//===----------------------------------------------------------------------===//
#include "AArch64.h"
+#include "AArch64CallingConvention.h"
#include "AArch64Subtarget.h"
#include "AArch64TargetMachine.h"
#include "MCTargetDesc/AArch64AddressingModes.h"
return Base.Reg;
}
void setOffsetReg(unsigned Reg) {
- assert(isRegBase() && "Invalid offset register access!");
OffsetReg = Reg;
}
unsigned getOffsetReg() const {
- assert(isRegBase() && "Invalid offset register access!");
return OffsetReg;
}
void setFI(unsigned FI) {
bool selectBitCast(const Instruction *I);
bool selectFRem(const Instruction *I);
bool selectSDiv(const Instruction *I);
+ bool selectGetElementPtr(const Instruction *I);
// Utility helper routines.
bool isTypeLegal(Type *Ty, MVT &VT);
bool foldXALUIntrinsic(AArch64CC::CondCode &CC, const Instruction *I,
const Value *Cond);
bool optimizeIntExtLoad(const Instruction *I, MVT RetVT, MVT SrcVT);
+ bool optimizeSelect(const SelectInst *SI);
+ std::pair<unsigned, bool> getRegForGEPIndex(const Value *Idx);
// Emit helper routines.
unsigned emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS,
unsigned emitAdd(MVT RetVT, const Value *LHS, const Value *RHS,
bool SetFlags = false, bool WantResult = true,
bool IsZExt = false);
+ unsigned emitAdd_ri_(MVT VT, unsigned Op0, bool Op0IsKill, int64_t Imm);
unsigned emitSub(MVT RetVT, const Value *LHS, const Value *RHS,
bool SetFlags = false, bool WantResult = true,
bool IsZExt = false);
unsigned fastMaterializeFloatZero(const ConstantFP* CF) override;
explicit AArch64FastISel(FunctionLoweringInfo &FuncInfo,
- const TargetLibraryInfo *LibInfo)
+ const TargetLibraryInfo *LibInfo)
: FastISel(FuncInfo, LibInfo, /*SkipTargetIndependentISel=*/true) {
- Subtarget = &TM.getSubtarget<AArch64Subtarget>();
+ Subtarget =
+ &static_cast<const AArch64Subtarget &>(FuncInfo.MF->getSubtarget());
Context = &FuncInfo.Fn->getContext();
}
CCAssignFn *AArch64FastISel::CCAssignFnForCall(CallingConv::ID CC) const {
if (CC == CallingConv::WebKit_JS)
return CC_AArch64_WebKit_JS;
+ if (CC == CallingConv::GHC)
+ return CC_AArch64_GHC;
return Subtarget->isTargetDarwin() ? CC_AArch64_DarwinPCS : CC_AArch64_AAPCS;
}
return fastEmitInst_i(Opc, TLI.getRegClassFor(VT), Imm);
}
+ // For the MachO large code model materialize the FP constant in code.
+ if (Subtarget->isTargetMachO() && TM.getCodeModel() == CodeModel::Large) {
+ unsigned Opc1 = Is64Bit ? AArch64::MOVi64imm : AArch64::MOVi32imm;
+ const TargetRegisterClass *RC = Is64Bit ?
+ &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
+
+ unsigned TmpReg = createResultReg(RC);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc1), TmpReg)
+ .addImm(CFP->getValueAPF().bitcastToAPInt().getZExtValue());
+
+ unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), ResultReg)
+ .addReg(TmpReg, getKillRegState(true));
+
+ return ResultReg;
+ }
+
// Materialize via constant pool. MachineConstantPool wants an explicit
// alignment.
unsigned Align = DL.getPrefTypeAlignment(CFP->getType());
if (Addr.getOffsetReg())
break;
- if (DL.getTypeSizeInBits(Ty) != 8)
+ if (!Ty || DL.getTypeSizeInBits(Ty) != 8)
break;
const Value *LHS = U->getOperand(0);
}
} // end switch
- if (Addr.getReg()) {
- if (!Addr.getOffsetReg()) {
- unsigned Reg = getRegForValue(Obj);
- if (!Reg)
- return false;
- Addr.setOffsetReg(Reg);
- return true;
- }
- return false;
+ if (Addr.isRegBase() && !Addr.getReg()) {
+ unsigned Reg = getRegForValue(Obj);
+ if (!Reg)
+ return false;
+ Addr.setReg(Reg);
+ return true;
}
- unsigned Reg = getRegForValue(Obj);
- if (!Reg)
- return false;
- Addr.setReg(Reg);
- return true;
+ if (!Addr.getOffsetReg()) {
+ unsigned Reg = getRegForValue(Obj);
+ if (!Reg)
+ return false;
+ Addr.setOffsetReg(Reg);
+ return true;
+ }
+
+ return false;
}
bool AArch64FastISel::computeCallAddress(const Value *V, Address &Addr) {
// Cannot encode an offset register and an immediate offset in the same
// instruction. Fold the immediate offset into the load/store instruction and
// emit an additonal add to take care of the offset register.
