#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineMemOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
-#include "llvm/CodeGen/PseudoSourceValue.h"
#include "llvm/Support/CallSite.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ErrorHandling.h"
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);
#include "ARMGenFastISel.inc"
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 SelectCall(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 isLoadTypeLegal(Type *Ty, MVT &VT);
bool ARMEmitCmp(const Value *Src1Value, const Value *Src2Value,
bool isZExt);
- bool ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr);
- bool ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr);
+ bool ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr,
+ unsigned Alignment = 0, bool isZExt = true,
+ bool allocReg = true);
+
+ bool ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr,
+ unsigned Alignment = 0);
bool ARMComputeAddress(const Value *Obj, Address &Addr);
- void ARMSimplifyAddress(Address &Addr, EVT VT);
+ void ARMSimplifyAddress(Address &Addr, EVT 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);
// Call handling routines.
private:
- bool FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT,
- unsigned &ResultReg);
CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool Return);
bool ProcessCallArgs(SmallVectorImpl<Value*> &Args,
SmallVectorImpl<unsigned> &ArgRegs,
const MachineInstrBuilder &AddOptionalDefs(const MachineInstrBuilder &MIB);
void AddLoadStoreOperands(EVT VT, Address &Addr,
const MachineInstrBuilder &MIB,
- unsigned Flags);
+ unsigned Flags, bool useAM3);
};
} // end anonymous namespace
// we don't care about implicit defs here, just places we'll need to add a
// default CCReg argument. Sets CPSR if we're setting CPSR instead of CCR.
bool ARMFastISel::DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR) {
- const MCInstrDesc &MCID = MI->getDesc();
- if (!MCID.hasOptionalDef())
+ if (!MI->hasOptionalDef())
return false;
// Look to see if our OptionalDef is defining CPSR or CCR.
// do so now.
const ConstantInt *CI = cast<ConstantInt>(C);
if (Subtarget->hasV6T2Ops() && isUInt<16>(CI->getZExtValue())) {
- EVT SrcVT = MVT::i32;
unsigned Opc = isThumb2 ? ARM::t2MOVi16 : ARM::MOVi16;
- unsigned ImmReg = createResultReg(TLI.getRegClassFor(SrcVT));
+ unsigned ImmReg = createResultReg(TLI.getRegClassFor(MVT::i32));
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(Opc), ImmReg)
.addImm(CI->getZExtValue()));
return ImmReg;
}
- // For now 32-bit only.
+ // Use MVN to emit negative constants.
+ if (VT == MVT::i32 && Subtarget->hasV6T2Ops() && CI->isNegative()) {
+ unsigned Imm = (unsigned)~(CI->getSExtValue());
+ bool UseImm = isThumb2 ? (ARM_AM::getT2SOImmVal(Imm) != -1) :
+ (ARM_AM::getSOImmVal(Imm) != -1);
+ if (UseImm) {
+ unsigned Opc = isThumb2 ? ARM::t2MVNi : ARM::MVNi;
+ unsigned ImmReg = createResultReg(TLI.getRegClassFor(MVT::i32));
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(Opc), ImmReg)
+ .addImm(Imm));
+ return ImmReg;
+ }
+ }
+
+ // Load from constant pool. For now 32-bit only.
if (VT != MVT::i32)
return false;
// TODO: Need more magic for ARM PIC.
if (!isThumb2 && (RelocM == Reloc::PIC_)) return 0;
- // MachineConstantPool wants an explicit alignment.
- unsigned Align = TD.getPrefTypeAlignment(GV->getType());
- if (Align == 0) {
- // TODO: Figure out if this is correct.
- Align = TD.getTypeAllocSize(GV->getType());
- }
-
- // Grab index.
- unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb() ? 4 : 8);
- unsigned Id = AFI->createPICLabelUId();
- ARMConstantPoolValue *CPV = ARMConstantPoolConstant::Create(GV, Id,
- ARMCP::CPValue,
- PCAdj);
- unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
-
- // Load value.
- MachineInstrBuilder MIB;
unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
- if (isThumb2) {
- unsigned Opc = (RelocM != Reloc::PIC_) ? ARM::t2LDRpci : ARM::t2LDRpci_pic;
- MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
- .addConstantPoolIndex(Idx);
- if (RelocM == Reloc::PIC_)
- MIB.addImm(Id);
+
+ // 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.
