void FastISel::startNewBlock() {
LocalValueMap.clear();
- // Start out as null, meaining no local-value instructions have
- // been emitted.
- LastLocalValue = 0;
+ EmitStartPt = 0;
- // Advance the last local value past any EH_LABEL instructions.
+ // Advance the emit start point past any EH_LABEL instructions.
MachineBasicBlock::iterator
I = FuncInfo.MBB->begin(), E = FuncInfo.MBB->end();
while (I != E && I->getOpcode() == TargetOpcode::EH_LABEL) {
- LastLocalValue = I;
+ EmitStartPt = I;
++I;
}
+ LastLocalValue = EmitStartPt;
+}
+
+void FastISel::flushLocalValueMap() {
+ LocalValueMap.clear();
+ LastLocalValue = EmitStartPt;
+ recomputeInsertPt();
}
bool FastISel::hasTrivialKill(const Value *V) const {
// of whether FastISel can handle them.
MVT VT = RealVT.getSimpleVT();
if (!TLI.isTypeLegal(VT)) {
- // Promote MVT::i1 to a legal type though, because it's common and easy.
- if (VT == MVT::i1)
+ // Handle integer promotions, though, because they're common and easy.
+ if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
VT = TLI.getTypeToTransformTo(V->getContext(), VT).getSimpleVT();
else
return 0;
(void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
APFloat::rmTowardZero, &isExact);
if (isExact) {
- APInt IntVal(IntBitWidth, 2, x);
+ APInt IntVal(IntBitWidth, x);
unsigned IntegerReg =
getRegForValue(ConstantInt::get(V->getContext(), IntVal));
/// NOTE: This is only necessary because we might select a block that uses
/// a value before we select the block that defines the value. It might be
/// possible to fix this by selecting blocks in reverse postorder.
-unsigned FastISel::UpdateValueMap(const Value *I, unsigned Reg) {
+void FastISel::UpdateValueMap(const Value *I, unsigned Reg, unsigned NumRegs) {
if (!isa<Instruction>(I)) {
LocalValueMap[I] = Reg;
- return Reg;
+ return;
}
unsigned &AssignedReg = FuncInfo.ValueMap[I];
AssignedReg = Reg;
else if (Reg != AssignedReg) {
// Arrange for uses of AssignedReg to be replaced by uses of Reg.
- FuncInfo.RegFixups[AssignedReg] = Reg;
+ for (unsigned i = 0; i < NumRegs; i++)
+ FuncInfo.RegFixups[AssignedReg+i] = Reg+i;
AssignedReg = Reg;
}
-
- return AssignedReg;
}
std::pair<unsigned, bool> FastISel::getRegForGEPIndex(const Value *Idx) {
bool NIsKill = hasTrivialKill(I->getOperand(0));
- const Type *Ty = I->getOperand(0)->getType();
+ Type *Ty = I->getOperand(0)->getType();
MVT VT = TLI.getPointerTy();
for (GetElementPtrInst::const_op_iterator OI = I->op_begin()+1,
E = I->op_end(); OI != E; ++OI) {
const Value *Idx = *OI;
- if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
+ if (StructType *StTy = dyn_cast<StructType>(Ty)) {
unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
if (Field) {
// N = N + Offset
unsigned Reg = 0;
unsigned Offset = 0;
if (const Argument *Arg = dyn_cast<Argument>(Address)) {
- if (Arg->hasByValAttr()) {
- // Byval arguments' frame index is recorded during argument lowering.
- // Use this info directly.
- Offset = FuncInfo.getByValArgumentFrameIndex(Arg);
- if (Offset)
- Reg = TRI.getFrameRegister(*FuncInfo.MF);
- }
+ // Some arguments' frame index is recorded during argument lowering.
