if (CanRevCond) { // Try the other path... - if (ValidTriangle(FalseBBI, TrueBBI, false, Dups) && - MeetIfcvtSizeLimit(FalseBBI.NonPredSize) && + if (ValidTriangle(FalseBBI, TrueBBI, false, Dups, + Prediction.getCompl()) && + MeetIfcvtSizeLimit(*FalseBBI.BB, + FalseBBI.NonPredSize + FalseBBI.ExtraCost, + FalseBBI.ExtraCost2, Prediction.getCompl()) && FeasibilityAnalysis(FalseBBI, RevCond, true)) { Tokens.push_back(new IfcvtToken(BBI, ICTriangleFalse, FNeedSub, Dups)); Enqueued = true; } - if (ValidTriangle(FalseBBI, TrueBBI, true, Dups) && - MeetIfcvtSizeLimit(FalseBBI.NonPredSize) && + if (ValidTriangle(FalseBBI, TrueBBI, true, Dups, + Prediction.getCompl()) && + MeetIfcvtSizeLimit(*FalseBBI.BB, + FalseBBI.NonPredSize + FalseBBI.ExtraCost, + FalseBBI.ExtraCost2, Prediction.getCompl()) && FeasibilityAnalysis(FalseBBI, RevCond, true, true)) { Tokens.push_back(new IfcvtToken(BBI, ICTriangleFRev, FNeedSub, Dups)); Enqueued = true; } - if (ValidSimple(FalseBBI, Dups) && - MeetIfcvtSizeLimit(FalseBBI.NonPredSize) && + if (ValidSimple(FalseBBI, Dups, Prediction.getCompl()) && + MeetIfcvtSizeLimit(*FalseBBI.BB, + FalseBBI.NonPredSize + FalseBBI.ExtraCost, + FalseBBI.ExtraCost2, Prediction.getCompl()) && FeasibilityAnalysis(FalseBBI, RevCond)) { Tokens.push_back(new IfcvtToken(BBI, ICSimpleFalse, FNeedSub, Dups)); Enqueued = true; @@ -847,13 +902,9 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, /// candidates. void IfConverter::AnalyzeBlocks(MachineFunction &MF, std::vector &Tokens) { - std::set Visited; - for (unsigned i = 0, e = Roots.size(); i != e; ++i) { - for (idf_ext_iterator I=idf_ext_begin(Roots[i],Visited), - E = idf_ext_end(Roots[i], Visited); I != E; ++I) { - MachineBasicBlock *BB = *I; - AnalyzeBlock(BB, Tokens); - } + for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { + MachineBasicBlock *BB = I; + AnalyzeBlock(BB, Tokens); } // Sort to favor more complex ifcvt scheme. @@ -920,9 +971,8 @@ static void InitPredRedefs(MachineBasicBlock *BB, SmallSet &Redefs, E = BB->livein_end(); I != E; ++I) { unsigned Reg = *I; Redefs.insert(Reg); - for (const unsigned *Subreg = TRI->getSubRegisters(Reg); - *Subreg; ++Subreg) - Redefs.insert(*Subreg); + for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs) + Redefs.insert(*SubRegs); } } @@ -941,21 +991,20 @@ static void UpdatePredRedefs(MachineInstr *MI, SmallSet &Redefs, Defs.push_back(Reg); else if (MO.isKill()) { Redefs.erase(Reg); - for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) - Redefs.erase(*SR); + for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs) + Redefs.erase(*SubRegs); } } + MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI); for (unsigned i = 0, e = Defs.size(); i != e; ++i) { unsigned Reg = Defs[i]; - if (Redefs.count(Reg)) { + if (!Redefs.insert(Reg)) { if (AddImpUse) // Treat predicated update as read + write. - MI->addOperand(MachineOperand::CreateReg(Reg, false/*IsDef*/, - true/*IsImp*/,false/*IsKill*/)); + MIB.addReg(Reg, RegState::Implicit | RegState::Undef); } else { - Redefs.insert(Reg); - for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) - Redefs.insert(*SR); + for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs) + Redefs.insert(*SubRegs); } } } @@ -990,11 +1039,15 @@ bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) { return false; } + if (CvtBBI->BB->hasAddressTaken()) + // Conservatively abort if-conversion if BB's address is taken. + return false; + if (Kind == ICSimpleFalse) if (TII->ReverseBranchCondition(Cond)) - assert(false && "Unable to reverse branch condition!"); + llvm_unreachable("Unable to reverse branch condition!"); - // Initialize liveins to the first BB. These are potentiall re-defined by + // Initialize liveins to the first BB. These are potentiall redefined by // predicated