X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FInlineSpiller.cpp;h=c6d1a18dbd066091405e483b4e4932b5a6efa573;hb=021e3b6444e9179751002db7b20de96455e03171;hp=4bd35c3812aa8ab932303357ca4125e7c2ad3d8f;hpb=9693d4c3f0d0e8c91e4748eb1c18799b1de7551c;p=oota-llvm.git diff --git a/lib/CodeGen/InlineSpiller.cpp b/lib/CodeGen/InlineSpiller.cpp index 4bd35c3812a..c6d1a18dbd0 100644 --- a/lib/CodeGen/InlineSpiller.cpp +++ b/lib/CodeGen/InlineSpiller.cpp @@ -14,39 +14,44 @@ #define DEBUG_TYPE "regalloc" #include "Spiller.h" -#include "LiveRangeEdit.h" -#include "VirtRegMap.h" #include "llvm/ADT/Statistic.h" +#include "llvm/ADT/TinyPtrVector.h" #include "llvm/Analysis/AliasAnalysis.h" #include "llvm/CodeGen/LiveIntervalAnalysis.h" +#include "llvm/CodeGen/LiveRangeEdit.h" #include "llvm/CodeGen/LiveStackAnalysis.h" #include "llvm/CodeGen/MachineDominators.h" #include "llvm/CodeGen/MachineFrameInfo.h" #include "llvm/CodeGen/MachineFunction.h" +#include "llvm/CodeGen/MachineInstrBundle.h" #include "llvm/CodeGen/MachineLoopInfo.h" #include "llvm/CodeGen/MachineRegisterInfo.h" -#include "llvm/Target/TargetMachine.h" -#include "llvm/Target/TargetInstrInfo.h" +#include "llvm/CodeGen/VirtRegMap.h" +#include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" +#include "llvm/Target/TargetInstrInfo.h" +#include "llvm/Target/TargetMachine.h" using namespace llvm; STATISTIC(NumSpilledRanges, "Number of spilled live ranges"); -STATISTIC(NumSnippets, "Number of snippets included in spills"); +STATISTIC(NumSnippets, "Number of spilled snippets"); STATISTIC(NumSpills, "Number of spills inserted"); +STATISTIC(NumSpillsRemoved, "Number of spills removed"); STATISTIC(NumReloads, "Number of reloads inserted"); +STATISTIC(NumReloadsRemoved, "Number of reloads removed"); STATISTIC(NumFolded, "Number of folded stack accesses"); STATISTIC(NumFoldedLoads, "Number of folded loads"); STATISTIC(NumRemats, "Number of rematerialized defs for spilling"); -STATISTIC(NumOmitReloadSpill, "Number of omitted spills after reloads"); -STATISTIC(NumHoistLocal, "Number of locally hoisted spills"); -STATISTIC(NumHoistGlobal, "Number of globally hoisted spills"); -STATISTIC(NumRedundantSpills, "Number of redundant spills identified"); +STATISTIC(NumOmitReloadSpill, "Number of omitted spills of reloads"); +STATISTIC(NumHoists, "Number of hoisted spills"); + +static cl::opt DisableHoisting("disable-spill-hoist", cl::Hidden, + cl::desc("Disable inline spill hoisting")); namespace { class InlineSpiller : public Spiller { - MachineFunctionPass &Pass; MachineFunction &MF; LiveIntervals &LIS; LiveStacks &LSS; @@ -75,26 +80,49 @@ class InlineSpiller : public Spiller { // Values that failed to remat at some point. SmallPtrSet UsedValues; +public: // Information about a value that was defined by a copy from a sibling // register. struct SibValueInfo { // True when all reaching defs were reloads: No spill is necessary. bool AllDefsAreReloads; + // True when value is defined by an original PHI not from splitting. + bool DefByOrigPHI; + + // True when the COPY defining this value killed its source. + bool KillsSource; + // The preferred register to spill. unsigned SpillReg; // The value of SpillReg that should be spilled. VNInfo *SpillVNI; + // The block where SpillVNI should be spilled. Currently, this must be the + // block containing SpillVNI->def. + MachineBasicBlock *SpillMBB; + // A defining instruction that is not a sibling copy or a reload, or NULL. // This can be used as a template for rematerialization. MachineInstr *DefMI; + // List of values that depend on this one. These values are actually the + // same, but live range splitting has placed them in different registers, + // or SSA update needed to insert PHI-defs to preserve SSA form. This is + // copies of the current value and phi-kills. Usually only phi-kills cause + // more than one dependent value. + TinyPtrVector Deps; + SibValueInfo(unsigned Reg, VNInfo *VNI) - : AllDefsAreReloads(false), SpillReg(Reg), SpillVNI(VNI), DefMI(0) {} + : AllDefsAreReloads(true), DefByOrigPHI(false), KillsSource(false), + SpillReg(Reg), SpillVNI(VNI), SpillMBB(0), DefMI(0) {} + + // Returns true when a def has been found. + bool hasDef() const { return DefByOrigPHI || DefMI; } }; +private: // Values in RegsToSpill defined by sibling copies. typedef DenseMap SibValueMap; SibValueMap SibValues; @@ -108,8 +136,7 @@ public: InlineSpiller(MachineFunctionPass &pass, MachineFunction &mf, VirtRegMap &vrm) - : Pass(pass), - MF(mf), + : MF(mf), LIS(pass.getAnalysis()), LSS(pass.getAnalysis()), AA(&pass.getAnalysis()), @@ -134,6 +161,7 @@ private: bool