//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "simpleregistercoalescing"
-#include "llvm/CodeGen/SimpleRegisterCoalescing.h"
-#include "llvm/CodeGen/LiveIntervalAnalysis.h"
+#define DEBUG_TYPE "regcoalescing"
+#include "SimpleRegisterCoalescing.h"
#include "VirtRegMap.h"
+#include "llvm/CodeGen/LiveIntervalAnalysis.h"
#include "llvm/Value.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/CodeGen/LiveVariables.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/CodeGen/SSARegMap.h"
+#include "llvm/CodeGen/RegisterCoalescer.h"
#include "llvm/Target/MRegisterInfo.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetMachine.h"
cl::desc("Coalesce copies (default=true)"),
cl::init(true));
+ static cl::opt<bool>
+ NewHeuristic("new-coalescer-heuristic",
+ cl::desc("Use new coalescer heuristic"),
+ cl::init(false));
+
RegisterPass<SimpleRegisterCoalescing>
- X("simple-register-coalescing",
- "Simple register coalescing to eliminate all possible register copies");
+ X("simple-register-coalescing", "Simple Register Coalescing");
+
+ // Declare that we implement the RegisterCoalescer interface
+ RegisterAnalysisGroup<RegisterCoalescer, true/*The Default*/> V(X);
}
const PassInfo *llvm::SimpleRegisterCoalescingID = X.getPassInfo();
void SimpleRegisterCoalescing::getAnalysisUsage(AnalysisUsage &AU) const {
- //AU.addPreserved<LiveVariables>();
AU.addPreserved<LiveIntervals>();
AU.addPreservedID(PHIEliminationID);
AU.addPreservedID(TwoAddressInstructionPassID);
// BValNo is a value number in B that is defined by a copy from A. 'B3' in
// the example above.
LiveInterval::iterator BLR = IntB.FindLiveRangeContaining(CopyIdx);
- unsigned BValNo = BLR->ValId;
+ VNInfo *BValNo = BLR->valno;
// Get the location that B is defined at. Two options: either this value has
// an unknown definition point or it is defined at CopyIdx. If unknown, we
// can't process it.
- unsigned BValNoDefIdx = IntB.getInstForValNum(BValNo);
- if (BValNoDefIdx == ~0U) return false;
- assert(BValNoDefIdx == CopyIdx &&
+ if (!BValNo->reg) return false;
+ assert(BValNo->def == CopyIdx &&
"Copy doesn't define the value?");
// AValNo is the value number in A that defines the copy, A0 in the example.
LiveInterval::iterator AValLR = IntA.FindLiveRangeContaining(CopyIdx-1);
- unsigned AValNo = AValLR->ValId;
+ VNInfo *AValNo = AValLR->valno;
// If AValNo is defined as a copy from IntB, we can potentially process this.
// Get the instruction that defines this value number.
- unsigned SrcReg = IntA.getSrcRegForValNum(AValNo);
+ unsigned SrcReg = AValNo->reg;
if (!SrcReg) return false; // Not defined by a copy.
// If the value number is not defined by a copy instruction, ignore it.
if (rep(SrcReg) != IntB.reg) return false;
// Get the LiveRange in IntB that this value number starts with.
- unsigned AValNoInstIdx = IntA.getInstForValNum(AValNo);
- LiveInterval::iterator ValLR = IntB.FindLiveRangeContaining(AValNoInstIdx-1);
+ LiveInterval::iterator ValLR = IntB.FindLiveRangeContaining(AValNo->def-1);
// Make sure that the end of the live range is inside the same block as
// CopyMI.
// live-range starts. If there are no intervening live ranges between them in
// IntB, we can merge them.
if (ValLR+1 != BLR) return false;
+
+ // If a live interval is a physical register, conservatively check if any
+ // of its sub-registers is overlapping the live interval of the virtual
+ // register. If so, do not coalesce.
+ if (MRegisterInfo::isPhysicalRegister(IntB.reg) &&
+ *mri_->getSubRegisters(IntB.reg)) {
+ for (const unsigned* SR = mri_->getSubRegisters(IntB.reg); *SR; ++SR)
+ if (li_->hasInterval(*SR) && IntA.overlaps(li_->getInterval(*SR))) {
+ DOUT << "Interfere with sub-register ";
+ DEBUG(li_->getInterval(*SR).print(DOUT, mri_));
+ return false;
+ }
+ }
DOUT << "\nExtending: "; IntB.print(DOUT, mri_);
+ unsigned FillerStart = ValLR->end, FillerEnd = BLR->start;
// We are about to delete CopyMI, so need to remove it as the 'instruction
- // that defines this value #'.
- IntB.setValueNumberInfo(BValNo, std::make_pair(~0U, 0));
+ // that defines this value #'. Update the the valnum with the new defining
+ // instruction #.
+ BValNo->def = FillerStart;
+ BValNo->reg = 0;
// Okay, we can merge them. We need to insert a new liverange:
// [ValLR.end, BLR.begin) of either value number, then we merge the
// two value numbers.
- unsigned FillerStart = ValLR->end, FillerEnd = BLR->start;
IntB.addRange(LiveRange(FillerStart, FillerEnd, BValNo));
// If the IntB live range is assigned to a physical register, and if that
for (const unsigned *AS = mri_->getSubRegisters(IntB.reg); *AS; ++AS) {
LiveInterval &AliasLI = li_->getInterval(*AS);
AliasLI.addRange(LiveRange(FillerStart, FillerEnd,
- AliasLI.getNextValue(~0U, 0)));
+ AliasLI.getNextValue(FillerStart, 0, li_->getVNInfoAllocator())));
}
}
// Okay, merge "B1" into the same value number as "B0".
- if (BValNo != ValLR->ValId)
- IntB.MergeValueNumberInto(BValNo, ValLR->ValId);
+ if (BValNo != ValLR->valno)
+ IntB.MergeValueNumberInto(BValNo, ValLR->valno);
DOUT << " result = "; IntB.print(DOUT, mri_);
DOUT << "\n";
if (UIdx != -1)
ValLREndInst->getOperand(UIdx).unsetIsKill();
- // Finally, delete the copy instruction.
- li_->RemoveMachineInstrFromMaps(CopyMI);
- CopyMI->eraseFromParent();
++numPeep;
return true;
}
+/// AddSubRegIdxPairs - Recursively mark all the registers represented by the
+/// specified register as sub-registers. The recursion level is expected to be
+/// shallow.
+void SimpleRegisterCoalescing::AddSubRegIdxPairs(unsigned Reg, unsigned SubIdx) {
+ std::vector<unsigned> &JoinedRegs = r2rRevMap_[Reg];
+ for (unsigned i = 0, e = JoinedRegs.size(); i != e; ++i) {
+ SubRegIdxes.push_back(std::make_pair(JoinedRegs[i], SubIdx));
+ AddSubRegIdxPairs(JoinedRegs[i], SubIdx);
+ }
+}
+
+/// isBackEdgeCopy - Returns true if CopyMI is a back edge copy.
