#include "ARM.h"
#include "ARMBaseInstrInfo.h"
#include "ARMCallingConv.h"
-#include "ARMRegisterInfo.h"
-#include "ARMTargetMachine.h"
-#include "ARMSubtarget.h"
#include "ARMConstantPoolValue.h"
+#include "ARMSubtarget.h"
+#include "ARMTargetMachine.h"
#include "MCTargetDesc/ARMAddressingModes.h"
-#include "llvm/CallingConv.h"
-#include "llvm/DerivedTypes.h"
-#include "llvm/GlobalVariable.h"
-#include "llvm/Instructions.h"
-#include "llvm/IntrinsicInst.h"
-#include "llvm/Module.h"
-#include "llvm/Operator.h"
+#include "llvm/ADT/STLExtras.h"
#include "llvm/CodeGen/Analysis.h"
#include "llvm/CodeGen/FastISel.h"
#include "llvm/CodeGen/FunctionLoweringInfo.h"
-#include "llvm/CodeGen/MachineInstrBuilder.h"
-#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineMemOperand.h"
+#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/IR/CallingConv.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/DerivedTypes.h"
+#include "llvm/IR/GlobalVariable.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/IntrinsicInst.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/Operator.h"
#include "llvm/Support/CallSite.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
-#include "llvm/Target/TargetData.h"
+#include "llvm/Support/MathExtras.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetOptions.h"
using namespace llvm;
-static cl::opt<bool>
-DisableARMFastISel("disable-arm-fast-isel",
- cl::desc("Turn off experimental ARM fast-isel support"),
- cl::init(false), cl::Hidden);
-
extern cl::opt<bool> EnableARMLongCalls;
namespace {
LLVMContext *Context;
public:
- explicit ARMFastISel(FunctionLoweringInfo &funcInfo)
- : FastISel(funcInfo),
+ explicit ARMFastISel(FunctionLoweringInfo &funcInfo,
+ const TargetLibraryInfo *libInfo)
+ : FastISel(funcInfo, libInfo),
TM(funcInfo.MF->getTarget()),
TII(*TM.getInstrInfo()),
TLI(*TM.getTargetLowering()) {
}
// Code from FastISel.cpp.
- virtual unsigned FastEmitInst_(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC);
- virtual unsigned FastEmitInst_r(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- unsigned Op0, bool Op0IsKill);
- virtual unsigned FastEmitInst_rr(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- unsigned Op0, bool Op0IsKill,
- unsigned Op1, bool Op1IsKill);
- virtual unsigned FastEmitInst_rrr(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- unsigned Op0, bool Op0IsKill,
- unsigned Op1, bool Op1IsKill,
- unsigned Op2, bool Op2IsKill);
- virtual unsigned FastEmitInst_ri(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- unsigned Op0, bool Op0IsKill,
- uint64_t Imm);
- virtual unsigned FastEmitInst_rf(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- unsigned Op0, bool Op0IsKill,
- const ConstantFP *FPImm);
- virtual unsigned FastEmitInst_rri(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- unsigned Op0, bool Op0IsKill,
- unsigned Op1, bool Op1IsKill,
- uint64_t Imm);
- virtual unsigned FastEmitInst_i(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- uint64_t Imm);
- virtual unsigned FastEmitInst_ii(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- uint64_t Imm1, uint64_t Imm2);
-
- virtual unsigned FastEmitInst_extractsubreg(MVT RetVT,
- unsigned Op0, bool Op0IsKill,
- uint32_t Idx);
+ private:
+ unsigned FastEmitInst_(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC);
+ unsigned FastEmitInst_r(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill);
+ unsigned FastEmitInst_rr(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill,
+ unsigned Op1, bool Op1IsKill);
+ unsigned FastEmitInst_rrr(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill,
+ unsigned Op1, bool Op1IsKill,
+ unsigned Op2, bool Op2IsKill);
+ unsigned FastEmitInst_ri(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill,
+ uint64_t Imm);
+ unsigned FastEmitInst_rf(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill,
+ const ConstantFP *FPImm);
+ unsigned FastEmitInst_rri(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill,
+ unsigned Op1, bool Op1IsKill,
+ uint64_t Imm);
+ unsigned FastEmitInst_i(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ uint64_t Imm);
+ unsigned FastEmitInst_ii(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ uint64_t Imm1, uint64_t Imm2);
+
+ unsigned FastEmitInst_extractsubreg(MVT RetVT,
+ unsigned Op0, bool Op0IsKill,
+ uint32_t Idx);
// Backend specific FastISel code.
+ private:
virtual bool TargetSelectInstruction(const Instruction *I);
virtual unsigned TargetMaterializeConstant(const Constant *C);
virtual unsigned TargetMaterializeAlloca(const AllocaInst *AI);
- virtual bool TryToFoldLoad(MachineInstr *MI, unsigned OpNo,
- const LoadInst *LI);
-
+ virtual bool tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
+ const LoadInst *LI);
+ virtual bool FastLowerArguments();
+ private:
#include "ARMGenFastISel.inc"
// Instruction selection routines.
bool SelectLoad(const Instruction *I);
bool SelectStore(const Instruction *I);
bool SelectBranch(const Instruction *I);
+ bool SelectIndirectBr(const Instruction *I);
bool SelectCmp(const Instruction *I);
bool SelectFPExt(const Instruction *I);
bool SelectFPTrunc(const Instruction *I);
- bool SelectBinaryOp(const Instruction *I, unsigned ISDOpcode);
- bool SelectSIToFP(const Instruction *I);
- bool SelectFPToSI(const Instruction *I);
- bool SelectSDiv(const Instruction *I);
- bool SelectSRem(const Instruction *I);
+ bool SelectBinaryIntOp(const Instruction *I, unsigned ISDOpcode);
+ bool SelectBinaryFPOp(const Instruction *I, unsigned ISDOpcode);
+ bool SelectIToFP(const Instruction *I, bool isSigned);
+ bool SelectFPToI(const Instruction *I, bool isSigned);
+ bool SelectDiv(const Instruction *I, bool isSigned);
+ bool SelectRem(const Instruction *I, bool isSigned);
bool SelectCall(const Instruction *I, const char *IntrMemName);
bool SelectIntrinsicCall(const IntrinsicInst &I);
bool SelectSelect(const Instruction *I);
bool SelectRet(const Instruction *I);
bool SelectTrunc(const Instruction *I);
bool SelectIntExt(const Instruction *I);
+ bool SelectShift(const Instruction *I, ARM_AM::ShiftOpc ShiftTy);
// Utility routines.
private:
bool isLoadTypeLegal(Type *Ty, MVT &VT);
bool ARMEmitCmp(const Value *Src1Value, const Value *Src2Value,
bool isZExt);
- bool ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr,
+ bool ARMEmitLoad(MVT VT, unsigned &ResultReg, Address &Addr,
unsigned Alignment = 0, bool isZExt = true,
bool allocReg = true);
-
- bool ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr,
+ bool ARMEmitStore(MVT VT, unsigned SrcReg, Address &Addr,
unsigned Alignment = 0);
bool ARMComputeAddress(const Value *Obj, Address &Addr);
- void ARMSimplifyAddress(Address &Addr, EVT VT, bool useAM3);
+ void ARMSimplifyAddress(Address &Addr, MVT VT, bool useAM3);
bool ARMIsMemCpySmall(uint64_t Len);
- bool ARMTryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len);
- unsigned ARMEmitIntExt(EVT SrcVT, unsigned SrcReg, EVT DestVT, bool isZExt);
- unsigned ARMMaterializeFP(const ConstantFP *CFP, EVT VT);
- unsigned ARMMaterializeInt(const Constant *C, EVT VT);
- unsigned ARMMaterializeGV(const GlobalValue *GV, EVT VT);
- unsigned ARMMoveToFPReg(EVT VT, unsigned SrcReg);
- unsigned ARMMoveToIntReg(EVT VT, unsigned SrcReg);
- unsigned ARMSelectCallOp(const GlobalValue *GV);
+ bool ARMTryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len,
+ unsigned Alignment);
+ unsigned ARMEmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, bool isZExt);
+ unsigned ARMMaterializeFP(const ConstantFP *CFP, MVT VT);
+ unsigned ARMMaterializeInt(const Constant *C, MVT VT);
+ unsigned ARMMaterializeGV(const GlobalValue *GV, MVT VT);
+ unsigned ARMMoveToFPReg(MVT VT, unsigned SrcReg);
+ unsigned ARMMoveToIntReg(MVT VT, unsigned SrcReg);
+ unsigned ARMSelectCallOp(bool UseReg);
+ unsigned ARMLowerPICELF(const GlobalValue *GV, unsigned Align, MVT VT);
// Call handling routines.
private:
- CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool Return);
+ CCAssignFn *CCAssignFnForCall(CallingConv::ID CC,
+ bool Return,
+ bool isVarArg);
bool ProcessCallArgs(SmallVectorImpl<Value*> &Args,
SmallVectorImpl<unsigned> &ArgRegs,
SmallVectorImpl<MVT> &ArgVTs,
SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
SmallVectorImpl<unsigned> &RegArgs,
CallingConv::ID CC,
- unsigned &NumBytes);
+ unsigned &NumBytes,
+ bool isVarArg);
+ unsigned getLibcallReg(const Twine &Name);
bool FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
const Instruction *I, CallingConv::ID CC,
- unsigned &NumBytes);
+ unsigned &NumBytes, bool isVarArg);
bool ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call);
// OptionalDef handling routines.
