#include "llvm/CodeGen/LiveVariables.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/MC/MCInstrItineraries.h"
#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetMachine.h"
STATISTIC(Num3AddrSunk, "Number of 3-address instructions sunk");
STATISTIC(NumReMats, "Number of instructions re-materialized");
STATISTIC(NumDeletes, "Number of dead instructions deleted");
+STATISTIC(NumReSchedUps, "Number of instructions re-scheduled up");
+STATISTIC(NumReSchedDowns, "Number of instructions re-scheduled down");
namespace {
class TwoAddressInstructionPass : public MachineFunctionPass {
const TargetInstrInfo *TII;
const TargetRegisterInfo *TRI;
+ const InstrItineraryData *InstrItins;
MachineRegisterInfo *MRI;
LiveVariables *LV;
AliasAnalysis *AA;
+ CodeGenOpt::Level OptLevel;
// DistanceMap - Keep track the distance of a MI from the start of the
// current basic block.
MachineFunction::iterator &mbbi,
unsigned RegB, unsigned RegC, unsigned Dist);
- bool isProfitableToConv3Addr(unsigned RegA);
+ bool isProfitableToConv3Addr(unsigned RegA, unsigned RegB);
bool ConvertInstTo3Addr(MachineBasicBlock::iterator &mi,
MachineBasicBlock::iterator &nmi,
MachineFunction::iterator &mbbi,
- unsigned RegB, unsigned Dist);
+ unsigned RegA, unsigned RegB, unsigned Dist);
typedef std::pair<std::pair<unsigned, bool>, MachineInstr*> NewKill;
bool canUpdateDeletedKills(SmallVector<unsigned, 4> &Kills,
MachineBasicBlock::iterator &nmi,
MachineFunction::iterator &mbbi, unsigned Dist);
+ bool isDefTooClose(unsigned Reg, unsigned Dist,
+ MachineInstr *MI, MachineBasicBlock *MBB);
+
+ bool RescheduleMIBelowKill(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator &mi,
+ MachineBasicBlock::iterator &nmi,
+ unsigned Reg);
+ bool RescheduleKillAboveMI(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator &mi,
+ MachineBasicBlock::iterator &nmi,
+ unsigned Reg);
+
bool TryInstructionTransform(MachineBasicBlock::iterator &mi,
MachineBasicBlock::iterator &nmi,
MachineFunction::iterator &mbbi,
unsigned SrcIdx, unsigned DstIdx,
- unsigned Dist);
+ unsigned Dist,
+ SmallPtrSet<MachineInstr*, 8> &Processed);
+
+ void ScanUses(unsigned DstReg, MachineBasicBlock *MBB,
+ SmallPtrSet<MachineInstr*, 8> &Processed);
void ProcessCopy(MachineInstr *MI, MachineBasicBlock *MBB,
SmallPtrSet<MachineInstr*, 8> &Processed);
/// of the de-ssa process. This replaces sources of REG_SEQUENCE as
/// sub-register references of the register defined by REG_SEQUENCE.
bool EliminateRegSequences();
+
public:
static char ID; // Pass identification, replacement for typeid
- TwoAddressInstructionPass() : MachineFunctionPass(&ID) {}
+ TwoAddressInstructionPass() : MachineFunctionPass(ID) {
+ initializeTwoAddressInstructionPassPass(*PassRegistry::getPassRegistry());
+ }
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesCFG();
AU.addPreserved<LiveVariables>();
AU.addPreservedID(MachineLoopInfoID);
AU.addPreservedID(MachineDominatorsID);
- if (StrongPHIElim)
- AU.addPreservedID(StrongPHIEliminationID);
- else
- AU.addPreservedID(PHIEliminationID);
+ AU.addPreservedID(PHIEliminationID);
MachineFunctionPass::getAnalysisUsage(AU);
}
}
char TwoAddressInstructionPass::ID = 0;
-static RegisterPass<TwoAddressInstructionPass>
-X("twoaddressinstruction", "Two-Address instruction pass");
+INITIALIZE_PASS_BEGIN(TwoAddressInstructionPass, "twoaddressinstruction",
+ "Two-Address instruction pass", false, false)
+INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
+INITIALIZE_PASS_END(TwoAddressInstructionPass, "twoaddressinstruction",
+ "Two-Address instruction pass", false, false)
-const PassInfo *const llvm::TwoAddressInstructionPassID = &X;
+char &llvm::TwoAddressInstructionPassID = TwoAddressInstructionPass::ID;
/// Sink3AddrInstruction - A two-address instruction has been converted to a
/// three-address instruction to avoid clobbering a register. Try to sink it
bool TwoAddressInstructionPass::Sink3AddrInstruction(MachineBasicBlock *MBB,
MachineInstr *MI, unsigned SavedReg,
MachineBasicBlock::iterator OldPos) {
+ // FIXME: Shouldn't we be trying to do this before we three-addressify the
+ // instruction? After this transformation is done, we no longer need
+ // the instruction to be in three-address form.
+
// Check if it's safe to move this instruction.
bool SeenStore = true; // Be conservative.
if (!MI->isSafeToMove(TII, AA, SeenStore))
break;
}
- if (!KillMI || KillMI->getParent() != MBB || KillMI == MI)
+ // If we find the instruction that kills SavedReg, and it is in an
+ // appropriate location, we can try to sink the current instruction
+ // past it.
+ if (!KillMI || KillMI->getParent() != MBB || KillMI == MI ||
+ KillMI->isTerminator())
return false;
// If any of the definitions are used by another instruction between the
/// isTwoAddrUse - Return true if the specified MI is using the specified
/// register as a two-address operand.
static bool isTwoAddrUse(MachineInstr *UseMI, unsigned Reg) {
- const TargetInstrDesc &TID = UseMI->getDesc();
- for (unsigned i = 0, e = TID.getNumOperands(); i != e; ++i) {
+ const MCInstrDesc &MCID = UseMI->getDesc();
+ for (unsigned i = 0, e = MCID.getNumOperands(); i != e; ++i) {
MachineOperand &MO = UseMI->getOperand(i);
if (MO.isReg() && MO.getReg() == Reg &&
(MO.isDef() || UseMI->isRegTiedToDefOperand(i)))
bool &IsSrcPhys, bool &IsDstPhys) {
SrcReg = 0;
DstReg = 0;
- unsigned SrcSubIdx, DstSubIdx;
- if (!TII->isMoveInstr(MI, SrcReg, DstReg, SrcSubIdx, DstSubIdx)) {
- if (MI.isExtractSubreg()) {
- DstReg = MI.getOperand(0).getReg();
- SrcReg = MI.getOperand(1).getReg();
- } else if (MI.isInsertSubreg()) {
- DstReg = MI.getOperand(0).getReg();
- SrcReg = MI.getOperand(2).getReg();
- } else if (MI.isSubregToReg()) {
- DstReg = MI.getOperand(0).getReg();
- SrcReg = MI.getOperand(2).getReg();
- }
- }
+ if (MI.isCopy()) {
+ DstReg = MI.getOperand(0).getReg();
+ SrcReg = MI.getOperand(1).getReg();
+ } else if (MI.isInsertSubreg() || MI.isSubregToReg()) {
+ DstReg = MI.getOperand(0).getReg();
+ SrcReg = MI.getOperand(2).getReg();
+ } else
+ return false;
- if (DstReg) {
- IsSrcPhys = TargetRegisterInfo::isPhysicalRegister(SrcReg);
- IsDstPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
- return true;
- }
- return false;
+ IsSrcPhys = TargetRegisterInfo::isPhysicalRegister(SrcReg);
+ IsDstPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
+ return true;
}
/// isKilled - Test if the given register value, which is used by the given
/// isTwoAddrUse - Return true if the specified MI uses the specified register
/// as a two-address use. If so, return the destination register by reference.
