X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FCalcSpillWeights.cpp;h=76bb3d148b0b0ef27ea72a6b5275cec7b0c069ec;hb=8701768ae2e93e8741106acfa4a29959e1439487;hp=4c121c5749d4618e9d32e4f2a0d02539abf78f2b;hpb=518bb53485df640d7b7e3f6b0544099020c42aa7;p=oota-llvm.git diff --git a/lib/CodeGen/CalcSpillWeights.cpp b/lib/CodeGen/CalcSpillWeights.cpp index 4c121c5749d..76bb3d148b0 100644 --- a/lib/CodeGen/CalcSpillWeights.cpp +++ b/lib/CodeGen/CalcSpillWeights.cpp @@ -20,13 +20,17 @@ #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" #include "llvm/Target/TargetInstrInfo.h" +#include "llvm/Target/TargetMachine.h" #include "llvm/Target/TargetRegisterInfo.h" - using namespace llvm; char CalculateSpillWeights::ID = 0; -static RegisterPass X("calcspillweights", - "Calculate spill weights"); +INITIALIZE_PASS_BEGIN(CalculateSpillWeights, "calcspillweights", + "Calculate spill weights", false, false) +INITIALIZE_PASS_DEPENDENCY(LiveIntervals) +INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) +INITIALIZE_PASS_END(CalculateSpillWeights, "calcspillweights", + "Calculate spill weights", false, false) void CalculateSpillWeights::getAnalysisUsage(AnalysisUsage &au) const { au.addRequired(); @@ -41,111 +45,183 @@ bool CalculateSpillWeights::runOnMachineFunction(MachineFunction &fn) { << "********** Function: " << fn.getFunction()->getName() << '\n'); - LiveIntervals *lis = &getAnalysis(); - MachineLoopInfo *loopInfo = &getAnalysis(); - const TargetInstrInfo *tii = fn.getTarget().getInstrInfo(); - MachineRegisterInfo *mri = &fn.getRegInfo(); - - SmallSet processed; - for (MachineFunction::iterator mbbi = fn.begin(), mbbe = fn.end(); - mbbi != mbbe; ++mbbi) { - MachineBasicBlock* mbb = mbbi; - SlotIndex mbbEnd = lis->getMBBEndIdx(mbb); - MachineLoop* loop = loopInfo->getLoopFor(mbb); - unsigned loopDepth = loop ? loop->getLoopDepth() : 0; - bool isExiting = loop ? loop->isLoopExiting(mbb) : false; - - for (MachineBasicBlock::const_iterator mii = mbb->begin(), mie = mbb->end(); - mii != mie; ++mii) { - const MachineInstr *mi = mii; - if (tii->isIdentityCopy(*mi) || mi->isImplicitDef() || mi->isDebugValue()) - continue; - - for (unsigned i = 0, e = mi->getNumOperands(); i != e; ++i) { - const MachineOperand &mopi = mi->getOperand(i); - if (!mopi.isReg() || mopi.getReg() == 0) - continue; - unsigned reg = mopi.getReg(); - if (!TargetRegisterInfo::isVirtualRegister(mopi.getReg())) - continue; - // Multiple uses of reg by the same instruction. It should not - // contribute to spill weight again. - if (!processed.insert(reg)) - continue; - - bool hasDef = mopi.isDef(); - bool hasUse = !hasDef; - for (unsigned j = i+1; j != e; ++j) { - const MachineOperand &mopj = mi->getOperand(j); - if (!mopj.isReg() || mopj.getReg() != reg) - continue; - hasDef |= mopj.isDef(); - hasUse |= mopj.isUse(); - if (hasDef && hasUse) - break; - } - - LiveInterval ®Int = lis->getInterval(reg); - float weight = lis->getSpillWeight(hasDef, hasUse, loopDepth); - if (hasDef && isExiting) { - // Looks like this is a loop count variable update. - SlotIndex defIdx = lis->getInstructionIndex(mi).getDefIndex(); - const LiveRange *dlr = - lis->getInterval(reg).getLiveRangeContaining(defIdx); - if (dlr->end >= mbbEnd) - weight *= 3.0F; - } - regInt.weight += weight; - } - processed.clear(); - } + LiveIntervals &lis = getAnalysis(); + VirtRegAuxInfo vrai(fn, lis, getAnalysis()); + for (LiveIntervals::iterator I = lis.begin(), E = lis.end(); I != E; ++I) { + LiveInterval &li = *I->second; + if (TargetRegisterInfo::isVirtualRegister(li.reg)) + vrai.CalculateWeightAndHint(li); + } + return false; +} + +// Return