- if (!ImmediateOffsetNeedsLowering && Addr.getOffset() && Addr.isRegBase() &&
- Addr.getOffsetReg())
+ if (!ImmediateOffsetNeedsLowering && Addr.getOffset() && Addr.getOffsetReg())
RegisterOffsetNeedsLowering = true;
// Cannot encode zero register as base.
// If this is a stack pointer and the offset needs to be simplified then put
// the alloca address into a register, set the base type back to register and
// continue. This should almost never happen.
- if (ImmediateOffsetNeedsLowering && Addr.isFIBase()) {
+ if ((ImmediateOffsetNeedsLowering || Addr.getOffsetReg()) && Addr.isFIBase())
+ {
unsigned ResultReg = createResultReg(&AArch64::GPR64spRegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ADDXri),
ResultReg)
// reg+offset into a register.
if (ImmediateOffsetNeedsLowering) {
unsigned ResultReg;
- if (Addr.getReg()) {
+ if (Addr.getReg())
// Try to fold the immediate into the add instruction.
- if (Offset < 0)
- ResultReg = emitAddSub_ri(/*UseAdd=*/false, MVT::i64, Addr.getReg(),
- /*IsKill=*/false, -Offset);
- else
- ResultReg = emitAddSub_ri(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
- /*IsKill=*/false, Offset);
- if (!ResultReg) {
- unsigned ImmReg = fastEmit_i(MVT::i64, MVT::i64, ISD::Constant, Offset);
- ResultReg = emitAddSub_rr(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
- /*IsKill=*/false, ImmReg, /*IsKill=*/true);
- }
- } else
+ ResultReg = emitAdd_ri_(MVT::i64, Addr.getReg(), /*IsKill=*/false, Offset);
+ else
ResultReg = fastEmit_i(MVT::i64, MVT::i64, ISD::Constant, Offset);
if (!ResultReg)
MIB.addReg(Addr.getOffsetReg());
MIB.addImm(IsSigned);
MIB.addImm(Addr.getShift() != 0);
- } else {
- MIB.addReg(Addr.getReg());
- MIB.addImm(Offset);
- }
+ } else
+ MIB.addReg(Addr.getReg()).addImm(Offset);
}
if (MMO)
RetVT.SimpleTy = std::max(RetVT.SimpleTy, MVT::i32);
// Canonicalize immediates to the RHS first.
- if (UseAdd && isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS))
+ if (UseAdd && isa<Constant>(LHS) && !isa<Constant>(RHS))
std::swap(LHS, RHS);
// Canonicalize mul by power of 2 to the RHS.
else
ResultReg = emitAddSub_ri(UseAdd, RetVT, LHSReg, LHSIsKill, Imm, SetFlags,
WantResult);
- }
+ } else if (const auto *C = dyn_cast<Constant>(RHS))
+ if (C->isNullValue())
+ ResultReg = emitAddSub_ri(UseAdd, RetVT, LHSReg, LHSIsKill, 0, SetFlags,
+ WantResult);
+
if (ResultReg)
return ResultReg;
IsZExt);
}
+/// \brief This method is a wrapper to simplify add emission.
+///
+/// First try to emit an add with an immediate operand using emitAddSub_ri. If
+/// that fails, then try to materialize the immediate into a register and use
+/// emitAddSub_rr instead.
