X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;ds=sidebyside;f=lib%2FCodeGen%2FMachineCSE.cpp;h=7eda8c129dc47dccb9b87ece8c0aa21d1ac58bce;hb=f6fb7ed53c786228445fc55e8d495ccead59b9ae;hp=597d51d4f370c658b8f76753b623be6e1e6d5e49;hpb=db8771af286bc84267e7b5bd17eab51cd4ea552f;p=oota-llvm.git diff --git a/lib/CodeGen/MachineCSE.cpp b/lib/CodeGen/MachineCSE.cpp index 597d51d4f37..7eda8c129dc 100644 --- a/lib/CodeGen/MachineCSE.cpp +++ b/lib/CodeGen/MachineCSE.cpp @@ -20,14 +20,20 @@ #include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/Analysis/AliasAnalysis.h" #include "llvm/Target/TargetInstrInfo.h" +#include "llvm/ADT/DenseMap.h" #include "llvm/ADT/ScopedHashTable.h" +#include "llvm/ADT/SmallSet.h" #include "llvm/ADT/Statistic.h" #include "llvm/Support/Debug.h" +#include "llvm/Support/RecyclingAllocator.h" using namespace llvm; STATISTIC(NumCoalesces, "Number of copies coalesced"); STATISTIC(NumCSEs, "Number of common subexpression eliminated"); +STATISTIC(NumPhysCSEs, + "Number of physreg referencing common subexpr eliminated"); +STATISTIC(NumCommutes, "Number of copies coalesced after commuting"); namespace { class MachineCSE : public MachineFunctionPass { @@ -38,7 +44,9 @@ namespace { MachineRegisterInfo *MRI; public: static char ID; // Pass identification - MachineCSE() : MachineFunctionPass(&ID), CurrVN(0) {} + MachineCSE() : MachineFunctionPass(ID), LookAheadLimit(5), CurrVN(0) { + initializeMachineCSEPass(*PassRegistry::getPassRegistry()); + } virtual bool runOnMachineFunction(MachineFunction &MF); @@ -46,30 +54,57 @@ namespace { AU.setPreservesCFG(); MachineFunctionPass::getAnalysisUsage(AU); AU.addRequired(); + AU.addPreservedID(MachineLoopInfoID); AU.addRequired(); AU.addPreserved(); } + virtual void releaseMemory() { + ScopeMap.clear(); + Exps.clear(); + } + private: - unsigned CurrVN; - ScopedHashTable VNT; + const unsigned LookAheadLimit; + typedef RecyclingAllocator > AllocatorTy; + typedef ScopedHashTable ScopedHTType; + typedef ScopedHTType::ScopeTy ScopeType; + DenseMap ScopeMap; + ScopedHTType VNT; SmallVector Exps; + unsigned CurrVN; bool PerformTrivialCoalescing(MachineInstr *MI, MachineBasicBlock *MBB); bool isPhysDefTriviallyDead(unsigned Reg, MachineBasicBlock::const_iterator I, - MachineBasicBlock::const_iterator E); - bool hasLivePhysRegDefUse(MachineInstr *MI, MachineBasicBlock *MBB); + MachineBasicBlock::const_iterator E) const ; + bool hasLivePhysRegDefUses(const MachineInstr *MI, + const MachineBasicBlock *MBB, + SmallSet &PhysRefs) const; + bool PhysRegDefsReach(MachineInstr *CSMI, MachineInstr *MI, + SmallSet &PhysRefs) const; bool isCSECandidate(MachineInstr *MI); bool isProfitableToCSE(unsigned CSReg, unsigned Reg, MachineInstr *CSMI, MachineInstr *MI); - bool ProcessBlock(MachineDomTreeNode *Node); + void EnterScope(MachineBasicBlock *MBB); + void ExitScope(MachineBasicBlock *MBB); + bool ProcessBlock(MachineBasicBlock *MBB); + void ExitScopeIfDone(MachineDomTreeNode *Node, + DenseMap &OpenChildren, + DenseMap &ParentMap); + bool PerformCSE(MachineDomTreeNode *Node); }; } // end anonymous namespace char MachineCSE::ID = 0; -static RegisterPass -X("machine-cse", "Machine Common Subexpression Elimination"); +INITIALIZE_PASS_BEGIN(MachineCSE, "machine-cse", + "Machine Common Subexpression Elimination", false, false) +INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) +INITIALIZE_AG_DEPENDENCY(AliasAnalysis) +INITIALIZE_PASS_END(MachineCSE, "machine-cse", + "Machine Common Subexpression Elimination", false, false) FunctionPass *llvm::createMachineCSEPass() { return new MachineCSE(); } @@ -81,42 +116,41 @@ bool MachineCSE::PerformTrivialCoalescing(MachineInstr *MI, if (!MO.isReg() || !MO.isUse()) continue; unsigned Reg = MO.getReg(); - if (!Reg || TargetRegisterInfo::isPhysicalRegister(Reg)) + if (!TargetRegisterInfo::isVirtualRegister(Reg)) continue; - if (!MRI->hasOneUse(Reg)) + if (!MRI->hasOneNonDBGUse(Reg)) // Only coalesce single use copies. This ensure the copy will be // deleted. continue; MachineInstr *DefMI = MRI->getVRegDef(Reg); if (DefMI->getParent() != MBB) continue; - unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx; - if (TII->isMoveInstr(*DefMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx) && - TargetRegisterInfo::isVirtualRegister(SrcReg) && - !SrcSubIdx && !DstSubIdx) { - const TargetRegisterClass *SRC = MRI->getRegClass(SrcReg); - const TargetRegisterClass *RC = MRI->getRegClass(Reg); - const TargetRegisterClass *NewRC = getCommonSubClass(RC, SRC); - if (!NewRC) - continue; - DEBUG(dbgs() << "Coalescing: " << *DefMI); - DEBUG(dbgs() << "*** to: " << *MI); - MO.setReg(SrcReg); - if (NewRC != SRC) - MRI->setRegClass(SrcReg, NewRC); - DefMI->eraseFromParent(); - ++NumCoalesces; - Changed = true; - } + if (!DefMI->isCopy()) + continue; + unsigned SrcReg = DefMI->getOperand(1).getReg(); + if (!TargetRegisterInfo::isVirtualRegister(SrcReg)) + continue; + if (DefMI->getOperand(0).getSubReg() || DefMI->getOperand(1).getSubReg()) + continue; + if (!MRI->constrainRegClass(SrcReg, MRI->getRegClass(Reg))) + continue; + DEBUG(dbgs() << "Coalescing: " << *DefMI); + DEBUG(dbgs() << "*** to: " << *MI); + MO.setReg(SrcReg); + MRI->clearKillFlags(SrcReg); + DefMI->eraseFromParent(); + ++NumCoalesces; + Changed = true; } return Changed; } -bool MachineCSE::isPhysDefTriviallyDead(unsigned Reg, - MachineBasicBlock::const_iterator I, - MachineBasicBlock::const_iterator E) { - unsigned LookAheadLeft = 5; +bool +MachineCSE::isPhysDefTriviallyDead(unsigned Reg, + MachineBasicBlock::const_iterator I, + MachineBasicBlock::const_iterator E) const { + unsigned LookAheadLeft = LookAheadLimit; while (LookAheadLeft) { // Skip over dbg_value's. while (I != E && I->isDebugValue()) @@ -134,6 +168,7 @@ bool MachineCSE::isPhysDefTriviallyDead(unsigned Reg, if (!TRI->regsOverlap(MO.getReg(), Reg)) continue; if (MO.isUse()) + // Found a use! return false; SeenDef = true; } @@ -148,46 +183,71 @@ bool MachineCSE::isPhysDefTriviallyDead(unsigned Reg, return false; } -/// hasLivePhysRegDefUse - Return true if the specified instruction read / write -/// physical registers (except for dead defs of physical registers). -bool MachineCSE::hasLivePhysRegDefUse(MachineInstr *MI, MachineBasicBlock *MBB){ - unsigned PhysDef = 0; +/// hasLivePhysRegDefUses - Return true if the specified instruction read/write +/// physical registers (except for dead defs of physical registers). It also +/// returns the physical register def by reference if it's the only one and the +/// instruction does not uses a physical register. +bool MachineCSE::hasLivePhysRegDefUses(const MachineInstr *MI, + const MachineBasicBlock *MBB, + SmallSet &PhysRefs) const { + MachineBasicBlock::const_iterator I = MI; I = llvm::next(I); for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { - MachineOperand &MO = MI->getOperand(i); + const MachineOperand &MO = MI->getOperand(i); if (!MO.isReg()) continue; unsigned Reg = MO.getReg(); if (!Reg) continue; - if (TargetRegisterInfo::isPhysicalRegister(Reg)) { - if (MO.isUse()) - // Can't touch anything to read a physical