bool isARMNEONPred(const MachineInstr *MI);
bool DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR);
const MachineInstrBuilder &AddOptionalDefs(const MachineInstrBuilder &MIB);
- void AddLoadStoreOperands(EVT VT, Address &Addr,
+ void AddLoadStoreOperands(MVT VT, Address &Addr,
const MachineInstrBuilder &MIB,
unsigned Flags, bool useAM3);
};
unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
- if (II.getNumDefs() >= 1)
+ if (II.getNumDefs() >= 1) {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill));
- else {
+ } else {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
.addReg(Op0, Op0IsKill * RegState::Kill));
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
- if (II.getNumDefs() >= 1)
+ if (II.getNumDefs() >= 1) {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addReg(Op1, Op1IsKill * RegState::Kill));
- else {
+ } else {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addReg(Op1, Op1IsKill * RegState::Kill));
unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
- if (II.getNumDefs() >= 1)
+ if (II.getNumDefs() >= 1) {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addReg(Op1, Op1IsKill * RegState::Kill)
.addReg(Op2, Op2IsKill * RegState::Kill));
- else {
+ } else {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addReg(Op1, Op1IsKill * RegState::Kill)
unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
- if (II.getNumDefs() >= 1)
+ if (II.getNumDefs() >= 1) {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addImm(Imm));
- else {
+ } else {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addImm(Imm));
unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
- if (II.getNumDefs() >= 1)
+ if (II.getNumDefs() >= 1) {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addFPImm(FPImm));
- else {
+ } else {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addFPImm(FPImm));
unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
- if (II.getNumDefs() >= 1)
+ if (II.getNumDefs() >= 1) {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addReg(Op1, Op1IsKill * RegState::Kill)
.addImm(Imm));
- else {
+ } else {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
.addReg(Op0, Op0IsKill * RegState::Kill)
.addReg(Op1, Op1IsKill * RegState::Kill)
unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
- if (II.getNumDefs() >= 1)
+ if (II.getNumDefs() >= 1) {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
.addImm(Imm));
- else {
+ } else {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
.addImm(Imm));
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
- if (II.getNumDefs() >= 1)
+ if (II.getNumDefs() >= 1) {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
.addImm(Imm1).addImm(Imm2));
- else {
+ } else {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
.addImm(Imm1).addImm(Imm2));
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
assert(TargetRegisterInfo::isVirtualRegister(Op0) &&
"Cannot yet extract from physregs");
+
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
- DL, TII.get(TargetOpcode::COPY), ResultReg)
- .addReg(Op0, getKillRegState(Op0IsKill), Idx));
+ DL, TII.get(TargetOpcode::COPY), ResultReg)
+ .addReg(Op0, getKillRegState(Op0IsKill), Idx));
return ResultReg;
}
// TODO: Don't worry about 64-bit now, but when this is fixed remove the
// checks from the various callers.
-unsigned ARMFastISel::ARMMoveToFPReg(EVT VT, unsigned SrcReg) {
+unsigned ARMFastISel::ARMMoveToFPReg(MVT VT, unsigned SrcReg) {
if (VT == MVT::f64) return 0;
unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(ARM::VMOVRS), MoveReg)
+ TII.get(ARM::VMOVSR), MoveReg)
.addReg(SrcReg));
return MoveReg;
}
-unsigned ARMFastISel::ARMMoveToIntReg(EVT VT, unsigned SrcReg) {
+unsigned ARMFastISel::ARMMoveToIntReg(MVT VT, unsigned SrcReg) {
if (VT == MVT::i64) return 0;
unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(ARM::VMOVSR), MoveReg)
+ TII.get(ARM::VMOVRS), MoveReg)
.addReg(SrcReg));
return MoveReg;
}
// For double width floating point we need to materialize two constants
// (the high and the low) into integer registers then use a move to get
// the combined constant into an FP reg.
-unsigned ARMFastISel::ARMMaterializeFP(const ConstantFP *CFP, EVT VT) {
+unsigned ARMFastISel::ARMMaterializeFP(const ConstantFP *CFP, MVT VT) {
const APFloat Val = CFP->getValueAPF();
bool is64bit = VT == MVT::f64;
return DestReg;
}
-unsigned ARMFastISel::ARMMaterializeInt(const Constant *C, EVT VT) {
+unsigned ARMFastISel::ARMMaterializeInt(const Constant *C, MVT VT) {
if (VT != MVT::i32 && VT != MVT::i16 && VT != MVT::i8 && VT != MVT::i1)
return false;
const ConstantInt *CI = cast<ConstantInt>(C);
if (Subtarget->hasV6T2Ops() && isUInt<16>(CI->getZExtValue())) {
unsigned Opc = isThumb2 ? ARM::t2MOVi16 : ARM::MOVi16;
- unsigned ImmReg = createResultReg(TLI.getRegClassFor(MVT::i32));
+ const TargetRegisterClass *RC = isThumb2 ? &ARM::rGPRRegClass :
+ &ARM::GPRRegClass;
+ unsigned ImmReg = createResultReg(RC);
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(Opc), ImmReg)
.addImm(CI->getZExtValue()));
return DestReg;
}
-unsigned ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, EVT VT) {
+unsigned ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, MVT VT) {
// For now 32-bit only.
if (VT != MVT::i32) return 0;
Reloc::Model RelocM = TM.getRelocationModel();
+ bool IsIndirect = Subtarget->GVIsIndirectSymbol(GV, RelocM);
+ const TargetRegisterClass *RC = isThumb2 ?
+ (const TargetRegisterClass*)&ARM::rGPRRegClass :
+ (const TargetRegisterClass*)&ARM::GPRRegClass;
+ unsigned DestReg = createResultReg(RC);
- // TODO: Need more magic for ARM PIC.
- if (!isThumb2 && (RelocM == Reloc::PIC_)) return 0;
-
- unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
+ // FastISel TLS support on non-Darwin is broken, punt to SelectionDAG.
+ const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
+ bool IsThreadLocal = GVar && GVar->isThreadLocal();
+ if (!Subtarget->isTargetDarwin() && IsThreadLocal) return 0;
// Use movw+movt when possible, it avoids constant pool entries.
// Darwin targets don't support movt with Reloc::Static, see
Align = TD.getTypeAllocSize(GV->getType());
}
+ if (Subtarget->isTargetELF() && RelocM == Reloc::PIC_)
+ return ARMLowerPICELF(GV, Align, VT);
+
// Grab index.
unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 :
(Subtarget->isThumb() ? 4 : 8);
.addConstantPoolIndex(Idx);
if (RelocM == Reloc::PIC_)
MIB.addImm(Id);
+ AddOptionalDefs(MIB);
} else {
// The extra immediate is for addrmode2.
MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp),
DestReg)
.addConstantPoolIndex(Idx)
.addImm(0);
+ AddOptionalDefs(MIB);
+
+ if (RelocM == Reloc::PIC_) {
+ unsigned Opc = IsIndirect ? ARM::PICLDR : ARM::PICADD;
+ unsigned NewDestReg = createResultReg(TLI.getRegClassFor(VT));
+
+ MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
+ DL, TII.get(Opc), NewDestReg)
+ .addReg(DestReg)
+ .addImm(Id);
+ AddOptionalDefs(MIB);
+ return NewDestReg;
+ }
}
- AddOptionalDefs(MIB);
}
- if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) {
+ if (IsIndirect) {
MachineInstrBuilder MIB;
unsigned NewDestReg = createResultReg(TLI.getRegClassFor(VT));
if (isThumb2)
}
unsigned ARMFastISel::TargetMaterializeConstant(const Constant *C) {
- EVT VT = TLI.getValueType(C->getType(), true);
+ EVT CEVT = TLI.getValueType(C->getType(), true);
// Only handle simple types.
- if (!VT.isSimple()) return 0;
+ if (!CEVT.isSimple()) return 0;
+ MVT VT = CEVT.getSimpleVT();
if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
return ARMMaterializeFP(CFP, VT);
if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;
MVT VT;
- if (!isLoadTypeLegal(AI->getType(), VT)) return false;
+ if (!isLoadTypeLegal(AI->getType(), VT)) return 0;
DenseMap<const AllocaInst*, int>::iterator SI =
FuncInfo.StaticAllocaMap.find(AI);
// This will get lowered later into the correct offsets and registers
// via rewriteXFrameIndex.
if (SI != FuncInfo.StaticAllocaMap.end()) {
- TargetRegisterClass* RC = TLI.getRegClassFor(VT);
+ const TargetRegisterClass* RC = TLI.getRegClassFor(VT);
unsigned ResultReg = createResultReg(RC);
unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri;
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
return Addr.Base.Reg != 0;
}
-void ARMFastISel::ARMSimplifyAddress(Address &Addr, EVT VT, bool useAM3) {
-
- assert(VT.isSimple() && "Non-simple types are invalid here!");
-
+void ARMFastISel::ARMSimplifyAddress(Address &Addr, MVT VT, bool useAM3) {
bool needsLowering = false;
- switch (VT.getSimpleVT().SimpleTy) {
- default:
- assert(false && "Unhandled load/store type!");
- break;
+ switch (VT.SimpleTy) {
+ default: llvm_unreachable("Unhandled load/store type!");
case MVT::i1:
case MVT::i8:
case MVT::i16:
// put the alloca address into a register, set the base type back to
// register and continue. This should almost never happen.
if (needsLowering && Addr.BaseType == Address::FrameIndexBase) {
- TargetRegisterClass *RC = isThumb2 ? ARM::tGPRRegisterClass :
- ARM::GPRRegisterClass;
+ const TargetRegisterClass *RC = isThumb2 ?
+ (const TargetRegisterClass*)&ARM::tGPRRegClass :
+ (const TargetRegisterClass*)&ARM::GPRRegClass;
unsigned ResultReg = createResultReg(RC);
unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri;
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
}
}
-void ARMFastISel::AddLoadStoreOperands(EVT VT, Address &Addr,
+void ARMFastISel::AddLoadStoreOperands(MVT VT, Address &Addr,
const MachineInstrBuilder &MIB,
unsigned Flags, bool useAM3) {
// addrmode5 output depends on the selection dag addressing dividing the
// offset by 4 that it then later multiplies. Do this here as well.
- if (VT.getSimpleVT().SimpleTy == MVT::f32 ||
- VT.getSimpleVT().SimpleTy == MVT::f64)
+ if (VT.SimpleTy == MVT::f32 || VT.SimpleTy == MVT::f64)
Addr.Offset /= 4;
// Frame base works a bit differently. Handle it separately.