+unsigned AArch64FastISel::emitAdd_ri_(MVT VT, unsigned Op0, bool Op0IsKill,
+ int64_t Imm) {
+ unsigned ResultReg;
+ if (Imm < 0)
+ ResultReg = emitAddSub_ri(false, VT, Op0, Op0IsKill, -Imm);
+ else
+ ResultReg = emitAddSub_ri(true, VT, Op0, Op0IsKill, Imm);
+
+ if (ResultReg)
+ return ResultReg;
+
+ unsigned CReg = fastEmit_i(VT, VT, ISD::Constant, Imm);
+ if (!CReg)
+ return 0;
+
+ ResultReg = emitAddSub_rr(true, VT, Op0, Op0IsKill, CReg, true);
+ return ResultReg;
+}
+
unsigned AArch64FastISel::emitSub(MVT RetVT, const Value *LHS, const Value *RHS,
bool SetFlags, bool WantResult, bool IsZExt) {
return emitAddSub(/*UseAdd=*/false, RetVT, LHS, RHS, SetFlags, WantResult,
// could select it. Emit a copy to subreg if necessary. FastISel will remove
// it when it selects the integer extend.
unsigned Reg = lookUpRegForValue(IntExtVal);
- if (!Reg) {
+ auto *MI = MRI.getUniqueVRegDef(Reg);
+ if (!MI) {
if (RetVT == MVT::i64 && VT <= MVT::i32) {
if (WantZExt) {
// Delete the last emitted instruction from emitLoad (SUBREG_TO_REG).
// The integer extend has already been emitted - delete all the instructions
// that have been emitted by the integer extend lowering code and use the
// result from the load instruction directly.
- while (Reg) {
- auto *MI = MRI.getUniqueVRegDef(Reg);
- if (!MI)
- break;
+ while (MI) {
Reg = 0;
for (auto &Opnd : MI->uses()) {
if (Opnd.isReg()) {
}
}
MI->eraseFromParent();
+ MI = nullptr;
+ if (Reg)
+ MI = MRI.getUniqueVRegDef(Reg);
}
updateValueMap(IntExtVal, ResultReg);
return true;
const Value *LHS = CI->getOperand(0);
const Value *RHS = CI->getOperand(1);
- Type *Ty = LHS->getType();
- if (!Ty->isIntegerTy())
- return false;
+ MVT VT;
+ if (!isTypeSupported(LHS->getType(), VT))
+ return false;
- unsigned BW = cast<IntegerType>(Ty)->getBitWidth();
- if (BW != 1 && BW != 8 && BW != 16 && BW != 32 && BW != 64)
+ unsigned BW = VT.getSizeInBits();
+ if (BW > 64)
return false;
MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
int TestBit = -1;
bool IsCmpNE;
- if ((Predicate == CmpInst::ICMP_EQ) || (Predicate == CmpInst::ICMP_NE)) {
- if (const auto *C = dyn_cast<ConstantInt>(LHS))
- if (C->isNullValue())
- std::swap(LHS, RHS);
-
- if (!isa<ConstantInt>(RHS))
- return false;
+ switch (Predicate) {
+ default:
+ return false;
+ case CmpInst::ICMP_EQ:
+ case CmpInst::ICMP_NE:
+ if (isa<Constant>(LHS) && cast<Constant>(LHS)->isNullValue())
+ std::swap(LHS, RHS);
- if (!cast<ConstantInt>(RHS)->isNullValue())
+ if (!isa<Constant>(RHS) || !cast<Constant>(RHS)->isNullValue())
return false;
if (const auto *AI = dyn_cast<BinaryOperator>(LHS))
- if (AI->getOpcode() == Instruction::And) {
+ if (AI->getOpcode() == Instruction::And && isValueAvailable(AI)) {
const Value *AndLHS = AI->getOperand(0);
const Value *AndRHS = AI->getOperand(1);
LHS = AndLHS;
}
}
- IsCmpNE = Predicate == CmpInst::ICMP_NE;
- } else if (Predicate == CmpInst::ICMP_SLT) {
- if (!isa<ConstantInt>(RHS))
- return false;
- if (!cast<ConstantInt>(RHS)->isNullValue())
+ if (VT == MVT::i1)
+ TestBit = 0;
+
+ IsCmpNE = Predicate == CmpInst::ICMP_NE;
+ break;
+ case CmpInst::ICMP_SLT:
+ case CmpInst::ICMP_SGE:
+ if (!isa<Constant>(RHS) || !cast<Constant>(RHS)->isNullValue())
return false;
TestBit = BW - 1;
- IsCmpNE = true;
- } else if (Predicate == CmpInst::ICMP_SGT) {
+ IsCmpNE = Predicate == CmpInst::ICMP_SLT;
+ break;
+ case CmpInst::ICMP_SGT:
+ case CmpInst::ICMP_SLE:
if (!isa<ConstantInt>(RHS))
return false;
- if (cast<ConstantInt>(RHS)->getValue() != -1)
+ if (cast<ConstantInt>(RHS)->getValue() != APInt(BW, -1, true))
return false;
TestBit = BW - 1;
- IsCmpNE = false;
- } else
- return false;
+ IsCmpNE = Predicate == CmpInst::ICMP_SLE;
+ break;
+ } // end switch
static const unsigned OpcTable[2][2][2] = {
{ {AArch64::CBZW, AArch64::CBZX },
bool Is64Bit = BW == 64;
if (TestBit < 32 && TestBit >= 0)
Is64Bit = false;
-
+
unsigned Opc = OpcTable[IsBitTest][IsCmpNE][Is64Bit];
const MCInstrDesc &II = TII.get(Opc);
SrcReg = fastEmitInst_extractsubreg(MVT::i32, SrcReg, SrcIsKill,
AArch64::sub_32);
+ if ((BW < 32) && !IsBitTest)
+ SrcReg = emitIntExt(VT, SrcReg, MVT::i32, /*IsZExt=*/true);
+
// Emit the combined compare and branch instruction.