+ if (Subtarget->useMovt() &&
+ Subtarget->isTargetDarwin() == (RelocM != Reloc::Static)) {
+ unsigned Opc;
+ 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));
} else {
- // The extra immediate is for addrmode2.
- MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp),
- DestReg)
- .addConstantPoolIndex(Idx)
- .addImm(0);
+ // MachineConstantPool wants an explicit alignment.
+ unsigned Align = TD.getPrefTypeAlignment(GV->getType());
+ if (Align == 0) {
+ // TODO: Figure out if this is correct.
+ Align = TD.getTypeAllocSize(GV->getType());
+ }
+
+ // Grab index.
+ unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 :
+ (Subtarget->isThumb() ? 4 : 8);
+ unsigned Id = AFI->createPICLabelUId();
+ ARMConstantPoolValue *CPV = ARMConstantPoolConstant::Create(GV, Id,
+ ARMCP::CPValue,
+ PCAdj);
+ unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);
+
+ // Load value.
+ MachineInstrBuilder MIB;
+ if (isThumb2) {
+ unsigned Opc = (RelocM!=Reloc::PIC_) ? ARM::t2LDRpci : ARM::t2LDRpci_pic;
+ MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
+ .addConstantPoolIndex(Idx);
+ if (RelocM == Reloc::PIC_)
+ MIB.addImm(Id);
+ } 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);
}
- AddOptionalDefs(MIB);
if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) {
+ MachineInstrBuilder MIB;
unsigned NewDestReg = createResultReg(TLI.getRegClassFor(VT));
if (isThumb2)
MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
return 0;
}
+// TODO: unsigned ARMFastISel::TargetMaterializeFloatZero(const ConstantFP *CF);
+
unsigned ARMFastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
// Don't handle dynamic allocas.
if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;
TargetRegisterClass* RC = TLI.getRegClassFor(VT);
unsigned ResultReg = createResultReg(RC);
unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri;
- AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(Opc), ResultReg)
.addFrameIndex(SI->second)
.addImm(0));
// If this is a type than can be sign or zero-extended to a basic operation
// go ahead and accept it now.
- if (VT == MVT::i8 || VT == MVT::i16)
+ if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
return true;
return false;
}
}
- // Materialize the global variable's address into a reg which can
- // then be used later to load the variable.
- if (const GlobalValue *GV = dyn_cast<GlobalValue>(Obj)) {
- unsigned Tmp = ARMMaterializeGV(GV, TLI.getValueType(Obj->getType()));
- if (Tmp == 0) return false;
-
- Addr.Base.Reg = Tmp;
- return true;
- }
-
// Try to get this in a register if nothing else has worked.
if (Addr.Base.Reg == 0) Addr.Base.Reg = getRegForValue(Obj);
return Addr.Base.Reg != 0;
}
-void ARMFastISel::ARMSimplifyAddress(Address &Addr, EVT VT) {
+void ARMFastISel::ARMSimplifyAddress(Address &Addr, EVT VT, bool useAM3) {
assert(VT.isSimple() && "Non-simple types are invalid here!");
bool needsLowering = false;
switch (VT.getSimpleVT().SimpleTy) {
- default:
- assert(false && "Unhandled load/store type!");
- case MVT::i16:
- if (isThumb2)
- // Integer loads/stores handle 12-bit offsets.
- needsLowering = ((Addr.Offset & 0xfff) != Addr.Offset);
- else
- // ARM i16 integer loads/stores handle +/-imm8 offsets.
- if (Addr.Offset > 255 || Addr.Offset < -255)
- needsLowering = true;
- break;
+ default: llvm_unreachable("Unhandled load/store type!");
case MVT::i1:
case MVT::i8:
+ case MVT::i16:
case MVT::i32:
- // Integer loads/stores handle 12-bit offsets.
- needsLowering = ((Addr.Offset & 0xfff) != Addr.Offset);
+ if (!useAM3) {
+ // Integer loads/stores handle 12-bit offsets.
+ needsLowering = ((Addr.Offset & 0xfff) != Addr.Offset);
+ // Handle negative offsets.
+ if (needsLowering && isThumb2)
+ needsLowering = !(Subtarget->hasV6T2Ops() && Addr.Offset < 0 &&
+ Addr.Offset > -256);
+ } else {
+ // ARM halfword load/stores and signed byte loads use +/-imm8 offsets.
