X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FBranchFolding.cpp;h=d4b1e90fd54c8f530ba483edca7c8265f69a3800;hb=c92da3882ee4e18153bb36fcdf33af393aba8259;hp=dc527b9fd54973b3d68fc5abd0d42d71ce4bd0fb;hpb=6acfe12dd6d52c801f78c240528b7cb42fa91159;p=oota-llvm.git diff --git a/lib/CodeGen/BranchFolding.cpp b/lib/CodeGen/BranchFolding.cpp index dc527b9fd54..d4b1e90fd54 100644 --- a/lib/CodeGen/BranchFolding.cpp +++ b/lib/CodeGen/BranchFolding.cpp @@ -16,65 +16,99 @@ // //===----------------------------------------------------------------------===// +#define DEBUG_TYPE "branchfolding" #include "llvm/CodeGen/Passes.h" -#include "llvm/CodeGen/MachineDebugInfo.h" +#include "llvm/CodeGen/MachineModuleInfo.h" #include "llvm/CodeGen/MachineFunctionPass.h" #include "llvm/CodeGen/MachineJumpTableInfo.h" +#include "llvm/CodeGen/RegisterScavenging.h" #include "llvm/Target/TargetInstrInfo.h" #include "llvm/Target/TargetMachine.h" +#include "llvm/Target/MRegisterInfo.h" #include "llvm/Support/CommandLine.h" +#include "llvm/Support/Debug.h" #include "llvm/ADT/Statistic.h" #include "llvm/ADT/STLExtras.h" +#include using namespace llvm; -static Statistic<> NumDeadBlocks("branchfold", "Number of dead blocks removed"); -static Statistic<> NumBranchOpts("branchfold", "Number of branches optimized"); -static Statistic<> NumTailMerge ("branchfold", "Number of block tails merged"); - +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); namespace { + // Throttle for huge numbers of predecessors (compile speed problems) + cl::opt + TailMergeThreshold("tail-merge-threshold", + cl::desc("Max number of predecessors to consider tail merging"), + cl::init(100), cl::Hidden); + struct BranchFolder : public MachineFunctionPass { + static char ID; + explicit BranchFolder(bool defaultEnableTailMerge) : + MachineFunctionPass((intptr_t)&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"; } const TargetInstrInfo *TII; - MachineDebugInfo *MDI; + MachineModuleInfo *MMI; 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); + std::vector > MergePotentials; + const MRegisterInfo *RegInfo; + RegScavenger *RS; // Branch optzn. bool OptimizeBranches(MachineFunction &MF); void OptimizeBlock(MachineBasicBlock *MBB); void RemoveDeadBlock(MachineBasicBlock *MBB); + + bool CanFallThrough(MachineBasicBlock *CurBB); + bool CanFallThrough(MachineBasicBlock *CurBB, bool BranchUnAnalyzable, + MachineBasicBlock *TBB, MachineBasicBlock *FBB, + const std::vector &Cond); }; + 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. void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { assert(MBB->pred_empty() && "MBB must be dead!"); + DOUT << "\nRemoving MBB: " << *MBB; MachineFunction *MF = MBB->getParent(); // drop all successors. while (!MBB->succ_empty()) MBB->removeSuccessor(MBB->succ_end()-1); - // If there is DWARF info to active, check to see if there are any DWARF_LABEL - // records in the basic block. If so, unregister them from MachineDebugInfo. - if (MDI && !MBB->empty()) { - unsigned DWARF_LABELOpc = TII->getDWARF_LABELOpcode(); - assert(DWARF_LABELOpc && - "Target supports dwarf but didn't implement getDWARF_LABELOpcode!"); - + // 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 (MMI && !MBB->empty()) { for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E; ++I) { - if ((unsigned)I->getOpcode() == DWARF_LABELOpc) { + if ((unsigned)I->getOpcode() == TargetInstrInfo::LABEL) { // The label ID # is always operand #0, an immediate. - MDI->InvalidateLabel(I->getOperand(0).getImm()); + MMI->InvalidateLabel(I->getOperand(0).getImm()); } } } @@ -87,9 +121,20 @@ bool BranchFolder::runOnMachineFunction(MachineFunction &MF) { TII = MF.getTarget().getInstrInfo(); if (!TII) return false; - MDI = 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; + std::vector Cond; + if (!TII->AnalyzeBranch(*MBB, TBB, FBB, Cond)) + EverMadeChange |= MBB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); + } + + RegInfo = MF.getTarget().getRegisterInfo(); + RS = RegInfo->requiresRegisterScavenging(MF) ? new RegScavenger() : NULL; + + MMI = getAnalysisToUpdate(); + bool MadeChangeThisIteration = true; while (MadeChangeThisIteration) { MadeChangeThisIteration = false; @@ -145,6 +190,7 @@ bool BranchFolder::runOnMachineFunction(MachineFunction &MF) { } } + delete