#include "ARM.h"
#include "ARMBaseInstrInfo.h"
#include "ARMCallingConv.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/DataLayout.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetMachine.h"
/// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
/// make the right decision when generating code for different targets.
const ARMSubtarget *Subtarget;
+ Module &M;
const TargetMachine &TM;
const TargetInstrInfo &TII;
const TargetLowering &TLI;
explicit ARMFastISel(FunctionLoweringInfo &funcInfo,
const TargetLibraryInfo *libInfo)
: FastISel(funcInfo, libInfo),
+ M(const_cast<Module&>(*funcInfo.Fn->getParent())),
TM(funcInfo.MF->getTarget()),
TII(*TM.getInstrInfo()),
TLI(*TM.getTargetLowering()) {
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"
// Utility routines.
private:
+ unsigned constrainOperandRegClass(const MCInstrDesc &II, unsigned OpNum,
+ unsigned Op);
bool isTypeLegal(Type *Ty, MVT &VT);
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);
+ 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, EVT VT);
+ unsigned ARMLowerPICELF(const GlobalValue *GV, unsigned Align, MVT VT);
// Call handling routines.
private:
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);
};
bool ARMFastISel::isARMNEONPred(const MachineInstr *MI) {
const MCInstrDesc &MCID = MI->getDesc();
- // If we're a thumb2 or not NEON function we were handled via isPredicable.
+ // If we're a thumb2 or not NEON function we'll be handled via isPredicable.
if ((MCID.TSFlags & ARMII::DomainMask) != ARMII::DomainNEON ||
AFI->isThumb2Function())
- return false;
+ return MI->isPredicable();
for (unsigned i = 0, e = MCID.getNumOperands(); i != e; ++i)
if (MCID.OpInfo[i].isPredicate())
// Do we use a predicate? or...
// Are we NEON in ARM mode and have a predicate operand? If so, I know
// we're not predicable but add it anyways.
- if (TII.isPredicable(MI) || isARMNEONPred(MI))
+ if (isARMNEONPred(MI))
AddDefaultPred(MIB);
// Do we optionally set a predicate? Preds is size > 0 iff the predicate
return MIB;
}
+unsigned ARMFastISel::constrainOperandRegClass(const MCInstrDesc &II,
+ unsigned Op, unsigned OpNum) {
+ if (TargetRegisterInfo::isVirtualRegister(Op)) {
+ const TargetRegisterClass *RegClass =
+ TII.getRegClass(II, OpNum, &TRI, *FuncInfo.MF);
+ if (!MRI.constrainRegClass(Op, RegClass)) {
+ // If it's not legal to COPY between the register classes, something
+ // has gone very wrong before we got here.
+ unsigned NewOp = createResultReg(RegClass);
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(TargetOpcode::COPY), NewOp).addReg(Op));
+ return NewOp;
+ }
+ }
+ return Op;
+}
+
unsigned ARMFastISel::FastEmitInst_(unsigned MachineInstOpcode,
const TargetRegisterClass* RC) {
unsigned ResultReg = createResultReg(RC);
unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
+ // Make sure the input operand is sufficiently constrained to be legal
+ // for this instruction.
+ Op0 = constrainOperandRegClass(II, Op0, 1);
if (II.getNumDefs() >= 1) {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill));
unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
+ // Make sure the input operands are sufficiently constrained to be legal
+ // for this instruction.
+ Op0 = constrainOperandRegClass(II, Op0, 1);
+ Op1 = constrainOperandRegClass(II, Op1, 2);
+
if (II.getNumDefs() >= 1) {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill)
unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
+ // Make sure the input operands are sufficiently constrained to be legal
+ // for this instruction.
+ Op0 = constrainOperandRegClass(II, Op0, 1);
+ Op1 = constrainOperandRegClass(II, Op1, 2);
+ Op2 = constrainOperandRegClass(II, Op1, 3);
+
if (II.getNumDefs() >= 1) {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill)
unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
+ // Make sure the input operand is sufficiently constrained to be legal
+ // for this instruction.
+ Op0 = constrainOperandRegClass(II, Op0, 1);
if (II.getNumDefs() >= 1) {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill)
unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
+ // Make sure the input operand is sufficiently constrained to be legal
+ // for this instruction.
