diff options
author | Lang Hames <lhames@gmail.com> | 2009-12-09 05:39:12 +0000 |
---|---|---|
committer | Lang Hames <lhames@gmail.com> | 2009-12-09 05:39:12 +0000 |
commit | 6194569d22003fddaf1a33acdbb84d5efe76e7d7 (patch) | |
tree | 774665726e58a7ee539ab6188a4feb277c374cdf /lib | |
parent | 164c7ba0ef6a7db1b2d58f959e879ddcdf276ad5 (diff) | |
download | external_llvm-6194569d22003fddaf1a33acdbb84d5efe76e7d7.zip external_llvm-6194569d22003fddaf1a33acdbb84d5efe76e7d7.tar.gz external_llvm-6194569d22003fddaf1a33acdbb84d5efe76e7d7.tar.bz2 |
Added a new "splitting" spiller.
When a call is placed to spill an interval this spiller will first try to
break the interval up into its component values. Single value intervals and
intervals which have already been split (or are the result of previous splits)
are spilled by the default spiller.
Splitting intervals as described above may improve the performance of generated
code in some circumstances. This work is experimental however, and it still
miscompiles many benchmarks. It's not recommended for general use yet.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@90951 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib')
-rw-r--r-- | lib/CodeGen/LiveInterval.cpp | 2 | ||||
-rw-r--r-- | lib/CodeGen/LiveIntervalAnalysis.cpp | 4 | ||||
-rw-r--r-- | lib/CodeGen/RegAllocLinearScan.cpp | 15 | ||||
-rw-r--r-- | lib/CodeGen/Spiller.cpp | 336 | ||||
-rw-r--r-- | lib/CodeGen/Spiller.h | 4 |
5 files changed, 339 insertions, 22 deletions
diff --git a/lib/CodeGen/LiveInterval.cpp b/lib/CodeGen/LiveInterval.cpp index 8d632cb..cc286aa 100644 --- a/lib/CodeGen/LiveInterval.cpp +++ b/lib/CodeGen/LiveInterval.cpp @@ -847,7 +847,7 @@ void LiveInterval::print(raw_ostream &OS, const TargetRegisterInfo *TRI) const { if (vni->isUnused()) { OS << "x"; } else { - if (!vni->isDefAccurate()) + if (!vni->isDefAccurate() && !vni->isPHIDef()) OS << "?"; else OS << vni->def; diff --git a/lib/CodeGen/LiveIntervalAnalysis.cpp b/lib/CodeGen/LiveIntervalAnalysis.cpp index 35337ef..c16ddc5 100644 --- a/lib/CodeGen/LiveIntervalAnalysis.cpp +++ b/lib/CodeGen/LiveIntervalAnalysis.cpp @@ -401,7 +401,7 @@ void LiveIntervals::handleVirtualRegisterDef(MachineBasicBlock *mbb, interval.removeRange(Start, End); assert(interval.ranges.size() == 1 && "Newly discovered PHI interval has >1 ranges."); - MachineBasicBlock *killMBB = getMBBFromIndex(interval.endIndex()); + MachineBasicBlock *killMBB = getMBBFromIndex(VNI->def); VNI->addKill(indexes_->getTerminatorGap(killMBB)); VNI->setHasPHIKill(true); DEBUG({ @@ -412,7 +412,7 @@ void LiveIntervals::handleVirtualRegisterDef(MachineBasicBlock *mbb, // Replace the interval with one of a NEW value number. Note that this // value number isn't actually defined by an instruction, weird huh? :) LiveRange LR(Start, End, - interval.getNextValue(SlotIndex(getMBBStartIdx(mbb), true), + interval.getNextValue(SlotIndex(getMBBStartIdx(Killer->getParent()), true), 0, false, VNInfoAllocator)); LR.valno->setIsPHIDef(true); DEBUG(errs() << " replace range with " << LR); diff --git a/lib/CodeGen/RegAllocLinearScan.cpp b/lib/CodeGen/RegAllocLinearScan.cpp index 4ff5129..c1f5875 100644 --- a/lib/CodeGen/RegAllocLinearScan.cpp +++ b/lib/CodeGen/RegAllocLinearScan.cpp @@ -1261,9 +1261,9 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) { // The earliest start of a Spilled interval indicates up to where // in handled we need to roll back + assert(!spillIs.empty() && "No spill intervals?"); + SlotIndex earliestStart = spillIs[0]->beginIndex(); - LiveInterval *earliestStartInterval = cur; - // Spill live intervals of virtual regs mapped to the physical register we // want to clear (and its aliases). We only spill those that overlap with the // current interval as the rest do not affect its allocation. we also keep @@ -1274,19 +1274,16 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) { LiveInterval *sli = spillIs.back(); spillIs.pop_back(); DEBUG(errs() << "\t\t\tspilling(a): " << *sli << '\n'); - earliestStartInterval = - (earliestStartInterval->beginIndex() < sli->beginIndex()) ? - earliestStartInterval : sli; + if (sli->beginIndex() < earliestStart) + earliestStart = sli->beginIndex(); std::vector<LiveInterval*> newIs; - newIs = spiller_->spill(sli, spillIs); + newIs = spiller_->spill(sli, spillIs, &earliestStart); addStackInterval(sli, ls_, li_, mri_, *vrm_); std::copy(newIs.begin(), newIs.end(), std::back_inserter(added)); spilled.insert(sli->reg); } - SlotIndex earliestStart = earliestStartInterval->beginIndex(); - DEBUG(errs() << "\t\trolling back to: " << earliestStart << '\n'); // Scan handled in reverse order up to the earliest start of a @@ -1295,7 +1292,7 @@ void RALinScan::assignRegOrStackSlotAtInterval(LiveInterval* cur) { while (!handled_.empty()) { LiveInterval* i = handled_.back(); // If this interval starts before t we are done. - if (i->beginIndex() < earliestStart) + if (!i->empty() && i->beginIndex() < earliestStart) break; DEBUG(errs() << "\t\t\tundo changes for: " << *i << '\n'); handled_.pop_back(); diff --git a/lib/CodeGen/Spiller.cpp b/lib/CodeGen/Spiller.cpp index 7466215..bc246c1 100644 --- a/lib/CodeGen/Spiller.cpp +++ b/lib/CodeGen/Spiller.cpp @@ -20,22 +20,25 @@ #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" +#include <set> using namespace llvm; namespace { - enum SpillerName { trivial, standard }; + enum SpillerName { trivial, standard, splitting }; } static cl::opt<SpillerName> spillerOpt("spiller", cl::desc("Spiller to use: (default: standard)"), cl::Prefix, - cl::values(clEnumVal(trivial, "trivial spiller"), - clEnumVal(standard, "default spiller"), + cl::values(clEnumVal(trivial, "trivial spiller"), + clEnumVal(standard, "default spiller"), + clEnumVal(splitting, "splitting spiller"), clEnumValEnd), cl::init(standard)); +// Spiller virtual destructor implementation. Spiller::~Spiller() {} namespace { @@ -170,7 +173,8 @@ public: : SpillerBase(mf, lis, vrm) {} std::vector<LiveInterval*> spill(LiveInterval *li, - SmallVectorImpl<LiveInterval*> &spillIs) { + SmallVectorImpl<LiveInterval*> &spillIs, + SlotIndex*) { // Ignore spillIs - we don't use it. return trivialSpillEverywhere(li); } @@ -179,23 +183,336 @@ public: /// Falls back on LiveIntervals::addIntervalsForSpills. class StandardSpiller : public Spiller { -private: +protected: LiveIntervals *lis; const MachineLoopInfo *loopInfo; VirtRegMap *vrm; public: - StandardSpiller(MachineFunction *mf, LiveIntervals *lis, - const MachineLoopInfo *loopInfo, VirtRegMap *vrm) + StandardSpiller(LiveIntervals *lis, const MachineLoopInfo *loopInfo, + VirtRegMap *vrm) : lis(lis), loopInfo(loopInfo), vrm(vrm) {} /// Falls back on LiveIntervals::addIntervalsForSpills. std::vector<LiveInterval*> spill(LiveInterval *li, - SmallVectorImpl<LiveInterval*> &spillIs) { + SmallVectorImpl<LiveInterval*> &spillIs, + SlotIndex*) { return lis->addIntervalsForSpills(*li, spillIs, loopInfo, *vrm); } }; +/// When a call to spill is placed this spiller will first try to break the +/// interval up into its component values (one new interval per value). +/// If this fails, or if a call is placed to spill a previously split interval +/// then the spiller falls back on the standard spilling mechanism. +class SplittingSpiller : public StandardSpiller { +public: + SplittingSpiller(MachineFunction *mf, LiveIntervals *lis, + const MachineLoopInfo *loopInfo, VirtRegMap *vrm) + : StandardSpiller(lis, loopInfo, vrm) { + + mri = &mf->getRegInfo(); + tii = mf->getTarget().getInstrInfo(); + tri = mf->getTarget().getRegisterInfo(); + } + + std::vector<LiveInterval*> spill(LiveInterval *li, + SmallVectorImpl<LiveInterval*> &spillIs, + SlotIndex *earliestStart) { + + if (worthTryingToSplit(li)) { + return tryVNISplit(li, earliestStart); + } + // else + return StandardSpiller::spill(li, spillIs, earliestStart); + } + +private: + + MachineRegisterInfo *mri; + const TargetInstrInfo *tii; + const TargetRegisterInfo *tri; + DenseSet<LiveInterval*> alreadySplit; + + bool worthTryingToSplit(LiveInterval *li) const { + return (!alreadySplit.count(li) && li->getNumValNums() > 1); + } + + /// Try to break a LiveInterval into its component values. + std::vector<LiveInterval*> tryVNISplit(LiveInterval *li, + SlotIndex *earliestStart) { + + DEBUG(errs() << "Trying VNI split of %reg" << *li << "\n"); + + std::vector<LiveInterval*> added; + SmallVector<VNInfo*, 4> vnis; + + std::copy(li->vni_begin(), li->vni_end(), std::back_inserter(vnis)); + + for (SmallVectorImpl<VNInfo*>::iterator vniItr = vnis.begin(), + vniEnd = vnis.end(); vniItr != vniEnd; ++vniItr) { + VNInfo *vni = *vniItr; + + // Skip unused VNIs, or VNIs with no kills. + if (vni->isUnused() || vni->kills.empty()) + continue; + + DEBUG(errs() << " Extracted Val #" << vni->id << " as "); + LiveInterval *splitInterval = extractVNI(li, vni); + + if (splitInterval != 0) { + DEBUG(errs() << *splitInterval << "\n"); + added.push_back(splitInterval); + alreadySplit.insert(splitInterval); + if (earliestStart != 0) { + if (splitInterval->beginIndex() < *earliestStart) + *earliestStart = splitInterval->beginIndex(); + } + } else { + DEBUG(errs() << "0\n"); + } + } + + DEBUG(errs() << "Original LI: " << *li << "\n"); + + // If there original interval still contains some live ranges + // add it to added and alreadySplit. + if (!li->empty()) { + added.push_back(li); + alreadySplit.insert(li); + if (earliestStart != 0) { + if (li->beginIndex() < *earliestStart) + *earliestStart = li->beginIndex(); + } + } + + return added; + } + + /// Extract the given value number from the