X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FIfConversion.cpp;h=8d212a606fdf305012b13e8e20decadcb19923c0;hb=b6e223a9e806921183da972253c49082a2e07944;hp=1302ec2cd2166ad4e498a096e1bd2d41b6c6073f;hpb=7e75ba802e4e5df9fe8f0d38a85e05adc05ea7f3;p=oota-llvm.git diff --git a/lib/CodeGen/IfConversion.cpp b/lib/CodeGen/IfConversion.cpp index 1302ec2cd21..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. // //===----------------------------------------------------------------------===// // @@ -23,41 +23,49 @@ #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 DisableSimpleFalse("disable-ifcvt-simple-false", +// 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); - cl::opt DisableTriangle("disable-ifcvt-triangle", - cl::init(false), cl::Hidden); - cl::opt DisableDiamond("disable-ifcvt-diamond", - cl::init(false), cl::Hidden); -} - -STATISTIC(NumSimple, "Number of simple if-conversions performed"); -STATISTIC(NumSimpleRev, "Number of simple (reversed) if-conversions performed"); -STATISTIC(NumTriangle, "Number of triangle if-conversions performed"); -STATISTIC(NumDiamonds, "Number of diamond if-conversions performed"); -STATISTIC(NumIfConvBBs, "Number of if-converted blocks"); +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"); +STATISTIC(NumTriangle, "Number of triangle if-conversions performed"); +STATISTIC(NumTriangleRev, "Number of triangle (R) if-conversions performed"); +STATISTIC(NumTriangleFalse,"Number of triangle (F) if-conversions performed"); +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"); namespace { - class IfConverter : public MachineFunctionPass { - enum BBICKind { - ICNotAnalyzed, // BB has not been analyzed. - ICReAnalyze, // BB must be re-analyzed. + class VISIBILITY_HIDDEN IfConverter : public MachineFunctionPass { + enum IfcvtKind { ICNotClassfied, // BB data valid, but not classified. - ICSimple, // BB is entry of an one split, no rejoin sub-CFG. ICSimpleFalse, // Same as ICSimple, but on the false path. + ICSimple, // BB is entry of an one split, no rejoin sub-CFG. + ICTriangleFRev, // Same as ICTriangleFalse, but false path rev condition. + ICTriangleRev, // Same as ICTriangle, but true path rev condition. + ICTriangleFalse, // Same as ICTriangle, but on the false path. ICTriangle, // BB is entry of a triangle sub-CFG. - ICDiamond, // BB is entry of a diamond sub-CFG. - ICChild, // BB is part of the sub-CFG that'll be predicated. - ICDead // BB cannot be if-converted again. + ICDiamond // BB is entry of a diamond sub-CFG. }; /// BBInfo - One per MachineBasicBlock, this is used to cache the result @@ -67,31 +75,62 @@ namespace { /// diamond shape), its size, whether it's predicable, and whether any /// instruction can clobber the 'would-be' predicate. /// - /// Kind - Type of block. See BBICKind. - /// NonPredSize - Number of non-predicated instructions. - /// IsAnalyzable - True if AnalyzeBranch() returns false. - /// ModifyPredicate - True if BB would modify the predicate (e.g. has + /// IsDone - True if BB is not to be considered for ifcvt. + /// IsBeingAnalyzed - True if BB is currently being analyzed. + /// IsAnalyzed - True if BB has been analyzed (info is still valid). + /// IsEnqueued - True if BB has been enqueued to be ifcvt'ed. + /// 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. + /// ClobbersPred - True if BB could modify predicates (e.g. has /// cmp, call, etc.) + /// NonPredSize - Number of non-predicated instructions. /// BB - Corresponding MachineBasicBlock. /// TrueBB / FalseBB- See AnalyzeBranch(). /// BrCond - Conditions for end of block conditional branches. /// Predicate - Predicate used in the BB. struct BBInfo { - BBICKind Kind; + bool IsDone : 1; + bool IsBeingAnalyzed : 1; + bool IsAnalyzed : 1; + bool IsEnqueued : 1; + bool IsBrAnalyzable : 1; + bool HasFallThrough : 1; + bool IsUnpredicable : 1; + bool CannotBeCopied : 1; + bool ClobbersPred : 1; unsigned NonPredSize; - bool IsAnalyzable; - bool hasFallThrough; - bool ModifyPredicate; MachineBasicBlock *BB; MachineBasicBlock *TrueBB; MachineBasicBlock *FalseBB; - MachineBasicBlock *TailBB; - std::vector BrCond; - std::vector Predicate; - BBInfo() : Kind(ICNotAnalyzed), NonPredSize(0), - IsAnalyzable(false), hasFallThrough(false), - ModifyPredicate(false), - BB(0), TrueBB(0), FalseBB(0), TailBB(0) {} + SmallVector BrCond; + SmallVector Predicate; + BBInfo() : IsDone(false), IsBeingAnalyzed(false), + IsAnalyzed(false), IsEnqueued(false), IsBrAnalyzable(false), + HasFallThrough(false), IsUnpredicable(false), + CannotBeCopied(false), ClobbersPred(false), NonPredSize(0), + BB(0), TrueBB(0), FalseBB(0) {} + }; + + /// IfcvtToken - Record information about pending if-conversions to attemp: + /// BBI - Corresponding BBInfo. + /// Kind - Type of block. See IfcvtKind. + /// NeedSubsumsion - True if the to be predicated BB has already been + /// predicated. + /// 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 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 @@ -107,48 +146,78 @@ 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) const; + bool ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const; bool ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, - bool FalseBranch = false) const; - bool ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI) const; + bool FalseBranch, unsigned &Dups) const; + bool ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI, + unsigned &Dups1, unsigned &Dups2) const; void ScanInstructions(BBInfo &BBI); - void AnalyzeBlock(MachineBasicBlock *BB); - bool FeasibilityAnalysis(BBInfo &BBI, std::vector &Cond, + BBInfo &AnalyzeBlock(MachineBasicBlock *BB, + std::vector &Tokens); + bool FeasibilityAnalysis(BBInfo &BBI, SmallVectorImpl &Cond, bool isTriangle = false, bool RevBranch = false); - bool AttemptRestructuring(BBInfo &BBI); bool AnalyzeBlocks(MachineFunction &MF, - std::vector &Candidates); - void ReTryPreds(MachineBasicBlock *BB); + std::vector &Tokens); + void InvalidatePreds(MachineBasicBlock *BB); void RemoveExtraEdges(BBInfo &BBI); - bool IfConvertSimple(BBInfo &BBI); - bool IfConvertTriangle(BBInfo &BBI); - bool IfConvertDiamond(BBInfo &BBI); + bool IfConvertSimple(BBInfo &BBI, IfcvtKind Kind); + bool IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind); + bool IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, + unsigned NumDups1, unsigned NumDups2); void PredicateBlock(BBInfo &BBI, - std::vector &Cond, - bool IgnoreTerm = false); - void MergeBlocks(BBInfo &TrueBBI, BBInfo &FalseBBI); + MachineBasicBlock::iterator E, + SmallVectorImpl &Cond); + void CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, + 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.IsAnalyzable && BBI.TrueBB == NULL; + return BBI.IsBrAnalyzable && BBI.TrueBB == NULL; } - // IfcvtCandidateCmp - Used to sort if-conversion candidates. - static bool IfcvtCandidateCmp(BBInfo* C1, BBInfo* C2){ - // Favor diamond over triangle, etc. - return (unsigned)C1->Kind < (unsigned)C2->Kind; + // IfcvtTokenCmp - Used to sort if-conversion candidates. + static bool IfcvtTokenCmp(IfcvtToken *C1, IfcvtToken *C2) { + 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 (Incr1 == Incr2) { + // Favors subsumsion. + if (C1->NeedSubsumsion == false && C2->NeedSubsumsion == true) + return true; + else if (C1->NeedSubsumsion == C2->NeedSubsumsion) { + // Favors diamond over triangle, etc. + if ((unsigned)C1->Kind < (unsigned)C2->Kind) + return true; + else if (C1->Kind == C2->Kind) + return C1->BBI.BB->getNumber() < C2->BBI.BB->getNumber(); + } + } + return false; } }; + char IfConverter::ID = 0; } +static RegisterPass +X("if-converter", "If Converter"); + FunctionPass *llvm::createIfConverterPass() { return new IfConverter(); } bool IfConverter::runOnMachineFunction(MachineFunction &MF) { @@ -171,67 +240,105 @@ 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 Candidates; + std::vector Tokens; MadeChange = false; - while (IfCvtLimit == -1 || (int)NumIfConvBBs < IfCvtLimit) { + unsigned NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + + NumTriangleRev + NumTriangleFalse + NumTriangleFRev + NumDiamonds; + while (IfCvtLimit == -1 || (int)NumIfCvts < IfCvtLimit) { // Do an intial analysis for each basic block and finding all the potential // candidates to perform if-convesion. - bool Change = AnalyzeBlocks(MF, Candidates); - while (!Candidates.empty()) { - BBInfo &BBI = *Candidates.back(); - Candidates.pop_back(); + bool Change = AnalyzeBlocks(MF, Tokens); + while (!Tokens.empty()) { + IfcvtToken *Token = Tokens.back(); + 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. + if (BBI.IsDone) + BBI.IsEnqueued = false; + if (!BBI.IsEnqueued) + continue; + + BBI.IsEnqueued = false; bool RetVal = false; - switch (BBI.Kind) { + switch (Kind) { default: assert(false && "Unexpected!"); break; - case ICReAnalyze: - // One or more of 'children' have been modified, abort! - case ICDead: - // Block has been already been if-converted, abort! - break; case ICSimple: case ICSimpleFalse: { - bool isRev = BBI.Kind == ICSimpleFalse; - if ((isRev && DisableSimpleFalse) || (!isRev && DisableSimple)) break; - DOUT << "Ifcvt (Simple" << (BBI.Kind == ICSimpleFalse ? " false" : "") + bool isFalse = Kind == ICSimpleFalse; + if ((isFalse && DisableSimpleF) || (!isFalse && DisableSimple)) break; + DOUT << "Ifcvt (Simple" << (Kind == ICSimpleFalse ? " false" :"") << "): BB#" << BBI.BB->getNumber() << " (" - << ((BBI.Kind == ICSimpleFalse) - ? BBI.FalseBB->getNumber() : BBI.TrueBB->getNumber()) << ") "; - RetVal = IfConvertSimple(BBI); + << ((Kind == ICSimpleFalse) + ? BBI.FalseBB->getNumber() + : BBI.TrueBB->getNumber()) << ") "; + RetVal = IfConvertSimple(BBI, Kind); DOUT << (RetVal ? "succeeded!" : "failed!") << "\n"; - if (RetVal) - if (isRev) NumSimpleRev++; - else NumSimple++; + if (RetVal) { + if (isFalse) NumSimpleFalse++; + else NumSimple++; + } break; } case ICTriangle: - if (DisableTriangle) break; - DOUT << "Ifcvt (Triangle): BB#" << BBI.BB->getNumber() << " (T:" - << BBI.TrueBB->getNumber() << ",F:" << BBI.FalseBB->getNumber() - << ") "; - RetVal = IfConvertTriangle(BBI); + case ICTriangleRev: + case ICTriangleFalse: + case ICTriangleFRev: { + bool isFalse = Kind == ICTriangleFalse; + bool isRev = (Kind == ICTriangleRev || Kind == ICTriangleFRev); + if (DisableTriangle && !isFalse && !isRev) break; + if (DisableTriangleR && !isFalse && isRev) break; + if (DisableTriangleF && isFalse && !isRev) break; + if (DisableTriangleFR && isFalse && isRev) break; + DOUT << "Ifcvt (Triangle"; + if (isFalse) + DOUT << " false"; + if (isRev) + DOUT << " rev"; + DOUT << "): BB#" << BBI.BB->getNumber() << " (T:" + << BBI.TrueBB->getNumber() << ",F:" + << BBI.FalseBB->getNumber() << ") "; + RetVal = IfConvertTriangle(BBI, Kind); DOUT << (RetVal ? "succeeded!" : "failed!") << "\n"; - if (RetVal) NumTriangle++; + if (RetVal) { + if (isFalse) { + if (isRev) NumTriangleFRev++; + else NumTriangleFalse++; + } else { + if (isRev) NumTriangleRev++; + else NumTriangle++; + } + } break; - case ICDiamond: + } + case ICDiamond: { if (DisableDiamond) break; DOUT << "Ifcvt (Diamond): BB#" << BBI.BB->getNumber() << " (T:" - << BBI.TrueBB->getNumber() << ",F:" << BBI.FalseBB->getNumber(); - if (BBI.TailBB) - DOUT << "," << BBI.TailBB->getNumber() ; - DOUT << ") "; - RetVal = IfConvertDiamond(BBI); + << BBI.TrueBB->getNumber() << ",F:" + << BBI.FalseBB->getNumber() << ") "; + RetVal = IfConvertDiamond(BBI, Kind, NumDups, NumDups2); DOUT << (RetVal ? "succeeded!" : "failed!") << "\n"; if (RetVal) NumDiamonds++; break; } + } + Change |= RetVal; - if (IfCvtLimit != -1 && (int)NumIfConvBBs > IfCvtLimit) + NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + NumTriangleRev + + NumTriangleFalse + NumTriangleFRev + NumDiamonds; + if (IfCvtLimit != -1 && (int)NumIfCvts >= IfCvtLimit) break; } @@ -240,12 +347,22 @@ bool IfConverter::runOnMachineFunction(MachineFunction &MF) { MadeChange |= Change; } + // Delete tokens in case of early exit. + while (!Tokens.empty()) { + IfcvtToken *Token = Tokens.back(); + Tokens.pop_back(); + delete Token; + } + + Tokens.clear(); Roots.clear(); BBAnalysis.clear(); return MadeChange; } +/// findFalseBlock - BB has a fallthrough. Find its 'false' successor given +/// its 'true' successor. static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB, MachineBasicBlock *TrueBB) { for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), @@ -257,6 +374,8 @@ static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB, return NULL; } +/// ReverseBranchCondition - Reverse the condition of the end of the block +/// branchs. Swap block's 'true' and 'false' successors. bool IfConverter::ReverseBranchCondition(BBInfo &BBI) { if (!TII->ReverseBranchCondition(BBI.BrCond)) { TII->RemoveBranch(*BBI.BB); @@ -267,20 +386,72 @@ bool IfConverter::ReverseBranchCondition(BBInfo &BBI) { return false; } +/// getNextBlock - Returns the next block in the function blocks ordering. If +/// it is the end, returns NULL. +static inline MachineBasicBlock *getNextBlock(MachineBasicBlock *BB) { + MachineFunction::iterator I = BB; + MachineFunction::iterator E = BB->getParent()->end(); + if (++I == E) + return NULL; + return I; +} + /// ValidSimple - Returns true if the 'true' block (along with its -/// predecessor) forms a valid simple shape for ifcvt. -bool IfConverter::ValidSimple(BBInfo &TrueBBI) const { - return !blockAlwaysFallThrough(TrueBBI) && - TrueBBI.BrCond.size() == 0 && TrueBBI.BB->pred_size() == 1; +/// 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 { + Dups = 0; + if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) + return false; + + if (TrueBBI.IsBrAnalyzable) + return false; + + if (TrueBBI.BB->pred_size() > 1) { + if (TrueBBI.CannotBeCopied || + TrueBBI.NonPredSize > TLI->getIfCvtDupBlockSizeLimit()) + return false; + Dups = TrueBBI.NonPredSize; + } + + return true; } /// ValidTriangle - Returns true if the 'true' and 'false' blocks (along /// with their common predecessor) forms a valid triangle shape for ifcvt. +/// If 'FalseBranch' is true, it checks if 'true' block's false branch +/// branches to the false branch rather than