AddOptionalDefs(MIB);
}
-bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr,
+bool ARMFastISel::ARMEmitLoad(MVT VT, unsigned &ResultReg, Address &Addr,
unsigned Alignment, bool isZExt, bool allocReg) {
- assert(VT.isSimple() && "Non-simple types are invalid here!");
unsigned Opc;
bool useAM3 = false;
bool needVMOV = false;
- TargetRegisterClass *RC;
- switch (VT.getSimpleVT().SimpleTy) {
+ const TargetRegisterClass *RC;
+ switch (VT.SimpleTy) {
// This is mostly going to be Neon/vector support.
default: return false;
case MVT::i1:
useAM3 = true;
}
}
- RC = ARM::GPRRegisterClass;
+ RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass;
break;
case MVT::i16:
+ if (Alignment && Alignment < 2 && !Subtarget->allowsUnalignedMem())
+ return false;
+
if (isThumb2) {
if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
Opc = isZExt ? ARM::t2LDRHi8 : ARM::t2LDRSHi8;
Opc = isZExt ? ARM::LDRH : ARM::LDRSH;
useAM3 = true;
}
- RC = ARM::GPRRegisterClass;
+ RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass;
break;
case MVT::i32:
+ if (Alignment && Alignment < 4 && !Subtarget->allowsUnalignedMem())
+ return false;
+
if (isThumb2) {
if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
Opc = ARM::t2LDRi8;
} else {
Opc = ARM::LDRi12;
}
- RC = ARM::GPRRegisterClass;
+ RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass;
break;
case MVT::f32:
if (!Subtarget->hasVFP2()) return false;
needVMOV = true;
VT = MVT::i32;
Opc = isThumb2 ? ARM::t2LDRi12 : ARM::LDRi12;
- RC = ARM::GPRRegisterClass;
+ RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass;
} else {
Opc = ARM::VLDRS;
RC = TLI.getRegClassFor(VT);
return true;
}
-bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr,
+bool ARMFastISel::ARMEmitStore(MVT VT, unsigned SrcReg, Address &Addr,
unsigned Alignment) {
unsigned StrOpc;
bool useAM3 = false;
- switch (VT.getSimpleVT().SimpleTy) {
+ switch (VT.SimpleTy) {
// This is mostly going to be Neon/vector support.
default: return false;
case MVT::i1: {
- unsigned Res = createResultReg(isThumb2 ? ARM::tGPRRegisterClass :
- ARM::GPRRegisterClass);
+ unsigned Res = createResultReg(isThumb2 ?
+ (const TargetRegisterClass*)&ARM::tGPRRegClass :
+ (const TargetRegisterClass*)&ARM::GPRRegClass);
unsigned Opc = isThumb2 ? ARM::t2ANDri : ARM::ANDri;
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(Opc), Res)
}
break;
case MVT::i16:
+ if (Alignment && Alignment < 2 && !Subtarget->allowsUnalignedMem())
+ return false;
+
if (isThumb2) {
if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
StrOpc = ARM::t2STRHi8;
}
break;
case MVT::i32:
+ if (Alignment && Alignment < 4 && !Subtarget->allowsUnalignedMem())
+ return false;
+
if (isThumb2) {
if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
StrOpc = ARM::t2STRi8;
return true;
}
+bool ARMFastISel::SelectIndirectBr(const Instruction *I) {
+ unsigned AddrReg = getRegForValue(I->getOperand(0));
+ if (AddrReg == 0) return false;
+
+ unsigned Opc = isThumb2 ? ARM::tBRIND : ARM::BX;
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc))
+ .addReg(AddrReg));
+
+ const IndirectBrInst *IB = cast<IndirectBrInst>(I);
+ for (unsigned i = 0, e = IB->getNumSuccessors(); i != e; ++i)
+ FuncInfo.MBB->addSuccessor(FuncInfo.MBBMap[IB->getSuccessor(i)]);
+
+ return true;
+}
+
bool ARMFastISel::ARMEmitCmp(const Value *Src1Value, const Value *Src2Value,
bool isZExt) {
Type *Ty = Src1Value->getType();
- EVT SrcVT = TLI.getValueType(Ty, true);
- if (!SrcVT.isSimple()) return false;
+ EVT SrcEVT = TLI.getValueType(Ty, true);
+ if (!SrcEVT.isSimple()) return false;
+ MVT SrcVT = SrcEVT.getSimpleVT();
bool isFloat = (Ty->isFloatTy() || Ty->isDoubleTy());
if (isFloat && !Subtarget->hasVFP2())
SrcVT == MVT::i1) {
const APInt &CIVal = ConstInt->getValue();
Imm = (isZExt) ? (int)CIVal.getZExtValue() : (int)CIVal.getSExtValue();
- if (Imm < 0) {
+ // For INT_MIN/LONG_MIN (i.e., 0x80000000) we need to use a cmp, rather
+ // then a cmn, because there is no way to represent 2147483648 as a
+ // signed 32-bit int.
+ if (Imm < 0 && Imm != (int)0x80000000) {
isNegativeImm = true;
Imm = -Imm;
}
unsigned CmpOpc;
bool isICmp = true;
bool needsExt = false;
- switch (SrcVT.getSimpleVT().SimpleTy) {
+ switch (SrcVT.SimpleTy) {
default: return false;
// TODO: Verify compares.
case MVT::f32:
if (!UseImm)
CmpOpc = ARM::t2CMPrr;
else
- CmpOpc = isNegativeImm ? ARM::t2CMNzri : ARM::t2CMPri;
+ CmpOpc = isNegativeImm ? ARM::t2CMNri : ARM::t2CMPri;
} else {
if (!UseImm)
CmpOpc = ARM::CMPrr;
else
- CmpOpc = isNegativeImm ? ARM::CMNzri : ARM::CMPri;
+ CmpOpc = isNegativeImm ? ARM::CMNri : ARM::CMPri;
}
break;
}
// We have i1, i8, or i16, we need to either zero extend or sign extend.
if (needsExt) {
- unsigned ResultReg;
- ResultReg = ARMEmitIntExt(SrcVT, SrcReg1, MVT::i32, isZExt);
- if (ResultReg == 0) return false;
- SrcReg1 = ResultReg;
+ SrcReg1 = ARMEmitIntExt(SrcVT, SrcReg1, MVT::i32, isZExt);
+ if (SrcReg1 == 0) return false;
if (!UseImm) {
- ResultReg = ARMEmitIntExt(SrcVT, SrcReg2, MVT::i32, isZExt);
- if (ResultReg == 0) return false;
- SrcReg2 = ResultReg;
+ SrcReg2 = ARMEmitIntExt(SrcVT, SrcReg2, MVT::i32, isZExt);
+ if (SrcReg2 == 0) return false;
}
}
bool ARMFastISel::SelectCmp(const Instruction *I) {
const CmpInst *CI = cast<CmpInst>(I);
- Type *Ty = CI->getOperand(0)->getType();
// Get the compare predicate.
ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());
// Now set a register based on the comparison. Explicitly set the predicates
// here.
unsigned MovCCOpc = isThumb2 ? ARM::t2MOVCCi : ARM::MOVCCi;
- TargetRegisterClass *RC = isThumb2 ? ARM::rGPRRegisterClass
- : ARM::GPRRegisterClass;
+ const TargetRegisterClass *RC = isThumb2 ?
+ (const TargetRegisterClass*)&ARM::rGPRRegClass :
+ (const TargetRegisterClass*)&ARM::GPRRegClass;
unsigned DestReg = createResultReg(RC);
Constant *Zero = ConstantInt::get(Type::getInt32Ty(*Context), 0);
unsigned ZeroReg = TargetMaterializeConstant(Zero);
- bool isFloat = (Ty->isFloatTy() || Ty->isDoubleTy());
- unsigned CondReg = isFloat ? ARM::FPSCR : ARM::CPSR;
+ // ARMEmitCmp emits a FMSTAT when necessary, so it's always safe to use CPSR.
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), DestReg)
.addReg(ZeroReg).addImm(1)
- .addImm(ARMPred).addReg(CondReg);
+ .addImm(ARMPred).addReg(ARM::CPSR);
UpdateValueMap(I, DestReg);
return true;
unsigned Op = getRegForValue(V);
if (Op == 0) return false;
- unsigned Result = createResultReg(ARM::DPRRegisterClass);
+ unsigned Result = createResultReg(&ARM::DPRRegClass);
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(ARM::VCVTDS), Result)
.addReg(Op));
unsigned Op = getRegForValue(V);
if (Op == 0) return false;
- unsigned Result = createResultReg(ARM::SPRRegisterClass);
+ unsigned Result = createResultReg(&ARM::SPRRegClass);
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(ARM::VCVTSD), Result)
.addReg(Op));
return true;
}
-bool ARMFastISel::SelectSIToFP(const Instruction *I) {
+bool ARMFastISel::SelectIToFP(const Instruction *I, bool isSigned) {
// Make sure we have VFP.
if (!Subtarget->hasVFP2()) return false;
return false;
Value *Src = I->getOperand(0);
- EVT SrcVT = TLI.getValueType(Src->getType(), true);
+ EVT SrcEVT = TLI.getValueType(Src->getType(), true);
+ if (!SrcEVT.isSimple())
+ return false;
+ MVT SrcVT = SrcEVT.getSimpleVT();
if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8)
return false;
// Handle sign-extension.
if (SrcVT == MVT::i16 || SrcVT == MVT::i8) {
- EVT DestVT = MVT::i32;
- unsigned ResultReg = ARMEmitIntExt(SrcVT, SrcReg, DestVT, /*isZExt*/ false);
- if (ResultReg == 0) return false;
- SrcReg = ResultReg;
+ SrcReg = ARMEmitIntExt(SrcVT, SrcReg, MVT::i32,
+ /*isZExt*/!isSigned);
+ if (SrcReg == 0) return false;
}
// The conversion routine works on fp-reg to fp-reg and the operand above
if (FP == 0) return false;
unsigned Opc;
- if (Ty->isFloatTy()) Opc = ARM::VSITOS;
- else if (Ty->isDoubleTy()) Opc = ARM::VSITOD;
+ if (Ty->isFloatTy()) Opc = isSigned ? ARM::VSITOS : ARM::VUITOS;
+ else if (Ty->isDoubleTy()) Opc = isSigned ? ARM::VSITOD : ARM::VUITOD;
else return false;
unsigned ResultReg = createResultReg(TLI.getRegClassFor(DstVT));
return true;
}
-bool ARMFastISel::SelectFPToSI(const Instruction *I) {
+bool ARMFastISel::SelectFPToI(const Instruction *I, bool isSigned) {
// Make sure we have VFP.
if (!Subtarget->hasVFP2()) return false;
unsigned Opc;
Type *OpTy = I->getOperand(0)->getType();
- if (OpTy->isFloatTy()) Opc = ARM::VTOSIZS;
- else if (OpTy->isDoubleTy()) Opc = ARM::VTOSIZD;
+ if (OpTy->isFloatTy()) Opc = isSigned ? ARM::VTOSIZS : ARM::VTOUIZS;
+ else if (OpTy->isDoubleTy()) Opc = isSigned ? ARM::VTOSIZD : ARM::VTOUIZD;
else return false;
- // f64->s32 or f32->s32 both need an intermediate f32 reg.
