X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FBranchFolding.cpp;h=d25152b1e436ca5ccd62dd36d1267a48172b40ba;hb=08ada178b8a9f2b380fb6c2bc2693c615532278e;hp=394fe7b81b1613dee78a0832cd2cf5483e7c8e8d;hpb=794fd75c67a2cdc128d67342c6d88a504d186896;p=oota-llvm.git diff --git a/lib/CodeGen/BranchFolding.cpp b/lib/CodeGen/BranchFolding.cpp index 394fe7b81b1..d25152b1e43 100644 --- a/lib/CodeGen/BranchFolding.cpp +++ b/lib/CodeGen/BranchFolding.cpp @@ -2,8 +2,8 @@ // // The LLVM Compiler Infrastructure // -// This file was developed by the LLVM research group 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. // //===----------------------------------------------------------------------===// // @@ -24,9 +24,11 @@ #include "llvm/CodeGen/RegisterScavenging.h" #include "llvm/Target/TargetInstrInfo.h" #include "llvm/Target/TargetMachine.h" -#include "llvm/Target/MRegisterInfo.h" +#include "llvm/Target/TargetRegisterInfo.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" +#include "llvm/Support/ErrorHandling.h" +#include "llvm/ADT/SmallSet.h" #include "llvm/ADT/Statistic.h" #include "llvm/ADT/STLExtras.h" #include @@ -35,12 +37,25 @@ using namespace llvm; STATISTIC(NumDeadBlocks, "Number of dead blocks removed"); STATISTIC(NumBranchOpts, "Number of branches optimized"); STATISTIC(NumTailMerge , "Number of block tails merged"); -static cl::opt EnableTailMerge("enable-tail-merge", cl::Hidden); +static cl::opt FlagEnableTailMerge("enable-tail-merge", + cl::init(cl::BOU_UNSET), cl::Hidden); +// Throttle for huge numbers of predecessors (compile speed problems) +static cl::opt +TailMergeThreshold("tail-merge-threshold", + cl::desc("Max number of predecessors to consider tail merging"), + cl::init(150), cl::Hidden); namespace { - struct BranchFolder : public MachineFunctionPass { - static const int ID; - BranchFolder() : MachineFunctionPass((intptr_t)&ID) {} + struct VISIBILITY_HIDDEN BranchFolder : public MachineFunctionPass { + static char ID; + explicit BranchFolder(bool defaultEnableTailMerge) : + MachineFunctionPass(&ID) { + switch (FlagEnableTailMerge) { + case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break; + case cl::BOU_TRUE: EnableTailMerge = true; break; + case cl::BOU_FALSE: EnableTailMerge = false; break; + } + } virtual bool runOnMachineFunction(MachineFunction &MF); virtual const char *getPassName() const { return "Control Flow Optimizer"; } @@ -49,28 +64,45 @@ namespace { bool MadeChange; private: // Tail Merging. + bool EnableTailMerge; bool TailMergeBlocks(MachineFunction &MF); + bool TryMergeBlocks(MachineBasicBlock* SuccBB, + MachineBasicBlock* PredBB); void ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, MachineBasicBlock *NewDest); MachineBasicBlock *SplitMBBAt(MachineBasicBlock &CurMBB, MachineBasicBlock::iterator BBI1); + unsigned ComputeSameTails(unsigned CurHash, unsigned minCommonTailLength); + void RemoveBlocksWithHash(unsigned CurHash, MachineBasicBlock* SuccBB, + MachineBasicBlock* PredBB); + unsigned CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, + unsigned maxCommonTailLength); + + typedef std::pair MergePotentialsElt; + typedef std::vector::iterator MPIterator; + std::vector MergePotentials; - const MRegisterInfo *RegInfo; + typedef std::pair SameTailElt; + std::vector SameTails; + + const TargetRegisterInfo *RegInfo; RegScavenger *RS; // Branch optzn. bool OptimizeBranches(MachineFunction &MF); void OptimizeBlock(MachineBasicBlock *MBB); void RemoveDeadBlock(MachineBasicBlock *MBB); + bool OptimizeImpDefsBlock(MachineBasicBlock *MBB); bool CanFallThrough(MachineBasicBlock *CurBB); bool CanFallThrough(MachineBasicBlock *CurBB, bool BranchUnAnalyzable, MachineBasicBlock *TBB, MachineBasicBlock *FBB, - const std::vector &Cond); + const SmallVectorImpl &Cond); }; - const int BranchFolder::ID = 0; + char BranchFolder::ID = 0; } -FunctionPass *llvm::createBranchFoldingPass() { return new BranchFolder(); } +FunctionPass *llvm::createBranchFoldingPass(bool DefaultEnableTailMerge) { + return new BranchFolder(DefaultEnableTailMerge); } /// RemoveDeadBlock - Remove the specified dead machine basic block from the /// function, updating the CFG. @@ -83,20 +115,70 @@ void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { while (!MBB->succ_empty()) MBB->removeSuccessor(MBB->succ_end()-1); - // If there is DWARF info to active, check to see if there are any LABEL - // records in the basic block. If so, unregister them from MachineModuleInfo. + // If there are any labels in the basic block, unregister them from + // MachineModuleInfo. if (MMI && !MBB->empty()) { for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E; ++I) { - if ((unsigned)I->getOpcode() == TargetInstrInfo::LABEL) { + if (I->isLabel()) // The label ID # is always operand #0, an immediate. MMI->InvalidateLabel(I->getOperand(0).getImm()); - } } } // Remove the block. - MF->getBasicBlockList().erase(MBB); + MF->erase(MBB); +} + +/// OptimizeImpDefsBlock - If a basic block is just a bunch of implicit_def +/// followed by terminators, and if the implicitly defined registers are not +/// used by the terminators, remove those implicit_def's. e.g. +/// BB1: +/// r0 = implicit_def +/// r1 = implicit_def +/// br +/// This block