X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FBranchFolding.cpp;h=8f519407ccd4e8c2890557861ab47989c01bf207;hb=22772214de79aa1c5ca38c4fb1da137d8fb30a05;hp=a91fc1ecf5a654fd3fab42130ce5114ae07a35e9;hpb=4e3f125e184f96ae72f2c44d16cafe0d44158283;p=oota-llvm.git diff --git a/lib/CodeGen/BranchFolding.cpp b/lib/CodeGen/BranchFolding.cpp index a91fc1ecf5a..8f519407ccd 100644 --- a/lib/CodeGen/BranchFolding.cpp +++ b/lib/CodeGen/BranchFolding.cpp @@ -32,6 +32,7 @@ #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/raw_ostream.h" #include "llvm/ADT/SmallSet.h" +#include "llvm/ADT/SetVector.h" #include "llvm/ADT/Statistic.h" #include "llvm/ADT/STLExtras.h" #include @@ -40,14 +41,36 @@ 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 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); +// Heuristic for tail merging (and, inversely, tail duplication). +// TODO: This should be replaced with a target query. +static cl::opt +TailMergeSize("tail-merge-size", + cl::desc("Min number of instructions to consider tail merging"), + cl::init(3), cl::Hidden); + +namespace { + /// BranchFolderPass - Wrap branch folder in a machine function pass. + class BranchFolderPass : public MachineFunctionPass, + public BranchFolder { + public: + static char ID; + explicit BranchFolderPass(bool defaultEnableTailMerge) + : MachineFunctionPass(&ID), BranchFolder(defaultEnableTailMerge) {} + + virtual bool runOnMachineFunction(MachineFunction &MF); + virtual const char *getPassName() const { return "Control Flow Optimizer"; } + }; +} char BranchFolderPass::ID = 0; @@ -75,24 +98,13 @@ BranchFolder::BranchFolder(bool defaultEnableTailMerge) { /// function, updating the CFG. void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { assert(MBB->pred_empty() && "MBB must be dead!"); - DEBUG(errs() << "\nRemoving MBB: " << *MBB); + DEBUG(dbgs() << "\nRemoving MBB: " << *MBB); MachineFunction *MF = MBB->getParent(); // drop all successors. while (!MBB->succ_empty()) MBB->removeSuccessor(MBB->succ_end()-1); - // 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 (I->isLabel()) - // The label ID # is always operand #0, an immediate. - MMI->InvalidateLabel(I->getOperand(0).getImm()); - } - } - // Remove the block. MF->erase(MBB); } @@ -110,7 +122,7 @@ bool BranchFolder::OptimizeImpDefsBlock(MachineBasicBlock *MBB) { SmallSet ImpDefRegs; MachineBasicBlock::iterator I = MBB->begin(); while (I != MBB->end()) { - if (I->getOpcode() != TargetInstrInfo::IMPLICIT_DEF) + if (!I->isImplicitDef()) break; unsigned Reg = I->getOperand(0).getReg(); ImpDefRegs.insert(Reg); @@ -172,7 +184,6 @@ bool BranchFolder::OptimizeFunction(MachineFunction &MF, MadeChange |= OptimizeImpDefsBlock(MBB); } - bool MadeChangeThisIteration = true; while (MadeChangeThisIteration) { MadeChangeThisIteration = false; @@ -181,56 +192,37 @@ bool BranchFolder::OptimizeFunction(MachineFunction &MF, MadeChange |= MadeChangeThisIteration; } - // See if any jump tables have become mergable or dead as the code generator + // See if any jump tables have become dead as the code generator // did its thing. MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); - const std::vector &JTs = JTI->getJumpTables(); - if (!JTs.empty()) { - // Figure out how these jump tables should be merged. - std::vector JTMapping; - JTMapping.reserve(JTs.size()); - - // We always keep the 0th jump table. - JTMapping.push_back(0); - - // Scan the jump tables, seeing if there are any duplicates. Note that this - // is N^2, which should be fixed someday. - for (unsigned i = 1, e = JTs.size(); i != e; ++i) { - if (JTs[i].MBBs.empty()) - JTMapping.push_back(i); - else - JTMapping.push_back(JTI->getJumpTableIndex(JTs[i].MBBs)); - } - - // 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. - 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.isJTI()) continue; - unsigned NewIdx = JTMapping[Op.getIndex()]; - Op.setIndex(NewIdx); - - // Remember that this JT is live. - JTIsLive.set(NewIdx); - } - } - - // Finally, remove dead jump tables. This happens either because the - // 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.test(i)) { - JTI->RemoveJumpTable(i); - MadeChange = true; + if (JTI == 0) { + delete RS; + return MadeChange; + } + + // Walk the function to find jump tables that are live. + BitVector JTIsLive(JTI->getJumpTables().