+ Offset = FuncInfo.getArgumentFrameIndex(Arg);
+ if (Offset)
+ Reg = TRI.getFrameRegister(*FuncInfo.MF);
}
if (!Reg)
Reg = getRegForValue(Address);
case Intrinsic::dbg_value: {
// This form of DBG_VALUE is target-independent.
const DbgValueInst *DI = cast<DbgValueInst>(Call);
- const TargetInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
+ const MCInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
const Value *V = DI->getValue();
if (!V) {
// Currently the optimizer can produce this; insert an undef to
.addReg(0U).addImm(DI->getOffset())
.addMetadata(DI->getVariable());
} else if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
- .addImm(CI->getZExtValue()).addImm(DI->getOffset())
- .addMetadata(DI->getVariable());
+ if (CI->getBitWidth() > 64)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
+ .addCImm(CI).addImm(DI->getOffset())
+ .addMetadata(DI->getVariable());
+ else
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
+ .addImm(CI->getZExtValue()).addImm(DI->getOffset())
+ .addMetadata(DI->getVariable());
} else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
.addFPImm(CF).addImm(DI->getOffset())
}
}
+ // Usually, it does not make sense to initialize a value,
+ // make an unrelated function call and use the value, because
+ // it tends to be spilled on the stack. So, we move the pointer
+ // to the last local value to the beginning of the block, so that
+ // all the values which have already been materialized,
+ // appear after the call. It also makes sense to skip intrinsics
+ // since they tend to be inlined.
+ if (!isa<IntrinsicInst>(F))
+ flushLocalValueMap();
+
// An arbitrary call. Bail.
return false;
}
// Unhandled type. Halt "fast" selection and bail.
return false;
- // Check if the destination type is legal. Or as a special case,
- // it may be i1 if we're doing a truncate because that's
- // easy and somewhat common.
+ // Check if the destination type is legal.
if (!TLI.isTypeLegal(DstVT))
- if (DstVT != MVT::i1 || Opcode != ISD::TRUNCATE)
- // Unhandled type. Halt "fast" selection and bail.
- return false;
+ return false;
- // Check if the source operand is legal. Or as a special case,
- // it may be i1 if we're doing zero-extension because that's
- // easy and somewhat common.
+ // Check if the source operand is legal.
if (!TLI.isTypeLegal(SrcVT))
- if (SrcVT != MVT::i1 || Opcode != ISD::ZERO_EXTEND)
- // Unhandled type. Halt "fast" selection and bail.
- return false;
+ return false;
unsigned InputReg = getRegForValue(I->getOperand(0));
if (!InputReg)
bool InputRegIsKill = hasTrivialKill(I->getOperand(0));
- // If the operand is i1, arrange for the high bits in the register to be zero.
- if (SrcVT == MVT::i1) {
- SrcVT = TLI.getTypeToTransformTo(I->getContext(), SrcVT);
- InputReg = FastEmitZExtFromI1(SrcVT.getSimpleVT(), InputReg, InputRegIsKill);
- if (!InputReg)
- return false;
- InputRegIsKill = true;
- }
- // If the result is i1, truncate to the target's type for i1 first.
- if (DstVT == MVT::i1)
- DstVT = TLI.getTypeToTransformTo(I->getContext(), DstVT);
-
unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(),
DstVT.getSimpleVT(),
Opcode,
if (!EVI)
return false;
- // Make sure we only try to handle extracts with a legal result.
+ // Make sure we only try to handle extracts with a legal result. But also
+ // allow i1 because it's easy.
EVT RealVT = TLI.getValueType(EVI->getType(), /*AllowUnknown=*/true);
if (!RealVT.isSimple())
return false;
MVT VT = RealVT.getSimpleVT();
- if (!TLI.isTypeLegal(VT))
+ if (!TLI.isTypeLegal(VT) && VT != MVT::i1)
return false;
const Value *Op0 = EVI->getOperand(0);
- const Type *AggTy = Op0->getType();
+ Type *AggTy = Op0->getType();
// Get the base result register.
unsigned ResultReg;
DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(Op0);
if (I != FuncInfo.ValueMap.end())
ResultReg = I->second;
- else
+ else if (isa<Instruction>(Op0))
ResultReg = FuncInfo.InitializeRegForValue(Op0);
+ else
+ return false; // fast-isel can't handle aggregate constants at the moment
// Get the actual result register, which is an offset from the base register.