+ Op0 = constrainOperandRegClass(II, Op0, 1);
if (II.getNumDefs() >= 1) {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill)
unsigned ResultReg = createResultReg(RC);
const MCInstrDesc &II = TII.get(MachineInstOpcode);
+ // Make sure the input operands are sufficiently constrained to be legal
+ // for this instruction.
+ Op0 = constrainOperandRegClass(II, Op0, 1);
+ Op1 = constrainOperandRegClass(II, Op1, 2);
if (II.getNumDefs() >= 1) {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
.addReg(Op0, Op0IsKill * RegState::Kill)
// 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));
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));
// 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()));
.addConstantPoolIndex(Idx));
else
// The extra immediate is for addrmode2.
+ DestReg = constrainOperandRegClass(TII.get(ARM::LDRcp), DestReg, 0);
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(ARM::LDRcp), DestReg)
.addConstantPoolIndex(Idx)
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;
(const TargetRegisterClass*)&ARM::GPRRegClass;
unsigned DestReg = createResultReg(RC);
+ // FastISel TLS support on non-MachO is broken, punt to SelectionDAG.
+ const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
+ bool IsThreadLocal = GVar && GVar->isThreadLocal();
+ if (!Subtarget->isTargetMachO() && IsThreadLocal) return 0;
+
// Use movw+movt when possible, it avoids constant pool entries.
- // Darwin targets don't support movt with Reloc::Static, see
- // ARMTargetLowering::LowerGlobalAddressDarwin. Other targets only support
- // static movt relocations.
+ // Non-darwin targets only support static movt relocations in FastISel.
if (Subtarget->useMovt() &&
- Subtarget->isTargetDarwin() == (RelocM != Reloc::Static)) {
+ (Subtarget->isTargetMachO() || RelocM == Reloc::Static)) {
unsigned Opc;
+ unsigned char TF = 0;
+ if (Subtarget->isTargetMachO())
+ TF = ARMII::MO_NONLAZY;
+
switch (RelocM) {
case Reloc::PIC_:
Opc = isThumb2 ? ARM::t2MOV_ga_pcrel : ARM::MOV_ga_pcrel;
break;
- case Reloc::DynamicNoPIC:
- Opc = isThumb2 ? ARM::t2MOV_ga_dyn : ARM::MOV_ga_dyn;
- break;
default:
Opc = isThumb2 ? ARM::t2MOVi32imm : ARM::MOVi32imm;
break;
}
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
- DestReg).addGlobalAddress(GV));
+ DestReg).addGlobalAddress(GV, 0, TF));
} else {
// MachineConstantPool wants an explicit alignment.
unsigned Align = TD.getPrefTypeAlignment(GV->getType());
AddOptionalDefs(MIB);
} else {
// The extra immediate is for addrmode2.
+ DestReg = constrainOperandRegClass(TII.get(ARM::LDRcp), DestReg, 0);
MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp),
DestReg)
.addConstantPoolIndex(Idx)
}
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);
// This will get lowered later into the correct offsets and registers
// via rewriteXFrameIndex.
if (SI != FuncInfo.StaticAllocaMap.end()) {
+ unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri;
const TargetRegisterClass* RC = TLI.getRegClassFor(VT);
unsigned ResultReg = createResultReg(RC);
- unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri;
+ ResultReg = constrainOperandRegClass(TII.get(Opc), ResultReg, 0);
+
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(Opc), ResultReg)
.addFrameIndex(SI->second)
switch (Opcode) {
default:
break;
- case Instruction::BitCast: {
+ case Instruction::BitCast:
// Look through bitcasts.
return ARMComputeAddress(U->getOperand(0), Addr);
- }
- case Instruction::IntToPtr: {
+ case Instruction::IntToPtr:
// Look past no-op inttoptrs.
if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
return ARMComputeAddress(U->getOperand(0), Addr);
break;
- }
- case Instruction::PtrToInt: {
+ case Instruction::PtrToInt:
// Look past no-op ptrtoints.
if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
return ARMComputeAddress(U->getOperand(0), Addr);
break;
- }
case Instruction::GetElementPtr: {
Address SavedAddr = Addr;
int TmpOffset = Addr.Offset;
TmpOffset += CI->getSExtValue() * S;
break;
}
- if (isa<AddOperator>(Op) &&
- (!isa<Instruction>(Op) ||
- FuncInfo.MBBMap[cast<Instruction>(Op)->getParent()]
- == FuncInfo.MBB) &&
- isa<ConstantInt>(cast<AddOperator>(Op)->getOperand(1))) {
- // An add (in the same block) with a constant operand. Fold the
- // constant.
