// If the "sub" instruction all ready sets (or could be modified to set) the
// same flag that the "cmp" instruction sets and that "bz" uses, then we can
// eliminate the "cmp" instruction.
-//
+//
+// Another instance, in this code:
+//
+// sub r1, r3 | sub r1, imm
+// cmp r3, r1 or cmp r1, r3 | cmp r1, imm
+// bge L1
+//
+// If the branch instruction can use flag from "sub", then we can replace
+// "sub" with "subs" and eliminate the "cmp" instruction.
+//
+// - Optimize Bitcast pairs:
+//
+// v1 = bitcast v0
+// v2 = bitcast v1
+// = v2
+// =>
+// v1 = bitcast v0
+// = v0
+//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "peephole-opt"
cl::desc("Disable the peephole optimizer"));
STATISTIC(NumReuse, "Number of extension results reused");
-STATISTIC(NumEliminated, "Number of compares eliminated");
-STATISTIC(NumImmFold, "Number of move immediate foled");
+STATISTIC(NumBitcasts, "Number of bitcasts eliminated");
+STATISTIC(NumCmps, "Number of compares eliminated");
+STATISTIC(NumImmFold, "Number of move immediate folded");
namespace {
class PeepholeOptimizer : public MachineFunctionPass {
}
private:
- bool OptimizeCmpInstr(MachineInstr *MI, MachineBasicBlock *MBB);
- bool OptimizeExtInstr(MachineInstr *MI, MachineBasicBlock *MBB,
+ bool optimizeBitcastInstr(MachineInstr *MI, MachineBasicBlock *MBB);
+ bool optimizeCmpInstr(MachineInstr *MI, MachineBasicBlock *MBB);
+ bool optimizeExtInstr(MachineInstr *MI, MachineBasicBlock *MBB,
SmallPtrSet<MachineInstr*, 8> &LocalMIs);
bool isMoveImmediate(MachineInstr *MI,
SmallSet<unsigned, 4> &ImmDefRegs,
DenseMap<unsigned, MachineInstr*> &ImmDefMIs);
- bool FoldImmediate(MachineInstr *MI, MachineBasicBlock *MBB,
+ bool foldImmediate(MachineInstr *MI, MachineBasicBlock *MBB,
SmallSet<unsigned, 4> &ImmDefRegs,
DenseMap<unsigned, MachineInstr*> &ImmDefMIs);
};
}
char PeepholeOptimizer::ID = 0;
+char &llvm::PeepholeOptimizerID = PeepholeOptimizer::ID;
INITIALIZE_PASS_BEGIN(PeepholeOptimizer, "peephole-opts",
"Peephole Optimizations", false, false)
INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
INITIALIZE_PASS_END(PeepholeOptimizer, "peephole-opts",
"Peephole Optimizations", false, false)
-FunctionPass *llvm::createPeepholeOptimizerPass() {
- return new PeepholeOptimizer();
-}
-
-/// OptimizeExtInstr - If instruction is a copy-like instruction, i.e. it reads
+/// optimizeExtInstr - If instruction is a copy-like instruction, i.e. it reads
/// a single register and writes a single register and it does not modify the
/// source, and if the source value is preserved as a sub-register of the
/// result, then replace all reachable uses of the source with the subreg of the
/// result.
-///
+///
/// Do not generate an EXTRACT that is used only in a debug use, as this changes
/// the code. Since this code does not currently share EXTRACTs, just ignore all
/// debug uses.
bool PeepholeOptimizer::
-OptimizeExtInstr(MachineInstr *MI, MachineBasicBlock *MBB,
+optimizeExtInstr(MachineInstr *MI, MachineBasicBlock *MBB,
SmallPtrSet<MachineInstr*, 8> &LocalMIs) {
unsigned SrcReg, DstReg, SubIdx;
if (!TII->isCoalescableExtInstr(*MI, SrcReg, DstReg, SubIdx))
return false;
-
+
if (TargetRegisterInfo::isPhysicalRegister(DstReg) ||
TargetRegisterInfo::isPhysicalRegister(SrcReg))
return false;
- MachineRegisterInfo::use_nodbg_iterator UI = MRI->use_nodbg_begin(SrcReg);
- if (++UI == MRI->use_nodbg_end())
+ if (MRI->hasOneNonDBGUse(SrcReg))
// No other uses.
return false;
+ // Ensure DstReg can get a register class that actually supports
+ // sub-registers. Don't change the class until we commit.