interval. + LiveInterval* extractVNI(LiveInterval *li, VNInfo *vni) const { + assert(vni->isDefAccurate() || vni->isPHIDef()); + assert(!vni->kills.empty()); + + // Create a new vreg and live interval, copy VNI kills & ranges over. + const TargetRegisterClass *trc = mri->getRegClass(li->reg); + unsigned newVReg = mri->createVirtualRegister(trc); + vrm->grow(); + LiveInterval *newLI = &lis->getOrCreateInterval(newVReg); + VNInfo *newVNI = newLI->createValueCopy(vni, lis->getVNInfoAllocator()); + + // Start by copying all live ranges in the VN to the new interval. + for (LiveInterval::iterator rItr = li->begin(), rEnd = li->end(); + rItr != rEnd; ++rItr) { + if (rItr->valno == vni) { + newLI->addRange(LiveRange(rItr->start, rItr->end, newVNI)); + } + } + + // Erase the old VNI & ranges. + li->removeValNo(vni); + + // Collect all current uses of the register belonging to the given VNI. + // We'll use this to rename the register after we've dealt with the def. + std::set<MachineInstr*> uses; + for (MachineRegisterInfo::use_iterator + useItr = mri->use_begin(li->reg), useEnd = mri->use_end(); + useItr != useEnd; ++useItr) { + uses.insert(&*useItr); + } + + // Process the def instruction for this VNI. + if (newVNI->isPHIDef()) { + // Insert a copy at the start of the MBB. The range proceeding the + // copy will be attached to the original LiveInterval. + MachineBasicBlock *defMBB = lis->getMBBFromIndex(newVNI->def); + tii->copyRegToReg(*defMBB, defMBB->begin(), newVReg, li->reg, trc, trc); + MachineInstr *copyMI = defMBB->begin(); + copyMI->addRegisterKilled(li->reg, tri); + SlotIndex copyIdx = lis->InsertMachineInstrInMaps(copyMI); + VNInfo *phiDefVNI = li->getNextValue(lis->getMBBStartIdx(defMBB), + 0, false, lis->getVNInfoAllocator()); + phiDefVNI->setIsPHIDef(true); + phiDefVNI->addKill(copyIdx.getDefIndex()); + li->addRange(LiveRange(phiDefVNI->def, copyIdx.getDefIndex(), phiDefVNI)); + LiveRange *oldPHIDefRange = + newLI->getLiveRangeContaining(lis->getMBBStartIdx(defMBB)); + + // If the old phi def starts in the middle of the range chop it up. + if (oldPHIDefRange->start < lis->getMBBStartIdx(defMBB)) { + LiveRange oldPHIDefRange2(copyIdx.getDefIndex(), oldPHIDefRange->end, + oldPHIDefRange->valno); + oldPHIDefRange->end = lis->getMBBStartIdx(defMBB); + newLI->addRange(oldPHIDefRange2); + } else if (oldPHIDefRange->start == lis->getMBBStartIdx(defMBB)) { + // Otherwise if it's at the start of the range just trim it. + oldPHIDefRange->start = copyIdx.getDefIndex(); + } else { + assert(false && "PHI def range doesn't cover PHI def?"); + } + + newVNI->def = copyIdx.getDefIndex(); + newVNI->setCopy(copyMI); + newVNI->setIsPHIDef(false); // not a PHI def anymore. + newVNI->setIsDefAccurate(true); + } else { + // non-PHI def. Rename the def. If it's two-addr that means renaming the use + // and inserting a new copy too. + MachineInstr *defInst = lis->getInstructionFromIndex(newVNI->def); + // We'll rename this now, so we can remove it from uses. + uses.erase(defInst); + unsigned defOpIdx = defInst->findRegisterDefOperandIdx(li->reg); + bool isTwoAddr = defInst->isRegTiedToUseOperand(defOpIdx), + twoAddrUseIsUndef = false; + + for (unsigned i = 0; i < defInst->getNumOperands(); ++i) { + MachineOperand &mo = defInst->getOperand(i); + if (mo.isReg() && (mo.isDef() || isTwoAddr) && (mo.getReg()==li->reg)) { + mo.setReg(newVReg); + if (isTwoAddr && mo.isUse() && mo.isUndef()) + twoAddrUseIsUndef = true; + } + } + + SlotIndex defIdx = lis->getInstructionIndex(defInst); + newVNI->def = defIdx.getDefIndex(); + + if (isTwoAddr && !twoAddrUseIsUndef) { + MachineBasicBlock *defMBB = defInst->getParent(); + tii->copyRegToReg(*defMBB, defInst, newVReg, li->reg, trc, trc); + MachineInstr *copyMI = prior(MachineBasicBlock::iterator(defInst)); + SlotIndex copyIdx = lis->InsertMachineInstrInMaps(copyMI); + copyMI->addRegisterKilled(li->reg, tri); + LiveRange *origUseRange = + li->getLiveRangeContaining(newVNI->def.getUseIndex()); + VNInfo *origUseVNI = origUseRange->valno; + origUseRange->end = copyIdx.getDefIndex(); + bool updatedKills = false; + for (unsigned k = 0; k < origUseVNI->kills.size(); ++k) { + if (origUseVNI->kills[k] == defIdx.getDefIndex()) { + origUseVNI->kills[k] = copyIdx.getDefIndex(); + updatedKills = true; + break; + } + } + assert(updatedKills && "Failed to update VNI kill list."); + VNInfo *copyVNI = newLI->getNextValue(copyIdx.getDefIndex(), copyMI, + true, lis->getVNInfoAllocator()); + copyVNI->addKill(defIdx.getDefIndex()); + LiveRange copyRange(copyIdx.getDefIndex(),defIdx.getDefIndex(),copyVNI); + newLI->addRange(copyRange); + } + } + + for (std::set<MachineInstr*>::iterator + usesItr = uses.begin(), usesEnd = uses.end(); + usesItr != usesEnd; ++usesItr) { + MachineInstr *useInst = *usesItr; + SlotIndex useIdx = lis->getInstructionIndex(useInst); + LiveRange *useRange = + newLI->getLiveRangeContaining(useIdx.getUseIndex()); + + // If this use doesn't belong to the new interval skip it. + if (useRange == 0) + continue; + + // This use doesn't belong to the VNI, skip it. + if (useRange->valno != newVNI) + continue; + + // Check if this instr is two address. + unsigned useOpIdx = useInst->findRegisterUseOperandIdx(li->reg); + bool isTwoAddress = useInst->isRegTiedToDefOperand(useOpIdx); + + // Rename uses (and defs for two-address instrs). + for (unsigned i = 0; i < useInst->getNumOperands(); ++i) { + MachineOperand &mo = useInst->getOperand(i); + if (mo.isReg() && (mo.isUse() || isTwoAddress) && + (mo.getReg() == li->reg)) { + mo.setReg(newVReg); + } + } + + // If this is a two address instruction we've got some extra work to do. + if (isTwoAddress) { + // We modified the def operand, so we need to copy back to the original + // reg. + MachineBasicBlock *useMBB = useInst->getParent(); + MachineBasicBlock::iterator useItr(useInst); + tii->copyRegToReg(*useMBB, next(useItr), li->reg, newVReg, trc, trc); + MachineInstr *copyMI = next(useItr); + copyMI->addRegisterKilled(newVReg, tri); + SlotIndex copyIdx = lis->InsertMachineInstrInMaps(copyMI); + + // Change the old two-address defined range & vni to start at + // (and be defined by) the copy. + LiveRange *origDefRange = + li->getLiveRangeContaining(useIdx.getDefIndex()); + origDefRange->start = copyIdx.getDefIndex(); + origDefRange->valno->def = copyIdx.getDefIndex(); + origDefRange->valno->setCopy(copyMI); + + // Insert a new