+ if (canFoldAddIntoGEP(U, Op)) {
+ // A compatible add with a constant operand. Fold the constant.
ConstantInt *CI =
cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
TmpOffset += CI->getSExtValue() * S;
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) {
+ switch (VT.SimpleTy) {
default: llvm_unreachable("Unhandled load/store type!");
case MVT::i1:
case MVT::i8:
}
}
-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;
const TargetRegisterClass *RC;
- switch (VT.getSimpleVT().SimpleTy) {
+ switch (VT.SimpleTy) {
// This is mostly going to be Neon/vector support.
default: return false;
case MVT::i1:
useAM3 = true;
}
}
- RC = &ARM::GPRRegClass;
+ RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass;
break;
case MVT::i16:
- if (Alignment && Alignment < 2 && (!Subtarget->allowsUnalignedMem() ||
- TM.Options.StrictAlign))
+ if (Alignment && Alignment < 2 && !Subtarget->allowsUnalignedMem())
return false;
if (isThumb2) {
Opc = isZExt ? ARM::LDRH : ARM::LDRSH;
useAM3 = true;
}
- RC = &ARM::GPRRegClass;
+ RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass;
break;
case MVT::i32:
- if (Alignment && Alignment < 4 && (!Subtarget->allowsUnalignedMem() ||
- TM.Options.StrictAlign))
+ if (Alignment && Alignment < 4 && !Subtarget->allowsUnalignedMem())
return false;
if (isThumb2) {
} else {
Opc = ARM::LDRi12;
}
- RC = &ARM::GPRRegClass;
+ 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::GPRRegClass;
+ 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: {
(const TargetRegisterClass*)&ARM::tGPRRegClass :
(const TargetRegisterClass*)&ARM::GPRRegClass);
unsigned Opc = isThumb2 ? ARM::t2ANDri : ARM::ANDri;
+ SrcReg = constrainOperandRegClass(TII.get(Opc), SrcReg, 1);
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(Opc), Res)
.addReg(SrcReg).addImm(1));
}
break;
case MVT::i16:
- if (Alignment && Alignment < 2 && (!Subtarget->allowsUnalignedMem() ||
- TM.Options.StrictAlign))
+ if (Alignment && Alignment < 2 && !Subtarget->allowsUnalignedMem())
return false;
if (isThumb2) {
}
break;
case MVT::i32:
- if (Alignment && Alignment < 4 && (!Subtarget->allowsUnalignedMem() ||
- TM.Options.StrictAlign))
+ if (Alignment && Alignment < 4 && !Subtarget->allowsUnalignedMem())
return false;
if (isThumb2) {
ARMSimplifyAddress(Addr, VT, useAM3);
// Create the base instruction, then add the operands.
+ SrcReg = constrainOperandRegClass(TII.get(StrOpc), SrcReg, 0);
MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(StrOpc))
.addReg(SrcReg);
(isLoadTypeLegal(TI->getOperand(0)->getType(), SourceVT))) {
unsigned TstOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
unsigned OpReg = getRegForValue(TI->getOperand(0));
+ OpReg = constrainOperandRegClass(TII.get(TstOpc), OpReg, 0);
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(TstOpc))
.addReg(OpReg).addImm(1));
// and it left a value for us in a virtual register. Ergo, we test
// the one-bit value left in the virtual register.
unsigned TstOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
+ CmpReg = constrainOperandRegClass(TII.get(TstOpc), CmpReg, 0);
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TstOpc))
.addReg(CmpReg).addImm(1));
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())
unsigned CmpOpc;
bool isICmp = true;
bool needsExt = false;
- switch (SrcVT.getSimpleVT().SimpleTy) {
+ switch (SrcVT.SimpleTy) {
default: return false;
// TODO: Verify compares.
case MVT::f32:
}
}
+ const MCInstrDesc &II = TII.get(CmpOpc);
+ SrcReg1 = constrainOperandRegClass(II, SrcReg1, 0);
if (!UseImm) {
- AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(CmpOpc))
+ SrcReg2 = constrainOperandRegClass(II, SrcReg2, 1);
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
.addReg(SrcReg1).addReg(SrcReg2));
} else {
MachineInstrBuilder MIB;
- MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
+ MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
.addReg(SrcReg1);
// Only add immediate for icmp as the immediate for fcmp is an implicit 0.0.