+ // f64->s32/u32 or f32->s32/u32 both need an intermediate f32 reg.
unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::f32));
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
ResultReg)
// Things need to be register sized for register moves.
if (VT != MVT::i32) return false;
- const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
unsigned CondReg = getRegForValue(I->getOperand(0));
if (CondReg == 0) return false;
.addReg(CondReg).addImm(0));
unsigned MovCCOpc;
+ const TargetRegisterClass *RC;
if (!UseImm) {
+ RC = isThumb2 ? &ARM::tGPRRegClass : &ARM::GPRRegClass;
MovCCOpc = isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr;
} else {
- if (!isNegativeImm) {
+ RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass;
+ if (!isNegativeImm)
MovCCOpc = isThumb2 ? ARM::t2MOVCCi : ARM::MOVCCi;
- } else {
+ else
MovCCOpc = isThumb2 ? ARM::t2MVNCCi : ARM::MVNCCi;
- }
}
unsigned ResultReg = createResultReg(RC);
if (!UseImm)
return true;
}
-bool ARMFastISel::SelectSDiv(const Instruction *I) {
+bool ARMFastISel::SelectDiv(const Instruction *I, bool isSigned) {
MVT VT;
Type *Ty = I->getType();
if (!isTypeLegal(Ty, VT))
// Otherwise emit a libcall.
RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
if (VT == MVT::i8)
- LC = RTLIB::SDIV_I8;
+ LC = isSigned ? RTLIB::SDIV_I8 : RTLIB::UDIV_I8;
else if (VT == MVT::i16)
- LC = RTLIB::SDIV_I16;
+ LC = isSigned ? RTLIB::SDIV_I16 : RTLIB::UDIV_I16;
else if (VT == MVT::i32)
- LC = RTLIB::SDIV_I32;
+ LC = isSigned ? RTLIB::SDIV_I32 : RTLIB::UDIV_I32;
else if (VT == MVT::i64)
- LC = RTLIB::SDIV_I64;
+ LC = isSigned ? RTLIB::SDIV_I64 : RTLIB::UDIV_I64;
else if (VT == MVT::i128)
- LC = RTLIB::SDIV_I128;
+ LC = isSigned ? RTLIB::SDIV_I128 : RTLIB::UDIV_I128;
assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SDIV!");
return ARMEmitLibcall(I, LC);
}
-bool ARMFastISel::SelectSRem(const Instruction *I) {
+bool ARMFastISel::SelectRem(const Instruction *I, bool isSigned) {
MVT VT;
Type *Ty = I->getType();
if (!isTypeLegal(Ty, VT))
RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
if (VT == MVT::i8)
- LC = RTLIB::SREM_I8;
+ LC = isSigned ? RTLIB::SREM_I8 : RTLIB::UREM_I8;
else if (VT == MVT::i16)
- LC = RTLIB::SREM_I16;
+ LC = isSigned ? RTLIB::SREM_I16 : RTLIB::UREM_I16;
else if (VT == MVT::i32)
- LC = RTLIB::SREM_I32;
+ LC = isSigned ? RTLIB::SREM_I32 : RTLIB::UREM_I32;
else if (VT == MVT::i64)
- LC = RTLIB::SREM_I64;
+ LC = isSigned ? RTLIB::SREM_I64 : RTLIB::UREM_I64;
else if (VT == MVT::i128)
- LC = RTLIB::SREM_I128;
+ LC = isSigned ? RTLIB::SREM_I128 : RTLIB::UREM_I128;
assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SREM!");
return ARMEmitLibcall(I, LC);
}
-bool ARMFastISel::SelectBinaryOp(const Instruction *I, unsigned ISDOpcode) {
- EVT VT = TLI.getValueType(I->getType(), true);
+bool ARMFastISel::SelectBinaryIntOp(const Instruction *I, unsigned ISDOpcode) {
+ EVT DestVT = TLI.getValueType(I->getType(), true);
+
+ // We can get here in the case when we have a binary operation on a non-legal
+ // type and the target independent selector doesn't know how to handle it.
+ if (DestVT != MVT::i16 && DestVT != MVT::i8 && DestVT != MVT::i1)
+ return false;
+
+ unsigned Opc;
+ switch (ISDOpcode) {
+ default: return false;
+ case ISD::ADD:
+ Opc = isThumb2 ? ARM::t2ADDrr : ARM::ADDrr;
+ break;
+ case ISD::OR:
+ Opc = isThumb2 ? ARM::t2ORRrr : ARM::ORRrr;
+ break;
+ case ISD::SUB:
+ Opc = isThumb2 ? ARM::t2SUBrr : ARM::SUBrr;
+ break;
+ }
+
+ unsigned SrcReg1 = getRegForValue(I->getOperand(0));
+ if (SrcReg1 == 0) return false;
+
+ // TODO: Often the 2nd operand is an immediate, which can be encoded directly
+ // in the instruction, rather then materializing the value in a register.
+ unsigned SrcReg2 = getRegForValue(I->getOperand(1));
+ if (SrcReg2 == 0) return false;
+
+ unsigned ResultReg = createResultReg(&ARM::GPRnopcRegClass);
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(Opc), ResultReg)
+ .addReg(SrcReg1).addReg(SrcReg2));
+ UpdateValueMap(I, ResultReg);
+ return true;
+}
+
+bool ARMFastISel::SelectBinaryFPOp(const Instruction *I, unsigned ISDOpcode) {
+ EVT FPVT = TLI.getValueType(I->getType(), true);
+ if (!FPVT.isSimple()) return false;
+ MVT VT = FPVT.getSimpleVT();
// We can get here in the case when we want to use NEON for our fp
// operations, but can't figure out how to. Just use the vfp instructions
unsigned Op2 = getRegForValue(I->getOperand(1));
if (Op2 == 0) return false;
- unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
+ unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT.SimpleTy));
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(Opc), ResultReg)
.addReg(Op1).addReg(Op2));
// Call Handling Code
-// This is largely taken directly from CCAssignFnForNode - we don't support
-// varargs in FastISel so that part has been removed.
+// This is largely taken directly from CCAssignFnForNode
// TODO: We may not support all of this.
-CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC, bool Return) {
+CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC,
+ bool Return,
+ bool isVarArg) {
switch (CC) {
default:
llvm_unreachable("Unsupported calling convention");
case CallingConv::Fast:
- // Ignore fastcc. Silence compiler warnings.
- (void)RetFastCC_ARM_APCS;
- (void)FastCC_ARM_APCS;
+ if (Subtarget->hasVFP2() && !isVarArg) {
+ if (!Subtarget->isAAPCS_ABI())
+ return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
+ // For AAPCS ABI targets, just use VFP variant of the calling convention.
+ return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
+ }
// Fallthrough
case CallingConv::C:
// Use target triple & subtarget features to do actual dispatch.
if (Subtarget->isAAPCS_ABI()) {
if (Subtarget->hasVFP2() &&
- TM.Options.FloatABIType == FloatABI::Hard)
+ TM.Options.FloatABIType == FloatABI::Hard && !isVarArg)
return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
else
return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
} else
return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
case CallingConv::ARM_AAPCS_VFP:
- return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
+ if (!isVarArg)
+ return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
+ // Fall through to soft float variant, variadic functions don't
+ // use hard floating point ABI.
case CallingConv::ARM_AAPCS:
return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
case CallingConv::ARM_APCS:
return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
+ case CallingConv::GHC:
+ if (Return)
+ llvm_unreachable("Can't return in GHC call convention");
+ else
+ return CC_ARM_APCS_GHC;
}
}
SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
SmallVectorImpl<unsigned> &RegArgs,
CallingConv::ID CC,
- unsigned &NumBytes) {
+ unsigned &NumBytes,
+ bool isVarArg) {
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(CC, false, *FuncInfo.MF, TM, ArgLocs, *Context);
- CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC, false));
+ CCState CCInfo(CC, isVarArg, *FuncInfo.MF, TM, ArgLocs, *Context);
+ CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags,
+ CCAssignFnForCall(CC, false, isVarArg));
+
+ // Check that we can handle all of the arguments. If we can't, then bail out
+ // now before we add code to the MBB.
+ for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
+ CCValAssign &VA = ArgLocs[i];
+ MVT ArgVT = ArgVTs[VA.getValNo()];
+
+ // We don't handle NEON/vector parameters yet.
+ if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64)
+ return false;
+
+ // Now copy/store arg to correct locations.
+ if (VA.isRegLoc() && !VA.needsCustom()) {
+ continue;
+ } else if (VA.needsCustom()) {
+ // TODO: We need custom lowering for vector (v2f64) args.
+ if (VA.getLocVT() != MVT::f64 ||
+ // TODO: Only handle register args for now.
+ !VA.isRegLoc() || !ArgLocs[++i].isRegLoc())
+ return false;
+ } else {
+ switch (static_cast<EVT>(ArgVT).getSimpleVT().SimpleTy) {
+ default:
+ return false;
+ case MVT::i1:
+ case MVT::i8:
+ case MVT::i16:
+ case MVT::i32:
+ break;
+ case MVT::f32:
+ if (!Subtarget->hasVFP2())
+ return false;
+ break;
+ case MVT::f64:
+ if (!Subtarget->hasVFP2())
+ return false;
+ break;
+ }
+ }
+ }
+
+ // At the point, we are able to handle the call's arguments in fast isel.
// Get a count of how many bytes are to be pushed on the stack.
NumBytes = CCInfo.getNextStackOffset();
unsigned Arg = ArgRegs[VA.getValNo()];
MVT ArgVT = ArgVTs[VA.getValNo()];
- // We don't handle NEON/vector parameters yet.