static bool isTwoAddrUse(MachineInstr &MI, unsigned Reg, unsigned &DstReg) {
- const TargetInstrDesc &TID = MI.getDesc();
- unsigned NumOps = MI.isInlineAsm() ? MI.getNumOperands():TID.getNumOperands();
+ const MCInstrDesc &MCID = MI.getDesc();
+ unsigned NumOps = MI.isInlineAsm()
+ ? MI.getNumOperands() : MCID.getNumOperands();
for (unsigned i = 0; i != NumOps; ++i) {
const MachineOperand &MO = MI.getOperand(i);
if (!MO.isReg() || !MO.isUse() || MO.getReg() != Reg)
return false;
}
+/// findLocalKill - Look for an instruction below MI in the MBB that kills the
+/// specified register. Returns null if there are any other Reg use between the
+/// instructions.
+static
+MachineInstr *findLocalKill(unsigned Reg, MachineBasicBlock *MBB,
+ MachineInstr *MI, MachineRegisterInfo *MRI,
+ DenseMap<MachineInstr*, unsigned> &DistanceMap) {
+ MachineInstr *KillMI = 0;
+ for (MachineRegisterInfo::use_nodbg_iterator
+ UI = MRI->use_nodbg_begin(Reg),
+ UE = MRI->use_nodbg_end(); UI != UE; ++UI) {
+ MachineInstr *UseMI = &*UI;
+ if (UseMI == MI || UseMI->getParent() != MBB)
+ continue;
+ if (DistanceMap.count(UseMI))
+ continue;
+ if (!UI.getOperand().isKill())
+ return 0;
+ if (KillMI)
+ return 0; // -O0 kill markers cannot be trusted?
+ KillMI = UseMI;
+ }
+
+ return KillMI;
+}
+
/// findOnlyInterestingUse - Given a register, if has a single in-basic block
/// use, return the use instruction if it's a copy or a two-address use.
static
TwoAddressInstructionPass::isProfitableToCommute(unsigned regB, unsigned regC,
MachineInstr *MI, MachineBasicBlock *MBB,
unsigned Dist) {
+ if (OptLevel == CodeGenOpt::None)
+ return false;
+
// Determine if it's profitable to commute this two address instruction. In
// general, we want no uses between this instruction and the definition of
// the two-address register.
unsigned FromRegC = getMappedReg(regC, SrcRegMap);
unsigned ToRegB = getMappedReg(regB, DstRegMap);
unsigned ToRegC = getMappedReg(regC, DstRegMap);
- if (!regsAreCompatible(FromRegB, ToRegB, TRI) &&
- (regsAreCompatible(FromRegB, ToRegC, TRI) ||
+ if ((FromRegB && ToRegB && !regsAreCompatible(FromRegB, ToRegB, TRI)) &&
+ ((!FromRegC && !ToRegC) ||
+ regsAreCompatible(FromRegB, ToRegC, TRI) ||
regsAreCompatible(FromRegC, ToRegB, TRI)))
return true;
/// isProfitableToConv3Addr - Return true if it is profitable to convert the
/// given 2-address instruction to a 3-address one.
bool
-TwoAddressInstructionPass::isProfitableToConv3Addr(unsigned RegA) {
+TwoAddressInstructionPass::isProfitableToConv3Addr(unsigned RegA,unsigned RegB){
// Look for situations like this:
// %reg1024<def> = MOV r1
// %reg1025<def> = MOV r0
// %reg1026<def> = ADD %reg1024, %reg1025
// r2 = MOV %reg1026
// Turn ADD into a 3-address instruction to avoid a copy.
- unsigned FromRegA = getMappedReg(RegA, SrcRegMap);
+ unsigned FromRegB = getMappedReg(RegB, SrcRegMap);
+ if (!FromRegB)
+ return false;
unsigned ToRegA = getMappedReg(RegA, DstRegMap);
- return (FromRegA && ToRegA && !regsAreCompatible(FromRegA, ToRegA, TRI));
+ return (ToRegA && !regsAreCompatible(FromRegB, ToRegA, TRI));
}
/// ConvertInstTo3Addr - Convert the specified two-address instruction into a
TwoAddressInstructionPass::ConvertInstTo3Addr(MachineBasicBlock::iterator &mi,
MachineBasicBlock::iterator &nmi,
MachineFunction::iterator &mbbi,
- unsigned RegB, unsigned Dist) {
+ unsigned RegA, unsigned RegB,
+ unsigned Dist) {
MachineInstr *NewMI = TII->convertToThreeAddress(mbbi, mi, LV);
if (NewMI) {
DEBUG(dbgs() << "2addr: CONVERTING 2-ADDR: " << *mi);
mi = NewMI;
nmi = llvm::next(mi);
}
+
+ // Update source and destination register maps.
+ SrcRegMap.erase(RegA);
+ DstRegMap.erase(RegB);
return true;
}
return false;
}
+/// ScanUses - Scan forward recursively for only uses, update maps if the use
+/// is a copy or a two-address instruction.
+void
+TwoAddressInstructionPass::ScanUses(unsigned DstReg, MachineBasicBlock *MBB,
+ SmallPtrSet<MachineInstr*, 8> &Processed) {
+ SmallVector<unsigned, 4> VirtRegPairs;
+ bool IsDstPhys;
+ bool IsCopy = false;
+ unsigned NewReg = 0;
+ unsigned Reg = DstReg;
+ while (MachineInstr *UseMI = findOnlyInterestingUse(Reg, MBB, MRI, TII,IsCopy,
+ NewReg, IsDstPhys)) {
+ if (IsCopy && !Processed.insert(UseMI))
+ break;
+
+ DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(UseMI);
+ if (DI != DistanceMap.end())
+ // Earlier in the same MBB.Reached via a back edge.