range & vni for the two-address-to-copy value. This + // will be attached to the new live interval. + VNInfo *copyVNI = + newLI->getNextValue(useIdx.getDefIndex(), 0, true, + lis->getVNInfoAllocator()); + copyVNI->addKill(copyIdx.getDefIndex()); + LiveRange copyRange(useIdx.getDefIndex(),copyIdx.getDefIndex(),copyVNI); + newLI->addRange(copyRange); + } + } + + // Iterate over any PHI kills - we'll need to insert new copies for them. + for (VNInfo::KillSet::iterator + killItr = newVNI->kills.begin(), killEnd = newVNI->kills.end(); + killItr != killEnd; ++killItr) { + SlotIndex killIdx(*killItr); + if (killItr->isPHI()) { + MachineBasicBlock *killMBB = lis->getMBBFromIndex(killIdx); + LiveRange *oldKillRange = + newLI->getLiveRangeContaining(killIdx); + + assert(oldKillRange != 0 && "No kill range?"); + + tii->copyRegToReg(*killMBB, killMBB->getFirstTerminator(), + li->reg, newVReg, trc, trc); + MachineInstr *copyMI = prior(killMBB->getFirstTerminator()); + copyMI->addRegisterKilled(newVReg, tri); + SlotIndex copyIdx = lis->InsertMachineInstrInMaps(copyMI); + + // Save the current end. We may need it to add a new range if the + // current range runs of the end of the MBB. + SlotIndex newKillRangeEnd = oldKillRange->end; + oldKillRange->end = copyIdx.getDefIndex(); + + if (newKillRangeEnd != lis->getMBBEndIdx(killMBB).getNextSlot()) { + assert(newKillRangeEnd > lis->getMBBEndIdx(killMBB).getNextSlot() && + "PHI kill range doesn't reach kill-block end. Not sane."); + newLI->addRange(LiveRange(lis->getMBBEndIdx(killMBB).getNextSlot(), + newKillRangeEnd, newVNI)); + } + + *killItr = oldKillRange->end; + VNInfo *newKillVNI = li->getNextValue(copyIdx.getDefIndex(), + copyMI, true, + lis->getVNInfoAllocator()); + newKillVNI->addKill(lis->getMBBTerminatorGap(killMBB)); + newKillVNI->setHasPHIKill(true); + li->addRange(LiveRange(copyIdx.getDefIndex(), + lis->getMBBEndIdx(killMBB).getNextSlot(), + newKillVNI)); + } + + } + + newVNI->setHasPHIKill(false); + + return newLI; + } + +}; + } llvm::Spiller* llvm::createSpiller(MachineFunction *mf, LiveIntervals *lis, @@ -203,7 +520,8 @@ llvm::Spiller* llvm::createSpiller(MachineFunction *mf, LiveIntervals *lis, VirtRegMap *vrm) { switch (spillerOpt) { case trivial: return new TrivialSpiller(mf, lis, vrm); break; - case standard: return new StandardSpiller(mf, lis, loopInfo, vrm); break; + case standard: return new StandardSpiller(lis, loopInfo, vrm); break; + case splitting: return new SplittingSpiller(mf, lis, loopInfo, vrm); break; default: llvm_unreachable("Unreachable!"); break; } } diff --git a/lib/CodeGen/Spiller.h b/lib/CodeGen/Spiller.h index c6bd985..dda52e8 100644 --- a/lib/CodeGen/Spiller.h +++ b/lib/CodeGen/Spiller.h @@ -21,6 +21,7 @@ namespace llvm { class MachineFunction; class MachineInstr; class MachineLoopInfo; + class SlotIndex; class VirtRegMap; class VNInfo; @@ -35,7 +36,8 @@ namespace llvm { /// Spill the given live range. The method used will depend on the Spiller /// implementation selected. virtual std::vector<LiveInterval*> spill(LiveInterval *li, - SmallVectorImpl<LiveInterval*> &spillIs) = 0; + SmallVectorImpl<LiveInterval*> &spillIs, + SlotIndex *earliestIndex = 0) = 0; }; |