- if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64)
- return false;
+ assert((!ArgVT.isVector() && ArgVT.getSizeInBits() <= 64) &&
+ "We don't handle NEON/vector parameters yet.");
// Handle arg promotion, etc.
switch (VA.getLocInfo()) {
case CCValAssign::Full: break;
case CCValAssign::SExt: {
MVT DestVT = VA.getLocVT();
- unsigned ResultReg = ARMEmitIntExt(ArgVT, Arg, DestVT,
- /*isZExt*/false);
- assert (ResultReg != 0 && "Failed to emit a sext");
- Arg = ResultReg;
+ Arg = ARMEmitIntExt(ArgVT, Arg, DestVT, /*isZExt*/false);
+ assert (Arg != 0 && "Failed to emit a sext");
ArgVT = DestVT;
break;
}
// Intentional fall-through. Handle AExt and ZExt.
case CCValAssign::ZExt: {
MVT DestVT = VA.getLocVT();
- unsigned ResultReg = ARMEmitIntExt(ArgVT, Arg, DestVT,
- /*isZExt*/true);
- assert (ResultReg != 0 && "Failed to emit a sext");
- Arg = ResultReg;
+ Arg = ARMEmitIntExt(ArgVT, Arg, DestVT, /*isZExt*/true);
+ assert (Arg != 0 && "Failed to emit a zext");
ArgVT = DestVT;
break;
}
RegArgs.push_back(VA.getLocReg());
} else if (VA.needsCustom()) {
// TODO: We need custom lowering for vector (v2f64) args.
- if (VA.getLocVT() != MVT::f64) return false;
+ assert(VA.getLocVT() == MVT::f64 &&
+ "Custom lowering for v2f64 args not available");
CCValAssign &NextVA = ArgLocs[++i];
- // TODO: Only handle register args for now.
- if(!(VA.isRegLoc() && NextVA.isRegLoc())) return false;
+ assert(VA.isRegLoc() && NextVA.isRegLoc() &&
+ "We only handle register args!");
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(ARM::VMOVRRD), VA.getLocReg())
Addr.Base.Reg = ARM::SP;
Addr.Offset = VA.getLocMemOffset();
- if (!ARMEmitStore(ArgVT, Arg, Addr)) return false;
+ bool EmitRet = ARMEmitStore(ArgVT, Arg, Addr); (void)EmitRet;
+ assert(EmitRet && "Could not emit a store for argument!");
}
}
+
return true;
}
bool ARMFastISel::FinishCall(MVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
const Instruction *I, CallingConv::ID CC,
- unsigned &NumBytes) {
+ unsigned &NumBytes, bool isVarArg) {
// Issue CALLSEQ_END
unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
// Now the return value.
if (RetVT != MVT::isVoid) {
SmallVector<CCValAssign, 16> RVLocs;
- CCState CCInfo(CC, false, *FuncInfo.MF, TM, RVLocs, *Context);
- CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true));
+ CCState CCInfo(CC, isVarArg, *FuncInfo.MF, TM, RVLocs, *Context);
+ CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true, isVarArg));
// Copy all of the result registers out of their specified physreg.
if (RVLocs.size() == 2 && RetVT == MVT::f64) {
// For this move we copy into two registers and then move into the
// double fp reg we want.
- EVT DestVT = RVLocs[0].getValVT();
- TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT);
+ MVT DestVT = RVLocs[0].getValVT();
+ const TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT);
unsigned ResultReg = createResultReg(DstRC);
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(ARM::VMOVDRR), ResultReg)
UpdateValueMap(I, ResultReg);
} else {
assert(RVLocs.size() == 1 &&"Can't handle non-double multi-reg retvals!");
- EVT CopyVT = RVLocs[0].getValVT();
+ MVT CopyVT = RVLocs[0].getValVT();
// Special handling for extended integers.
if (RetVT == MVT::i1 || RetVT == MVT::i8 || RetVT == MVT::i16)
CopyVT = MVT::i32;
- TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
+ const TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
unsigned ResultReg = createResultReg(DstRC);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
if (!FuncInfo.CanLowerReturn)
return false;
- if (F.isVarArg())
- return false;
+ // Build a list of return value registers.
+ SmallVector<unsigned, 4> RetRegs;
CallingConv::ID CC = F.getCallingConv();
if (Ret->getNumOperands() > 0) {
SmallVector<ISD::OutputArg, 4> Outs;
- GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
- Outs, TLI);
+ GetReturnInfo(F.getReturnType(), F.getAttributes(), Outs, TLI);
// Analyze operands of the call, assigning locations to each operand.
SmallVector<CCValAssign, 16> ValLocs;
CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs,I->getContext());
- CCInfo.AnalyzeReturn(Outs, CCAssignFnForCall(CC, true /* is Ret */));
+ CCInfo.AnalyzeReturn(Outs, CCAssignFnForCall(CC, true /* is Ret */,
+ F.isVarArg()));
const Value *RV = Ret->getOperand(0);
unsigned Reg = getRegForValue(RV);
return false;
unsigned SrcReg = Reg + VA.getValNo();
- EVT RVVT = TLI.getValueType(RV->getType());
- EVT DestVT = VA.getValVT();
+ EVT RVEVT = TLI.getValueType(RV->getType());
+ if (!RVEVT.isSimple()) return false;
+ MVT RVVT = RVEVT.getSimpleVT();
+ MVT DestVT = VA.getValVT();
// Special handling for extended integers.
if (RVVT != DestVT) {
if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
return false;
- if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
- return false;
-
assert(DestVT == MVT::i32 && "ARM should always ext to i32");
- bool isZExt = Outs[0].Flags.isZExt();
- unsigned ResultReg = ARMEmitIntExt(RVVT, SrcReg, DestVT, isZExt);
- if (ResultReg == 0) return false;
- SrcReg = ResultReg;
+ // Perform extension if flagged as either zext or sext. Otherwise, do
+ // nothing.
+ if (Outs[0].Flags.isZExt() || Outs[0].Flags.isSExt()) {
+ SrcReg = ARMEmitIntExt(RVVT, SrcReg, DestVT, Outs[0].Flags.isZExt());
+ if (SrcReg == 0) return false;
+ }
}
// Make the copy.
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
DstReg).addReg(SrcReg);
- // Mark the register as live out of the function.
- MRI.addLiveOut(VA.getLocReg());
+ // Add register to return instruction.
+ RetRegs.push_back(VA.getLocReg());
}
unsigned RetOpc = isThumb2 ? ARM::tBX_RET : ARM::BX_RET;
- AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(RetOpc)));
+ MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(RetOpc));
+ AddOptionalDefs(MIB);
+ for (unsigned i = 0, e = RetRegs.size(); i != e; ++i)
+ MIB.addReg(RetRegs[i], RegState::Implicit);
return true;
}
-unsigned ARMFastISel::ARMSelectCallOp(const GlobalValue *GV) {
+unsigned ARMFastISel::ARMSelectCallOp(bool UseReg) {
+ if (UseReg)
+ return isThumb2 ? ARM::tBLXr : ARM::BLX;
+ else
+ return isThumb2 ? ARM::tBL : ARM::BL;
+}
- // iOS needs the r9 versions of the opcodes.
- bool isiOS = Subtarget->isTargetIOS();
- if (isThumb2) {
- return isiOS ? ARM::tBLr9 : ARM::tBL;
- } else {
- return isiOS ? ARM::BLr9 : ARM::BL;
- }
+unsigned ARMFastISel::getLibcallReg(const Twine &Name) {
+ GlobalValue *GV = new GlobalVariable(Type::getInt32Ty(*Context), false,
+ GlobalValue::ExternalLinkage, 0, Name);
+ EVT LCREVT = TLI.getValueType(GV->getType());
+ if (!LCREVT.isSimple()) return 0;
+ return ARMMaterializeGV(GV, LCREVT.getSimpleVT());
}
// A quick function that will emit a call for a named libcall in F with the
else if (!isTypeLegal(RetTy, RetVT))
return false;
- // TODO: For now if we have long calls specified we don't handle the call.
- if (EnableARMLongCalls) return false;
+ // Can't handle non-double multi-reg retvals.
+ if (RetVT != MVT::isVoid && RetVT != MVT::i32) {
+ SmallVector<CCValAssign, 16> RVLocs;
+ CCState CCInfo(CC, false, *FuncInfo.MF, TM, RVLocs, *Context);
+ CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true, false));
+ if (RVLocs.size() >= 2 && RetVT != MVT::f64)
+ return false;
+ }
// Set up the argument vectors.
SmallVector<Value*, 8> Args;
// Handle the arguments now that we've gotten them.
SmallVector<unsigned, 4> RegArgs;
unsigned NumBytes;
- if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
+ if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags,
+ RegArgs, CC, NumBytes, false))
return false;
- // Issue the call, BLr9 for iOS, BL otherwise.
- // TODO: Turn this into the table of arm call ops.
- MachineInstrBuilder MIB;
- unsigned CallOpc = ARMSelectCallOp(NULL);
- if(isThumb2)
- // Explicitly adding the predicate here.
- MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(CallOpc)))
- .addExternalSymbol(TLI.getLibcallName(Call));
+ unsigned CalleeReg = 0;
+ if (EnableARMLongCalls) {
+ CalleeReg = getLibcallReg(TLI.getLibcallName(Call));
+ if (CalleeReg == 0) return false;
+ }
+
+ // Issue the call.
+ unsigned CallOpc = ARMSelectCallOp(EnableARMLongCalls);
+ MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
+ DL, TII.get(CallOpc));
+ // BL / BLX don't take a predicate, but tBL / tBLX do.
+ if (isThumb2)
+ AddDefaultPred(MIB);
+ if (EnableARMLongCalls)
+ MIB.addReg(CalleeReg);
else
- // Explicitly adding the predicate here.
- MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(CallOpc))
- .addExternalSymbol(TLI.getLibcallName(Call)));
+ MIB.addExternalSymbol(TLI.getLibcallName(Call));
// Add implicit physical register uses to the call.
for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
- MIB.addReg(RegArgs[i]);
+ MIB.addReg(RegArgs[i], RegState::Implicit);
+
+ // Add a register mask with the call-preserved registers.
+ // Proper defs for return values will be added by setPhysRegsDeadExcept().