+ break;
+
+ if (IsDstPhys) {
+ VirtRegPairs.push_back(NewReg);
+ break;
+ }
+ bool isNew = SrcRegMap.insert(std::make_pair(NewReg, Reg)).second;
+ if (!isNew)
+ assert(SrcRegMap[NewReg] == Reg && "Can't map to two src registers!");
+ VirtRegPairs.push_back(NewReg);
+ Reg = NewReg;
+ }
+
+ if (!VirtRegPairs.empty()) {
+ unsigned ToReg = VirtRegPairs.back();
+ VirtRegPairs.pop_back();
+ while (!VirtRegPairs.empty()) {
+ unsigned FromReg = VirtRegPairs.back();
+ VirtRegPairs.pop_back();
+ bool isNew = DstRegMap.insert(std::make_pair(FromReg, ToReg)).second;
+ if (!isNew)
+ assert(DstRegMap[FromReg] == ToReg &&"Can't map to two dst registers!");
+ ToReg = FromReg;
+ }
+ bool isNew = DstRegMap.insert(std::make_pair(DstReg, ToReg)).second;
+ if (!isNew)
+ assert(DstRegMap[DstReg] == ToReg && "Can't map to two dst registers!");
+ }
+}
+
/// ProcessCopy - If the specified instruction is not yet processed, process it
/// if it's a copy. For a copy instruction, we find the physical registers the
/// source and destination registers might be mapped to. These are kept in
assert(SrcRegMap[DstReg] == SrcReg &&
"Can't map to two src physical registers!");
- SmallVector<unsigned, 4> VirtRegPairs;
- bool IsCopy = false;
- unsigned NewReg = 0;
- while (MachineInstr *UseMI = findOnlyInterestingUse(DstReg, MBB, MRI,TII,
- IsCopy, NewReg, IsDstPhys)) {
- if (IsCopy) {
- if (!Processed.insert(UseMI))
- break;
- }
-
- DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(UseMI);
- if (DI != DistanceMap.end())
- // Earlier in the same MBB.Reached via a back edge.
- break;
-
- if (IsDstPhys) {
- VirtRegPairs.push_back(NewReg);
- break;
- }
- bool isNew = SrcRegMap.insert(std::make_pair(NewReg, DstReg)).second;
- if (!isNew)
- assert(SrcRegMap[NewReg] == DstReg &&
- "Can't map to two src physical registers!");
- VirtRegPairs.push_back(NewReg);
- DstReg = NewReg;
- }
-
- if (!VirtRegPairs.empty()) {
- unsigned ToReg = VirtRegPairs.back();
- VirtRegPairs.pop_back();
- while (!VirtRegPairs.empty()) {
- unsigned FromReg = VirtRegPairs.back();
- VirtRegPairs.pop_back();
- bool isNew = DstRegMap.insert(std::make_pair(FromReg, ToReg)).second;
- if (!isNew)
- assert(DstRegMap[FromReg] == ToReg &&
- "Can't map to two dst physical registers!");
- ToReg = FromReg;
- }
- }
+ ScanUses(DstReg, MBB, Processed);
}
Processed.insert(MI);
+ return;
}
/// isSafeToDelete - If the specified instruction does not produce any side
static bool isSafeToDelete(MachineInstr *MI,
const TargetInstrInfo *TII,
SmallVector<unsigned, 4> &Kills) {
- const TargetInstrDesc &TID = MI->getDesc();
- if (TID.mayStore() || TID.isCall())
+ if (MI->mayStore() || MI->isCall())
return false;
- if (TID.isTerminator() || TID.hasUnmodeledSideEffects())
+ if (MI->isTerminator() || MI->hasUnmodeledSideEffects())
return false;
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
if (!LastKill)
return false;
- bool isModRef = LastKill->modifiesRegister(Kill);
+ bool isModRef = LastKill->definesRegister(Kill);
NewKills.push_back(std::make_pair(std::make_pair(Kill, isModRef),
LastKill));
}
return true;
}
+/// RescheduleMIBelowKill - If there is one more local instruction that reads
+/// 'Reg' and it kills 'Reg, consider moving the instruction below the kill
+/// instruction in order to eliminate the need for the copy.
+bool
+TwoAddressInstructionPass::RescheduleMIBelowKill(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator &mi,
+ MachineBasicBlock::iterator &nmi,
+ unsigned Reg) {
+ MachineInstr *MI = &*mi;
+ DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(MI);
+ if (DI == DistanceMap.end())
+ // Must be created from unfolded load. Don't waste time trying this.
+ return false;
+
+ MachineInstr *KillMI = findLocalKill(Reg, MBB, mi, MRI, DistanceMap);
+ if (!KillMI || KillMI->isCopy() || KillMI->isCopyLike())
+ // Don't mess with copies, they may be coalesced later.
+ return false;
+
+ if (KillMI->hasUnmodeledSideEffects() || KillMI->isCall() ||
+ KillMI->isBranch() || KillMI->isTerminator())
+ // Don't move pass calls, etc.
+ return false;
+
+ unsigned DstReg;
+ if (isTwoAddrUse(*KillMI, Reg, DstReg))
+ return false;
+
+ bool SeenStore = true;
+ if (!MI->isSafeToMove(TII, AA, SeenStore))
+ return false;
+
+ if (TII->getInstrLatency(InstrItins, MI) > 1)
+ // FIXME: Needs more sophisticated heuristics.
+ return false;
+
+ SmallSet<unsigned, 2> Uses;
+ SmallSet<unsigned, 2> Kills;
+ SmallSet<unsigned, 2> Defs;
+ for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = MI->getOperand(i);
+ if (!MO.isReg())
+ continue;
+ unsigned MOReg = MO.getReg();
+ if (!MOReg)
+ continue;
+ if (MO.isDef())
+ Defs.insert(MOReg);
+ else {
+ Uses.insert(MOReg);
+ if (MO.isKill() && MOReg != Reg)
+ Kills.insert(MOReg);
+ }
+ }
+
+ // Move the copies connected to MI down as well.
+ MachineBasicBlock::iterator From = MI;
+ MachineBasicBlock::iterator To = llvm::next(From);
+ while (To->isCopy() && Defs.count(To->getOperand(1).getReg())) {
+ Defs.insert(To->getOperand(0).getReg());
+ ++To;
+ }
+
+ // Check if the reschedule will not break depedencies.
+ unsigned NumVisited = 0;
+ MachineBasicBlock::iterator KillPos = KillMI;
+ ++KillPos;
+ for (MachineBasicBlock::iterator I = To; I != KillPos; ++I) {
+ MachineInstr *OtherMI = I;
+ // DBG_VALUE cannot be counted against the limit.
+ if (OtherMI->isDebugValue())
+ continue;
+ if (NumVisited > 10) // FIXME: Arbitrary limit to reduce compile time cost.
+ return false;
+ ++NumVisited;
+ if (OtherMI->hasUnmodeledSideEffects() || OtherMI->isCall() ||
+ OtherMI->isBranch() || OtherMI->isTerminator())
+ // Don't move pass calls, etc.
+ return false;
+ for (unsigned i = 0, e = OtherMI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = OtherMI->getOperand(i);
+ if (!MO.isReg())
+ continue;
+ unsigned MOReg = MO.getReg();
+ if (!MOReg)
+ continue;
+ if (MO.isDef()) {
+ if (Uses.count(MOReg))
+ // Physical register use would be clobbered.
+ return false;
+ if (!MO.isDead() && Defs.count(MOReg))
+ // May clobber a physical register def.
+ // FIXME: This may be too conservative. It's ok if the instruction
+ // is sunken completely below the use.
+ return false;
+ } else {
+ if (Defs.count(MOReg))
+ return false;
+ if (MOReg != Reg &&
+ ((MO.isKill() && Uses.count(MOReg)) || Kills.count(MOReg)))
+ // Don't want to extend other live ranges and update kills.
+ return false;
+ }
+ }
+ }
+
+ // Move debug info as well.
+ while (From != MBB->begin() && llvm::prior(From)->isDebugValue())
+ --From;
+
+ // Copies following MI may have been moved as well.
