X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FPreAllocSplitting.cpp;h=580a6d1d9d8e6605f3cf04f9cfe90503d27015fd;hb=5657c01949dca6c012ac60d242d1a8d2ffdf5603;hp=f4540889f6aa04758a84676911940d8b85449164;hpb=45e6855ea2d1df527e0468469d693fa0c276f235;p=oota-llvm.git diff --git a/lib/CodeGen/PreAllocSplitting.cpp b/lib/CodeGen/PreAllocSplitting.cpp index f4540889f6a..580a6d1d9d8 100644 --- a/lib/CodeGen/PreAllocSplitting.cpp +++ b/lib/CodeGen/PreAllocSplitting.cpp @@ -15,6 +15,7 @@ //===----------------------------------------------------------------------===// #define DEBUG_TYPE "pre-alloc-split" +#include "VirtRegMap.h" #include "llvm/CodeGen/LiveIntervalAnalysis.h" #include "llvm/CodeGen/LiveStackAnalysis.h" #include "llvm/CodeGen/MachineDominators.h" @@ -30,6 +31,7 @@ #include "llvm/Target/TargetRegisterInfo.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" +#include "llvm/Support/ErrorHandling.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/DepthFirstIterator.h" #include "llvm/ADT/SmallPtrSet.h" @@ -38,10 +40,12 @@ using namespace llvm; static cl::opt PreSplitLimit("pre-split-limit", cl::init(-1), cl::Hidden); static cl::opt DeadSplitLimit("dead-split-limit", cl::init(-1), cl::Hidden); +static cl::opt RestoreFoldLimit("restore-fold-limit", cl::init(-1), cl::Hidden); STATISTIC(NumSplits, "Number of intervals split"); STATISTIC(NumRemats, "Number of intervals split by rematerialization"); STATISTIC(NumFolds, "Number of intervals split with spill folding"); +STATISTIC(NumRestoreFolds, "Number of intervals split with restore folding"); STATISTIC(NumRenumbers, "Number of intervals renumbered into new registers"); STATISTIC(NumDeadSpills, "Number of dead spills removed"); @@ -50,10 +54,12 @@ namespace { MachineFunction *CurrMF; const TargetMachine *TM; const TargetInstrInfo *TII; + const TargetRegisterInfo* TRI; MachineFrameInfo *MFI; MachineRegisterInfo *MRI; LiveIntervals *LIs; LiveStacks *LSs; + VirtRegMap *VRM; // Barrier - Current barrier being processed. MachineInstr *Barrier; @@ -97,8 +103,10 @@ namespace { AU.addPreservedID(PHIEliminationID); AU.addRequired(); AU.addRequired(); + AU.addRequired(); AU.addPreserved(); AU.addPreserved(); + AU.addPreserved(); MachineFunctionPass::getAnalysisUsage(AU); } @@ -158,21 +166,35 @@ namespace { MachineBasicBlock* MBB, int& SS, SmallPtrSet& RefsInMBB); + MachineInstr* FoldRestore(unsigned vreg, + const TargetRegisterClass* RC, + MachineInstr* Barrier, + MachineBasicBlock* MBB, + int SS, + SmallPtrSet& RefsInMBB); void RenumberValno(VNInfo* VN); void ReconstructLiveInterval(LiveInterval* LI); bool removeDeadSpills(SmallPtrSet& split); unsigned getNumberOfNonSpills(SmallPtrSet& MIs, unsigned Reg, int FrameIndex, bool& TwoAddr); - VNInfo* PerformPHIConstruction(MachineBasicBlock::iterator use, - MachineBasicBlock* MBB, - LiveInterval* LI, + VNInfo* PerformPHIConstruction(MachineBasicBlock::iterator Use, + MachineBasicBlock* MBB, LiveInterval* LI, SmallPtrSet& Visited, DenseMap >& Defs, DenseMap >& Uses, DenseMap& NewVNs, DenseMap& LiveOut, DenseMap& Phis, - bool toplevel, bool intrablock); + bool IsTopLevel, bool IsIntraBlock); + VNInfo* PerformPHIConstructionFallBack(MachineBasicBlock::iterator Use, + MachineBasicBlock* MBB, LiveInterval* LI, + SmallPtrSet& Visited, + DenseMap >& Defs, + DenseMap >& Uses, + DenseMap& NewVNs, + DenseMap& LiveOut, + DenseMap& Phis, + bool IsTopLevel, bool IsIntraBlock); }; } // end anonymous namespace @@ -211,32 +233,38 @@ PreAllocSplitting::findSpillPoint(MachineBasicBlock *MBB, MachineInstr *MI, unsigned &SpillIndex) { MachineBasicBlock::iterator Pt = MBB->begin(); - // Go top down if RefsInMBB is empty. - if (RefsInMBB.empty() && !DefMI) { - MachineBasicBlock::iterator MII = MBB->begin(); - MachineBasicBlock::iterator EndPt = MI; - do { - ++MII; - unsigned Index = LIs->getInstructionIndex(MII); - unsigned Gap = LIs->findGapBeforeInstr(Index); - if (Gap) { - Pt = MII; - SpillIndex = Gap; - break; - } - } while (MII != EndPt); - } else { - MachineBasicBlock::iterator MII = MI; - MachineBasicBlock::iterator EndPt = DefMI - ? MachineBasicBlock::iterator(DefMI) : MBB->begin(); - while (MII != EndPt && !RefsInMBB.count(MII)) { - unsigned Index = LIs->getInstructionIndex(MII); - if (LIs->hasGapBeforeInstr(Index)) { - Pt = MII; - SpillIndex = LIs->findGapBeforeInstr(Index, true); + MachineBasicBlock::iterator MII = MI; + MachineBasicBlock::iterator EndPt = DefMI + ? MachineBasicBlock::iterator(DefMI) : MBB->begin(); + + while (MII != EndPt && !RefsInMBB.count(MII) && + MII->getOpcode() != TRI->getCallFrameSetupOpcode()) + --MII; + if (MII == EndPt || RefsInMBB.count(MII)) return Pt; + + while (MII != EndPt && !RefsInMBB.count(MII)) { + unsigned Index = LIs->getInstructionIndex(MII); + + // We can't insert the spill between the barrier (a call), and its + // corresponding call frame setup. + if (MII->getOpcode() == TRI->getCallFrameDestroyOpcode()) { + while (MII->getOpcode() != TRI->getCallFrameSetupOpcode()) { + --MII; + if (MII == EndPt) { + return Pt; + } } - --MII; + continue; + } else if (LIs->hasGapBeforeInstr(Index)) { + Pt = MII; + SpillIndex = LIs->findGapBeforeInstr(Index, true); } + + if (RefsInMBB.count(MII)) + return Pt; + + + --MII; } return Pt; @@ -254,42 +282,44 @@ PreAllocSplitting::findRestorePoint(MachineBasicBlock *MBB, MachineInstr *MI, // FIXME: Allow spill to be inserted to the beginning of the mbb. Update mbb // begin index accordingly. MachineBasicBlock::iterator Pt = MBB->end(); - unsigned EndIdx = LIs->getMBBEndIdx(MBB); + MachineBasicBlock::iterator EndPt = MBB->getFirstTerminator(); - // Go bottom up if RefsInMBB is empty and the end of the mbb isn't beyond - // the last index in the live range. - if (RefsInMBB.empty() && LastIdx >= EndIdx) { - MachineBasicBlock::iterator MII = MBB->getFirstTerminator(); - MachineBasicBlock::iterator EndPt = MI; - --MII; - do { - unsigned Index = LIs->getInstructionIndex(MII); - unsigned Gap = LIs->findGapBeforeInstr(Index); - if (Gap) { - Pt = MII; - RestoreIndex = Gap; - break; - } - --MII; - } while (MII != EndPt); - } else { - MachineBasicBlock::iterator MII = MI; - MII = ++MII; - // FIXME: Limit the number of instructions to examine to reduce - // compile time? - while (MII != MBB->end()) { - unsigned Index = LIs->getInstructionIndex(MII); - if (Index > LastIdx) - break; - unsigned Gap = LIs->findGapBeforeInstr(Index); - if (Gap) { - Pt = MII; - RestoreIndex = Gap; - } - if (RefsInMBB.count(MII)) - break; - ++MII; + // We start at the call, so walk forward until we find the call frame teardown + // since we can't insert restores before that. Bail if we encounter a use + // during this time. + MachineBasicBlock::iterator MII = MI; + if (MII == EndPt) return Pt; + + while (MII != EndPt && !RefsInMBB.count(MII) && + MII->getOpcode() != TRI->getCallFrameDestroyOpcode()) + ++MII; + if (MII == EndPt || RefsInMBB.count(MII)) return Pt; + ++MII; + + // FIXME: Limit the number of instructions to examine to reduce + // compile time? + while (MII != EndPt) { + unsigned Index = LIs->getInstructionIndex(MII); + if (Index > LastIdx) + break; + unsigned Gap = LIs->findGapBeforeInstr(Index); + + // We can't insert a restore between the barrier (a call) and its + // corresponding call frame teardown. + if (MII->getOpcode() == TRI->getCallFrameSetupOpcode()) { + do { + if (MII == EndPt || RefsInMBB.count(MII)) return Pt; + ++MII; + } while (MII->getOpcode() != TRI->getCallFrameDestroyOpcode()); + } else if (Gap) { + Pt = MII; + RestoreIndex = Gap; } + + if (RefsInMBB.count(MII)) + return Pt; + + ++MII; } return Pt; @@ -310,11 +340,11 @@ int PreAllocSplitting::CreateSpillStackSlot(unsigned Reg, } // Create live interval for stack slot. - CurrSLI = &LSs->getOrCreateInterval(SS); + CurrSLI = &LSs->getOrCreateInterval(SS, RC); if (CurrSLI->hasAtLeastOneValue()) CurrSValNo = CurrSLI->getValNumInfo(0); else - CurrSValNo = CurrSLI->getNextValue(~0U, 0, LSs->getVNInfoAllocator()); + CurrSValNo = CurrSLI->getNextValue(0, 0, false, LSs->getVNInfoAllocator()); return SS; } @@ -411,94 +441,35 @@ PreAllocSplitting::UpdateSpillSlotInterval(VNInfo *ValNo, unsigned SpillIndex, /// PerformPHIConstruction - From properly set up use and def lists, use a PHI /// construction algorithm to compute the ranges and valnos for an interval. -VNInfo* PreAllocSplitting::PerformPHIConstruction( - MachineBasicBlock::iterator use, - MachineBasicBlock* MBB, - LiveInterval* LI, +VNInfo* +PreAllocSplitting::PerformPHIConstruction(MachineBasicBlock::iterator UseI, + MachineBasicBlock* MBB, LiveInterval* LI, SmallPtrSet& Visited, DenseMap >& Defs, DenseMap >& Uses, DenseMap& NewVNs, DenseMap& LiveOut, DenseMap& Phis, - bool toplevel, bool intrablock) { + bool IsTopLevel, bool IsIntraBlock) { // Return memoized result if it's available. - if (toplevel && Visited.count(use) && NewVNs.count(use)) - return NewVNs[use]; - else if (!toplevel && intrablock && NewVNs.count(use)) - return NewVNs[use]; - else