+ needsLowering = (Addr.Offset > 255 || Addr.Offset < -255);
+ }
break;
case MVT::f32:
case MVT::f64:
ARM::GPRRegisterClass;
unsigned ResultReg = createResultReg(RC);
unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri;
- AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(Opc), ResultReg)
.addFrameIndex(Addr.Base.FI)
.addImm(0));
void ARMFastISel::AddLoadStoreOperands(EVT VT, Address &Addr,
const MachineInstrBuilder &MIB,
- unsigned Flags) {
+ 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 ||
// Now add the rest of the operands.
MIB.addFrameIndex(FI);
- // ARM halfword load/stores need an additional operand.
- if (!isThumb2 && VT.getSimpleVT().SimpleTy == MVT::i16) MIB.addReg(0);
-
- MIB.addImm(Addr.Offset);
+ // ARM halfword load/stores and signed byte loads need an additional
+ // operand.
+ if (useAM3) {
+ signed Imm = (Addr.Offset < 0) ? (0x100 | -Addr.Offset) : Addr.Offset;
+ MIB.addReg(0);
+ MIB.addImm(Imm);
+ } else {
+ MIB.addImm(Addr.Offset);
+ }
MIB.addMemOperand(MMO);
} else {
// Now add the rest of the operands.
MIB.addReg(Addr.Base.Reg);
- // ARM halfword load/stores need an additional operand.
- if (!isThumb2 && VT.getSimpleVT().SimpleTy == MVT::i16) MIB.addReg(0);
-
- MIB.addImm(Addr.Offset);
+ // ARM halfword load/stores and signed byte loads need an additional
+ // operand.
+ if (useAM3) {
+ signed Imm = (Addr.Offset < 0) ? (0x100 | -Addr.Offset) : Addr.Offset;
+ MIB.addReg(0);
+ MIB.addImm(Imm);
+ } else {
+ MIB.addImm(Addr.Offset);
+ }
}
AddOptionalDefs(MIB);
}
-bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr) {
-
+bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr,
+ unsigned Alignment, bool isZExt, bool allocReg) {
assert(VT.isSimple() && "Non-simple types are invalid here!");
unsigned Opc;
- TargetRegisterClass *RC;
+ bool useAM3 = false;
+ bool needVMOV = false;
+ TargetRegisterClass *RC;
switch (VT.getSimpleVT().SimpleTy) {
// This is mostly going to be Neon/vector support.
default: return false;
+ case MVT::i1:
case MVT::i8:
- Opc = isThumb2 ? ARM::t2LDRBi12 : ARM::LDRBi12;
+ if (isThumb2) {
+ if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
+ Opc = isZExt ? ARM::t2LDRBi8 : ARM::t2LDRSBi8;
+ else
+ Opc = isZExt ? ARM::t2LDRBi12 : ARM::t2LDRSBi12;
+ } else {
+ if (isZExt) {
+ Opc = ARM::LDRBi12;
+ } else {
+ Opc = ARM::LDRSB;
+ useAM3 = true;
+ }
+ }
RC = ARM::GPRRegisterClass;
break;
case MVT::i16:
- Opc = isThumb2 ? ARM::t2LDRHi12 : ARM::LDRH;
+ if (isThumb2) {
+ if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
+ Opc = isZExt ? ARM::t2LDRHi8 : ARM::t2LDRSHi8;
+ else
+ Opc = isZExt ? ARM::t2LDRHi12 : ARM::t2LDRSHi12;
+ } else {
+ Opc = isZExt ? ARM::LDRH : ARM::LDRSH;
+ useAM3 = true;
+ }
RC = ARM::GPRRegisterClass;
break;
case MVT::i32:
- Opc = isThumb2 ? ARM::t2LDRi12 : ARM::LDRi12;
+ if (isThumb2) {
+ if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
+ Opc = ARM::t2LDRi8;
+ else
+ Opc = ARM::t2LDRi12;
+ } else {
+ Opc = ARM::LDRi12;
+ }
RC = ARM::GPRRegisterClass;
break;
case MVT::f32:
- Opc = ARM::VLDRS;
- RC = TLI.getRegClassFor(VT);
+ if (!Subtarget->hasVFP2()) return false;
+ // Unaligned loads need special handling. Floats require word-alignment.