RS; return EverMadeChange; } @@ -187,10 +233,15 @@ 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. @@ -198,7 +249,7 @@ 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; @@ -220,7 +271,13 @@ 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; } @@ -250,21 +307,160 @@ void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, ++NumTailMerge; } -bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { - MadeChange = false; +/// SplitMBBAt - Given a machine basic block and an iterator into it, split the +/// MBB so that the part before the iterator falls into the part starting at the +/// iterator. This returns the new MBB. +MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, + MachineBasicBlock::iterator BBI1) { + // Create the fall-through block. + MachineFunction::iterator MBBI = &CurMBB; + MachineBasicBlock *NewMBB = new MachineBasicBlock(CurMBB.getBasicBlock()); + CurMBB.getParent()->getBasicBlockList().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); + } + + // Add an edge from CurMBB to NewMBB for the fall-through. + CurMBB.addSuccessor(NewMBB); - return false; + // Splice the code over. + NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); + + // For targets that use the register scavenger, we must maintain LiveIns. + if (RS) { + RS->enterBasicBlock(&CurMBB); + if (!CurMBB.empty()) + RS->forward(prior(CurMBB.end())); + BitVector RegsLiveAtExit(RegInfo->getNumRegs()); + RS->getRegsUsed(RegsLiveAtExit, false); + for (unsigned int i=0, e=RegInfo->getNumRegs(); i!=e; i++) + if (RegsLiveAtExit[i]) + NewMBB->addLiveIn(i); + } + + return NewMBB; +} + +/// 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) { + unsigned Time = 0; + for (; I != E; ++I) { + const TargetInstrDescriptor &TID = TII->get(I->getOpcode()); + if (TID.Flags & M_CALL_FLAG) + Time += 10; + else if (TID.Flags & (M_LOAD_FLAG|M_STORE_FLAG)) + Time += 2; + else + ++Time; + } + 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, + MachineBasicBlock *PredBB) { + // If one block is the entry block, split the other one; we can't generate + // a branch to the entry block, as its label is not emitted. + MachineBasicBlock *Entry = MBB1->getParent()->begin(); + if (MBB1 == Entry) + return false; + if (MBB2 == Entry) + return true; + + // If one block falls through into the common successor, choose that + // one to split; it is one instruction less to do that. + if (PredBB) { + if (MBB1 == PredBB) + return true; + else if (MBB2 == PredBB) + return false; + } + // 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); - // 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)); + // 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; + std::vector Cond; + if (I != MF->end() && + !TII->AnalyzeBranch(*CurMBB, TBB, FBB, Cond)) { + MachineBasicBlock *NextBB = I; + if (TBB == NextBB && Cond.size() && !FBB) { + if (!TII->ReverseBranchCondition(Cond)) { + TII->RemoveBranch(*CurMBB); + TII->InsertBranch(*CurMBB, SuccBB, NULL, Cond); + return; + } + } } + TII->InsertBranch(*CurMBB, SuccBB, NULL, std::vector()); +} + +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 + assert(0 && "Predecessor appears twice"); +#endif + return(false); + } +} + +// 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) { + unsigned minCommonTailLength = (SuccBB ? 1 : 2); + MadeChange = false; // 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) { @@ -275,69 +471,216 @@ bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { // If there is nothing that matches the hash of the current basic block, // give up. if (CurHash != PrevHash) { + if (SuccBB && CurMBB != PredBB) + FixTail(CurMBB, SuccBB, TII); MergePotentials.pop_back(); 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); + // Look through all the pairs of blocks that have the same hash as this + // one, and find the pair that has the largest number of instructions in + // common. + // Since instructions may get combined later (e.g. single stores into + // store multiple) this measure is not particularly accurate. + MachineBasicBlock::iterator 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, + unsigned FoundI = ~0U, FoundJ = ~0U; + unsigned maxCommonTailLength = 0U; + for (int i = MergePotentials.size()-1; + i != -1 && MergePotentials[i].first == CurHash; --i) { + for (int j = i-1; + j != -1 && MergePotentials[j].first == CurHash; --j) { + MachineBasicBlock::iterator TrialBBI1, TrialBBI2; + unsigned CommonTailLen = ComputeCommonTailLength( MergePotentials[i].second, - BBI1, BBI2); - if (CommonTailLen >= 2) { - FoundMatch = i; - break; + MergePotentials[j].second, + TrialBBI1, TrialBBI2); + if (CommonTailLen >= minCommonTailLength && + CommonTailLen > maxCommonTailLength) { + FoundI = i; + FoundJ = j; + maxCommonTailLength = CommonTailLen; + BBI1 = TrialBBI1; + BBI2 = TrialBBI2; } } - - // If we didn't find anything that has at least two instructions matching - // this one, bail out. - if (FoundMatch == ~0U) { + } + + // If we didn't find any pair that has at least minCommonTailLength + // instructions in common, bail out. All entries with this + // hash code can go away now. + if (FoundI == ~0U) { + for (int i = MergePotentials.size()-1; + i != -1 && MergePotentials[i].first == CurHash; --i) { + // Put the unconditional branch back, if we need one. + CurMBB = MergePotentials[i].second; + if (SuccBB && CurMBB != PredBB) + FixTail(CurMBB, SuccBB, TII); + MergePotentials.pop_back(); + } + continue; + } + + // Otherwise, move the block(s) to the right position(s). So that + // BBI1/2 will be valid, the last must be I and the next-to-last J. + if (FoundI != MergePotentials.size()-1) + std::swap(MergePotentials[FoundI], *(MergePotentials.end()-1)); + if (FoundJ != MergePotentials.size()-2) + std::swap(MergePotentials[FoundJ], *(MergePotentials.end()-2)); + + CurMBB = (MergePotentials.end()-1)->second; + MachineBasicBlock *MBB2 = (MergePotentials.end()-2)->second; + + // If neither block is the entire common tail, split the tail of one block + // to make it redundant with the other tail. Also, we cannot jump to the + // entry block, so if one block is the entry block, split the other one. + MachineBasicBlock *Entry = CurMBB->getParent()->begin(); + if (CurMBB->begin() == BBI1 && CurMBB != Entry) + ; // CurMBB is common tail + else if (MBB2->begin() == BBI2 && MBB2 != Entry) + ; // MBB2 is common tail + else { + if (0) { // Enable this to disable partial tail merges. MergePotentials.pop_back(); continue; } - // Otherwise, move the matching block to the right position. - std::swap(MergePotentials[FoundMatch], *(MergePotentials.end()-2)); + // Decide whether we want to split CurMBB or MBB2. + if (ShouldSplitFirstBlock(CurMBB, BBI1, MBB2, BBI2, TII, PredBB)) { + CurMBB = SplitMBBAt(*CurMBB, BBI1); + BBI1 = CurMBB->begin(); + MergePotentials.back().second = CurMBB; + } else { + MBB2 = SplitMBBAt(*MBB2, BBI2); + BBI2 = MBB2->begin(); + (MergePotentials.end()-2)->second = MBB2; + } } - // If either block is the entire common tail, make the longer one branch to - // the shorter one. - MachineBasicBlock *MBB2 = (MergePotentials.end()-2)->second; - if (CurMBB->begin() == BBI1) { - // 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); - MadeChange = true; - continue; - } else if (MBB2->begin() == BBI2) { + if (MBB2->begin() == BBI2 && MBB2 != Entry) { // 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(); - MadeChange = true; - continue; + } else { + assert(CurMBB->begin() == BBI1 && CurMBB != Entry && + "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); } - - MergePotentials.pop_back(); + 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) + 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->succ_empty() && 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; + std::vector Cond; + if (!TII->AnalyzeBranch(*PBB, TBB, FBB, Cond)) { + // Failing case: IBB is the target of a cbr, and + // we cannot reverse the branch. + std::vector NewCond(Cond); + if (Cond.size() && 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.size() == 0) { + 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.size()==0 || FBB)) { + TII->RemoveBranch(*PBB); + if (Cond.size()) + // 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 be either an original single predecessor, or 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 @@ -346,6 +689,9 @@ bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { bool BranchFolder::OptimizeBranches(MachineFunction &MF) { MadeChange = false; + // Make sure blocks are numbered in order + MF.RenumberBlocks(); + for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { MachineBasicBlock *MBB = I++; OptimizeBlock(MBB); @@ -361,101 +707,37 @@ 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; +/// 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 std::vector &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 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; - } - } + // 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; - 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 { - // 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 of the specified branch condition can transfer -/// control to FallthroughBlock, the block immediately after the branch. -static bool CanFallThrough(MachineBasicBlock *TBB, - MachineBasicBlock *FBB, - const std::vector &Cond, - MachineFunction::iterator FallthroughBlock) { // 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) == FallthroughBlock || - MachineFunction::iterator(FBB) == FallthroughBlock) + if (MachineFunction::iterator(TBB) == Fallthrough || + MachineFunction::iterator(FBB) == Fallthrough) return true; // If it's an unconditional branch to some block not the fall through, it @@ -467,6 +749,38 @@ static bool CanFallThrough(MachineBasicBlock *TBB, return FBB == 0; } +/// 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; + std::vector Cond; + bool CurUnAnalyzable = TII->AnalyzeBranch(*CurBB, TBB, FBB, Cond); + return CanFallThrough(CurBB, CurUnAnalyzable, TBB, FBB, Cond); +} + +/// IsBetterFallthrough - Return true if it would be clearly better to +/// fall-through to MBB1 than to fall through into MBB2. This has to return +/// 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) { + // 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; + + MachineInstr *MBB1I = --MBB1->end(); + MachineInstr *MBB2I = --MBB2->end(); + return TII.isCall(MBB2I->getOpcode()) && !TII.isCall(MBB1I->getOpcode()); +} + /// OptimizeBlock - Analyze and optimize control flow related to the specified /// block. This is never called on the entry block. void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { @@ -474,8 +788,9 @@ 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; @@ -486,7 +801,7 @@ 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 MBB was the target of a jump table, update jump tables to go to the @@ -504,12 +819,12 @@ void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { MachineBasicBlock *PriorTBB = 0, *PriorFBB = 0; std::vector PriorCond; - bool PriorUnAnalyzable = false; - PriorUnAnalyzable = TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); + bool PriorUnAnalyzable = + TII->AnalyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond); 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 @@ -556,17 +871,90 @@ void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { return OptimizeBlock(MBB); } } + + // If this block doesn't fall through (e.g. it ends with an uncond branch or + // has no successors) and if the pred falls through into this block, and if + // it would otherwise fall through into the block after this, move this + // block to the end of the function. + // + // We consider it more likely that execution will stay in the function (e.g. + // due to loops) than it is to exit it. This asserts in loops etc, moving + // the assert condition out of the loop body. + if (!PriorCond.empty() && PriorFBB == 0 && + MachineFunction::iterator(PriorTBB) == FallThrough && + !CanFallThrough(MBB)) { + bool DoTransform = true; + + // We have to be careful that the succs of PredBB aren't both no-successor + // blocks. If neither have successors and if PredBB is the second from + // last block in the function, we'd just keep swapping the two blocks for + // 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)) + 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. + std::vector NewPriorCond(PriorCond); + if (!TII->ReverseBranchCondition(NewPriorCond)) { + DOUT << "\nMoving MBB: " << *MBB; + DOUT << "To make fallthrough to: " << *PriorTBB << "\n"; + + TII->RemoveBranch(PrevBB); + TII->InsertBranch(PrevBB, MBB, 0, NewPriorCond); + + // Move this block to the end of the function. + MBB->moveAfter(--MBB->getParent()->end()); + MadeChange = true; + ++NumBranchOpts; + return; + } + } + } } // Analyze the branch in the current block. MachineBasicBlock *CurTBB = 0, *CurFBB = 0; std::vector CurCond; - if (!TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond)) { + bool CurUnAnalyzable = TII->AnalyzeBranch(*MBB, CurTBB, CurFBB, CurCond); + 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: + // Loop: xxx; jcc Out; jmp Loop + // we want: + // Loop: xxx; jncc Loop; jmp Out + if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { + std::vector NewCond(CurCond); + if (!TII->ReverseBranchCondition(NewCond)) { + TII->RemoveBranch(*MBB); + TII->InsertBranch(*MBB, CurFBB, CurTBB, NewCond); + MadeChange = true; + ++NumBranchOpts; + return OptimizeBlock(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 && @@ -602,17 +990,18 @@ 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); } } @@ -630,62 +1019,81 @@ void BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { // Add the branch back if the block is more than just an uncond branch. TII->InsertBranch(*MBB, CurTBB, 0, CurCond); } - - // 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 (!PriorUnAnalyzable && !CanFallThrough(PriorTBB, PriorFBB, - PriorCond, MBB)) { - // Now we know that there was no fall-through into this block, check to - // see if it has fall-throughs. - if (!CanFallThrough(CurTBB, CurFBB, CurCond, FallThrough)) { - - // Check all the predecessors of this block. If one of them has no fall - // throughs, move this block right after it. - for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), - E = MBB->pred_end(); PI != E; ++PI) { - // Analyze the branch at the end of the pred. - MachineBasicBlock *PredBB = *PI; - MachineFunction::iterator PredFallthrough = PredBB; ++PredFallthrough; - MachineBasicBlock *PredTBB = 0, *PredFBB = 0; - std::vector PredCond; - if (PredBB != MBB && - !TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond) && - !CanFallThrough(PredTBB, PredFBB, PredCond, PredFallthrough)) { - MBB->moveAfter(PredBB); - MadeChange = true; - return OptimizeBlock(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)) { + // 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); + + if (!MBB->isLandingPad()) { + // Check all the predecessors of this block. If one of them has no fall + // throughs, move this block right after it. + for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), + E = MBB->pred_end(); PI != E; ++PI) { + // Analyze the branch at the end of the pred. + 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(); + TII->InsertBranch(*MBB, NextBB, 0, CurCond); } + MBB->moveAfter(PredBB); + MadeChange = true; + return OptimizeBlock(MBB); } + } + } - // Check all successors to see if we can move this block before it. - for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), - E = MBB->succ_end(); SI != E; ++SI) { - // Analyze the branch at the end of the block before the succ. - MachineBasicBlock *SuccBB = *SI; - MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; - MachineBasicBlock *SuccPrevTBB = 0, *SuccPrevFBB = 0; - std::vector SuccPrevCond; - if (SuccBB != MBB && - !TII->AnalyzeBranch(*SuccPrev, SuccPrevTBB, SuccPrevFBB, - SuccPrevCond) && - !CanFallThrough(SuccPrevTBB, SuccPrevFBB, SuccPrevCond, SuccBB)) { - MBB->moveBefore(SuccBB); - MadeChange = true; - return OptimizeBlock(MBB); - } - } + if (!CurFallsThru) { + // Check all successors to see if we can move this block before it. + for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), + E = MBB->succ_end(); SI != E; ++SI) { + // Analyze the branch at the end of the block before the succ. + MachineBasicBlock *SuccBB = *SI; + MachineFunction::iterator SuccPrev = SuccBB; --SuccPrev; + std::vector SuccPrevCond; - // 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. - if (FallThrough != MBB->getParent()->end() && - CanFallThrough(PriorTBB, PriorFBB, PriorCond, FallThrough)) { - MBB->moveAfter(--MBB->getParent()->end()); + // If this block doesn't already fall-through to that successor, and if + // 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) && + !SuccBB->isLandingPad()) { + MBB->moveBefore(SuccBB); MadeChange = true; - return; + return OptimizeBlock(MBB); } } + + // 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. + if (FallThrough != MBB->getParent()->end() && + PrevBB.isSuccessor(FallThrough)) { + MBB->moveAfter(--MBB->getParent()->end()); + MadeChange = true; + return; + } } } }