+///
+bool SimpleRegisterCoalescing::isBackEdgeCopy(MachineInstr *CopyMI,
+ unsigned DstReg) {
+ MachineBasicBlock *MBB = CopyMI->getParent();
+ const BasicBlock *BB = MBB->getBasicBlock();
+ const Loop *L = loopInfo->getLoopFor(BB);
+ if (!L)
+ return false;
+ if (BB != L->getLoopLatch())
+ return false;
+
+ DstReg = rep(DstReg);
+ LiveInterval &LI = li_->getInterval(DstReg);
+ unsigned DefIdx = li_->getInstructionIndex(CopyMI);
+ LiveInterval::const_iterator DstLR =
+ LI.FindLiveRangeContaining(li_->getDefIndex(DefIdx));
+ if (DstLR == LI.end())
+ return false;
+ unsigned KillIdx = li_->getInstructionIndex(&MBB->back()) + InstrSlots::NUM-1;
+ if (DstLR->valno->kills.size() == 1 && DstLR->valno->kills[0] == KillIdx)
+ return true;
+ return false;
+}
+
/// JoinCopy - Attempt to join intervals corresponding to SrcReg/DstReg,
/// which are the src/dst of the copy instruction CopyMI. This returns true
-/// if the copy was successfully coalesced away, or if it is never possible
-/// to coalesce this copy, due to register constraints. It returns
-/// false if it is not currently possible to coalesce this interval, but
-/// it may be possible if other things get coalesced.
-bool SimpleRegisterCoalescing::JoinCopy(MachineInstr *CopyMI,
- unsigned SrcReg, unsigned DstReg, bool PhysOnly) {
+/// if the copy was successfully coalesced away. If it is not currently
+/// possible to coalesce this interval, but it may be possible if other
+/// things get coalesced, then it returns true by reference in 'Again'.
+bool SimpleRegisterCoalescing::JoinCopy(CopyRec TheCopy, bool &Again) {
+ MachineInstr *CopyMI = TheCopy.MI;
+
+ Again = false;
+ if (JoinedCopies.count(CopyMI))
+ return false; // Already done.
+
DOUT << li_->getInstructionIndex(CopyMI) << '\t' << *CopyMI;
// Get representative registers.
+ unsigned SrcReg = TheCopy.SrcReg;
+ unsigned DstReg = TheCopy.DstReg;
unsigned repSrcReg = rep(SrcReg);
unsigned repDstReg = rep(DstReg);
// If they are already joined we continue.
if (repSrcReg == repDstReg) {
DOUT << "\tCopy already coalesced.\n";
- return true; // Not coalescable.
+ return false; // Not coalescable.
}
bool SrcIsPhys = MRegisterInfo::isPhysicalRegister(repSrcReg);
bool DstIsPhys = MRegisterInfo::isPhysicalRegister(repDstReg);
- if (PhysOnly && !SrcIsPhys && !DstIsPhys)
- // Only joining physical registers with virtual registers in this round.
- return true;
// If they are both physical registers, we cannot join them.
if (SrcIsPhys && DstIsPhys) {
DOUT << "\tCan not coalesce physregs.\n";
- return true; // Not coalescable.
+ return false; // Not coalescable.
}
// We only join virtual registers with allocatable physical registers.
if (SrcIsPhys && !allocatableRegs_[repSrcReg]) {
DOUT << "\tSrc reg is unallocatable physreg.\n";
- return true; // Not coalescable.
+ return false; // Not coalescable.
}
if (DstIsPhys && !allocatableRegs_[repDstReg]) {
DOUT << "\tDst reg is unallocatable physreg.\n";
- return true; // Not coalescable.
+ return false; // Not coalescable.
}
-
- // If they are not of the same register class, we cannot join them.
- if (differingRegisterClasses(repSrcReg, repDstReg)) {
+
+ bool isExtSubReg = CopyMI->getOpcode() == TargetInstrInfo::EXTRACT_SUBREG;
+ unsigned RealDstReg = 0;
+ if (isExtSubReg) {
+ unsigned SubIdx = CopyMI->getOperand(2).getImm();
+ if (SrcIsPhys)
+ // r1024 = EXTRACT_SUBREG EAX, 0 then r1024 is really going to be
+ // coalesced with AX.
+ repSrcReg = mri_->getSubReg(repSrcReg, SubIdx);
+ else if (DstIsPhys) {
+ // If this is a extract_subreg where dst is a physical register, e.g.
+ // cl = EXTRACT_SUBREG reg1024, 1
+ // then create and update the actual physical register allocated to RHS.
+ const TargetRegisterClass *RC=mf_->getSSARegMap()->getRegClass(repSrcReg);
+ for (const unsigned *SRs = mri_->getSuperRegisters(repDstReg);
+ unsigned SR = *SRs; ++SRs) {
+ if (repDstReg == mri_->getSubReg(SR, SubIdx) &&
+ RC->contains(SR)) {
+ RealDstReg = SR;
+ break;
+ }
+ }
+ assert(RealDstReg && "Invalid extra_subreg instruction!");
+
+ // For this type of EXTRACT_SUBREG, conservatively
+ // check if the live interval of the source register interfere with the
+ // actual super physical register we are trying to coalesce with.
+ LiveInterval &RHS = li_->getInterval(repSrcReg);
+ if (li_->hasInterval(RealDstReg) &&
+ RHS.overlaps(li_->getInterval(RealDstReg))) {
+ DOUT << "Interfere with register ";
+ DEBUG(li_->getInterval(RealDstReg).print(DOUT, mri_));
+ return false; // Not coalescable
+ }
+ for (const unsigned* SR = mri_->getSubRegisters(RealDstReg); *SR; ++SR)
+ if (li_->hasInterval(*SR) && RHS.overlaps(li_->getInterval(*SR))) {
+ DOUT << "Interfere with sub-register ";
+ DEBUG(li_->getInterval(*SR).print(DOUT, mri_));
+ return false; // Not coalescable
+ }
+ } else {
+ unsigned SrcSize= li_->getInterval(repSrcReg).getSize() / InstrSlots::NUM;
+ unsigned DstSize= li_->getInterval(repDstReg).getSize() / InstrSlots::NUM;
+ const TargetRegisterClass *RC=mf_->getSSARegMap()->getRegClass(repDstReg);
+ unsigned Threshold = allocatableRCRegs_[RC].count();
+ // Be conservative. If both sides are virtual registers, do not coalesce
+ // if this will cause a high use density interval to target a smaller set
+ // of registers.
+ if (DstSize > Threshold || SrcSize > Threshold) {
+ LiveVariables::VarInfo &svi = lv_->getVarInfo(repSrcReg);
+ LiveVariables::VarInfo &dvi = lv_->getVarInfo(repDstReg);
+ if ((float)dvi.NumUses / DstSize < (float)svi.NumUses / SrcSize) {
+ Again = true; // May be possible to coalesce later.
+ return false;
+ }
+ }
+ }
+ } else if (differingRegisterClasses(repSrcReg, repDstReg)) {
+ // If they are not of the same register class, we cannot join them.
DOUT << "\tSrc/Dest are different register classes.\n";
- return true; // Not coalescable.
+ // Allow the coalescer to try again in case either side gets coalesced to
+ // a physical register that's compatible with the other side. e.g.
+ // r1024 = MOV32to32_ r1025
+ // but later r1024 is assigned EAX then r1025 may be coalesced with EAX.
+ Again = true; // May be possible to coalesce later.
+ return false;
}
LiveInterval &SrcInt = li_->getInterval(repSrcReg);
// virtual register. Once the coalescing is done, it cannot be broken and
// these are not spillable! If the destination interval uses are far away,
// think twice about coalescing them!