+ if (Alignment && Alignment < 4) {
+ needVMOV = true;
+ VT = MVT::i32;
+ Opc = isThumb2 ? ARM::t2LDRi12 : ARM::LDRi12;
+ RC = ARM::GPRRegisterClass;
+ } else {
+ Opc = ARM::VLDRS;
+ RC = TLI.getRegClassFor(VT);
+ }
break;
case MVT::f64:
+ if (!Subtarget->hasVFP2()) return false;
+ // FIXME: Unaligned loads need special handling. Doublewords require
+ // word-alignment.
+ if (Alignment && Alignment < 4)
+ return false;
+
Opc = ARM::VLDRD;
RC = TLI.getRegClassFor(VT);
break;
}
// Simplify this down to something we can handle.
- ARMSimplifyAddress(Addr, VT);
+ ARMSimplifyAddress(Addr, VT, useAM3);
// Create the base instruction, then add the operands.
- ResultReg = createResultReg(RC);
+ if (allocReg)
+ ResultReg = createResultReg(RC);
+ assert (ResultReg > 255 && "Expected an allocated virtual register.");
MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(Opc), ResultReg);
- AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOLoad);
+ AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOLoad, useAM3);
+
+ // If we had an unaligned load of a float we've converted it to an regular
+ // load. Now we must move from the GRP to the FP register.
+ if (needVMOV) {
+ unsigned MoveReg = createResultReg(TLI.getRegClassFor(MVT::f32));
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(ARM::VMOVSR), MoveReg)
+ .addReg(ResultReg));
+ ResultReg = MoveReg;
+ }
return true;
}
if (!ARMComputeAddress(I->getOperand(0), Addr)) return false;
unsigned ResultReg;
- if (!ARMEmitLoad(VT, ResultReg, Addr)) return false;
+ if (!ARMEmitLoad(VT, ResultReg, Addr, cast<LoadInst>(I)->getAlignment()))
+ return false;
UpdateValueMap(I, ResultReg);
return true;
}
-bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr) {
+bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr,
+ unsigned Alignment) {
unsigned StrOpc;
+ bool useAM3 = false;
switch (VT.getSimpleVT().SimpleTy) {
// This is mostly going to be Neon/vector support.
default: return false;
SrcReg = Res;
} // Fallthrough here.
case MVT::i8:
- StrOpc = isThumb2 ? ARM::t2STRBi12 : ARM::STRBi12;
+ if (isThumb2) {
+ if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
+ StrOpc = ARM::t2STRBi8;
+ else
+ StrOpc = ARM::t2STRBi12;
+ } else {
+ StrOpc = ARM::STRBi12;
+ }
break;
case MVT::i16:
- StrOpc = isThumb2 ? ARM::t2STRHi12 : ARM::STRH;
+ if (isThumb2) {
+ if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
+ StrOpc = ARM::t2STRHi8;
+ else
+ StrOpc = ARM::t2STRHi12;
+ } else {
+ StrOpc = ARM::STRH;
+ useAM3 = true;
+ }
break;
case MVT::i32:
- StrOpc = isThumb2 ? ARM::t2STRi12 : ARM::STRi12;
+ if (isThumb2) {
+ if (Addr.Offset < 0 && Addr.Offset > -256 && Subtarget->hasV6T2Ops())
+ StrOpc = ARM::t2STRi8;
+ else
+ StrOpc = ARM::t2STRi12;
+ } else {
+ StrOpc = ARM::STRi12;
+ }
break;
case MVT::f32:
if (!Subtarget->hasVFP2()) return false;
- StrOpc = ARM::VSTRS;
+ // Unaligned stores need special handling. Floats require word-alignment.
+ if (Alignment && Alignment < 4) {
+ unsigned MoveReg = createResultReg(TLI.getRegClassFor(MVT::i32));
+ AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(ARM::VMOVRS), MoveReg)
+ .addReg(SrcReg));
+ SrcReg = MoveReg;
+ VT = MVT::i32;
+ StrOpc = isThumb2 ? ARM::t2STRi12 : ARM::STRi12;
+ } else {
+ StrOpc = ARM::VSTRS;
+ }
break;
case MVT::f64:
if (!Subtarget->hasVFP2()) return false;
+ // FIXME: Unaligned stores need special handling. Doublewords require
+ // word-alignment.
+ if (Alignment && Alignment < 4)
+ return false;
+
StrOpc = ARM::VSTRD;
break;
}
// Simplify this down to something we can handle.
