X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FIfConversion.cpp;h=3bddc771ab1daa4ee2c80bd55bf58e1dd1fc0f22;hb=524825b626e8cf9b323a9dd985dcf0f74a99c46e;hp=a490d3b27f67a93ece65500c52fcce8559949d09;hpb=9618bcaf7d3e6fd4a26fab47efb0bd2409b69e0f;p=oota-llvm.git diff --git a/lib/CodeGen/IfConversion.cpp b/lib/CodeGen/IfConversion.cpp index a490d3b27f6..3bddc771ab1 100644 --- a/lib/CodeGen/IfConversion.cpp +++ b/lib/CodeGen/IfConversion.cpp @@ -23,6 +23,7 @@ #include "llvm/Support/Debug.h" #include "llvm/ADT/DepthFirstIterator.h" #include "llvm/ADT/Statistic.h" +#include "llvm/ADT/STLExtras.h" using namespace llvm; namespace { @@ -36,6 +37,8 @@ namespace { 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", @@ -47,20 +50,23 @@ namespace { 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 { + 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. - ICTriangle, // BB is entry of a triangle sub-CFG. - ICTriangleFalse, // Same as ICTriangle, 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. }; @@ -71,7 +77,6 @@ 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. /// 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). @@ -79,7 +84,7 @@ 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. @@ -87,7 +92,6 @@ namespace { /// 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; @@ -95,19 +99,40 @@ namespace { bool IsBrAnalyzable : 1; bool HasFallThrough : 1; bool IsUnpredicable : 1; + bool CannotBeCopied : 1; bool ClobbersPred : 1; unsigned NonPredSize; MachineBasicBlock *BB; MachineBasicBlock *TrueBB; MachineBasicBlock *FalseBB; - MachineBasicBlock *TailBB; std::vector BrCond; std::vector Predicate; - BBInfo() : Kind(ICNotClassfied), IsDone(false), IsBeingAnalyzed(false), + BBInfo() : IsDone(false), IsBeingAnalyzed(false), IsAnalyzed(false), IsEnqueued(false), IsBrAnalyzable(false), HasFallThrough(false), IsUnpredicable(false), - ClobbersPred(false), NonPredSize(0), - BB(0), TrueBB(0), FalseBB(0), TailBB(0) {} + 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 @@ -130,38 +155,65 @@ namespace { 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); - BBInfo &AnalyzeBlock(MachineBasicBlock *BB); + BBInfo &AnalyzeBlock(MachineBasicBlock *BB, + std::vector &Tokens); bool FeasibilityAnalysis(BBInfo &BBI, std::vector &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, + std::vector &Cond); + void CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, + std::vector &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; } - // 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; } @@ -190,34 +242,43 @@ bool IfConverter::runOnMachineFunction(MachineFunction &MF) { if (I->succ_size() == 0) 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; // 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.IsEnqueued || BBI.IsDone) + 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 ICSimple: case ICSimpleFalse: { - bool isFalse = BBI.Kind == ICSimpleFalse; + bool isFalse = Kind == ICSimpleFalse; if ((isFalse && DisableSimpleF) || (!isFalse && DisableSimple)) break; - DOUT << "Ifcvt (Simple" << (BBI.Kind == ICSimpleFalse ? " false" : "") + 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 (isFalse) NumSimpleFalse++; @@ -225,45 +286,53 @@ bool IfConverter::runOnMachineFunction(MachineFunction &MF) { break; } case ICTriangle: + case ICTriangleRev: case ICTriangleFalse: case ICTriangleFRev: { - bool isFalse = BBI.Kind == ICTriangleFalse; - bool isFalseRev = BBI.Kind == ICTriangleFRev; - if (DisableTriangle && !isFalse && !isFalseRev) break; - if (DisableTriangleF && isFalse) break; - if (DisableTriangleFR && isFalseRev) break; + 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 (isFalseRev) - DOUT << " false/rev"; + if (isRev) + DOUT << " rev"; DOUT << "): BB#" << BBI.BB->getNumber() << " (T:" - << BBI.TrueBB->getNumber() << ",F:" << BBI.FalseBB->getNumber() - << ") "; - RetVal = IfConvertTriangle(BBI); + << BBI.TrueBB->getNumber() << ",F:" + << BBI.FalseBB->getNumber() << ") "; + RetVal = IfConvertTriangle(BBI, Kind); DOUT << (RetVal ? "succeeded!" : "failed!") << "\n"; if (RetVal) { - if (isFalseRev) NumTriangleFRev++; - else if (isFalse) NumTriangleFalse++; - else NumTriangle++; + 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, Token->NumDups, Token->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; } @@ -272,12 +341,22 @@ 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(), @@ -289,6 +368,8 @@ 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); @@ -310,24 +391,60 @@ static inline MachineBasicBlock *getNextBlock(MachineBasicBlock *BB) { } /// ValidSimple - Returns true if the 'true' block (along with its -/// predecessor) forms a valid simple shape for ifcvt. -bool IfConverter::ValidSimple(BBInfo &TrueBBI) const { - if (TrueBBI.IsBeingAnalyzed) +/// 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; - return !blockAlwaysFallThrough(TrueBBI) && - TrueBBI.BrCond.size() == 0 && TrueBBI.BB->pred_size() == 1; + 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.IsBeingAnalyzed) + bool FalseBranch, unsigned &Dups) const { + Dups = 0; + if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) return false; - if (TrueBBI.BB->pred_size() != 1) - 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.size() == 0) + // 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)) { @@ -339,10 +456,26 @@ 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; + const TargetInstrDescriptor *TID = I->getInstrDescriptor(); + if ((TID->Flags & M_BRANCH_FLAG) == 0) + 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; @@ -356,10 +489,35 @@ 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 @@ -371,8 +529,9 @@ 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; + BBI.TrueBB = BBI.FalseBB = NULL; BBI.BrCond.clear(); BBI.IsBrAnalyzable = !TII->AnalyzeBranch(*BBI.BB, BBI.TrueBB, BBI.FalseBB, BBI.BrCond); @@ -393,12 +552,25 @@ void IfConverter::ScanInstructions(BBInfo &BBI) { for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end(); I != E; ++I) { const TargetInstrDescriptor *TID = I->getInstrDescriptor(); + if ((TID->Flags & M_NOT_DUPLICABLE) != 0) + BBI.CannotBeCopied = true; + 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++; + 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 @@ -411,15 +583,18 @@ 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 (!I->isPredicable()) { + if ((TID->Flags & M_PREDICABLE) == 0) { BBI.IsUnpredicable = true; return; } @@ -431,17 +606,8 @@ void IfConverter::ScanInstructions(BBInfo &BBI) { bool IfConverter::FeasibilityAnalysis(BBInfo &BBI, std::vector &Pred, bool isTriangle, bool RevBranch) { - // Forget about it if it's unpredicable. - if (BBI.IsUnpredicable) - return false; - - // 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.IsDone || BBI.IsEnqueued) - return false; - - // Check predication threshold. - if (BBI.NonPredSize == 0 || BBI.NonPredSize > TLI->getIfCvtBlockSizeLimit()) + // 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 @@ -471,7 +637,8 @@ bool IfConverter::FeasibilityAnalysis(BBInfo &BBI, /// 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. -IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB) { +IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, + std::vector &Tokens) { BBInfo &BBI = BBAnalysis[BB->getNumber()]; if (BBI.IsAnalyzed || BBI.IsBeingAnalyzed) @@ -479,7 +646,6 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB) { BBI.BB = BB; BBI.IsBeingAnalyzed = true; - BBI.Kind = ICNotClassfied; ScanInstructions(BBI); @@ -497,8 +663,8 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB) { return BBI; } - BBInfo &TrueBBI = AnalyzeBlock(BBI.TrueBB); - BBInfo &FalseBBI = AnalyzeBlock(BBI.FalseBB); + BBInfo &TrueBBI = AnalyzeBlock(BBI.TrueBB, Tokens); + BBInfo &FalseBBI = AnalyzeBlock(BBI.FalseBB, Tokens); if (TrueBBI.IsDone && FalseBBI.IsDone) { BBI.IsBeingAnalyzed = false; @@ -509,8 +675,14 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB) { std::vector RevCond(BBI.BrCond); bool CanRevCond = !TII->ReverseBranchCondition(RevCond); - if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI) && - !(TrueBBI.ClobbersPred && FalseBBI.ClobbersPred) && + 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: @@ -521,105 +693,93 @@ IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB) { // \ / // TailBB // Note TailBB can be empty. - BBI.Kind = ICDiamond; - 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; - } else if (ValidSimple(TrueBBI) && - FeasibilityAnalysis(TrueBBI, BBI.BrCond)) { - // Simple (split, no rejoin): - // EBB - // | \_ - // | | - // | TBB---> exit - // | - // FBB - BBI.Kind = ICSimple; - } else if (CanRevCond) { - // Try the other path... - if (ValidTriangle(FalseBBI, TrueBBI) && - FeasibilityAnalysis(FalseBBI, RevCond, true)) { - BBI.Kind = ICTriangleFalse; - } else if (ValidTriangle(FalseBBI, TrueBBI, true) && - FeasibilityAnalysis(FalseBBI, RevCond, true, true)) { - BBI.Kind = ICTriangleFRev; - } else if (ValidSimple(FalseBBI) && - FeasibilityAnalysis(FalseBBI, RevCond)) { - BBI.Kind = ICSimpleFalse; - } + 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 + // | \_ + // | | + // | 