#include <algorithm>
using namespace llvm;
-STATISTIC(NumSpills, "Number of register spills");
-STATISTIC(NumPSpills,"Number of physical register spills");
-STATISTIC(NumReMats, "Number of re-materialization");
-STATISTIC(NumDRM , "Number of re-materializable defs elided");
-STATISTIC(NumStores, "Number of stores added");
-STATISTIC(NumLoads , "Number of loads added");
-STATISTIC(NumReused, "Number of values reused");
-STATISTIC(NumDSE , "Number of dead stores elided");
-STATISTIC(NumDCE , "Number of copies elided");
-STATISTIC(NumDSS , "Number of dead spill slots removed");
+STATISTIC(NumSpills , "Number of register spills");
+STATISTIC(NumPSpills , "Number of physical register spills");
+STATISTIC(NumReMats , "Number of re-materialization");
+STATISTIC(NumDRM , "Number of re-materializable defs elided");
+STATISTIC(NumStores , "Number of stores added");
+STATISTIC(NumLoads , "Number of loads added");
+STATISTIC(NumReused , "Number of values reused");
+STATISTIC(NumDSE , "Number of dead stores elided");
+STATISTIC(NumDCE , "Number of copies elided");
+STATISTIC(NumDSS , "Number of dead spill slots removed");
+STATISTIC(NumCommutes, "Number of instructions commuted");
namespace {
enum SpillerName { simple, local };
AvailableSpills &Spills, BitVector &RegKills,
std::vector<MachineOperand*> &KillOps,
VirtRegMap &VRM);
+ bool CommuteToFoldReload(MachineBasicBlock &MBB,
+ MachineBasicBlock::iterator &MII,
+ unsigned VirtReg, unsigned SrcReg, int SS,
+ BitVector &RegKills,
+ std::vector<MachineOperand*> &KillOps,
+ const TargetRegisterInfo *TRI,
+ VirtRegMap &VRM);
void SpillRegToStackSlot(MachineBasicBlock &MBB,
MachineBasicBlock::iterator &MII,
int Idx, unsigned PhysReg, int StackSlot,
return false;
bool FoundUse = false, Done = false;
- MachineBasicBlock::iterator E = NewDef;
+ MachineBasicBlock::iterator E = &NewDef;
++I; ++E;
for (; !Done && I != E; ++I) {
MachineInstr *NMI = I;
/// This enables unfolding optimization for a subsequent instruction which will
/// also eliminate the newly introduced store instruction.
bool LocalSpiller::PrepForUnfoldOpti(MachineBasicBlock &MBB,
- MachineBasicBlock::iterator &MII,
+ MachineBasicBlock::iterator &MII,
std::vector<MachineInstr*> &MaybeDeadStores,
- AvailableSpills &Spills,
- BitVector &RegKills,
- std::vector<MachineOperand*> &KillOps,
- VirtRegMap &VRM) {
+ AvailableSpills &Spills,
+ BitVector &RegKills,
+ std::vector<MachineOperand*> &KillOps,
+ VirtRegMap &VRM) {
MachineFunction &MF = *MBB.getParent();
MachineInstr &MI = *MII;
unsigned UnfoldedOpc = 0;
MBB.erase(&MI);
return true;
}
- delete NewMI;
+ MF.DeleteMachineInstr(NewMI);
}
}
return false;
}
+/// CommuteToFoldReload -
+/// Look for
+/// r1 = load fi#1
+/// r1 = op r1, r2<kill>
+/// store r1, fi#1
+///
+/// If op is commutable and r2 is killed, then we can xform these to
+/// r2 = op r2, fi#1
+/// store r2, fi#1
+bool LocalSpiller::CommuteToFoldReload(MachineBasicBlock &MBB,
+ MachineBasicBlock::iterator &MII,
+ unsigned VirtReg, unsigned SrcReg, int SS,
+ BitVector &RegKills,
+ std::vector<MachineOperand*> &KillOps,
+ const TargetRegisterInfo *TRI,
+ VirtRegMap &VRM) {
+ if (MII == MBB.begin() || !MII->killsRegister(SrcReg))
+ return false;
+
+ MachineFunction &MF = *MBB.getParent();
+ MachineInstr &MI = *MII;
+ MachineBasicBlock::iterator DefMII = prior(MII);
+ MachineInstr *DefMI = DefMII;
+ const TargetInstrDesc &TID = DefMI->getDesc();
+ unsigned NewDstIdx;
+ if (DefMII != MBB.begin() &&
+ TID.isCommutable() &&
+ TII->CommuteChangesDestination(DefMI, NewDstIdx)) {
+ MachineOperand &NewDstMO = DefMI->getOperand(NewDstIdx);
+ unsigned NewReg = NewDstMO.getReg();
+ if (!NewDstMO.isKill() || TRI->regsOverlap(NewReg, SrcReg))
+ return false;
+ MachineInstr *ReloadMI = prior(DefMII);
+ int FrameIdx;
+ unsigned DestReg = TII->isLoadFromStackSlot(ReloadMI, FrameIdx);
+ if (DestReg != SrcReg || FrameIdx != SS)
+ return false;
+ int UseIdx = DefMI->findRegisterUseOperandIdx(DestReg, false);
+ if (UseIdx == -1)
+ return false;
+ int DefIdx = TID.getOperandConstraint(UseIdx, TOI::TIED_TO);
+ if (DefIdx == -1)
+ return false;
+ assert(DefMI->getOperand(DefIdx).isRegister() &&
+ DefMI->getOperand(DefIdx).getReg() == SrcReg);
+
+ // Now commute def instruction.