- ARMSimplifyAddress(Addr, VT);
+ ARMSimplifyAddress(Addr, VT, useAM3);
// Create the base instruction, then add the operands.
MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(StrOpc))
- .addReg(SrcReg, getKillRegState(true));
- AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOStore);
+ .addReg(SrcReg);
+ AddLoadStoreOperands(VT, Addr, MIB, MachineMemOperand::MOStore, useAM3);
return true;
}
if (!ARMComputeAddress(I->getOperand(1), Addr))
return false;
- if (!ARMEmitStore(VT, SrcReg, Addr)) return false;
+ if (!ARMEmitStore(VT, SrcReg, Addr, cast<StoreInst>(I)->getAlignment()))
+ return false;
return true;
}
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));
+ return true;
+}
+
bool ARMFastISel::ARMEmitCmp(const Value *Src1Value, const Value *Src2Value,
bool isZExt) {
Type *Ty = Src1Value->getType();
// Check to see if the 2nd operand is a constant that we can encode directly
// in the compare.
- uint64_t Imm;
- int EncodedImm = 0;
- bool EncodeImm = false;
+ int Imm = 0;
+ bool UseImm = false;
bool isNegativeImm = false;
+ // FIXME: At -O0 we don't have anything that canonicalizes operand order.
+ // Thus, Src1Value may be a ConstantInt, but we're missing it.
if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(Src2Value)) {
if (SrcVT == MVT::i32 || SrcVT == MVT::i16 || SrcVT == MVT::i8 ||
SrcVT == MVT::i1) {
const APInt &CIVal = ConstInt->getValue();
-
- isNegativeImm = CIVal.isNegative();
- Imm = (isNegativeImm) ? (-CIVal).getZExtValue() : CIVal.getZExtValue();
- EncodedImm = (int)Imm;
- EncodeImm = isThumb2 ? (ARM_AM::getT2SOImmVal(EncodedImm) != -1) :
- (ARM_AM::getSOImmVal(EncodedImm) != -1);
+ Imm = (isZExt) ? (int)CIVal.getZExtValue() : (int)CIVal.getSExtValue();
+ if (Imm < 0) {
+ isNegativeImm = true;
+ Imm = -Imm;
+ }
+ UseImm = isThumb2 ? (ARM_AM::getT2SOImmVal(Imm) != -1) :
+ (ARM_AM::getSOImmVal(Imm) != -1);
}
} else if (const ConstantFP *ConstFP = dyn_cast<ConstantFP>(Src2Value)) {
if (SrcVT == MVT::f32 || SrcVT == MVT::f64)
if (ConstFP->isZero() && !ConstFP->isNegative())
- EncodeImm = true;
+ UseImm = true;
}
unsigned CmpOpc;
// TODO: Verify compares.
case MVT::f32:
isICmp = false;
- CmpOpc = EncodeImm ? ARM::VCMPEZS : ARM::VCMPES;
+ CmpOpc = UseImm ? ARM::VCMPEZS : ARM::VCMPES;
break;
case MVT::f64:
isICmp = false;
- CmpOpc = EncodeImm ? ARM::VCMPEZD : ARM::VCMPED;
+ CmpOpc = UseImm ? ARM::VCMPEZD : ARM::VCMPED;
break;
case MVT::i1:
case MVT::i8:
// Intentional fall-through.
case MVT::i32:
if (isThumb2) {
- if (!EncodeImm)
+ if (!UseImm)
CmpOpc = ARM::t2CMPrr;
else
CmpOpc = isNegativeImm ? ARM::t2CMNzri : ARM::t2CMPri;
} else {
- if (!EncodeImm)
+ if (!UseImm)
CmpOpc = ARM::CMPrr;
else
CmpOpc = isNegativeImm ? ARM::CMNzri : ARM::CMPri;
unsigned SrcReg1 = getRegForValue(Src1Value);
if (SrcReg1 == 0) return false;
- unsigned SrcReg2;
- if (!EncodeImm) {
+ unsigned SrcReg2 = 0;
+ if (!UseImm) {
SrcReg2 = getRegForValue(Src2Value);
if (SrcReg2 == 0) return false;
}
ResultReg = ARMEmitIntExt(SrcVT, SrcReg1, MVT::i32, isZExt);
if (ResultReg == 0) return false;
SrcReg1 = ResultReg;
- if (!EncodeImm) {
+ if (!UseImm) {
ResultReg = ARMEmitIntExt(SrcVT, SrcReg2, MVT::i32, isZExt);
if (ResultReg == 0) return false;
SrcReg2 = ResultReg;
}
}
- if (!EncodeImm) {
+ if (!UseImm) {
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(CmpOpc))
.addReg(SrcReg1).addReg(SrcReg2));
// Only add immediate for icmp as the immediate for fcmp is an implicit 0.0.