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.isJTI()) continue; + + // Remember that this JT is live. + JTIsLive.set(Op.getIndex()); } } + // Finally, remove dead jump tables. This happens when the + // indirect jump was unreachable (and thus deleted). + for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) + if (!JTIsLive.test(i)) { + JTI->RemoveJumpTable(i); + MadeChange = true; + } + delete RS; return MadeChange; } @@ -285,12 +277,23 @@ static unsigned HashEndOfMBB(const MachineBasicBlock *MBB, return 0; // Empty MBB. --I; + // Skip debug info so it will not affect codegen. + while (I->isDebugValue()) { + if (I==MBB->begin()) + return 0; // MBB empty except for debug info. + --I; + } unsigned Hash = HashMachineInstr(I); if (I == MBB->begin() || minCommonTailLength == 1) return Hash; // Single instr MBB. --I; + while (I->isDebugValue()) { + if (I==MBB->begin()) + return Hash; // MBB with single non-debug instr. + --I; + } // Hash in the second-to-last instruction. Hash ^= HashMachineInstr(I) << 2; return Hash; @@ -309,18 +312,66 @@ static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, unsigned TailLen = 0; while (I1 != MBB1->begin() && I2 != MBB2->begin()) { --I1; --I2; + // Skip debugging pseudos; necessary to avoid changing the code. + while (I1->isDebugValue()) { + if (I1==MBB1->begin()) { + while (I2->isDebugValue()) { + if (I2==MBB2->begin()) + // I1==DBG at begin; I2==DBG at begin + return TailLen; + --I2; + } + ++I2; + // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin + return TailLen; + } + --I1; + } + // I1==first (untested) non-DBG preceding known match + while (I2->isDebugValue()) { + if (I2==MBB2->begin()) { + ++I1; + // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin + return TailLen; + } + --I2; + } + // I1, I2==first (untested) non-DBGs preceding known match 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->isInlineAsm()) { ++I1; ++I2; break; } ++TailLen; } + // Back past possible debugging pseudos at beginning of block. This matters + // when one block differs from the other only by whether debugging pseudos + // are present at the beginning. (This way, the various checks later for + // I1==MBB1->begin() work as expected.) + if (I1 == MBB1->begin() && I2 != MBB2->begin()) { + --I2; + while (I2->isDebugValue()) { + if (I2 == MBB2->begin()) { + return TailLen; + } + --I2; + } + ++I2; + } + if (I2 == MBB2->begin() && I1 != MBB1->begin()) { + --I1; + while (I1->isDebugValue()) { + if (I1 == MBB1->begin()) + return TailLen; + --I1; + } + ++I1; + } return TailLen; } @@ -373,7 +424,7 @@ MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, RS->forward(prior(CurMBB.end())); BitVector RegsLiveAtExit(TRI->getNumRegs()); RS->getRegsUsed(RegsLiveAtExit, false); - for (unsigned int i=0, e=TRI->getNumRegs(); i!=e; i++) + for (unsigned int i = 0, e = TRI->getNumRegs(); i != e; i++) if (RegsLiveAtExit[i]) NewMBB->addLiveIn(i); } @@ -387,6 +438,8 @@ static unsigned EstimateRuntime(MachineBasicBlock::iterator I, MachineBasicBlock::iterator E) { unsigned Time = 0; for (; I != E; ++I) { + if (I->isDebugValue()) + continue; const TargetInstrDesc &TID = I->getDesc(); if (TID.isCall()) Time += 10; @@ -402,10 +455,10 @@ static unsigned EstimateRuntime(MachineBasicBlock::iterator I, // 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, +static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB, const TargetInstrInfo *TII) { MachineFunction *MF = CurMBB->getParent(); - MachineFunction::iterator I = next(MachineFunction::iterator(CurMBB)); + MachineFunction::iterator I = llvm::next(MachineFunction::iterator(CurMBB)); MachineBasicBlock *TBB = 0, *FBB = 0; SmallVector Cond; if (I != MF->end() && @@ -422,24 +475,43 @@ static void FixTail(MachineBasicBlock* CurMBB, MachineBasicBlock *SuccBB, TII->InsertBranch(*CurMBB, SuccBB, NULL, SmallVector()); } -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. +bool +BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const { + if (getHash() < o.getHash()) + return true; + else if (getHash() > o.getHash()) + return false; + else if (getBlock()->getNumber() < o.getBlock()->getNumber()) + return true; + else if (getBlock()->getNumber() > o.getBlock()->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"); + llvm_unreachable("Predecessor appears twice"); #endif - return false; + return false; + } +} + +/// CountTerminators - Count the number of terminators in the given +/// block