the other way around. It also +/// 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) const { - if (TrueBBI.BB->pred_size() != 1) + bool FalseBranch, unsigned &Dups) const { + Dups = 0; + if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) return false; + if (TrueBBI.BB->pred_size() > 1) { + if (TrueBBI.CannotBeCopied) + return false; + + unsigned Size = TrueBBI.NonPredSize; + if (TrueBBI.IsBrAnalyzable) { + if (TrueBBI.TrueBB && TrueBBI.BrCond.empty()) + // End with an unconditional branch. It will be removed. + --Size; + else { + MachineBasicBlock *FExit = FalseBranch + ? TrueBBI.TrueBB : TrueBBI.FalseBB; + if (FExit) + // Require a conditional branch + ++Size; + } + } + if (Size > TLI->getIfCvtDupBlockSizeLimit()) + return false; + Dups = Size; + } + MachineBasicBlock *TExit = FalseBranch ? TrueBBI.FalseBB : TrueBBI.TrueBB; if (!TExit && blockAlwaysFallThrough(TrueBBI)) { MachineFunction::iterator I = TrueBBI.BB; @@ -291,69 +462,231 @@ 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 { - // FIXME: Also look for fallthrough - return (TrueBBI.TrueBB == FalseBBI.TrueBB && - TrueBBI.BB->pred_size() == 1 && - FalseBBI.BB->pred_size() == 1 && - !TrueBBI.FalseBB && !FalseBBI.FalseBB); +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; + MachineBasicBlock *FT = FalseBBI.TrueBB; + + if (!TT && blockAlwaysFallThrough(TrueBBI)) + TT = getNextBlock(TrueBBI.BB); + if (!FT && blockAlwaysFallThrough(FalseBBI)) + FT = getNextBlock(FalseBBI.BB); + if (TT != FT) + return false; + if (TT == NULL && (TrueBBI.IsBrAnalyzable || FalseBBI.IsBrAnalyzable)) + return false; + 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 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(); + BBI.IsBrAnalyzable = + !TII->AnalyzeBranch(*BBI.BB, BBI.TrueBB, BBI.FalseBB, BBI.BrCond); + BBI.HasFallThrough = BBI.IsBrAnalyzable && BBI.FalseBB == NULL; + + if (BBI.BrCond.size()) { + // No false branch. This BB must end with a conditional branch and a + // fallthrough. + if (!BBI.FalseBB) + BBI.FalseBB = findFalseBlock(BBI.BB, BBI.TrueBB); + assert(BBI.FalseBB && "Expected to find the fallthrough block!"); + } + + // Then scan all the instructions. + BBI.NonPredSize = 0; + BBI.ClobbersPred = false; + bool SeenCondBr = false; + for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end(); + I != E; ++I) { + const TargetInstrDesc &TID = I->getDesc(); + if (TID.isNotDuplicable()) + BBI.CannotBeCopied = true; + + bool isPredicated = TII->isPredicated(I); + 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 + // already predicated instructions and end of block branches). + if (isCondBr) { + SeenCondBr = true; + + // Conditional branches is not predicable. But it may be eliminated. + continue; + } + + // Predicate may have been modified, the subsequent (currently) + // unpredicated instructions cannot be correctly predicated. + BBI.IsUnpredicable = true; + return; + } + + // 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.isPredicable()) { + BBI.IsUnpredicable = true; + return; + } + } +} + +/// FeasibilityAnalysis - Determine if the block is a suitable candidate to be +/// predicated by the specified predicate. +bool IfConverter::FeasibilityAnalysis(BBInfo &BBI, + 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; + + // If it is already predicated, check if its predicate subsumes the new + // predicate. + if (BBI.Predicate.size() && !TII->SubsumesPredicate(BBI.Predicate, Pred)) + return false; + + if (BBI.BrCond.size()) { + if (!isTriangle) + return false; + + // Test predicate subsumsion. + SmallVector RevPred(Pred.begin(), Pred.end()); + SmallVector Cond(BBI.BrCond.begin(), BBI.BrCond.end()); + if (RevBranch) { + if (TII->ReverseBranchCondition(Cond)) + return false; + } + if (TII->ReverseBranchCondition(RevPred) || + !TII->SubsumesPredicate(Cond, RevPred)) + return false; + } + + return true; } /// AnalyzeBlock - Analyze the structure of the sub-CFG starting from /// the specified block. Record its successors and whether it looks like an /// if-conversion candidate. -void IfConverter::AnalyzeBlock(MachineBasicBlock *BB) { +IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, + std::vector &Tokens) { BBInfo &BBI = BBAnalysis[BB->getNumber()]; - if (BBI.Kind == ICReAnalyze) { - BBI.BrCond.clear(); - BBI.TrueBB = BBI.FalseBB = NULL; - } else { - if (BBI.Kind != ICNotAnalyzed) - return; // Already analyzed. - BBI.BB = BB; - BBI.NonPredSize = std::distance(BB->begin(), BB->end()); - } + if (BBI.IsAnalyzed || BBI.IsBeingAnalyzed) + return BBI; - // Look for 'root' of a simple (non-nested) triangle or diamond. - BBI.Kind = ICNotClassfied; - BBI.IsAnalyzable = - !TII->AnalyzeBranch(*BB, BBI.TrueBB, BBI.FalseBB, BBI.BrCond); - BBI.hasFallThrough = BBI.IsAnalyzable && BBI.FalseBB == NULL; - // Unanalyable or ends with fallthrough or unconditional branch. - if (!BBI.IsAnalyzable || BBI.BrCond.size() == 0) - return; - // Do not ifcvt if either path is a back edge to the entry block. - if (BBI.TrueBB == BB || BBI.FalseBB == BB) - return; + BBI.BB = BB; + BBI.IsBeingAnalyzed = true; - AnalyzeBlock(BBI.TrueBB); - BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; + ScanInstructions(BBI); - // No false branch. This BB must end with a conditional branch and a - // fallthrough. - if (!BBI.FalseBB) - BBI.FalseBB = findFalseBlock(BB, BBI.TrueBB); - assert(BBI.FalseBB && "Expected to find the fallthrough block!"); + // Unanalyable or ends with fallthrough or unconditional branch. + if (!BBI.IsBrAnalyzable || BBI.BrCond.empty()) { + BBI.IsBeingAnalyzed = false; + BBI.IsAnalyzed = true; + return BBI; + } - AnalyzeBlock(BBI.FalseBB); - BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; + // Do not ifcvt if either path is a back edge to the entry block. + if (BBI.TrueBB == BB || BBI.FalseBB == BB) { + BBI.IsBeingAnalyzed = false; + BBI.IsAnalyzed = true; + return BBI; + } - // If both paths are dead, then forget about it. - if (TrueBBI.Kind == ICDead && FalseBBI.Kind == ICDead) { - BBI.Kind = ICDead; - return; + BBInfo &TrueBBI = AnalyzeBlock(BBI.TrueBB, Tokens); + BBInfo &FalseBBI = AnalyzeBlock(BBI.FalseBB, Tokens); + + if (TrueBBI.IsDone && FalseBBI.IsDone) { + BBI.IsBeingAnalyzed = false; + BBI.IsAnalyzed = true; + return BBI; } - // Look for more opportunities to if-convert a triangle. Try to restructure - // the CFG to form a triangle with the 'false' path. - std::vector RevCond(BBI.BrCond); + SmallVector RevCond(BBI.BrCond.begin(), BBI.BrCond.end()); bool CanRevCond = !TII->ReverseBranchCondition(RevCond); - if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI) && - !(TrueBBI.ModifyPredicate && FalseBBI.ModifyPredicate) && + 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, Dups, Dups2) && + MeetIfcvtSizeLimit(TrueBBI.NonPredSize - (Dups + Dups2)) && + MeetIfcvtSizeLimit(FalseBBI.NonPredSize - (Dups + Dups2)) && FeasibilityAnalysis(TrueBBI, BBI.BrCond) && FeasibilityAnalysis(FalseBBI, RevCond)) { // Diamond: @@ -364,181 +697,93 @@ void IfConverter::AnalyzeBlock(MachineBasicBlock *BB) { // \ / // TailBB // Note TailBB can be empty. - BBI.Kind = ICDiamond; - TrueBBI.Kind = FalseBBI.Kind = ICChild; - BBI.TailBB = TrueBBI.TrueBB; - } else { - // FIXME: Consider duplicating if BB is small. - if (ValidTriangle(TrueBBI, FalseBBI) && - FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) { - // Triangle: - // EBB - // | \_ - // | | - // | TBB - // | / - // FBB - BBI.Kind = ICTriangle; - TrueBBI.Kind = ICChild; - } else if (ValidSimple(TrueBBI) && - FeasibilityAnalysis(TrueBBI, BBI.BrCond)) { - // Simple (split, no rejoin): - // EBB - // | \_ - // | | - // | TBB---> exit - // | - // FBB - BBI.Kind = ICSimple; - TrueBBI.Kind = ICChild; - } else if (CanRevCond) { - // Try the other path... - if (ValidTriangle(FalseBBI, TrueBBI) && - FeasibilityAnalysis(FalseBBI, RevCond, true)) { - // Reverse 'true' and 'false' paths. - ReverseBranchCondition(BBI); - BBI.Kind = ICTriangle; - FalseBBI.Kind = ICChild; - } else if (ValidTriangle(FalseBBI, TrueBBI, true) && - FeasibilityAnalysis(FalseBBI, RevCond, true, true)) { - ReverseBranchCondition(FalseBBI); - ReverseBranchCondition(BBI); - BBI.Kind = ICTriangle; - FalseBBI.Kind = ICChild; - } else if (ValidSimple(FalseBBI) && - FeasibilityAnalysis(FalseBBI, RevCond)) { - BBI.Kind = ICSimpleFalse; - FalseBBI.Kind = ICChild; - } - } + Tokens.push_back(new IfcvtToken(BBI, ICDiamond, TNeedSub|FNeedSub, Dups, + Dups2)); + Enqueued = true; } - return; -} - -/// FeasibilityAnalysis - 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 instruction which can clobber a -/// predicate (e.g. condition code register). If so, the block is not -/// predicable unless it's the last instruction. -bool IfConverter::FeasibilityAnalysis(BBInfo &BBI, - std::vector &Pred, - bool isTriangle, bool RevBranch) { - // If the block is dead, or it is going to be the entry block of a sub-CFG - // that will be if-converted, then it cannot be predicated. - if (BBI.Kind != ICNotAnalyzed && - BBI.Kind != ICNotClassfied && - BBI.Kind != ICChild) - return false; - // Check predication threshold. - if (BBI.NonPredSize == 0 || BBI.NonPredSize > TLI->getIfCvtBlockSizeLimit()) - return false; + if (ValidTriangle(TrueBBI, FalseBBI, false, Dups) && + MeetIfcvtSizeLimit(TrueBBI.NonPredSize) && + FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) { + // Triangle: + // EBB + // | \_ + // | | + // | TBB + // | / + // FBB + Tokens.push_back(new