if (isICmp)
- MIB.addImm(EncodedImm);
+ MIB.addImm(Imm);
AddOptionalDefs(MIB);
}
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;
// Handle sign-extension.
if (SrcVT == MVT::i16 || SrcVT == MVT::i8) {
EVT DestVT = MVT::i32;
- unsigned ResultReg = ARMEmitIntExt(SrcVT, SrcReg, DestVT, /*isZExt*/ false);
+ unsigned ResultReg = ARMEmitIntExt(SrcVT, SrcReg, DestVT,
+ /*isZExt*/!isSigned);
if (ResultReg == 0) return false;
SrcReg = ResultReg;
}
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)
if (CondReg == 0) return false;
unsigned Op1Reg = getRegForValue(I->getOperand(1));
if (Op1Reg == 0) return false;
- unsigned Op2Reg = getRegForValue(I->getOperand(2));
- if (Op2Reg == 0) return false;
- unsigned CmpOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri;
+ // Check to see if we can use an immediate in the conditional move.
+ int Imm = 0;
+ bool UseImm = false;
+ bool isNegativeImm = false;
+ if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(I->getOperand(2))) {
+ assert (VT == MVT::i32 && "Expecting an i32.");
+ Imm = (int)ConstInt->getValue().getZExtValue();
+ if (Imm < 0) {
+ isNegativeImm = true;
+ Imm = ~Imm;
+ }
+ UseImm = isThumb2 ? (ARM_AM::getT2SOImmVal(Imm) != -1) :
+ (ARM_AM::getSOImmVal(Imm) != -1);
+ }
+
+ unsigned Op2Reg = 0;
+ if (!UseImm) {
+ Op2Reg = getRegForValue(I->getOperand(2));
+ if (Op2Reg == 0) return false;
+ }
+
+ unsigned CmpOpc = isThumb2 ? ARM::t2CMPri : ARM::CMPri;
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
- .addReg(CondReg).addImm(1));
+ .addReg(CondReg).addImm(0));
+
+ unsigned MovCCOpc;
+ if (!UseImm) {
+ MovCCOpc = isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr;
+ } else {
+ if (!isNegativeImm) {
+ MovCCOpc = isThumb2 ? ARM::t2MOVCCi : ARM::MOVCCi;
+ } else {
+ MovCCOpc = isThumb2 ? ARM::t2MVNCCi : ARM::MVNCCi;
+ }
+ }
unsigned ResultReg = createResultReg(RC);
- unsigned MovCCOpc = isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr;
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
- .addReg(Op1Reg).addReg(Op2Reg)
- .addImm(ARMCC::EQ).addReg(ARM::CPSR);
+ if (!UseImm)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
+ .addReg(Op2Reg).addReg(Op1Reg).addImm(ARMCC::NE).addReg(ARM::CPSR);
+ else
+ 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;
}
-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) {
+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(TLI.getRegClassFor(MVT::i32));
+ 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 VT = TLI.getValueType(I->getType(), true);
// We can get here in the case when we want to use NEON for our fp
if (isFloat && !Subtarget->hasVFP2())
return false;
- unsigned Op1 = getRegForValue(I->getOperand(0));
- if (Op1 == 0) return false;
-
- unsigned Op2 = getRegForValue(I->getOperand(1));
- if (Op2 == 0) return false;
-
unsigned Opc;
bool is64bit = VT == MVT::f64 || VT == MVT::i64;
switch (ISDOpcode) {
Opc = is64bit ? ARM::VMULD : ARM::VMULS;
break;
}
+ unsigned Op1 = getRegForValue(I->getOperand(0));
+ if (Op1 == 0) return false;
+
+ unsigned Op2 = getRegForValue(I->getOperand(1));
+ if (Op2 == 0) return false;
+
unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(Opc), ResultReg)
// Call Handling Code
-bool ARMFastISel::FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src,
- EVT SrcVT, unsigned &ResultReg) {
- unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
- Src, /*TODO: Kill=*/false);
-
- if (RR != 0) {
- ResultReg = RR;
- return true;
- } else
- return false;
-}
-
// This is largely taken directly from CCAssignFnForNode - we don't support
// varargs in FastISel so that part has been removed.