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;
- SrcReg = ARMEmitIntExt(SrcVT, SrcReg, DestVT,
+ SrcReg = ARMEmitIntExt(SrcVT, SrcReg, MVT::i32,
/*isZExt*/!isSigned);
if (SrcReg == 0) return false;
}
// 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;
}
unsigned CmpOpc = isThumb2 ? ARM::t2CMPri : ARM::CMPri;
+ CondReg = constrainOperandRegClass(TII.get(CmpOpc), CondReg, 0);
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
.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)
+ if (!UseImm) {
+ Op2Reg = constrainOperandRegClass(TII.get(MovCCOpc), Op2Reg, 1);
+ Op1Reg = constrainOperandRegClass(TII.get(MovCCOpc), Op1Reg, 2);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
.addReg(Op2Reg).addReg(Op1Reg).addImm(ARMCC::NE).addReg(ARM::CPSR);
- else
+ } else {
+ Op1Reg = constrainOperandRegClass(TII.get(MovCCOpc), Op1Reg, 1);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
.addReg(Op1Reg).addImm(Imm).addImm(ARMCC::EQ).addReg(ARM::CPSR);
+ }
UpdateValueMap(I, ResultReg);
return true;
}
unsigned SrcReg2 = getRegForValue(I->getOperand(1));
if (SrcReg2 == 0) return false;
- unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::i32));
+ unsigned ResultReg = createResultReg(&ARM::GPRnopcRegClass);
+ SrcReg1 = constrainOperandRegClass(TII.get(Opc), SrcReg1, 1);
+ SrcReg2 = constrainOperandRegClass(TII.get(Opc), SrcReg2, 2);
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(Opc), ResultReg)
.addReg(SrcReg1).addReg(SrcReg2));
}
bool ARMFastISel::SelectBinaryFPOp(const Instruction *I, unsigned ISDOpcode) {
- EVT VT = TLI.getValueType(I->getType(), true);
+ 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));
!VA.isRegLoc() || !ArgLocs[++i].isRegLoc())
return false;
} else {
- switch (static_cast<EVT>(ArgVT).getSimpleVT().SimpleTy) {
+ switch (ArgVT.SimpleTy) {
default:
return false;
case MVT::i1:
case CCValAssign::ZExt: {
MVT DestVT = VA.getLocVT();
Arg = ARMEmitIntExt(ArgVT, Arg, DestVT, /*isZExt*/true);
- assert (Arg != 0 && "Failed to emit a sext");
+ assert (Arg != 0 && "Failed to emit a zext");
ArgVT = DestVT;
break;
}
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();
+ MVT DestVT = RVLocs[0].getValVT();
const TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT);
unsigned ResultReg = createResultReg(DstRC);
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
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)
if (!FuncInfo.CanLowerReturn)
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;
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)
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::getLibcallReg(const Twine &Name) {
- GlobalValue *GV = new GlobalVariable(Type::getInt32Ty(*Context), false,
+ // Manually compute the global's type to avoid building it when unnecessary.
+ Type *GVTy = Type::getInt32PtrTy(*Context, /*AS=*/0);
+ EVT LCREVT = TLI.getValueType(GVTy);
+ if (!LCREVT.isSimple()) return 0;
+
+ GlobalValue *GV = new GlobalVariable(M, Type::getInt32Ty(*Context), false,
GlobalValue::ExternalLinkage, 0, Name);
- return ARMMaterializeGV(GV, TLI.getValueType(GV->getType()));
+ assert(GV->getType() == GVTy && "We miscomputed the type for the global!");
+ return ARMMaterializeGV(GV, LCREVT.getSimpleVT());
}
// A quick function that will emit a call for a named libcall in F with the
// Can't handle inline asm.
if (isa<InlineAsm>(Callee)) return false;
+ // Allow SelectionDAG isel to handle tail calls.
+ if (CI->isTailCall()) return false;
+
// Check the calling convention.
ImmutableCallSite CS(CI);
CallingConv::ID CC = CS.getCallingConv();
ISD::ArgFlagsTy Flags;
unsigned AttrInd = i - CS.arg_begin() + 1;
- if (CS.paramHasAttr(AttrInd, Attributes::SExt))
+ if (CS.paramHasAttr(AttrInd, Attribute::SExt))
Flags.setSExt();
- if (CS.paramHasAttr(AttrInd, Attributes::ZExt))
+ if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
Flags.setZExt();
// FIXME: Only handle *easy* calls for now.