+ nmi = To;
+ MBB->splice(KillPos, MBB, From, To);
+ DistanceMap.erase(DI);
+
+ if (LV) {
+ // Update live variables
+ LV->removeVirtualRegisterKilled(Reg, KillMI);
+ LV->addVirtualRegisterKilled(Reg, MI);
+ } else {
+ for (unsigned i = 0, e = KillMI->getNumOperands(); i != e; ++i) {
+ MachineOperand &MO = KillMI->getOperand(i);
+ if (!MO.isReg() || !MO.isUse() || MO.getReg() != Reg)
+ continue;
+ MO.setIsKill(false);
+ }
+ MI->addRegisterKilled(Reg, 0);
+ }
+
+ return true;
+}
+
+/// isDefTooClose - Return true if the re-scheduling will put the given
+/// instruction too close to the defs of its register dependencies.
+bool TwoAddressInstructionPass::isDefTooClose(unsigned Reg, unsigned Dist,
+ MachineInstr *MI,
+ MachineBasicBlock *MBB) {
+ for (MachineRegisterInfo::def_iterator DI = MRI->def_begin(Reg),
+ DE = MRI->def_end(); DI != DE; ++DI) {
+ MachineInstr *DefMI = &*DI;
+ if (DefMI->getParent() != MBB || DefMI->isCopy() || DefMI->isCopyLike())
+ continue;
+ if (DefMI == MI)
+ return true; // MI is defining something KillMI uses
+ DenseMap<MachineInstr*, unsigned>::iterator DDI = DistanceMap.find(DefMI);
+ if (DDI == DistanceMap.end())
+ return true; // Below MI
+ unsigned DefDist = DDI->second;
+ assert(Dist > DefDist && "Visited def already?");
+ if (TII->getInstrLatency(InstrItins, DefMI) > (int)(Dist - DefDist))
+ return true;
+ }
+ return false;
+}
+
+/// RescheduleKillAboveMI - If there is one more local instruction that reads
+/// 'Reg' and it kills 'Reg, consider moving the kill instruction above the
+/// current two-address instruction in order to eliminate the need for the
+/// copy.
+bool
+TwoAddressInstructionPass::RescheduleKillAboveMI(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator &mi,
+ MachineBasicBlock::iterator &nmi,
+ unsigned Reg) {
+ MachineInstr *MI = &*mi;
+ DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(MI);
+ if (DI == DistanceMap.end())
+ // Must be created from unfolded load. Don't waste time trying this.
+ return false;
+
+ MachineInstr *KillMI = findLocalKill(Reg, MBB, mi, MRI, DistanceMap);
+ if (!KillMI || KillMI->isCopy() || KillMI->isCopyLike())
+ // Don't mess with copies, they may be coalesced later.
+ return false;
+
+ unsigned DstReg;
+ if (isTwoAddrUse(*KillMI, Reg, DstReg))
+ return false;
+
+ bool SeenStore = true;
+ if (!KillMI->isSafeToMove(TII, AA, SeenStore))
+ return false;
+
+ SmallSet<unsigned, 2> Uses;
+ SmallSet<unsigned, 2> Kills;
+ SmallSet<unsigned, 2> Defs;
+ SmallSet<unsigned, 2> LiveDefs;
+ for (unsigned i = 0, e = KillMI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = KillMI->getOperand(i);
+ if (!MO.isReg())
+ continue;
+ unsigned MOReg = MO.getReg();
+ if (MO.isUse()) {
+ if (!MOReg)
+ continue;
+ if (isDefTooClose(MOReg, DI->second, MI, MBB))
+ return false;
+ Uses.insert(MOReg);
+ if (MO.isKill() && MOReg != Reg)
+ Kills.insert(MOReg);
+ } else if (TargetRegisterInfo::isPhysicalRegister(MOReg)) {
+ Defs.insert(MOReg);
+ if (!MO.isDead())
+ LiveDefs.insert(MOReg);
+ }
+ }
+
+ // Check if the reschedule will not break depedencies.
+ unsigned NumVisited = 0;
+ MachineBasicBlock::iterator KillPos = KillMI;
+ for (MachineBasicBlock::iterator I = mi; I != KillPos; ++I) {
+ MachineInstr *OtherMI = I;
+ // DBG_VALUE cannot be counted against the limit.
+ if (OtherMI->isDebugValue())
+ continue;
+ if (NumVisited > 10) // FIXME: Arbitrary limit to reduce compile time cost.
+ return false;
+ ++NumVisited;
+ if (OtherMI->hasUnmodeledSideEffects() || OtherMI->isCall() ||
+ OtherMI->isBranch() || OtherMI->isTerminator())
+ // Don't move pass calls, etc.
+ return false;
+ SmallVector<unsigned, 2> OtherDefs;
+ for (unsigned i = 0, e = OtherMI->getNumOperands(); i != e; ++i) {
+ const MachineOperand &MO = OtherMI->getOperand(i);
+ if (!MO.isReg())
+ continue;
+ unsigned MOReg = MO.getReg();
+ if (!MOReg)
+ continue;
+ if (MO.isUse()) {
+ if (Defs.count(MOReg))
+ // Moving KillMI can clobber the physical register if the def has
+ // not been seen.
+ return false;
+ if (Kills.count(MOReg))
+ // Don't want to extend other live ranges and update kills.
+ return false;
+ } else {
+ OtherDefs.push_back(MOReg);
+ }
+ }
+
+ for (unsigned i = 0, e = OtherDefs.size(); i != e; ++i) {
+ unsigned MOReg = OtherDefs[i];
+ if (Uses.count(MOReg))
+ return false;
+ if (TargetRegisterInfo::isPhysicalRegister(MOReg) &&
+ LiveDefs.count(MOReg))
+ return false;
+ // Physical register def is seen.
+ Defs.erase(MOReg);
+ }
+ }
+
+ // Move the old kill above MI, don't forget to move debug info as well.
+ MachineBasicBlock::iterator InsertPos = mi;
+ while (InsertPos != MBB->begin() && llvm::prior(InsertPos)->isDebugValue())
+ --InsertPos;
+ MachineBasicBlock::iterator From = KillMI;
+ MachineBasicBlock::iterator To = llvm::next(From);
+ while (llvm::prior(From)->isDebugValue())
+ --From;
+ MBB->splice(InsertPos, MBB, From, To);
+
+ nmi = llvm::prior(InsertPos); // Backtrack so we process the moved instr.
+ DistanceMap.erase(DI);
+
+ if (LV) {
+ // Update live variables
+ LV->removeVirtualRegisterKilled(Reg, KillMI);
+ LV->addVirtualRegisterKilled(Reg, MI);
+ } else {
+ for (unsigned i = 0, e = KillMI->getNumOperands(); i != e; ++i) {
+ MachineOperand &MO = KillMI->getOperand(i);
+ if (!MO.isReg() || !MO.isUse() || MO.getReg() != Reg)
+ continue;
+ MO.setIsKill(false);
+ }
+ MI->addRegisterKilled(Reg, 0);
+ }
+ return true;
+}
+
/// TryInstructionTransform - For the case where an instruction has a single
/// pair of tied register operands, attempt some transformations that may
/// either eliminate the tied operands or improve the opportunities for
TryInstructionTransform(MachineBasicBlock::iterator &mi,
MachineBasicBlock::iterator &nmi,
MachineFunction::iterator &mbbi,
- unsigned SrcIdx, unsigned DstIdx, unsigned Dist) {
- const TargetInstrDesc &TID = mi->getDesc();
- unsigned regA = mi->getOperand(DstIdx).getReg();
- unsigned regB = mi->getOperand(SrcIdx).getReg();
+ unsigned SrcIdx, unsigned DstIdx, unsigned Dist,
+ SmallPtrSet<MachineInstr*, 8> &Processed) {
+ if (OptLevel == CodeGenOpt::None)
+ return false;
+
+ MachineInstr &MI = *mi;
+ unsigned regA = MI.getOperand(DstIdx).getReg();
+ unsigned regB = MI.getOperand(SrcIdx).getReg();
assert(TargetRegisterInfo::isVirtualRegister(regB) &&
"cannot make instruction into two-address form");
// If regA is dead and the instruction can be deleted, just delete
// it so it doesn't clobber regB.