isSibling(unsigned Reg); MachineInstr *traceSiblingValue(unsigned, VNInfo*, VNInfo*); + void propagateSiblingValue(SibValueMap::iterator, VNInfo *VNI = 0); void analyzeSiblingValues(); bool hoistSpill(LiveInterval &SpillLI, MachineInstr *CopyMI); @@ -144,8 +172,7 @@ private: void reMaterializeAll(); bool coalesceStackAccess(MachineInstr *MI, unsigned Reg); - bool foldMemoryOperand(MachineBasicBlock::iterator MI, - const SmallVectorImpl &Ops, + bool foldMemoryOperand(ArrayRef >, MachineInstr *LoadMI = 0); void insertReload(LiveInterval &NewLI, SlotIndex, MachineBasicBlock::iterator MI); @@ -282,6 +309,156 @@ bool InlineSpiller::isSibling(unsigned Reg) { VRM.getOriginal(Reg) == Original; } +#ifndef NDEBUG +static raw_ostream &operator<<(raw_ostream &OS, + const InlineSpiller::SibValueInfo &SVI) { + OS << "spill " << PrintReg(SVI.SpillReg) << ':' + << SVI.SpillVNI->id << '@' << SVI.SpillVNI->def; + if (SVI.SpillMBB) + OS << " in BB#" << SVI.SpillMBB->getNumber(); + if (SVI.AllDefsAreReloads) + OS << " all-reloads"; + if (SVI.DefByOrigPHI) + OS << " orig-phi"; + if (SVI.KillsSource) + OS << " kill"; + OS << " deps["; + for (unsigned i = 0, e = SVI.Deps.size(); i != e; ++i) + OS << ' ' << SVI.Deps[i]->id << '@' << SVI.Deps[i]->def; + OS << " ]"; + if (SVI.DefMI) + OS << " def: " << *SVI.DefMI; + else + OS << '\n'; + return OS; +} +#endif + +/// propagateSiblingValue - Propagate the value in SVI to dependents if it is +/// known. Otherwise remember the dependency for later. +/// +/// @param SVI SibValues entry to propagate. +/// @param VNI Dependent value, or NULL to propagate to all saved dependents. +void InlineSpiller::propagateSiblingValue(SibValueMap::iterator SVI, + VNInfo *VNI) { + // When VNI is non-NULL, add it to SVI's deps, and only propagate to that. + TinyPtrVector FirstDeps; + if (VNI) { + FirstDeps.push_back(VNI); + SVI->second.Deps.push_back(VNI); + } + + // Has the value been completely determined yet? If not, defer propagation. + if (!SVI->second.hasDef()) + return; + + // Work list of values to propagate. It would be nice to use a SetVector + // here, but then we would be forced to use a SmallSet. + SmallVector WorkList(1, SVI); + SmallPtrSet WorkSet; + + do { + SVI = WorkList.pop_back_val(); + WorkSet.erase(SVI->first); + TinyPtrVector *Deps = VNI ? &FirstDeps : &SVI->second.Deps; + VNI = 0; + + SibValueInfo &SV = SVI->second; + if (!SV.SpillMBB) + SV.SpillMBB = LIS.getMBBFromIndex(SV.SpillVNI->def); + + DEBUG(dbgs() << " prop to " << Deps->size() << ": " + << SVI->first->id << '@' << SVI->first->def << ":\t" << SV); + + assert(SV.hasDef() && "Propagating undefined value"); + + // Should this value be propagated as a preferred spill candidate? We don't + // propagate values of registers that are about to spill. + bool PropSpill = !DisableHoisting && !isRegToSpill(SV.SpillReg); + unsigned SpillDepth = ~0u; + + for (TinyPtrVector::iterator DepI = Deps->begin(), + DepE = Deps->end(); DepI != DepE; ++DepI) { + SibValueMap::iterator DepSVI = SibValues.find(*DepI); + assert(DepSVI != SibValues.end() && "Dependent value not in SibValues"); + SibValueInfo &DepSV = DepSVI->second; + if (!DepSV.SpillMBB) + DepSV.SpillMBB = LIS.getMBBFromIndex(DepSV.SpillVNI->def); + + bool Changed = false; + + // Propagate defining instruction. + if (!DepSV.hasDef()) { + Changed = true; + DepSV.DefMI = SV.DefMI; + DepSV.DefByOrigPHI = SV.DefByOrigPHI; + } + + // Propagate AllDefsAreReloads. For PHI values, this computes an AND of + // all predecessors. + if (!SV.AllDefsAreReloads && DepSV.AllDefsAreReloads) { + Changed = true; + DepSV.AllDefsAreReloads = false; + } + + // Propagate best spill value. + if (PropSpill && SV.SpillVNI != DepSV.SpillVNI) { + if (SV.SpillMBB == DepSV.SpillMBB) { + // DepSV is in the same block. Hoist when dominated. + if (DepSV.KillsSource && SV.SpillVNI->def < DepSV.SpillVNI->def) { + // This is an alternative def earlier in the same MBB. + // Hoist the spill as far as possible in SpillMBB. This can ease + // register pressure: + // + // x = def + // y = use x + // s = copy x + // + // Hoisting the spill of s to immediately after the def removes the + // interference between x and y: + // + // x = def + // spill x + // y = use x + // + // This hoist only helps when the DepSV copy kills its source. + Changed = true; + DepSV.SpillReg = SV.SpillReg; + DepSV.SpillVNI = SV.SpillVNI; + DepSV.SpillMBB = SV.SpillMBB; + } + } else { + // DepSV is in a different block. + if (SpillDepth == ~0u) + SpillDepth = Loops.getLoopDepth(SV.SpillMBB); + + // Also hoist spills to blocks with smaller loop depth, but make sure + // that the