instructions. SmallSet Redefs; InitPredRedefs(CvtBBI->BB, Redefs, TRI); @@ -1005,6 +1058,10 @@ bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) { // Copy instructions in the true block, predicate them, and add them to // the entry block. CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs); + + // RemoveExtraEdges won't work if the block has an unanalyzable branch, so + // explicitly remove CvtBBI as a successor. + BBI.BB->removeSuccessor(CvtBBI->BB); } else { PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs); @@ -1063,9 +1120,13 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { return false; } + if (CvtBBI->BB->hasAddressTaken()) + // Conservatively abort if-conversion if BB's address is taken. + return false; + if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) if (TII->ReverseBranchCondition(Cond)) - assert(false && "Unable to reverse branch condition!"); + llvm_unreachable("Unable to reverse branch condition!"); if (Kind == ICTriangleRev || Kind == ICTriangleFRev) { if (ReverseBranchCondition(*CvtBBI)) { @@ -1085,19 +1146,22 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { } } - // Initialize liveins to the first BB. These are potentiall re-defined by + // Initialize liveins to the first BB. These are potentially redefined by // predicated instructions. SmallSet Redefs; InitPredRedefs(CvtBBI->BB, Redefs, TRI); InitPredRedefs(NextBBI->BB, Redefs, TRI); bool HasEarlyExit = CvtBBI->FalseBB != NULL; - bool DupBB = CvtBBI->BB->pred_size() > 1; - if (DupBB) { + if (CvtBBI->BB->pred_size() > 1) { BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); // Copy instructions in the true block, predicate them, and add them to // the entry block. CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs, true); + + // RemoveExtraEdges won't work if the block has an unanalyzable branch, so + // explicitly remove CvtBBI as a successor. + BBI.BB->removeSuccessor(CvtBBI->BB); } else { // Predicate the 'true' block after removing its branch. CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB); @@ -1105,7 +1169,7 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { // Now merge the entry of the triangle with the true block. BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); - MergeBlocks(BBI, *CvtBBI); + MergeBlocks(BBI, *CvtBBI, false); } // If 'true' block has a 'false' successor, add an exit branch to it. @@ -1113,7 +1177,7 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { SmallVector RevCond(CvtBBI->BrCond.begin(), CvtBBI->BrCond.end()); if (TII->ReverseBranchCondition(RevCond)) - assert(false && "Unable to reverse branch condition!"); + llvm_unreachable("Unable to reverse branch condition!"); TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, NULL, RevCond, dl); BBI.BB->addSuccessor(CvtBBI->FalseBB); } @@ -1128,7 +1192,8 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { // block. By not merging them, we make it possible to iteratively // ifcvt the blocks. if (!HasEarlyExit && - NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough) { + NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough && + !NextBBI->BB->hasAddressTaken()) { MergeBlocks(BBI, *NextBBI); FalseBBDead = true; } else { @@ -1178,14 +1243,18 @@ bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, return false; } - // Merge the 'true' and 'false' blocks by copying the instructions - // from the 'false' block to the 'true' block. That is, unless the true - // block would clobber the predicate, in that case, do the opposite. + if (TrueBBI.BB->hasAddressTaken() || FalseBBI.BB->hasAddressTaken()) + // Conservatively abort if-conversion if either BB has its address taken. + return false; + + // Put the predicated instructions from the 'true' block before the + // instructions from the 'false' block, unless the true