// TODO: We may not support all of this.
// Use target triple & subtarget features to do actual dispatch.
if (Subtarget->isAAPCS_ABI()) {
if (Subtarget->hasVFP2() &&
- FloatABIType == FloatABI::Hard)
+ TM.Options.FloatABIType == FloatABI::Hard)
return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
else
return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
switch (VA.getLocInfo()) {
case CCValAssign::Full: break;
case CCValAssign::SExt: {
- EVT DestVT = VA.getLocVT();
- unsigned ResultReg = ARMEmitIntExt(ArgVT, Arg, DestVT,
- /*isZExt*/false);
- assert (ResultReg != 0 && "Failed to emit a sext");
- Arg = ResultReg;
+ MVT DestVT = VA.getLocVT();
+ Arg = ARMEmitIntExt(ArgVT, Arg, DestVT, /*isZExt*/false);
+ assert (Arg != 0 && "Failed to emit a sext");
+ ArgVT = DestVT;
break;
}
case CCValAssign::AExt:
// Intentional fall-through. Handle AExt and ZExt.
case CCValAssign::ZExt: {
- EVT DestVT = VA.getLocVT();
- unsigned ResultReg = ARMEmitIntExt(ArgVT, Arg, DestVT,
- /*isZExt*/true);
- assert (ResultReg != 0 && "Failed to emit a sext");
- Arg = ResultReg;
+ MVT DestVT = VA.getLocVT();
+ Arg = ARMEmitIntExt(ArgVT, Arg, DestVT, /*isZExt*/true);
+ assert (Arg != 0 && "Failed to emit a sext");
+ ArgVT = DestVT;
break;
}
case CCValAssign::BCvt: {
unsigned ARMFastISel::ARMSelectCallOp(const GlobalValue *GV) {
- // Darwin needs the r9 versions of the opcodes.
- bool isDarwin = Subtarget->isTargetDarwin();
+ // iOS needs the r9 versions of the opcodes.
+ bool isiOS = Subtarget->isTargetIOS();
if (isThumb2) {
- return isDarwin ? ARM::tBLr9 : ARM::tBL;
+ return isiOS ? ARM::tBLr9 : ARM::tBL;
} else {
- return isDarwin ? ARM::BLr9 : ARM::BL;
+ return isiOS ? ARM::BLr9 : ARM::BL;
}
}
if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
return false;
- // Issue the call, BLr9 for darwin, BL otherwise.
+ // Issue the call, BLr9 for iOS, BL otherwise.
// TODO: Turn this into the table of arm call ops.
MachineInstrBuilder MIB;
unsigned CallOpc = ARMSelectCallOp(NULL);
return true;
}
-bool ARMFastISel::SelectCall(const Instruction *I) {
+bool ARMFastISel::SelectCall(const Instruction *I,
+ const char *IntrMemName = 0) {
const CallInst *CI = cast<CallInst>(I);
const Value *Callee = CI->getCalledValue();
- // Can't handle inline asm or worry about intrinsics yet.
- if (isa<InlineAsm>(Callee) || isa<IntrinsicInst>(CI)) return false;
+ // Can't handle inline asm.
+ if (isa<InlineAsm>(Callee)) return false;
// Only handle global variable Callees.
const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
ArgFlags.reserve(CS.arg_size());
for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
i != e; ++i) {
- unsigned Arg = getRegForValue(*i);
+ // If we're lowering a memory intrinsic instead of a regular call, skip the
+ // last two arguments, which shouldn't be passed to the underlying function.
+ if (IntrMemName && e-i <= 2)
+ break;
- if (Arg == 0)
- return false;
ISD::ArgFlagsTy Flags;
unsigned AttrInd = i - CS.arg_begin() + 1;
if (CS.paramHasAttr(AttrInd, Attribute::SExt))
if (!isTypeLegal(ArgTy, ArgVT) && ArgVT != MVT::i16 && ArgVT != MVT::i8 &&
ArgVT != MVT::i1)
return false;
+
+ unsigned Arg = getRegForValue(*i);
+ if (Arg == 0)
+ return false;
+
unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
Flags.setOrigAlign(OriginalAlignment);
if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
return false;
- // Issue the call, BLr9 for darwin, BL otherwise.
