X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;ds=sidebyside;f=lib%2FCodeGen%2FTailDuplication.cpp;h=8ebfbcae785b5f2c54bfa2e992c0b31eee8efd6c;hb=d25303533909f62eae18ee53ed0ff692174b092c;hp=76bae2ac6529e11ed4d070c28f58fe687671c3de;hpb=9a9a3a5c0fff9647a7a6454d974c058a4fb187d7;p=oota-llvm.git diff --git a/lib/CodeGen/TailDuplication.cpp b/lib/CodeGen/TailDuplication.cpp index 76bae2ac652..8ebfbcae785 100644 --- a/lib/CodeGen/TailDuplication.cpp +++ b/lib/CodeGen/TailDuplication.cpp @@ -25,6 +25,7 @@ #include "llvm/Support/Debug.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/raw_ostream.h" +#include "llvm/ADT/DenseSet.h" #include "llvm/ADT/SmallSet.h" #include "llvm/ADT/SetVector.h" #include "llvm/ADT/Statistic.h" @@ -55,10 +56,10 @@ typedef std::vector > AvailableValsTy; namespace { /// TailDuplicatePass - Perform tail duplication. class TailDuplicatePass : public MachineFunctionPass { - bool PreRegAlloc; const TargetInstrInfo *TII; MachineModuleInfo *MMI; MachineRegisterInfo *MRI; + bool PreRegAlloc; // SSAUpdateVRs - A list of virtual registers for which to update SSA form. SmallVector SSAUpdateVRs; @@ -69,11 +70,10 @@ namespace { public: static char ID; - explicit TailDuplicatePass(bool PreRA) : - MachineFunctionPass(ID), PreRegAlloc(PreRA) {} + explicit TailDuplicatePass() : + MachineFunctionPass(ID), PreRegAlloc(false) {} virtual bool runOnMachineFunction(MachineFunction &MF); - virtual const char *getPassName() const { return "Tail Duplication"; } private: void AddSSAUpdateEntry(unsigned OrigReg, unsigned NewReg, @@ -95,24 +95,38 @@ namespace { SmallSetVector &Succs); bool TailDuplicateBlocks(MachineFunction &MF); bool shouldTailDuplicate(const MachineFunction &MF, - MachineBasicBlock &TailBB); - bool TailDuplicate(MachineBasicBlock *TailBB, MachineFunction &MF, + bool IsSimple, MachineBasicBlock &TailBB); + bool isSimpleBB(MachineBasicBlock *TailBB); + bool canCompletelyDuplicateBB(MachineBasicBlock &BB); + bool duplicateSimpleBB(MachineBasicBlock *TailBB, + SmallVector &TDBBs, + const DenseSet &RegsUsedByPhi, + SmallVector &Copies); + bool TailDuplicate(MachineBasicBlock *TailBB, + bool IsSimple, + MachineFunction &MF, SmallVector &TDBBs, SmallVector &Copies); + bool TailDuplicateAndUpdate(MachineBasicBlock *MBB, + bool IsSimple, + MachineFunction &MF); + void RemoveDeadBlock(MachineBasicBlock *MBB); }; char TailDuplicatePass::ID = 0; } -FunctionPass *llvm::createTailDuplicatePass(bool PreRegAlloc) { - return new TailDuplicatePass(PreRegAlloc); -} +char &llvm::TailDuplicateID = TailDuplicatePass::ID; + +INITIALIZE_PASS(TailDuplicatePass, "tailduplication", "Tail Duplication", + false, false) bool TailDuplicatePass::runOnMachineFunction(MachineFunction &MF) { TII = MF.getTarget().getInstrInfo(); MRI = &MF.getRegInfo(); MMI = getAnalysisIfAvailable(); + PreRegAlloc = MRI->isSSA(); bool MadeChange = false; while (TailDuplicateBlocks(MF)) @@ -169,6 +183,109 @@ static void VerifyPHIs(MachineFunction &MF, bool CheckExtra) { } } +/// TailDuplicateAndUpdate - Tail duplicate the block and cleanup. +bool +TailDuplicatePass::TailDuplicateAndUpdate(MachineBasicBlock *MBB, + bool IsSimple, + MachineFunction &MF) { + // Save the successors list. + SmallSetVector Succs(MBB->succ_begin(), + MBB->succ_end()); + + SmallVector TDBBs; + SmallVector Copies; + if (!TailDuplicate(MBB, IsSimple, MF, TDBBs, Copies)) + return false; + + ++NumTails; + + SmallVector