- if (!mopd->isDead() && (SrcIsPhys || DstIsPhys)) {
+ if (!mopd->isDead() && (SrcIsPhys || DstIsPhys) && !isExtSubReg) {
LiveInterval &JoinVInt = SrcIsPhys ? DstInt : SrcInt;
unsigned JoinVReg = SrcIsPhys ? repDstReg : repSrcReg;
unsigned JoinPReg = SrcIsPhys ? repSrcReg : repDstReg;
const TargetRegisterClass *RC = mf_->getSSARegMap()->getRegClass(JoinVReg);
unsigned Threshold = allocatableRCRegs_[RC].count();
+ if (TheCopy.isBackEdge)
+ Threshold *= 2; // Favors back edge copies.
- // If the virtual register live interval is long has it has low use desity,
+ // If the virtual register live interval is long but it has low use desity,
// do not join them, instead mark the physical register as its allocation
// preference.
unsigned Length = JoinVInt.getSize() / InstrSlots::NUM;
JoinVInt.preference = JoinPReg;
++numAborts;
DOUT << "\tMay tie down a physical register, abort!\n";
+ Again = true; // May be possible to coalesce later.
return false;
}
}
// Otherwise, if one of the intervals being joined is a physreg, this method
// always canonicalizes DstInt to be it. The output "SrcInt" will not have
// been modified, so we can use this information below to update aliases.
- if (JoinIntervals(DstInt, SrcInt)) {
+ bool Swapped = false;
+ if (JoinIntervals(DstInt, SrcInt, Swapped)) {
if (isDead) {
// Result of the copy is dead. Propagate this property.
if (SrcStart == 0) {
}
if (isShorten || isDead) {
- // Shorten the live interval.
- LiveInterval &LiveInInt = (repSrcReg == DstInt.reg) ? DstInt : SrcInt;
- LiveInInt.removeRange(RemoveStart, RemoveEnd);
+ // Shorten the destination live interval.
+ if (Swapped)
+ SrcInt.removeRange(RemoveStart, RemoveEnd);
}
} else {
// Coalescing failed.
// If we can eliminate the copy without merging the live ranges, do so now.
- if (AdjustCopiesBackFrom(SrcInt, DstInt, CopyMI))
+ if (!isExtSubReg && AdjustCopiesBackFrom(SrcInt, DstInt, CopyMI)) {
+ JoinedCopies.insert(CopyMI);
return true;
+ }
// Otherwise, we are unable to join the intervals.
DOUT << "Interference!\n";
+ Again = true; // May be possible to coalesce later.
return false;
}
- bool Swapped = repSrcReg == DstInt.reg;
- if (Swapped)
+ LiveInterval *ResSrcInt = &SrcInt;
+ LiveInterval *ResDstInt = &DstInt;
+ if (Swapped) {
std::swap(repSrcReg, repDstReg);
+ std::swap(ResSrcInt, ResDstInt);
+ }
assert(MRegisterInfo::isVirtualRegister(repSrcReg) &&
"LiveInterval::join didn't work right!");
// have clobbered values for this range.
if (MRegisterInfo::isPhysicalRegister(repDstReg)) {
// Unset unnecessary kills.
- if (!DstInt.containsOneValue()) {
- for (LiveInterval::Ranges::const_iterator I = SrcInt.begin(),
- E = SrcInt.end(); I != E; ++I)
+ if (!ResDstInt->containsOneValue()) {
+ for (LiveInterval::Ranges::const_iterator I = ResSrcInt->begin(),
+ E = ResSrcInt->end(); I != E; ++I)
unsetRegisterKills(I->start, I->end, repDstReg);
}
+ // If this is a extract_subreg where dst is a physical register, e.g.
+ // cl = EXTRACT_SUBREG reg1024, 1
+ // then create and update the actual physical register allocated to RHS.
+ if (RealDstReg) {
+ LiveInterval &RealDstInt = li_->getOrCreateInterval(RealDstReg);
+ SmallSet<const VNInfo*, 4> CopiedValNos;
+ for (LiveInterval::Ranges::const_iterator I = ResSrcInt->ranges.begin(),
+ E = ResSrcInt->ranges.end(); I != E; ++I) {
+ LiveInterval::const_iterator DstLR =
+ ResDstInt->FindLiveRangeContaining(I->start);
+ assert(DstLR != ResDstInt->end() && "Invalid joined interval!");
+ const VNInfo *DstValNo = DstLR->valno;
+ if (CopiedValNos.insert(DstValNo)) {
+ VNInfo *ValNo = RealDstInt.getNextValue(DstValNo->def, DstValNo->reg,
+ li_->getVNInfoAllocator());
+ RealDstInt.addKills(ValNo, DstValNo->kills);
+ RealDstInt.MergeValueInAsValue(*ResDstInt, DstValNo, ValNo);
+ }
+ }
+ repDstReg = RealDstReg;
+ }
+
// Update the liveintervals of sub-registers.
for (const unsigned *AS = mri_->getSubRegisters(repDstReg); *AS; ++AS)
- li_->getInterval(*AS).MergeInClobberRanges(SrcInt);
+ li_->getOrCreateInterval(*AS).MergeInClobberRanges(*ResSrcInt,
+ li_->getVNInfoAllocator());
} else {
// Merge use info if the destination is a virtual register.
LiveVariables::VarInfo& dVI = lv_->getVarInfo(repDstReg);
dVI.NumUses += sVI.NumUses;
}
- DOUT << "\n\t\tJoined. Result = "; DstInt.print(DOUT, mri_);
- DOUT << "\n";
-
// Remember these liveintervals have been joined.
JoinedLIs.set(repSrcReg - MRegisterInfo::FirstVirtualRegister);
if (MRegisterInfo::isVirtualRegister(repDstReg))
JoinedLIs.set(repDstReg - MRegisterInfo::FirstVirtualRegister);
- // If the intervals were swapped by Join, swap them back so that the register
- // mapping (in the r2i map) is correct.
- if (Swapped) SrcInt.swap(DstInt);
+ if (isExtSubReg && !SrcIsPhys && !DstIsPhys) {
+ if (!Swapped) {
+ // Make sure we allocate the larger super-register.
+ ResSrcInt->Copy(*ResDstInt, li_->getVNInfoAllocator());
+ std::swap(repSrcReg, repDstReg);
+ std::swap(ResSrcInt, ResDstInt);
+ }
+ unsigned SubIdx = CopyMI->getOperand(2).getImm();
+ SubRegIdxes.push_back(std::make_pair(repSrcReg, SubIdx));
+ AddSubRegIdxPairs(repSrcReg, SubIdx);
+ }
+
+ if (NewHeuristic) {
+ for (LiveInterval::const_vni_iterator i = ResSrcInt->vni_begin(),
+ e = ResSrcInt->vni_end(); i != e; ++i) {
+ const VNInfo *vni = *i;
+ if (vni->def && vni->def != ~1U && vni->def != ~0U) {
+ MachineInstr *CopyMI = li_->getInstructionFromIndex(vni->def);
+ unsigned SrcReg, DstReg;
+ if (CopyMI && tii_->isMoveInstr(*CopyMI, SrcReg, DstReg) &&
+ JoinedCopies.count(CopyMI) == 0) {
+ unsigned LoopDepth =
+ loopInfo->getLoopDepth(CopyMI->getParent()->getBasicBlock());
+ JoinQueue->push(CopyRec(CopyMI, SrcReg, DstReg, LoopDepth,
+ isBackEdgeCopy(CopyMI, DstReg)));
+ }
+ }
+ }
+ }
+
+ DOUT << "\n\t\tJoined. Result = "; ResDstInt->print(DOUT, mri_);
+ DOUT << "\n";
// repSrcReg is guarateed to be the register whose live interval that is
// being merged.
li_->removeInterval(repSrcReg);
r2rMap_[repSrcReg] = repDstReg;
+ r2rRevMap_[repDstReg].push_back(repSrcReg);
// Finally, delete the copy instruction.
