X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FPHIElimination.cpp;h=af65f13bf065eb202152145480a6bf8db7480724;hb=96cc1d0dfbcf9c7ffffc65f0aa008ff532d444f4;hp=d11e01a0c30d11647d8fe8bd28957c14e2be7a7c;hpb=74215fc29fa748e006c0309671555d5873bac56a;p=oota-llvm.git diff --git a/lib/CodeGen/PHIElimination.cpp b/lib/CodeGen/PHIElimination.cpp index d11e01a0c30..af65f13bf06 100644 --- a/lib/CodeGen/PHIElimination.cpp +++ b/lib/CodeGen/PHIElimination.cpp @@ -14,13 +14,15 @@ //===----------------------------------------------------------------------===// #define DEBUG_TYPE "phielim" -#include "PHIElimination.h" +#include "PHIEliminationUtils.h" #include "llvm/CodeGen/LiveVariables.h" #include "llvm/CodeGen/Passes.h" #include "llvm/CodeGen/MachineDominators.h" #include "llvm/CodeGen/MachineInstr.h" #include "llvm/CodeGen/MachineInstrBuilder.h" +#include "llvm/CodeGen/MachineLoopInfo.h" #include "llvm/CodeGen/MachineRegisterInfo.h" +#include "llvm/Target/TargetInstrInfo.h" #include "llvm/Function.h" #include "llvm/Target/TargetMachine.h" #include "llvm/ADT/SmallPtrSet.h" @@ -30,79 +32,133 @@ #include "llvm/Support/Compiler.h" #include "llvm/Support/Debug.h" #include -#include using namespace llvm; +static cl::opt +DisableEdgeSplitting("disable-phi-elim-edge-splitting", cl::init(false), + cl::Hidden, cl::desc("Disable critical edge splitting " + "during PHI elimination")); + +namespace { + class PHIElimination : public MachineFunctionPass { + MachineRegisterInfo *MRI; // Machine register information + + public: + static char ID; // Pass identification, replacement for typeid + PHIElimination() : MachineFunctionPass(ID) { + initializePHIEliminationPass(*PassRegistry::getPassRegistry()); + } + + virtual bool runOnMachineFunction(MachineFunction &Fn); + virtual void getAnalysisUsage(AnalysisUsage &AU) const; + + private: + /// EliminatePHINodes - Eliminate phi nodes by inserting copy instructions + /// in predecessor basic blocks. + /// + bool EliminatePHINodes(MachineFunction &MF, MachineBasicBlock &MBB); + void LowerAtomicPHINode(MachineBasicBlock &MBB, + MachineBasicBlock::iterator AfterPHIsIt); + + /// analyzePHINodes - Gather information about the PHI nodes in + /// here. In particular, we want to map the number of uses of a virtual + /// register which is used in a PHI node. We map that to the BB the + /// vreg is coming from. This is used later to determine when the vreg + /// is killed in the BB. + /// + void analyzePHINodes(const MachineFunction& Fn); + + /// Split critical edges where necessary for good coalescer performance. + bool SplitPHIEdges(MachineFunction &MF, MachineBasicBlock &MBB, + LiveVariables &LV, MachineLoopInfo *MLI); + + typedef std::pair BBVRegPair; + typedef DenseMap VRegPHIUse; + + VRegPHIUse VRegPHIUseCount; + + // Defs of PHI sources which are implicit_def. + SmallPtrSet ImpDefs; + + // Map reusable lowered PHI node -> incoming join register. + typedef DenseMap LoweredPHIMap; + LoweredPHIMap LoweredPHIs; + }; +} + STATISTIC(NumAtomic, "Number of atomic phis lowered"); -STATISTIC(NumSplits, "Number of critical edges split on demand"); +STATISTIC(NumCriticalEdgesSplit, "Number