NewPHIs; + MachineSSAUpdater SSAUpdate(MF, &NewPHIs); + + // TailBB's immediate successors are now successors of those predecessors + // which duplicated TailBB. Add the predecessors as sources to the PHI + // instructions. + bool isDead = MBB->pred_empty() && !MBB->hasAddressTaken(); + if (PreRegAlloc) + UpdateSuccessorsPHIs(MBB, isDead, TDBBs, Succs); + + // If it is dead, remove it. + if (isDead) { + NumInstrDups -= MBB->size(); + RemoveDeadBlock(MBB); + ++NumDeadBlocks; + } + + // Update SSA form. + if (!SSAUpdateVRs.empty()) { + for (unsigned i = 0, e = SSAUpdateVRs.size(); i != e; ++i) { + unsigned VReg = SSAUpdateVRs[i]; + SSAUpdate.Initialize(VReg); + + // If the original definition is still around, add it as an available + // value. + MachineInstr *DefMI = MRI->getVRegDef(VReg); + MachineBasicBlock *DefBB = 0; + if (DefMI) { + DefBB = DefMI->getParent(); + SSAUpdate.AddAvailableValue(DefBB, VReg); + } + + // Add the new vregs as available values. + DenseMap::iterator LI = + SSAUpdateVals.find(VReg); + for (unsigned j = 0, ee = LI->second.size(); j != ee; ++j) { + MachineBasicBlock *SrcBB = LI->second[j].first; + unsigned SrcReg = LI->second[j].second; + SSAUpdate.AddAvailableValue(SrcBB, SrcReg); + } + + // Rewrite uses that are outside of the original def's block. + MachineRegisterInfo::use_iterator UI = MRI->use_begin(VReg); + while (UI != MRI->use_end()) { + MachineOperand &UseMO = UI.getOperand(); + MachineInstr *UseMI = &*UI; + ++UI; + if (UseMI->isDebugValue()) { + // SSAUpdate can replace the use with an undef. That creates + // a debug instruction that is a kill. + // FIXME: Should it SSAUpdate job to delete debug instructions + // instead of replacing the use with undef? + UseMI->eraseFromParent(); + continue; + } + if (UseMI->getParent() == DefBB && !UseMI->isPHI()) + continue; + SSAUpdate.RewriteUse(UseMO); + } + } + + SSAUpdateVRs.clear(); + SSAUpdateVals.clear(); + } + + // Eliminate some of the copies inserted by tail duplication to maintain + // SSA form. + for (unsigned i = 0, e = Copies.size(); i != e; ++i) { + MachineInstr *Copy = Copies[i]; + if (!Copy->isCopy()) + continue; + unsigned Dst = Copy->getOperand(0).getReg(); + unsigned Src = Copy->getOperand(1).getReg(); + MachineRegisterInfo::use_iterator UI = MRI->use_begin(Src); + if (++UI == MRI->use_end()) { + // Copy is the only use. Do trivial copy propagation here. + MRI->replaceRegWith(Dst, Src); + Copy->eraseFromParent(); + } + } + + if (NewPHIs.size()) + NumAddedPHIs += NewPHIs.size(); + + return true; +} + /// TailDuplicateBlocks - Look for small blocks that are unconditionally /// branched to and do not fall through. Tail-duplicate their instructions /// into their predecessors to eliminate (dynamic) branches. @@ -180,100 +297,22 @@ bool TailDuplicatePass::TailDuplicateBlocks(MachineFunction &MF) { VerifyPHIs(MF, true); } - SmallVector NewPHIs; - MachineSSAUpdater SSAUpdate(MF, &NewPHIs); - for (MachineFunction::iterator I = ++MF.begin(), E = MF.end(); I != E; ) { MachineBasicBlock *MBB = I++; if (NumTails == TailDupLimit) break; - // Save the successors list. - SmallSetVector Succs(MBB->succ_begin(), - MBB->succ_end()); - - SmallVector TDBBs; - SmallVector Copies; - if (TailDuplicate(MBB, MF, TDBBs, Copies)) { - ++NumTails; + bool IsSimple = isSimpleBB(MBB); - // TailBB's immediate successors are now successors of those predecessors - // which