- li_->RemoveMachineInstrFromMaps(CopyMI);
- CopyMI->eraseFromParent();
+ JoinedCopies.insert(CopyMI);
++numPeep;
++numJoins;
return true;
/// ThisFromOther[x] - If x is defined as a copy from the other interval, this
/// contains the value number the copy is from.
///
-static unsigned ComputeUltimateVN(unsigned VN,
- SmallVector<std::pair<unsigned,
- unsigned>, 16> &ValueNumberInfo,
- SmallVector<int, 16> &ThisFromOther,
- SmallVector<int, 16> &OtherFromThis,
+static unsigned ComputeUltimateVN(VNInfo *VNI,
+ SmallVector<VNInfo*, 16> &NewVNInfo,
+ DenseMap<VNInfo*, VNInfo*> &ThisFromOther,
+ DenseMap<VNInfo*, VNInfo*> &OtherFromThis,
SmallVector<int, 16> &ThisValNoAssignments,
- SmallVector<int, 16> &OtherValNoAssignments,
- LiveInterval &ThisLI, LiveInterval &OtherLI) {
+ SmallVector<int, 16> &OtherValNoAssignments) {
+ unsigned VN = VNI->id;
+
// If the VN has already been computed, just return it.
if (ThisValNoAssignments[VN] >= 0)
return ThisValNoAssignments[VN];
// assert(ThisValNoAssignments[VN] != -2 && "Cyclic case?");
-
+
// If this val is not a copy from the other val, then it must be a new value
// number in the destination.
- int OtherValNo = ThisFromOther[VN];
- if (OtherValNo == -1) {
- ValueNumberInfo.push_back(ThisLI.getValNumInfo(VN));
- return ThisValNoAssignments[VN] = ValueNumberInfo.size()-1;
+ DenseMap<VNInfo*, VNInfo*>::iterator I = ThisFromOther.find(VNI);
+ if (I == ThisFromOther.end()) {
+ NewVNInfo.push_back(VNI);
+ return ThisValNoAssignments[VN] = NewVNInfo.size()-1;
}
+ VNInfo *OtherValNo = I->second;
// Otherwise, this *is* a copy from the RHS. If the other side has already
// been computed, return it.
- if (OtherValNoAssignments[OtherValNo] >= 0)
- return ThisValNoAssignments[VN] = OtherValNoAssignments[OtherValNo];
+ if (OtherValNoAssignments[OtherValNo->id] >= 0)
+ return ThisValNoAssignments[VN] = OtherValNoAssignments[OtherValNo->id];
// Mark this value number as currently being computed, then ask what the
// ultimate value # of the other value is.
ThisValNoAssignments[VN] = -2;
unsigned UltimateVN =
- ComputeUltimateVN(OtherValNo, ValueNumberInfo,
- OtherFromThis, ThisFromOther,
- OtherValNoAssignments, ThisValNoAssignments,
- OtherLI, ThisLI);
+ ComputeUltimateVN(OtherValNo, NewVNInfo, OtherFromThis, ThisFromOther,
+ OtherValNoAssignments, ThisValNoAssignments);
return ThisValNoAssignments[VN] = UltimateVN;
}
-static bool InVector(unsigned Val, const SmallVector<unsigned, 8> &V) {
+static bool InVector(VNInfo *Val, const SmallVector<VNInfo*, 8> &V) {
return std::find(V.begin(), V.end(), Val) != V.end();
}
if (RHSIt != RHS.begin()) --RHSIt;
}
- SmallVector<unsigned, 8> EliminatedLHSVals;
+ SmallVector<VNInfo*, 8> EliminatedLHSVals;
while (1) {
// Determine if these live intervals overlap.
// coalesce these live ranges and we bail out.
if (Overlaps) {
// If we haven't already recorded that this value # is safe, check it.
- if (!InVector(LHSIt->ValId, EliminatedLHSVals)) {
+ if (!InVector(LHSIt->valno, EliminatedLHSVals)) {
// Copy from the RHS?
- unsigned SrcReg = LHS.getSrcRegForValNum(LHSIt->ValId);
+ unsigned SrcReg = LHSIt->valno->reg;
if (rep(SrcReg) != RHS.reg)
return false; // Nope, bail out.
- EliminatedLHSVals.push_back(LHSIt->ValId);
+ EliminatedLHSVals.push_back(LHSIt->valno);
}
// We know this entire LHS live range is okay, so skip it now.
// want to notice this copy (so that it gets coalesced away) even though
// the live ranges don't actually overlap.
if (LHSIt->start == RHSIt->end) {
- if (InVector(LHSIt->ValId, EliminatedLHSVals)) {
+ if (InVector(LHSIt->valno, EliminatedLHSVals)) {
// We already know that this value number is going to be merged in
// if coalescing succeeds. Just skip the liverange.
if (++LHSIt == LHSEnd) break;
} else {
// Otherwise, if this is a copy from the RHS, mark it as being merged
// in.
- if (rep(LHS.getSrcRegForValNum(LHSIt->ValId)) == RHS.reg) {
- EliminatedLHSVals.push_back(LHSIt->ValId);
+ if (rep(LHSIt->valno->reg) == RHS.reg) {
+ EliminatedLHSVals.push_back(LHSIt->valno);
// We know this entire LHS live range is okay, so skip it now.
if (++LHSIt == LHSEnd) break;
// optimize for it: if there is more than one value, we merge them all into
// the lowest numbered one, then handle the interval as if we were merging
// with one value number.
- unsigned LHSValNo;
+ VNInfo *LHSValNo;
if (EliminatedLHSVals.size() > 1) {
// Loop through all the equal value numbers merging them into the smallest
// one.
- unsigned Smallest = EliminatedLHSVals[0];
+ VNInfo *Smallest = EliminatedLHSVals[0];
for (unsigned i = 1, e = EliminatedLHSVals.size(); i != e; ++i) {
- if (EliminatedLHSVals[i] < Smallest) {
+ if (EliminatedLHSVals[i]->id < Smallest->id) {
// Merge the current notion of the smallest into the smaller one.
LHS.MergeValueNumberInto(Smallest, EliminatedLHSVals[i]);
Smallest = EliminatedLHSVals[i];
// Okay, now that there is a single LHS value number that we're merging the
// RHS into, update the value number info for the LHS to indicate that the
// value number is defined where the RHS value number was.
- LHS.setValueNumberInfo(LHSValNo, RHS.getValNumInfo(0));
+ const VNInfo *VNI = RHS.getValNumInfo(0);
+ LHSValNo->def = VNI->def;
+ LHSValNo->reg = VNI->reg;
// Okay, the final step is to loop over the RHS live intervals, adding them to
// the LHS.