of critical edges split"); STATISTIC(NumReused, "Number of reused lowered phis"); char PHIElimination::ID = 0; -static RegisterPass -X("phi-node-elimination", "Eliminate PHI nodes for register allocation"); +INITIALIZE_PASS(PHIElimination, "phi-node-elimination", + "Eliminate PHI nodes for register allocation", false, false) -const PassInfo *const llvm::PHIEliminationID = &X; +char& llvm::PHIEliminationID = PHIElimination::ID; -void llvm::PHIElimination::getAnalysisUsage(AnalysisUsage &AU) const { +void PHIElimination::getAnalysisUsage(AnalysisUsage &AU) const { AU.addPreserved(); AU.addPreserved(); - // rdar://7401784 This would be nice: - // AU.addPreservedID(MachineLoopInfoID); + AU.addPreserved(); MachineFunctionPass::getAnalysisUsage(AU); } -bool llvm::PHIElimination::runOnMachineFunction(MachineFunction &Fn) { - MRI = &Fn.getRegInfo(); +bool PHIElimination::runOnMachineFunction(MachineFunction &MF) { + MRI = &MF.getRegInfo(); - PHIDefs.clear(); - PHIKills.clear(); bool Changed = false; // Split critical edges to help the coalescer - if (LiveVariables *LV = getAnalysisIfAvailable()) - for (MachineFunction::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) - Changed |= SplitPHIEdges(Fn, *I, *LV); + if (!DisableEdgeSplitting) { + if (LiveVariables *LV = getAnalysisIfAvailable()) { + MachineLoopInfo *MLI = getAnalysisIfAvailable(); + for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) + Changed |= SplitPHIEdges(MF, *I, *LV, MLI); + } + } // Populate VRegPHIUseCount - analyzePHINodes(Fn); + analyzePHINodes(MF); // Eliminate PHI instructions by inserting copies into predecessor blocks. - for (MachineFunction::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) - Changed |= EliminatePHINodes(Fn, *I); + for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) + Changed |= EliminatePHINodes(MF, *I); // Remove dead IMPLICIT_DEF instructions. - for (SmallPtrSet::iterator I = ImpDefs.begin(), + for (SmallPtrSet::iterator I = ImpDefs.begin(), E = ImpDefs.end(); I != E; ++I) { MachineInstr *DefMI = *I; unsigned DefReg = DefMI->getOperand(0).getReg(); - if (MRI->use_empty(DefReg)) + if (MRI->use_nodbg_empty(DefReg)) DefMI->eraseFromParent(); } // Clean up the lowered PHI instructions. for (LoweredPHIMap::iterator I = LoweredPHIs.begin(), E = LoweredPHIs.end(); I != E; ++I) - Fn.DeleteMachineInstr(I->first); - LoweredPHIs.clear(); + MF.DeleteMachineInstr(I->first); + LoweredPHIs.clear(); ImpDefs.clear(); VRegPHIUseCount.clear(); + return Changed; } /// EliminatePHINodes - Eliminate phi nodes by inserting copy instructions in /// predecessor basic blocks. /// -bool llvm::PHIElimination::EliminatePHINodes(MachineFunction &MF, +bool PHIElimination::EliminatePHINodes(MachineFunction &MF, MachineBasicBlock &MBB) { - if (MBB.empty() || MBB.front().getOpcode() != TargetInstrInfo::PHI) + if (MBB.empty() || !MBB.front().isPHI()) return false; // Quick exit for basic blocks without PHIs. // Get an iterator to the first instruction after the last PHI node (this may // also be the end of the basic block). - MachineBasicBlock::iterator AfterPHIsIt = SkipPHIsAndLabels(MBB, MBB.begin()); + MachineBasicBlock::iterator