can be optimized away later if the implicit instructions are +/// removed. +bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) { + SmallSet ImpDefRegs; + MachineBasicBlock::iterator I = MBB->begin(); + while (I != MBB->end()) { + if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF) + break; + unsigned Reg = I->getOperand(0).getReg(); + ImpDefRegs.insert(Reg); + for (const unsigned *SubRegs = RegInfo->getSubRegisters(Reg); + unsigned SubReg = *SubRegs; ++SubRegs) + ImpDefRegs.insert(SubReg); + ++I; + } + if (ImpDefRegs.empty()) + return false; + + MachineBasicBlock::iterator FirstTerm = I; + while (I != MBB->end()) { + if (!TII->isUnpredicatedTerminator(I)) + return false; + // See if it uses any of the implicitly defined registers. + for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { + MachineOperand &MO = I->getOperand(i); + if (!MO.isReg() || !MO.isUse()) + continue; + unsigned Reg = MO.getReg(); + if (ImpDefRegs.count(Reg)) + return false; + } + ++I; + } + + I = MBB->begin(); + while (I != FirstTerm) { + MachineInstr *ImpDefMI = &*I; + ++I; + MBB->erase(ImpDefMI); + } + + return true; } bool BranchFolder::runOnMachineFunction(MachineFunction &MF) { @@ -104,11 +186,21 @@ bool BranchFolder::runOnMachineFunction(MachineFunction &MF) { if (!TII) return false; RegInfo = MF.getTarget().getRegisterInfo(); - RS = RegInfo->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; - MMI = getAnalysisToUpdate(); - + // Fix CFG. The later algorithms expect it to be right. bool EverMadeChange = false; + for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; I++) { + MachineBasicBlock *MBB = I, *TBB = 0, *FBB = 0; + SmallVector Cond; + if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond, true)) + EverMadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); + EverMadeChange |= OptimizeImpDefsBlock(MBB); + } + + RS = RegInfo->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; + + MMI = getAnalysisIfAvailable(); + bool MadeChangeThisIteration = true; while (MadeChangeThisIteration) { MadeChangeThisIteration = false; @@ -137,20 +229,19 @@ bool BranchFolder::runOnMachineFunction(MachineFunction &MF) { // If a jump table was merge with another one, walk the function rewriting // references to jump tables to reference the new JT ID's. Keep track of // whether we see a jump table idx, if not, we can delete the JT. - std::vector JTIsLive; - JTIsLive.resize(JTs.size()); + BitVector JTIsLive(JTs.size()); for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); BB != E; ++BB) { for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) for (unsigned op = 0, e = I->getNumOperands(); op != e; ++op) { MachineOperand &Op = I->getOperand(op); - if (!Op.isJumpTableIndex()) continue; - unsigned NewIdx = JTMapping[Op.getJumpTableIndex()]; - Op.setJumpTableIndex(NewIdx); + if (!Op.isJTI()) continue; + unsigned NewIdx = JTMapping[Op.getIndex()]; + Op.setIndex(NewIdx); // Remember that this JT is live. - JTIsLive[NewIdx] = true; + JTIsLive.set(NewIdx); } } @@ -158,7 +249,7 @@ bool BranchFolder::runOnMachineFunction(MachineFunction &MF) { // indirect jump was unreachable (and thus deleted) or because the jump // table was merged with some other one. for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) - if (!JTIsLive[i]) { + if (!JTIsLive.test(i)) { JTI->RemoveJumpTable(i); EverMadeChange = true; } @@ -184,14 +275,12 @@ static unsigned HashMachineInstr(const MachineInstr *MI) { case MachineOperand::MO_Register: OperandHash = Op.getReg(); break; case MachineOperand::MO_Immediate: OperandHash = Op.getImm(); break; case MachineOperand::MO_MachineBasicBlock: - OperandHash = Op.getMachineBasicBlock()->getNumber(); + OperandHash = Op.getMBB()->getNumber(); break; - case MachineOperand::MO_FrameIndex: OperandHash = Op.getFrameIndex(); break; + case MachineOperand::MO_FrameIndex: case MachineOperand::MO_ConstantPoolIndex: - OperandHash = Op.getConstantPoolIndex(); - break; case MachineOperand::MO_JumpTableIndex: - OperandHash = Op.getJumpTableIndex(); + OperandHash = Op.getIndex(); break; case MachineOperand::MO_GlobalAddress: case MachineOperand::MO_ExternalSymbol: @@ -207,10 +296,15 @@ static unsigned HashMachineInstr(const MachineInstr *MI) { return Hash; } -/// HashEndOfMBB - Hash the last two instructions in the MBB. We hash two -/// instructions, because cross-jumping only saves code when at least two -/// instructions are removed (since a branch must be inserted). -static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) { +/// HashEndOfMBB - Hash the last few instructions in the MBB. For blocks +/// with no successors, we hash two instructions, because cross-jumping +/// only saves code when at least two instructions are removed (since a +/// branch must be inserted). For blocks with a successor, one of the +/// two blocks to be tail-merged will end with a branch already, so +/// it gains to cross-jump even for one instruction. + +static unsigned HashEndOfMBB(const MachineBasicBlock *MBB, + unsigned minCommonTailLength) { MachineBasicBlock::const_iterator I = MBB->end(); if (I == MBB->begin()) return 0; // Empty MBB. @@ -218,7 +312,7 @@ static unsigned HashEndOfMBB(const MachineBasicBlock *MBB) { --I; unsigned Hash = HashMachineInstr(I); - if (I == MBB->begin()) + if (I == MBB->begin() || minCommonTailLength == 1) return Hash; // Single instr MBB. --I; @@ -240,7 +334,13 @@ static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, unsigned TailLen = 0; while (I1 != MBB1->begin() && I2 != MBB2->begin()) { --I1; --I2; - if (!I1->isIdenticalTo(I2)) { + if (!I1->isIdenticalTo(I2) || + // FIXME: This check is dubious. It's used to get around a problem where + // people incorrectly expect inline asm directives to remain in the same + // relative order. This is untenable because normal compiler + // optimizations (like this one) may reorder and/or merge these + // directives. + I1->getOpcode() == TargetInstrInfo::INLINEASM) { ++I1; ++I2; break; } @@ -265,7 +365,7 @@ void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, // If OldBB isn't immediately before OldBB, insert a branch to it. if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest)) - TII->InsertBranch(*OldBB, NewDest, 0, std::vector()); + TII->InsertBranch(*OldBB, NewDest, 0, SmallVector()); OldBB->addSuccessor(NewDest); ++NumTailMerge; } @@ -275,17 +375,15 @@ void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, /// iterator. This returns the new MBB. MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, MachineBasicBlock::iterator BBI1) { + MachineFunction &MF = *CurMBB.getParent(); + // Create the fall-through block. MachineFunction::iterator MBBI = &CurMBB; - MachineBasicBlock *NewMBB = new MachineBasicBlock(CurMBB.getBasicBlock()); - CurMBB.getParent()->getBasicBlockList().insert(++MBBI, NewMBB); + MachineBasicBlock *NewMBB =MF.CreateMachineBasicBlock(CurMBB.getBasicBlock()); + CurMBB.getParent()->insert(++MBBI, NewMBB); // Move all the successors of this block to the specified block. - while (!CurMBB.succ_empty()) { - MachineBasicBlock *S = *(CurMBB.succ_end()-1); - NewMBB->addSuccessor(S); - CurMBB.removeSuccessor(S); - } + NewMBB->transferSuccessors(&CurMBB); // Add an edge from CurMBB to NewMBB for the fall-through. CurMBB.addSuccessor(NewMBB); @@ -311,14 +409,13 @@ MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, /// EstimateRuntime - Make a rough estimate for how long it will take to run /// the specified code. static unsigned EstimateRuntime(MachineBasicBlock::iterator I, - MachineBasicBlock::iterator E, - const TargetInstrInfo *TII) { + MachineBasicBlock::iterator E) { unsigned Time = 0; for (; I != E; ++I) { - const TargetInstrDescriptor &TID = TII->get(I->getOpcode()); - if (TID.Flags & M_CALL_FLAG) + const TargetInstrDesc &TID = I->getDesc(); + if (TID.isCall()) Time += 10; - else if (TID.Flags & (M_LOAD_FLAG|M_STORE_FLAG)) + else if (TID.mayLoad() || TID.mayStore()) Time += 2; else ++Time; @@ -326,129 +423,350 @@ static unsigned EstimateRuntime(MachineBasicBlock::iterator I, return Time; } -/// ShouldSplitFirstBlock - We need to either split MBB1 at MBB1I or MBB2 at -/// MBB2I and then insert an unconditional branch in the other block. Determine -/// which is the best to split -static bool ShouldSplitFirstBlock(MachineBasicBlock *MBB1, - MachineBasicBlock::iterator MBB1I, - MachineBasicBlock *MBB2, - MachineBasicBlock::iterator MBB2I, - const TargetInstrInfo *TII) { - // TODO: if we had some notion of which block was hotter, we could split - // the hot block, so it is the fall-through. Since we don't have profile info - // make a decision based on which will hurt most to split. - unsigned MBB1Time = EstimateRuntime(MBB1->begin(), MBB1I, TII); - unsigned MBB2Time = EstimateRuntime(MBB2->begin(), MBB2I, TII); - - // If the MBB1 prefix takes "less time" to run than the MBB2 prefix, split the - // MBB1 block so it falls through. This will penalize the MBB2 path, but will - // have a lower overall impact on the program execution. - return MBB1Time < MBB2Time; +// CurMBB needs to add an unconditional branch to SuccMBB (we removed these +// branches temporarily for tail merging). In the case where CurMBB ends +// with a conditional branch to the next block, optimize by reversing the +// test and conditionally branching to SuccMBB instead. + +static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB, + const TargetInstrInfo *TII) { + MachineFunction *MF = CurMBB->getParent(); + MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB)); + MachineBasicBlock *TBB = 0, *FBB = 0; + SmallVector Cond; + if (I != MF->end() && + !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond, true)) { + MachineBasicBlock *NextBB = I; + if (TBB == NextBB && !Cond.empty() && !FBB) { + if (!TII->ReverseBranchCondition(Cond)) { + TII->RemoveBranch(*CurMBB); + TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond); + return; + } + } + } + TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector()); } -bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { - MadeChange = false; - - if (!EnableTailMerge) return false; - - // Find blocks with no successors. - std::vector > MergePotentials; - for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { - if (I->succ_empty()) - MergePotentials.push_back(std::make_pair(HashEndOfMBB(I), I)); +static bool MergeCompare(const std::pair &p, + const std::pair &q) { + if (p.first < q.first) + return true; + else if (p.first > q.first) + return