if (!intrablock && LiveOut.count(MBB)) + if (IsTopLevel && Visited.count(UseI) && NewVNs.count(UseI)) + return NewVNs[UseI]; + else if (!IsTopLevel && IsIntraBlock && NewVNs.count(UseI)) + return NewVNs[UseI]; + else if (!IsIntraBlock && LiveOut.count(MBB)) return LiveOut[MBB]; - typedef DenseMap > RegMap; - // Check if our block contains any uses or defs. bool ContainsDefs = Defs.count(MBB); bool ContainsUses = Uses.count(MBB); - VNInfo* ret = 0; + VNInfo* RetVNI = 0; // Enumerate the cases of use/def contaning blocks. if (!ContainsDefs && !ContainsUses) { - Fallback: - // NOTE: Because this is the fallback case from other cases, we do NOT - // assume that we are not intrablock here. - if (Phis.count(MBB)) return Phis[MBB]; - - unsigned StartIndex = LIs->getMBBStartIdx(MBB); - - if (MBB->pred_size() == 1) { - Phis[MBB] = ret = PerformPHIConstruction((*MBB->pred_begin())->end(), - *(MBB->pred_begin()), LI, Visited, - Defs, Uses, NewVNs, LiveOut, Phis, - false, false); - unsigned EndIndex = 0; - if (intrablock) { - EndIndex = LIs->getInstructionIndex(use); - EndIndex = LiveIntervals::getUseIndex(EndIndex); - } else - EndIndex = LIs->getMBBEndIdx(MBB); - - LI->addRange(LiveRange(StartIndex, EndIndex+1, ret)); - if (intrablock) - LI->addKill(ret, EndIndex); - } else { - Phis[MBB] = ret = LI->getNextValue(~0U, /*FIXME*/ 0, - LIs->getVNInfoAllocator()); - if (!intrablock) LiveOut[MBB] = ret; - - // If there are no uses or defs between our starting point and the - // beginning of the block, then recursive perform phi construction - // on our predecessors. - DenseMap IncomingVNs; - for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), - PE = MBB->pred_end(); PI != PE; ++PI) { - VNInfo* Incoming = PerformPHIConstruction((*PI)->end(), *PI, LI, - Visited, Defs, Uses, NewVNs, - LiveOut, Phis, false, false); - if (Incoming != 0) - IncomingVNs[*PI] = Incoming; - } - - // Otherwise, merge the incoming VNInfos with a phi join. Create a new - // VNInfo to represent the joined value. - for (DenseMap::iterator I = - IncomingVNs.begin(), E = IncomingVNs.end(); I != E; ++I) { - I->second->hasPHIKill = true; - unsigned KillIndex = LIs->getMBBEndIdx(I->first); - LI->addKill(I->second, KillIndex); - } - - unsigned EndIndex = 0; - if (intrablock) { - EndIndex = LIs->getInstructionIndex(use); - EndIndex = LiveIntervals::getUseIndex(EndIndex); - } else - EndIndex = LIs->getMBBEndIdx(MBB); - LI->addRange(LiveRange(StartIndex, EndIndex+1, ret)); - if (intrablock) - LI->addKill(ret, EndIndex); - } + return PerformPHIConstructionFallBack(UseI, MBB, LI, Visited, Defs, Uses, + NewVNs, LiveOut, Phis, + IsTopLevel, IsIntraBlock); } else if (ContainsDefs && !ContainsUses) { SmallPtrSet& BlockDefs = Defs[MBB]; @@ -506,71 +477,75 @@ VNInfo* PreAllocSplitting::PerformPHIConstruction( // instruction we care about, go to the fallback case. Note that that // should never happen: this cannot be intrablock, so use should // always be an end() iterator. - assert(use == MBB->end() && "No use marked in intrablock"); + assert(UseI == MBB->end() && "No use marked in intrablock"); - MachineBasicBlock::iterator walker = use; - --walker; - while (walker != MBB->begin()) - if (BlockDefs.count(walker)) { + MachineBasicBlock::iterator Walker = UseI; + --Walker; + while (Walker != MBB->begin()) { + if (BlockDefs.count(Walker)) break; - } else - --walker; + --Walker; + } // Once we've found it, extend its VNInfo to our instruction. - unsigned DefIndex = LIs->getInstructionIndex(walker); + unsigned DefIndex = LIs->getInstructionIndex(Walker); DefIndex = LiveIntervals::getDefIndex(DefIndex); unsigned EndIndex = LIs->getMBBEndIdx(MBB); - ret = NewVNs[walker]; - LI->addRange(LiveRange(DefIndex, EndIndex+1, ret)); + RetVNI = NewVNs[Walker]; + LI->addRange(LiveRange(DefIndex, EndIndex+1, RetVNI)); } else if (!ContainsDefs && ContainsUses) { SmallPtrSet& BlockUses = Uses[MBB]; // Search for the use in this block that precedes the instruction we care - // about, going to the fallback case if we don't find it. - - if (use == MBB->begin()) - goto Fallback; + // about, going to the fallback case if we don't find it. + if (UseI == MBB->begin()) + return PerformPHIConstructionFallBack(UseI, MBB, LI, Visited, Defs, + Uses, NewVNs, LiveOut, Phis, + IsTopLevel, IsIntraBlock); - MachineBasicBlock::iterator walker = use; - --walker; + MachineBasicBlock::iterator Walker = UseI; + --Walker; bool found = false; - while (walker != MBB->begin()) - if (BlockUses.count(walker)) { + while (Walker != MBB->begin()) { + if (BlockUses.count(Walker)) { found = true; break; - } else - --walker; + } + --Walker; + } // Must check begin() too. if (!found) { - if (BlockUses.count(walker)) + if (BlockUses.count(Walker)) found = true; else - goto Fallback; + return PerformPHIConstructionFallBack(UseI, MBB, LI, Visited, Defs, + Uses, NewVNs, LiveOut, Phis, + IsTopLevel, IsIntraBlock); } - unsigned UseIndex = LIs->getInstructionIndex(walker); + unsigned UseIndex = LIs->getInstructionIndex(Walker); UseIndex = LiveIntervals::getUseIndex(UseIndex); unsigned EndIndex = 0; - if (intrablock) { - EndIndex = LIs->getInstructionIndex(use); + if (IsIntraBlock) { + EndIndex = LIs->getInstructionIndex(UseI); EndIndex = LiveIntervals::getUseIndex(EndIndex); } else EndIndex = LIs->getMBBEndIdx(MBB); // Now, recursively phi construct the VNInfo for the use we found, // and then extend it to include the instruction we care about - ret = PerformPHIConstruction(walker, MBB, LI, Visited, Defs, Uses, - NewVNs, LiveOut, Phis, false, true); + RetVNI = PerformPHIConstruction(Walker, MBB, LI, Visited, Defs, Uses, + NewVNs, LiveOut, Phis, false, true); - LI->addRange(LiveRange(UseIndex, EndIndex+1, ret)); + LI->addRange(LiveRange(UseIndex, EndIndex+1, RetVNI)); // FIXME: Need to set kills properly for inter-block stuff. - if (LI->isKill(ret, UseIndex)) LI->removeKill(ret, UseIndex); - if (intrablock) - LI->addKill(ret, EndIndex); - } else if (ContainsDefs && ContainsUses){ + if (LI->isKill(RetVNI, UseIndex)) LI->removeKill(RetVNI, UseIndex); + if (IsIntraBlock) + LI->addKill(RetVNI, EndIndex, false); + } else if (ContainsDefs && ContainsUses) { SmallPtrSet& BlockDefs = Defs[MBB]; SmallPtrSet& BlockUses = Uses[MBB]; @@ -578,67 +553,160 @@ VNInfo* PreAllocSplitting::PerformPHIConstruction( // special note that checking for defs must take precedence over checking // for uses, because of two-address instructions. - if (use == MBB->begin()) - goto Fallback; + if (UseI == MBB->begin()) + return PerformPHIConstructionFallBack(UseI, MBB, LI, Visited, Defs, Uses, + NewVNs, LiveOut, Phis, + IsTopLevel, IsIntraBlock); - MachineBasicBlock::iterator walker = use; - --walker; + MachineBasicBlock::iterator Walker = UseI; + --Walker; bool foundDef = false; bool foundUse = false; - while (walker != MBB->begin()) - if (BlockDefs.count(walker)) { + while (Walker != MBB->begin()) { + if (BlockDefs.count(Walker)) { foundDef = true; break; - } else if (BlockUses.count(walker)) { + } else if (BlockUses.count(Walker)) { foundUse = true; break; - } else - --walker; + } + --Walker; + } // Must check begin() too. if (!foundDef && !foundUse) { - if (BlockDefs.count(walker)) + if (BlockDefs.count(Walker)) foundDef = true; - else if (BlockUses.count(walker)) + else if (BlockUses.count(Walker)) foundUse = true; else - goto Fallback; + return PerformPHIConstructionFallBack(UseI, MBB, LI, Visited, Defs, + Uses, NewVNs, LiveOut, Phis, + IsTopLevel, IsIntraBlock); } - unsigned StartIndex = LIs->getInstructionIndex(walker); + unsigned StartIndex = LIs->getInstructionIndex(Walker); StartIndex = foundDef ? LiveIntervals::getDefIndex(StartIndex) : LiveIntervals::getUseIndex(StartIndex); unsigned EndIndex = 0; - if (intrablock) { - EndIndex = LIs->getInstructionIndex(use); + if (IsIntraBlock) { + EndIndex = LIs->getInstructionIndex(UseI); EndIndex = LiveIntervals::getUseIndex(EndIndex); } else EndIndex = LIs->getMBBEndIdx(MBB); if (foundDef) - ret = NewVNs[walker]; + RetVNI = NewVNs[Walker]; else - ret = PerformPHIConstruction(walker, MBB, LI, Visited, Defs, Uses, - NewVNs, LiveOut, Phis, false, true); + RetVNI = PerformPHIConstruction(Walker, MBB, LI, Visited, Defs, Uses, + NewVNs, LiveOut, Phis, false, true); - LI->addRange(LiveRange(StartIndex, EndIndex+1, ret)); + LI->addRange(LiveRange(StartIndex, EndIndex+1, RetVNI)); - if (foundUse && LI->isKill(ret, StartIndex)) - LI->removeKill(ret, StartIndex); - if (intrablock) { - LI->addKill(ret, EndIndex); + if (foundUse && LI->isKill(RetVNI, StartIndex)) + LI->removeKill(RetVNI, StartIndex); + if (IsIntraBlock) { + LI->addKill(RetVNI, EndIndex, false); } } // Memoize results so we don't have to recompute them. - if (!intrablock) LiveOut[MBB] = ret; + if (!IsIntraBlock) LiveOut[MBB] = RetVNI; + else { + if (!NewVNs.count(UseI)) + NewVNs[UseI] = RetVNI; + Visited.insert(UseI); + } + + return RetVNI; +} + +/// PerformPHIConstructionFallBack - PerformPHIConstruction fall back