the preferred allocation register for reg, given a COPY instruction. +static unsigned copyHint(const MachineInstr *mi, unsigned reg, + const TargetRegisterInfo &tri, + const MachineRegisterInfo &mri) { + unsigned sub, hreg, hsub; + if (mi->getOperand(0).getReg() == reg) { + sub = mi->getOperand(0).getSubReg(); + hreg = mi->getOperand(1).getReg(); + hsub = mi->getOperand(1).getSubReg(); + } else { + sub = mi->getOperand(1).getSubReg(); + hreg = mi->getOperand(0).getReg(); + hsub = mi->getOperand(0).getSubReg(); } - for (LiveIntervals::iterator I = lis->begin(), E = lis->end(); I != E; ++I) { - LiveInterval &li = *I->second; - if (TargetRegisterInfo::isVirtualRegister(li.reg)) { - // If the live interval length is essentially zero, i.e. in every live - // range the use follows def immediately, it doesn't make sense to spill - // it and hope it will be easier to allocate for this li. - if (isZeroLengthInterval(&li)) { - li.weight = HUGE_VALF; - continue; - } - - bool isLoad = false; - SmallVector spillIs; - if (lis->isReMaterializable(li, spillIs, isLoad)) { - // If all of the definitions of the interval are re-materializable, - // it is a preferred candidate for spilling. If non of the defs are - // loads, then it's potentially very cheap to re-materialize. - // FIXME: this gets much more complicated once we support non-trivial - // re-materialization. - if (isLoad) - li.weight *= 0.9F; - else - li.weight *= 0.5F; - } - - // Slightly prefer live interval that has been assigned a preferred reg. - std::pair Hint = mri->getRegAllocationHint(li.reg); - if (Hint.first || Hint.second) - li.weight *= 1.01F; - - // Divide the weight of the interval by its size. This encourages - // spilling of intervals that are large and have few uses, and - // discourages spilling of small intervals with many uses. - li.weight /= lis->getApproximateInstructionCount(li) * SlotIndex::NUM; + if (!hreg) + return 0; + + if (TargetRegisterInfo::isVirtualRegister(hreg)) + return sub == hsub ? hreg : 0; + + const TargetRegisterClass *rc = mri.getRegClass(reg); + + // Only allow physreg hints in rc. + if (sub == 0) + return rc->contains(hreg) ? hreg : 0; + + // reg:sub should match the physreg hreg. + return tri.getMatchingSuperReg(hreg, sub, rc); +} + +void VirtRegAuxInfo::CalculateWeightAndHint(LiveInterval &li) { + MachineRegisterInfo &mri = mf_.getRegInfo(); + const TargetRegisterInfo &tri = *mf_.getTarget().getRegisterInfo(); + MachineBasicBlock *mbb = 0; + MachineLoop *loop = 0; + unsigned loopDepth = 0; + bool isExiting = false; + float totalWeight = 0; + SmallPtrSet visited; + + // Find the best physreg hist and the best virtreg hint. + float bestPhys = 0, bestVirt = 0; + unsigned hintPhys = 0, hintVirt = 0; + + // Don't recompute a target specific hint. + bool noHint = mri.getRegAllocationHint(li.reg).first != 0; + + for (MachineRegisterInfo::reg_iterator I = mri.reg_begin(li.reg); + MachineInstr *mi = I.skipInstruction();) { + if (mi->isIdentityCopy() || mi->isImplicitDef() || mi->isDebugValue()) + continue; + if (!visited.insert(mi)) + continue; + + // Get loop info for mi. + if (mi->getParent() != mbb) { + mbb = mi->getParent(); + loop = loops_.getLoopFor(mbb); + loopDepth = loop ? loop->getLoopDepth() : 0; + isExiting = loop ? loop->isLoopExiting(mbb) : false; + } + + // Calculate instr weight. + bool reads, writes; + tie(reads, writes) = mi->readsWritesVirtualRegister(li.reg); + float weight = LiveIntervals::getSpillWeight(writes, reads, loopDepth); + + // Give extra weight to what looks like a loop induction variable update. + if (writes && isExiting && lis_.isLiveOutOfMBB(li, mbb)) + weight *= 3; + + totalWeight += weight; + + // Get allocation hints from copies. + if (noHint || !mi->isCopy()) + continue; + unsigned hint = copyHint(mi, li.reg, tri, mri); + if (!hint) + continue; + float hweight = hint_[hint] += weight; + if (TargetRegisterInfo::isPhysicalRegister(hint)) { + if (hweight > bestPhys && lis_.isAllocatable(hint)) + bestPhys = hweight, hintPhys = hint; + } else { + if (hweight > bestVirt) + bestVirt = hweight, hintVirt = hint; } } - - return false; + + hint_.clear(); + + // Always prefer the physreg hint. + if (unsigned hint = hintPhys ? hintPhys : hintVirt) { + mri.setRegAllocationHint(li.reg, 0, hint); + // Weakly boost the spill weifght of hinted registers. + totalWeight *= 1.01F; + } + + // Mark li as unspillable if all live ranges are tiny. + if (li.isZeroLength()) { + li.markNotSpillable(); + return; + } + + // If all of the definitions of the interval are re-materializable, + // it is a preferred candidate for spilling. If none of the defs are + // loads, then it's potentially very cheap to re-materialize. + // FIXME: this gets much more complicated once we support non-trivial + // re-materialization. + bool isLoad = false; + SmallVector spillIs; + if (lis_.isReMaterializable(li, spillIs, isLoad)) { + if (isLoad) + totalWeight *= 0.9F; + else + totalWeight *= 0.5F; + } + + li.weight = normalizeSpillWeight(totalWeight, li.getSize()); } -/// Returns true if the given live interval is zero length. -bool CalculateSpillWeights::isZeroLengthInterval(LiveInterval *li) const { - for (LiveInterval::Ranges::const_iterator - i = li->ranges.begin(), e = li->ranges.end(); i != e; ++i) - if (i->end.getPrevIndex() > i->start) - return false; - return true; +void VirtRegAuxInfo::CalculateRegClass(unsigned reg) { + MachineRegisterInfo &mri = mf_.getRegInfo(); + const TargetRegisterInfo *tri = mf_.getTarget().getRegisterInfo(); + const TargetRegisterClass *orc = mri.getRegClass(reg); + SmallPtrSet rcs; + + for (MachineRegisterInfo::reg_nodbg_iterator I = mri.reg_nodbg_begin(reg), + E = mri.reg_nodbg_end(); I != E; ++I) { + // The targets don't have accurate enough regclass descriptions that we can + // handle subregs. We need something similar to + // TRI::getMatchingSuperRegClass, but returning a super class instead of a + // sub class. + if (I.getOperand().getSubReg()) { + DEBUG(dbgs() << "Cannot handle subregs: " << I.getOperand() << '\n'); + return; + } + if (const TargetRegisterClass *rc = + I->getDesc().getRegClass(I.getOperandNo(), tri)) + rcs.insert(rc); + } + + // If we found no regclass constraints, just leave reg as is. + // In theory, we could inflate to the largest superclass of reg's existing + // class, but that might not be legal for the current cpu setting. + // This could happen if reg is only used by COPY instructions, so we may need + // to improve on this. + if (rcs.empty()) { + return; + } + + // Compute the intersection of all classes in rcs. + // This ought to be independent of iteration order, but if the target register + // classes don't form a proper algebra, it is possible to get different + // results. The solution is to make sure the intersection of any two register + // classes is also a register class or the null set. + const TargetRegisterClass *rc = 0; + for (SmallPtrSet::iterator I = rcs.begin(), + E = rcs.end(); I != E; ++I) { + rc = rc ? getCommonSubClass(rc, *I) : *I; + assert(rc && "Incompatible regclass constraints found"); + } + + if (rc == orc) + return; + DEBUG(dbgs() << "Inflating " << orc->getName() << ':' << PrintReg(reg) + << " to " << rc->getName() <<".\n"); + mri.setRegClass(reg, rc); }