- bool regBKilled = isKilled(*mi, regB, MRI, TII);
- if (!regBKilled && mi->getOperand(DstIdx).isDead() &&
+ bool regBKilled = isKilled(MI, regB, MRI, TII);
+ if (!regBKilled && MI.getOperand(DstIdx).isDead() &&
DeleteUnusedInstr(mi, nmi, mbbi, Dist)) {
++NumDeletes;
return true; // Done with this instruction.
unsigned regCIdx = ~0U;
bool TryCommute = false;
bool AggressiveCommute = false;
- if (TID.isCommutable() && mi->getNumOperands() >= 3 &&
- TII->findCommutedOpIndices(mi, SrcOp1, SrcOp2)) {
+ if (MI.isCommutable() && MI.getNumOperands() >= 3 &&
+ TII->findCommutedOpIndices(&MI, SrcOp1, SrcOp2)) {
if (SrcIdx == SrcOp1)
regCIdx = SrcOp2;
else if (SrcIdx == SrcOp2)
regCIdx = SrcOp1;
if (regCIdx != ~0U) {
- regC = mi->getOperand(regCIdx).getReg();
- if (!regBKilled && isKilled(*mi, regC, MRI, TII))
+ regC = MI.getOperand(regCIdx).getReg();
+ if (!regBKilled && isKilled(MI, regC, MRI, TII))
// If C dies but B does not, swap the B and C operands.
// This makes the live ranges of A and C joinable.
TryCommute = true;
- else if (isProfitableToCommute(regB, regC, mi, mbbi, Dist)) {
+ else if (isProfitableToCommute(regB, regC, &MI, mbbi, Dist)) {
TryCommute = true;
AggressiveCommute = true;
}
return false;
}
- if (TID.isConvertibleTo3Addr()) {
+ // If there is one more use of regB later in the same MBB, consider
+ // re-schedule this MI below it.
+ if (RescheduleMIBelowKill(mbbi, mi, nmi, regB)) {
+ ++NumReSchedDowns;
+ return true;
+ }
+
+ if (TargetRegisterInfo::isVirtualRegister(regA))
+ ScanUses(regA, &*mbbi, Processed);
+
+ if (MI.isConvertibleTo3Addr()) {
// This instruction is potentially convertible to a true
// three-address instruction. Check if it is profitable.
- if (!regBKilled || isProfitableToConv3Addr(regA)) {
+ if (!regBKilled || isProfitableToConv3Addr(regA, regB)) {
// Try to convert it.
- if (ConvertInstTo3Addr(mi, nmi, mbbi, regB, Dist)) {
+ if (ConvertInstTo3Addr(mi, nmi, mbbi, regA, regB, Dist)) {
++NumConvertedTo3Addr;
return true; // Done with this instruction.
}
}
}
+
+ // If there is one more use of regB later in the same MBB, consider
+ // re-schedule it before this MI if it's legal.
+ if (RescheduleKillAboveMI(mbbi, mi, nmi, regB)) {
+ ++NumReSchedUps;
+ return true;
+ }
+
+ // If this is an instruction with a load folded into it, try unfolding
+ // the load, e.g. avoid this:
+ // movq %rdx, %rcx
+ // addq (%rax), %rcx
+ // in favor of this:
+ // movq (%rax), %rcx
+ // addq %rdx, %rcx
+ // because it's preferable to schedule a load than a register copy.
+ if (MI.mayLoad() && !regBKilled) {
+ // Determine if a load can be unfolded.
+ unsigned LoadRegIndex;
+ unsigned NewOpc =
+ TII->getOpcodeAfterMemoryUnfold(MI.getOpcode(),
+ /*UnfoldLoad=*/true,
+ /*UnfoldStore=*/false,
+ &LoadRegIndex);
+ if (NewOpc != 0) {
+ const MCInstrDesc &UnfoldMCID = TII->get(NewOpc);
+ if (UnfoldMCID.getNumDefs() == 1) {
+ MachineFunction &MF = *mbbi->getParent();
+
+ // Unfold the load.
+ DEBUG(dbgs() << "2addr: UNFOLDING: " << MI);
+ const TargetRegisterClass *RC =
+ TII->getRegClass(UnfoldMCID, LoadRegIndex, TRI);
+ unsigned Reg = MRI->createVirtualRegister(RC);
+ SmallVector<MachineInstr *, 2> NewMIs;
+ if (!TII->unfoldMemoryOperand(MF, &MI, Reg,
+ /*UnfoldLoad=*/true,/*UnfoldStore=*/false,
+ NewMIs)) {
+ DEBUG(dbgs() << "2addr: ABANDONING UNFOLD\n");
+ return false;
+ }
+ assert(NewMIs.size() == 2 &&
+ "Unfolded a load into multiple instructions!");
+ // The load was previously folded, so this is the only use.
+ NewMIs[1]->addRegisterKilled(Reg, TRI);
+
+ // Tentatively insert the instructions into the block so that they
+ // look "normal" to the transformation logic.
+ mbbi->insert(mi, NewMIs[0]);
+ mbbi->insert(mi, NewMIs[1]);
+
+ DEBUG(dbgs() << "2addr: NEW LOAD: " << *NewMIs[0]
+ << "2addr: NEW INST: " << *NewMIs[1]);
+
+ // Transform the instruction, now that it no longer has a load.
+ unsigned NewDstIdx = NewMIs[1]->findRegisterDefOperandIdx(regA);
+ unsigned NewSrcIdx = NewMIs[1]->findRegisterUseOperandIdx(regB);
+ MachineBasicBlock::iterator NewMI = NewMIs[1];
+ bool TransformSuccess =
+ TryInstructionTransform(NewMI, mi, mbbi,
+ NewSrcIdx, NewDstIdx, Dist, Processed);
+ if (TransformSuccess ||
+ NewMIs[1]->getOperand(NewSrcIdx).isKill()) {
+ // Success, or at least we made an improvement. Keep the unfolded
+ // instructions and discard the original.