SrcReg = constrainOperandRegClass(II, SrcReg, II.getNumDefs());
MachineInstrBuilder MIB =
return true;
}
-bool AArch64FastISel::selectSelect(const Instruction *I) {
- const SelectInst *SI = cast<SelectInst>(I);
+/// \brief Optimize selects of i1 if one of the operands has a 'true' or 'false'
+/// value.
+bool AArch64FastISel::optimizeSelect(const SelectInst *SI) {
+ if (!SI->getType()->isIntegerTy(1))
+ return false;
- EVT DestEVT = TLI.getValueType(SI->getType(), true);
- if (!DestEVT.isSimple())
+ const Value *Src1Val, *Src2Val;
+ unsigned Opc = 0;
+ bool NeedExtraOp = false;
+ if (auto *CI = dyn_cast<ConstantInt>(SI->getTrueValue())) {
+ if (CI->isOne()) {
+ Src1Val = SI->getCondition();
+ Src2Val = SI->getFalseValue();
+ Opc = AArch64::ORRWrr;
+ } else {
+ assert(CI->isZero());
+ Src1Val = SI->getFalseValue();
+ Src2Val = SI->getCondition();
+ Opc = AArch64::BICWrr;
+ }
+ } else if (auto *CI = dyn_cast<ConstantInt>(SI->getFalseValue())) {
+ if (CI->isOne()) {
+ Src1Val = SI->getCondition();
+ Src2Val = SI->getTrueValue();
+ Opc = AArch64::ORRWrr;
+ NeedExtraOp = true;
+ } else {
+ assert(CI->isZero());
+ Src1Val = SI->getCondition();
+ Src2Val = SI->getTrueValue();
+ Opc = AArch64::ANDWrr;
+ }
+ }
+
+ if (!Opc)
return false;
- MVT DestVT = DestEVT.getSimpleVT();
- if (DestVT != MVT::i32 && DestVT != MVT::i64 && DestVT != MVT::f32 &&
- DestVT != MVT::f64)
+ unsigned Src1Reg = getRegForValue(Src1Val);
+ if (!Src1Reg)
return false;
+ bool Src1IsKill = hasTrivialKill(Src1Val);
- unsigned SelectOpc;
- const TargetRegisterClass *RC = nullptr;
- switch (DestVT.SimpleTy) {
- default: return false;
+ unsigned Src2Reg = getRegForValue(Src2Val);
+ if (!Src2Reg)
+ return false;
+ bool Src2IsKill = hasTrivialKill(Src2Val);
+
+ if (NeedExtraOp) {
+ Src1Reg = emitLogicalOp_ri(ISD::XOR, MVT::i32, Src1Reg, Src1IsKill, 1);
+ Src1IsKill = true;
+ }
+ unsigned ResultReg = fastEmitInst_rr(Opc, &AArch64::GPR32spRegClass, Src1Reg,
+ Src1IsKill, Src2Reg, Src2IsKill);
+ updateValueMap(SI, ResultReg);
+ return true;
+}
+
+bool AArch64FastISel::selectSelect(const Instruction *I) {
+ assert(isa<SelectInst>(I) && "Expected a select instruction.");
+ MVT VT;
+ if (!isTypeSupported(I->getType(), VT))
+ return false;
+
+ unsigned Opc;
+ const TargetRegisterClass *RC;
+ switch (VT.SimpleTy) {
+ default:
+ return false;
+ case MVT::i1:
+ case MVT::i8:
+ case MVT::i16:
case MVT::i32:
- SelectOpc = AArch64::CSELWr; RC = &AArch64::GPR32RegClass; break;
+ Opc = AArch64::CSELWr;
+ RC = &AArch64::GPR32RegClass;
+ break;
case MVT::i64:
- SelectOpc = AArch64::CSELXr; RC = &AArch64::GPR64RegClass; break;
+ Opc = AArch64::CSELXr;