- unsigned VTIndex = ComputeLinearIndex(AggTy, EVI->idx_begin(), EVI->idx_end());
+ unsigned VTIndex = ComputeLinearIndex(AggTy, EVI->getIndices());
SmallVector<EVT, 4> AggValueVTs;
ComputeValueVTs(TLI, AggTy, AggValueVTs);
if (MaterialReg == 0) {
// This is a bit ugly/slow, but failing here means falling out of
// fast-isel, which would be very slow.
- const IntegerType *ITy = IntegerType::get(FuncInfo.Fn->getContext(),
+ IntegerType *ITy = IntegerType::get(FuncInfo.Fn->getContext(),
VT.getSizeInBits());
MaterialReg = getRegForValue(ConstantInt::get(ITy, Imm));
}
unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode,
const TargetRegisterClass* RC) {
unsigned ResultReg = createResultReg(RC);
- const TargetInstrDesc &II = TII.get(MachineInstOpcode);
+ const MCInstrDesc &II = TII.get(MachineInstOpcode);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg);
return ResultReg;
const TargetRegisterClass *RC,
unsigned Op0, bool Op0IsKill) {
unsigned ResultReg = createResultReg(RC);
- const TargetInstrDesc &II = TII.get(MachineInstOpcode);
+ const MCInstrDesc &II = TII.get(MachineInstOpcode);
if (II.getNumDefs() >= 1)
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
unsigned Op0, bool Op0IsKill,
unsigned Op1, bool Op1IsKill) {
unsigned ResultReg = createResultReg(RC);
- const TargetInstrDesc &II = TII.get(MachineInstOpcode);
+ const MCInstrDesc &II = TII.get(MachineInstOpcode);
if (II.getNumDefs() >= 1)
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
unsigned Op1, bool Op1IsKill,
unsigned Op2, bool Op2IsKill) {
unsigned ResultReg = createResultReg(RC);
- const TargetInstrDesc &II = TII.get(MachineInstOpcode);
+ const MCInstrDesc &II = TII.get(MachineInstOpcode);
if (II.getNumDefs() >= 1)
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
unsigned Op0, bool Op0IsKill,
uint64_t Imm) {
unsigned ResultReg = createResultReg(RC);
- const TargetInstrDesc &II = TII.get(MachineInstOpcode);
+ const MCInstrDesc &II = TII.get(MachineInstOpcode);
if (II.getNumDefs() >= 1)
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
unsigned Op0, bool Op0IsKill,
uint64_t Imm1, uint64_t Imm2) {
unsigned ResultReg = createResultReg(RC);
- const TargetInstrDesc &II = TII.get(MachineInstOpcode);
+ const MCInstrDesc &II = TII.get(MachineInstOpcode);
if (II.getNumDefs() >= 1)
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
unsigned Op0, bool Op0IsKill,
const ConstantFP *FPImm) {
unsigned ResultReg = createResultReg(RC);
- const TargetInstrDesc &II = TII.get(MachineInstOpcode);
+ const MCInstrDesc &II = TII.get(MachineInstOpcode);
if (II.getNumDefs() >= 1)
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
unsigned Op1, bool Op1IsKill,
uint64_t Imm) {
unsigned ResultReg = createResultReg(RC);
- const TargetInstrDesc &II = TII.get(MachineInstOpcode);
+ const MCInstrDesc &II = TII.get(MachineInstOpcode);
if (II.getNumDefs() >= 1)
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
const TargetRegisterClass *RC,
uint64_t Imm) {
unsigned ResultReg = createResultReg(RC);
- const TargetInstrDesc &II = TII.get(MachineInstOpcode);
+ const MCInstrDesc &II = TII.get(MachineInstOpcode);
if (II.getNumDefs() >= 1)
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg).addImm(Imm);
const TargetRegisterClass *RC,
uint64_t Imm1, uint64_t Imm2) {
unsigned ResultReg = createResultReg(RC);
- const TargetInstrDesc &II = TII.get(MachineInstOpcode);
+ const MCInstrDesc &II = TII.get(MachineInstOpcode);
if (II.getNumDefs() >= 1)
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)