and set I to the position of the first non-terminator, if there +/// is one, or MBB->end() otherwise. +static unsigned CountTerminators(MachineBasicBlock *MBB, + MachineBasicBlock::iterator &I) { + I = MBB->end(); + unsigned NumTerms = 0; + for (;;) { + if (I == MBB->begin()) { + I = MBB->end(); + break; } + --I; + if (!I->getDesc().isTerminator()) break; + ++NumTerms; + } + return NumTerms; } /// ProfitableToMerge - Check if two machine basic blocks have a common tail @@ -451,21 +523,52 @@ static bool ProfitableToMerge(MachineBasicBlock *MBB1, unsigned minCommonTailLength, unsigned &CommonTailLen, MachineBasicBlock::iterator &I1, - MachineBasicBlock::iterator &I2) { + MachineBasicBlock::iterator &I2, + MachineBasicBlock *SuccBB, + MachineBasicBlock *PredBB) { CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); MachineFunction *MF = MBB1->getParent(); - if (CommonTailLen >= minCommonTailLength) - return true; - if (CommonTailLen == 0) return false; - // If we are optimizing for code size, 1 instruction in common is enough if - // we don't have to split a block. 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. - if (MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) && + // It's almost always profitable to merge any number of non-terminator + // instructions with the block that falls through into the common successor. + if (MBB1 == PredBB || MBB2 == PredBB) { + MachineBasicBlock::iterator I; + unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I); + if (CommonTailLen > NumTerms) + return true; + } + + // If one of the blocks can be completely merged and happens to be in + // a position where the other could fall through into it, merge any number + // of instructions, because it can be done without a branch. + // TODO: If the blocks are not adjacent, move one of them so that they are? + if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin()) + return true; + if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin()) + return true; + + // If both blocks have an unconditional branch temporarily stripped out, + // count that as an additional common instruction for the following + // heuristics. + unsigned EffectiveTailLen = CommonTailLen; + if (SuccBB && MBB1 != PredBB && MBB2 != PredBB && + !MBB1->back().getDesc().isBarrier() && + !MBB2->back().getDesc().isBarrier()) + ++EffectiveTailLen; + + // Check if the common tail is long enough to be worthwhile. + if (EffectiveTailLen >= minCommonTailLength) + return true; + + // If we are optimizing for code size, 2 instructions in common is enough if + // we don't have to split a block. At worst we will be introducing 1 new + // branch instruction, which is likely to be smaller than the 2 + // instructions that would be deleted in the merge. + if (EffectiveTailLen >= 2 && + MF->getFunction()->hasFnAttr(Attribute::OptimizeForSize) && (I1 == MBB1->begin() || I2 == MBB2->begin())) return true; @@ -483,28 +586,32 @@ static bool ProfitableToMerge(MachineBasicBlock *MBB1, /// those blocks appear in MergePotentials (where they are not necessarily /// consecutive). unsigned BranchFolder::ComputeSameTails(unsigned CurHash, - unsigned minCommonTailLength) { + unsigned minCommonTailLength, + MachineBasicBlock *SuccBB, + MachineBasicBlock *PredBB) { 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 != B && CurMPIter->getHash() == CurHash; --CurMPIter) { - for (MPIterator I = prior(CurMPIter); I->first == CurHash ; --I) { + for (MPIterator I = prior(CurMPIter); I->getHash() == CurHash ; --I) { unsigned CommonTailLen; - if (ProfitableToMerge(CurMPIter->second, I->second, minCommonTailLength, - CommonTailLen, TrialBBI1, TrialBBI2)) { + if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(), + minCommonTailLength, + CommonTailLen, TrialBBI1, TrialBBI2, + SuccBB, PredBB)) { if (CommonTailLen > maxCommonTailLength) { SameTails.clear(); maxCommonTailLength = CommonTailLen; HighestMPIter = CurMPIter; - SameTails.push_back(std::make_pair(CurMPIter, TrialBBI1)); + SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1)); } if (HighestMPIter == CurMPIter && CommonTailLen == maxCommonTailLength) - SameTails.push_back(std::make_pair(I, TrialBBI2)); + SameTails.push_back(SameTailElt(I, TrialBBI2)); } if (I == B) break; @@ -516,20 +623,20 @@ unsigned BranchFolder::ComputeSameTails(unsigned CurHash, /// 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) { + MachineBasicBlock *SuccBB, + MachineBasicBlock *PredBB) { MPIterator CurMPIter, B; for (CurMPIter = prior(MergePotentials.end()), B = MergePotentials.begin(); - CurMPIter->first == CurHash; + CurMPIter->getHash() == CurHash; --CurMPIter) { // Put the unconditional branch back, if we need one. - MachineBasicBlock *CurMBB = CurMPIter->second; + MachineBasicBlock *CurMBB = CurMPIter->getBlock(); if (SuccBB && CurMBB != PredBB) FixTail(CurMBB, SuccBB, TII); if (CurMPIter == B) break; } - if (CurMPIter->first!