new value dominates. Non-phi dependents are always + // dominated, phis need checking. + if ((Loops.getLoopDepth(DepSV.SpillMBB) > SpillDepth) && + (!DepSVI->first->isPHIDef() || + MDT.dominates(SV.SpillMBB, DepSV.SpillMBB))) { + Changed = true; + DepSV.SpillReg = SV.SpillReg; + DepSV.SpillVNI = SV.SpillVNI; + DepSV.SpillMBB = SV.SpillMBB; + } + } + } + + if (!Changed) + continue; + + // Something changed in DepSVI. Propagate to dependents. + if (WorkSet.insert(DepSVI->first)) + WorkList.push_back(DepSVI); + + DEBUG(dbgs() << " update " << DepSVI->first->id << '@' + << DepSVI->first->def << " to:\t" << DepSV); + } + } while (!WorkList.empty()); +} + /// traceSiblingValue - Trace a value that is about to be spilled back to the /// real defining instructions by looking through sibling copies. Always stay /// within the range of OrigVNI so the registers are known to carry the same @@ -294,59 +471,101 @@ bool InlineSpiller::isSibling(unsigned Reg) { /// MachineInstr *InlineSpiller::traceSiblingValue(unsigned UseReg, VNInfo *UseVNI, VNInfo *OrigVNI) { + // Check if a cached value already exists. + SibValueMap::iterator SVI; + bool Inserted; + tie(SVI, Inserted) = + SibValues.insert(std::make_pair(UseVNI, SibValueInfo(UseReg, UseVNI))); + if (!Inserted) { + DEBUG(dbgs() << "Cached value " << PrintReg(UseReg) << ':' + << UseVNI->id << '@' << UseVNI->def << ' ' << SVI->second); + return SVI->second.DefMI; + } + DEBUG(dbgs() << "Tracing value " << PrintReg(UseReg) << ':' << UseVNI->id << '@' << UseVNI->def << '\n'); - SmallPtrSet Visited; + + // List of (Reg, VNI) that have been inserted into SibValues, but need to be + // processed. SmallVector, 8> WorkList; WorkList.push_back(std::make_pair(UseReg, UseVNI)); - // Best spill candidate seen so far. This must dominate UseVNI. - SibValueInfo SVI(UseReg, UseVNI); - MachineBasicBlock *UseMBB = LIS.getMBBFromIndex(UseVNI->def); - unsigned SpillDepth = Loops.getLoopDepth(UseMBB); - bool SeenOrigPHI = false; // Original PHI met. - do { unsigned Reg; VNInfo *VNI; tie(Reg, VNI) = WorkList.pop_back_val(); - if (!Visited.insert(VNI)) - continue; + DEBUG(dbgs() << " " << PrintReg(Reg) << ':' << VNI->id << '@' << VNI->def + << ":\t"); - // Is this value a better spill candidate? - if (!isRegToSpill(Reg)) { - MachineBasicBlock *MBB = LIS.getMBBFromIndex(VNI->def); - if (MBB != UseMBB && MDT.dominates(MBB, UseMBB)) { - // This is a valid spill location dominating UseVNI. - // Prefer to spill at a smaller loop depth. - unsigned Depth = Loops.getLoopDepth(MBB); - if (Depth < SpillDepth) { - DEBUG(dbgs() << " spill depth " << Depth << ": " << PrintReg(Reg) - << ':' << VNI->id << '@' << VNI->def << '\n'); - SVI.SpillReg = Reg; - SVI.SpillVNI = VNI; - SpillDepth = Depth; - } - } - } + // First check if this value has already been computed. + SVI = SibValues.find(VNI); + assert(SVI != SibValues.end() && "Missing SibValues entry"); // Trace through PHI-defs created by live range splitting. if (VNI->isPHIDef()) { + // Stop at original PHIs. We don't know the value at the predecessors. if (VNI->def == OrigVNI->def) { - DEBUG(dbgs() << " orig phi value " << PrintReg(Reg) << ':' - << VNI->id << '@' << VNI->def << '\n'); - SeenOrigPHI = true; + DEBUG(dbgs() << "orig phi value\n"); + SVI->second.DefByOrigPHI = true; + SVI->second.AllDefsAreReloads = false; + propagateSiblingValue(SVI); continue; } - // Get values live-out of predecessors. + + // This is a PHI inserted by live range splitting. We could trace the + // live-out value from predecessor blocks, but that search can be very + // expensive if there are many predecessors and many more PHIs as + // generated by tail-dup when it sees an indirectbr. Instead, look at + // all the non-PHI defs that have the same value as OrigVNI. They must + // jointly dominate VNI->def. This is not optimal since VNI may actually + // be jointly dominated by a smaller subset of defs, so there is a change + // we will miss a AllDefsAreReloads optimization. + + // Separate all values dominated by OrigVNI into PHIs and non-PHIs. + SmallVector PHIs, NonPHIs; LiveInterval &LI = LIS.getInterval(Reg); - MachineBasicBlock *MBB = LIS.getMBBFromIndex(VNI->def); - for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), - PE = MBB->pred_end(); PI != PE; ++PI) { - VNInfo *PVNI = LI.getVNInfoAt(LIS.getMBBEndIdx(*PI).getPrevSlot()); - if (PVNI) - WorkList.push_back(std::make_pair(Reg, PVNI)); + LiveInterval &OrigLI = LIS.getInterval(Original); + + for (LiveInterval::vni_iterator VI = LI.vni_begin(), VE = LI.vni_end(); + VI != VE; ++VI) { + VNInfo *VNI2 = *VI; + if (VNI2->isUnused()) + continue; + if (!OrigLI.containsOneValue() && + OrigLI.getVNInfoAt(VNI2->def) != OrigVNI) + continue; + if (VNI2->isPHIDef() && VNI2->def != OrigVNI->def) + PHIs.push_back(VNI2); + else + NonPHIs.push_back(VNI2); } + DEBUG(dbgs() << "split phi value, checking " << PHIs.size() + << " phi-defs, and " << NonPHIs.size() + << " non-phi/orig defs\n"); + + // Create entries for all the PHIs. Don't add them to the worklist, we + // are processing all of them in one go here. + for (unsigned i = 0, e = PHIs.size(); i != e; ++i) + SibValues.insert(std::make_pair(PHIs[i], SibValueInfo(Reg, PHIs[i]))); + + // Add every PHI as a dependent of all the non-PHIs. + for (unsigned i = 0, e = NonPHIs.size(); i != e; ++i) { + VNInfo *NonPHI = NonPHIs[i]; + // Known value? Try an insertion. + tie(SVI, Inserted) = + SibValues.insert(std::make_pair(NonPHI, SibValueInfo(Reg, NonPHI))); + // Add all the PHIs as dependents of NonPHI. + for (unsigned pi = 0, pe = PHIs.size(); pi != pe; ++pi) + SVI->second.Deps.push_back(PHIs[pi]); + // This is the first time we see NonPHI, add it to the worklist. + if (Inserted) + WorkList.push_back(std::make_pair(Reg, NonPHI)); + else + // Propagate to all inserted PHIs, not just VNI. + propagateSiblingValue(SVI); + } + + // Next work list item. continue; } @@ -357,48 +576,49 @@ MachineInstr *InlineSpiller::traceSiblingValue(unsigned UseReg, VNInfo *UseVNI, if (unsigned SrcReg = isFullCopyOf(MI, Reg)) { if (isSibling(SrcReg)) { LiveInterval &SrcLI = LIS.getInterval(SrcReg); - VNInfo *SrcVNI = SrcLI.getVNInfoAt(VNI->def.getUseIndex()); - assert(SrcVNI && "Copy from non-existing value"); - DEBUG(dbgs() << " copy of " << PrintReg(SrcReg) << ':' - << SrcVNI->id << '@' << SrcVNI->def << '\n'); - WorkList.push_back(std::make_pair(SrcReg, SrcVNI)); + LiveRangeQuery SrcQ(SrcLI, VNI->def); + assert(SrcQ.valueIn() && "Copy from non-existing value"); + // Check if this COPY kills its source. + SVI->second.KillsSource = SrcQ.isKill(); + VNInfo *SrcVNI = SrcQ.valueIn(); + DEBUG(dbgs() << "copy of " << PrintReg(SrcReg) << ':' + << SrcVNI->id << '@' << SrcVNI->def + << " kill=" << unsigned(SVI->second.KillsSource) << '\n'); + // Known sibling source value? Try an insertion. + tie(SVI, Inserted) = SibValues.insert(std::make_pair(SrcVNI, + SibValueInfo(SrcReg, SrcVNI))); + // This is the first time we see Src, add it to the worklist. + if (Inserted) + WorkList.push_back(std::make_pair(SrcReg, SrcVNI)); + propagateSiblingValue(SVI, VNI); + // Next work list item. continue; } } // Track reachable reloads. + SVI->second.DefMI = MI; + SVI->second.SpillMBB = MI->getParent(); int FI; if (Reg == TII.isLoadFromStackSlot(MI, FI) && FI == StackSlot) { - DEBUG(dbgs() << " reload " << PrintReg(Reg) << ':' - << VNI->id << "@" << VNI->def << '\n'); - SVI.AllDefsAreReloads = true; + DEBUG(dbgs() << "reload\n"); + propagateSiblingValue(SVI); + // Next work list item. continue; } - // We have an 'original' def. Don't record trivial cases. - if (VNI == UseVNI) { - DEBUG(dbgs() << "Not a sibling copy.\n"); - return MI; - } - // Potential remat candidate. - DEBUG(dbgs() << " def " << PrintReg(Reg) << ':' - << VNI->id << '@' << VNI->def << '\t' << *MI); - SVI.DefMI = MI; + DEBUG(dbgs() << "def " << *MI); + SVI->second.AllDefsAreReloads = false; + propagateSiblingValue(SVI); } while (!WorkList.empty()); - if (SeenOrigPHI || SVI.DefMI) - SVI.AllDefsAreReloads = false; - - DEBUG({ - if (SVI.AllDefsAreReloads) - dbgs() << "All defs are reloads.\n"; - else - dbgs() << "Prefer to spill " << PrintReg(SVI.SpillReg) << ':' - << SVI.SpillVNI->id << '@' << SVI.SpillVNI->def << '\n'; - }); - SibValues.insert(std::make_pair(UseVNI, SVI)); - return SVI.DefMI; + // Look up the value we were looking for. We already did this lookup at the + // top of the function, but SibValues may have been invalidated. + SVI = SibValues.find(UseVNI); + assert(SVI != SibValues.end() && "Didn't compute requested info"); + DEBUG(dbgs() << " traced to:\t" << SVI->second); + return SVI->second.DefMI; } /// analyzeSiblingValues - Trace values defined by sibling copies back to @@ -422,16 +642,18 @@ void InlineSpiller::analyzeSiblingValues() { if (VNI->isUnused()) continue; MachineInstr *DefMI = 0; + if (!VNI->isPHIDef()) { + DefMI = LIS.getInstructionFromIndex(VNI->def); + assert(DefMI && "No defining instruction"); + } // Check possible sibling copies. - if (VNI->isPHIDef() || VNI->getCopy()) { + if (VNI->isPHIDef() || DefMI->isCopy()) { VNInfo *OrigVNI = OrigLI.getVNInfoAt(VNI->def); assert(OrigVNI && "Def outside original live range"); if (OrigVNI->def != VNI->def) DefMI = traceSiblingValue(Reg, VNI, OrigVNI); } - if (!DefMI && !VNI->isPHIDef()) - DefMI = LIS.getInstructionFromIndex(VNI->def); - if (DefMI && Edit->checkRematerializable(VNI, DefMI, TII, AA)) { + if (DefMI && Edit->checkRematerializable(VNI, DefMI, AA)) { DEBUG(dbgs() << "Value " << PrintReg(Reg) << ':' << VNI->id << '@' << VNI->def << " may remat from " << *DefMI); } @@ -443,8 +665,8 @@ void InlineSpiller::analyzeSiblingValues() { /// a spill at a better location. bool InlineSpiller::hoistSpill(LiveInterval &SpillLI, MachineInstr *CopyMI) { SlotIndex Idx = LIS.getInstructionIndex(CopyMI); - VNInfo *VNI = SpillLI.getVNInfoAt(Idx.getDefIndex()); - assert(VNI && VNI->def == Idx.getDefIndex() && "Not defined by copy"); + VNInfo *VNI = SpillLI.getVNInfoAt(Idx.getRegSlot()); + assert(VNI && VNI->def == Idx.getRegSlot() && "Not defined by copy"); SibValueMap::iterator I = SibValues.find(VNI); if (I == SibValues.end()) return false; @@ -481,6 +703,7 @@ bool InlineSpiller::hoistSpill(LiveInterval &SpillLI, MachineInstr *CopyMI) { // Already spilled everywhere. if (SVI.AllDefsAreReloads) { + DEBUG(dbgs() << "\tno spill needed: " << SVI); ++NumOmitReloadSpill; return true; } @@ -503,13 +726,10 @@ bool InlineSpiller::hoistSpill(LiveInterval &SpillLI, MachineInstr *CopyMI) { MRI.getRegClass(SVI.SpillReg), &TRI); --MII; // Point to store instruction. LIS.InsertMachineInstrInMaps(MII); - VRM.addSpillSlotUse(StackSlot, MII); DEBUG(dbgs() << "\thoisted: " << SVI.SpillVNI->def << '\t' << *MII); - if (MBB == CopyMI->getParent()) - ++NumHoistLocal; - else - ++NumHoistGlobal; + ++NumSpills; + ++NumHoists; return true; } @@ -539,7 +759,7 @@ void InlineSpiller::eliminateRedundantSpills(LiveInterval &SLI, VNInfo *VNI) { // Find all spills and copies of VNI. for (MachineRegisterInfo::use_nodbg_iterator UI = MRI.use_nodbg_begin(Reg); MachineInstr *MI = UI.skipInstruction();) { - if (!MI->isCopy() && !MI->getDesc().mayStore()) + if (!MI->isCopy() && !MI->mayStore()) continue; SlotIndex Idx = LIS.getInstructionIndex(MI); if (LI->getVNInfoAt(Idx) != VNI) @@ -549,9 +769,9 @@ void InlineSpiller::eliminateRedundantSpills(LiveInterval &SLI, VNInfo *VNI) { if (unsigned DstReg = isFullCopyOf(MI, Reg)) { if (isSibling(DstReg)) { LiveInterval &DstLI = LIS.getInterval(DstReg); - VNInfo *DstVNI = DstLI.getVNInfoAt(Idx.getDefIndex()); + VNInfo *DstVNI = DstLI.getVNInfoAt(Idx.getRegSlot()); assert(DstVNI && "Missing defined value"); - assert(DstVNI->def == Idx.getDefIndex() && "Wrong copy def slot"); + assert(DstVNI->def == Idx.getRegSlot() && "Wrong copy def slot"); WorkList.push_back(std::make_pair(&DstLI, DstVNI)); } continue; @@ -564,7 +784,8 @@ void InlineSpiller::eliminateRedundantSpills(LiveInterval &SLI, VNInfo *VNI) { // eliminateDeadDefs won't normally remove stores, so switch opcode. MI->setDesc(TII.get(TargetOpcode::KILL)); DeadDefs.push_back(MI); - ++NumRedundantSpills; + ++NumSpillsRemoved; + --NumSpills; } } } while (!WorkList.empty()); @@ -589,7 +810,7 @@ void InlineSpiller::markValueUsed(LiveInterval *LI, VNInfo *VNI) { MachineBasicBlock *MBB = LIS.getMBBFromIndex(VNI->def); for (MachineBasicBlock::pred_iterator PI = MBB->pred_begin(), PE = MBB->pred_end(); PI != PE; ++PI) { - VNInfo *PVNI = LI->getVNInfoAt(LIS.getMBBEndIdx(*PI).getPrevSlot()); + VNInfo *PVNI = LI->getVNInfoBefore(LIS.getMBBEndIdx(*PI)); if (PVNI) WorkList.push_back(std::make_pair(LI, PVNI)); } @@ -602,7 +823,7 @@ void InlineSpiller::markValueUsed(LiveInterval *LI, VNInfo *VNI) { continue; LiveInterval &SnipLI = LIS.getInterval(MI->getOperand(1).getReg()); assert(isRegToSpill(SnipLI.reg) && "Unexpected register in copy"); - VNInfo *SnipVNI = SnipLI.getVNInfoAt(VNI->def.getUseIndex()); + VNInfo *SnipVNI = SnipLI.getVNInfoAt(VNI->def.getRegSlot(true)); assert(SnipVNI && "Snippet undefined before copy"); WorkList.push_back(std::make_pair(&SnipLI, SnipVNI)); } while (!WorkList.empty()); @@ -611,8 +832,8 @@ void InlineSpiller::markValueUsed(LiveInterval *LI, VNInfo *VNI) { /// reMaterializeFor - Attempt to rematerialize before MI instead of reloading. bool InlineSpiller::reMaterializeFor(LiveInterval &VirtReg, MachineBasicBlock::iterator MI) { - SlotIndex UseIdx = LIS.getInstructionIndex(MI).getUseIndex(); - VNInfo *ParentVNI = VirtReg.getVNInfoAt(UseIdx); + SlotIndex UseIdx = LIS.getInstructionIndex(MI).getRegSlot(true); + VNInfo *ParentVNI = VirtReg.getVNInfoAt(UseIdx.getBaseIndex()); if (!ParentVNI) { DEBUG(dbgs() << "\tadding flags: "); @@ -633,7 +854,7 @@ bool InlineSpiller::reMaterializeFor(LiveInterval &VirtReg, SibValueMap::const_iterator SibI = SibValues.find(ParentVNI); if (SibI != SibValues.end()) RM.OrigMI = SibI->second.DefMI; - if (!Edit->canRematerializeAt(RM, UseIdx, false, LIS)) { + if (!Edit->canRematerializeAt(RM, UseIdx, false)) { markValueUsed(&VirtReg, ParentVNI); DEBUG(dbgs() << "\tcannot remat for " << UseIdx << '\t' << *MI); return false; @@ -641,42 +862,37 @@ bool InlineSpiller::reMaterializeFor(LiveInterval &VirtReg, // If the instruction also writes VirtReg.reg, it had better not require the // same register for uses and defs. - bool Reads, Writes; - SmallVector Ops; - tie(Reads, Writes) = MI->readsWritesVirtualRegister(VirtReg.reg, &Ops); - if (Writes) { - for (unsigned i = 0, e = Ops.size(); i != e; ++i) { - MachineOperand &MO = MI->getOperand(Ops[i]); - if (MO.isUse() ? MI->isRegTiedToDefOperand(Ops[i]) : MO.getSubReg()) { - markValueUsed(&VirtReg, ParentVNI); - DEBUG(dbgs() << "\tcannot remat tied reg: " << UseIdx << '\t' << *MI); - return false; - } - } + SmallVector, 8> Ops; + MIBundleOperands::VirtRegInfo RI = + MIBundleOperands(MI).analyzeVirtReg(VirtReg.reg, &Ops); + if (RI.Tied) { + markValueUsed(&VirtReg, ParentVNI); + DEBUG(dbgs() << "\tcannot remat tied reg: " << UseIdx << '\t' << *MI); + return false; } // Before rematerializing into a register for a single instruction, try to // fold a load into the instruction. That avoids allocating a new register. - if (RM.OrigMI->getDesc().canFoldAsLoad() && - foldMemoryOperand(MI, Ops, RM.OrigMI)) { + if (RM.OrigMI->canFoldAsLoad() && + foldMemoryOperand(Ops, RM.OrigMI)) { Edit->markRematerialized(RM.ParentVNI); ++NumFoldedLoads; return true; } // Alocate a new register for the remat. - LiveInterval &NewLI = Edit->createFrom(Original, LIS, VRM); + LiveInterval &NewLI = Edit->createFrom(Original); NewLI.markNotSpillable(); // Finally we can rematerialize OrigMI before MI. SlotIndex DefIdx = Edit->rematerializeAt(*MI->getParent(), MI, NewLI.reg, RM, - LIS, TII, TRI); + TRI); DEBUG(dbgs() << "\tremat: " << DefIdx << '\t' << *LIS.getInstructionFromIndex(DefIdx)); // Replace operands for (unsigned i = 0, e = Ops.size(); i != e; ++i) { - MachineOperand &MO = MI->getOperand(Ops[i]); + MachineOperand &MO = MI->getOperand(Ops[i].second); if (MO.isReg() && MO.isUse() && MO.getReg() == VirtReg.reg) { MO.setReg(NewLI.reg); MO.setIsKill(); @@ -684,8 +900,8 @@ bool InlineSpiller::reMaterializeFor(LiveInterval &VirtReg, } DEBUG(dbgs() << "\t " << UseIdx << '\t' << *MI); - VNInfo *DefVNI = NewLI.getNextValue(DefIdx, 0, LIS.getVNInfoAllocator()); - NewLI.addRange(LiveRange(DefIdx, UseIdx.getDefIndex(), DefVNI)); + VNInfo *DefVNI = NewLI.getNextValue(DefIdx, LIS.getVNInfoAllocator()); + NewLI.addRange(LiveRange(DefIdx, UseIdx.getRegSlot(), DefVNI)); DEBUG(dbgs() << "\tinterval: " << NewLI << '\n'); ++NumRemats; return true; @@ -695,7 +911,7 @@ bool InlineSpiller::reMaterializeFor(LiveInterval &VirtReg, /// and trim the live ranges after. void InlineSpiller::reMaterializeAll() { // analyzeSiblingValues has already tested all relevant defining instructions. - if (!Edit->anyRematerializable(LIS, TII, AA)) + if (!Edit->anyRematerializable(AA)) return; UsedValues.clear(); @@ -707,7 +923,7 @@ void InlineSpiller::reMaterializeAll() { LiveInterval &LI = LIS.getInterval(Reg); for (MachineRegisterInfo::use_nodbg_iterator RI = MRI.use_nodbg_begin(Reg); - MachineInstr *MI = RI.skipInstruction();) + MachineInstr *MI = RI.skipBundle();) anyRemat |= reMaterializeFor(LI, MI); } if (!anyRemat) @@ -736,7 +952,7 @@ void InlineSpiller::reMaterializeAll() { if (DeadDefs.empty()) return; DEBUG(dbgs() << "Remat created " << DeadDefs.size() << " dead defs.\n"); - Edit->eliminateDeadDefs(DeadDefs, LIS, VRM, TII); + Edit->eliminateDeadDefs(DeadDefs, RegsToSpill); // Get rid of deleted and empty intervals. for (unsigned i = RegsToSpill.size(); i != 0; --i) { @@ -748,7 +964,7 @@ void InlineSpiller::reMaterializeAll() { LiveInterval &LI = LIS.getInterval(Reg); if (!LI.empty()) continue; - Edit->eraseVirtReg(Reg, LIS); + Edit->eraseVirtReg(Reg); RegsToSpill.erase(RegsToSpill.begin() + (i - 1)); } DEBUG(dbgs() << RegsToSpill.size() << " registers to spill after remat.\n"); @@ -762,10 +978,10 @@ void InlineSpiller::reMaterializeAll() { /// If MI is a load or store of StackSlot, it can be removed. bool InlineSpiller::coalesceStackAccess(MachineInstr *MI, unsigned Reg) { int FI = 0; - unsigned InstrReg; - if (!(InstrReg = TII.isLoadFromStackSlot(MI, FI)) && - !