X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FIfConversion.cpp;h=8d212a606fdf305012b13e8e20decadcb19923c0;hb=a29c13086a3add78a3a79f744573fe09eaa9dc88;hp=3846284ec10d6f5b57f25e15d9a3f501eab3e697;hpb=e882fca902ba6b1a9e0c361c5781084f79eb6216;p=oota-llvm.git diff --git a/lib/CodeGen/IfConversion.cpp b/lib/CodeGen/IfConversion.cpp index 3846284ec10..8d212a606fd 100644 --- a/lib/CodeGen/IfConversion.cpp +++ b/lib/CodeGen/IfConversion.cpp @@ -2,8 +2,8 @@ // // The LLVM Compiler Infrastructure // -// This file was developed by the Evan Cheng and is distributed under -// the University of Illinois Open Source License. See LICENSE.TXT for details. +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// // @@ -23,28 +23,27 @@ #include "llvm/Support/Debug.h" #include "llvm/ADT/DepthFirstIterator.h" #include "llvm/ADT/Statistic.h" +#include "llvm/ADT/STLExtras.h" using namespace llvm; -namespace { - // Hidden options for help debugging. - cl::opt IfCvtFnStart("ifcvt-fn-start", cl::init(-1), cl::Hidden); - cl::opt IfCvtFnStop("ifcvt-fn-stop", cl::init(-1), cl::Hidden); - cl::opt IfCvtLimit("ifcvt-limit", cl::init(-1), cl::Hidden); - cl::opt DisableSimple("disable-ifcvt-simple", - cl::init(false), cl::Hidden); - cl::opt DisableSimpleF("disable-ifcvt-simple-false", - cl::init(false), cl::Hidden); - cl::opt DisableTriangle("disable-ifcvt-triangle", - cl::init(false), cl::Hidden); - cl::opt DisableTriangleR("disable-ifcvt-triangle-rev", - cl::init(false), cl::Hidden); - cl::opt DisableTriangleF("disable-ifcvt-triangle-false", - cl::init(false), cl::Hidden); - cl::opt DisableTriangleFR("disable-ifcvt-triangle-false-rev", - cl::init(false), cl::Hidden); - cl::opt DisableDiamond("disable-ifcvt-diamond", - cl::init(false), cl::Hidden); -} +// Hidden options for help debugging. +static cl::opt IfCvtFnStart("ifcvt-fn-start", cl::init(-1), cl::Hidden); +static cl::opt IfCvtFnStop("ifcvt-fn-stop", cl::init(-1), cl::Hidden); +static cl::opt IfCvtLimit("ifcvt-limit", cl::init(-1), cl::Hidden); +static cl::opt DisableSimple("disable-ifcvt-simple", + cl::init(false), cl::Hidden); +static cl::opt DisableSimpleF("disable-ifcvt-simple-false", + cl::init(false), cl::Hidden); +static cl::opt DisableTriangle("disable-ifcvt-triangle", + cl::init(false), cl::Hidden); +static cl::opt DisableTriangleR("disable-ifcvt-triangle-rev", + cl::init(false), cl::Hidden); +static cl::opt DisableTriangleF("disable-ifcvt-triangle-false", + cl::init(false), cl::Hidden); +static cl::opt DisableTriangleFR("disable-ifcvt-triangle-false-rev", + cl::init(false), cl::Hidden); +static cl::opt DisableDiamond("disable-ifcvt-diamond", + cl::init(false), cl::Hidden); STATISTIC(NumSimple, "Number of simple if-conversions performed"); STATISTIC(NumSimpleFalse, "Number of simple (F) if-conversions performed"); @@ -57,7 +56,7 @@ STATISTIC(NumIfConvBBs, "Number of if-converted blocks"); STATISTIC(NumDupBBs, "Number of duplicated blocks"); namespace { - class IfConverter : public MachineFunctionPass { + class VISIBILITY_HIDDEN IfConverter : public MachineFunctionPass { enum IfcvtKind { ICNotClassfied, // BB data valid, but not classified. ICSimpleFalse, // Same as ICSimple, but on the false path. @@ -83,7 +82,7 @@ namespace { /// IsBrAnalyzable - True if AnalyzeBranch() returns false. /// HasFallThrough - True if BB may fallthrough to the following BB. /// IsUnpredicable - True if BB is known to be unpredicable. - /// ClobbersPredicate- True if BB would modify the predicate (e.g. has + /// ClobbersPred - True if BB could modify predicates (e.g. has /// cmp, call, etc.) /// NonPredSize - Number of non-predicated instructions. /// BB - Corresponding MachineBasicBlock. @@ -104,8 +103,8 @@ namespace { MachineBasicBlock *BB; MachineBasicBlock *TrueBB; MachineBasicBlock *FalseBB; - std::vector BrCond; - std::vector Predicate; + SmallVector BrCond; + SmallVector Predicate; BBInfo() : IsDone(false), IsBeingAnalyzed(false), IsAnalyzed(false), IsEnqueued(false), IsBrAnalyzable(false), HasFallThrough(false), IsUnpredicable(false), @@ -118,15 +117,20 @@ namespace { /// Kind - Type of block. See IfcvtKind. /// NeedSubsumsion - True if the to be predicated BB has already been /// predicated. - /// Duplicates - Number of instructions that would be duplicated due - /// to this if-conversion. + /// NumDups - Number of instructions that would be duplicated due + /// to this if-conversion. (For