path. +/// +VNInfo* +PreAllocSplitting::PerformPHIConstructionFallBack(MachineBasicBlock::iterator UseI, + MachineBasicBlock* MBB, LiveInterval* LI, + SmallPtrSet& Visited, + DenseMap >& Defs, + DenseMap >& Uses, + DenseMap& NewVNs, + DenseMap& LiveOut, + DenseMap& Phis, + bool IsTopLevel, bool IsIntraBlock) { + // NOTE: Because this is the fallback case from other cases, we do NOT + // assume that we are not intrablock here. + if (Phis.count(MBB)) return Phis[MBB]; + + unsigned StartIndex = LIs->getMBBStartIdx(MBB); + VNInfo *RetVNI = Phis[MBB] = + LI->getNextValue(0, /*FIXME*/ 0, false, LIs->getVNInfoAllocator()); + + if (!IsIntraBlock) LiveOut[MBB] = RetVNI; + + // If there are no uses or defs between our starting point and the + // beginning of the block, then recursive perform phi construction + // on our predecessors. + DenseMap IncomingVNs; + for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), + PE = MBB->pred_end(); PI != PE; ++PI) { + VNInfo* Incoming = PerformPHIConstruction((*PI)->end(), *PI, LI, + Visited, Defs, Uses, NewVNs, + LiveOut, Phis, false, false); + if (Incoming != 0) + IncomingVNs[*PI] = Incoming; + } + + if (MBB->pred_size() == 1 && !RetVNI->hasPHIKill()) { + VNInfo* OldVN = RetVNI; + VNInfo* NewVN = IncomingVNs.begin()->second; + VNInfo* MergedVN = LI->MergeValueNumberInto(OldVN, NewVN); + if (MergedVN == OldVN) std::swap(OldVN, NewVN); + + for (DenseMap::iterator LOI = LiveOut.begin(), + LOE = LiveOut.end(); LOI != LOE; ++LOI) + if (LOI->second == OldVN) + LOI->second = MergedVN; + for (DenseMap::iterator NVI = NewVNs.begin(), + NVE = NewVNs.end(); NVI != NVE; ++NVI) + if (NVI->second == OldVN) + NVI->second = MergedVN; + for (DenseMap::iterator PI = Phis.begin(), + PE = Phis.end(); PI != PE; ++PI) + if (PI->second == OldVN) + PI->second = MergedVN; + RetVNI = MergedVN; + } else { + // Otherwise, merge the incoming VNInfos with a phi join. Create a new + // VNInfo to represent the joined value. + for (DenseMap::iterator I = + IncomingVNs.begin(), E = IncomingVNs.end(); I != E; ++I) { + I->second->setHasPHIKill(true); + unsigned KillIndex = LIs->getMBBEndIdx(I->first); + if (!LiveInterval::isKill(I->second, KillIndex)) + LI->addKill(I->second, KillIndex, false); + } + } + + unsigned EndIndex = 0; + if (IsIntraBlock) { + EndIndex = LIs->getInstructionIndex(UseI); + EndIndex = LiveIntervals::getUseIndex(EndIndex); + } else + EndIndex = LIs->getMBBEndIdx(MBB); + LI->addRange(LiveRange(StartIndex, EndIndex+1, RetVNI)); + if (IsIntraBlock) + LI->addKill(RetVNI, EndIndex, false); + + // Memoize results so we don't have to recompute them. + if (!IsIntraBlock) + LiveOut[MBB] = RetVNI; else { - if (!NewVNs.count(use)) - NewVNs[use] = ret; - Visited.insert(use); + if (!NewVNs.count(UseI)) + NewVNs[UseI] = RetVNI; + Visited.insert(UseI); } - return ret; + return RetVNI; } /// ReconstructLiveInterval - Recompute a live interval from scratch. @@ -664,7 +732,9 @@ void PreAllocSplitting::ReconstructLiveInterval(LiveInterval* LI) { unsigned DefIdx = LIs->getInstructionIndex(&*DI); DefIdx = LiveIntervals::getDefIndex(DefIdx); - VNInfo* NewVN = LI->getNextValue(DefIdx, 0, Alloc); + assert(DI->getOpcode() != TargetInstrInfo::PHI && + "Following NewVN isPHIDef flag incorrect. Fix me!"); + VNInfo* NewVN = LI->getNextValue(DefIdx, 0, true, Alloc); // If the def is a move, set the copy field. unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx; @@ -702,7 +772,7 @@ void PreAllocSplitting::ReconstructLiveInterval(LiveInterval* LI) { VNInfo* DeadVN = NewVNs[&*DI]; LI->addRange(LiveRange(DefIdx, DefIdx+1, DeadVN)); - LI->addKill(DeadVN, DefIdx); + LI->addKill(DeadVN, DefIdx, false); } } @@ -727,19 +797,20 @@ void PreAllocSplitting::RenumberValno(VNInfo* VN) { // Bail out if we ever encounter a valno that has a PHI kill. We can't // renumber these. - if (OldVN->hasPHIKill) return; + if (OldVN->hasPHIKill()) return; VNsToCopy.push_back(OldVN); // Locate two-address redefinitions - for (SmallVector::iterator KI = OldVN->kills.begin(), + for (VNInfo::KillSet::iterator KI = OldVN->kills.begin(), KE = OldVN->kills.end(); KI != KE; ++KI) { - MachineInstr* MI = LIs->getInstructionFromIndex(*KI); + assert(!KI->isPHIKill && "VN previously reported having no PHI kills."); + MachineInstr* MI = LIs->getInstructionFromIndex(KI->killIdx); unsigned DefIdx = MI->findRegisterDefOperandIdx(CurrLI->reg); if (DefIdx == ~0U) continue; - if (MI->isRegReDefinedByTwoAddr(DefIdx)) { + if (MI->isRegTiedToUseOperand(DefIdx)) { VNInfo* NextVN = - CurrLI->findDefinedVNInfo(LiveIntervals::getDefIndex(*KI)); + CurrLI->findDefinedVNInfo(LiveIntervals::getDefIndex(KI->killIdx)); if (NextVN == OldVN) continue; Stack.push_back(NextVN); } @@ -757,9 +828,7 @@ void PreAllocSplitting::RenumberValno(VNInfo* VN) { VNInfo* OldVN = *OI; // Copy the valno over - VNInfo* NewVN = NewLI.getNextValue(OldVN->def, OldVN->copy, - LIs->getVNInfoAllocator()); - NewLI.copyValNumInfo(NewVN, OldVN); + VNInfo* NewVN = NewLI.createValueCopy(OldVN, LIs->getVNInfoAllocator()); NewLI.MergeValueInAsValue(*CurrLI, OldVN, NewVN); // Remove the valno from the old interval @@ -788,6 +857,9 @@ void PreAllocSplitting::RenumberValno(VNInfo* VN) { MO.setReg(NewVReg); } + // Grow the VirtRegMap, since we've created a new vreg. + VRM->grow(); + // The renumbered vreg shares a stack slot with the old register. if (IntervalSSMap.count(CurrLI->reg)) IntervalSSMap[NewVReg] = IntervalSSMap[CurrLI->reg]; @@ -795,7 +867,7 @@ void PreAllocSplitting::RenumberValno(VNInfo* VN) { NumRenumbers++; } -bool PreAllocSplitting::Rematerialize(unsigned vreg, VNInfo* ValNo, +bool PreAllocSplitting::Rematerialize(unsigned VReg, VNInfo* ValNo, MachineInstr* DefMI, MachineBasicBlock::iterator RestorePt, unsigned RestoreIdx, @@ -804,7 +876,7 @@ bool PreAllocSplitting::Rematerialize(unsigned vreg, VNInfo* ValNo, MachineBasicBlock::iterator KillPt = BarrierMBB->end(); unsigned KillIdx = 0; - if (ValNo->def == ~0U || DefMI->getParent() == BarrierMBB) + if (!ValNo->isDefAccurate() || DefMI->getParent() == BarrierMBB) KillPt = findSpillPoint(BarrierMBB, Barrier, NULL, RefsInMBB, KillIdx); else KillPt = findNextEmptySlot(DefMI->getParent(), DefMI, KillIdx); @@ -812,7 +884,7 @@ bool PreAllocSplitting::Rematerialize(unsigned vreg, VNInfo* ValNo, if (KillPt == DefMI->getParent()->end()) return false; - TII->reMaterialize(MBB, RestorePt, vreg, DefMI); + TII->reMaterialize(MBB, RestorePt, VReg, 0, DefMI); LIs->InsertMachineInstrInMaps(prior(RestorePt), RestoreIdx); ReconstructLiveInterval(CurrLI); @@ -857,8 +929,7 @@ MachineInstr* PreAllocSplitting::FoldSpill(unsigned vreg, if (I != IntervalSSMap.end()) { SS = I->second; } else { - SS = MFI->CreateStackObject(RC->getSize(), RC->getAlignment()); - + SS = MFI->CreateStackObject(RC->getSize(), RC->getAlignment()); } MachineInstr* FMI = TII->foldMemoryOperand(*MBB->getParent(), @@ -870,11 +941,84 @@ MachineInstr* PreAllocSplitting::FoldSpill(unsigned vreg, ++NumFolds; IntervalSSMap[vreg] = SS; - CurrSLI = &LSs->getOrCreateInterval(SS); + CurrSLI = &LSs->getOrCreateInterval(SS, RC); if (CurrSLI->hasAtLeastOneValue()) CurrSValNo = CurrSLI->getValNumInfo(0); else - CurrSValNo = CurrSLI->getNextValue(~0U, 0, LSs->getVNInfoAllocator()); + CurrSValNo = CurrSLI->getNextValue(0, 0, false, LSs->getVNInfoAllocator()); + } + + return FMI; +} + +MachineInstr* PreAllocSplitting::FoldRestore(unsigned vreg, + const TargetRegisterClass* RC, + MachineInstr* Barrier, + MachineBasicBlock* MBB, + int SS, + SmallPtrSet& RefsInMBB) { + if ((int)RestoreFoldLimit != -1 && RestoreFoldLimit == (int)NumRestoreFolds) + return 0; + + // Go top down if RefsInMBB is empty. + if (RefsInMBB.empty()) + return 0; + + // Can't fold a restore between a call stack setup and teardown. + MachineBasicBlock::iterator FoldPt = Barrier; + + // Advance from barrier to call frame teardown. + while (FoldPt != MBB->getFirstTerminator() && + FoldPt->getOpcode() != TRI->getCallFrameDestroyOpcode()) { + if (RefsInMBB.count(FoldPt)) + return 0; + + ++FoldPt; + } + + if (FoldPt == MBB->getFirstTerminator()) + return 0; + else + ++FoldPt; + + // Now find the restore point. + while (FoldPt != MBB->getFirstTerminator() && !RefsInMBB.count(FoldPt)) { + if (FoldPt->getOpcode() == TRI->getCallFrameSetupOpcode()) { + while (FoldPt != MBB->getFirstTerminator() && + FoldPt->getOpcode() != TRI->getCallFrameDestroyOpcode()) { + if (RefsInMBB.count(FoldPt)) + return 0; + + ++FoldPt; + } + + if (FoldPt == MBB->getFirstTerminator()) + return 0; + } + + ++FoldPt; + } + + if (FoldPt == MBB->getFirstTerminator()) + return 0; + + int OpIdx = FoldPt->findRegisterUseOperandIdx(vreg, true); + if (OpIdx == -1) + return 0; + + SmallVector Ops; + Ops.push_back(OpIdx); + + if (!TII->canFoldMemoryOperand(FoldPt, Ops)) + return 0; + + MachineInstr* FMI = TII->foldMemoryOperand(*MBB->getParent(), + FoldPt, Ops, SS); + + if (FMI) { + LIs->ReplaceMachineInstrInMaps(FoldPt, FMI); + FMI = MBB->insert(MBB->erase(FoldPt), FMI); + ++NumRestoreFolds; } return