IfcvtToken(BBI, ICTriangle, TNeedSub, Dups)); + Enqueued = true; + } + + 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 it is already predicated, check if its predicate subsumes the new - // predicate. - if (BBI.Predicate.size() && !TII->SubsumesPredicate(BBI.Predicate, Pred)) - return false; + if (ValidSimple(TrueBBI, Dups) && + MeetIfcvtSizeLimit(TrueBBI.NonPredSize) && + FeasibilityAnalysis(TrueBBI, BBI.BrCond)) { + // Simple (split, no rejoin): + // EBB + // | \_ + // | | + // | TBB---> exit + // | + // FBB + Tokens.push_back(new IfcvtToken(BBI, ICSimple, TNeedSub, Dups)); + Enqueued = true; + } - bool SeenPredMod = false; - bool SeenCondBr = false; - for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end(); - I != E; ++I) { - const TargetInstrDescriptor *TID = I->getInstrDescriptor(); - if (SeenPredMod) { - // Predicate modification instruction should end the block (except for - // already predicated instructions and end of block branches). - if (!TII->isPredicated(I)) { - // This is the 'true' block of a triangle, i.e. its 'true' block is - // the same as the 'false' block of the entry. So false positive - // is ok. - if (isTriangle && !SeenCondBr && BBI.IsAnalyzable && - (TID->Flags & M_BRANCH_FLAG) != 0 && - (TID->Flags & M_BARRIER_FLAG) == 0) { - // This is the first conditional branch, test predicate subsumsion. - std::vector RevPred(Pred); - std::vector Cond(BBI.BrCond); - if (RevBranch) { - if (TII->ReverseBranchCondition(Cond)) - return false; - } - if (TII->ReverseBranchCondition(RevPred) || - !TII->SubsumesPredicate(Cond, RevPred)) - return false; - SeenCondBr = true; - continue; // Conditional branches is not predicable. - } - return false; - } + 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 (TID->Flags & M_CLOBBERS_PRED) { - BBI.ModifyPredicate = true; - SeenPredMod = 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 (!I->isPredicable()) - return false; + if (ValidSimple(FalseBBI, Dups) && + MeetIfcvtSizeLimit(FalseBBI.NonPredSize) && + FeasibilityAnalysis(FalseBBI, RevCond)) { + Tokens.push_back(new IfcvtToken(BBI, ICSimpleFalse, FNeedSub, Dups)); + Enqueued = true; + } } - return true; -} - -/// AttemptRestructuring - Restructure the sub-CFG rooted in the given block to -/// expose more if-conversion opportunities. e.g. -/// -/// cmp -/// b le BB1 -/// / \____ -/// / | -/// cmp | -/// b eq BB1 | -/// / \____ | -/// / \ | -/// BB1 -/// ==> -/// -/// cmp -/// b eq BB1 -/// / \____ -/// / | -/// cmp | -/// b le BB1 | -/// / \____ | -/// / \ | -/// BB1 -bool IfConverter::AttemptRestructuring(BBInfo &BBI) { - return false; + BBI.IsEnqueued = Enqueued; + BBI.IsBeingAnalyzed = false; + BBI.IsAnalyzed = true; + return BBI; } /// AnalyzeBlocks - Analyze all blocks and find entries for all if-conversion /// candidates. It returns true if any CFG restructuring is done to expose more /// if-conversion opportunities. bool IfConverter::AnalyzeBlocks(MachineFunction &MF, - std::vector &Candidates) { + std::vector &Tokens) { bool Change = false; 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); - BBInfo &BBI = BBAnalysis[BB->getNumber()]; - switch (BBI.Kind) { - case ICSimple: - case ICSimpleFalse: - case ICTriangle: - case ICDiamond: - Candidates.push_back(&BBI); - break; - default: - Change |= AttemptRestructuring(BBI); - break; - } + AnalyzeBlock(BB, Tokens); } } // Sort to favor more complex ifcvt scheme. - std::stable_sort(Candidates.begin(), Candidates.end(), IfcvtCandidateCmp); + std::stable_sort(Tokens.begin(), Tokens.end(), IfcvtTokenCmp); return Change; } @@ -556,24 +801,17 @@ static bool canFallThroughTo(MachineBasicBlock *BB, MachineBasicBlock *ToBB) { return true; } -/// getNextBlock - Returns the next block in the function blocks ordering. If -/// it is the end, returns NULL. -static inline MachineBasicBlock *getNextBlock(MachineBasicBlock *BB) { - MachineFunction::iterator I = BB; - MachineFunction::iterator E = BB->getParent()->end(); - if (++I == E) - return NULL; - return I; -} - -/// ReTryPreds - Invalidate predecessor BB info so it would be re-analyzed -/// to determine if it can be if-converted. -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()]; - if (PBBI.Kind == ICNotClassfied) - PBBI.Kind = ICReAnalyze; + if (PBBI.IsDone || PBBI.BB == BB) + continue; + PBBI.IsAnalyzed = false; + PBBI.IsEnqueued = false; } } @@ -581,7 +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); } @@ -589,43 +827,52 @@ 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. /// -bool IfConverter::IfConvertSimple(BBInfo &BBI) { +bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) { BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; BBInfo *CvtBBI = &TrueBBI; BBInfo *NextBBI = &FalseBBI; - std::vector