=CurHash) + if (CurMPIter->getHash() != CurHash) CurMPIter++; MergePotentials.erase(CurMPIter, MergePotentials.end()); } @@ -538,33 +645,36 @@ void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, /// only of the common tail. Create a block that does by splitting one. unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, unsigned maxCommonTailLength) { - unsigned i, commonTailIndex; + unsigned commonTailIndex = 0; unsigned TimeEstimate = ~0U; - for (i=0, commonTailIndex=0; isecond == PredBB) { + if (SameTails[i].getBlock() == 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); + unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(), + SameTails[i].getTailStartPos()); if (t <= TimeEstimate) { TimeEstimate = t; commonTailIndex = i; } } - MachineBasicBlock::iterator BBI = SameTails[commonTailIndex].second; - MachineBasicBlock *MBB = SameTails[commonTailIndex].first->second; + MachineBasicBlock::iterator BBI = + SameTails[commonTailIndex].getTailStartPos(); + MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); - DEBUG(errs() << "\nSplitting BB#" << MBB->getNumber() << ", size " + // If the common tail includes any debug info we will take it pretty + // randomly from one of the inputs. Might be better to remove it? + DEBUG(dbgs() << "\nSplitting BB#" << MBB->getNumber() << ", size " << maxCommonTailLength); MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI); - SameTails[commonTailIndex].first->second = newMBB; - SameTails[commonTailIndex].second = newMBB->begin(); + SameTails[commonTailIndex].setBlock(newMBB); + SameTails[commonTailIndex].setTailStartPos(newMBB->begin()); // If we split PredBB, newMBB is the new predecessor. if (PredBB == MBB) @@ -581,29 +691,43 @@ unsigned BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, // The lone predecessor of Succ that falls through into Succ, // if any, is given in PredBB. -bool BranchFolder::TryMergeBlocks(MachineBasicBlock *SuccBB, - MachineBasicBlock* PredBB) { +bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB, + MachineBasicBlock *PredBB) { bool MadeChange = false; - // 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; - - DEBUG(errs() << "\nTryMergeBlocks " << MergePotentials.size() << '\n'); + // Except for the special cases below, tail-merge if there are at least + // this many instructions in common. + unsigned minCommonTailLength = TailMergeSize; + + DEBUG(dbgs() << "\nTryTailMergeBlocks: "; + for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) + dbgs() << "BB#" << MergePotentials[i].getBlock()->getNumber() + << (i == e-1 ? "" : ", "); + dbgs() << "\n"; + if (SuccBB) { + dbgs() << " with successor BB#" << SuccBB->getNumber() << '\n'; + if (PredBB) + dbgs() << " which has fall-through from BB#" + << PredBB->getNumber() << "\n"; + } + dbgs() << "Looking for common tails of at least " + << minCommonTailLength << " instruction" + << (minCommonTailLength == 1 ? "" : "s") << '\n'; + ); // Sort by hash value so that blocks with identical end sequences sort // together. - std::stable_sort(MergePotentials.begin(), MergePotentials.end(),MergeCompare); + std::stable_sort(MergePotentials.begin(), MergePotentials.end()); // Walk through equivalence sets looking for actual exact matches. while (MergePotentials.size() > 1) { - unsigned CurHash = MergePotentials.back().first; + unsigned CurHash = MergePotentials.back().getHash(); // 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); + minCommonTailLength, + SuccBB, PredBB); // If we didn't find any pair that has at least minCommonTailLength // instructions in common, remove all blocks with this hash code and retry. @@ -616,38 +740,60 @@ bool BranchFolder::TryMergeBlocks(MachineBasicBlock *SuccBB, // 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) + MachineBasicBlock *EntryBB = MergePotentials.begin()->getBlock()-> + getParent()->begin(); + unsigned commonTailIndex = SameTails.size(); + // If there are two blocks, check to see if one can be made to fall through + // into the other. + if (SameTails.size() == 2 && + SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) && + SameTails[1].tailIsWholeBlock()) + commonTailIndex = 1; + else if (SameTails.size() == 2 && + SameTails[1].getBlock()->isLayoutSuccessor( + SameTails[0].getBlock()) && + SameTails[0].tailIsWholeBlock()) + commonTailIndex = 0; + else { + // Otherwise just pick one, favoring the fall-through predecessor if + // there is one. + for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { + MachineBasicBlock *MBB = SameTails[i].getBlock(); + if (MBB == EntryBB && SameTails[i].tailIsWholeBlock()) + continue; + if (MBB == PredBB) { + commonTailIndex = i; break; + } + if (SameTails[i].tailIsWholeBlock()) + commonTailIndex = i; } } - if (commonTailIndex == SameTails.size()) { + if (commonTailIndex == SameTails.size() || + (SameTails[commonTailIndex].getBlock() == PredBB && + !SameTails[commonTailIndex].tailIsWholeBlock())) { // None of the blocks consist entirely of the common tail. // Split a block so that one does. - commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength); + commonTailIndex = CreateCommonTailOnlyBlock(PredBB, maxCommonTailLength); } - MachineBasicBlock *MBB = SameTails[commonTailIndex].first->second; + MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); // MBB is common tail. Adjust all other BB's to jump to this one. // Traversal must be forwards so erases work. - DEBUG(errs() << "\nUsing common tail BB#" << MBB->getNumber() << " for "); - for (unsigned int i=0; igetNumber() + << " for "); + for (unsigned int i=0, e = SameTails.size(); i != e; ++i) { if (commonTailIndex == i) continue; - DEBUG(errs() << "BB#" << SameTails[i].first->second->getNumber() << ", "); + DEBUG(dbgs() << "BB#" << SameTails[i].getBlock()->getNumber() + << (i == e-1 ? "" : ", ")); // Hack the end off BB i, making it jump to BB commonTailIndex instead. - ReplaceTailWithBranchTo(SameTails[i].second, MBB); + ReplaceTailWithBranchTo(SameTails[i].getTailStartPos(), MBB); // BB i is no longer a predecessor of SuccBB; remove it from the worklist. - MergePotentials.erase(SameTails[i].first); + MergePotentials.erase(SameTails[i].getMPIter()); } - DEBUG(errs() << "\n"); + DEBUG(dbgs() << "\n"); // We leave commonTailIndex in the worklist in case there are other blocks // that match it with a smaller number of instructions. MadeChange = true; @@ -665,13 +811,13 @@ bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { 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)); + MergePotentials.push_back(MergePotentialsElt(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); + MadeChange |= TryTailMergeBlocks(NULL, NULL); // Look at blocks (IBB) with multiple predecessors (PBB). // We change each predecessor to a canonical form, by @@ -692,7 +838,8 @@ bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { // a compile-time infinite loop repeatedly doing and undoing the same // transformations.) - for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { + for (MachineFunction::iterator I = llvm::next(MF.begin()), E = MF.end(); + I != E; ++I) { if (I->pred_size() >= 2 && I->pred_size() < TailMergeThreshold) { SmallPtrSet UniquePreds; MachineBasicBlock *IBB = I; @@ -701,7 +848,7 @@ bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { for (MachineBasicBlock::pred_iterator P = I->pred_begin(), E2 = I->pred_end(); P != E2; ++P) { - MachineBasicBlock* PBB = *P; + MachineBasicBlock *PBB = *P; // Skip blocks that loop to themselves, can't tail merge these. if (PBB == IBB) continue; @@ -719,27 +866,27 @@ bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { continue; // This is the QBB case described above if (!FBB) - FBB = next(MachineFunction::iterator(PBB)); + FBB = llvm::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()) + MachineBasicBlock *PredNextBB = NULL; + if (IP != MF.end()) PredNextBB = IP; if (TBB == NULL) { - if (IBB!=PredNextBB) // fallthrough + if (IBB != PredNextBB) // fallthrough continue; } else if (FBB) { - if (TBB!=IBB && FBB!=IBB) // cbr then ubr + if (TBB != IBB && FBB != IBB) // cbr then ubr continue; } else if (Cond.empty()) { - if (TBB!=IBB) // ubr + if (TBB != IBB) // ubr continue; } else { - if (TBB!=IBB && IBB!