FMI; @@ -891,13 +1035,9 @@ bool PreAllocSplitting::SplitRegLiveInterval(LiveInterval *LI) { CurrLI->FindLiveRangeContaining(LIs->getUseIndex(BarrierIdx)); VNInfo *ValNo = LR->valno; - if (ValNo->def == ~1U) { - // Defined by a dead def? How can this be? - assert(0 && "Val# is defined by a dead def?"); - abort(); - } + assert(!ValNo->isUnused() && "Val# is defined by a dead def?"); - MachineInstr *DefMI = (ValNo->def != ~0U) + MachineInstr *DefMI = ValNo->isDefAccurate() ? LIs->getInstructionFromIndex(ValNo->def) : NULL; // If this would create a new join point, do not split. @@ -931,8 +1071,8 @@ bool PreAllocSplitting::SplitRegLiveInterval(LiveInterval *LI) { unsigned SpillIndex = 0; MachineInstr *SpillMI = NULL; int SS = -1; - if (ValNo->def == ~0U) { - // If it's defined by a phi, we must split just before the barrier. + if (!ValNo->isDefAccurate()) { + // If we don't know where the def is we must split just before the barrier. if ((SpillMI = FoldSpill(LI->reg, RC, 0, Barrier, BarrierMBB, SS, RefsInMBB))) { SpillIndex = LIs->getInstructionIndex(SpillMI); @@ -987,18 +1127,29 @@ bool PreAllocSplitting::SplitRegLiveInterval(LiveInterval *LI) { Def2SpillMap[ValNo->def] = SpillIndex; // Add restore. - TII->loadRegFromStackSlot(*BarrierMBB, RestorePt, CurrLI->reg, SS, RC); - MachineInstr *LoadMI = prior(RestorePt); - LIs->InsertMachineInstrInMaps(LoadMI, RestoreIndex); + bool FoldedRestore = false; + if (MachineInstr* LMI = FoldRestore(CurrLI->reg, RC, Barrier, + BarrierMBB, SS, RefsInMBB)) { + RestorePt = LMI; + RestoreIndex = LIs->getInstructionIndex(RestorePt); + FoldedRestore = true; + } else { + TII->loadRegFromStackSlot(*BarrierMBB, RestorePt, CurrLI->reg, SS, RC); + MachineInstr *LoadMI = prior(RestorePt); + LIs->InsertMachineInstrInMaps(LoadMI, RestoreIndex); + } // Update spill stack slot live interval. UpdateSpillSlotInterval(ValNo, LIs->getUseIndex(SpillIndex)+1, LIs->getDefIndex(RestoreIndex)); ReconstructLiveInterval(CurrLI); - unsigned RestoreIdx = LIs->getInstructionIndex(prior(RestorePt)); - RestoreIdx = LiveIntervals::getDefIndex(RestoreIdx); - RenumberValno(CurrLI->findDefinedVNInfo(RestoreIdx)); + + if (!FoldedRestore) { + unsigned RestoreIdx = LIs->getInstructionIndex(prior(RestorePt)); + RestoreIdx = LiveIntervals::getDefIndex(RestoreIdx); + RenumberValno(CurrLI->findDefinedVNInfo(RestoreIdx)); + } ++NumSplits; return true; @@ -1013,7 +1164,10 @@ PreAllocSplitting::SplitRegLiveIntervals(const TargetRegisterClass **RCs, // by the current barrier. SmallVector Intervals; for (const TargetRegisterClass **RC = RCs; *RC; ++RC) { - if (TII->IgnoreRegisterClassBarriers(*RC)) + // FIXME: If it's not safe to move any instruction that defines the barrier + // register class, then it means there are some special dependencies which + // codegen is not modelling. Ignore these barriers for now. + if (!TII->isSafeToMoveRegClassDefs(*RC)) continue; std::vector &VRs = MRI->getRegClassVirtRegs(*RC); for (unsigned i = 0, e = VRs.size(); i != e; ++i) { @@ -1058,7 +1212,7 @@ unsigned PreAllocSplitting::getNumberOfNonSpills( NonSpills++; int DefIdx = (*UI)->findRegisterDefOperandIdx(Reg); - if (DefIdx != -1 && (*UI)->isRegReDefinedByTwoAddr(DefIdx)) + if (DefIdx != -1 && (*UI)->isRegTiedToUseOperand(DefIdx)) FeedsTwoAddr = true; } @@ -1070,10 +1224,14 @@ unsigned PreAllocSplitting::getNumberOfNonSpills( bool PreAllocSplitting::removeDeadSpills(SmallPtrSet& split) { bool changed = false; + // Walk over all of the live intervals that were touched by the splitter, + // and see if we can do any DCE and/or folding. for (SmallPtrSet::iterator LI = split.begin(), LE = split.end(); LI != LE; ++LI) { DenseMap > VNUseCount; + // First, collect all the uses of the vreg, and sort them by their + // reaching definition (VNInfo). for (MachineRegisterInfo::use_iterator UI = MRI->use_begin((*LI)->reg), UE = MRI->use_end(); UI != UE; ++UI) { unsigned index = LIs->getInstructionIndex(&*UI); @@ -1083,6 +1241,8 @@ bool PreAllocSplitting::removeDeadSpills(SmallPtrSet& split) { VNUseCount[LR->valno].insert(&*UI); } + // Now, take the definitions (VNInfo's) one at a time and try to DCE + // and/or fold them away. for (LiveInterval::vni_iterator VI = (*LI)->vni_begin(), VE = (*LI)->vni_end(); VI != VE; ++VI) { @@ -1090,30 +1250,44 @@ bool PreAllocSplitting::removeDeadSpills(SmallPtrSet& split) { return changed; VNInfo* CurrVN = *VI; - if (CurrVN->hasPHIKill) continue; - unsigned DefIdx = CurrVN->def; - if (DefIdx == ~0U || DefIdx == ~1U) continue; + // We don't currently try to handle definitions with PHI kills, because + // it would involve processing more than one VNInfo at once. + if (CurrVN->hasPHIKill()) continue; + + // We also don't try to handle the results of PHI joins, since there's + // no defining instruction to analyze. + if (!CurrVN->isDefAccurate() || CurrVN->isUnused()) continue; - MachineInstr* DefMI = LIs->getInstructionFromIndex(DefIdx); + // We're only interested in eliminating cruft introduced by the splitter, + // is of the form load-use or load-use-store. First, check that the + // definition is a load, and remember what stack slot we loaded it from. + MachineInstr* DefMI = LIs->getInstructionFromIndex(CurrVN->def); int FrameIndex; if (!TII->isLoadFromStackSlot(DefMI, FrameIndex)) continue; + // If the definition has no uses at all, just DCE it. if (VNUseCount[CurrVN].size() == 0) { LIs->RemoveMachineInstrFromMaps(DefMI); (*LI)->removeValNo(CurrVN); DefMI->eraseFromParent(); + VNUseCount.erase(CurrVN); NumDeadSpills++; changed = true; continue; } + // Second, get the number of non-store uses of the definition, as well as + // a flag indicating whether it feeds into a later two-address definition. bool FeedsTwoAddr = false; unsigned NonSpillCount = getNumberOfNonSpills(VNUseCount[CurrVN], (*LI)->reg, FrameIndex, FeedsTwoAddr); + // If there's one non-store use and it doesn't feed a two-addr, then + // this is a load-use-store case that we can try to fold. if (NonSpillCount == 1 && !FeedsTwoAddr) { + // Start by finding the non-store use MachineInstr. SmallPtrSet::iterator UI = VNUseCount[CurrVN].begin(); int StoreFrameIndex; unsigned StoreVReg = TII->isStoreToStackSlot(*UI, StoreFrameIndex); @@ -1123,17 +1297,15 @@ bool PreAllocSplitting::removeDeadSpills(SmallPtrSet& split) { if (UI != VNUseCount[CurrVN].end()) StoreVReg = TII->isStoreToStackSlot(*UI, StoreFrameIndex); } - if (UI == VNUseCount[CurrVN].end()) continue; MachineInstr* use = *UI; + // Attempt to fold it away! int OpIdx = use->findRegisterUseOperandIdx((*LI)->reg, false); if (OpIdx == -1) continue; - SmallVector Ops; Ops.push_back(OpIdx); - if (!TII->canFoldMemoryOperand(use, Ops)) continue; MachineInstr* NewMI = @@ -1142,6 +1314,7 @@ bool PreAllocSplitting::removeDeadSpills(SmallPtrSet& split) { if (!NewMI) continue; + // Update relevant analyses. LIs->RemoveMachineInstrFromMaps(DefMI); LIs->ReplaceMachineInstrInMaps(use, NewMI); (*LI)->removeValNo(CurrVN); @@ -1151,12 +1324,21 @@ bool PreAllocSplitting::removeDeadSpills(SmallPtrSet& split) { NewMI = MBB->insert(MBB->erase(use), NewMI); VNUseCount[CurrVN].erase(use); + // Remove deleted instructions. Note that we need to remove them from + // the VNInfo->use map as well, just to be safe. for (SmallPtrSet::iterator II = VNUseCount[CurrVN].begin(), IE = VNUseCount[CurrVN].end(); II != IE; ++II) { + for (DenseMap >::iterator + VNI = VNUseCount.begin(), VNE = VNUseCount.end(); VNI != VNE; + ++VNI) + if (VNI->first != CurrVN) + VNI->second.erase(*II); LIs->RemoveMachineInstrFromMaps(*II); (*II)->eraseFromParent(); } + + VNUseCount.erase(CurrVN); for (DenseMap >::iterator VI = VNUseCount.begin(), VE = VNUseCount.end(); VI != VE; ++VI) @@ -1168,8 +1350,11 @@ bool PreAllocSplitting::removeDeadSpills(SmallPtrSet& split) { continue; } + // If there's more than one non-store instruction, we can't profitably + // fold it, so bail. if (NonSpillCount) continue; + // Otherwise, this is a load-store case, so DCE them. for (SmallPtrSet::iterator UI = VNUseCount[CurrVN].begin(), UE = VNUseCount[CurrVN].end(); UI != UI; ++UI) { @@ -1177,6 +1362,8 @@ bool PreAllocSplitting::removeDeadSpills(SmallPtrSet& split) { (*UI)->eraseFromParent(); } + VNUseCount.erase(CurrVN); + LIs->RemoveMachineInstrFromMaps(DefMI); (*LI)->removeValNo(CurrVN); DefMI->eraseFromParent(); @@ -1194,7 +1381,7 @@ bool PreAllocSplitting::createsNewJoin(LiveRange* LR, if (DefMBB == BarrierMBB) return false; - if (LR->valno->hasPHIKill) + if (LR->valno->hasPHIKill()) return false; unsigned MBBEnd = LIs->getMBBEndIdx(BarrierMBB); @@ -1256,11 +1443,13 @@ bool PreAllocSplitting::createsNewJoin(LiveRange* LR, bool PreAllocSplitting::runOnMachineFunction(MachineFunction &MF) { CurrMF = &MF; TM = &MF.getTarget(); + TRI = TM->getRegisterInfo(); TII = TM->getInstrInfo(); MFI = MF.getFrameInfo(); MRI = &MF.getRegInfo(); LIs = &getAnalysis(); LSs = &getAnalysis(); + VRM = &getAnalysis(); bool MadeChange = false;