+ const TargetRegisterClass *DstRC = MRI->getRegClass(DstReg);
+ DstRC = TM->getRegisterInfo()->getSubClassWithSubReg(DstRC, SubIdx);
+ if (!DstRC)
+ return false;
+
+ // The ext instr may be operating on a sub-register of SrcReg as well.
+ // PPC::EXTSW is a 32 -> 64-bit sign extension, but it reads a 64-bit
+ // register.
+ // If UseSrcSubIdx is Set, SubIdx also applies to SrcReg, and only uses of
+ // SrcReg:SubIdx should be replaced.
+ bool UseSrcSubIdx = TM->getRegisterInfo()->
+ getSubClassWithSubReg(MRI->getRegClass(SrcReg), SubIdx) != 0;
+
// The source has other uses. See if we can replace the other uses with use of
// the result of the extension.
SmallPtrSet<MachineBasicBlock*, 4> ReachedBBs;
- UI = MRI->use_nodbg_begin(DstReg);
- for (MachineRegisterInfo::use_nodbg_iterator UE = MRI->use_nodbg_end();
+ for (MachineRegisterInfo::use_nodbg_iterator
+ UI = MRI->use_nodbg_begin(DstReg), UE = MRI->use_nodbg_end();
UI != UE; ++UI)
ReachedBBs.insert(UI->getParent());
SmallVector<MachineOperand*, 8> ExtendedUses;
bool ExtendLife = true;
- UI = MRI->use_nodbg_begin(SrcReg);
- for (MachineRegisterInfo::use_nodbg_iterator UE = MRI->use_nodbg_end();
+ for (MachineRegisterInfo::use_nodbg_iterator
+ UI = MRI->use_nodbg_begin(SrcReg), UE = MRI->use_nodbg_end();
UI != UE; ++UI) {
MachineOperand &UseMO = UI.getOperand();
MachineInstr *UseMI = &*UI;
continue;
}
+ // Only accept uses of SrcReg:SubIdx.
+ if (UseSrcSubIdx && UseMO.getSubReg() != SubIdx)
+ continue;
+
// It's an error to translate this:
//
// %reg1025 = <sext> %reg1024
// Look for PHI uses of the extended result, we don't want to extend the
// liveness of a PHI input. It breaks all kinds of assumptions down
// stream. A PHI use is expected to be the kill of its source values.
- UI = MRI->use_nodbg_begin(DstReg);
for (MachineRegisterInfo::use_nodbg_iterator
- UE = MRI->use_nodbg_end(); UI != UE; ++UI)
+ UI = MRI->use_nodbg_begin(DstReg), UE = MRI->use_nodbg_end();
+ UI != UE; ++UI)
if (UI->isPHI())
PHIBBs.insert(UI->getParent());
if (PHIBBs.count(UseMBB))
continue;
+ // About to add uses of DstReg, clear DstReg's kill flags.
+ if (!Changed) {
+ MRI->clearKillFlags(DstReg);
+ MRI->constrainRegClass(DstReg, DstRC);
+ }
+
unsigned NewVR = MRI->createVirtualRegister(RC);
- BuildMI(*UseMBB, UseMI, UseMI->getDebugLoc(),
- TII->get(TargetOpcode::COPY), NewVR)
+ MachineInstr *Copy = BuildMI(*UseMBB, UseMI, UseMI->getDebugLoc(),
+ TII->get(TargetOpcode::COPY), NewVR)
.addReg(DstReg, 0, SubIdx);
-
+ // SubIdx applies to both SrcReg and DstReg when UseSrcSubIdx is set.