+ if (LV) {
+ for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
+ MachineOperand &MO = MI.getOperand(i);
+ if (MO.isReg() &&
+ TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
+ if (MO.isUse()) {
+ if (MO.isKill()) {
+ if (NewMIs[0]->killsRegister(MO.getReg()))
+ LV->replaceKillInstruction(MO.getReg(), &MI, NewMIs[0]);
+ else {
+ assert(NewMIs[1]->killsRegister(MO.getReg()) &&
+ "Kill missing after load unfold!");
+ LV->replaceKillInstruction(MO.getReg(), &MI, NewMIs[1]);
+ }
+ }
+ } else if (LV->removeVirtualRegisterDead(MO.getReg(), &MI)) {
+ if (NewMIs[1]->registerDefIsDead(MO.getReg()))
+ LV->addVirtualRegisterDead(MO.getReg(), NewMIs[1]);
+ else {
+ assert(NewMIs[0]->registerDefIsDead(MO.getReg()) &&
+ "Dead flag missing after load unfold!");
+ LV->addVirtualRegisterDead(MO.getReg(), NewMIs[0]);
+ }
+ }
+ }
+ }
+ LV->addVirtualRegisterKilled(Reg, NewMIs[1]);
+ }
+ MI.eraseFromParent();
+ mi = NewMIs[1];
+ if (TransformSuccess)
+ return true;
+ } else {
+ // Transforming didn't eliminate the tie and didn't lead to an
+ // improvement. Clean up the unfolded instructions and keep the
+ // original.
+ DEBUG(dbgs() << "2addr: ABANDONING UNFOLD\n");
+ NewMIs[0]->eraseFromParent();
+ NewMIs[1]->eraseFromParent();
+ }
+ }
+ }
+ }
+
return false;
}
MRI = &MF.getRegInfo();
TII = TM.getInstrInfo();
TRI = TM.getRegisterInfo();
+ InstrItins = TM.getInstrItineraryData();
LV = getAnalysisIfAvailable<LiveVariables>();
AA = &getAnalysis<AliasAnalysis>();
+ OptLevel = TM.getOptLevel();
bool MadeChange = false;
DEBUG(dbgs() << "********** Function: "
<< MF.getFunction()->getName() << '\n');
+ // This pass takes the function out of SSA form.
+ MRI->leaveSSA();
+
// ReMatRegs - Keep track of the registers whose def's are remat'ed.
- BitVector ReMatRegs;
- ReMatRegs.resize(MRI->getLastVirtReg()+1);
+ BitVector ReMatRegs(MRI->getNumVirtRegs());
typedef DenseMap<unsigned, SmallVector<std::pair<unsigned, unsigned>, 4> >
TiedOperandMap;
if (mi->isRegSequence())
RegSequences.push_back(&*mi);
- const TargetInstrDesc &TID = mi->getDesc();
+ const MCInstrDesc &MCID = mi->getDesc();
bool FirstTied = true;
DistanceMap.insert(std::make_pair(mi, ++Dist));
// First scan through all the tied register uses in this instruction
// and record a list of pairs of tied operands for each register.
unsigned NumOps = mi->isInlineAsm()
- ? mi->getNumOperands() : TID.getNumOperands();
+ ? mi->getNumOperands() : MCID.getNumOperands();
for (unsigned SrcIdx = 0; SrcIdx < NumOps; ++SrcIdx) {
unsigned DstIdx = 0;
if (!mi->isRegTiedToDefOperand(SrcIdx, &DstIdx))
"two address instruction invalid");
unsigned regB = mi->getOperand(SrcIdx).getReg();
- TiedOperandMap::iterator OI = TiedOperands.find(regB);
- if (OI == TiedOperands.end()) {
- SmallVector<std::pair<unsigned, unsigned>, 4> TiedPair;
- OI = TiedOperands.insert(std::make_pair(regB, TiedPair)).first;
- }
- OI->second.push_back(std::make_pair(SrcIdx, DstIdx));
+ TiedOperands[regB].push_back(std::make_pair(SrcIdx, DstIdx));
}
// Now iterate over the information collected above.
mi->getOperand(DstIdx).getReg())
break; // Done with this instruction.
- if (TryInstructionTransform(mi, nmi, mbbi, SrcIdx, DstIdx, Dist))
+ if (TryInstructionTransform(mi, nmi, mbbi, SrcIdx, DstIdx, Dist,
+ Processed))
break; // The tied operands have been eliminated.
}
+ bool IsEarlyClobber = false;
bool RemovedKillFlag = false;
bool AllUsesCopied = true;
unsigned LastCopiedReg = 0;
for (unsigned tpi = 0, tpe = TiedPairs.size(); tpi != tpe; ++tpi) {
unsigned SrcIdx = TiedPairs[tpi].first;
unsigned DstIdx = TiedPairs[tpi].second;
- unsigned regA = mi->getOperand(DstIdx).getReg();
+
+ const MachineOperand &DstMO = mi->getOperand(DstIdx);
+ unsigned regA = DstMO.getReg();
+ IsEarlyClobber |= DstMO.isEarlyClobber();
+
// Grab regB from the instruction because it may have changed if the
// instruction was commuted.
regB = mi->getOperand(SrcIdx).getReg();
// If it's safe and profitable, remat the definition instead of
// copying it.
if (DefMI &&
- DefMI->getDesc().isAsCheapAsAMove() &&
+ DefMI->isAsCheapAsAMove() &&
DefMI->isSafeToReMat(TII, AA, regB) &&
isProfitableToReMat(regB, rc, mi, DefMI, mbbi, Dist)){
DEBUG(dbgs() << "2addr: REMATTING : " << *DefMI << "\n");
unsigned regASubIdx = mi->getOperand(DstIdx).getSubReg();
- TII->reMaterialize(*mbbi, mi, regA, regASubIdx, DefMI, TRI);
- ReMatRegs.set(regB);
+ TII->reMaterialize(*mbbi, mi, regA, regASubIdx, DefMI, *TRI);
+ ReMatRegs.set(TargetRegisterInfo::virtReg2Index(regB));
++NumReMats;
} else {
- bool Emitted = TII->copyRegToReg(*mbbi, mi, regA, regB, rc, rc,
- mi->getDebugLoc());
- (void)Emitted;
- assert(Emitted && "Unable to issue a copy instruction!\n");
+ BuildMI(*mbbi, mi, mi->getDebugLoc(), TII->get(TargetOpcode::COPY),
+ regA).addReg(regB);
}
MachineBasicBlock::iterator prevMI = prior(mi);
}
if (AllUsesCopied) {
- // Replace other (un-tied) uses of regB with LastCopiedReg.
- for (unsigned i = 0, e = mi->getNumOperands(); i != e; ++i) {
- MachineOperand &MO = mi->getOperand(i);
- if (MO.isReg() && MO.getReg() == regB && MO.isUse()) {
- if (MO.isKill()) {
- MO.setIsKill(false);
- RemovedKillFlag = true;
+ if (!IsEarlyClobber) {
+ // Replace other (un-tied) uses of regB with LastCopiedReg.
+ for (unsigned i = 0, e = mi->getNumOperands(); i != e; ++i) {
+ MachineOperand &MO = mi->getOperand(i);
+ if (MO.isReg() && MO.getReg() == regB && MO.isUse()) {
+ if (MO.isKill()) {
+ MO.setIsKill(false);
+ RemovedKillFlag = true;
+ }
+ MO.setReg(LastCopiedReg);
}
- MO.setReg(LastCopiedReg);
}
}
}
}
}
-
+
+ // Schedule the source copy / remat inserted to form two-address
+ // instruction. FIXME: Does it matter the distance map may not be
+ // accurate after it's scheduled?