- if (CS.paramHasAttr(AttrInd, Attributes::InReg) ||
- CS.paramHasAttr(AttrInd, Attributes::StructRet) ||
- CS.paramHasAttr(AttrInd, Attributes::Nest) ||
- CS.paramHasAttr(AttrInd, Attributes::ByVal))
+ if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
+ CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
+ CS.paramHasAttr(AttrInd, Attribute::Nest) ||
+ CS.paramHasAttr(AttrInd, Attribute::ByVal))
return false;
Type *ArgTy = (*i)->getType();
MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
DL, TII.get(CallOpc));
+ unsigned char OpFlags = 0;
+
+ // Add MO_PLT for global address or external symbol in the PIC relocation
+ // model.
+ if (Subtarget->isTargetELF() && TM.getRelocationModel() == Reloc::PIC_)
+ OpFlags = ARMII::MO_PLT;
+
// 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);
+ MIB.addGlobalAddress(GV, 0, OpFlags);
else
- MIB.addExternalSymbol(IntrMemName, 0);
+ MIB.addExternalSymbol(IntrMemName, OpFlags);
// Add implicit physical register uses to the call.
for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
}
bool ARMFastISel::ARMTryEmitSmallMemCpy(Address Dest, Address Src,
- uint64_t Len) {
+ 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;
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;
}
}
return SelectCall(&I, "memset");
}
case Intrinsic::trap: {
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::TRAP));
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(
+ Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP));
return true;
}
}
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 InstructionTable {
+ uint32_t Opc : 16;
+ uint32_t hasS : 1; // Some instructions have an S bit, always set it to 0.
+ uint32_t Shift : 7; // For shift operand addressing mode, used by MOVsi.
+ uint32_t Imm : 8; // All instructions have either a shift or a mask.
+ } IT[2][2][3][2] = {
+ { // Two instructions (first is left shift, second is in this table).
+ { // ARM Opc S Shift Imm
+ /* 1 bit sext */ { { ARM::MOVsi , 1, ARM_AM::asr , 31 },
+ /* 1 bit zext */ { ARM::MOVsi , 1, ARM_AM::lsr , 31 } },
+ /* 8 bit sext */ { { ARM::MOVsi , 1, ARM_AM::asr , 24 },
+ /* 8 bit zext */ { ARM::MOVsi , 1, ARM_AM::lsr , 24 } },
+ /* 16 bit sext */ { { ARM::MOVsi , 1, ARM_AM::asr , 16 },
+ /* 16 bit zext */ { ARM::MOVsi , 1, ARM_AM::lsr , 16 } }
+ },
+ { // Thumb Opc S Shift Imm
+ /* 1 bit sext */ { { ARM::tASRri , 0, ARM_AM::no_shift, 31 },
+ /* 1 bit zext */ { ARM::tLSRri , 0, ARM_AM::no_shift, 31 } },
+ /* 8 bit sext */ { { ARM::tASRri , 0, ARM_AM::no_shift, 24 },
+ /* 8 bit zext */ { ARM::tLSRri , 0, ARM_AM::no_shift, 24 } },
+ /* 16 bit sext */ { { ARM::tASRri , 0, ARM_AM::no_shift, 16 },
+ /* 16 bit zext */ { ARM::tLSRri , 0, ARM_AM::no_shift, 16 } }
+ }
+ },
+ { // Single instruction.