+ LHS.addKills(LHSValNo, VNI->kills);
LHS.MergeRangesInAsValue(RHS, LHSValNo);
LHS.weight += RHS.weight;
if (RHS.preference && !LHS.preference)
/// physreg, this method always canonicalizes LHS to be it. The output
/// "RHS" will not have been modified, so we can use this information
/// below to update aliases.
-bool SimpleRegisterCoalescing::JoinIntervals(LiveInterval &LHS, LiveInterval &RHS) {
+bool SimpleRegisterCoalescing::JoinIntervals(LiveInterval &LHS,
+ LiveInterval &RHS, bool &Swapped) {
// Compute the final value assignment, assuming that the live ranges can be
// coalesced.
SmallVector<int, 16> LHSValNoAssignments;
SmallVector<int, 16> RHSValNoAssignments;
- SmallVector<std::pair<unsigned,unsigned>, 16> ValueNumberInfo;
+ DenseMap<VNInfo*, VNInfo*> LHSValsDefinedFromRHS;
+ DenseMap<VNInfo*, VNInfo*> RHSValsDefinedFromLHS;
+ SmallVector<VNInfo*, 16> NewVNInfo;
// If a live interval is a physical register, conservatively check if any
// of its sub-registers is overlapping the live interval of the virtual
// often RHS is small and LHS is large (e.g. a physreg).
// Find out if the RHS is defined as a copy from some value in the LHS.
+ int RHSVal0DefinedFromLHS = -1;
int RHSValID = -1;
- std::pair<unsigned,unsigned> RHSValNoInfo;
- unsigned RHSSrcReg = RHS.getSrcRegForValNum(0);
+ VNInfo *RHSValNoInfo = NULL;
+ VNInfo *RHSValNoInfo0 = RHS.getValNumInfo(0);
+ unsigned RHSSrcReg = RHSValNoInfo0->reg;
if ((RHSSrcReg == 0 || rep(RHSSrcReg) != LHS.reg)) {
// If RHS is not defined as a copy from the LHS, we can use simpler and
// faster checks to see if the live ranges are coalescable. This joiner
if (!MRegisterInfo::isPhysicalRegister(RHS.reg)) {
return SimpleJoin(LHS, RHS);
} else {
- RHSValNoInfo = RHS.getValNumInfo(0);
+ RHSValNoInfo = RHSValNoInfo0;
}
} else {
// It was defined as a copy from the LHS, find out what value # it is.
- unsigned ValInst = RHS.getInstForValNum(0);
- RHSValID = LHS.getLiveRangeContaining(ValInst-1)->ValId;
- RHSValNoInfo = LHS.getValNumInfo(RHSValID);
+ RHSValNoInfo = LHS.getLiveRangeContaining(RHSValNoInfo0->def-1)->valno;
+ RHSValID = RHSValNoInfo->id;
+ RHSVal0DefinedFromLHS = RHSValID;
}
LHSValNoAssignments.resize(LHS.getNumValNums(), -1);
RHSValNoAssignments.resize(RHS.getNumValNums(), -1);
- ValueNumberInfo.resize(LHS.getNumValNums());
+ NewVNInfo.resize(LHS.getNumValNums(), NULL);
// Okay, *all* of the values in LHS that are defined as a copy from RHS
// should now get updated.
- for (unsigned VN = 0, e = LHS.getNumValNums(); VN != e; ++VN) {
- if (unsigned LHSSrcReg = LHS.getSrcRegForValNum(VN)) {
+ for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
+ i != e; ++i) {
+ VNInfo *VNI = *i;
+ unsigned VN = VNI->id;
+ if (unsigned LHSSrcReg = VNI->reg) {
if (rep(LHSSrcReg) != RHS.reg) {
// If this is not a copy from the RHS, its value number will be
// unmodified by the coalescing.
- ValueNumberInfo[VN] = LHS.getValNumInfo(VN);
+ NewVNInfo[VN] = VNI;
LHSValNoAssignments[VN] = VN;
} else if (RHSValID == -1) {
// Otherwise, it is a copy from the RHS, and we don't already have a
// value# for it. Keep the current value number, but remember it.
LHSValNoAssignments[VN] = RHSValID = VN;
- ValueNumberInfo[VN] = RHSValNoInfo;
+ NewVNInfo[VN] = RHSValNoInfo;
+ LHSValsDefinedFromRHS[VNI] = RHSValNoInfo0;
} else {
// Otherwise, use the specified value #.
LHSValNoAssignments[VN] = RHSValID;
- if (VN != (unsigned)RHSValID)
- ValueNumberInfo[VN].first = ~1U;
- else
- ValueNumberInfo[VN] = RHSValNoInfo;
+ if (VN == (unsigned)RHSValID) { // Else this val# is dead.
+ NewVNInfo[VN] = RHSValNoInfo;
+ LHSValsDefinedFromRHS[VNI] = RHSValNoInfo0;
+ }
}
} else {
- ValueNumberInfo[VN] = LHS.getValNumInfo(VN);
+ NewVNInfo[VN] = VNI;
LHSValNoAssignments[VN] = VN;
}
}
assert(RHSValID != -1 && "Didn't find value #?");
RHSValNoAssignments[0] = RHSValID;
-
+ if (RHSVal0DefinedFromLHS != -1) {
+ // This path doesn't go through ComputeUltimateVN so just set
+ // it to anything.
+ RHSValsDefinedFromLHS[RHSValNoInfo0] = (VNInfo*)1;
+ }
} else {
// Loop over the value numbers of the LHS, seeing if any are defined from
// the RHS.
- SmallVector<int, 16> LHSValsDefinedFromRHS;
- LHSValsDefinedFromRHS.resize(LHS.getNumValNums(), -1);
- for (unsigned VN = 0, e = LHS.getNumValNums(); VN != e; ++VN) {
- unsigned ValSrcReg = LHS.getSrcRegForValNum(VN);
- if (ValSrcReg == 0) // Src not defined by a copy?
+ for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
+ i != e; ++i) {
+ VNInfo *VNI = *i;
+ unsigned ValSrcReg = VNI->reg;
+ if (VNI->def == ~1U ||ValSrcReg == 0) // Src not defined by a copy?
continue;
// DstReg is known to be a register in the LHS interval. If the src is
continue;
// Figure out the value # from the RHS.
- unsigned ValInst = LHS.getInstForValNum(VN);
- LHSValsDefinedFromRHS[VN] = RHS.getLiveRangeContaining(ValInst-1)->ValId;
+ LHSValsDefinedFromRHS[VNI] = RHS.getLiveRangeContaining(VNI->def-1)->valno;
}
// Loop over the value numbers of the RHS, seeing if any are defined from
// the LHS.
- SmallVector<int, 16> RHSValsDefinedFromLHS;
- RHSValsDefinedFromLHS.resize(RHS.getNumValNums(), -1);
- for (unsigned VN = 0, e = RHS.getNumValNums(); VN != e; ++VN) {
- unsigned ValSrcReg = RHS.getSrcRegForValNum(VN);
- if (ValSrcReg == 0) // Src not defined by a copy?
+ for (LiveInterval::vni_iterator i = RHS.vni_begin(), e = RHS.vni_end();
+ i != e; ++i) {
+ VNInfo *VNI = *i;
+ unsigned ValSrcReg = VNI->reg;
+ if (VNI->def == ~1U || ValSrcReg == 0) // Src not defined by a copy?
continue;
// DstReg is known to be a register in the RHS interval. If the src is
continue;
// Figure out the value # from the LHS.