duplicated TailBB. Add the predecessors as sources to the PHI - // instructions. - bool isDead = MBB->pred_empty(); - if (PreRegAlloc) - UpdateSuccessorsPHIs(MBB, isDead, TDBBs, Succs); - - // If it is dead, remove it. - if (isDead) { - NumInstrDups -= MBB->size(); - RemoveDeadBlock(MBB); - ++NumDeadBlocks; - } - - // Update SSA form. - if (!SSAUpdateVRs.empty()) { - for (unsigned i = 0, e = SSAUpdateVRs.size(); i != e; ++i) { - unsigned VReg = SSAUpdateVRs[i]; - SSAUpdate.Initialize(VReg); - - // If the original definition is still around, add it as an available - // value. - MachineInstr *DefMI = MRI->getVRegDef(VReg); - MachineBasicBlock *DefBB = 0; - if (DefMI) { - DefBB = DefMI->getParent(); - SSAUpdate.AddAvailableValue(DefBB, VReg); - } - - // Add the new vregs as available values. - DenseMap::iterator LI = - SSAUpdateVals.find(VReg); - for (unsigned j = 0, ee = LI->second.size(); j != ee; ++j) { - MachineBasicBlock *SrcBB = LI->second[j].first; - unsigned SrcReg = LI->second[j].second; - SSAUpdate.AddAvailableValue(SrcBB, SrcReg); - } - - // Rewrite uses that are outside of the original def's block. - MachineRegisterInfo::use_iterator UI = MRI->use_begin(VReg); - while (UI != MRI->use_end()) { - MachineOperand &UseMO = UI.getOperand(); - MachineInstr *UseMI = &*UI; - ++UI; - if (UseMI->getParent() == DefBB && !UseMI->isPHI()) - continue; - SSAUpdate.RewriteUse(UseMO); - } - } - - SSAUpdateVRs.clear(); - SSAUpdateVals.clear(); - } - - // Eliminate some of the copies inserted by tail duplication to maintain - // SSA form. - for (unsigned i = 0, e = Copies.size(); i != e; ++i) { - MachineInstr *Copy = Copies[i]; - if (!Copy->isCopy()) - continue; - unsigned Dst = Copy->getOperand(0).getReg(); - unsigned Src = Copy->getOperand(1).getReg(); - MachineRegisterInfo::use_iterator UI = MRI->use_begin(Src); - if (++UI == MRI->use_end()) { - // Copy is the only use. Do trivial copy propagation here. - MRI->replaceRegWith(Dst, Src); - Copy->eraseFromParent(); - } - } + if (!shouldTailDuplicate(MF, IsSimple, *MBB)) + continue; - if (PreRegAlloc && TailDupVerify) - VerifyPHIs(MF, false); - MadeChange = true; - } + MadeChange |= TailDuplicateAndUpdate(MBB, IsSimple, MF); } - NumAddedPHIs += NewPHIs.size(); + + if (PreRegAlloc && TailDupVerify) + VerifyPHIs(MF, false); return MadeChange; } @@ -283,6 +322,8 @@ static bool isDefLiveOut(unsigned Reg, MachineBasicBlock *BB, for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(Reg), UE = MRI->use_end(); UI != UE; ++UI) { MachineInstr *UseMI = &*UI; + if (UseMI->isDebugValue()) + continue; if (UseMI->getParent() != BB) return true; } @@ -301,7 +342,7 @@ static unsigned getPHISrcRegOpIdx(MachineInstr *MI, MachineBasicBlock *SrcBB) { // used to determine which registers are liveout while modifying the // block (which is why we need to copy the information). static void getRegsUsedByPHIs(const MachineBasicBlock &BB, - DenseSet *UsedByPhi) { + DenseSet *UsedByPhi) { for(MachineBasicBlock::const_iterator I = BB.begin(), E = BB.end(); I != E; ++I) { const MachineInstr &MI = *I; @@ -337,7 +378,7 @@ void TailDuplicatePass::ProcessPHI(MachineInstr *MI, MachineBasicBlock *PredBB, DenseMap &LocalVRMap, SmallVector, 4> &Copies, - const DenseSet &RegsUsedByPhi, + const DenseSet &RegsUsedByPhi, bool Remove) { unsigned DefReg = MI->getOperand(0).getReg(); unsigned SrcOpIdx = getPHISrcRegOpIdx(MI, PredBB); @@ -392,7 +433,7 @@ void TailDuplicatePass::DuplicateInstruction(MachineInstr *MI, MO.setReg(VI->second); } } - PredBB->insert(PredBB->end(), NewMI); + PredBB->insert(PredBB->instr_end(), NewMI); } /// UpdateSuccessorsPHIs - After FromBB is tail duplicated into its predecessor @@ -485,11 +526,16 @@ TailDuplicatePass::UpdateSuccessorsPHIs(MachineBasicBlock *FromBB, bool isDead, /// shouldTailDuplicate - Determine if it is profitable to duplicate this block. bool TailDuplicatePass::shouldTailDuplicate(const MachineFunction &MF, + bool IsSimple, MachineBasicBlock &TailBB) { // Only duplicate blocks that end with unconditional branches. if (TailBB.canFallThrough()) return false; + // Don't try to tail-duplicate single-block loops. + if (TailBB.isSuccessor(&TailBB)) + return false; + // Set the limit on the cost to duplicate. When optimizing for size, // duplicate only one, because one branch instruction can be eliminated to // compensate for the duplication. @@ -500,73 +546,207 @@ TailDuplicatePass::shouldTailDuplicate(const MachineFunction &MF, else MaxDuplicateCount = TailDuplicateSize; - if (PreRegAlloc) { - if (TailBB.empty()) - return false; - const TargetInstrDesc &TID = TailBB.back().getDesc(); - // Pre-regalloc tail duplication hurts compile time and doesn't help - // much except for indirect branches. - if (!TID.isIndirectBranch()) - return false; - // If the target has hardware branch prediction that can handle indirect - // branches, duplicating them can often make them predictable when there - // are common paths through the code. The limit needs to be high enough - // to allow undoing the effects of tail merging and other optimizations - // that rearrange the predecessors of the indirect branch. - MaxDuplicateCount = 20; - } + // If the target has hardware branch prediction that can handle indirect + // branches, duplicating them can often make them predictable when there + // are common paths through the code. The limit needs to be high enough + // to allow undoing the effects of tail merging and other optimizations + // that rearrange the predecessors of the indirect branch. - // Don't try to tail-duplicate single-block loops. - if (TailBB.isSuccessor(&TailBB)) - return false; + bool HasIndirectbr = false; + if (!TailBB.empty()) + HasIndirectbr = TailBB.back().isIndirectBranch(); + + if (HasIndirectbr && PreRegAlloc) + MaxDuplicateCount = 20; // Check the instructions in the block to determine whether tail-duplication // is invalid or unlikely to be profitable. unsigned InstrCount = 0; - bool HasCall = false; - for (MachineBasicBlock::const_iterator I = TailBB.begin(); I != TailBB.end(); - ++I) { + for (MachineBasicBlock::iterator I = TailBB.begin(); I != TailBB.end(); ++I) { // Non-duplicable things shouldn't be tail-duplicated. - if (I->getDesc().isNotDuplicable()) return false; + if (I->isNotDuplicable()) + return false; + // Do not duplicate 'return' instructions if this is a pre-regalloc run. // A return may expand into a lot more instructions (e.g. reload of callee // saved registers) after PEI. - if (PreRegAlloc && I->getDesc().isReturn()) return false; - // Don't duplicate more than the threshold. - if (InstrCount == MaxDuplicateCount) return false; - // Remember if we saw a call. - if (I->getDesc().isCall()) HasCall = true; + if (PreRegAlloc && I->isReturn()) + return false; + + // Avoid duplicating calls before register allocation. Calls