+ MIB.addRegMask(TRI.getCallPreservedMask(CC));
// Finish off the call including any return values.
SmallVector<unsigned, 4> UsedRegs;
- if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
+ if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes, false)) return false;
// Set all unused physreg defs as dead.
static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
// Can't handle inline asm.
if (isa<InlineAsm>(Callee)) return false;
- // Only handle global variable Callees.
- const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
- if (!GV)
- return false;
+ // Allow SelectionDAG isel to handle tail calls.
+ if (CI->isTailCall()) return false;
// Check the calling convention.
ImmutableCallSite CS(CI);
// TODO: Avoid some calling conventions?
- // Let SDISel handle vararg functions.
PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
FunctionType *FTy = cast<FunctionType>(PT->getElementType());
- if (FTy->isVarArg())
- return false;
+ bool isVarArg = FTy->isVarArg();
// Handle *simple* calls for now.
Type *RetTy = I->getType();
RetVT != MVT::i8 && RetVT != MVT::i1)
return false;
- // TODO: For now if we have long calls specified we don't handle the call.
- if (EnableARMLongCalls) return false;
+ // Can't handle non-double multi-reg retvals.
+ if (RetVT != MVT::isVoid && RetVT != MVT::i1 && RetVT != MVT::i8 &&
+ RetVT != MVT::i16 && RetVT != MVT::i32) {
+ SmallVector<CCValAssign, 16> RVLocs;
+ CCState CCInfo(CC, isVarArg, *FuncInfo.MF, TM, RVLocs, *Context);
+ CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true, isVarArg));
+ if (RVLocs.size() >= 2 && RetVT != MVT::f64)
+ return false;
+ }
// Set up the argument vectors.
SmallVector<Value*, 8> Args;
SmallVector<unsigned, 8> ArgRegs;
SmallVector<MVT, 8> ArgVTs;
SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
- Args.reserve(CS.arg_size());
- ArgRegs.reserve(CS.arg_size());
- ArgVTs.reserve(CS.arg_size());
- ArgFlags.reserve(CS.arg_size());
+ unsigned arg_size = CS.arg_size();
+ Args.reserve(arg_size);
+ ArgRegs.reserve(arg_size);
+ ArgVTs.reserve(arg_size);
+ ArgFlags.reserve(arg_size);
for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
i != e; ++i) {
// If we're lowering a memory intrinsic instead of a regular call, skip the
// Handle the arguments now that we've gotten them.
SmallVector<unsigned, 4> RegArgs;
unsigned NumBytes;
- if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
+ if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags,
+ RegArgs, CC, NumBytes, isVarArg))
return false;
- // Issue the call, BLr9 for iOS, BL otherwise.
- // TODO: Turn this into the table of arm call ops.
- MachineInstrBuilder MIB;
- unsigned CallOpc = ARMSelectCallOp(GV);
- // Explicitly adding the predicate here.
- if(isThumb2) {
- // Explicitly adding the predicate here.
- MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(CallOpc)));
- if (!IntrMemName)
- MIB.addGlobalAddress(GV, 0, 0);
- else
- MIB.addExternalSymbol(IntrMemName, 0);
- } else {
- if (!IntrMemName)
- // Explicitly adding the predicate here.
- MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(CallOpc))
- .addGlobalAddress(GV, 0, 0));
+ bool UseReg = false;
+ const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
+ if (!GV || EnableARMLongCalls) UseReg = true;
+
+ unsigned CalleeReg = 0;
+ if (UseReg) {
+ if (IntrMemName)
+ CalleeReg = getLibcallReg(IntrMemName);
else
- MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(CallOpc))
- .addExternalSymbol(IntrMemName, 0));
+ CalleeReg = getRegForValue(Callee);
+
+ if (CalleeReg == 0) return false;
}
-
+
+ // Issue the call.
+ unsigned CallOpc = ARMSelectCallOp(UseReg);
+ MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
+ DL, TII.get(CallOpc));
+
+ // ARM calls don't take a predicate, but tBL / tBLX do.
+ if(isThumb2)
+ AddDefaultPred(MIB);
+ if (UseReg)
+ MIB.addReg(CalleeReg);
+ else if (!IntrMemName)
+ MIB.addGlobalAddress(GV, 0, 0);
+ else
+ MIB.addExternalSymbol(IntrMemName, 0);
+
// Add implicit physical register uses to the call.
for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
- MIB.addReg(RegArgs[i]);
+ MIB.addReg(RegArgs[i], RegState::Implicit);
+
+ // Add a register mask with the call-preserved registers.
+ // Proper defs for return values will be added by setPhysRegsDeadExcept().
+ MIB.addRegMask(TRI.getCallPreservedMask(CC));
// Finish off the call including any return values.
SmallVector<unsigned, 4> UsedRegs;
- if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;
+ if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes, isVarArg))
+ return false;
// Set all unused physreg defs as dead.
static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
return Len <= 16;
}
-bool ARMFastISel::ARMTryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len) {
+bool ARMFastISel::ARMTryEmitSmallMemCpy(Address Dest, Address Src,
+ uint64_t Len, unsigned Alignment) {
// Make sure we don't bloat code by inlining very large memcpy's.
if (!ARMIsMemCpySmall(Len))
return false;
- // We don't care about alignment here since we just emit integer accesses.
while (Len) {
MVT VT;
- if (Len >= 4)
- VT = MVT::i32;
- else if (Len >= 2)
- VT = MVT::i16;
- else {
- assert(Len == 1);
- VT = MVT::i8;
+ if (!Alignment || Alignment >= 4) {
+ if (Len >= 4)
+ VT = MVT::i32;
+ else if (Len >= 2)
+ VT = MVT::i16;
+ else {
+ assert (Len == 1 && "Expected a length of 1!");
+ VT = MVT::i8;
+ }
+ } else {
+ // Bound based on alignment.
+ if (Len >= 2 && Alignment == 2)
+ VT = MVT::i16;
+ else {
+ VT = MVT::i8;
+ }
}
bool RV;
assert (RV == true && "Should be able to handle this load.");
RV = ARMEmitStore(VT, ResultReg, Dest);
assert (RV == true && "Should be able to handle this store.");
+ (void)RV;
unsigned Size = VT.getSizeInBits()/8;
Len -= Size;
// FIXME: Handle more intrinsics.
switch (I.getIntrinsicID()) {
default: return false;
+ case Intrinsic::frameaddress: {
+ MachineFrameInfo *MFI = FuncInfo.MF->getFrameInfo();
+ MFI->setFrameAddressIsTaken(true);
+
+ unsigned LdrOpc;
+ const TargetRegisterClass *RC;
+ if (isThumb2) {
+ LdrOpc = ARM::t2LDRi12;
+ RC = (const TargetRegisterClass*)&ARM::tGPRRegClass;
+ } else {
+ LdrOpc = ARM::LDRi12;
+ RC = (const TargetRegisterClass*)&ARM::GPRRegClass;
+ }
+
+ const ARMBaseRegisterInfo *RegInfo =
+ static_cast<const ARMBaseRegisterInfo*>(TM.getRegisterInfo());
+ unsigned FramePtr = RegInfo->getFrameRegister(*(FuncInfo.MF));
+ unsigned SrcReg = FramePtr;
+
+ // Recursively load frame address
+ // ldr r0 [fp]
+ // ldr r0 [r0]
+ // ldr r0 [r0]
+ // ...
+ unsigned DestReg;
+ unsigned Depth = cast<ConstantInt>(I.getOperand(0))->getZExtValue();
+ while (Depth--) {
+ DestReg = createResultReg(RC);
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(LdrOpc), DestReg)
+ .addReg(SrcReg).addImm(0));
+ SrcReg = DestReg;
+ }
+ UpdateValueMap(&I, SrcReg);
+ return true;
+ }
case Intrinsic::memcpy:
case Intrinsic::memmove: {
const MemTransferInst &MTI = cast<MemTransferInst>(I);
if (!ARMComputeAddress(MTI.getRawDest(), Dest) ||
!ARMComputeAddress(MTI.getRawSource(), Src))
return false;
- if (ARMTryEmitSmallMemCpy(Dest, Src, Len))
+ unsigned Alignment = MTI.getAlignment();
+ if (ARMTryEmitSmallMemCpy(Dest, Src, Len, Alignment))
return true;
}
}
-
+
if (!MTI.getLength()->getType()->isIntegerTy(32))
return false;
-
+
if (MTI.getSourceAddressSpace() > 255 || MTI.getDestAddressSpace() > 255)
return false;
// Don't handle volatile.
if (MSI.isVolatile())
return false;
-
+
if (!MSI.getLength()->getType()->isIntegerTy(32))
return false;
-
+
if (MSI.getDestAddressSpace() > 255)
return false;
-
+
return SelectCall(&I, "memset");
}
+ case Intrinsic::trap: {
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(
+ Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP));
+ return true;
+ }
}
}
bool ARMFastISel::SelectTrunc(const Instruction *I) {
- // The high bits for a type smaller than the register size are assumed to be
+ // The high bits for a type smaller than the register size are assumed to be
// undefined.
Value *Op = I->getOperand(0);
return true;
}
-unsigned ARMFastISel::ARMEmitIntExt(EVT SrcVT, unsigned SrcReg, EVT DestVT,
+unsigned ARMFastISel::ARMEmitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
bool isZExt) {
if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8)
return 0;
+ if (SrcVT != MVT::i16 && SrcVT != MVT::i8 && SrcVT != MVT::i1)
+ return 0;
- unsigned Opc;
- bool isBoolZext = false;
- if (!SrcVT.isSimple()) return 0;
- switch (SrcVT.getSimpleVT().SimpleTy) {
- default: return 0;
- case MVT::i16:
- if (!Subtarget->hasV6Ops()) return 0;
- if (isZExt)
- Opc = isThumb2 ? ARM::t2UXTH : ARM::UXTH;
- else
- Opc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
- break;
- case MVT::i8:
- if (!Subtarget->hasV6Ops()) return 0;
- if (isZExt)
- Opc = isThumb2 ? ARM::t2UXTB : ARM::UXTB;
- else
- Opc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
- break;
- case MVT::i1:
- if (isZExt) {
- Opc = isThumb2 ? ARM::t2ANDri : ARM::ANDri;
- isBoolZext = true;
- break;
+ // Table of which combinations can be emitted as a single instruction,
+ // and which will require two.