- unsigned ValInst = RHS.getInstForValNum(VN);
- RHSValsDefinedFromLHS[VN] = LHS.getLiveRangeContaining(ValInst-1)->ValId;
+ RHSValsDefinedFromLHS[VNI]= LHS.getLiveRangeContaining(VNI->def-1)->valno;
}
LHSValNoAssignments.resize(LHS.getNumValNums(), -1);
RHSValNoAssignments.resize(RHS.getNumValNums(), -1);
- ValueNumberInfo.reserve(LHS.getNumValNums() + RHS.getNumValNums());
+ NewVNInfo.reserve(LHS.getNumValNums() + RHS.getNumValNums());
- for (unsigned VN = 0, e = LHS.getNumValNums(); VN != e; ++VN) {
- if (LHSValNoAssignments[VN] >= 0 || LHS.getInstForValNum(VN) == ~2U)
+ for (LiveInterval::vni_iterator i = LHS.vni_begin(), e = LHS.vni_end();
+ i != e; ++i) {
+ VNInfo *VNI = *i;
+ unsigned VN = VNI->id;
+ if (LHSValNoAssignments[VN] >= 0 || VNI->def == ~1U)
continue;
- ComputeUltimateVN(VN, ValueNumberInfo,
+ ComputeUltimateVN(VNI, NewVNInfo,
LHSValsDefinedFromRHS, RHSValsDefinedFromLHS,
- LHSValNoAssignments, RHSValNoAssignments, LHS, RHS);
+ LHSValNoAssignments, RHSValNoAssignments);
}
- for (unsigned VN = 0, e = RHS.getNumValNums(); VN != e; ++VN) {
- if (RHSValNoAssignments[VN] >= 0 || RHS.getInstForValNum(VN) == ~2U)
+ for (LiveInterval::vni_iterator i = RHS.vni_begin(), e = RHS.vni_end();
+ i != e; ++i) {
+ VNInfo *VNI = *i;
+ unsigned VN = VNI->id;
+ if (RHSValNoAssignments[VN] >= 0 || VNI->def == ~1U)
continue;
// If this value number isn't a copy from the LHS, it's a new number.
- if (RHSValsDefinedFromLHS[VN] == -1) {
- ValueNumberInfo.push_back(RHS.getValNumInfo(VN));
- RHSValNoAssignments[VN] = ValueNumberInfo.size()-1;
+ if (RHSValsDefinedFromLHS.find(VNI) == RHSValsDefinedFromLHS.end()) {
+ NewVNInfo.push_back(VNI);
+ RHSValNoAssignments[VN] = NewVNInfo.size()-1;
continue;
}
- ComputeUltimateVN(VN, ValueNumberInfo,
+ ComputeUltimateVN(VNI, NewVNInfo,
RHSValsDefinedFromLHS, LHSValsDefinedFromRHS,
- RHSValNoAssignments, LHSValNoAssignments, RHS, LHS);
+ RHSValNoAssignments, LHSValNoAssignments);
}
}
if (Overlaps) {
// If the live range overlap will map to the same value number in the
// result liverange, we can still coalesce them. If not, we can't.
- if (LHSValNoAssignments[I->ValId] != RHSValNoAssignments[J->ValId])
+ if (LHSValNoAssignments[I->valno->id] !=
+ RHSValNoAssignments[J->valno->id])
return false;
}
}
}
+ // Update kill info. Some live ranges are extended due to copy coalescing.
+ for (DenseMap<VNInfo*, VNInfo*>::iterator I = LHSValsDefinedFromRHS.begin(),
+ E = LHSValsDefinedFromRHS.end(); I != E; ++I) {
+ VNInfo *VNI = I->first;
+ unsigned LHSValID = LHSValNoAssignments[VNI->id];
+ LiveInterval::removeKill(NewVNInfo[LHSValID], VNI->def);
+ RHS.addKills(NewVNInfo[LHSValID], VNI->kills);
+ }
+
+ // Update kill info. Some live ranges are extended due to copy coalescing.
+ for (DenseMap<VNInfo*, VNInfo*>::iterator I = RHSValsDefinedFromLHS.begin(),
+ E = RHSValsDefinedFromLHS.end(); I != E; ++I) {
+ VNInfo *VNI = I->first;
+ unsigned RHSValID = RHSValNoAssignments[VNI->id];
+ LiveInterval::removeKill(NewVNInfo[RHSValID], VNI->def);
+ LHS.addKills(NewVNInfo[RHSValID], VNI->kills);
+ }
+
// If we get here, we know that we can coalesce the live ranges. Ask the
// intervals to coalesce themselves now.
- LHS.join(RHS, &LHSValNoAssignments[0], &RHSValNoAssignments[0],
- ValueNumberInfo);
+ if ((RHS.ranges.size() > LHS.ranges.size() &&
+ MRegisterInfo::isVirtualRegister(LHS.reg)) ||
+ MRegisterInfo::isPhysicalRegister(RHS.reg)) {
+ RHS.join(LHS, &RHSValNoAssignments[0], &LHSValNoAssignments[0], NewVNInfo);
+ Swapped = true;
+ } else {
+ LHS.join(RHS, &LHSValNoAssignments[0], &RHSValNoAssignments[0], NewVNInfo);
+ Swapped = false;
+ }
return true;
}
};
}
+/// getRepIntervalSize - Returns the size of the interval that represents the
+/// specified register.
+template<class SF>
+unsigned JoinPriorityQueue<SF>::getRepIntervalSize(unsigned Reg) {
+ return Rc->getRepIntervalSize(Reg);
+}
+
+/// CopyRecSort::operator - Join priority queue sorting function.
+///
+bool CopyRecSort::operator()(CopyRec left, CopyRec right) const {
+ // Inner loops first.
+ if (left.LoopDepth > right.LoopDepth)
+ return false;
+ else if (left.LoopDepth == right.LoopDepth) {
+ if (left.isBackEdge && !right.isBackEdge)
+ return false;
+ else if (left.isBackEdge == right.isBackEdge) {
+ // Join virtuals to physical registers first.
+ bool LDstIsPhys = MRegisterInfo::isPhysicalRegister(left.DstReg);
+ bool LSrcIsPhys = MRegisterInfo::isPhysicalRegister(left.SrcReg);
+ bool LIsPhys = LDstIsPhys || LSrcIsPhys;
+ bool RDstIsPhys = MRegisterInfo::isPhysicalRegister(right.DstReg);
+ bool RSrcIsPhys = MRegisterInfo::isPhysicalRegister(right.SrcReg);
+ bool RIsPhys = RDstIsPhys || RSrcIsPhys;
+ if (LIsPhys && !RIsPhys)
+ return false;
+ else if (LIsPhys == RIsPhys) {
+ // Join shorter intervals first.