block would clobber + // the predicate, in which case, do the opposite. BBInfo *BBI1 = &TrueBBI; BBInfo *BBI2 = &FalseBBI; SmallVector RevCond(BBI.BrCond.begin(), BBI.BrCond.end()); if (TII->ReverseBranchCondition(RevCond)) - assert(false && "Unable to reverse branch condition!"); + llvm_unreachable("Unable to reverse branch condition!"); SmallVector *Cond1 = &BBI.BrCond; SmallVector *Cond2 = &RevCond; @@ -1205,7 +1274,7 @@ bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, // Remove the conditional branch from entry to the blocks. BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); - // Initialize liveins to the first BB. These are potentiall re-defined by + // Initialize liveins to the first BB. These are potentially redefined by // predicated instructions. SmallSet Redefs; InitPredRedefs(BBI1->BB, Redefs, TRI); @@ -1222,17 +1291,24 @@ bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, ++DI2; BBI1->NonPredSize -= NumDups1; BBI2->NonPredSize -= NumDups1; + + // Skip past the dups on each side separately since there may be + // differing dbg_value entries. + for (unsigned i = 0; i < NumDups1; ++DI1) { + if (!DI1->isDebugValue()) + ++i; + } while (NumDups1 != 0) { - ++DI1; ++DI2; - --NumDups1; + if (!DI2->isDebugValue()) + --NumDups1; } UpdatePredRedefs(BBI1->BB->begin(), DI1, Redefs, TRI); BBI.BB->splice(BBI.BB->end(), BBI1->BB, BBI1->BB->begin(), DI1); BBI2->BB->erase(BBI2->BB->begin(), DI2); - // Predicate the 'true' block after removing its branch. + // Remove branch from 'true' block and remove duplicated instructions. BBI1->NonPredSize -= TII->RemoveBranch(*BBI1->BB); DI1 = BBI1->BB->end(); for (unsigned i = 0; i != NumDups2; ) { @@ -1245,9 +1321,8 @@ bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, ++i; } BBI1->BB->erase(DI1, BBI1->BB->end()); - PredicateBlock(*BBI1, BBI1->BB->end(), *Cond1, Redefs); - // Predicate the 'false' block. + // Remove 'false' block branch and find the last instruction to predicate. BBI2->NonPredSize -= TII->RemoveBranch(*BBI2->BB); DI2 = BBI2->BB->end(); while (NumDups2 != 0) { @@ -1259,28 +1334,94 @@ bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, if (!DI2->isDebugValue()) --NumDups2; } + + // Remember which registers would later be defined by the false block. + // This allows us not to predicate instructions in the true block that would + // later be re-defined. That is, rather than + // subeq r0, r1, #1 + // addne r0, r1, #1 + // generate: + // sub r0, r1, #1 + // addne r0, r1, #1 + SmallSet RedefsByFalse; + SmallSet ExtUses; + if (TII->isProfitableToUnpredicate(*BBI1->BB, *BBI2->BB)) { + for (MachineBasicBlock::iterator FI = BBI2->BB->begin(); FI != DI2; ++FI) { + if (FI->isDebugValue()) + continue; + SmallVector Defs; + for (unsigned i = 0, e = FI->getNumOperands(); i != e; ++i) { + const MachineOperand &MO = FI->getOperand(i); + if (!MO.isReg()) + continue; + unsigned Reg = MO.getReg(); + if (!Reg) + continue; + if (MO.isDef()) { + Defs.push_back(Reg); + } else if (!RedefsByFalse.count(Reg)) { + // These are defined before ctrl flow reach the 'false' instructions. + // They cannot be modified by the 'true' instructions. + ExtUses.insert(Reg); + for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs) + ExtUses.insert(*SubRegs); + } + } + + for (unsigned i = 0, e = Defs.size(); i != e; ++i) { + unsigned Reg = Defs[i]; + if (!ExtUses.count(Reg)) { + RedefsByFalse.insert(Reg); + for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs) + RedefsByFalse.insert(*SubRegs); + } + } + } + } + + // Predicate the 'true' block. + PredicateBlock(*BBI1, BBI1->BB->end(), *Cond1, Redefs, &RedefsByFalse); + + // Predicate the 'false' block. PredicateBlock(*BBI2, DI2, *Cond2, Redefs); // Merge the true block into the entry of the diamond. - MergeBlocks(BBI, *BBI1); - MergeBlocks(BBI, *BBI2); + MergeBlocks(BBI, *BBI1, TailBB == 0); + MergeBlocks(BBI, *BBI2, TailBB == 0); // If the if-converted block falls through or unconditionally branches into // the tail block, and the tail block does not have other predecessors, then // fold the tail block in as well. Otherwise, unless it falls through to the // tail, add a unconditional branch to it. if (TailBB) { - BBInfo TailBBI = BBAnalysis[TailBB->getNumber()]; - if (TailBB->pred_size() == 1 && !TailBBI.HasFallThrough) { - BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); + BBInfo &TailBBI = BBAnalysis[TailBB->getNumber()]; + bool CanMergeTail = !TailBBI.HasFallThrough && + !TailBBI.BB->hasAddressTaken(); + // There may still be a fall-through edge from BBI1 or BBI2 to TailBB; + // check if there are any other predecessors besides those. + unsigned NumPreds = TailBB->pred_size(); + if (NumPreds > 1) + CanMergeTail = false; + else if (NumPreds == 1 && CanMergeTail) { + MachineBasicBlock::pred_iterator PI = TailBB->pred_begin(); + if (*PI != BBI1->BB && *PI != BBI2->BB) + CanMergeTail = false; + } + if (CanMergeTail) { MergeBlocks(BBI, TailBBI); TailBBI.IsDone = true; } else { + BBI.BB->addSuccessor(TailBB); InsertUncondBranch(BBI.BB, TailBB, TII); BBI.HasFallThrough = false; } } + // RemoveExtraEdges won't work if the block has an unanalyzable branch, + // which can happen here if TailBB is unanalyzable and is merged, so + // explicitly remove BBI1 and BBI2 as successors. + BBI.BB->removeSuccessor(BBI1->BB); + BBI.BB->removeSuccessor(BBI2->BB); RemoveExtraEdges(BBI); // Update block info. @@ -1291,15 +1432,49 @@ bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, return true; } +static bool MaySpeculate(const MachineInstr *MI, + SmallSet &LaterRedefs, + const TargetInstrInfo *TII) { + bool SawStore = true; + if (!MI->isSafeToMove(TII, 0, SawStore)) + return false; + + for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { + const MachineOperand &MO = MI->getOperand(i); + if (!MO.isReg()) + continue; + unsigned Reg = MO.getReg(); + if (!Reg) + continue; + if (MO.isDef() && !LaterRedefs.count(Reg)) + return false; + } + + return true; +} + /// PredicateBlock - Predicate instructions from the start of the block to the /// specified end with the specified condition. void IfConverter::PredicateBlock(BBInfo &BBI, MachineBasicBlock::iterator E, SmallVectorImpl &Cond, - SmallSet &Redefs) { + SmallSet &Redefs, + SmallSet *LaterRedefs) { + bool AnyUnpred = false; + bool MaySpec = LaterRedefs != 0; for (MachineBasicBlock::iterator I = BBI.BB->begin(); I != E; ++I) { if (I->isDebugValue() || TII->isPredicated(I)) continue; + // It may be possible not to predicate an instruction if it's the 'true' + // side of a diamond and the 'false' side may re-define the instruction's + // defs. + if (MaySpec && MaySpeculate(I, *LaterRedefs, TII)) { + AnyUnpred = true; + continue; + } + // If any instruction is predicated, then every instruction after it must + // be predicated. + MaySpec = false; if (!TII->PredicateInstruction(I, Cond)) { #ifndef NDEBUG dbgs() << "Unable to predicate " << *I << "!\n"; @@ -1307,7 +1482,7 @@ void IfConverter::PredicateBlock(BBInfo &BBI, llvm_unreachable(0); } - // If the predicated instruction now re-defines a register as the result of + // If the predicated instruction now redefines a register as the result of // if-conversion, add an implicit kill. UpdatePredRedefs(I, Redefs, TRI, true); } @@ -1317,7 +1492,9 @@ void IfConverter::PredicateBlock(BBInfo &BBI, BBI.IsAnalyzed = false; BBI.NonPredSize = 0; - NumIfConvBBs++; + ++NumIfConvBBs; + if (AnyUnpred) + ++NumUnpred; } /// CopyAndPredicateBlock - Copy and predicate instructions from source BB to @@ -1330,14 +1507,18 @@ void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, for (MachineBasicBlock::iterator I = FromBBI.BB->begin(), E = FromBBI.BB->end(); I != E; ++I) { - const TargetInstrDesc &TID = I->getDesc(); // Do not copy the end of the block branches. - if (IgnoreBr && TID.isBranch()) + if (IgnoreBr && I->isBranch()) break; MachineInstr *MI = MF.CloneMachineInstr(I); ToBBI.BB->insert(ToBBI.BB->end(), MI); ToBBI.NonPredSize++; + unsigned ExtraPredCost = 0; + unsigned NumCycles = TII->getInstrLatency(InstrItins, &*I, &ExtraPredCost); + if (NumCycles > 1) + ToBBI.ExtraCost += NumCycles-1; + ToBBI.ExtraCost2 += ExtraPredCost; if (!TII->isPredicated(I) && !MI->isDebugValue()) { if (!TII->PredicateInstruction(MI, Cond)) { @@ -1348,22 +1529,24 @@ void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, } } - // If the predicated instruction now re-defines a register as the result of + // If the predicated instruction now redefines a register as the result of // if-conversion, add an implicit kill. UpdatePredRedefs(MI, Redefs, TRI, true); } - std::vector Succs(FromBBI.BB->succ_begin(), - FromBBI.BB->succ_end()); - MachineBasicBlock *NBB = getNextBlock(FromBBI.BB); - MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL; + if (!IgnoreBr) { + std::vector Succs(FromBBI.BB->succ_begin(), + FromBBI.BB->succ_end()); + MachineBasicBlock *NBB = getNextBlock(FromBBI.BB); + MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL; - for (unsigned i = 0, e = Succs.size(); i != e; ++i) { - MachineBasicBlock *Succ = Succs[i]; - // Fallthrough edge can't be transferred. - if (Succ == FallThrough) - continue; - ToBBI.BB->addSuccessor(Succ); + for (unsigned i = 0, e = Succs.size(); i != e; ++i) { + MachineBasicBlock *Succ = Succs[i]; + // Fallthrough edge can't be transferred. + if (Succ == FallThrough) + continue; + ToBBI.BB->addSuccessor(Succ); + } } std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(), @@ -1373,25 +1556,21 @@ void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, ToBBI.ClobbersPred |= FromBBI.ClobbersPred; ToBBI.IsAnalyzed = false; - NumDupBBs++; + ++NumDupBBs; } /// MergeBlocks - Move all instructions from FromBB to the end of ToBB. -/// -void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI) { +/// This will leave FromBB as an empty block, so remove all of its +/// successor edges except for the fall-through edge. If AddEdges is true, +/// i.e., when FromBBI's branch is being moved, add those successor edges to +/// ToBBI. +void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges) { + assert(!FromBBI.BB->hasAddressTaken() && + "Removing a BB whose address is taken!"); + ToBBI.BB->splice(ToBBI.BB->end(), FromBBI.BB, FromBBI.BB->begin(), FromBBI.BB->end()); - // Redirect all branches to FromBB to ToBB. - std::vector Preds(FromBBI.BB->pred_begin(), - FromBBI.BB->pred_end()); - for (unsigned i = 0, e = Preds.size(); i != e; ++i) { - MachineBasicBlock *Pred = Preds[i]; - if (Pred == ToBBI.BB) - continue; - Pred->ReplaceUsesOfBlockWith(FromBBI.BB, ToBBI.BB); - } - std::vector Succs(FromBBI.BB->succ_begin(), FromBBI.BB->succ_end()); MachineBasicBlock *NBB = getNextBlock(FromBBI.BB); @@ -1403,7 +1582,8 @@ void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI) { if (Succ == FallThrough) continue; FromBBI.BB->removeSuccessor(Succ); - ToBBI.BB->addSuccessor(Succ); + if (AddEdges && !ToBBI.BB->isSuccessor(Succ)) + ToBBI.BB->addSuccessor(Succ); } // Now FromBBI always falls through to the next block! @@ -1415,7 +1595,11 @@ void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI) { FromBBI.Predicate.clear(); ToBBI.NonPredSize += FromBBI.NonPredSize; + ToBBI.ExtraCost += FromBBI.ExtraCost; + ToBBI.ExtraCost2 += FromBBI.ExtraCost2; FromBBI.NonPredSize = 0; + FromBBI.ExtraCost = 0; + FromBBI.ExtraCost2 = 0; ToBBI.ClobbersPred |= FromBBI.ClobbersPred; ToBBI.HasFallThrough = FromBBI.HasFallThrough;