+ RC = &AArch64::GPR64RegClass;
+ break;
case MVT::f32:
- SelectOpc = AArch64::FCSELSrrr; RC = &AArch64::FPR32RegClass; break;
+ Opc = AArch64::FCSELSrrr;
+ RC = &AArch64::FPR32RegClass;
+ break;
case MVT::f64:
- SelectOpc = AArch64::FCSELDrrr; RC = &AArch64::FPR64RegClass; break;
+ Opc = AArch64::FCSELDrrr;
+ RC = &AArch64::FPR64RegClass;
+ break;
}
+ const SelectInst *SI = cast<SelectInst>(I);
const Value *Cond = SI->getCondition();
- bool NeedTest = true;
AArch64CC::CondCode CC = AArch64CC::NE;
- if (foldXALUIntrinsic(CC, I, Cond))
- NeedTest = false;
+ AArch64CC::CondCode ExtraCC = AArch64CC::AL;
- unsigned CondReg = getRegForValue(Cond);
- if (!CondReg)
- return false;
- bool CondIsKill = hasTrivialKill(Cond);
+ if (optimizeSelect(SI))
+ return true;
- if (NeedTest) {
- unsigned ANDReg = emitAnd_ri(MVT::i32, CondReg, CondIsKill, 1);
- assert(ANDReg && "Unexpected AND instruction emission failure.");
- emitICmp_ri(MVT::i32, ANDReg, /*IsKill=*/true, 0);
+ // Try to pickup the flags, so we don't have to emit another compare.
+ if (foldXALUIntrinsic(CC, I, Cond)) {
+ // Fake request the condition to force emission of the XALU intrinsic.
+ unsigned CondReg = getRegForValue(Cond);
+ if (!CondReg)
+ return false;
+ } else if (isa<CmpInst>(Cond) && cast<CmpInst>(Cond)->hasOneUse() &&
+ isValueAvailable(Cond)) {
+ const auto *Cmp = cast<CmpInst>(Cond);
+ // Try to optimize or fold the cmp.
+ CmpInst::Predicate Predicate = optimizeCmpPredicate(Cmp);
+ const Value *FoldSelect = nullptr;
+ switch (Predicate) {
+ default:
+ break;
+ case CmpInst::FCMP_FALSE:
+ FoldSelect = SI->getFalseValue();
+ break;
+ case CmpInst::FCMP_TRUE:
+ FoldSelect = SI->getTrueValue();
+ break;
+ }
+
+ if (FoldSelect) {
+ unsigned SrcReg = getRegForValue(FoldSelect);
+ if (!SrcReg)
+ return false;
+ unsigned UseReg = lookUpRegForValue(SI);
+ if (UseReg)
+ MRI.clearKillFlags(UseReg);
+
+ updateValueMap(I, SrcReg);
+ return true;
+ }
+
+ // Emit the cmp.
+ if (!emitCmp(Cmp->getOperand(0), Cmp->getOperand(1), Cmp->isUnsigned()))
+ return false;
+
+ // FCMP_UEQ and FCMP_ONE cannot be checked with a single select instruction.
+ CC = getCompareCC(Predicate);
+ switch (Predicate) {
+ default:
+ break;
+ case CmpInst::FCMP_UEQ:
+ ExtraCC = AArch64CC::EQ;
+ CC = AArch64CC::VS;
+ break;
+ case CmpInst::FCMP_ONE:
+ ExtraCC = AArch64CC::MI;
+ CC = AArch64CC::GT;
+ break;
+ }
+ assert((CC != AArch64CC::AL) && "Unexpected condition code.");
+ } else {
+ unsigned CondReg = getRegForValue(Cond);
+ if (!CondReg)
+ return false;
+ bool CondIsKill = hasTrivialKill(Cond);
+
+ // Emit a TST instruction (ANDS wzr, reg, #imm).