+ TII->scheduleTwoAddrSource(prior(mi), mi, *TRI);
+
MadeChange = true;
DEBUG(dbgs() << "\t\trewrite to:\t" << *mi);
}
+ // Rewrite INSERT_SUBREG as COPY now that we no longer need SSA form.
+ if (mi->isInsertSubreg()) {
+ // From %reg = INSERT_SUBREG %reg, %subreg, subidx
+ // To %reg:subidx = COPY %subreg
+ unsigned SubIdx = mi->getOperand(3).getImm();
+ mi->RemoveOperand(3);
+ assert(mi->getOperand(0).getSubReg() == 0 && "Unexpected subreg idx");
+ mi->getOperand(0).setSubReg(SubIdx);
+ mi->RemoveOperand(1);
+ mi->setDesc(TII->get(TargetOpcode::COPY));
+ DEBUG(dbgs() << "\t\tconvert to:\t" << *mi);
+ }
+
// Clear TiedOperands here instead of at the top of the loop
// since most instructions do not have tied operands.
TiedOperands.clear();
}
// Some remat'ed instructions are dead.
- int VReg = ReMatRegs.find_first();
- while (VReg != -1) {
+ for (int i = ReMatRegs.find_first(); i != -1; i = ReMatRegs.find_next(i)) {
+ unsigned VReg = TargetRegisterInfo::index2VirtReg(i);
if (MRI->use_nodbg_empty(VReg)) {
MachineInstr *DefMI = MRI->getVRegDef(VReg);
DefMI->eraseFromParent();
}
- VReg = ReMatRegs.find_next(VReg);
}
// Eliminate REG_SEQUENCE instructions. Their whole purpose was to preseve
static void UpdateRegSequenceSrcs(unsigned SrcReg,
unsigned DstReg, unsigned SubIdx,
- MachineRegisterInfo *MRI) {
+ MachineRegisterInfo *MRI,
+ const TargetRegisterInfo &TRI) {
for (MachineRegisterInfo::reg_iterator RI = MRI->reg_begin(SrcReg),
RE = MRI->reg_end(); RI != RE; ) {
MachineOperand &MO = RI.getOperand();
++RI;
- MO.setReg(DstReg);
- assert(MO.getSubReg() == 0);
- MO.setSubReg(SubIdx);
+ MO.substVirtReg(DstReg, SubIdx, TRI);
}
}
if (!Seen.insert(SrcReg))
continue;
- // If there are no other uses than extract_subreg which feed into
+ // Check that the instructions are all in the same basic block.
+ MachineInstr *SrcDefMI = MRI->getVRegDef(SrcReg);
+ MachineInstr *DstDefMI = MRI->getVRegDef(DstReg);
+ if (SrcDefMI->getParent() != DstDefMI->getParent())
+ continue;
+
+ // If there are no other uses than copies which feed into
// the reg_sequence, then we might be able to coalesce them.
bool CanCoalesce = true;
- SmallVector<unsigned, 4> SubIndices;
+ SmallVector<unsigned, 4> SrcSubIndices, DstSubIndices;
for (MachineRegisterInfo::use_nodbg_iterator
UI = MRI->use_nodbg_begin(SrcReg),
UE = MRI->use_nodbg_end(); UI != UE; ++UI) {
MachineInstr *UseMI = &*UI;
- if (!UseMI->isExtractSubreg() ||
- UseMI->getOperand(0).getReg() != DstReg) {
+ if (!UseMI->isCopy() || UseMI->getOperand(0).getReg() != DstReg) {
CanCoalesce = false;
break;
}
- SubIndices.push_back(UseMI->getOperand(2).getImm());
+ SrcSubIndices.push_back(UseMI->getOperand(1).getSubReg());
+ DstSubIndices.push_back(UseMI->getOperand(0).getSubReg());
}
- if (!CanCoalesce || SubIndices.size() < 2)
+ if (!CanCoalesce || SrcSubIndices.size() < 2)
continue;
- std::sort(SubIndices.begin(), SubIndices.end());
- unsigned NewSubIdx = 0;
- if (TRI->canCombinedSubRegIndex(MRI->getRegClass(SrcReg), SubIndices,
- NewSubIdx)) {
- bool Proceed = true;
- if (NewSubIdx)
- for (MachineRegisterInfo::reg_iterator RI = MRI->reg_begin(SrcReg),
- RE = MRI->reg_end(); RI != RE; ) {
- MachineOperand &MO = RI.getOperand();
- ++RI;
- // FIXME: If the sub-registers do not combine to the whole
- // super-register, i.e. NewSubIdx != 0, and any of the use has a
- // sub-register index, then abort the coalescing attempt.
- if (MO.getSubReg()) {
- Proceed = false;
- break;
- }
- MO.setReg(DstReg);
- MO.setSubReg(NewSubIdx);
- }
- if (Proceed)
- for (MachineRegisterInfo::reg_iterator RI = MRI->reg_begin(SrcReg),
- RE = MRI->reg_end(); RI != RE; ) {
- MachineOperand &MO = RI.getOperand();
- ++RI;
- MO.setReg(DstReg);
- if (NewSubIdx)
- MO.setSubReg(NewSubIdx);
- }
+ // Check that the source subregisters can be combined.
+ std::sort(SrcSubIndices.begin(), SrcSubIndices.end());
+ unsigned NewSrcSubIdx = 0;
+ if (!TRI->canCombineSubRegIndices(MRI->getRegClass(SrcReg), SrcSubIndices,
+ NewSrcSubIdx))
+ continue;
+
+ // Check that the destination subregisters can also be combined.
+ std::sort(DstSubIndices.begin(), DstSubIndices.end());
+ unsigned NewDstSubIdx = 0;
+ if (!TRI->canCombineSubRegIndices(MRI->getRegClass(DstReg), DstSubIndices,
+ NewDstSubIdx))
+ continue;
+
+ // If neither source nor destination can be combined to the full register,
+ // just give up. This could be improved if it ever matters.
+ if (NewSrcSubIdx != 0 && NewDstSubIdx != 0)
+ continue;
+
+ // Now that we know that all the uses are extract_subregs and that those
+ // subregs can somehow be combined, scan all the extract_subregs again to
+ // make sure the subregs are in the right order and can be composed.
+ MachineInstr *SomeMI = 0;
+ CanCoalesce = true;
+ for (MachineRegisterInfo::use_nodbg_iterator
+ UI = MRI->use_nodbg_begin(SrcReg),
+ UE = MRI->use_nodbg_end(); UI != UE; ++UI) {
+ MachineInstr *UseMI = &*UI;
+ assert(UseMI->isCopy());
+ unsigned DstSubIdx = UseMI->getOperand(0).getSubReg();
+ unsigned SrcSubIdx = UseMI->getOperand(1).getSubReg();
+ assert(DstSubIdx != 0 && "missing subreg from RegSequence elimination");
+ if ((NewDstSubIdx == 0 &&
+ TRI->composeSubRegIndices(NewSrcSubIdx, DstSubIdx) != SrcSubIdx) ||
+ (NewSrcSubIdx == 0 &&
+ TRI->composeSubRegIndices(NewDstSubIdx, SrcSubIdx) != DstSubIdx)) {
+ CanCoalesce = false;
+ break;
+ }
+ // Keep track of one of the uses.