diamonds, the number of + /// identical instructions at the beginnings of both + /// paths). + /// NumDups2 - For diamonds, the number of identical instructions + /// at the ends of both paths. struct IfcvtToken { BBInfo &BBI; IfcvtKind Kind; bool NeedSubsumsion; - unsigned Duplicates; - IfcvtToken(BBInfo &b, IfcvtKind k, bool s, unsigned d) - : BBI(b), Kind(k), NeedSubsumsion(s), Duplicates(d) {} + unsigned NumDups; + unsigned NumDups2; + IfcvtToken(BBInfo &b, IfcvtKind k, bool s, unsigned d, unsigned d2 = 0) + : BBI(b), Kind(k), NeedSubsumsion(s), NumDups(d), NumDups2(d2) {} }; /// Roots - Basic blocks that do not have successors. These are the starting @@ -142,37 +146,43 @@ namespace { bool MadeChange; public: static char ID; - IfConverter() : MachineFunctionPass((intptr_t)&ID) {} + IfConverter() : MachineFunctionPass(&ID) {} virtual bool runOnMachineFunction(MachineFunction &MF); - virtual const char *getPassName() const { return "If converter"; } + virtual const char *getPassName() const { return "If Converter"; } private: bool ReverseBranchCondition(BBInfo &BBI); bool ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const; bool ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, bool FalseBranch, unsigned &Dups) const; - bool ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI) const; + bool ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI, + unsigned &Dups1, unsigned &Dups2) const; void ScanInstructions(BBInfo &BBI); BBInfo &AnalyzeBlock(MachineBasicBlock *BB, std::vector &Tokens); - bool FeasibilityAnalysis(BBInfo &BBI, std::vector &Cond, + bool FeasibilityAnalysis(BBInfo &BBI, SmallVectorImpl &Cond, bool isTriangle = false, bool RevBranch = false); bool AnalyzeBlocks(MachineFunction &MF, std::vector &Tokens); - void ReTryPreds(MachineBasicBlock *BB); + void InvalidatePreds(MachineBasicBlock *BB); void RemoveExtraEdges(BBInfo &BBI); bool IfConvertSimple(BBInfo &BBI, IfcvtKind Kind); bool IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind); - bool IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind); + bool IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, + unsigned NumDups1, unsigned NumDups2); void PredicateBlock(BBInfo &BBI, - std::vector &Cond, - bool IgnoreTerm = false); + MachineBasicBlock::iterator E, + SmallVectorImpl &Cond); void CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, - std::vector &Cond, + SmallVectorImpl &Cond, bool IgnoreBr = false); void MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI); + bool MeetIfcvtSizeLimit(unsigned Size) const { + return Size > 0 && Size <= TLI->getIfCvtBlockSizeLimit(); + } + // blockAlwaysFallThrough - Block ends without a terminator. bool blockAlwaysFallThrough(BBInfo &BBI) const { return BBI.IsBrAnalyzable && BBI.TrueBB == NULL; @@ -180,13 +190,17 @@ namespace { // IfcvtTokenCmp - Used to sort if-conversion candidates. static bool IfcvtTokenCmp(IfcvtToken *C1, IfcvtToken *C2) { - // Favors subsumsion. - if (C1->NeedSubsumsion == false && C2->NeedSubsumsion == true) + int Incr1 = (C1->Kind == ICDiamond) + ? -(int)(C1->NumDups + C1->NumDups2) : (int)C1->NumDups; + int Incr2 = (C2->Kind == ICDiamond) + ? -(int)(C2->NumDups + C2->NumDups2) : (int)C2->NumDups; + if (Incr1 > Incr2) return true; - else if (C1->NeedSubsumsion == C2->NeedSubsumsion) { - if (C1->Duplicates > C2->Duplicates) + else if (Incr1 == Incr2) { + // Favors subsumsion. + if (C1->NeedSubsumsion == false && C2->NeedSubsumsion == true) return true; - else if (C1->Duplicates == C2->Duplicates) { + else if (C1->NeedSubsumsion == C2->NeedSubsumsion) { // Favors diamond over triangle, etc. if ((unsigned)C1->Kind < (unsigned)C2->Kind) return true; @@ -201,6 +215,9 @@ namespace { char IfConverter::ID = 0; } +static RegisterPass +X("if-converter", "If Converter"); + FunctionPass *llvm::createIfConverterPass() { return new IfConverter(); } bool IfConverter::runOnMachineFunction(MachineFunction &MF) { @@ -223,7 +240,7 @@ bool IfConverter::runOnMachineFunction(MachineFunction &MF) { // Look for root nodes, i.e. blocks without successors. for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) - if (I->succ_size() == 0) + if (I->succ_empty()) Roots.push_back(I); std::vector Tokens; @@ -239,6 +256,10 @@ bool IfConverter::runOnMachineFunction(MachineFunction &MF) { Tokens.pop_back(); BBInfo &BBI = Token->BBI; IfcvtKind Kind = Token->Kind; + unsigned NumDups = Token->NumDups; + unsigned NumDups2 = Token->NumDups2; + + delete Token; // If the block has been evicted out of the queue or it has already been // marked dead (due to it being predicated), then skip it. @@ -264,9 +285,10 @@ bool IfConverter::runOnMachineFunction(MachineFunction &MF) { : BBI.TrueBB->getNumber()) << ") "; RetVal = IfConvertSimple(BBI, Kind); DOUT << (RetVal ? "succeeded!" : "failed!") << "\n"; - if (RetVal) + if (RetVal) { if (isFalse) NumSimpleFalse++; else NumSimple++; + } break; } case ICTriangle: @@ -305,7 +327,7 @@ bool IfConverter::runOnMachineFunction(MachineFunction &MF) { DOUT << "Ifcvt (Diamond): BB#" << BBI.BB->getNumber() << " (T:" << BBI.TrueBB->getNumber() << ",F:" << BBI.FalseBB->getNumber() << ") "; - RetVal = IfConvertDiamond(BBI, Kind); + RetVal = IfConvertDiamond(BBI, Kind, NumDups, NumDups2); DOUT << (RetVal ? "succeeded!" : "failed!") << "\n"; if (RetVal) NumDiamonds++; break; @@ -380,7 +402,10 @@ static inline MachineBasicBlock *getNextBlock(MachineBasicBlock *BB) { /// in Dups. bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const { Dups = 0; - if (TrueBBI.IsBeingAnalyzed) + if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) + return false; + + if (TrueBBI.IsBrAnalyzable) return false; if (TrueBBI.BB->pred_size() > 1) { @@ -390,7 +415,7 @@ bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const { Dups = TrueBBI.NonPredSize; } - return !blockAlwaysFallThrough(TrueBBI) && TrueBBI.BrCond.size() == 0; + return true; } /// ValidTriangle - Returns true if the 'true' and 'false' blocks (along @@ -402,7 +427,7 @@ bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const { bool IfConverter::ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, bool FalseBranch, unsigned &Dups) const { Dups = 0; - if (TrueBBI.IsBeingAnalyzed) + if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) return false; if (TrueBBI.BB->pred_size() > 1) { @@ -411,7 +436,7 @@ bool IfConverter::ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, unsigned Size = TrueBBI.NonPredSize; if (TrueBBI.IsBrAnalyzable) { - if (TrueBBI.TrueBB && TrueBBI.BrCond.size() == 0) + if (TrueBBI.TrueBB && TrueBBI.BrCond.empty()) // End with an unconditional branch. It will be removed. --Size; else { @@ -437,10 +462,25 @@ 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) const { - if (TrueBBI.IsBeingAnalyzed || FalseBBI.IsBeingAnalyzed) +bool IfConverter::ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI, + unsigned &Dups1, unsigned &Dups2) const { + Dups1 = Dups2 = 0; + if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone || + FalseBBI.IsBeingAnalyzed || FalseBBI.IsDone) return false; MachineBasicBlock *TT = TrueBBI.TrueBB; @@ -454,21 +494,47 @@ bool IfConverter::ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI) const { return false; if (TT == NULL && (TrueBBI.IsBrAnalyzable || FalseBBI.IsBrAnalyzable)) return false; - // FIXME: Allow false block to have an early exit? - return (TrueBBI.BB->pred_size() == 1 && - FalseBBI.BB->pred_size() == 1 && - !TrueBBI.FalseBB && !FalseBBI.FalseBB); + if (TrueBBI.BB->pred_size() > 1 || FalseBBI.BB->pred_size() > 1) + return false; + + // FIXME: Allow true block to have an early exit? + if (TrueBBI.FalseBB || FalseBBI.FalseBB || + (TrueBBI.ClobbersPred && FalseBBI.ClobbersPred)) + return false; + + MachineBasicBlock::iterator TI = TrueBBI.BB->begin(); + MachineBasicBlock::iterator FI = FalseBBI.BB->begin(); + while (TI != TrueBBI.BB->end() && FI != FalseBBI.BB->end()) { + if (!TI->isIdenticalTo(FI)) + break; + ++Dups1; + ++TI; + ++FI; + } + + TI = firstNonBranchInst(TrueBBI.BB, TII); + FI = firstNonBranchInst(FalseBBI.BB, TII); + while (TI != TrueBBI.BB->begin() && FI != FalseBBI.BB->begin()) { + if (!TI->isIdenticalTo(FI)) + break; + ++Dups2; + --TI; + --FI; + } + + return true; } /// ScanInstructions - Scan all the instructions in the block to determine if /// the block is predicable. In most cases, that means all the instructions -/// in the block has M_PREDICABLE flag. Also checks if the block contains any +/// in the block are isPredicable(). Also checks if the block contains any /// instruction which can clobber a predicate (e.g. condition code register). /// If so, the block is not predicable unless it's the last instruction. void IfConverter::ScanInstructions(BBInfo &BBI) { if (BBI.IsDone) return; + bool AlreadyPredicated = BBI.Predicate.size() > 