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::ANDSWri),
+ AArch64::WZR)
+ .addReg(CondReg, getKillRegState(CondIsKill))
+ .addImm(AArch64_AM::encodeLogicalImmediate(1, 32));
}
- unsigned TrueReg = getRegForValue(SI->getTrueValue());
- bool TrueIsKill = hasTrivialKill(SI->getTrueValue());
+ unsigned Src1Reg = getRegForValue(SI->getTrueValue());
+ bool Src1IsKill = hasTrivialKill(SI->getTrueValue());
- unsigned FalseReg = getRegForValue(SI->getFalseValue());
- bool FalseIsKill = hasTrivialKill(SI->getFalseValue());
+ unsigned Src2Reg = getRegForValue(SI->getFalseValue());
+ bool Src2IsKill = hasTrivialKill(SI->getFalseValue());
- if (!TrueReg || !FalseReg)
+ if (!Src1Reg || !Src2Reg)
return false;
- unsigned ResultReg = fastEmitInst_rri(SelectOpc, RC, TrueReg, TrueIsKill,
- FalseReg, FalseIsKill, CC);
+ if (ExtraCC != AArch64CC::AL) {
+ Src2Reg = fastEmitInst_rri(Opc, RC, Src1Reg, Src1IsKill, Src2Reg,
+ Src2IsKill, ExtraCC);
+ Src2IsKill = true;
+ }
+ unsigned ResultReg = fastEmitInst_rri(Opc, RC, Src1Reg, Src1IsKill, Src2Reg,
+ Src2IsKill, CC);
updateValueMap(I, ResultReg);
return true;
}
// Copy all of the result registers out of their specified physreg.
MVT CopyVT = RVLocs[0].getValVT();
+
+ // TODO: Handle big-endian results
+ if (CopyVT.isVector() && !Subtarget->isLittleEndian())
+ return false;
+
unsigned ResultReg = createResultReg(TLI.getRegClassFor(CopyVT));
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(TargetOpcode::COPY), ResultReg)
// Add a register mask with the call-preserved registers.
// Proper defs for return values will be added by setPhysRegsDeadExcept().
- MIB.addRegMask(TRI.getCallPreservedMask(CC));
+ MIB.addRegMask(TRI.getCallPreservedMask(*FuncInfo.MF, CC));
CLI.Call = MIB;
MFI->setFrameAddressIsTaken(true);
const AArch64RegisterInfo *RegInfo =
- static_cast<const AArch64RegisterInfo *>(
- TM.getSubtargetImpl()->getRegisterInfo());
+ static_cast<const AArch64RegisterInfo *>(Subtarget->getRegisterInfo());
unsigned FramePtr = RegInfo->getFrameRegister(*(FuncInfo.MF));
unsigned SrcReg = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
updateValueMap(II, CLI.ResultReg);
return true;
}
+ case Intrinsic::fabs: {
+ MVT VT;
+ if (!isTypeLegal(II->getType(), VT))
+ return false;
+
+ unsigned Opc;
+ switch (VT.SimpleTy) {
+ default:
+ return false;
+ case MVT::f32:
+ Opc = AArch64::FABSSr;
+ break;
+ case MVT::f64:
+ Opc = AArch64::FABSDr;
+ break;
+ }
+ unsigned SrcReg = getRegForValue(II->getOperand(0));
+ if (!SrcReg)
+ return false;
+ bool SrcRegIsKill = hasTrivialKill(II->getOperand(0));
+ unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
+ .addReg(SrcReg, getKillRegState(SrcRegIsKill));
+ updateValueMap(II, ResultReg);
+ return true;
+ }
case Intrinsic::trap: {
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AArch64::BRK))
.addImm(1);
unsigned AArch64FastISel::emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
bool Op0IsKill, uint64_t Shift,
- bool IsZext) {
+ bool IsZExt) {
assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
"Unexpected source/return type pair.");
assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 ||
unsigned RegSize = Is64Bit ? 64 : 32;
unsigned DstBits = RetVT.getSizeInBits();
unsigned SrcBits = SrcVT.getSizeInBits();
+ const TargetRegisterClass *RC =
+ Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
+
+ // Just emit a copy for "zero" shifts.
+ if (Shift == 0) {
+ if (RetVT == SrcVT) {
+ unsigned ResultReg = createResultReg(RC);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), ResultReg)
+ .addReg(Op0, getKillRegState(Op0IsKill));
+ return ResultReg;
+ } else
+ return emitIntExt(SrcVT, Op0, RetVT, IsZExt);
+ }
// Don't deal with undefined shifts.