+ static const uint8_t isSingleInstrTbl[3][2][2][2] = {
+ // ARM Thumb
+ // !hasV6Ops hasV6Ops !hasV6Ops hasV6Ops
+ // ext: s z s z s z s z
+ /* 1 */ { { { 0, 1 }, { 0, 1 } }, { { 0, 0 }, { 0, 1 } } },
+ /* 8 */ { { { 0, 1 }, { 1, 1 } }, { { 0, 0 }, { 1, 1 } } },
+ /* 16 */ { { { 0, 0 }, { 1, 1 } }, { { 0, 0 }, { 1, 1 } } }
+ };
+
+ // Target registers for:
+ // - For ARM can never be PC.
+ // - For 16-bit Thumb are restricted to lower 8 registers.
+ // - For 32-bit Thumb are restricted to non-SP and non-PC.
+ static const TargetRegisterClass *RCTbl[2][2] = {
+ // Instructions: Two Single
+ /* ARM */ { &ARM::GPRnopcRegClass, &ARM::GPRnopcRegClass },
+ /* Thumb */ { &ARM::tGPRRegClass, &ARM::rGPRRegClass }
+ };
+
+ // Table governing the instruction(s) to be emitted.
+ static const struct {
+ // First entry for each of the following is sext, second zext.
+ uint16_t Opc[2];
+ uint8_t Imm[2]; // All instructions have either a shift or a mask.
+ uint8_t hasS[2]; // Some instructions have an S bit, always set it to 0.
+ } OpcTbl[2][2][3] = {
+ { // Two instructions (first is left shift, second is in this table).
+ { // ARM
+ /* 1 */ { { ARM::ASRi, ARM::LSRi }, { 31, 31 }, { 1, 1 } },
+ /* 8 */ { { ARM::ASRi, ARM::LSRi }, { 24, 24 }, { 1, 1 } },
+ /* 16 */ { { ARM::ASRi, ARM::LSRi }, { 16, 16 }, { 1, 1 } }
+ },
+ { // Thumb
+ /* 1 */ { { ARM::tASRri, ARM::tLSRri }, { 31, 31 }, { 0, 0 } },
+ /* 8 */ { { ARM::tASRri, ARM::tLSRri }, { 24, 24 }, { 0, 0 } },
+ /* 16 */ { { ARM::tASRri, ARM::tLSRri }, { 16, 16 }, { 0, 0 } }
+ }
+ },
+ { // Single instruction.
+ { // ARM
+ /* 1 */ { { ARM::KILL, ARM::ANDri }, { 0, 1 }, { 0, 1 } },
+ /* 8 */ { { ARM::SXTB, ARM::ANDri }, { 0, 255 }, { 0, 1 } },
+ /* 16 */ { { ARM::SXTH, ARM::UXTH }, { 0, 0 }, { 0, 0 } }
+ },
+ { // Thumb
+ /* 1 */ { { ARM::KILL, ARM::t2ANDri }, { 0, 1 }, { 0, 1 } },
+ /* 8 */ { { ARM::t2SXTB, ARM::t2ANDri }, { 0, 255 }, { 0, 1 } },
+ /* 16 */ { { ARM::t2SXTH, ARM::t2UXTH }, { 0, 0 }, { 0, 0 } }
+ }
}
- return 0;
+ };
+
+ unsigned SrcBits = SrcVT.getSizeInBits();
+ unsigned DestBits = DestVT.getSizeInBits();
+ (void) DestBits;
+ assert((SrcBits < DestBits) && "can only extend to larger types");
+ assert((DestBits == 32 || DestBits == 16 || DestBits == 8) &&
+ "other sizes unimplemented");
+ assert((SrcBits == 16 || SrcBits == 8 || SrcBits == 1) &&
+ "other sizes unimplemented");
+
+ bool hasV6Ops = Subtarget->hasV6Ops();
+ unsigned Bitness = countTrailingZeros(SrcBits) >> 1; // {1,8,16}=>{0,1,2}
+ assert((Bitness < 3) && "sanity-check table bounds");
+
+ bool isSingleInstr = isSingleInstrTbl[Bitness][isThumb2][hasV6Ops][isZExt];
+ const TargetRegisterClass *RC = RCTbl[isThumb2][isSingleInstr];
+ unsigned Opc = OpcTbl[isSingleInstr][isThumb2][Bitness].Opc[isZExt];
+ assert(ARM::KILL != Opc && "Invalid table entry");
+ unsigned Imm = OpcTbl[isSingleInstr][isThumb2][Bitness].Imm[isZExt];
+ unsigned hasS = OpcTbl[isSingleInstr][isThumb2][Bitness].hasS[isZExt];
+
+ // 16-bit Thumb instructions always set CPSR (unless they're in an IT block).
+ bool setsCPSR = &ARM::tGPRRegClass == RC;
+ unsigned LSLOpc = isThumb2 ? ARM::tLSLri : ARM::LSLi;
+ unsigned ResultReg;
+
+ // Either one or two instructions are emitted.
+ // They're always of the form:
+ // dst = in OP imm
+ // CPSR is set only by 16-bit Thumb instructions.
+ // Predicate, if any, is AL.
+ // S bit, if available, is always 0.
+ // When two are emitted the first's result will feed as the second's input,
+ // that value is then dead.
+ unsigned NumInstrsEmitted = isSingleInstr ? 1 : 2;
+ for (unsigned Instr = 0; Instr != NumInstrsEmitted; ++Instr) {
+ ResultReg = createResultReg(RC);
+ unsigned Opcode = ((0 == Instr) && !isSingleInstr) ? LSLOpc : Opc;
+ bool isKill = 1 == Instr;
+ MachineInstrBuilder MIB = BuildMI(
+ *FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opcode), ResultReg);
+ if (setsCPSR)
+ MIB.addReg(ARM::CPSR, RegState::Define);
+ AddDefaultPred(MIB.addReg(SrcReg, isKill * RegState::Kill).addImm(Imm));
+ if (hasS)
+ AddDefaultCC(MIB);
+ // Second instruction consumes the first's result.
+ SrcReg = ResultReg;
}
- unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::i32));
- MachineInstrBuilder MIB;
- MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg)
- .addReg(SrcReg);
- if (isBoolZext)
- MIB.addImm(1);
- else
- MIB.addImm(0);
- AddOptionalDefs(MIB);
return ResultReg;
}
Value *Src = I->getOperand(0);
Type *SrcTy = Src->getType();
- EVT SrcVT, DestVT;
- SrcVT = TLI.getValueType(SrcTy, true);
- DestVT = TLI.getValueType(DestTy, true);
-
bool isZExt = isa<ZExtInst>(I);
unsigned SrcReg = getRegForValue(Src);
if (!SrcReg) return false;
+ EVT SrcEVT, DestEVT;
+ SrcEVT = TLI.getValueType(SrcTy, true);
+ DestEVT = TLI.getValueType(DestTy, true);
+ if (!SrcEVT.isSimple()) return false;
+ if (!DestEVT.isSimple()) return false;
+
+ MVT SrcVT = SrcEVT.getSimpleVT();
+ MVT DestVT = DestEVT.getSimpleVT();
unsigned ResultReg = ARMEmitIntExt(SrcVT, SrcReg, DestVT, isZExt);
if (ResultReg == 0) return false;
UpdateValueMap(I, ResultReg);
return true;
}
+bool ARMFastISel::SelectShift(const Instruction *I,
+ ARM_AM::ShiftOpc ShiftTy) {
+ // We handle thumb2 mode by target independent selector
+ // or SelectionDAG ISel.
+ if (isThumb2)
+ return false;
+
+ // Only handle i32 now.
+ EVT DestVT = TLI.getValueType(I->getType(), true);
+ if (DestVT != MVT::i32)
+ return false;
+
+ unsigned Opc = ARM::MOVsr;
+ unsigned ShiftImm;
+ Value *Src2Value = I->getOperand(1);
+ if (const ConstantInt *CI = dyn_cast<ConstantInt>(Src2Value)) {
+ ShiftImm = CI->getZExtValue();
+
+ // Fall back to selection DAG isel if the shift amount
+ // is zero or greater than the width of the value type.
+ if (ShiftImm == 0 || ShiftImm >=32)
+ return false;
+
+ Opc = ARM::MOVsi;
+ }
+
+ Value *Src1Value = I->getOperand(0);
+ unsigned Reg1 = getRegForValue(Src1Value);
+ if (Reg1 == 0) return false;
+
+ unsigned Reg2 = 0;
+ if (Opc == ARM::MOVsr) {
+ Reg2 = getRegForValue(Src2Value);
+ if (Reg2 == 0) return false;
+ }
+
+ unsigned ResultReg = createResultReg(&ARM::GPRnopcRegClass);
+ if(ResultReg == 0) return false;
+
+ MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(Opc), ResultReg)
+ .addReg(Reg1);
+
+ if (Opc == ARM::MOVsi)
+ MIB.addImm(ARM_AM::getSORegOpc(ShiftTy, ShiftImm));
+ else if (Opc == ARM::MOVsr) {
+ MIB.addReg(Reg2);
+ MIB.addImm(ARM_AM::getSORegOpc(ShiftTy, 0));
+ }
+
+ AddOptionalDefs(MIB);
+ UpdateValueMap(I, ResultReg);
+ return true;
+}
+
// TODO: SoftFP support.
bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
return SelectStore(I);
case Instruction::Br:
return SelectBranch(I);
+ case Instruction::IndirectBr:
+ return SelectIndirectBr(I);
case Instruction::ICmp:
case Instruction::FCmp:
return SelectCmp(I);
case Instruction::FPTrunc:
return SelectFPTrunc(I);
case Instruction::SIToFP:
- return SelectSIToFP(I);
+ return SelectIToFP(I, /*isSigned*/ true);
+ case Instruction::UIToFP:
+ return SelectIToFP(I, /*isSigned*/ false);
case Instruction::FPToSI:
- return SelectFPToSI(I);
+ return SelectFPToI(I, /*isSigned*/ true);
+ case Instruction::FPToUI:
+ return SelectFPToI(I, /*isSigned*/ false);
+ case Instruction::Add:
+ return SelectBinaryIntOp(I, ISD::ADD);
+ case Instruction::Or:
+ return SelectBinaryIntOp(I, ISD::OR);
+ case Instruction::Sub:
+ return SelectBinaryIntOp(I, ISD::SUB);
case Instruction::FAdd:
- return SelectBinaryOp(I, ISD::FADD);
+ return SelectBinaryFPOp(I, ISD::FADD);
case Instruction::FSub:
- return SelectBinaryOp(I, ISD::FSUB);
+ return SelectBinaryFPOp(I, ISD::FSUB);
case Instruction::FMul:
- return SelectBinaryOp(I, ISD::FMUL);
+ return SelectBinaryFPOp(I, ISD::FMUL);
case Instruction::SDiv:
- return SelectSDiv(I);
+ return SelectDiv(I, /*isSigned*/ true);
+ case Instruction::UDiv:
+ return SelectDiv(I, /*isSigned*/ false);
case Instruction::SRem:
- return SelectSRem(I);
+ return SelectRem(I, /*isSigned*/ true);
+ case Instruction::URem:
+ return SelectRem(I, /*isSigned*/ false);
case Instruction::Call:
if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
return SelectIntrinsicCall(*II);
case Instruction::ZExt:
case Instruction::SExt:
return SelectIntExt(I);
+ case Instruction::Shl:
+ return SelectShift(I, ARM_AM::lsl);
+ case Instruction::LShr:
+ return SelectShift(I, ARM_AM::lsr);
+ case Instruction::AShr:
+ return SelectShift(I, ARM_AM::asr);
default: break;
}
return false;
}
-/// TryToFoldLoad - The specified machine instr operand is a vreg, and that
+namespace {
+// This table describes sign- and zero-extend instructions which can be
+// folded into a preceding load. All of these extends have an immediate
+// (sometimes a mask and sometimes a shift) that's applied after
+// extension.
+const struct FoldableLoadExtendsStruct {
+ uint16_t Opc[2]; // ARM, Thumb.
+ uint8_t ExpectedImm;
+ uint8_t isZExt : 1;
+ uint8_t ExpectedVT : 7;
+} FoldableLoadExtends[] = {
+ { { ARM::SXTH, ARM::t2SXTH }, 0, 0, MVT::i16 },
+ { { ARM::UXTH, ARM::t2UXTH }, 0, 1, MVT::i16 },
+ { { ARM::ANDri, ARM::t2ANDri }, 255, 1, MVT::i8 },
+ { { ARM::SXTB, ARM::t2SXTB }, 0, 0, MVT::i8 },
+ { { ARM::UXTB, ARM::t2UXTB }, 0, 1, MVT::i8 }
+};
+}
+
+/// \brief The specified machine instr operand is a vreg, and that
/// vreg is being provided by the specified load instruction. If possible,
/// try to fold the load as an operand to the instruction, returning true if
/// successful.
-bool ARMFastISel::TryToFoldLoad(MachineInstr *MI, unsigned OpNo,
- const LoadInst *LI) {
+bool ARMFastISel::tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
+ const LoadInst *LI) {
// Verify we have a legal type before going any further.
MVT VT;
if (!isLoadTypeLegal(LI->getType(), VT))
// ldrb r1, [r0] ldrb r1, [r0]
// uxtb r2, r1 =>
// mov r3, r2 mov r3, r1
- bool isZExt = true;
- switch(MI->getOpcode()) {
- default: return false;
- case ARM::SXTH:
- case ARM::t2SXTH:
- isZExt = false;
- case ARM::UXTH:
- case ARM::t2UXTH:
- if (VT != MVT::i16)
- return false;
- break;
- case ARM::SXTB:
- case ARM::t2SXTB:
- isZExt = false;
- case ARM::UXTB:
- case ARM::t2UXTB:
- if (VT != MVT::i8)
- return false;
- break;
+ if (MI->getNumOperands() < 3 || !MI->getOperand(2).isImm())
+ return false;
+ const uint64_t Imm = MI->getOperand(2).getImm();
+
+ bool Found = false;
+ bool isZExt;
+ for (unsigned i = 0, e = array_lengthof(FoldableLoadExtends);
+ i != e; ++i) {
+ if (FoldableLoadExtends[i].Opc[isThumb2] == MI->getOpcode() &&
+ (uint64_t)FoldableLoadExtends[i].ExpectedImm == Imm &&
+ MVT((MVT::SimpleValueType)FoldableLoadExtends[i].ExpectedVT) == VT) {
+ Found = true;
+ isZExt = FoldableLoadExtends[i].isZExt;
+ }
}
+ if (!Found) return false;
+
// See if we can handle this address.
Address Addr;
if (!ARMComputeAddress(LI->getOperand(0), Addr)) return false;
-
+
unsigned ResultReg = MI->getOperand(0).getReg();
if (!ARMEmitLoad(VT, ResultReg, Addr, LI->getAlignment(), isZExt, false))
return false;
return true;
}
+unsigned ARMFastISel::ARMLowerPICELF(const GlobalValue *GV,
+ unsigned Align, MVT VT) {
+ bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
+ ARMConstantPoolConstant *CPV =
+ ARMConstantPoolConstant::Create(GV, UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
+ unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
+
+ unsigned Opc;
+ unsigned DestReg1 = createResultReg(TLI.getRegClassFor(VT));
+ // Load value.
+ if (isThumb2) {
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(ARM::t2LDRpci), DestReg1)
+ .addConstantPoolIndex(Idx));
+ Opc = UseGOTOFF ? ARM::t2ADDrr : ARM::t2LDRs;
+ } else {
+ // The extra immediate is for addrmode2.
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
+ DL, TII.get(ARM::LDRcp), DestReg1)
+ .addConstantPoolIndex(Idx).addImm(0));
+ Opc = UseGOTOFF ? ARM::ADDrr : ARM::LDRrs;
+ }
+
+ unsigned GlobalBaseReg = AFI->getGlobalBaseReg();
+ if (GlobalBaseReg == 0) {
+ GlobalBaseReg = MRI.createVirtualRegister(TLI.getRegClassFor(VT));
+ AFI->setGlobalBaseReg(GlobalBaseReg);
+ }
+
+ unsigned DestReg2 = createResultReg(TLI.getRegClassFor(VT));
+ MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
+ DL, TII.get(Opc), DestReg2)
+ .addReg(DestReg1)
+ .addReg(GlobalBaseReg);
+ if (!UseGOTOFF)
+ MIB.addImm(0);
+ AddOptionalDefs(MIB);
+
+ return DestReg2;
+}
+
+bool ARMFastISel::FastLowerArguments() {
+ if (!FuncInfo.CanLowerReturn)
+ return false;
+
+ const Function *F = FuncInfo.Fn;
+ if (F->isVarArg())
+ return false;
+
+ CallingConv::ID CC = F->getCallingConv();
+ switch (CC) {
+ default:
+ return false;
+ case CallingConv::Fast:
+ case CallingConv::C:
+ case CallingConv::ARM_AAPCS_VFP:
+ case CallingConv::ARM_AAPCS:
+ case CallingConv::ARM_APCS:
+ break;
+ }
+
+ // Only handle simple cases. i.e. Up to 4 i8/i16/i32 scalar arguments
+ // which are passed in r0 - r3.
+ unsigned Idx = 1;
+ for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
+ I != E; ++I, ++Idx) {
+ if (Idx > 4)
+ return false;
+
+ if (F->getAttributes().hasAttribute(Idx, Attribute::InReg) ||
+ F->getAttributes().hasAttribute(Idx, Attribute::StructRet) ||
+ F->getAttributes().hasAttribute(Idx, Attribute::ByVal))
+ return false;
+
+ Type *ArgTy = I->getType();
+ if (ArgTy->isStructTy() || ArgTy->isArrayTy() || ArgTy->isVectorTy())
+ return false;
+
+ EVT ArgVT = TLI.getValueType(ArgTy);
+ if (!ArgVT.isSimple()) return false;
+ switch (ArgVT.getSimpleVT().SimpleTy) {
+ case MVT::i8:
+ case MVT::i16:
+ case MVT::i32:
+ break;
+ default:
+ return false;
+ }
+ }
+
+
+ static const uint16_t GPRArgRegs[] = {
+ ARM::R0, ARM::R1, ARM::R2, ARM::R3
+ };
+
+ const TargetRegisterClass *RC = TLI.getRegClassFor(MVT::i32);
+ Idx = 0;
+ for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
+ I != E; ++I, ++Idx) {
+ unsigned SrcReg = GPRArgRegs[Idx];
+ unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC);
+ // FIXME: Unfortunately it's necessary to emit a copy from the livein copy.
+ // Without this, EmitLiveInCopies may eliminate the livein if its only
+ // use is a bitcast (which isn't turned into an instruction).
+ unsigned ResultReg = createResultReg(RC);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
+ ResultReg).addReg(DstReg, getKillRegState(true));
+ UpdateValueMap(I, ResultReg);
+ }
+
+ return true;
+}
+
namespace llvm {
- llvm::FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo) {
- // Completely untested on non-iOS.
+ FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo,
+ const TargetLibraryInfo *libInfo) {
const TargetMachine &TM = funcInfo.MF->getTarget();
- // Darwin and thumb1 only for now.
const ARMSubtarget *Subtarget = &TM.getSubtarget<ARMSubtarget>();
- if (Subtarget->isTargetIOS() && !Subtarget->isThumb1Only() &&
- !DisableARMFastISel)
- return new ARMFastISel(funcInfo);
+ // Thumb2 support on iOS; ARM support on iOS, Linux and NaCl.
+ bool UseFastISel = false;
+ UseFastISel |= Subtarget->isTargetIOS() && !Subtarget->isThumb1Only();
+ UseFastISel |= Subtarget->isTargetLinux() && !Subtarget->isThumb();
+ UseFastISel |= Subtarget->isTargetNaCl() && !Subtarget->isThumb();
+
+ if (UseFastISel) {
+ // iOS always has a FP for backtracking, force other targets
+ // to keep their FP when doing FastISel. The emitted code is
+ // currently superior, and in cases like test-suite's lencod
+ // FastISel isn't quite correct when FP is eliminated.
+ TM.Options.NoFramePointerElim = true;
+ return new ARMFastISel(funcInfo, libInfo);
+ }
return 0;
}
}