false; + else if (p.second->getNumber() < q.second->getNumber()) + return true; + else if (p.second->getNumber() > q.second->getNumber()) + return false; + else { + // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing + // an object with itself. +#ifndef _GLIBCXX_DEBUG + llvm_unreachable("Predecessor appears twice"); +#endif + return false; + } +} + +/// ComputeSameTails - Look through all the blocks in MergePotentials that have +/// hash CurHash (guaranteed to match the last element). Build the vector +/// SameTails of all those that have the (same) largest number of instructions +/// in common of any pair of these blocks. SameTails entries contain an +/// iterator into MergePotentials (from which the MachineBasicBlock can be +/// found) and a MachineBasicBlock::iterator into that MBB indicating the +/// instruction where the matching code sequence begins. +/// Order of elements in SameTails is the reverse of the order in which +/// those blocks appear in MergePotentials (where they are not necessarily +/// consecutive). +unsigned BranchFolder::ComputeSameTails(unsigned CurHash, + unsigned minCommonTailLength) { + unsigned maxCommonTailLength = 0U; + SameTails.clear(); + MachineBasicBlock::iterator TrialBBI1, TrialBBI2; + MPIterator HighestMPIter = prior(MergePotentials.end()); + for (MPIterator CurMPIter = prior(MergePotentials.end()), + B = MergePotentials.begin(); + CurMPIter!=B && CurMPIter->first==CurHash; + --CurMPIter) { + for (MPIterator I = prior(CurMPIter); I->first==CurHash ; --I) { + unsigned CommonTailLen = ComputeCommonTailLength( + CurMPIter->second, + I->second, + TrialBBI1, TrialBBI2); + // If we will have to split a block, there should be at least + // minCommonTailLength instructions in common; if not, at worst + // we will be replacing a fallthrough into the common tail with a + // branch, which at worst breaks even with falling through into + // the duplicated common tail, so 1 instruction in common is enough. + // We will always pick a block we do not have to split as the common + // tail if there is one. + // (Empty blocks will get forwarded and need not be considered.) + if (CommonTailLen >= minCommonTailLength || + (CommonTailLen > 0 && + (TrialBBI1==CurMPIter->second->begin() || + TrialBBI2==I->second->begin()))) { + if (CommonTailLen > maxCommonTailLength) { + SameTails.clear(); + maxCommonTailLength = CommonTailLen; + HighestMPIter = CurMPIter; + SameTails.push_back(std::make_pair(CurMPIter, TrialBBI1)); + } + if (HighestMPIter == CurMPIter && + CommonTailLen == maxCommonTailLength) + SameTails.push_back(std::make_pair(I, TrialBBI2)); + } + if (I==B) + break; + } } + return maxCommonTailLength; +} + +/// RemoveBlocksWithHash - Remove all blocks with hash CurHash from +/// MergePotentials, restoring branches at ends of blocks as appropriate. +void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, + MachineBasicBlock* SuccBB, + MachineBasicBlock* PredBB) { + MPIterator CurMPIter, B; + for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); + CurMPIter->first==CurHash; + --CurMPIter) { + // Put the unconditional branch back, if we need one. + MachineBasicBlock *CurMBB = CurMPIter->second; + if (SuccBB && CurMBB != PredBB) + FixTail(CurMBB, SuccBB, TII); + if (CurMPIter==B) + break; + } + if (CurMPIter->first!=CurHash) + CurMPIter++; + MergePotentials.erase(CurMPIter, MergePotentials.end()); +} + +/// CreateCommonTailOnlyBlock - None of the blocks to be tail-merged consist +/// only of the common tail. Create a block that does by splitting one. +unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, + unsigned maxCommonTailLength) { + unsigned i, commonTailIndex; + unsigned TimeEstimate = ~0U; + for (i=0, commonTailIndex=0; isecond==PredBB) { + commonTailIndex = i; + break; + } + // Otherwise, make a (fairly bogus) choice based on estimate of + // how long it will take the various blocks to execute. + unsigned t = EstimateRuntime(SameTails[i].first->second->begin(), + SameTails[i].second); + if (t<=TimeEstimate) { + TimeEstimate = t; + commonTailIndex = i; + } + } + + MachineBasicBlock::iterator BBI = SameTails[commonTailIndex].second; + MachineBasicBlock *MBB = SameTails[commonTailIndex].first->second; + + DOUT << "\nSplitting " << MBB->getNumber() << ", size " << + maxCommonTailLength; + + MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); + SameTails[commonTailIndex].first->second = newMBB; + SameTails[commonTailIndex].second = newMBB->begin(); + // If we split PredBB, newMBB is the new predecessor. + if (PredBB==MBB) + PredBB = newMBB; + + return commonTailIndex; +} + +// See if any of the blocks in MergePotentials (which all have a common single +// successor, or all have no successor) can be tail-merged. If there is a +// successor, any blocks in MergePotentials that are not tail-merged and +// are not immediately before Succ must have an unconditional branch to +// Succ added (but the predecessor/successor lists need no adjustment). +// The lone predecessor of Succ that falls through into Succ, +// if any, is given in PredBB. + +bool BranchFolder::TryMergeBlocks(MachineBasicBlock *SuccBB, + MachineBasicBlock* PredBB) { + // It doesn't make sense to save a single instruction since tail merging + // will add a jump. + // FIXME: Ask the target to provide the threshold? + unsigned minCommonTailLength = (SuccBB ? 