register. - return true; - if (MO.isDead()) - // If the def is dead, it's ok. - continue; - // Ok, this is a physical register def that's not marked "dead". That's - // common since this pass is run before livevariables. We can scan - // forward a few instructions and check if it is obviously dead. - if (PhysDef) - // Multiple physical register defs. These are rare, forget about it. - return true; - PhysDef = Reg; - } + if (TargetRegisterInfo::isVirtualRegister(Reg)) + continue; + // If the def is dead, it's ok. But the def may not marked "dead". That's + // common since this pass is run before livevariables. We can scan + // forward a few instructions and check if it is obviously dead. + if (MO.isDef() && + (MO.isDead() || isPhysDefTriviallyDead(Reg, I, MBB->end()))) + continue; + PhysRefs.insert(Reg); + for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) + PhysRefs.insert(*Alias); } - if (PhysDef) { - MachineBasicBlock::iterator I = MI; I = llvm::next(I); - if (!isPhysDefTriviallyDead(PhysDef, I, MBB->end())) + return !PhysRefs.empty(); +} + +bool MachineCSE::PhysRegDefsReach(MachineInstr *CSMI, MachineInstr *MI, + SmallSet &PhysRefs) const { + // For now conservatively returns false if the common subexpression is + // not in the same basic block as the given instruction. + MachineBasicBlock *MBB = MI->getParent(); + if (CSMI->getParent() != MBB) + return false; + MachineBasicBlock::const_iterator I = CSMI; I = llvm::next(I); + MachineBasicBlock::const_iterator E = MI; + unsigned LookAheadLeft = LookAheadLimit; + while (LookAheadLeft) { + // Skip over dbg_value's. + while (I != E && I->isDebugValue()) + ++I; + + if (I == E) return true; + + for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { + const MachineOperand &MO = I->getOperand(i); + if (!MO.isReg() || !MO.isDef()) + continue; + unsigned MOReg = MO.getReg(); + if (TargetRegisterInfo::isVirtualRegister(MOReg)) + continue; + if (PhysRefs.count(MOReg)) + return false; + } + + --LookAheadLeft; + ++I; } - return false; -} -static bool isCopy(const MachineInstr *MI, const TargetInstrInfo *TII) { - unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx; - return TII->isMoveInstr(*MI, SrcReg, DstReg, SrcSubIdx, DstSubIdx) || - MI->isExtractSubreg() || MI->isInsertSubreg() || MI->isSubregToReg(); + return false; } bool MachineCSE::isCSECandidate(MachineInstr *MI) { @@ -196,16 +256,16 @@ bool MachineCSE::isCSECandidate(MachineInstr *MI) { return false; // Ignore copies. - if (isCopy(MI, TII)) + if (MI->isCopyLike()) return false; // Ignore stuff that we obviously can't move. - const TargetInstrDesc &TID = MI->getDesc(); - if (TID.mayStore() || TID.isCall() || TID.isTerminator() || - TID.hasUnmodeledSideEffects()) + const MCInstrDesc &MCID = MI->getDesc(); + if (MCID.mayStore() || MCID.isCall() || MCID.isTerminator() || + MI->hasUnmodeledSideEffects()) return false; - if (TID.mayLoad()) { + if (MCID.mayLoad()) { // Okay, this instruction does a load. As a refinement, we allow the target // to decide whether the loaded value is actually a constant. If so, we can // actually use it as a load. @@ -224,14 +284,13 @@ bool MachineCSE::isProfitableToCSE(unsigned CSReg, unsigned Reg, MachineInstr *CSMI, MachineInstr *MI) { // FIXME: Heuristics that works around the lack the live range splitting. - // Heuristics #1: Don't cse "cheap" computating if the def is not local or in an - // immediate predecessor. We don't want to increase register pressure and end up - // causing other computation to be spilled. + // Heuristics #1: Don't CSE "cheap" computation if the def is not local or in + // an immediate predecessor. We don't want to increase register pressure and + // end up causing other computation to be spilled. if (MI->getDesc().isAsCheapAsAMove()) { MachineBasicBlock *CSBB = CSMI->getParent(); MachineBasicBlock *BB = MI->getParent(); - if (CSBB != BB && - find(CSBB->succ_begin(), CSBB->succ_end(), BB) == CSBB->succ_end()) + if (CSBB != BB && !CSBB->isSuccessor(BB)) return false; } @@ -240,7 +299,7 @@ bool MachineCSE::isProfitableToCSE(unsigned CSReg, unsigned Reg, bool HasVRegUse = false; for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { const MachineOperand &MO = MI->getOperand(i); - if (MO.isReg() && MO.isUse() && MO.getReg() && + if (MO.isReg() && MO.isUse() && TargetRegisterInfo::isVirtualRegister(MO.getReg())) { HasVRegUse = true; break; @@ -252,7 +311,7 @@ bool MachineCSE::isProfitableToCSE(unsigned CSReg, unsigned Reg, E = MRI->use_nodbg_end(); I != E; ++I) { MachineInstr *Use = &*I; // Ignore copies. - if (!isCopy(Use, TII)) { + if (!Use->isCopyLike()) { HasNonCopyUse = true; break; } @@ -277,13 +336,24 @@ bool MachineCSE::isProfitableToCSE(unsigned CSReg, unsigned Reg, return CSBBs.count(MI->getParent()); } -bool MachineCSE::ProcessBlock(MachineDomTreeNode *Node) { +void MachineCSE::EnterScope(MachineBasicBlock *MBB) { + DEBUG(dbgs() << "Entering: " << MBB->getName() << '\n'); + ScopeType *Scope = new ScopeType(VNT); + ScopeMap[MBB] = Scope; +} + +void MachineCSE::ExitScope(MachineBasicBlock *MBB) { + DEBUG(dbgs() << "Exiting: " << MBB->getName() << '\n'); + DenseMap::iterator SI = ScopeMap.find(MBB); + assert(SI != ScopeMap.end()); + ScopeMap.erase(SI); + delete SI->second; +} + +bool MachineCSE::ProcessBlock(MachineBasicBlock *MBB) { bool Changed = false; SmallVector, 8> CSEPairs; - ScopedHashTableScope VNTS(VNT); - MachineBasicBlock *MBB = Node->getBlock(); for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E; ) { MachineInstr *MI = &*I; ++I; @@ -295,19 +365,48 @@ bool MachineCSE::ProcessBlock(MachineDomTreeNode *Node) { if (!FoundCSE) { // Look for trivial copy coalescing opportunities. if (PerformTrivialCoalescing(MI, MBB)) { + Changed = true; + // After coalescing MI itself may become a copy. - if (isCopy(MI, TII)) + if (MI->isCopyLike()) continue; FoundCSE = VNT.count(MI); } } - // FIXME: commute commutable instructions? - // If the instruction defines a physical register and the value *may* be + // Commute commutable instructions. + bool Commuted = false; + if (!FoundCSE && MI->getDesc().isCommutable()) { + MachineInstr *NewMI = TII->commuteInstruction(MI); + if (NewMI) { + Commuted = true; + FoundCSE = VNT.count(NewMI); + if (NewMI != MI) { + // New instruction. It doesn't need to be kept. + NewMI->eraseFromParent(); + Changed = true; + } else if (!FoundCSE) + // MI was changed but it didn't help, commute it back! + (void)TII->commuteInstruction(MI); + } + } + + // If the instruction defines physical registers and the values *may* be // used, then it's not safe to replace it with a common subexpression. - if (FoundCSE && hasLivePhysRegDefUse(MI, MBB)) + // It's also not safe if the instruction uses physical registers. + SmallSet PhysRefs; + if (FoundCSE && hasLivePhysRegDefUses(MI, MBB, PhysRefs)) { FoundCSE = false; + // ... Unless the CS is local and it also defines the physical register + // which is not clobbered in between and the physical register uses + // were not clobbered. + unsigned CSVN = VNT.lookup(MI); + MachineInstr *CSMI = Exps[CSVN]; + if (PhysRegDefsReach(CSMI, MI, PhysRefs)) + FoundCSE = true; + } + if (!FoundCSE) { VNT.insert(MI, CurrVN++); Exps.push_back(MI); @@ -331,23 +430,41 @@ bool MachineCSE::ProcessBlock(MachineDomTreeNode *Node) { unsigned NewReg = CSMI->getOperand(i).getReg(); if (OldReg == NewReg) continue; + assert(TargetRegisterInfo::isVirtualRegister(OldReg) && TargetRegisterInfo::isVirtualRegister(NewReg) && "Do not CSE physical register defs!"); + if (!isProfitableToCSE(NewReg, OldReg, CSMI, MI)) { DoCSE = false; break; } + + // Don't perform CSE if the result of the old instruction cannot exist + // within the register class of the new instruction. + const TargetRegisterClass *OldRC = MRI->getRegClass(OldReg); + if (!MRI->constrainRegClass(NewReg, OldRC)) { + DoCSE = false; + break; + } + CSEPairs.push_back(std::make_pair(OldReg, NewReg)); --NumDefs; } // Actually perform the elimination. if (DoCSE) { - for (unsigned i = 0, e = CSEPairs.size(); i != e; ++i) + for (unsigned i = 0, e = CSEPairs.size(); i != e; ++i) { MRI->replaceRegWith(CSEPairs[i].first, CSEPairs[i].second); + MRI->clearKillFlags(CSEPairs[i].second); + } MI->eraseFromParent(); ++NumCSEs; + if (!PhysRefs.empty()) + ++NumPhysCSEs; + if (Commuted) + ++NumCommutes; + Changed = true; } else { DEBUG(dbgs() << "*** Not profitable, avoid CSE!\n"); VNT.insert(MI, CurrVN++); @@ -356,10 +473,65 @@ bool MachineCSE::ProcessBlock(MachineDomTreeNode *Node) { CSEPairs.clear(); } - // Recursively call ProcessBlock with childred. - const std::vector &Children = Node->getChildren(); - for (unsigned i = 0, e = Children.size(); i != e; ++i) - Changed |= ProcessBlock(Children[i]); + return Changed; +} + +/// ExitScopeIfDone - Destroy scope for the MBB that corresponds to the given +/// dominator tree node if its a leaf or all of its children are done. Walk +/// up the dominator tree to destroy ancestors which are now done. +void +MachineCSE::ExitScopeIfDone(MachineDomTreeNode *Node, + DenseMap &OpenChildren, + DenseMap &ParentMap) { + if (OpenChildren[Node]) + return; + + // Pop scope. + ExitScope(Node->getBlock()); + + // Now traverse upwards to pop ancestors whose offsprings are all done. + while (MachineDomTreeNode *Parent = ParentMap[Node]) { + unsigned Left = --OpenChildren[Parent]; + if (Left != 0) + break; + ExitScope(Parent->getBlock()); + Node = Parent; + } +} + +bool MachineCSE::PerformCSE(MachineDomTreeNode *Node) { + SmallVector Scopes; + SmallVector WorkList; + DenseMap ParentMap; + DenseMap OpenChildren; + + CurrVN = 0; + + // Perform a DFS walk to determine the order of visit. + WorkList.push_back(Node); + do { + Node = WorkList.pop_back_val(); + Scopes.push_back(Node); + const std::vector &Children = Node->getChildren(); + unsigned NumChildren = Children.size(); + OpenChildren[Node] = NumChildren; + for (unsigned i = 0; i != NumChildren; ++i) { + MachineDomTreeNode *Child = Children[i]; + ParentMap[Child] = Node; + WorkList.push_back(Child); + } + } while (!WorkList.empty()); + + // Now perform CSE. + bool Changed = false; + for (unsigned i = 0, e = Scopes.size(); i != e; ++i) { + MachineDomTreeNode *Node = Scopes[i]; + MachineBasicBlock *MBB = Node->getBlock(); + EnterScope(MBB); + Changed |= ProcessBlock(MBB); + // If it's a leaf node, it's done. Traverse upwards to pop ancestors. + ExitScopeIfDone(Node, OpenChildren, ParentMap); + } return Changed; } @@ -370,5 +542,5 @@ bool MachineCSE::runOnMachineFunction(MachineFunction &MF) { MRI = &MF.getRegInfo(); AA = &getAnalysis(); DT = &getAnalysis(); - return ProcessBlock(DT->getRootNode()); + return PerformCSE(DT->getRootNode()); }