0; // First analyze the end of BB branches. BBI.TrueBB = BBI.FalseBB = NULL; BBI.BrCond.clear(); @@ -490,16 +556,25 @@ void IfConverter::ScanInstructions(BBInfo &BBI) { bool SeenCondBr = false; for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end(); I != E; ++I) { - if (!BBI.CannotBeCopied && !TII->CanBeDuplicated(I)) + const TargetInstrDesc &TID = I->getDesc(); + if (TID.isNotDuplicable()) BBI.CannotBeCopied = true; - const TargetInstrDescriptor *TID = I->getInstrDescriptor(); bool isPredicated = TII->isPredicated(I); - bool isCondBr = BBI.IsBrAnalyzable && - (TID->Flags & M_BRANCH_FLAG) != 0 && (TID->Flags & M_BARRIER_FLAG) == 0; - - if (!isPredicated && !isCondBr) - BBI.NonPredSize++; + bool isCondBr = BBI.IsBrAnalyzable && TID.isConditionalBranch(); + + if (!isCondBr) { + if (!isPredicated) + BBI.NonPredSize++; + 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. + BBI.IsUnpredicable = true; + return; + } + + } if (BBI.ClobbersPred && !isPredicated) { // Predicate modification instruction should end the block (except for @@ -512,15 +587,18 @@ void IfConverter::ScanInstructions(BBInfo &BBI) { } // Predicate may have been modified, the subsequent (currently) - // unpredocated instructions cannot be correctly predicated. + // unpredicated instructions cannot be correctly predicated. BBI.IsUnpredicable = true; return; } - if (TID->Flags & M_CLOBBERS_PRED) + // FIXME: Make use of PredDefs? e.g. ADDC, SUBC sets predicates but are + // still potentially predicable. + std::vector PredDefs; + if (TII->DefinesPredicate(I, PredDefs)) BBI.ClobbersPred = true; - if ((TID->Flags & M_PREDICABLE) == 0) { + if (!TID.isPredicable()) { BBI.IsUnpredicable = true; return; } @@ -530,16 +608,12 @@ void IfConverter::ScanInstructions(BBInfo &BBI) { /// FeasibilityAnalysis - Determine if the block is a suitable candidate to be /// predicated by the specified predicate. bool IfConverter::FeasibilityAnalysis(BBInfo &BBI, - std::vector &Pred, + SmallVectorImpl &Pred, bool isTriangle, bool RevBranch) { // If the block is dead or unpredicable, then it cannot be predicated. if (BBI.IsDone || BBI.IsUnpredicable) return false; - // Check predication threshold. - if (BBI.NonPredSize == 0 || BBI.NonPredSize > TLI->getIfCvtBlockSizeLimit()) - return false; - // If it is already predicated, check if its predicate subsumes the new // predicate. if (BBI.Predicate.size() && !TII->SubsumesPredicate(BBI.Predicate, Pred)) @@ -550,8 +624,8 @@ bool IfConverter::FeasibilityAnalysis(BBInfo &BBI, return false; // Test predicate subsumsion. - std::vector RevPred(Pred); - std::vector Cond(BBI.BrCond); + SmallVector RevPred(Pred.begin(), Pred.end()); + SmallVector Cond(BBI.BrCond.begin(), BBI.BrCond.end()); if (RevBranch) { if (TII->ReverseBranchCondition(Cond)) return false; @@ -580,7 +654,7 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, ScanInstructions(BBI); // Unanalyable or ends with fallthrough or unconditional branch. - if (!BBI.IsBrAnalyzable || BBI.BrCond.size() == 0) { + if (!BBI.IsBrAnalyzable || BBI.BrCond.empty()) { BBI.IsBeingAnalyzed = false; BBI.IsAnalyzed = true; return BBI; @@ -602,15 +676,17 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, return BBI; } - std::vector RevCond(BBI.BrCond); + SmallVector RevCond(BBI.BrCond.begin(), BBI.BrCond.end()); bool CanRevCond = !TII->ReverseBranchCondition(RevCond); unsigned Dups = 0; + unsigned Dups2 = 0; bool TNeedSub = TrueBBI.Predicate.size() > 0; bool FNeedSub = FalseBBI.Predicate.size() > 0; bool Enqueued = false; - if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI) && - !(TrueBBI.ClobbersPred && FalseBBI.ClobbersPred) && + if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI, Dups, Dups2) && + MeetIfcvtSizeLimit(TrueBBI.NonPredSize - (Dups + Dups2)) && + MeetIfcvtSizeLimit(FalseBBI.NonPredSize - (Dups + Dups2)) && FeasibilityAnalysis(TrueBBI, BBI.BrCond) && FeasibilityAnalysis(FalseBBI, RevCond)) { // Diamond: @@ -621,11 +697,13 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, // \ / // TailBB // Note TailBB can be empty. - Tokens.push_back(new IfcvtToken(BBI, ICDiamond, TNeedSub|FNeedSub, Dups)); + Tokens.push_back(new IfcvtToken(BBI, ICDiamond, TNeedSub|FNeedSub, Dups, + Dups2)); Enqueued = true; } if (ValidTriangle(TrueBBI, FalseBBI, false, Dups) && + MeetIfcvtSizeLimit(TrueBBI.NonPredSize) && FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) { // Triangle: // EBB @@ -639,12 +717,14 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, } if (ValidTriangle(TrueBBI, FalseBBI, true, Dups) && + MeetIfcvtSizeLimit(TrueBBI.NonPredSize) && FeasibilityAnalysis(TrueBBI, BBI.BrCond, true, true)) { Tokens.push_back(new IfcvtToken(BBI, ICTriangleRev, TNeedSub, Dups)); Enqueued = true; } if (ValidSimple(TrueBBI, Dups) && + MeetIfcvtSizeLimit(TrueBBI.NonPredSize) && FeasibilityAnalysis(TrueBBI, BBI.BrCond)) { // Simple (split, no rejoin): // EBB @@ -660,18 +740,21 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, if (CanRevCond) { // Try the other path... if (ValidTriangle(FalseBBI, TrueBBI, false, Dups) && + MeetIfcvtSizeLimit(FalseBBI.NonPredSize) && FeasibilityAnalysis(FalseBBI, RevCond, true)) { Tokens.push_back(new IfcvtToken(BBI, ICTriangleFalse, FNeedSub, Dups)); Enqueued = true; } if (ValidTriangle(FalseBBI, TrueBBI, true, Dups) && + MeetIfcvtSizeLimit(FalseBBI.NonPredSize) && FeasibilityAnalysis(FalseBBI, RevCond, true, true)) { Tokens.push_back(new IfcvtToken(BBI, ICTriangleFRev, FNeedSub, Dups)); Enqueued = true; } if (ValidSimple(FalseBBI, Dups) && + MeetIfcvtSizeLimit(FalseBBI.NonPredSize) && FeasibilityAnalysis(FalseBBI, RevCond)) { Tokens.push_back(new IfcvtToken(BBI, ICSimpleFalse, FNeedSub, Dups)); Enqueued = true; @@ -718,10 +801,10 @@ static bool canFallThroughTo(MachineBasicBlock *BB, MachineBasicBlock *ToBB) { return true; } -/// ReTryPreds - Invalidate predecessor BB info so it would be re-analyzed -/// to determine if it can be if-converted. If predecessor is already -/// enqueud, dequeue it! -void IfConverter::ReTryPreds(MachineBasicBlock *BB) { +/// InvalidatePreds - Invalidate predecessor BB info so it would be re-analyzed +/// to determine if it can be if-converted. If predecessor is already enqueued, +/// dequeue it! +void IfConverter::InvalidatePreds(MachineBasicBlock *BB) { for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(), E = BB->pred_end(); PI != E; ++PI) { BBInfo &PBBI = BBAnalysis[(*PI)->getNumber()]; @@ -736,7 +819,7 @@ void IfConverter::ReTryPreds(MachineBasicBlock *BB) { /// static void InsertUncondBranch(MachineBasicBlock *BB, MachineBasicBlock *ToBB, const TargetInstrInfo *TII) { - std::vector NoCond; + SmallVector NoCond; TII->InsertBranch(*BB, ToBB, NULL, NoCond); } @@ -744,17 +827,9 @@ static void InsertUncondBranch(MachineBasicBlock *BB, MachineBasicBlock *ToBB, /// successors. void IfConverter::RemoveExtraEdges(BBInfo &BBI) { MachineBasicBlock *TBB = NULL, *FBB = NULL; - std::vector Cond; - bool isAnalyzable = !TII->AnalyzeBranch(*BBI.BB, TBB, FBB, Cond); - bool CanFallthrough = isAnalyzable && (TBB == NULL || FBB == NULL); - if (BBI.TrueBB && BBI.BB->isSuccessor(BBI.TrueBB)) - if (!(BBI.TrueBB == TBB || BBI.TrueBB == FBB || - (CanFallthrough && getNextBlock(BBI.BB) == BBI.TrueBB))) - BBI.BB->removeSuccessor(BBI.TrueBB); - if (BBI.FalseBB && BBI.BB->isSuccessor(BBI.FalseBB)) - if (!(BBI.FalseBB == TBB || BBI.FalseBB == FBB || - (CanFallthrough && getNextBlock(BBI.BB) == BBI.FalseBB))) - BBI.BB->removeSuccessor(BBI.FalseBB); + SmallVector Cond; + if (!TII->AnalyzeBranch(*BBI.BB, TBB, FBB, Cond)) + BBI.BB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); } /// IfConvertSimple - If convert a simple (split, no rejoin) sub-CFG. @@ -765,11 +840,12 @@ bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) { BBInfo *CvtBBI = &TrueBBI; BBInfo *NextBBI = &FalseBBI; - std::vector Cond(BBI.BrCond); + SmallVector Cond(BBI.BrCond.begin(), BBI.BrCond.end()); if (Kind == ICSimpleFalse) std::swap(CvtBBI, NextBBI); - if (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1) { + if (CvtBBI->IsDone || + (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) { // Something has changed. It's no longer safe to predicate this block. BBI.IsAnalyzed = false; CvtBBI->IsAnalyzed = false; @@ -777,7 +853,8 @@ bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) { } if (Kind == ICSimpleFalse) - TII->ReverseBranchCondition(Cond); + if (TII->ReverseBranchCondition(Cond)) + assert(false && "Unable to reverse branch condition!"); if (CvtBBI->BB->pred_size() > 1) { BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); @@ -785,7 +862,7 @@ bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) { // the entry block. CopyAndPredicateBlock(BBI, *CvtBBI, Cond); } else { - PredicateBlock(*CvtBBI, Cond); + PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond); // Merge