AfterPHIsIt = MBB.SkipPHIsAndLabels(MBB.begin()); - while (MBB.front().getOpcode() == TargetInstrInfo::PHI) + while (MBB.front().isPHI()) LowerAtomicPHINode(MBB, AfterPHIsIt); return true; @@ -115,64 +171,20 @@ static bool isSourceDefinedByImplicitDef(const MachineInstr *MPhi, for (unsigned i = 1; i != MPhi->getNumOperands(); i += 2) { unsigned SrcReg = MPhi->getOperand(i).getReg(); const MachineInstr *DefMI = MRI->getVRegDef(SrcReg); - if (!DefMI || DefMI->getOpcode() != TargetInstrInfo::IMPLICIT_DEF) + if (!DefMI || !DefMI->isImplicitDef()) return false; } return true; } -// FindCopyInsertPoint - Find a safe place in MBB to insert a copy from SrcReg -// when following the CFG edge to SuccMBB. This needs to be after any def of -// SrcReg, but before any subsequent point where control flow might jump out of -// the basic block. -MachineBasicBlock::iterator -llvm::PHIElimination::FindCopyInsertPoint(MachineBasicBlock &MBB, - MachineBasicBlock &SuccMBB, - unsigned SrcReg) { - // Handle the trivial case trivially. - if (MBB.empty()) - return MBB.begin(); - - // Usually, we just want to insert the copy before the first terminator - // instruction. However, for the edge going to a landing pad, we must insert - // the copy before the call/invoke instruction. - if (!SuccMBB.isLandingPad()) - return MBB.getFirstTerminator(); - - // Discover any defs/uses in this basic block. - SmallPtrSet DefUsesInMBB; - for (MachineRegisterInfo::reg_iterator RI = MRI->reg_begin(SrcReg), - RE = MRI->reg_end(); RI != RE; ++RI) { - MachineInstr *DefUseMI = &*RI; - if (DefUseMI->getParent() == &MBB) - DefUsesInMBB.insert(DefUseMI); - } - MachineBasicBlock::iterator InsertPoint; - if (DefUsesInMBB.empty()) { - // No defs. Insert the copy at the start of the basic block. - InsertPoint = MBB.begin(); - } else if (DefUsesInMBB.size() == 1) { - // Insert the copy immediately after the def/use. - InsertPoint = *DefUsesInMBB.begin(); - ++InsertPoint; - } else { - // Insert the copy immediately after the last def/use. - InsertPoint = MBB.end(); - while (!DefUsesInMBB.count(&*--InsertPoint)) {} - ++InsertPoint; - } - - // Make sure the copy goes after any phi nodes however. - return SkipPHIsAndLabels(MBB, InsertPoint); -} /// LowerAtomicPHINode - Lower the PHI node at the top of the specified block, /// under the assuption that it needs to be lowered in a way that supports /// atomic execution of PHIs. This lowering method is always correct all of the /// time. /// -void llvm::PHIElimination::LowerAtomicPHINode( +void PHIElimination::LowerAtomicPHINode( MachineBasicBlock &MBB, MachineBasicBlock::iterator AfterPHIsIt) { ++NumAtomic; @@ -181,11 +193,11 @@ void llvm::PHIElimination::LowerAtomicPHINode( unsigned NumSrcs = (MPhi->getNumOperands() - 1) / 2; unsigned DestReg = MPhi->getOperand(0).getReg(); + assert(MPhi->getOperand(0).getSubReg() == 0 && "Can't handle sub-reg PHIs"); bool isDead = MPhi->getOperand(0).isDead(); // Create a new register for the incoming PHI arguments. MachineFunction &MF = *MBB.getParent(); - const TargetRegisterClass *RC = MF.getRegInfo().getRegClass(DestReg); unsigned IncomingReg = 0; bool