Cond(BBI.BrCond); - if (BBI.Kind == ICSimpleFalse) { + SmallVector Cond(BBI.BrCond.begin(), BBI.BrCond.end()); + if (Kind == ICSimpleFalse) std::swap(CvtBBI, NextBBI); - TII->ReverseBranchCondition(Cond); + + 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; + return false; } - PredicateBlock(*CvtBBI, Cond); + if (Kind == ICSimpleFalse) + if (TII->ReverseBranchCondition(Cond)) + assert(false && "Unable to reverse branch condition!"); - // Merge converted block into entry block. - BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); - MergeBlocks(BBI, *CvtBBI); + if (CvtBBI->BB->pred_size() > 1) { + BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); + // Copy instructions in the true block, predicate them add them to + // the entry block. + CopyAndPredicateBlock(BBI, *CvtBBI, Cond); + } else { + PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond); + + // Merge converted block into entry block. + BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); + MergeBlocks(BBI, *CvtBBI); + } bool IterIfcvt = true; if (!canFallThroughTo(BBI.BB, NextBBI->BB)) { InsertUncondBranch(BBI.BB, NextBBI->BB, TII); - BBI.hasFallThrough = false; + BBI.HasFallThrough = false; // Now ifcvt'd block will look like this: // BB: // ... @@ -642,12 +889,10 @@ bool IfConverter::IfConvertSimple(BBInfo &BBI) { RemoveExtraEdges(BBI); // Update block info. BB can be iteratively if-converted. - if (IterIfcvt) - BBI.Kind = ICReAnalyze; - else - BBI.Kind = ICDead; - ReTryPreds(BBI.BB); - CvtBBI->Kind = ICDead; + if (!IterIfcvt) + BBI.IsDone = true; + InvalidatePreds(BBI.BB); + CvtBBI->IsDone = true; // FIXME: Must maintain LiveIns. return true; @@ -655,42 +900,91 @@ bool IfConverter::IfConvertSimple(BBInfo &BBI) { /// IfConvertTriangle - If convert a triangle sub-CFG. /// -bool IfConverter::IfConvertTriangle(BBInfo &BBI) { +bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; + BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; + BBInfo *CvtBBI = &TrueBBI; + BBInfo *NextBBI = &FalseBBI; - // Predicate the 'true' block after removing its branch. - TrueBBI.NonPredSize -= TII->RemoveBranch(*BBI.TrueBB); - PredicateBlock(TrueBBI, BBI.BrCond); + SmallVector Cond(BBI.BrCond.begin(), BBI.BrCond.end()); + if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) + std::swap(CvtBBI, NextBBI); + + 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; + return false; + } + + if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) + if (TII->ReverseBranchCondition(Cond)) + assert(false && "Unable to reverse branch condition!"); + + if (Kind == ICTriangleRev || Kind == ICTriangleFRev) { + 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; + } + } + } + } + + bool HasEarlyExit = CvtBBI->FalseBB != NULL; + bool DupBB = CvtBBI->BB->pred_size() > 1; + if (DupBB) { + BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); + // Copy instructions in the true block, predicate them add them to + // the entry block. + CopyAndPredicateBlock(BBI, *CvtBBI, Cond, true); + } else { + // Predicate the 'true' block after removing its branch. + CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB); + PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond); + } + + if (!DupBB) { + // Now merge the entry of the triangle with the true block. + BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); + MergeBlocks(BBI, *CvtBBI); + } // If 'true' block has a 'false' successor, add an exit branch to it. - bool HasEarlyExit = TrueBBI.FalseBB != NULL; if (HasEarlyExit) { - std::vector RevCond(TrueBBI.BrCond); + SmallVector RevCond(CvtBBI->BrCond.begin(), + CvtBBI->BrCond.end()); if (TII->ReverseBranchCondition(RevCond)) assert(false && "Unable to reverse branch condition!"); - TII->InsertBranch(*BBI.TrueBB, TrueBBI.FalseBB, NULL, RevCond); + TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, NULL, RevCond); + BBI.BB->addSuccessor(CvtBBI->FalseBB); } - // Now merge the entry of the triangle with the true block. - BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); - MergeBlocks(BBI, TrueBBI); - // Merge in the 'false' block if the 'false' block has no other // predecessors. Otherwise, add a unconditional branch from to 'false'. - BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; bool FalseBBDead = false; bool IterIfcvt = true; - bool isFallThrough = canFallThroughTo(BBI.BB, FalseBBI.BB); + bool isFallThrough = canFallThroughTo(BBI.BB, NextBBI->BB); if (!isFallThrough) { // 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 && FalseBBI.BB->pred_size() == 1) { - MergeBlocks(BBI, FalseBBI); + if (!HasEarlyExit && + NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough) { + MergeBlocks(BBI, *NextBBI); FalseBBDead = true; } else { - InsertUncondBranch(BBI.BB, FalseBBI.BB, TII); - TrueBBI.hasFallThrough = false; + InsertUncondBranch(BBI.BB, NextBBI->BB, TII); + BBI.HasFallThrough = false; } // Mixed predicated and unpredicated code. This cannot be iteratively // predicated. @@ -700,14 +994,12 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI) { RemoveExtraEdges(BBI); // Update block info. BB can be iteratively if-converted. - if (IterIfcvt) - BBI.Kind = ICReAnalyze; - else - BBI.Kind = ICDead; - ReTryPreds(BBI.BB); - TrueBBI.Kind = ICDead; + if (!IterIfcvt) + BBI.IsDone = true; + InvalidatePreds(BBI.BB); + CvtBBI->IsDone = true; if (FalseBBDead) - FalseBBI.Kind = ICDead; + NextBBI->IsDone = true; // FIXME: Must maintain LiveIns. return true; @@ -715,141 +1007,121 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI) { /// IfConvertDiamond - If convert a diamond sub-CFG. /// -bool IfConverter::IfConvertDiamond(BBInfo &BBI) { +bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, + unsigned NumDups1, unsigned NumDups2) { BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; - - SmallVector Dups; - if (!BBI.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->isPredicable() && !FT->isPredicable()) - 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->isPredicable()) { - return false; // Can't if-convert. Abort! - } - ++TT; - } - while (FT != BBI.FalseBB->end()) { - if (!FT->isPredicable()) { - return false; // Can't if-convert. Abort! - } - ++FT; - } + MachineBasicBlock *TailBB = TrueBBI.TrueBB; + // True block must fall through or ended with unanalyzable terminator. + if (!TailBB) { + 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(); - bool NeedBr2 = !BBI2->TrueBB && BBI2->BB->succ_size(); - - if ((TrueBBI.ModifyPredicate && !FalseBBI.ModifyPredicate) || - (!TrueBBI.ModifyPredicate && !FalseBBI.ModifyPredicate && - NeedBr1 && !NeedBr2)) { + 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 (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 (BBI.TailBB && - BBI.TailBB->pred_size() == 1 && CvtBBI->BB->succ_size() == 1) { - CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB); - BBInfo TailBBI = BBAnalysis[BBI.TailBB->getNumber()]; - MergeBlocks(*CvtBBI, TailBBI); - TailBBI.Kind = ICDead; + // 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); + MergeBlocks(BBI, TailBBI); + TailBBI.IsDone = true; + } else { + InsertUncondBranch(BBI.BB, TailBB, TII); + BBI.HasFallThrough = false; + } } RemoveExtraEdges(BBI); // Update block info. - BBI.Kind = ICDead; - TrueBBI.Kind = ICDead; - FalseBBI.Kind = ICDead; + 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)) { @@ -858,12 +1130,64 @@ void IfConverter::PredicateBlock(BBInfo &BBI, } } - BBI.NonPredSize = 0; std::copy(Cond.begin(), Cond.end(), std::back_inserter(BBI.Predicate)); + BBI.IsAnalyzed = false; + BBI.NonPredSize = 0; + NumIfConvBBs++; } +/// 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, + SmallVectorImpl &Cond, + bool IgnoreBr) { + MachineFunction &MF = *ToBBI.BB->getParent(); + + for (MachineBasicBlock::iterator I = FromBBI.BB->begin(), + E = FromBBI.BB->end(); I != E; ++I) { + const TargetInstrDesc &TID = I->getDesc(); + bool isPredicated = TII->isPredicated(I); + // Do not copy the end of the block branches. + if (IgnoreBr && !isPredicated && TID.isBranch()) + break; + + MachineInstr *MI = MF.CloneMachineInstr(I); + ToBBI.BB->insert(ToBBI.BB->end(), MI); + ToBBI.NonPredSize++; + + if (!isPredicated) + if (!TII->PredicateInstruction(MI, Cond)) { + cerr << "Unable to predicate " << *MI << "!\n"; + abort(); + } + } + + 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)); + + ToBBI.ClobbersPred |= FromBBI.ClobbersPred; + ToBBI.IsAnalyzed = false; + + NumDupBBs++; +} + /// MergeBlocks - Move all instructions from FromBB to the end of ToBB. /// void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI) { @@ -883,7 +1207,7 @@ void IfConverter::MergeBlocks(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; + MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL; for (unsigned i = 0, e = Succs.size(); i != e; ++i) { MachineBasicBlock *Succ = Succs[i]; @@ -896,16 +1220,18 @@ 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); - ToBBI.NonPredSize += FromBBI.NonPredSize; - FromBBI.NonPredSize = 0; - - ToBBI.ModifyPredicate |= FromBBI.ModifyPredicate; - ToBBI.hasFallThrough = FromBBI.hasFallThrough; - std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(), std::back_inserter(ToBBI.Predicate)); FromBBI.Predicate.clear(); + + ToBBI.NonPredSize += FromBBI.NonPredSize; + FromBBI.NonPredSize = 0; + + ToBBI.ClobbersPred |= FromBBI.ClobbersPred; + ToBBI.HasFallThrough = FromBBI.HasFallThrough; + ToBBI.IsAnalyzed = false; + FromBBI.IsAnalyzed = false; }