converted block into entry block. BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); @@ -814,7 +891,7 @@ bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) { // Update block info. BB can be iteratively if-converted. if (!IterIfcvt) BBI.IsDone = true; - ReTryPreds(BBI.BB); + InvalidatePreds(BBI.BB); CvtBBI->IsDone = true; // FIXME: Must maintain LiveIns. @@ -829,11 +906,12 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { BBInfo *CvtBBI = &TrueBBI; BBInfo *NextBBI = &FalseBBI; - std::vector Cond(BBI.BrCond); + SmallVector Cond(BBI.BrCond.begin(), BBI.BrCond.end()); if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) std::swap(CvtBBI, NextBBI); - if (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1) { + if (CvtBBI->IsDone || + (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) { // Something has changed. It's no longer safe to predicate this block. BBI.IsAnalyzed = false; CvtBBI->IsAnalyzed = false; @@ -841,21 +919,23 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { } if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) - TII->ReverseBranchCondition(Cond); + if (TII->ReverseBranchCondition(Cond)) + assert(false && "Unable to reverse branch condition!"); if (Kind == ICTriangleRev || Kind == ICTriangleFRev) { - ReverseBranchCondition(*CvtBBI); - // BB has been changed, modify its predecessors (except for this - // one) so they don't get ifcvt'ed based on bad intel. - for (MachineBasicBlock::pred_iterator PI = CvtBBI->BB->pred_begin(), - E = CvtBBI->BB->pred_end(); PI != E; ++PI) { - MachineBasicBlock *PBB = *PI; - if (PBB == BBI.BB) - continue; - BBInfo &PBBI = BBAnalysis[PBB->getNumber()]; - if (PBBI.IsEnqueued) { - PBBI.IsAnalyzed = false; - PBBI.IsEnqueued = false; + if (ReverseBranchCondition(*CvtBBI)) { + // BB has been changed, modify its predecessors (except for this + // one) so they don't get ifcvt'ed based on bad intel. + for (MachineBasicBlock::pred_iterator PI = CvtBBI->BB->pred_begin(), + E = CvtBBI->BB->pred_end(); PI != E; ++PI) { + MachineBasicBlock *PBB = *PI; + if (PBB == BBI.BB) + continue; + BBInfo &PBBI = BBAnalysis[PBB->getNumber()]; + if (PBBI.IsEnqueued) { + PBBI.IsAnalyzed = false; + PBBI.IsEnqueued = false; + } } } } @@ -870,20 +950,7 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { } else { // Predicate the 'true' block after removing its branch. CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB); - PredicateBlock(*CvtBBI, Cond); - } - - // If 'true' block has a 'false' successor, add an exit branch to it. - if (HasEarlyExit) { - std::vector RevCond(CvtBBI->BrCond); - if (TII->ReverseBranchCondition(RevCond)) - assert(false && "Unable to reverse branch condition!"); - if (DupBB) { - TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, NULL, RevCond); - BBI.BB->addSuccessor(CvtBBI->FalseBB); - } else { - TII->InsertBranch(*CvtBBI->BB, CvtBBI->FalseBB, NULL, RevCond); - } + PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond); } if (!DupBB) { @@ -892,6 +959,16 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { MergeBlocks(BBI, *CvtBBI); } + // If 'true' block has a 'false' successor, add an exit branch to it. + if (HasEarlyExit) { + SmallVector RevCond(CvtBBI->BrCond.begin(), + CvtBBI->BrCond.end()); + if (TII->ReverseBranchCondition(RevCond)) + assert(false && "Unable to reverse branch condition!"); + TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, NULL, RevCond); + BBI.BB->addSuccessor(CvtBBI->FalseBB); + } + // Merge in the 'false' block if the 'false' block has no other // predecessors. Otherwise, add a unconditional branch from to 'false'. bool FalseBBDead = false; @@ -901,7 +978,8 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { // Only merge them if the true block does not fallthrough to the false // block. By not merging them, we make it possible to iteratively // ifcvt the blocks. - if (!HasEarlyExit && NextBBI->BB->pred_size() == 1) { + if (!HasEarlyExit && + NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough) { MergeBlocks(BBI, *NextBBI); FalseBBDead = true; } else { @@ -918,7 +996,7 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { // Update block info. BB can be iteratively if-converted. if (!IterIfcvt) BBI.IsDone = true; - ReTryPreds(BBI.BB); + InvalidatePreds(BBI.BB); CvtBBI->IsDone = true; if (FalseBBDead) NextBBI->IsDone = true; @@ -929,149 +1007,121 