+ SomeMI = UseMI;
+ }
+ if (!CanCoalesce)
+ continue;
+
+ // Insert a copy to replace the original.
+ MachineInstr *CopyMI = BuildMI(*SomeMI->getParent(), SomeMI,
+ SomeMI->getDebugLoc(),
+ TII->get(TargetOpcode::COPY))
+ .addReg(DstReg, RegState::Define, NewDstSubIdx)
+ .addReg(SrcReg, 0, NewSrcSubIdx);
+
+ // Remove all the old extract instructions.
+ for (MachineRegisterInfo::use_nodbg_iterator
+ UI = MRI->use_nodbg_begin(SrcReg),
+ UE = MRI->use_nodbg_end(); UI != UE; ) {
+ MachineInstr *UseMI = &*UI;
+ ++UI;
+ if (UseMI == CopyMI)
+ continue;
+ assert(UseMI->isCopy());
+ // Move any kills to the new copy or extract instruction.
+ if (UseMI->getOperand(1).isKill()) {
+ CopyMI->getOperand(1).setIsKill();
+ if (LV)
+ // Update live variables
+ LV->replaceKillInstruction(SrcReg, UseMI, &*CopyMI);
}
+ UseMI->eraseFromParent();
+ }
+ }
+}
+
+static bool HasOtherRegSequenceUses(unsigned Reg, MachineInstr *RegSeq,
+ MachineRegisterInfo *MRI) {
+ for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(Reg),
+ UE = MRI->use_end(); UI != UE; ++UI) {
+ MachineInstr *UseMI = &*UI;
+ if (UseMI != RegSeq && UseMI->isRegSequence())
+ return true;
}
+ return false;
}
/// EliminateRegSequences - Eliminate REG_SEQUENCE instructions as part
llvm_unreachable(0);
}
+ bool IsImpDef = true;
SmallVector<unsigned, 4> RealSrcs;
SmallSet<unsigned, 4> Seen;
for (unsigned i = 1, e = MI->getNumOperands(); i < e; i += 2) {
unsigned SrcReg = MI->getOperand(i).getReg();
+ unsigned SubIdx = MI->getOperand(i+1).getImm();
if (MI->getOperand(i).getSubReg() ||
TargetRegisterInfo::isPhysicalRegister(SrcReg)) {
DEBUG(dbgs() << "Illegal REG_SEQUENCE instruction:" << *MI);
DefMI->eraseFromParent();
continue;
}
+ IsImpDef = false;
- // Remember EXTRACT_SUBREG sources. These might be candidate for
- // coalescing.
- if (DefMI->isExtractSubreg())
+ // Remember COPY sources. These might be candidate for coalescing.
+ if (DefMI->isCopy() && DefMI->getOperand(1).getSubReg())
RealSrcs.push_back(DefMI->getOperand(1).getReg());
- if (!Seen.insert(SrcReg) || MI->getParent() != DefMI->getParent()) {
+ bool isKill = MI->getOperand(i).isKill();
+ if (!Seen.insert(SrcReg) || MI->getParent() != DefMI->getParent() ||
+ !isKill || HasOtherRegSequenceUses(SrcReg, MI, MRI) ||
+ !TRI->getMatchingSuperRegClass(MRI->getRegClass(DstReg),
+ MRI->getRegClass(SrcReg), SubIdx)) {
// REG_SEQUENCE cannot have duplicated operands, add a copy.
- // Also add an copy if the source if live-in the block. We don't want
+ // Also add an copy if the source is live-in the block. We don't want
// to end up with a partial-redef of a livein, e.g.
// BB0:
// reg1051:10<def> =
// BB2:
// reg1051:9<def> =
// LiveIntervalAnalysis won't like it.
- const TargetRegisterClass *RC = MRI->getRegClass(SrcReg);
- unsigned NewReg = MRI->createVirtualRegister(RC);
+ //
+ // If the REG_SEQUENCE doesn't kill its source, keeping live variables
+ // correctly up to date becomes very difficult. Insert a copy.
+
+ // Defer any kill flag to the last operand using SrcReg. Otherwise, we
+ // might insert a COPY that uses SrcReg after is was killed.
+ if (isKill)
+ for (unsigned j = i + 2; j < e; j += 2)
+ if (MI->getOperand(j).getReg() == SrcReg) {
+ MI->getOperand(j).setIsKill();
+ isKill = false;
+ break;
+ }
+
MachineBasicBlock::iterator InsertLoc = MI;
- bool Emitted =
- TII->copyRegToReg(*MI->getParent(), InsertLoc, NewReg, SrcReg, RC, RC,
- MI->getDebugLoc());
- (void)Emitted;
- assert(Emitted && "Unable to issue a copy instruction!\n");
- MI->getOperand(i).setReg(NewReg);
- if (MI->getOperand(i).isKill()) {
- MachineBasicBlock::iterator CopyMI = prior(InsertLoc);
- MachineOperand *KillMO = CopyMI->findRegisterUseOperand(SrcReg);
- KillMO->setIsKill();
- if (LV)
- // Update live variables
- LV->replaceKillInstruction(SrcReg, MI, &*CopyMI);
- }
+ MachineInstr *CopyMI = BuildMI(*MI->getParent(), InsertLoc,
+ MI->getDebugLoc(), TII->get(TargetOpcode::COPY))
+ .addReg(DstReg, RegState::Define, SubIdx)
+ .addReg(SrcReg, getKillRegState(isKill));
+ MI->getOperand(i).setReg(0);
+ if (LV && isKill)
+ LV->replaceKillInstruction(SrcReg, MI, CopyMI);
+ DEBUG(dbgs() << "Inserted: " << *CopyMI);
}
}
for (unsigned i = 1, e = MI->getNumOperands(); i < e; i += 2) {
unsigned SrcReg = MI->getOperand(i).getReg();
+ if (!SrcReg) continue;
unsigned SubIdx = MI->getOperand(i+1).getImm();
- UpdateRegSequenceSrcs(SrcReg, DstReg, SubIdx, MRI);
+ UpdateRegSequenceSrcs(SrcReg, DstReg, SubIdx, MRI, *TRI);
}
- DEBUG(dbgs() << "Eliminated: " << *MI);
- MI->eraseFromParent();
+ if (IsImpDef) {
+ DEBUG(dbgs() << "Turned: " << *MI << " into an IMPLICIT_DEF");
+ MI->setDesc(TII->get(TargetOpcode::IMPLICIT_DEF));
+ for (int j = MI->getNumOperands() - 1, ee = 0; j > ee; --j)
+ MI->RemoveOperand(j);
+ } else {
+ DEBUG(dbgs() << "Eliminated: " << *MI);
+ MI->eraseFromParent();
+ }
- // Try coalescing some EXTRACT_SUBREG instructions.
- CoalesceExtSubRegs(RealSrcs, DstReg);
+ // Try coalescing some EXTRACT_SUBREG instructions. This can create
+ // INSERT_SUBREG instructions that must have <undef> flags added by
+ // LiveIntervalAnalysis, so only run it when LiveVariables is available.
+ if (LV)
+ CoalesceExtSubRegs(RealSrcs, DstReg);
}
RegSequences.clear();