1 : 2) + 1; + MadeChange = false; + DOUT << "\nTryMergeBlocks " << MergePotentials.size() << '\n'; + // Sort by hash value so that blocks with identical end sequences sort // together. - std::stable_sort(MergePotentials.begin(), MergePotentials.end()); + std::stable_sort(MergePotentials.begin(), MergePotentials.end(),MergeCompare); // Walk through equivalence sets looking for actual exact matches. while (MergePotentials.size() > 1) { - unsigned CurHash = (MergePotentials.end()-1)->first; - unsigned PrevHash = (MergePotentials.end()-2)->first; - MachineBasicBlock *CurMBB = (MergePotentials.end()-1)->second; + unsigned CurHash = prior(MergePotentials.end())->first; - // If there is nothing that matches the hash of the current basic block, - // give up. - if (CurHash != PrevHash) { - MergePotentials.pop_back(); + // Build SameTails, identifying the set of blocks with this hash code + // and with the maximum number of instructions in common. + unsigned maxCommonTailLength = ComputeSameTails(CurHash, + minCommonTailLength); + + // If we didn't find any pair that has at least minCommonTailLength + // instructions in common, remove all blocks with this hash code and retry. + if (SameTails.empty()) { + RemoveBlocksWithHash(CurHash, SuccBB, PredBB); continue; } - - // Determine the actual length of the shared tail between these two basic - // blocks. Because the hash can have collisions, it's possible that this is - // less than 2. - MachineBasicBlock::iterator BBI1, BBI2; - unsigned CommonTailLen = - ComputeCommonTailLength(CurMBB, (MergePotentials.end()-2)->second, - BBI1, BBI2); - - // If the tails don't have at least two instructions in common, see if there - // is anything else in the equivalence class that does match. - if (CommonTailLen < 2) { - unsigned FoundMatch = ~0U; - for (int i = MergePotentials.size()-2; - i != -1 && MergePotentials[i].first == CurHash; --i) { - CommonTailLen = ComputeCommonTailLength(CurMBB, - MergePotentials[i].second, - BBI1, BBI2); - if (CommonTailLen >= 2) { - FoundMatch = i; + + // If one of the blocks is the entire common tail (and not the entry + // block, which we can't jump to), we can treat all blocks with this same + // tail at once. Use PredBB if that is one of the possibilities, as that + // will not introduce any extra branches. + MachineBasicBlock *EntryBB = MergePotentials.begin()->second-> + getParent()->begin(); + unsigned int commonTailIndex, i; + for (commonTailIndex=SameTails.size(), i=0; isecond; + if (MBB->begin() == SameTails[i].second && MBB != EntryBB) { + commonTailIndex = i; + if (MBB==PredBB) break; - } - } - - // If we didn't find anything that has at least two instructions matching - // this one, bail out. - if (FoundMatch == ~0U) { - MergePotentials.pop_back(); - continue; } - - // Otherwise, move the matching block to the right position. - std::swap(MergePotentials[FoundMatch], *(MergePotentials.end()-2)); } - MachineBasicBlock *MBB2 = (MergePotentials.end()-2)->second; + if (commonTailIndex==SameTails.size()) { + // None of the blocks consist entirely of the common tail. + // Split a block so that one does. + commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength); + } - // If neither block is the entire common tail, split the tail of one block - // to make it redundant with the other tail. - if (CurMBB->begin() != BBI1 && MBB2->begin() != BBI2) { - if (0) { // Enable this to disable partial tail merges. - MergePotentials.pop_back(); + MachineBasicBlock *MBB = SameTails[commonTailIndex].first->second; + // MBB is common tail. Adjust all other BB's to jump to this one. + // Traversal must be forwards so erases work. + DOUT << "\nUsing common tail " << MBB->getNumber() << " for "; + for (unsigned int i=0; ibegin(); - MergePotentials.back().second = CurMBB; - } else { - MBB2 = SplitMBBAt(*MBB2, BBI2); - BBI2 = MBB2->begin(); - (MergePotentials.end()-2)->second = MBB2; - } - } - - if (MBB2->begin() == BBI2) { - // Hack the end off CurMBB, making it jump to MBBI@ instead. - ReplaceTailWithBranchTo(BBI1, MBB2); - // This modifies CurMBB, so remove it from the worklist. - MergePotentials.pop_back(); - } else { - assert(CurMBB->begin() == BBI1 && "Didn't split block correctly?"); - // Hack the end off MBB2, making it jump to CurMBB instead. - ReplaceTailWithBranchTo(BBI2, CurMBB); - // This modifies MBB2, so remove it from the worklist. - MergePotentials.erase(MergePotentials.end()-2); + DOUT << SameTails[i].first->second->getNumber() << ","; + // Hack the end off BB i, making it jump to BB commonTailIndex instead. + ReplaceTailWithBranchTo(SameTails[i].second, MBB); + // BB i is no longer a predecessor of SuccBB; remove it from the worklist. + MergePotentials.erase(SameTails[i].first); } + DOUT << "\n"; + // We leave commonTailIndex in the worklist in case there are other blocks + // that match it with a smaller number of instructions. MadeChange = true; } - return MadeChange; } +bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { + + if (!EnableTailMerge) return false; + + MadeChange = false; + + // First find blocks with no