presents a + // barrier to register allocation so duplicating them may end up increasing + // spills. + if (PreRegAlloc && I->isCall()) + return false; + if (!I->isPHI() && !I->isDebugValue()) InstrCount += 1; + + if (InstrCount > MaxDuplicateCount) + return false; } - // Don't tail-duplicate calls before register allocation. Calls presents a - // barrier to register allocation so duplicating them may end up increasing - // spills. - if (InstrCount > 1 && (PreRegAlloc && HasCall)) + + if (HasIndirectbr && PreRegAlloc) + return true; + + if (IsSimple) + return true; + + if (!PreRegAlloc) + return true; + + return canCompletelyDuplicateBB(TailBB); +} + +/// isSimpleBB - True if this BB has only one unconditional jump. +bool +TailDuplicatePass::isSimpleBB(MachineBasicBlock *TailBB) { + if (TailBB->succ_size() != 1) + return false; + if (TailBB->pred_empty()) return false; + MachineBasicBlock::iterator I = TailBB->begin(); + MachineBasicBlock::iterator E = TailBB->end(); + while (I != E && I->isDebugValue()) + ++I; + if (I == E) + return true; + return I->isUnconditionalBranch(); +} +static bool +bothUsedInPHI(const MachineBasicBlock &A, + SmallPtrSet SuccsB) { + for (MachineBasicBlock::const_succ_iterator SI = A.succ_begin(), + SE = A.succ_end(); SI != SE; ++SI) { + MachineBasicBlock *BB = *SI; + if (SuccsB.count(BB) && !BB->empty() && BB->begin()->isPHI()) + return true; + } + + return false; +} + +bool +TailDuplicatePass::canCompletelyDuplicateBB(MachineBasicBlock &BB) { + SmallPtrSet Succs(BB.succ_begin(), BB.succ_end()); + + for (MachineBasicBlock::pred_iterator PI = BB.pred_begin(), + PE = BB.pred_end(); PI != PE; ++PI) { + MachineBasicBlock *PredBB = *PI; + + if (PredBB->succ_size() > 1) + return false; + + MachineBasicBlock *PredTBB = NULL, *PredFBB = NULL; + SmallVector PredCond; + if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)) + return false; + + if (!PredCond.empty()) + return false; + } return true; } +bool +TailDuplicatePass::duplicateSimpleBB(MachineBasicBlock *TailBB, + SmallVector &TDBBs, + const DenseSet &UsedByPhi, + SmallVector &Copies) { + SmallPtrSet Succs(TailBB->succ_begin(), + TailBB->succ_end()); + SmallVector Preds(TailBB->pred_begin(), + TailBB->pred_end()); + bool Changed = false; + for (SmallSetVector::iterator PI = Preds.begin(), + PE = Preds.end(); PI != PE; ++PI) { + MachineBasicBlock *PredBB = *PI; + + if (PredBB->getLandingPadSuccessor()) + continue; + + if (bothUsedInPHI(*PredBB, Succs)) + continue; + + MachineBasicBlock *PredTBB = NULL, *PredFBB = NULL; + SmallVector PredCond; + if (TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true)) + continue; + + Changed = true; + DEBUG(dbgs() << "\nTail-duplicating into PredBB: " << *PredBB + << "From simple Succ: " << *TailBB); + + MachineBasicBlock *NewTarget = *TailBB->succ_begin(); + MachineBasicBlock *NextBB = llvm::next(MachineFunction::iterator(PredBB)); + + // Make PredFBB explicit. + if (PredCond.empty()) + PredFBB = PredTBB; + + // Make fall through explicit. + if (!PredTBB) + PredTBB = NextBB; + if (!PredFBB) + PredFBB = NextBB; + + // Redirect + if (PredFBB == TailBB) + PredFBB = NewTarget; + if (PredTBB == TailBB) + PredTBB = NewTarget; + + // Make the branch unconditional if possible + if (PredTBB == PredFBB) { + PredCond.clear(); + PredFBB = NULL; + } + + // Avoid adding fall through branches. + if (PredFBB == NextBB) + PredFBB = NULL; + if (PredTBB == NextBB && PredFBB == NULL) + PredTBB = NULL; + + TII->RemoveBranch(*PredBB); + + if (PredTBB) + TII->InsertBranch(*PredBB, PredTBB, PredFBB, PredCond, DebugLoc()); + + PredBB->removeSuccessor(TailBB); + unsigned NumSuccessors = PredBB->succ_size(); + assert(NumSuccessors <= 1); + if (NumSuccessors == 0 || *PredBB->succ_begin() != NewTarget) + PredBB->addSuccessor(NewTarget); + + TDBBs.push_back(PredBB); + } + return Changed; +} + /// TailDuplicate - If it is profitable, duplicate TailBB's contents in each /// of its predecessors. bool -TailDuplicatePass::TailDuplicate(MachineBasicBlock *TailBB, MachineFunction &MF, +TailDuplicatePass::TailDuplicate(MachineBasicBlock *TailBB, + bool IsSimple, + MachineFunction &MF, SmallVector &TDBBs, SmallVector &Copies) { - if (!shouldTailDuplicate(MF, *TailBB)) - return false; - DEBUG(dbgs() << "\n*** Tail-duplicating BB#" << TailBB->getNumber() << '\n'); + DenseSet UsedByPhi; + getRegsUsedByPHIs(*TailBB, &UsedByPhi); + + if (IsSimple) + return duplicateSimpleBB(TailBB, TDBBs, UsedByPhi, Copies); + // Iterate through all the unique predecessors and tail-duplicate this // block into them, if possible. Copying the list ahead of time also // avoids trouble with the predecessor list reallocating. bool Changed = false; SmallSetVector Preds(TailBB->pred_begin(), TailBB->pred_end()); - DenseSet UsedByPhi; - getRegsUsedByPHIs(*TailBB, &UsedByPhi); for (SmallSetVector::iterator PI = Preds.begin(), PE = Preds.end(); PI != PE; ++PI) { MachineBasicBlock *PredBB = *PI; @@ -598,8 +778,10 @@ TailDuplicatePass::TailDuplicate(MachineBasicBlock *TailBB, MachineFunction &MF, // Clone the contents of TailBB into PredBB. DenseMap LocalVRMap; SmallVector, 4> CopyInfos; - MachineBasicBlock::iterator I = TailBB->begin(); - while (I != TailBB->end()) { + // Use instr_iterator here to properly handle bundles, e.g. + // ARM Thumb2 IT block. + MachineBasicBlock::instr_iterator I = TailBB->instr_begin(); + while (I != TailBB->instr_end()) { MachineInstr *MI = &*I; ++I; if (MI->isPHI()) { @@ -618,6 +800,10 @@ TailDuplicatePass::TailDuplicate(MachineBasicBlock *TailBB, MachineFunction &MF, TII->get(TargetOpcode::COPY), CopyInfos[i].first).addReg(CopyInfos[i].second)); } + + // Simplify + TII->AnalyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true); + NumInstrDups += TailBB->size() - 1; // subtract one for removed branch // Update the CFG. @@ -640,7 +826,7 @@ TailDuplicatePass::TailDuplicate(MachineBasicBlock *TailBB, MachineFunction &MF, SmallVector PriorCond; // This has to check PrevBB->succ_size() because EH edges are ignored by // AnalyzeBranch. - if (PrevBB->succ_size() == 1 && + if (PrevBB->succ_size() == 1 && !TII->AnalyzeBranch(*PrevBB, PriorTBB, PriorFBB, PriorCond, true) && PriorCond.empty() && !PriorTBB && TailBB->pred_size() == 1 && !TailBB->hasAddressTaken()) { @@ -665,6 +851,7 @@ TailDuplicatePass::TailDuplicate(MachineBasicBlock *TailBB, MachineFunction &MF, // Replace def of virtual registers with new registers, and update // uses with PHI source register or the new registers. MachineInstr *MI = &*I++; + assert(!MI->isBundle() && "Not expecting bundles before regalloc!"); DuplicateInstruction(MI, TailBB, PrevBB, MF, LocalVRMap, UsedByPhi); MI->eraseFromParent(); } @@ -753,4 +940,3 @@ void TailDuplicatePass::RemoveDeadBlock(MachineBasicBlock *MBB) { // Remove the block. MBB->eraseFromParent(); } -