reusedIncoming = false; // Is IncomingReg reused from an earlier PHI? @@ -197,7 +209,7 @@ void llvm::PHIElimination::LowerAtomicPHINode( // If all sources of a PHI node are implicit_def, just emit an // implicit_def instead of a copy. BuildMI(MBB, AfterPHIsIt, MPhi->getDebugLoc(), - TII->get(TargetInstrInfo::IMPLICIT_DEF), DestReg); + TII->get(TargetOpcode::IMPLICIT_DEF), DestReg); else { // Can we reuse an earlier PHI node? This only happens for critical edges, // typically those created by tail duplication. @@ -207,17 +219,16 @@ void llvm::PHIElimination::LowerAtomicPHINode( IncomingReg = entry; reusedIncoming = true; ++NumReused; - DEBUG(errs() << "Reusing %reg" << IncomingReg << " for " << *MPhi); + DEBUG(dbgs() << "Reusing " << PrintReg(IncomingReg) << " for " << *MPhi); } else { + const TargetRegisterClass *RC = MF.getRegInfo().getRegClass(DestReg); entry = IncomingReg = MF.getRegInfo().createVirtualRegister(RC); } - TII->copyRegToReg(MBB, AfterPHIsIt, DestReg, IncomingReg, RC, RC); + BuildMI(MBB, AfterPHIsIt, MPhi->getDebugLoc(), + TII->get(TargetOpcode::COPY), DestReg) + .addReg(IncomingReg); } - // Record PHI def. - assert(!hasPHIDef(DestReg) && "Vreg has multiple phi-defs?"); - PHIDefs[DestReg] = &MBB; - // Update live variable information if there is any. LiveVariables *LV = getAnalysisIfAvailable(); if (LV) { @@ -228,13 +239,14 @@ void llvm::PHIElimination::LowerAtomicPHINode( // Increment use count of the newly created virtual register. VI.NumUses++; + LV->setPHIJoin(IncomingReg); // When we are reusing the incoming register, it may already have been // killed in this block. The old kill will also have been inserted at // AfterPHIsIt, so it appears before the current PHICopy. if (reusedIncoming) if (MachineInstr *OldKill = VI.findKill(&MBB)) { - DEBUG(errs() << "Remove old kill from " << *OldKill); + DEBUG(dbgs() << "Remove old kill from " << *OldKill); LV->removeVirtualRegisterKilled(IncomingReg, OldKill); DEBUG(MBB.dump()); } @@ -268,6 +280,8 @@ void llvm::PHIElimination::LowerAtomicPHINode( SmallPtrSet MBBsInsertedInto; for (int i = NumSrcs - 1; i >= 0; --i) { unsigned SrcReg = MPhi->getOperand(i*2+1).getReg(); + unsigned SrcSubReg = MPhi->getOperand(i*2+1).getSubReg(); + assert(TargetRegisterInfo::isVirtualRegister(SrcReg) && "Machine PHI Operands must all be virtual registers!"); @@ -275,13 +289,10 @@ void llvm::PHIElimination::LowerAtomicPHINode( // path the PHI. MachineBasicBlock &opBlock = *MPhi->getOperand(i*2+2).getMBB(); - // Record the kill. - PHIKills[SrcReg].insert(&opBlock); - // If source is defined by an implicit def, there is no need to insert a // copy. MachineInstr *DefMI = MRI->getVRegDef(SrcReg); - if (DefMI->getOpcode() == TargetInstrInfo::IMPLICIT_DEF) { + if (DefMI->isImplicitDef()) { ImpDefs.insert(DefMI); continue; } @@ -295,11 +306,12 @@ void llvm::PHIElimination::LowerAtomicPHINode( // Find a safe location to insert the copy, this may be the first terminator // in the block (or end()). MachineBasicBlock::iterator InsertPos = - FindCopyInsertPoint(opBlock, MBB, SrcReg); + findPHICopyInsertPoint(&opBlock, &MBB, SrcReg); // Insert the copy. if (!reusedIncoming && IncomingReg) - TII->copyRegToReg(opBlock, InsertPos, IncomingReg, SrcReg, RC, RC); + BuildMI(opBlock, InsertPos, MPhi->getDebugLoc(), + TII->get(TargetOpcode::COPY), IncomingReg).addReg(SrcReg, 0, SrcSubReg); // Now update live variable information if we have it. Otherwise we're done if (!LV) continue; @@ -335,6 +347,8 @@ void llvm::PHIElimination::LowerAtomicPHINode( #ifndef NDEBUG for (MachineBasicBlock::iterator TI = llvm::next(Term); TI != opBlock.end(); ++TI) { + if (TI->isDebugValue()) + continue; assert(!TI->readsRegister(SrcReg) && "Terminator instructions cannot use virtual registers unless" "they are the first terminator in a block!"); @@ -343,9 +357,13 @@ void llvm::PHIElimination::LowerAtomicPHINode( } else if (reusedIncoming || !IncomingReg) { // We may have to rewind a bit if we didn't insert a copy this time. KillInst = Term; - while (KillInst != opBlock.begin()) - if ((--KillInst)->readsRegister(SrcReg)) + while (KillInst != opBlock.begin()) { + --KillInst; + if (KillInst->isDebugValue()) + continue; + if (KillInst->readsRegister(SrcReg)) break; + } } else { // We just inserted this copy. KillInst = prior(InsertPos); @@ -371,113 +389,46 @@ void llvm::PHIElimination::LowerAtomicPHINode( /// used in a PHI node. We map that to the BB the vreg is coming from. This is /// used later to determine when the vreg is killed in the BB. /// -void llvm::PHIElimination::analyzePHINodes(const MachineFunction& Fn) { - for (MachineFunction::const_iterator I = Fn.begin(), E = Fn.end(); +void PHIElimination::analyzePHINodes(const MachineFunction& MF) { + for (MachineFunction::const_iterator I = MF.begin(), E = MF.end(); I != E; ++I) for (MachineBasicBlock::const_iterator BBI = I->begin(), BBE = I->end(); - BBI != BBE && BBI->getOpcode() == TargetInstrInfo::PHI; ++BBI) + BBI != BBE && BBI->isPHI(); ++BBI) for (unsigned i = 1, e = BBI->getNumOperands(); i != e; i += 2) ++VRegPHIUseCount[BBVRegPair(BBI->getOperand(i+1).getMBB()->getNumber(), BBI->getOperand(i).getReg())]; } -bool llvm::PHIElimination::SplitPHIEdges(MachineFunction &MF, - MachineBasicBlock &MBB, - LiveVariables &LV) { - if (MBB.empty() || MBB.front().getOpcode() != TargetInstrInfo::PHI) +bool PHIElimination::SplitPHIEdges(MachineFunction &MF, + MachineBasicBlock &MBB, + LiveVariables &LV, + MachineLoopInfo *MLI) { + if (MBB.empty() || !MBB.front().isPHI() || MBB.isLandingPad()) return false; // Quick exit for basic blocks without PHIs. + bool Changed = false; for (MachineBasicBlock::const_iterator BBI = MBB.begin(), BBE = MBB.end(); - BBI != BBE && BBI->getOpcode() == TargetInstrInfo::PHI; ++BBI) { + BBI != BBE && BBI->isPHI(); ++BBI) { for (unsigned i = 1, e = BBI->getNumOperands(); i != e; i += 2) { unsigned Reg = BBI->getOperand(i).getReg(); MachineBasicBlock *PreMBB = BBI->getOperand(i+1).getMBB(); // We break edges when registers are live out from the predecessor block // (not considering PHI nodes). If the register is live in to this block // anyway, we would gain nothing from splitting. - if (!LV.isLiveIn(Reg, MBB) && LV.isLiveOut(Reg, *PreMBB)) - SplitCriticalEdge(PreMBB, &MBB); + // Avoid splitting backedges of loops. It would introduce small + // out-of-line blocks into the loop which is very bad for code placement. + if (PreMBB != &MBB && + !LV.isLiveIn(Reg, MBB) && LV.isLiveOut(Reg, *PreMBB)) { + if (!MLI || + !