+ unsigned LSize = 0;
+ unsigned RSize = 0;
+ if (LIsPhys) {
+ LSize = LDstIsPhys ? 0 : JPQ->getRepIntervalSize(left.DstReg);
+ LSize += LSrcIsPhys ? 0 : JPQ->getRepIntervalSize(left.SrcReg);
+ RSize = RDstIsPhys ? 0 : JPQ->getRepIntervalSize(right.DstReg);
+ RSize += RSrcIsPhys ? 0 : JPQ->getRepIntervalSize(right.SrcReg);
+ } else {
+ LSize = std::min(JPQ->getRepIntervalSize(left.DstReg),
+ JPQ->getRepIntervalSize(left.SrcReg));
+ RSize = std::min(JPQ->getRepIntervalSize(right.DstReg),
+ JPQ->getRepIntervalSize(right.SrcReg));
+ }
+ if (LSize < RSize)
+ return false;
+ }
+ }
+ }
+ return true;
+}
+
void SimpleRegisterCoalescing::CopyCoalesceInMBB(MachineBasicBlock *MBB,
- std::vector<CopyRec> *TryAgain, bool PhysOnly) {
+ std::vector<CopyRec> &TryAgain) {
DOUT << ((Value*)MBB->getBasicBlock())->getName() << ":\n";
-
+
+ std::vector<CopyRec> VirtCopies;
+ std::vector<CopyRec> PhysCopies;
+ unsigned LoopDepth = loopInfo->getLoopDepth(MBB->getBasicBlock());
for (MachineBasicBlock::iterator MII = MBB->begin(), E = MBB->end();
MII != E;) {
MachineInstr *Inst = MII++;
- // If this isn't a copy, we can't join intervals.
+ // If this isn't a copy nor a extract_subreg, we can't join intervals.
unsigned SrcReg, DstReg;
- if (!tii_->isMoveInstr(*Inst, SrcReg, DstReg)) continue;
-
- if (TryAgain && !JoinCopy(Inst, SrcReg, DstReg, PhysOnly))
- TryAgain->push_back(getCopyRec(Inst, SrcReg, DstReg));
+ if (Inst->getOpcode() == TargetInstrInfo::EXTRACT_SUBREG) {
+ DstReg = Inst->getOperand(0).getReg();
+ SrcReg = Inst->getOperand(1).getReg();
+ } else if (!tii_->isMoveInstr(*Inst, SrcReg, DstReg))
+ continue;
+
+ unsigned repSrcReg = rep(SrcReg);
+ unsigned repDstReg = rep(DstReg);
+ bool SrcIsPhys = MRegisterInfo::isPhysicalRegister(repSrcReg);
+ bool DstIsPhys = MRegisterInfo::isPhysicalRegister(repDstReg);
+ if (NewHeuristic) {
+ JoinQueue->push(CopyRec(Inst, SrcReg, DstReg, LoopDepth,
+ isBackEdgeCopy(Inst, DstReg)));
+ } else {
+ if (SrcIsPhys || DstIsPhys)
+ PhysCopies.push_back(CopyRec(Inst, SrcReg, DstReg, 0, false));
+ else
+ VirtCopies.push_back(CopyRec(Inst, SrcReg, DstReg, 0, false));
+ }
+ }
+
+ if (NewHeuristic)
+ return;
+
+ // Try coalescing physical register + virtual register first.
+ for (unsigned i = 0, e = PhysCopies.size(); i != e; ++i) {
+ CopyRec &TheCopy = PhysCopies[i];
+ bool Again = false;
+ if (!JoinCopy(TheCopy, Again))
+ if (Again)
+ TryAgain.push_back(TheCopy);
+ }
+ for (unsigned i = 0, e = VirtCopies.size(); i != e; ++i) {
+ CopyRec &TheCopy = VirtCopies[i];
+ bool Again = false;
+ if (!JoinCopy(TheCopy, Again))
+ if (Again)
+ TryAgain.push_back(TheCopy);
}
}
void SimpleRegisterCoalescing::joinIntervals() {
DOUT << "********** JOINING INTERVALS ***********\n";
+ if (NewHeuristic)
+ JoinQueue = new JoinPriorityQueue<CopyRecSort>(this);
+
JoinedLIs.resize(li_->getNumIntervals());
JoinedLIs.reset();
std::vector<CopyRec> TryAgainList;
- const LoopInfo &LI = getAnalysis<LoopInfo>();
- if (LI.begin() == LI.end()) {
+ if (loopInfo->begin() == loopInfo->end()) {
// If there are no loops in the function, join intervals in function order.
for (MachineFunction::iterator I = mf_->begin(), E = mf_->end();
I != E; ++I)
- CopyCoalesceInMBB(I, &TryAgainList);
+ CopyCoalesceInMBB(I, TryAgainList);
} else {
// Otherwise, join intervals in inner loops before other intervals.
// Unfortunately we can't just iterate over loop hierarchy here because
// registers with virtual registers before the intervals got too long.
std::vector<std::pair<unsigned, MachineBasicBlock*> > MBBs;
for (MachineFunction::iterator I = mf_->begin(), E = mf_->end(); I != E;++I)
- MBBs.push_back(std::make_pair(LI.getLoopDepth(I->getBasicBlock()), I));
+ MBBs.push_back(std::make_pair(loopInfo->
+ getLoopDepth(I->getBasicBlock()), I));
// Sort by loop depth.
std::sort(MBBs.begin(), MBBs.end(), DepthMBBCompare());
// Finally, join intervals in loop nest order.
for (unsigned i = 0, e = MBBs.size(); i != e; ++i)
- CopyCoalesceInMBB(MBBs[i].second, NULL, true);
- for (unsigned i = 0, e = MBBs.size(); i != e; ++i)
- CopyCoalesceInMBB(MBBs[i].second, &TryAgainList, false);
+ CopyCoalesceInMBB(MBBs[i].second, TryAgainList);
}
// Joining intervals can allow other intervals to be joined. Iteratively join
// until we make no progress.
- bool ProgressMade = true;
- while (ProgressMade) {
- ProgressMade = false;
-
- for (unsigned i = 0, e = TryAgainList.size(); i != e; ++i) {
- CopyRec &TheCopy = TryAgainList[i];
- if (TheCopy.MI &&
- JoinCopy(TheCopy.MI, TheCopy.SrcReg, TheCopy.DstReg)) {
- TheCopy.MI = 0; // Mark this one as done.
- ProgressMade = true;
+ if (NewHeuristic) {
+ SmallVector<CopyRec, 16> TryAgain;
+ bool ProgressMade = true;
+ while (ProgressMade) {
+ ProgressMade = false;
+ while (!JoinQueue->empty()) {
+ CopyRec R = JoinQueue->pop();
+ bool Again = false;
+ bool Success = JoinCopy(R, Again);
+ if (Success)
+ ProgressMade = true;
+ else if (Again)
+ TryAgain.push_back(R);
+ }
+
+ if (ProgressMade) {
+ while (!TryAgain.empty()) {
+ JoinQueue->push(TryAgain.back());
+ TryAgain.pop_back();
+ }
+ }
+ }
+ } else {
+ bool ProgressMade = true;
+ while (ProgressMade) {
+ ProgressMade = false;
+
+ for (unsigned i = 0, e = TryAgainList.size(); i != e; ++i) {
+ CopyRec &TheCopy = TryAgainList[i];
+ if (TheCopy.MI) {
+ bool Again = false;
+ bool Success = JoinCopy(TheCopy, Again);
+ if (Success || !Again) {
+ TheCopy.MI = 0; // Mark this one as done.