if (Shift >= DstBits)
{AArch64::SBFMWri, AArch64::SBFMXri},
{AArch64::UBFMWri, AArch64::UBFMXri}
};
- unsigned Opc = OpcTable[IsZext][Is64Bit];
- const TargetRegisterClass *RC =
- Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
+ unsigned Opc = OpcTable[IsZExt][Is64Bit];
if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
unsigned TmpReg = MRI.createVirtualRegister(RC);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
bool IsZExt) {
assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
"Unexpected source/return type pair.");
- assert((SrcVT == MVT::i8 || SrcVT == MVT::i16 || SrcVT == MVT::i32 ||
- SrcVT == MVT::i64) && "Unexpected source value type.");
+ assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 ||
+ SrcVT == MVT::i32 || SrcVT == MVT::i64) &&
+ "Unexpected source value type.");
assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
RetVT == MVT::i64) && "Unexpected return value type.");
unsigned RegSize = Is64Bit ? 64 : 32;
unsigned DstBits = RetVT.getSizeInBits();
unsigned SrcBits = SrcVT.getSizeInBits();
+ const TargetRegisterClass *RC =
+ Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
+
+ // Just emit a copy for "zero" shifts.
+ if (Shift == 0) {
+ if (RetVT == SrcVT) {
+ unsigned ResultReg = createResultReg(RC);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), ResultReg)
+ .addReg(Op0, getKillRegState(Op0IsKill));
+ return ResultReg;
+ } else
+ return emitIntExt(SrcVT, Op0, RetVT, IsZExt);
+ }
// Don't deal with undefined shifts.
if (Shift >= DstBits)
{AArch64::UBFMWri, AArch64::UBFMXri}
};
unsigned Opc = OpcTable[IsZExt][Is64Bit];
- const TargetRegisterClass *RC =
- Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
unsigned TmpReg = MRI.createVirtualRegister(RC);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
bool IsZExt) {
assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
"Unexpected source/return type pair.");
- assert((SrcVT == MVT::i8 || SrcVT == MVT::i16 || SrcVT == MVT::i32 ||
- SrcVT == MVT::i64) && "Unexpected source value type.");
+ assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 ||
+ SrcVT == MVT::i32 || SrcVT == MVT::i64) &&
+ "Unexpected source value type.");
assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
RetVT == MVT::i64) && "Unexpected return value type.");
unsigned RegSize = Is64Bit ? 64 : 32;
unsigned DstBits = RetVT.getSizeInBits();
unsigned SrcBits = SrcVT.getSizeInBits();
+ const TargetRegisterClass *RC =
+ Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
+
+ // Just emit a copy for "zero" shifts.
+ if (Shift == 0) {
+ if (RetVT == SrcVT) {
+ unsigned ResultReg = createResultReg(RC);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), ResultReg)
+ .addReg(Op0, getKillRegState(Op0IsKill));
+ return ResultReg;
+ } else
+ return emitIntExt(SrcVT, Op0, RetVT, IsZExt);
+ }
// Don't deal with undefined shifts.
if (Shift >= DstBits)
{AArch64::UBFMWri, AArch64::UBFMXri}
};
unsigned Opc = OpcTable[IsZExt][Is64Bit];
- const TargetRegisterClass *RC =
- Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
unsigned TmpReg = MRI.createVirtualRegister(RC);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
.addImm(AArch64::sub_32);
SrcReg = ResultReg;
}
+ // Conservatively clear all kill flags from all uses, because we are
+ // replacing a sign-/zero-extend instruction at IR level with a nop at MI
+ // level. The result of the instruction at IR level might have been
+ // trivially dead, which is now not longer true.
+ unsigned UseReg = lookUpRegForValue(I);
+ if (UseReg)
+ MRI.clearKillFlags(UseReg);
+
updateValueMap(I, SrcReg);
return true;
}
unsigned ResultReg = 0;
uint64_t ShiftVal = C->getZExtValue();
MVT SrcVT = RetVT;
- bool IsZExt = (I->getOpcode() == Instruction::AShr) ? false : true;
+ bool IsZExt = I->getOpcode() != Instruction::AShr;
const Value *Op0 = I->getOperand(0);
if (const auto *ZExt = dyn_cast<ZExtInst>(Op0)) {
if (!isIntExtFree(ZExt)) {
return true;
}
- unsigned Pow2MinusOne = (1 << Lg2) - 1;
- unsigned AddReg = emitAddSub_ri(/*UseAdd=*/true, VT, Src0Reg,
- /*IsKill=*/false, Pow2MinusOne);
+ int64_t Pow2MinusOne = (1ULL << Lg2) - 1;
+ unsigned AddReg = emitAdd_ri_(VT, Src0Reg, /*IsKill=*/false, Pow2MinusOne);
if (!AddReg)
return false;
return true;
}
+/// This is mostly a copy of the existing FastISel getRegForGEPIndex code. We
+/// have to duplicate it for AArch64, because otherwise we would fail during the
+/// sign-extend emission.