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { /// IfConvertDiamond - If convert a diamond sub-CFG. /// -bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind) { +bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, + unsigned NumDups1, unsigned NumDups2) { BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; MachineBasicBlock *TailBB = TrueBBI.TrueBB; - - SmallVector Dups; + // True block must fall through or ended with unanalyzable terminator. if (!TailBB) { - // No common merge block. Check if the terminators (e.g. return) are - // the same or predicable. - MachineBasicBlock::iterator TT = BBI.TrueBB->getFirstTerminator(); - MachineBasicBlock::iterator FT = BBI.FalseBB->getFirstTerminator(); - while (TT != BBI.TrueBB->end() && FT != BBI.FalseBB->end()) { - if (TT->isIdenticalTo(FT)) - Dups.push_back(TT); // Will erase these later. - else if ((TT->getInstrDescriptor()->Flags & M_PREDICABLE) == 0 || - (FT->getInstrDescriptor()->Flags & M_PREDICABLE) == 0) - return false; // Can't if-convert. Abort! - ++TT; - ++FT; - } - - // One of the two pathes have more terminators, make sure they are - // all predicable. - while (TT != BBI.TrueBB->end()) { - if ((TT->getInstrDescriptor()->Flags & M_PREDICABLE) == 0) - return false; // Can't if-convert. Abort! - ++TT; - } - while (FT != BBI.FalseBB->end()) { - if ((FT->getInstrDescriptor()->Flags & M_PREDICABLE) == 0) - return false; // Can't if-convert. Abort! - ++FT; - } + if (blockAlwaysFallThrough(TrueBBI)) + TailBB = FalseBBI.TrueBB; + assert((TailBB || !TrueBBI.IsBrAnalyzable) && "Unexpected!"); } - // Remove the duplicated instructions from the 'true' block. - for (unsigned i = 0, e = Dups.size(); i != e; ++i) { - Dups[i]->eraseFromParent(); - --TrueBBI.NonPredSize; + if (TrueBBI.IsDone || FalseBBI.IsDone || + TrueBBI.BB->pred_size() > 1 || + FalseBBI.BB->pred_size() > 1) { + // Something has changed. It's no longer safe to predicate these blocks. + BBI.IsAnalyzed = false; + TrueBBI.IsAnalyzed = false; + FalseBBI.IsAnalyzed = false; + 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. BBInfo *BBI1 = &TrueBBI; BBInfo *BBI2 = &FalseBBI; - std::vector RevCond(BBI.BrCond); - TII->ReverseBranchCondition(RevCond); - std::vector *Cond1 = &BBI.BrCond; - std::vector *Cond2 = &RevCond; - // Check the 'true' and 'false' blocks if either isn't ended with a branch. - // Either the block fallthrough to another block or it ends with a - // return. If it's the former, add a branch to its successor. - bool NeedBr1 = !BBI1->TrueBB && BBI1->BB->succ_size() > 0; - bool NeedBr2 = !BBI2->TrueBB && BBI2->BB->succ_size() > 0; + SmallVector RevCond(BBI.BrCond.begin(), BBI.BrCond.end()); + if (TII->ReverseBranchCondition(RevCond)) + assert(false && "Unable to reverse branch condition!"); + SmallVector *Cond1 = &BBI.BrCond; + SmallVector *Cond2 = &RevCond; // Figure out the more profitable ordering. bool DoSwap = false; if (TrueBBI.ClobbersPred && !FalseBBI.ClobbersPred) DoSwap = true; else if (TrueBBI.ClobbersPred == FalseBBI.ClobbersPred) { - if (!NeedBr1 && NeedBr2) + if (TrueBBI.NonPredSize > FalseBBI.NonPredSize) DoSwap = true; - else if (NeedBr1 == NeedBr2) { - if (TrueBBI.NonPredSize > FalseBBI.NonPredSize) - DoSwap = true; - } } if (DoSwap) { std::swap(BBI1, BBI2); std::swap(Cond1, Cond2); - std::swap(NeedBr1, NeedBr2); } + // Remove the conditional branch from entry to the blocks. + BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); + + // Remove the duplicated instructions at the beginnings of both paths. + MachineBasicBlock::iterator DI1 = BBI1->BB->begin(); + MachineBasicBlock::iterator DI2 = BBI2->BB->begin(); + BBI1->NonPredSize -= NumDups1; + BBI2->NonPredSize -= NumDups1; + while (NumDups1 != 0) { + ++DI1; + ++DI2; + --NumDups1; + } + 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. BBI1->NonPredSize -= TII->RemoveBranch(*BBI1->BB); - PredicateBlock(*BBI1, *Cond1); - - // Add an early exit branch if needed. - if (NeedBr1) - TII->InsertBranch(*BBI1->BB, *BBI1->BB->succ_begin(), NULL, *Cond1); + DI1 = BBI1->BB->end(); + for (unsigned i = 0; i != NumDups2; ++i) + --DI1; + BBI1->BB->erase(DI1, BBI1->BB->end()); + PredicateBlock(*BBI1, BBI1->BB->end(), *Cond1); // Predicate the 'false' block. - PredicateBlock(*BBI2, *Cond2, true); - - // Add an unconditional branch from 'false' to to 'false' successor if it - // will not be the fallthrough block. - if (NeedBr2 && !NeedBr1) { - // If BBI2 isn't going to be merged in, then the existing fallthrough - // or branch is fine. - if (!canFallThroughTo(BBI.BB, *BBI2->BB->succ_begin())) { - InsertUncondBranch(BBI2->BB, *BBI2->BB->succ_begin(), TII); - BBI2->HasFallThrough = false; - } + BBI2->NonPredSize -= TII->RemoveBranch(*BBI2->BB); + DI2 = BBI2->BB->end(); + while (NumDups2 != 0) { + --DI2; + --NumDups2; } + PredicateBlock(*BBI2, DI2, *Cond2); - // Keep them as two separate blocks if there is an early exit. - if (!NeedBr1) - MergeBlocks(*BBI1, *BBI2); - - // Remove the conditional branch from entry to the blocks. - BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); - - // Merge the combined block into the entry of the diamond. + // Merge the true block into the entry of the diamond. MergeBlocks(BBI, *BBI1); - - // 'True' and 'false' aren't combined, see if we need to add a unconditional - // branch to the 'false' block. - if (NeedBr1 && !canFallThroughTo(BBI.BB, BBI2->BB)) { - InsertUncondBranch(BBI.BB, BBI2->BB, TII); - BBI1->HasFallThrough = false; - } + MergeBlocks(BBI, *BBI2); // If the if-converted block fallthrough or unconditionally branch into the // tail block, and the tail block does not have other predecessors, then - // fold the tail block in as well. - BBInfo *CvtBBI = NeedBr1 ? BBI2 : &BBI; - if (TailBB && - TailBB->pred_size() == 1 && CvtBBI->BB->succ_size() == 1) { - CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB); + // 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()]; - MergeBlocks(*CvtBBI, TailBBI); - TailBBI.IsDone = true; + if (TailBB->pred_size() == 1 && !TailBBI.HasFallThrough) { + BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); + MergeBlocks(BBI, TailBBI); + TailBBI.IsDone = true; + } else { + InsertUncondBranch(BBI.BB, TailBB, TII); + BBI.HasFallThrough = false; + } } RemoveExtraEdges(BBI); // Update block info. BBI.IsDone = TrueBBI.IsDone = FalseBBI.IsDone = true; + InvalidatePreds(BBI.BB); // FIXME: Must maintain LiveIns. return true; } -/// PredicateBlock - Predicate every instruction in the block with the specified -/// condition. If IgnoreTerm is true, skip over all terminator instructions. +/// PredicateBlock - Predicate instructions from the start of the block to the +/// specified end with the specified condition. void IfConverter::PredicateBlock(BBInfo &BBI, - std::vector &Cond, - bool IgnoreTerm) { - for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end(); - I != E; ++I) { - if (IgnoreTerm && TII->isTerminatorInstr(I->getOpcode())) - continue; + MachineBasicBlock::iterator E, + SmallVectorImpl &Cond) { + for (MachineBasicBlock::iterator I = BBI.BB->begin(); I != E; ++I) { if (TII->isPredicated(I)) continue; if (!TII->PredicateInstruction(I, Cond)) { @@ -1091,17 +1141,19 @@ void IfConverter::PredicateBlock(BBInfo &BBI, /// CopyAndPredicateBlock - Copy and predicate instructions from source BB to /// the destination block. Skip end of block branches if IgnoreBr is true. void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, - std::vector &Cond, + SmallVectorImpl &Cond, bool IgnoreBr) { + MachineFunction &MF = *ToBBI.BB->getParent(); + for (MachineBasicBlock::iterator I = FromBBI.BB->begin(), E = FromBBI.BB->end(); I != E; ++I) { - const TargetInstrDescriptor *TID = I->getInstrDescriptor(); + const TargetInstrDesc &TID = I->getDesc(); bool isPredicated = TII->isPredicated(I); // Do not copy the end of the block branches. - if (IgnoreBr && !isPredicated && (TID->Flags & M_BRANCH_FLAG) != 0) + if (IgnoreBr && !isPredicated && TID.isBranch()) break; - MachineInstr *MI = I->clone(); + MachineInstr *MI = MF.CloneMachineInstr(I); ToBBI.BB->insert(ToBBI.BB->end(), MI); ToBBI.NonPredSize++; @@ -1112,6 +1164,20 @@ void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, } } + 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; + if (!ToBBI.BB->isSuccessor(Succ)) + ToBBI.BB->addSuccessor(Succ); + } + std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(), std::back_inserter(ToBBI.Predicate)); std::copy(Cond.begin(), Cond.end(), std::back_inserter(ToBBI.Predicate)); @@ -1154,7 +1220,7 @@ void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI) { } // Now FromBBI always fall through to the next block! - if (NBB) + if (NBB && !FromBBI.BB->isSuccessor(NBB)) FromBBI.BB->addSuccessor(NBB); std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(),