+ MachineInstr *CommutedMI = TII->commuteInstruction(DefMI, true);
+ if (!CommutedMI)
+ return false;
+ SmallVector<unsigned, 2> Ops;
+ Ops.push_back(NewDstIdx);
+ MachineInstr *FoldedMI = TII->foldMemoryOperand(MF, CommutedMI, Ops, SS);
+ // Not needed since foldMemoryOperand returns new MI.
+ MF.DeleteMachineInstr(CommutedMI);
+ if (!FoldedMI)
+ return false;
+
+ VRM.addSpillSlotUse(SS, FoldedMI);
+ VRM.virtFolded(VirtReg, FoldedMI, VirtRegMap::isRef);
+ // Insert new def MI and spill MI.
+ const TargetRegisterClass* RC = MF.getRegInfo().getRegClass(VirtReg);
+ TII->storeRegToStackSlot(MBB, &MI, NewReg, true, SS, RC);
+ MII = prior(MII);
+ MachineInstr *StoreMI = MII;
+ VRM.addSpillSlotUse(SS, StoreMI);
+ VRM.virtFolded(VirtReg, StoreMI, VirtRegMap::isMod);
+ MII = MBB.insert(MII, FoldedMI); // Update MII to backtrack.
+
+ // Delete all 3 old instructions.
+ InvalidateKills(*ReloadMI, RegKills, KillOps);
+ VRM.RemoveMachineInstrFromMaps(ReloadMI);
+ MBB.erase(ReloadMI);
+ InvalidateKills(*DefMI, RegKills, KillOps);
+ VRM.RemoveMachineInstrFromMaps(DefMI);
+ MBB.erase(DefMI);
+ InvalidateKills(MI, RegKills, KillOps);
+ VRM.RemoveMachineInstrFromMaps(&MI);
+ MBB.erase(&MI);
+
+ ++NumCommutes;
+ return true;
+ }
+
+ return false;
+}
+
/// findSuperReg - Find the SubReg's super-register of given register class
/// where its SubIdx sub-register is SubReg.
static unsigned findSuperReg(const TargetRegisterClass *RC, unsigned SubReg,
unsigned RReg = SubIdx ? TRI->getSubReg(Phys, SubIdx) : Phys;
MI.getOperand(i).setReg(RReg);
if (VRM.isImplicitlyDefined(VirtReg))
- BuildMI(MBB, MI, TII->get(TargetInstrInfo::IMPLICIT_DEF), RReg);
+ BuildMI(MBB, &MI, TII->get(TargetInstrInfo::IMPLICIT_DEF), RReg);
continue;
}
if (unsigned SrcReg = TII->isStoreToStackSlot(&MI, StackSlot)) {
assert(TargetRegisterInfo::isPhysicalRegister(SrcReg) &&
"Src hasn't been allocated yet?");
+
+ if (CommuteToFoldReload(MBB, MII, VirtReg, SrcReg, StackSlot,
+ RegKills, KillOps, TRI, VRM)) {
+ NextMII = next(MII);
+ BackTracked = true;
+ goto ProcessNextInst;
+ }
+
// Okay, this is certainly a store of SrcReg to [StackSlot]. Mark
// this as a potentially dead store in case there is a subsequent
// store into the stack slot without a read from it.
MaybeDeadStores[StackSlot] = &MI;
// If the stack slot value was previously available in some other
- // register, change it now. Otherwise, make the register available,
- // in PhysReg.
- Spills.addAvailable(StackSlot, &MI, SrcReg, false/*don't clobber*/);
+ // register, change it now. Otherwise, make the register
+ // available in PhysReg.
+ Spills.addAvailable(StackSlot, &MI, SrcReg, false/*!clobber*/);
}
}
}
ProcessNextInst:
DistanceMap.insert(std::make_pair(&MI, Dist++));
if (!Erased && !BackTracked) {
- for (MachineBasicBlock::iterator II = MI; II != NextMII; ++II)
+ for (MachineBasicBlock::iterator II = &MI; II != NextMII; ++II)
UpdateKills(*II, RegKills, KillOps);
}
MII = NextMII;
projects / oota-llvm.git / blobdiff