(InstrReg = TII.isStoreToStackSlot(MI, FI))) - return false; + unsigned InstrReg = TII.isLoadFromStackSlot(MI, FI); + bool IsLoad = InstrReg; + if (!IsLoad) + InstrReg = TII.isStoreToStackSlot(MI, FI); // We have a stack access. Is it the right register and slot? if (InstrReg != Reg || FI != StackSlot) @@ -774,25 +990,47 @@ bool InlineSpiller::coalesceStackAccess(MachineInstr *MI, unsigned Reg) { DEBUG(dbgs() << "Coalescing stack access: " << *MI); LIS.RemoveMachineInstrFromMaps(MI); MI->eraseFromParent(); + + if (IsLoad) { + ++NumReloadsRemoved; + --NumReloads; + } else { + ++NumSpillsRemoved; + --NumSpills; + } + return true; } -/// foldMemoryOperand - Try folding stack slot references in Ops into MI. -/// @param MI Instruction using or defining the current register. -/// @param Ops Operand indices from readsWritesVirtualRegister(). +/// foldMemoryOperand - Try folding stack slot references in Ops into their +/// instructions. +/// +/// @param Ops Operand indices from analyzeVirtReg(). /// @param LoadMI Load instruction to use instead of stack slot when non-null. -/// @return True on success, and MI will be erased. -bool InlineSpiller::foldMemoryOperand(MachineBasicBlock::iterator MI, - const SmallVectorImpl &Ops, - MachineInstr *LoadMI) { +/// @return True on success. +bool InlineSpiller:: +foldMemoryOperand(ArrayRef > Ops, + MachineInstr *LoadMI) { + if (Ops.empty()) + return false; + // Don't attempt folding in bundles. + MachineInstr *MI = Ops.front().first; + if (Ops.back().first != MI || MI->isBundled()) + return false; + + bool WasCopy = MI->isCopy(); + unsigned ImpReg = 0; + // TargetInstrInfo::foldMemoryOperand only expects explicit, non-tied // operands. SmallVector FoldOps; for (unsigned i = 0, e = Ops.size(); i != e; ++i) { - unsigned Idx = Ops[i]; + unsigned Idx = Ops[i].second; MachineOperand &MO = MI->getOperand(Idx); - if (MO.isImplicit()) + if (MO.isImplicit()) { + ImpReg = MO.getReg(); continue; + } // FIXME: Teach targets to deal with subregs. if (MO.getSubReg()) return false; @@ -810,11 +1048,27 @@ bool InlineSpiller::foldMemoryOperand(MachineBasicBlock::iterator MI, if (!FoldMI) return false; LIS.ReplaceMachineInstrInMaps(MI, FoldMI); - if (!LoadMI) - VRM.addSpillSlotUse(StackSlot, FoldMI); MI->eraseFromParent(); - DEBUG(dbgs() << "\tfolded: " << *FoldMI); - ++NumFolded; + + // TII.foldMemoryOperand may have left some implicit operands on the + // instruction. Strip them. + if (ImpReg) + for (unsigned i = FoldMI->getNumOperands(); i; --i) { + MachineOperand &MO = FoldMI->getOperand(i - 1); + if (!MO.isReg() || !MO.isImplicit()) + break; + if (MO.getReg() == ImpReg) + FoldMI->RemoveOperand(i - 1); + } + + DEBUG(dbgs() << "\tfolded: " << LIS.getInstructionIndex(FoldMI) << '\t' + << *FoldMI); + if (!WasCopy) + ++NumFolded; + else if (Ops.front().second == 0) + ++NumSpills; + else + ++NumReloads; return true; } @@ -826,11 +1080,13 @@ void InlineSpiller::insertReload(LiveInterval &NewLI, TII.loadRegFromStackSlot(MBB, MI, NewLI.reg, StackSlot, MRI.getRegClass(NewLI.reg), &TRI); --MI; // Point to load instruction. - SlotIndex LoadIdx = LIS.InsertMachineInstrInMaps(MI).getDefIndex(); - VRM.addSpillSlotUse(StackSlot, MI); + SlotIndex LoadIdx = LIS.InsertMachineInstrInMaps(MI).getRegSlot(); + // Some (out-of-tree) targets have EC reload instructions. + if (MachineOperand *MO = MI->findRegisterDefOperand(NewLI.reg)) + if (MO->isEarlyClobber()) + LoadIdx = LoadIdx.getRegSlot(true); DEBUG(dbgs() << "\treload: " << LoadIdx << '\t' << *MI); - VNInfo *LoadVNI = NewLI.getNextValue(LoadIdx, 0, - LIS.getVNInfoAllocator()); + VNInfo *LoadVNI = NewLI.getNextValue(LoadIdx, LIS.getVNInfoAllocator()); NewLI.addRange(LiveRange(LoadIdx, Idx, LoadVNI)); ++NumReloads; } @@ -842,10 +1098,9 @@ void InlineSpiller::insertSpill(LiveInterval &NewLI, const LiveInterval &OldLI, TII.storeRegToStackSlot(MBB, ++MI, NewLI.reg, true, StackSlot, MRI.getRegClass(NewLI.reg), &TRI); --MI; // Point to store instruction. - SlotIndex StoreIdx = LIS.InsertMachineInstrInMaps(MI).getDefIndex(); - VRM.addSpillSlotUse(StackSlot, MI); + SlotIndex StoreIdx = LIS.InsertMachineInstrInMaps(MI).getRegSlot(); DEBUG(dbgs() << "\tspilled: " << StoreIdx << '\t' << *MI); - VNInfo *StoreVNI = NewLI.getNextValue(Idx, 0, LIS.getVNInfoAllocator()); + VNInfo *StoreVNI = NewLI.getNextValue(Idx, LIS.getVNInfoAllocator()); NewLI.addRange(LiveRange(Idx, StoreIdx, StoreVNI)); ++NumSpills; } @@ -856,8 +1111,8 @@ void InlineSpiller::spillAroundUses(unsigned Reg) { LiveInterval &OldLI = LIS.getInterval(Reg); // Iterate over instructions using Reg. - for (MachineRegisterInfo::reg_iterator RI = MRI.reg_begin(Reg); - MachineInstr *MI = RI.skipInstruction();) { + for (MachineRegisterInfo::reg_iterator RegI = MRI.reg_begin(Reg); + MachineInstr *MI = RegI.skipBundle();) { // Debug values are not allowed to affect codegen. if (MI->isDebugValue()) { @@ -886,14 +1141,14 @@ void InlineSpiller::spillAroundUses(unsigned Reg) { continue; // Analyze instruction. - bool Reads, Writes; - SmallVector Ops; - tie(Reads, Writes) = MI->readsWritesVirtualRegister(Reg, &Ops); + SmallVector, 8> Ops; + MIBundleOperands::VirtRegInfo RI = + MIBundleOperands(MI).analyzeVirtReg(Reg, &Ops); // Find the slot index where this instruction reads and writes OldLI. // This is usually the def slot, except for tied early clobbers. - SlotIndex Idx = LIS.getInstructionIndex(MI).getDefIndex(); - if (VNInfo *VNI = OldLI.getVNInfoAt(Idx.getUseIndex())) + SlotIndex Idx = LIS.getInstructionIndex(MI).getRegSlot(); + if (VNInfo *VNI = OldLI.getVNInfoAt(Idx.getRegSlot(true))) if (SlotIndex::isSameInstr(Idx, VNI->def)) Idx = VNI->def; @@ -906,7 +1161,7 @@ void InlineSpiller::spillAroundUses(unsigned Reg) { SnippetCopies.insert(MI); continue; } - if (Writes) { + if (RI.Writes) { // Hoist the spill of a sib-reg copy. if (hoistSpill(OldLI, MI)) { // This COPY is now dead, the value is already in the stack slot. @@ -923,24 +1178,24 @@ void InlineSpiller::spillAroundUses(unsigned Reg) { } // Attempt to fold memory ops. - if (foldMemoryOperand(MI, Ops)) + if (foldMemoryOperand(Ops)) continue; // Allocate interval around instruction. // FIXME: Infer regclass from instruction alone. - LiveInterval &NewLI = Edit->createFrom(Reg, LIS, VRM); + LiveInterval &NewLI = Edit->createFrom(Reg); NewLI.markNotSpillable(); - if (Reads) + if (RI.Reads) insertReload(NewLI, Idx, MI); // Rewrite instruction operands. bool hasLiveDef = false; for (unsigned i = 0, e = Ops.size(); i != e; ++i) { - MachineOperand &MO = MI->getOperand(Ops[i]); + MachineOperand &MO = Ops[i].first->getOperand(Ops[i].second); MO.setReg(NewLI.reg); if (MO.isUse()) { - if (!MI->isRegTiedToDefOperand(Ops[i])) + if (!Ops[i].first->isRegTiedToDefOperand(Ops[i].second)) MO.setIsKill(); } else { if (!MO.isDead()) @@ -950,8 +1205,16 @@ void InlineSpiller::spillAroundUses(unsigned Reg) { DEBUG(dbgs() << "\trewrite: " << Idx << '\t' << *MI); // FIXME: Use a second vreg if instruction has no tied ops. - if (Writes && hasLiveDef) - insertSpill(NewLI, OldLI, Idx, MI); + if (RI.Writes) { + if (hasLiveDef) + insertSpill(NewLI, OldLI, Idx, MI); + else { + // This instruction defines a dead value. We don't need to spill it, + // but do create a live range for the dead value. + VNInfo *VNI = NewLI.getNextValue(Idx, LIS.getVNInfoAllocator()); + NewLI.addRange(LiveRange(Idx, Idx.getDeadSlot(), VNI)); + } + } DEBUG(dbgs() << "\tinterval: " << NewLI << '\n'); } @@ -963,7 +1226,7 @@ void InlineSpiller::spillAll() { if (StackSlot == VirtRegMap::NO_STACK_SLOT) { StackSlot = VRM.assignVirt2StackSlot(Original); StackInt = &LSS.getOrCreateInterval(StackSlot, MRI.getRegClass(Original)); - StackInt->getNextValue(SlotIndex(), 0, LSS.getVNInfoAllocator()); + StackInt->getNextValue(SlotIndex(), LSS.getVNInfoAllocator()); } else StackInt = &LSS.getInterval(StackSlot); @@ -983,7 +1246,7 @@ void InlineSpiller::spillAll() { // Hoisted spills may cause dead code. if (!DeadDefs.empty()) { DEBUG(dbgs() << "Eliminating " << DeadDefs.size() << " dead defs\n"); - Edit->eliminateDeadDefs(DeadDefs, LIS, VRM, TII); + Edit->eliminateDeadDefs(DeadDefs, RegsToSpill); } // Finally delete the SnippetCopies. @@ -992,7 +1255,6 @@ void InlineSpiller::spillAll() { MachineInstr *MI = RI.skipInstruction();) { assert(SnippetCopies.count(MI) && "Remaining use wasn't a snippet copy"); // FIXME: Do this with a LiveRangeEdit callback. - VRM.RemoveMachineInstrFromMaps(MI); LIS.RemoveMachineInstrFromMaps(MI); MI->eraseFromParent(); } @@ -1000,7 +1262,7 @@ void InlineSpiller::spillAll() { // Delete all spilled registers. for (unsigned i = 0, e = RegsToSpill.size(); i != e; ++i) - Edit->eraseVirtReg(RegsToSpill[i], LIS); + Edit->eraseVirtReg(RegsToSpill[i]); } void InlineSpiller::spill(LiveRangeEdit &edit) { @@ -1015,8 +1277,8 @@ void InlineSpiller::spill(LiveRangeEdit &edit) { DEBUG(dbgs() << "Inline spilling " << MRI.getRegClass(edit.getReg())->getName() - << ':' << edit.getParent() << "\nFrom original " - << LIS.getInterval(Original) << '\n'); + << ':' << PrintReg(edit.getReg()) << ' ' << edit.getParent() + << "\nFrom original " << LIS.getInterval(Original) << '\n'); assert(edit.getParent().isSpillable() && "Attempting to spill already spilled value."); assert(DeadDefs.empty() && "Previous spill didn't remove dead defs"); @@ -1029,5 +1291,5 @@ void InlineSpiller::spill(LiveRangeEdit &edit) { if (!RegsToSpill.empty()) spillAll(); - Edit->calculateRegClassAndHint(MF, LIS, Loops); + Edit->calculateRegClassAndHint(MF, Loops); }