=PredNextBB) // cbr + if (TBB != IBB && IBB != PredNextBB) // cbr continue; } } @@ -750,17 +897,18 @@ bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { // 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)); + MergePotentials.push_back(MergePotentialsElt(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); + if (MergePotentials.size() >= 2) + MadeChange |= TryTailMergeBlocks(IBB, PredBB); + // Reinsert an unconditional branch if needed. + // The 1 below can occur as a result of removing blocks in TryTailMergeBlocks. + PredBB = prior(I); // this may have been changed in TryTailMergeBlocks + if (MergePotentials.size() == 1 && + MergePotentials.begin()->getBlock() != PredBB) + FixTail(MergePotentials.begin()->getBlock(), IBB, TII); } } return MadeChange; @@ -790,62 +938,28 @@ bool BranchFolder::OptimizeBranches(MachineFunction &MF) { return MadeChange; } - -/// 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 -/// block after it, but it uses an explicit branch to do so (e.g. a table jump). -/// -/// True is a conservative answer. -/// -bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB, - bool BranchUnAnalyzable, - 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. - if (Fallthrough == CurBB->getParent()->end()) - return false; - - // If FallthroughBlock isn't a successor of CurBB, no fallthrough is possible. - if (!CurBB->isSuccessor(Fallthrough)) - return false; - - // If we couldn't analyze the branch, assume it could fall through. - if (BranchUnAnalyzable) return true; - - // If there is no branch, control always falls through. - if (TBB == 0) return true; - - // If there is some explicit branch to the fallthrough block, it can obviously - // reach, even though the branch should get folded to fall through implicitly. - if (MachineFunction::iterator(TBB) == Fallthrough || - MachineFunction::iterator(FBB) == Fallthrough) +// Blocks should be considered empty if they contain only debug info; +// else the debug info would affect codegen. +static bool IsEmptyBlock(MachineBasicBlock *MBB) { + if (MBB->empty()) return true; - - // If it's an unconditional branch to some block not the fall through, it - // doesn't fall through. - if (Cond.empty()) return false; - - // Otherwise, if it is conditional and has no explicit false block, it falls - // through. - return FBB == 0; + for (MachineBasicBlock::iterator MBBI = MBB->begin(), MBBE = MBB->end(); + MBBI!=MBBE; ++MBBI) { + if (!MBBI->isDebugValue()) + return false; + } + return true; } -/// CanFallThrough - Return true if the specified 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 block after it, but it uses an explicit -/// branch to do so (e.g. a table jump). -/// -/// True is a conservative answer. -/// -bool BranchFolder::CanFallThrough(MachineBasicBlock *CurBB) { - MachineBasicBlock *TBB = 0, *FBB = 0; - SmallVector Cond; - bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond, true); - return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond); +// Blocks with only debug info and branches should be considered the same +// as blocks with only branches. +static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) { + MachineBasicBlock::iterator MBBI, MBBE; + for (MBBI = MBB->begin(), MBBE = MBB->end(); MBBI!=MBBE; ++MBBI) { + if (!MBBI->isDebugValue()) + break; + } + return (MBBI->getDesc().isBranch()); } /// IsBetterFallthrough - Return true if it would be clearly better to @@ -858,15 +972,21 @@ static bool IsBetterFallthrough(MachineBasicBlock *MBB1, // 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 (IsEmptyBlock(MBB1) || IsEmptyBlock(MBB2)) 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(); + // Neither block consists entirely of debug info (per IsEmptyBlock check), + // so we needn't test for falling off the beginning here. + MachineBasicBlock::iterator MBB1I = --MBB1->end(); + while (MBB1I->isDebugValue()) + --MBB1I; + MachineBasicBlock::iterator MBB2I = --MBB2->end(); + while (MBB2I->isDebugValue()) + --MBB2I; return MBB2I->getDesc().isCall() && !MBB1I->getDesc().isCall(); } @@ -875,6 +995,7 @@ static bool IsBetterFallthrough(MachineBasicBlock *MBB1, bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { bool MadeChange = false; MachineFunction &MF = *MBB->getParent(); +ReoptimizeBlock: MachineFunction::iterator FallThrough = MBB; ++FallThrough; @@ -883,7 +1004,7 @@ bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { // explicitly. Landing pads should