+ if (UseSrcSubIdx) {
+ Copy->getOperand(0).setSubReg(SubIdx);
+ Copy->getOperand(0).setIsUndef();
+ }
UseMO->setReg(NewVR);
++NumReuse;
Changed = true;
return Changed;
}
-/// OptimizeCmpInstr - If the instruction is a compare and the previous
+/// optimizeBitcastInstr - If the instruction is a bitcast instruction A that
+/// cannot be optimized away during isel (e.g. ARM::VMOVSR, which bitcast
+/// a value cross register classes), and the source is defined by another
+/// bitcast instruction B. And if the register class of source of B matches
+/// the register class of instruction A, then it is legal to replace all uses
+/// of the def of A with source of B. e.g.
+/// %vreg0<def> = VMOVSR %vreg1
+/// %vreg3<def> = VMOVRS %vreg0
+/// Replace all uses of vreg3 with vreg1.
+
+bool PeepholeOptimizer::optimizeBitcastInstr(MachineInstr *MI,
+ MachineBasicBlock *MBB) {
+ unsigned NumDefs = MI->getDesc().getNumDefs();
+ unsigned NumSrcs = MI->getDesc().getNumOperands() - NumDefs;
+ if (NumDefs != 1)
+ return false;
+
+ unsigned Def = 0;
+ unsigned Src = 0;
+ for (unsigned i = 0, e = NumDefs + NumSrcs; i != e; ++i) {
+ const MachineOperand &MO = MI->getOperand(i);
+ if (!MO.isReg())
+ continue;
+ unsigned Reg = MO.getReg();
+ if (!Reg)
+ continue;
+ if (MO.isDef())
+ Def = Reg;
+ else if (Src)
+ // Multiple sources?
+ return false;
+ else
+ Src = Reg;
+ }
+
+ assert(Def && Src && "Malformed bitcast instruction!");
+
+ MachineInstr *DefMI = MRI->getVRegDef(Src);
+ if (!DefMI || !DefMI->isBitcast())
+ return false;
+
+ unsigned SrcSrc = 0;
+ NumDefs = DefMI->getDesc().getNumDefs();
+ NumSrcs = DefMI->getDesc().getNumOperands() - NumDefs;
+ if (NumDefs != 1)
+ return false;
+ for (unsigned i = 0, e = NumDefs + NumSrcs; i != e; ++i) {
+ const MachineOperand &MO = DefMI->getOperand(i);
+ if (!MO.isReg() || MO.isDef())
+ continue;
+ unsigned Reg = MO.getReg();
+ if (!Reg)
+ continue;
+ if (!MO.isDef()) {
+ if (SrcSrc)
+ // Multiple sources?
+ return false;
+ else
+ SrcSrc = Reg;
+ }
+ }
+
+ if (MRI->getRegClass(SrcSrc) != MRI->getRegClass(Def))
+ return false;
+
+ MRI->replaceRegWith(Def, SrcSrc);
+ MRI->clearKillFlags(SrcSrc);
+ MI->eraseFromParent();
+ ++NumBitcasts;
+ return true;
+}
+
+/// optimizeCmpInstr - If the instruction is a compare and the previous
/// instruction it's comparing against all ready sets (or could be modified to
/// set) the same flag as the compare, then we can remove the comparison and use
/// the flag from the previous instruction.
-bool PeepholeOptimizer::OptimizeCmpInstr(MachineInstr *MI,
- MachineBasicBlock *MBB){
+bool PeepholeOptimizer::optimizeCmpInstr(MachineInstr *MI,
+ MachineBasicBlock *MBB) {
// If this instruction is a comparison against zero and isn't comparing a
// physical register, we can try to optimize it.
unsigned SrcReg;
// Attempt to optimize the comparison instruction.