successors. + MergePotentials.clear(); + for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { + if (I->succ_empty()) + MergePotentials.push_back(std::make_pair(HashEndOfMBB(I, 2U), I)); + } + // See if we can do any tail merging on those. + if (MergePotentials.size() < TailMergeThreshold && + MergePotentials.size() >= 2) + MadeChange |= TryMergeBlocks(NULL, NULL); + + // Look at blocks (IBB) with multiple predecessors (PBB). + // We change each predecessor to a canonical form, by + // (1) temporarily removing any unconditional branch from the predecessor + // to IBB, and + // (2) alter conditional branches so they branch to the other block + // not IBB; this may require adding back an unconditional branch to IBB + // later, where there wasn't one coming in. E.g. + // Bcc IBB + // fallthrough to QBB + // here becomes + // Bncc QBB + // with a conceptual B to IBB after that, which never actually exists. + // With those changes, we see whether the predecessors' tails match, + // and merge them if so. We change things out of canonical form and + // back to the way they were later in the process. (OptimizeBranches + // would undo some of this, but we can't use it, because we'd get into + // a compile-time infinite loop repeatedly doing and undoing the same + // transformations.) + + for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { + if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) { + MachineBasicBlock *IBB = I; + MachineBasicBlock *PredBB = prior(I); + MergePotentials.clear(); + for (MachineBasicBlock::pred_iterator P = I->pred_begin(), + E2 = I->pred_end(); + P != E2; ++P) { + MachineBasicBlock* PBB = *P; + // Skip blocks that loop to themselves, can't tail merge these. + if (PBB==IBB) + continue; + MachineBasicBlock *TBB = 0, *FBB = 0; + SmallVector Cond; + if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond, true)) { + // Failing case: IBB is the target of a cbr, and + // we cannot reverse the branch. + SmallVector NewCond(Cond); + if (!Cond.empty() && TBB==IBB) { + if (TII->ReverseBranchCondition(NewCond)) + continue; + // This is the QBB case described above + if (!FBB) + FBB = next(MachineFunction::iterator(PBB)); + } + // Failing case: the only way IBB can be reached from PBB is via + // exception handling. Happens for landing pads. Would be nice + // to have a bit in the edge so we didn't have to do all this. + if (IBB->isLandingPad()) { + MachineFunction::iterator IP = PBB; IP++; + MachineBasicBlock* PredNextBB = NULL; + if (IP!=MF.end()) + PredNextBB = IP; + if (TBB==NULL) { + if (IBB!=PredNextBB) // fallthrough + continue; + } else if (FBB) { + if (TBB!=IBB && FBB!=IBB) // cbr then ubr + continue; + } else if (Cond.empty()) { + if (TBB!=IBB) // ubr + continue; + } else { + if (TBB!=IBB && IBB!=PredNextBB) // cbr + continue; + } + } + // Remove the unconditional branch at the end, if any. + if (TBB && (Cond.empty() || FBB)) { + TII->RemoveBranch(*PBB); + if (!Cond.empty()) + // reinsert conditional branch only, for now + TII->InsertBranch(*PBB, (TBB==IBB) ? FBB : TBB, 0, NewCond); + } + MergePotentials.push_back(std::make_pair(HashEndOfMBB(PBB, 1U), *P)); + } + } + if (MergePotentials.size() >= 2) + MadeChange |= TryMergeBlocks(I, PredBB); + // Reinsert an unconditional branch if needed. + // The 1 below can occur as a result of removing blocks in TryMergeBlocks. + PredBB = prior(I); // this may have been changed in TryMergeBlocks + if (MergePotentials.size()==1 && + MergePotentials.begin()->second != PredBB) + FixTail(MergePotentials.begin()->second, I, TII); + } + } + return MadeChange; +} //===----------------------------------------------------------------------===// // Branch Optimization @@ -475,90 +793,6 @@ bool BranchFolder::OptimizeBranches(MachineFunction &MF) { } -/// CorrectExtraCFGEdges - Various pieces of code can cause excess edges in the -/// CFG to be inserted. If we have proven that MBB can only branch to DestA and -/// DestB, remove any other MBB successors from the CFG. DestA and DestB can -/// be null. -static bool CorrectExtraCFGEdges(MachineBasicBlock &MBB, - MachineBasicBlock *DestA, - MachineBasicBlock *DestB, - bool isCond, - MachineFunction::iterator FallThru) { - bool MadeChange = false; - bool AddedFallThrough = false; - - // If this block ends with a conditional branch that falls through to its - // successor, set DestB as the successor. - if (isCond) { - if (DestB == 0 && FallThru != MBB.getParent()->end()) { - DestB = FallThru; - AddedFallThrough = true; - } - } else { - // If this is an unconditional branch with no explicit dest, it must just be - // a fallthrough into DestB. - if (DestA == 0 && FallThru != MBB.getParent()->end()) { - DestA = FallThru; - AddedFallThrough = true; - } - } - - MachineBasicBlock::pred_iterator SI = MBB.succ_begin(); - while (SI != MBB.succ_end()) { - if (*SI == DestA) { - DestA = 0; - ++SI; - } else if (*SI == DestB) { - DestB = 0; - ++SI; - } else if ((*SI)->isLandingPad()) { - ++SI; - } else { - // Otherwise, this is a superfluous edge, remove it. - MBB.removeSuccessor(SI); - MadeChange = true; - } - } - if (!AddedFallThrough) { - assert(DestA == 0 && DestB == 0 && - "MachineCFG is missing edges!"); - } else if (isCond) { - assert(DestA == 