+std::pair<unsigned, bool> AArch64FastISel::getRegForGEPIndex(const Value *Idx) {
+ unsigned IdxN = getRegForValue(Idx);
+ if (IdxN == 0)
+ // Unhandled operand. Halt "fast" selection and bail.
+ return std::pair<unsigned, bool>(0, false);
+
+ bool IdxNIsKill = hasTrivialKill(Idx);
+
+ // If the index is smaller or larger than intptr_t, truncate or extend it.
+ MVT PtrVT = TLI.getPointerTy();
+ EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false);
+ if (IdxVT.bitsLT(PtrVT)) {
+ IdxN = emitIntExt(IdxVT.getSimpleVT(), IdxN, PtrVT, /*IsZExt=*/false);
+ IdxNIsKill = true;
+ } else if (IdxVT.bitsGT(PtrVT))
+ llvm_unreachable("AArch64 FastISel doesn't support types larger than i64");
+ return std::pair<unsigned, bool>(IdxN, IdxNIsKill);
+}
+
+/// This is mostly a copy of the existing FastISel GEP code, but we have to
+/// duplicate it for AArch64, because otherwise we would bail out even for
+/// simple cases. This is because the standard fastEmit functions don't cover
+/// MUL at all and ADD is lowered very inefficientily.
+bool AArch64FastISel::selectGetElementPtr(const Instruction *I) {
+ unsigned N = getRegForValue(I->getOperand(0));
+ if (!N)
+ return false;
+ bool NIsKill = hasTrivialKill(I->getOperand(0));
+
+ // Keep a running tab of the total offset to coalesce multiple N = N + Offset
+ // into a single N = N + TotalOffset.
+ uint64_t TotalOffs = 0;
+ Type *Ty = I->getOperand(0)->getType();
+ MVT VT = TLI.getPointerTy();
+ for (auto OI = std::next(I->op_begin()), E = I->op_end(); OI != E; ++OI) {
+ const Value *Idx = *OI;
+ if (auto *StTy = dyn_cast<StructType>(Ty)) {
+ unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
+ // N = N + Offset
+ if (Field)
+ TotalOffs += DL.getStructLayout(StTy)->getElementOffset(Field);
+ Ty = StTy->getElementType(Field);
+ } else {
+ Ty = cast<SequentialType>(Ty)->getElementType();
+ // If this is a constant subscript, handle it quickly.
+ if (const auto *CI = dyn_cast<ConstantInt>(Idx)) {
+ if (CI->isZero())
+ continue;
+ // N = N + Offset
+ TotalOffs +=
+ DL.getTypeAllocSize(Ty) * cast<ConstantInt>(CI)->getSExtValue();
+ continue;
+ }
+ if (TotalOffs) {
+ N = emitAdd_ri_(VT, N, NIsKill, TotalOffs);
+ if (!N)
+ return false;
+ NIsKill = true;
+ TotalOffs = 0;
+ }
+
+ // N = N + Idx * ElementSize;
+ uint64_t ElementSize = DL.getTypeAllocSize(Ty);
+ std::pair<unsigned, bool> Pair = getRegForGEPIndex(Idx);
+ unsigned IdxN = Pair.first;
+ bool IdxNIsKill = Pair.second;
+ if (!IdxN)
+ return false;
+
+ if (ElementSize != 1) {
+ unsigned C = fastEmit_i(VT, VT, ISD::Constant, ElementSize);
+ if (!C)
+ return false;
+ IdxN = emitMul_rr(VT, IdxN, IdxNIsKill, C, true);
+ if (!IdxN)
+ return false;
+ IdxNIsKill = true;
+ }
+ N = fastEmit_rr(VT, VT, ISD::ADD, N, NIsKill, IdxN, IdxNIsKill);
+ if (!N)
+ return false;
+ }
+ }
+ if (TotalOffs) {
+ N = emitAdd_ri_(VT, N, NIsKill, TotalOffs);
+ if (!N)
+ return false;
+ }
+ updateValueMap(I, N);
+ return true;
+}
+
bool AArch64FastISel::fastSelectInstruction(const Instruction *I) {
switch (I->getOpcode()) {
default:
return selectRet(I);
case Instruction::FRem:
return selectFRem(I);
+ case Instruction::GetElementPtr:
+ return selectGetElementPtr(I);
}
// fall-back to target-independent instruction selection.
projects / oota-llvm.git / blobdiff