+ ProgressMade = true;
+ }
+ }
}
}
}
}
RegNum = JoinedLIs.find_next(RegNum);
}
+
+ if (NewHeuristic)
+ delete JoinQueue;
DOUT << "*** Register mapping ***\n";
- for (int i = 0, e = r2rMap_.size(); i != e; ++i)
+ for (unsigned i = 0, e = r2rMap_.size(); i != e; ++i)
if (r2rMap_[i]) {
DOUT << " reg " << i << " -> ";
DEBUG(printRegName(r2rMap_[i]));
/// Return true if the two specified registers belong to different register
/// classes. The registers may be either phys or virt regs.
bool SimpleRegisterCoalescing::differingRegisterClasses(unsigned RegA,
- unsigned RegB) const {
+ unsigned RegB) const {
// Get the register classes for the first reg.
if (MRegisterInfo::isPhysicalRegister(RegA)) {
for (unsigned i = 0, NumOps = MI->getNumOperands(); i != NumOps; ++i) {
MachineOperand &MO = MI->getOperand(i);
- if (MO.isReg() && MO.isUse() && MO.getReg() &&
+ if (MO.isRegister() && MO.isUse() && MO.getReg() &&
mri_->regsOverlap(rep(MO.getReg()), Reg)) {
MOU = &MO;
return MI;
MachineOperand *SimpleRegisterCoalescing::findDefOperand(MachineInstr *MI, unsigned Reg) {
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
- if (MO.isReg() && MO.isDef() &&
+ if (MO.isRegister() && MO.isDef() &&
mri_->regsOverlap(rep(MO.getReg()), Reg))
return &MO;
}
void SimpleRegisterCoalescing::unsetRegisterKill(MachineInstr *MI, unsigned Reg) {
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
- if (MO.isReg() && MO.isKill() && MO.getReg() &&
+ if (MO.isRegister() && MO.isKill() && MO.getReg() &&
mri_->regsOverlap(rep(MO.getReg()), Reg))
MO.unsetIsKill();
}
for (unsigned i = 0, NumOps = MI->getNumOperands(); i != NumOps; ++i) {
MachineOperand &MO = MI->getOperand(i);
- if (MO.isReg() && MO.isKill() && MO.getReg() &&
+ if (MO.isRegister() && MO.isKill() && MO.getReg() &&
mri_->regsOverlap(rep(MO.getReg()), Reg)) {
MO.unsetIsKill();
}
bool SimpleRegisterCoalescing::hasRegisterDef(MachineInstr *MI, unsigned Reg) {
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
- if (MO.isReg() && MO.isDef() &&
+ if (MO.isRegister() && MO.isDef() &&
mri_->regsOverlap(rep(MO.getReg()), Reg))
return true;
}
}
void SimpleRegisterCoalescing::releaseMemory() {
- r2rMap_.clear();
- JoinedLIs.clear();
+ for (unsigned i = 0, e = r2rMap_.size(); i != e; ++i)
+ r2rRevMap_[i].clear();
+ r2rRevMap_.clear();
+ r2rMap_.clear();
+ JoinedLIs.clear();
+ SubRegIdxes.clear();
+ JoinedCopies.clear();
}
static bool isZeroLengthInterval(LiveInterval *li) {
tii_ = tm_->getInstrInfo();
li_ = &getAnalysis<LiveIntervals>();
lv_ = &getAnalysis<LiveVariables>();
+ loopInfo = &getAnalysis<LoopInfo>();
DOUT << "********** SIMPLE REGISTER COALESCING **********\n"
<< "********** Function: "
E = mri_->regclass_end(); I != E; ++I)
allocatableRCRegs_.insert(std::make_pair(*I,mri_->getAllocatableSet(fn, *I)));
- r2rMap_.grow(mf_->getSSARegMap()->getLastVirtReg());
+ SSARegMap *RegMap = mf_->getSSARegMap();
+ r2rMap_.grow(RegMap->getLastVirtReg());
+ r2rRevMap_.grow(RegMap->getLastVirtReg());
// Join (coalesce) intervals if requested.
+ IndexedMap<unsigned, VirtReg2IndexFunctor> RegSubIdxMap;
if (EnableJoining) {
joinIntervals();
DOUT << "********** INTERVALS POST JOINING **********\n";
I->second.print(DOUT, mri_);
DOUT << "\n";
}
+
+ // Delete all coalesced copies.
+ for (SmallPtrSet<MachineInstr*,32>::iterator I = JoinedCopies.begin(),
+ E = JoinedCopies.end(); I != E; ++I) {
+ li_->RemoveMachineInstrFromMaps(*I);
+ (*I)->eraseFromParent();
+ }
+
+ // Transfer sub-registers info to SSARegMap now that coalescing information
+ // is complete.
+ RegSubIdxMap.grow(mf_->getSSARegMap()->getLastVirtReg()+1);
+ while (!SubRegIdxes.empty()) {
+ std::pair<unsigned, unsigned> RI = SubRegIdxes.back();
+ SubRegIdxes.pop_back();
+ RegSubIdxMap[RI.first] = RI.second;
+ }
}
// perform a final pass over the instructions and compute spill
// weights, coalesce virtual registers and remove identity moves.
- const LoopInfo &loopInfo = getAnalysis<LoopInfo>();
-
for (MachineFunction::iterator mbbi = mf_->begin(), mbbe = mf_->end();
mbbi != mbbe; ++mbbi) {
MachineBasicBlock* mbb = mbbi;
- unsigned loopDepth = loopInfo.getLoopDepth(mbb->getBasicBlock());
+ unsigned loopDepth = loopInfo->getLoopDepth(mbb->getBasicBlock());
for (MachineBasicBlock::iterator mii = mbb->begin(), mie = mbb->end();
mii != mie; ) {
if (mop.isRegister() && mop.getReg() &&
MRegisterInfo::isVirtualRegister(mop.getReg())) {
// replace register with representative register
- unsigned reg = rep(mop.getReg());
- mii->getOperand(i).setReg(reg);
+ unsigned OrigReg = mop.getReg();
+ unsigned reg = rep(OrigReg);
+ unsigned SubIdx = RegSubIdxMap[OrigReg];
+ if (SubIdx && MRegisterInfo::isPhysicalRegister(reg))
+ mii->getOperand(i).setReg(mri_->getSubReg(reg, SubIdx));
+ else {
+ mii->getOperand(i).setReg(reg);
+ mii->getOperand(i).setSubReg(SubIdx);
+ }
// Multiple uses of reg by the same instruction. It should not
// contribute to spill weight again.
if (UniqueUses.count(reg) != 0)
continue;
LiveInterval &RegInt = li_->getInterval(reg);
- float w = (mop.isUse()+mop.isDef()) * powf(10.0F, (float)loopDepth);
- // If the definition instruction is re-materializable, its spill
- // weight is half of what it would have been normally unless it's
- // a load from fixed stack slot.
- int Dummy;
- if (RegInt.remat && !tii_->isLoadFromStackSlot(RegInt.remat, Dummy))
- w /= 2;
- RegInt.weight += w;
+ RegInt.weight += li_->getSpillWeight(mop, loopDepth);
UniqueUses.insert(reg);
}
}
void SimpleRegisterCoalescing::print(std::ostream &O, const Module* m) const {
li_->print(O, m);
}
+
+RegisterCoalescer* llvm::createSimpleRegisterCoalescer() {
+ return new SimpleRegisterCoalescing();
+}
+
+// Make sure that anything that uses RegisterCoalescer pulls in this file...
+DEFINING_FILE_FOR(SimpleRegisterCoalescing)