not do this since the landing-pad table // points to this block. Blocks with their addresses taken shouldn't be // optimized away. - if (MBB->empty() && !MBB->isLandingPad() && !MBB->hasAddressTaken()) { + if (IsEmptyBlock(MBB) && !MBB->isLandingPad() && !MBB->hasAddressTaken()) { // Dead block? Leave for cleanup later. if (MBB->pred_empty()) return MadeChange; @@ -898,7 +1019,8 @@ bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { } // If MBB was the target of a jump table, update jump tables to go to the // fallthrough instead. - MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, FallThrough); + if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) + MJTI->ReplaceMBBInJumpTables(MBB, FallThrough); MadeChange = true; } return MadeChange; @@ -927,7 +1049,27 @@ bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); MadeChange = true; ++NumBranchOpts; - return OptimizeBlock(MBB); + goto ReoptimizeBlock; + } + + // If the previous block unconditionally falls through to this block and + // this block has no other predecessors, move the contents of this block + // into the prior block. This doesn't usually happen when SimplifyCFG + // has been used, but it can happen if tail merging splits a fall-through + // predecessor of a block. + // This has to check PrevBB->succ_size() because EH edges are ignored by + // AnalyzeBranch. + if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 && + PrevBB.succ_size() == 1 && + !MBB->hasAddressTaken()) { + DEBUG(dbgs() << "\nMerging into block: " << PrevBB + << "From MBB: " << *MBB); + PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end()); + PrevBB.removeSuccessor(PrevBB.succ_begin());; + assert(PrevBB.succ_empty()); + PrevBB.transferSuccessors(MBB); + MadeChange = true; + return MadeChange; } // If the previous branch *only* branches to *this* block (conditional or @@ -936,7 +1078,7 @@ bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { TII->RemoveBranch(PrevBB); MadeChange = true; ++NumBranchOpts; - return OptimizeBlock(MBB); + goto ReoptimizeBlock; } // If the prior block branches somewhere else on the condition and here if @@ -946,7 +1088,7 @@ bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { TII->InsertBranch(PrevBB, PriorTBB, 0, PriorCond); MadeChange = true; ++NumBranchOpts; - return OptimizeBlock(MBB); + goto ReoptimizeBlock; } // If the prior block branches here on true and somewhere else on false, and @@ -959,7 +1101,7 @@ bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { TII->InsertBranch(PrevBB, PriorFBB, 0, NewPriorCond); MadeChange = true; ++NumBranchOpts; - return OptimizeBlock(MBB); + goto ReoptimizeBlock; } } @@ -973,7 +1115,7 @@ bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { // the assert condition out of the loop body. if (MBB->succ_empty() && !PriorCond.empty() && PriorFBB == 0 && MachineFunction::iterator(PriorTBB) == FallThrough && - !CanFallThrough(MBB)) { + !MBB->canFallThrough()) { bool DoTransform = true; // We have to be careful that the succs of PredBB aren't both no-successor @@ -985,27 +1127,11 @@ bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { !IsBetterFallthrough(PriorTBB, MBB)) DoTransform = false; - // We don't want to do this transformation if we have control flow like: - // br cond BB2 - // BB1: - // .. - // jmp BBX - // BB2: - // .. - // ret - // - // In this case, we could actually be moving the return block *into* a - // loop! - if (DoTransform && !MBB->succ_empty() && - (!CanFallThrough(PriorTBB) || PriorTBB->empty())) - DoTransform = false; - - if (DoTransform) { // Reverse the branch so we will fall through on the previous true cond. SmallVector NewPriorCond(PriorCond); if (!TII->ReverseBranchCondition(NewPriorCond)) { - DEBUG(errs() << "\nMoving MBB: " << *MBB + DEBUG(dbgs() << "\nMoving MBB: " << *MBB << "To make fallthrough to: " << *PriorTBB << "\n"); TII->RemoveBranch(PrevBB); @@ -1041,28 +1167,43 @@ bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond); MadeChange = true; ++NumBranchOpts; - return OptimizeBlock(MBB); + goto ReoptimizeBlock; } } - // 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 && - MBB->begin()->getDesc().isBranch() && CurTBB != MBB && + IsBranchOnlyBlock(MBB) && CurTBB != MBB && !MBB->hasAddressTaken()) { // 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. TII->RemoveBranch(*MBB); - + // If the only things remaining in the block are