0 && "MachineCFG is missing edges!"); - } - return MadeChange; -} - - -/// ReplaceUsesOfBlockWith - Given a machine basic block 'BB' that branched to -/// 'Old', change the code and CFG so that it branches to 'New' instead. -static void ReplaceUsesOfBlockWith(MachineBasicBlock *BB, - MachineBasicBlock *Old, - MachineBasicBlock *New, - const TargetInstrInfo *TII) { - assert(Old != New && "Cannot replace self with self!"); - - MachineBasicBlock::iterator I = BB->end(); - while (I != BB->begin()) { - --I; - if (!TII->isTerminatorInstr(I->getOpcode())) break; - - // Scan the operands of this machine instruction, replacing any uses of Old - // with New. - for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) - if (I->getOperand(i).isMachineBasicBlock() && - I->getOperand(i).getMachineBasicBlock() == Old) - I->getOperand(i).setMachineBasicBlock(New); - } - - // Update the successor information. - std::vector Succs(BB->succ_begin(), BB->succ_end()); - for (int i = Succs.size()-1; i >= 0; --i) - if (Succs[i] == Old) { - BB->removeSuccessor(Old); - BB->addSuccessor(New); - } -} - /// CanFallThrough - Return true if the specified block (with the specified /// branch condition) can implicitly transfer control to the block after it by /// falling off the end of it. This should return false if it can reach the @@ -568,8 +802,9 @@ static void ReplaceUsesOfBlockWith(MachineBasicBlock *BB, /// bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB, bool BranchUnAnalyzable, - MachineBasicBlock *TBB, MachineBasicBlock *FBB, - const std::vector &Cond) { + MachineBasicBlock *TBB, + MachineBasicBlock *FBB, + const SmallVectorImpl &Cond) { MachineFunction::iterator Fallthrough = CurBB; ++Fallthrough; // If FallthroughBlock is off the end of the function, it can't fall through. @@ -610,8 +845,8 @@ bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB, /// bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) { MachineBasicBlock *TBB = 0, *FBB = 0; - std::vector Cond; - bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond); + SmallVector Cond; + bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond, true); return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond); } @@ -620,17 +855,21 @@ bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) { /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will /// result in infinite loops. static bool IsBetterFallthrough(MachineBasicBlock *MBB1, - MachineBasicBlock *MBB2, - const TargetInstrInfo &TII) { + MachineBasicBlock *MBB2) { // Right now, we use a simple heuristic. If MBB2 ends with a call, and // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to // optimize branches that branch to either a return block or an assert block // into a fallthrough to the return. if (MBB1->empty() || MBB2->empty()) return false; + + // If there is a clear successor ordering we make sure that one block + // will fall through to the next + if (MBB1->isSuccessor(MBB2)) return true; + if (MBB2->isSuccessor(MBB1)) return false; MachineInstr *MBB1I = --MBB1->end(); MachineInstr *MBB2I = --MBB2->end(); - return TII.isCall(MBB2I->getOpcode()) && !TII.isCall(MBB1I->getOpcode()); + return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall(); } /// OptimizeBlock - Analyze and optimize control flow related to the specified @@ -640,8 +879,9 @@ void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { ++FallThrough; // If this block is empty, make everyone use its fall-through, not the block - // explicitly. - if (MBB->empty()) { + // explicitly. Landing pads should not do this since the landing-pad table + // points to this block. + if (MBB->empty() && !MBB->isLandingPad()) { // Dead block? Leave for cleanup later. if (MBB->pred_empty()) return; @@ -652,9 +892,25 @@ void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { // instead. while (!MBB->pred_empty()) { MachineBasicBlock *Pred = *(MBB->pred_end()-1); - ReplaceUsesOfBlockWith(Pred, MBB, FallThrough, TII); + Pred->ReplaceUsesOfBlockWith(MBB, FallThrough); + // If this resulted in a predecessor with true and false edges + // both going to the fallthrough block, clean up; + // BranchFolding doesn't like this. + MachineBasicBlock::succ_iterator SI = Pred->succ_begin(); + bool found = false; + while (SI != Pred->succ_end()) { + if (*SI == FallThrough) { + if (!found) { + found = true; + ++SI; + } else { + SI = Pred->removeSuccessor(SI); + } + } else { + ++SI; + } + } } - // If MBB was the target of a jump table, update jump tables to go to the // fallthrough instead. MBB->getParent()->getJumpTableInfo()-> @@ -669,13 +925,13 @@ void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { MachineBasicBlock &PrevBB = *prior(MachineFunction::iterator(MBB)); MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; - std::vector PriorCond; + SmallVector PriorCond; bool PriorUnAnalyzable = - TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); + TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); if (!PriorUnAnalyzable) { // If the CFG for the prior block has extra edges, remove them. - MadeChange |= CorrectExtraCFGEdges(PrevBB, PriorTBB, PriorFBB, - !PriorCond.empty(), MBB); + MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, + !PriorCond.empty()); // If