(MLI->getLoopFor(PreMBB) == MLI->getLoopFor(&MBB) && + MLI->isLoopHeader(&MBB))) { + if (PreMBB->SplitCriticalEdge(&MBB, this)) { + Changed = true; + ++NumCriticalEdgesSplit; + } + } + } } } - return true; -} - -MachineBasicBlock *PHIElimination::SplitCriticalEdge(MachineBasicBlock *A, - MachineBasicBlock *B) { - assert(A && B && "Missing MBB end point"); - - MachineFunction *MF = A->getParent(); - - // We may need to update A's terminator, but we can't do that if AnalyzeBranch - // fails. If A uses a jump table, we won't touch it. - const TargetInstrInfo *TII = MF->getTarget().getInstrInfo(); - MachineBasicBlock *TBB = 0, *FBB = 0; - SmallVector Cond; - if (TII->AnalyzeBranch(*A, TBB, FBB, Cond)) - return NULL; - - ++NumSplits; - - MachineBasicBlock *NMBB = MF->CreateMachineBasicBlock(); - MF->insert(llvm::next(MachineFunction::iterator(A)), NMBB); - DEBUG(errs() << "PHIElimination splitting critical edge:" - " BB#" << A->getNumber() - << " -- BB#" << NMBB->getNumber() - << " -- BB#" << B->getNumber() << '\n'); - - A->ReplaceUsesOfBlockWith(B, NMBB); - A->updateTerminator(); - - // Insert unconditional "jump B" instruction in NMBB if necessary. - NMBB->addSuccessor(B); - if (!NMBB->isLayoutSuccessor(B)) { - Cond.clear(); - MF->getTarget().getInstrInfo()->InsertBranch(*NMBB, B, NULL, Cond); - } - - // Fix PHI nodes in B so they refer to NMBB instead of A - for (MachineBasicBlock::iterator i = B->begin(), e = B->end(); - i != e && i->getOpcode() == TargetInstrInfo::PHI; ++i) - for (unsigned ni = 1, ne = i->getNumOperands(); ni != ne; ni += 2) - if (i->getOperand(ni+1).getMBB() == A) - i->getOperand(ni+1).setMBB(NMBB); - - if (LiveVariables *LV=getAnalysisIfAvailable()) - LV->addNewBlock(NMBB, A, B); - - if (MachineDominatorTree *MDT=getAnalysisIfAvailable()) - MDT->addNewBlock(NMBB, A); - - return NMBB; -} - -unsigned -PHIElimination::PHINodeTraits::getHashValue(const MachineInstr *MI) { - if (!MI || MI==getEmptyKey() || MI==getTombstoneKey()) - return DenseMapInfo::getHashValue(MI); - unsigned hash = 0; - for (unsigned ni = 1, ne = MI->getNumOperands(); ni != ne; ni += 2) - hash = hash*37 + DenseMapInfo:: - getHashValue(BBVRegPair(MI->getOperand(ni+1).getMBB()->getNumber(), - MI->getOperand(ni).getReg())); - return hash; -} - -bool PHIElimination::PHINodeTraits::isEqual(const MachineInstr *LHS, - const MachineInstr *RHS) { - const MachineInstr *EmptyKey = getEmptyKey(); - const MachineInstr *TombstoneKey = getTombstoneKey(); - if (!LHS || !RHS || LHS==EmptyKey || RHS==EmptyKey || - LHS==TombstoneKey || RHS==TombstoneKey) - return LHS==RHS; - - unsigned ne = LHS->getNumOperands(); - if (ne != RHS->getNumOperands()) - return false; - // Ignore operand 0, the defined register. - for (unsigned ni = 1; ni != ne; ni += 2) - if (LHS->getOperand(ni).getReg() != RHS->getOperand(ni).getReg() || - LHS->getOperand(ni+1).getMBB() != RHS->getOperand(ni+1).getMBB()) - return false; - return true; + return Changed; }