debug info, remove these + // as well, so this will behave the same as an empty block in non-debug + // mode. + if (!MBB->empty()) { + bool NonDebugInfoFound = false; + for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); + I != E; ++I) { + if (!I->isDebugValue()) { + NonDebugInfoFound = true; + break; + } + } + if (!NonDebugInfoFound) + // Make the block empty, losing the debug info (we could probably + // improve this in some cases.) + MBB->erase(MBB->begin(), MBB->end()); + } // If this block is just an unconditional branch to CurTBB, we can // usually completely eliminate the block. The only case we cannot // completely eliminate the block is when the block before this one // falls through into MBB and we can't understand the prior block's branch // condition. if (MBB->empty()) { - bool PredHasNoFallThrough = TII->BlockHasNoFallThrough(PrevBB); + bool PredHasNoFallThrough = !PrevBB.canFallThrough(); if (PredHasNoFallThrough || !PriorUnAnalyzable || !PrevBB.isSuccessor(MBB)) { // If the prior block falls through into us, turn it into an @@ -1107,13 +1248,14 @@ bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { TII->InsertBranch(*PMBB, NewCurTBB, 0, NewCurCond); MadeChange = true; ++NumBranchOpts; - PMBB->CorrectExtraCFGEdges(NewCurTBB, NewCurFBB, false); + PMBB->CorrectExtraCFGEdges(NewCurTBB, 0, false); } } } // Change any jumptables to go to the new MBB. - MF.getJumpTableInfo()->ReplaceMBBInJumpTables(MBB, CurTBB); + if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) + MJTI->ReplaceMBBInJumpTables(MBB, CurTBB); if (DidChange) { ++NumBranchOpts; MadeChange = true; @@ -1130,12 +1272,11 @@ bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { // If the prior block doesn't fall through into this block, and if this // block doesn't fall through into some other block, see if we can find a // place to move this block where a fall-through will happen. - if (!CanFallThrough(&PrevBB, PriorUnAnalyzable, - PriorTBB, PriorFBB, PriorCond)) { + if (!PrevBB.canFallThrough()) { + // Now we know that there was no fall-through into this block, check to // see if it has a fall-through into its successor. - bool CurFallsThru = CanFallThrough(MBB, CurUnAnalyzable, CurTBB, CurFBB, - CurCond); + bool CurFallsThru = MBB->canFallThrough(); if (!MBB->isLandingPad()) { // Check all the predecessors of this block. If one of them has no fall @@ -1145,7 +1286,10 @@ bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { // Analyze the branch at the end of the pred. MachineBasicBlock *PredBB = *PI; MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; - if (PredBB != MBB && !CanFallThrough(PredBB) + MachineBasicBlock *PredTBB = 0, *PredFBB = 0; + SmallVector PredCond; + if (PredBB != MBB && !PredBB->canFallThrough() && + !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) && (!CurFallsThru || !CurTBB || !CurFBB) && (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { // If the current block doesn't fall through, just move it. @@ -1159,13 +1303,13 @@ bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { // B elsewhere // next: if (CurFallsThru) { - MachineBasicBlock *NextBB = next(MachineFunction::iterator(MBB)); + MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(MBB)); CurCond.clear(); TII->InsertBranch(*MBB, NextBB, 0, CurCond); } MBB->moveAfter(PredBB); MadeChange = true; - return OptimizeBlock(MBB); + goto ReoptimizeBlock; } } } @@ -1182,18 +1326,22 @@ bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { // the succ doesn't already have a block that can fall through into it, // and if the successor isn't an EH destination, we can arrange for the // fallthrough to happen. - if (SuccBB != MBB && !CanFallThrough(SuccPrev) && + if (SuccBB != MBB && &*SuccPrev != MBB && + !SuccPrev->canFallThrough() && !CurUnAnalyzable && !SuccBB->isLandingPad()) { MBB->moveBefore(SuccBB); MadeChange = true; - return OptimizeBlock(MBB); + goto ReoptimizeBlock; } } // Okay, there is no really great place to put this block. If, however, // the block before this one would be a fall-through if this block were // removed, move this block to the end of the function. + MachineBasicBlock *PrevTBB = 0, *PrevFBB = 0; + SmallVector PrevCond; if (FallThrough != MF.end() && + !TII->AnalyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) && PrevBB.isSuccessor(FallThrough)) { MBB->moveAfter(--MF.end()); MadeChange = true;