if (TII->OptimizeCompareInstr(MI, SrcReg, CmpMask, CmpValue, MRI)) {
- ++NumEliminated;
+ ++NumCmps;
return true;
}
bool PeepholeOptimizer::isMoveImmediate(MachineInstr *MI,
SmallSet<unsigned, 4> &ImmDefRegs,
DenseMap<unsigned, MachineInstr*> &ImmDefMIs) {
- const TargetInstrDesc &TID = MI->getDesc();
- if (!TID.isMoveImmediate())
+ const MCInstrDesc &MCID = MI->getDesc();
+ if (!MI->isMoveImmediate())
return false;
- if (TID.getNumDefs() != 1)
+ if (MCID.getNumDefs() != 1)
return false;
unsigned Reg = MI->getOperand(0).getReg();
if (TargetRegisterInfo::isVirtualRegister(Reg)) {
ImmDefRegs.insert(Reg);
return true;
}
-
+
return false;
}
-/// FoldImmediate - Try folding register operands that are defined by move
+/// foldImmediate - Try folding register operands that are defined by move
/// immediate instructions, i.e. a trivial constant folding optimization, if
/// and only if the def and use are in the same BB.
-bool PeepholeOptimizer::FoldImmediate(MachineInstr *MI, MachineBasicBlock *MBB,
+bool PeepholeOptimizer::foldImmediate(MachineInstr *MI, MachineBasicBlock *MBB,
SmallSet<unsigned, 4> &ImmDefRegs,
DenseMap<unsigned, MachineInstr*> &ImmDefMIs) {
for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
if (!MO.isReg() || MO.isDef())
continue;
unsigned Reg = MO.getReg();
- if (!Reg || TargetRegisterInfo::isPhysicalRegister(Reg))
+ if (!TargetRegisterInfo::isVirtualRegister(Reg))
continue;
if (ImmDefRegs.count(Reg) == 0)
continue;
bool PeepholeOptimizer::runOnMachineFunction(MachineFunction &MF) {
if (DisablePeephole)
return false;
-
+
TM = &MF.getTarget();
TII = TM->getInstrInfo();
MRI = &MF.getRegInfo();
DenseMap<unsigned, MachineInstr*> ImmDefMIs;
for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
MachineBasicBlock *MBB = &*I;
-
+
bool SeenMoveImm = false;
LocalMIs.clear();
ImmDefRegs.clear();
ImmDefMIs.clear();
+ bool First = true;
+ MachineBasicBlock::iterator PMII;
for (MachineBasicBlock::iterator
MII = I->begin(), MIE = I->end(); MII != MIE; ) {
- MachineInstr *MI = &*MII++;
+ MachineInstr *MI = &*MII;
LocalMIs.insert(MI);
- if (MI->getDesc().hasUnmodeledSideEffects())
+ if (MI->isLabel() || MI->isPHI() || MI->isImplicitDef() ||
+ MI->isKill() || MI->isInlineAsm() || MI->isDebugValue() ||
+ MI->hasUnmodeledSideEffects()) {
+ ++MII;
continue;
+ }
- if (MI->getDesc().isCompare()) {
- Changed |= OptimizeCmpInstr(MI, MBB);
- } else if (isMoveImmediate(MI, ImmDefRegs, ImmDefMIs)) {
+ if (MI->isBitcast()) {
+ if (optimizeBitcastInstr(MI, MBB)) {
+ // MI is deleted.
+ LocalMIs.erase(MI);
+ Changed = true;
+ MII = First ? I->begin() : llvm::next(PMII);
+ continue;
+ }
+ } else if (MI->isCompare()) {
+ if (optimizeCmpInstr(MI, MBB)) {
+ // MI is deleted.
+ LocalMIs.erase(MI);
+ Changed = true;
+ MII = First ? I->begin() : llvm::next(PMII);
+ continue;
+ }
+ }
+
+ if (isMoveImmediate(MI, ImmDefRegs, ImmDefMIs)) {
SeenMoveImm = true;
} else {
- Changed |= OptimizeExtInstr(MI, MBB, LocalMIs);
+ Changed |= optimizeExtInstr(MI, MBB, LocalMIs);
if (SeenMoveImm)
- Changed |= FoldImmediate(MI, MBB, ImmDefRegs, ImmDefMIs);
+ Changed |= foldImmediate(MI, MBB, ImmDefRegs, ImmDefMIs);
}
+
+ First = false;
+ PMII = MII;
+ ++MII;
}
}
projects / oota-llvm.git / blobdiff