+ { // ARM Opc S Shift Imm
+ /* 1 bit sext */ { { ARM::KILL , 0, ARM_AM::no_shift, 0 },
+ /* 1 bit zext */ { ARM::ANDri , 1, ARM_AM::no_shift, 1 } },
+ /* 8 bit sext */ { { ARM::SXTB , 0, ARM_AM::no_shift, 0 },
+ /* 8 bit zext */ { ARM::ANDri , 1, ARM_AM::no_shift, 255 } },
+ /* 16 bit sext */ { { ARM::SXTH , 0, ARM_AM::no_shift, 0 },
+ /* 16 bit zext */ { ARM::UXTH , 0, ARM_AM::no_shift, 0 } }
+ },
+ { // Thumb Opc S Shift Imm
+ /* 1 bit sext */ { { ARM::KILL , 0, ARM_AM::no_shift, 0 },
+ /* 1 bit zext */ { ARM::t2ANDri, 1, ARM_AM::no_shift, 1 } },
+ /* 8 bit sext */ { { ARM::t2SXTB , 0, ARM_AM::no_shift, 0 },
+ /* 8 bit zext */ { ARM::t2ANDri, 1, ARM_AM::no_shift, 255 } },
+ /* 16 bit sext */ { { ARM::t2SXTH , 0, ARM_AM::no_shift, 0 },
+ /* 16 bit zext */ { ARM::t2UXTH , 0, ARM_AM::no_shift, 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 = SrcBits / 8; // {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];
+ const InstructionTable *ITP = &IT[isSingleInstr][isThumb2][Bitness][isZExt];
+ unsigned Opc = ITP->Opc;
+ assert(ARM::KILL != Opc && "Invalid table entry");
+ unsigned hasS = ITP->hasS;
+ ARM_AM::ShiftOpc Shift = (ARM_AM::ShiftOpc) ITP->Shift;
+ assert(((Shift == ARM_AM::no_shift) == (Opc != ARM::MOVsi)) &&
+ "only MOVsi has shift operand addressing mode");
+ unsigned Imm = ITP->Imm;
+
+ // 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::MOVsi;
+ unsigned ResultReg;
+ // MOVsi encodes shift and immediate in shift operand addressing mode.
+ // The following condition has the same value when emitting two
+ // instruction sequences: both are shifts.
+ bool ImmIsSO = (Shift != ARM_AM::no_shift);
+
+ // 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);
+ bool isLsl = (0 == Instr) && !isSingleInstr;
+ unsigned Opcode = isLsl ? LSLOpc : Opc;
+ ARM_AM::ShiftOpc ShiftAM = isLsl ? ARM_AM::lsl : Shift;
+ unsigned ImmEnc = ImmIsSO ? ARM_AM::getSORegOpc(ShiftAM, Imm) : Imm;
+ bool isKill = 1 == Instr;
+ MachineInstrBuilder MIB = BuildMI(
+ *FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opcode), ResultReg);
+ if (setsCPSR)
+ MIB.addReg(ARM::CPSR, RegState::Define);
+ SrcReg = constrainOperandRegClass(TII.get(Opcode), SrcReg, 1 + setsCPSR);
+ AddDefaultPred(MIB.addReg(SrcReg, isKill * RegState::Kill).addImm(ImmEnc));
+ 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);
if (Reg2 == 0) return false;
}
- unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::i32));
+ unsigned ResultReg = createResultReg(&ARM::GPRnopcRegClass);
if(ResultReg == 0) return false;
MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
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 ARMFastISel::ARMLowerPICELF(const GlobalValue *GV,
- unsigned Align, EVT VT) {
+ unsigned Align, MVT VT) {
bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
ARMConstantPoolConstant *CPV =
ARMConstantPoolConstant::Create(GV, UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
unsigned DestReg1 = createResultReg(TLI.getRegClassFor(VT));
// Load value.
if (isThumb2) {
+ DestReg1 = constrainOperandRegClass(TII.get(ARM::t2LDRpci), DestReg1, 0);
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.
+ DestReg1 = constrainOperandRegClass(TII.get(ARM::LDRcp), DestReg1, 0);
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
DL, TII.get(ARM::LDRcp), DestReg1)
.addConstantPoolIndex(Idx).addImm(0));
}
unsigned DestReg2 = createResultReg(TLI.getRegClassFor(VT));
+ DestReg2 = constrainOperandRegClass(TII.get(Opc), DestReg2, 0);
+ DestReg1 = constrainOperandRegClass(TII.get(Opc), DestReg1, 1);
+ GlobalBaseReg = constrainOperandRegClass(TII.get(Opc), GlobalBaseReg, 2);
MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
DL, TII.get(Opc), DestReg2)
.addReg(DestReg1)
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 = &ARM::rGPRRegClass;
+ 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 {
FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo,
const TargetLibraryInfo *libInfo) {
- // Completely untested on non-iOS.
const TargetMachine &TM = funcInfo.MF->getTarget();
- // Darwin and thumb1 only for now.
const ARMSubtarget *Subtarget = &TM.getSubtarget<ARMSubtarget>();
- if (Subtarget->isTargetIOS() && !Subtarget->isThumb1Only())
+ // Thumb2 support on iOS; ARM support on iOS, Linux and NaCl.
+ bool UseFastISel = false;
+ UseFastISel |= Subtarget->isTargetMachO() && !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;
}
}