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 (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; + } + + 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; } } + BBI.IsEnqueued = Enqueued; BBI.IsBeingAnalyzed = false; BBI.IsAnalyzed = true; return BBI; } -/// 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; -} - /// 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; - BBInfo &BBI = AnalyzeBlock(BB); - 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; } @@ -637,16 +797,17 @@ 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. -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.IsDone && PBBI.Kind == ICNotClassfied) { - assert(PBBI.IsEnqueued && "Unexpected"); - PBBI.IsAnalyzed = false; - } + if (PBBI.IsDone || PBBI.BB == BB) + continue; + PBBI.IsAnalyzed = false; + PBBI.IsEnqueued = false; } } @@ -663,37 +824,45 @@ static void InsertUncondBranch(MachineBasicBlock *BB, MachineBasicBlock *ToBB, 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); + 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) { + 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) + TII->ReverseBranchCondition(Cond); - // 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)) { @@ -717,7 +886,7 @@ bool IfConverter::IfConvertSimple(BBInfo &BBI) { // 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. @@ -726,49 +895,72 @@ 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; std::vector Cond(BBI.BrCond); - if (BBI.Kind == ICTriangleFalse || BBI.Kind == ICTriangleFRev) { - if (BBI.Kind == ICTriangleFRev) { - ReverseBranchCondition(*NextBBI); - // False 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 = NextBBI->BB->pred_begin(), - E = NextBBI->BB->pred_end(); PI != E; ++PI) { - MachineBasicBlock *PBB = *PI; - if (PBB == BBI.BB) - continue; - BBInfo &PBBI = BBAnalysis[PBB->getNumber()]; - if (PBBI.IsEnqueued) - PBBI.IsEnqueued = false; - } - } + 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) TII->ReverseBranchCondition(Cond); + + 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; + } + } } - // Predicate the 'true' block after removing its branch. - TrueBBI.NonPredSize -= TII->RemoveBranch(*CvtBBI->BB); - PredicateBlock(*CvtBBI, Cond); + 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 = CvtBBI->FalseBB != NULL; if (HasEarlyExit) { std::vector RevCond(CvtBBI->BrCond); if (TII->ReverseBranchCondition(RevCond)) assert(false && "Unable to reverse branch condition!"); - TII->InsertBranch(*CvtBBI->BB, CvtBBI->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, *CvtBBI); - // Merge in the 'false' block if the 'false' block has no other // predecessors. Otherwise, add a unconditional branch from to 'false'. bool FalseBBDead = false; @@ -778,7 +970,8 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI) { // 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 { @@ -795,7 +988,7 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI) { // 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; @@ -806,47 +999,28 @@ 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. @@ -856,89 +1030,89 @@ bool IfConverter::IfConvertDiamond(BBInfo &BBI) { 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.ClobbersPred && !FalseBBI.ClobbersPred) || - (!TrueBBI.ClobbersPred && !FalseBBI.ClobbersPred && - NeedBr1 && !NeedBr2)) { + + // 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.IsDone = true; + // 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.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, + std::vector &Cond) { + for (MachineBasicBlock::iterator I = BBI.BB->begin(); I != E; ++I) { if (TII->isPredicated(I)) continue; if (!TII->PredicateInstruction(I, Cond)) { @@ -947,12 +1121,62 @@ 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, + std::vector &Cond, + bool IgnoreBr) { + for (MachineBasicBlock::iterator I = FromBBI.BB->begin(), + E = FromBBI.BB->end(); I != E; ++I) { + const TargetInstrDescriptor *TID = I->getInstrDescriptor(); + bool isPredicated = TII->isPredicated(I); + // Do not copy the end of the block branches. + if (IgnoreBr && !isPredicated && (TID->Flags & M_BRANCH_FLAG) != 0) + break; + + MachineInstr *MI = I->clone(); + 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) { @@ -985,16 +1209,18 @@ 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); + 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; - - std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(), - std::back_inserter(ToBBI.Predicate)); - FromBBI.Predicate.clear(); + ToBBI.IsAnalyzed = false; + FromBBI.IsAnalyzed = false; }