the previous branch is conditional and both conditions go to the same // destination, remove the branch, replacing it with an unconditional one or @@ -713,7 +969,7 @@ void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { // if the branch condition is reversible, reverse the branch to create a // fall-through. if (PriorTBB == MBB) { - std::vector NewPriorCond(PriorCond); + SmallVector NewPriorCond(PriorCond); if (!TII->ReverseBranchCondition(NewPriorCond)) { TII->RemoveBranch(PrevBB); TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond); @@ -742,7 +998,7 @@ void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { // last. Only do the swap if one is clearly better to fall through than // the other. if (FallThrough == --MBB->getParent()->end() && - !IsBetterFallthrough(PriorTBB, MBB, *TII)) + !IsBetterFallthrough(PriorTBB, MBB)) DoTransform = false; // We don't want to do this transformation if we have control flow like: @@ -763,7 +1019,7 @@ void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { if (DoTransform) { // Reverse the branch so we will fall through on the previous true cond. - std::vector NewPriorCond(PriorCond); + SmallVector NewPriorCond(PriorCond); if (!TII->ReverseBranchCondition(NewPriorCond)) { DOUT << "\nMoving MBB: " << *MBB; DOUT << "To make fallthrough to: " << *PriorTBB << "\n"; @@ -783,13 +1039,11 @@ void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { // Analyze the branch in the current block. MachineBasicBlock *CurTBB = 0, *CurFBB = 0; - std::vector CurCond; - bool CurUnAnalyzable = TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond); + SmallVector CurCond; + bool CurUnAnalyzable= TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); if (!CurUnAnalyzable) { // If the CFG for the prior block has extra edges, remove them. - MadeChange |= CorrectExtraCFGEdges(*MBB, CurTBB, CurFBB, - !CurCond.empty(), - ++MachineFunction::iterator(MBB)); + MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); // If this is a two-way branch, and the FBB branches to this block, reverse // the condition so the single-basic-block loop is faster. Instead of: @@ -797,7 +1051,7 @@ void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { // we want: // Loop: xxx; jncc Loop; jmp Out if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { - std::vector NewCond(CurCond); + SmallVector NewCond(CurCond); if (!TII->ReverseBranchCondition(NewCond)) { TII->RemoveBranch(*MBB); TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond); @@ -811,7 +1065,7 @@ void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { // If this branch is the only thing in its block, see if we can forward // other blocks across it. if (CurTBB && CurCond.empty() && CurFBB == 0 && - TII->isBranch(MBB->begin()->getOpcode()) && CurTBB != MBB) { + MBB->begin()->getDesc().isBranch() && CurTBB != MBB) { // This block may contain just an unconditional branch. Because there can // be 'non-branch terminators' in the block, try removing the branch and // then seeing if the block is empty. @@ -843,17 +1097,33 @@ void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { } // Iterate through all the predecessors, revectoring each in-turn. - MachineBasicBlock::pred_iterator PI = MBB->pred_begin(); + size_t PI = 0; bool DidChange = false; bool HasBranchToSelf = false; - while (PI != MBB->pred_end()) { - if (*PI == MBB) { + while(PI != MBB->pred_size()) { + MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); + if (PMBB == MBB) { // If this block has an uncond branch to itself, leave it. ++PI; HasBranchToSelf = true; } else { DidChange = true; - ReplaceUsesOfBlockWith(*PI, MBB, CurTBB, TII); + PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); + // If this change resulted in PMBB ending in a conditional + // branch where both conditions go to the same destination, + // change this to an unconditional branch (and fix the CFG). + MachineBasicBlock *NewCurTBB = 0, *NewCurFBB = 0; + SmallVector NewCurCond; + bool NewCurUnAnalyzable = TII->AnalyzeBranch(*PMBB, NewCurTBB, + NewCurFBB, NewCurCond, true); + if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { + TII->RemoveBranch(*PMBB); + NewCurCond.clear(); + TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond); + MadeChange = true; + ++NumBranchOpts; + PMBB->CorrectExtraCFGEdges(NewCurTBB, NewCurFBB, false); + } } } @@ -892,12 +1162,18 @@ void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { MachineBasicBlock *PredBB = *PI; MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; if (PredBB != MBB && !CanFallThrough(PredBB) + && (!CurFallsThru || !CurTBB || !CurFBB) && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { // If the current block doesn't fall through, just move it. // If the current block can fall through and does not end with a // conditional branch, we need to append an unconditional jump to // the (current) next block. To avoid a possible compile-time // infinite loop, move blocks only backward in this case. + // Also, if there are already 2 branches here, we cannot add a third; + // this means we have the case + // Bcc next + // B elsewhere + // next: if (CurFallsThru) { MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB)); CurCond.clear();