+ // 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)
+ if(isThumb2) {
// Explicitly adding the predicate here.
MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(CallOpc)))
- .addGlobalAddress(GV, 0, 0);
- else
- // Explicitly adding the predicate here.
- MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(CallOpc))
- .addGlobalAddress(GV, 0, 0));
-
+ 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));
+ else
+ MIB = AddDefaultPred(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(CallOpc))
+ .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]);
return true;
}
+bool ARMFastISel::ARMIsMemCpySmall(uint64_t Len) {
+ return Len <= 16;
+}
+
+bool ARMFastISel::ARMTryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len) {
+ // 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;
+ }
+
+ bool RV;
+ unsigned ResultReg;
+ RV = ARMEmitLoad(VT, ResultReg, Src);
+ 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;
+ Dest.Offset += Size;
+ Src.Offset += Size;
+ }
+
+ return true;
+}
+
+bool ARMFastISel::SelectIntrinsicCall(const IntrinsicInst &I) {
+ // FIXME: Handle more intrinsics.
+ switch (I.getIntrinsicID()) {
+ default: return false;
+ case Intrinsic::memcpy:
+ case Intrinsic::memmove: {
+ const MemTransferInst &MTI = cast<MemTransferInst>(I);
+ // Don't handle volatile.
+ if (MTI.isVolatile())
+ return false;
+
+ // Disable inlining for memmove before calls to ComputeAddress. Otherwise,
+ // we would emit dead code because we don't currently handle memmoves.
+ bool isMemCpy = (I.getIntrinsicID() == Intrinsic::memcpy);
+ if (isa<ConstantInt>(MTI.getLength()) && isMemCpy) {
+ // Small memcpy's are common enough that we want to do them without a call
+ // if possible.
+ uint64_t Len = cast<ConstantInt>(MTI.getLength())->getZExtValue();
+ if (ARMIsMemCpySmall(Len)) {
+ Address Dest, Src;
+ if (!ARMComputeAddress(MTI.getRawDest(), Dest) ||
+ !ARMComputeAddress(MTI.getRawSource(), Src))
+ return false;
+ if (ARMTryEmitSmallMemCpy(Dest, Src, Len))
+ return true;
+ }
+ }
+
+ if (!MTI.getLength()->getType()->isIntegerTy(32))
+ return false;
+
+ if (MTI.getSourceAddressSpace() > 255 || MTI.getDestAddressSpace() > 255)
+ return false;
+
+ const char *IntrMemName = isa<MemCpyInst>(I) ? "memcpy" : "memmove";
+ return SelectCall(&I, IntrMemName);
+ }
+ case Intrinsic::memset: {
+ const MemSetInst &MSI = cast<MemSetInst>(I);
+ // 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");
+ }
+ }
+}
+
bool ARMFastISel::SelectTrunc(const Instruction *I) {
// The high bits for a type smaller than the register size are assumed to be
// undefined.
bool ARMFastISel::SelectIntExt(const Instruction *I) {
// On ARM, in general, integer casts don't involve legal types; this code
// handles promotable integers.
- // FIXME: We could save an instruction in many cases by special-casing
- // load instructions.
Type *DestTy = I->getType();
Value *Src = I->getOperand(0);
Type *SrcTy = Src->getType();
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);
return SelectCall(I);
case Instruction::Select:
return SelectSelect(I);
return false;
}
+/// TryToFoldLoad - 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) {
+ // Verify we have a legal type before going any further.
+ MVT VT;
+ if (!isLoadTypeLegal(LI->getType(), VT))
+ return false;
+
+ // Combine load followed by zero- or sign-extend.
+ // 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;
+ }
+ // 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;
+ MI->eraseFromParent();
+ return true;
+}
+
namespace llvm {
llvm::FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo) {
- // Completely untested on non-darwin.
+ // 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->isTargetDarwin() && !Subtarget->isThumb1Only() &&
+ if (Subtarget->isTargetIOS() && !Subtarget->isThumb1Only() &&
!DisableARMFastISel)
return new ARMFastISel(funcInfo);
return 0;
projects / oota-llvm.git / blobdiff