#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"
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 isZExt,
- bool allocReg);
+ 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);
+ bool ARMEmitStore(EVT VT, unsigned SrcReg, Address &Addr,
+ unsigned Alignment = 0);
bool ARMComputeAddress(const Value *Obj, Address &Addr);
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,
// 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.
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;
case MVT::i8:
case MVT::i16:
case MVT::i32:
- if (!useAM3)
+ if (!useAM3) {
// Integer loads/stores handle 12-bit offsets.
needsLowering = ((Addr.Offset & 0xfff) != Addr.Offset);
- else
+ // 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:
// Now add the rest of the operands.
MIB.addFrameIndex(FI);
- // ARM halfword load/stores and signed byte loads need an additional operand.
+ // 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);
// Now add the rest of the operands.
MIB.addReg(Addr.Base.Reg);
- // ARM halfword load/stores and signed byte loads need an additional operand.
+ // 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);
}
bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg, Address &Addr,
- bool isZExt = true, bool allocReg = true) {
+ 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) {
// This is mostly going to be Neon/vector support.
default: return false;
case MVT::i1:
case MVT::i8:
- if (isZExt) {
- 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 {
- Opc = isThumb2 ? ARM::t2LDRSBi12 : ARM::LDRSB;
- if (!isThumb2) useAM3 = true;
+ if (isZExt) {
+ Opc = ARM::LDRBi12;
+ } else {
+ Opc = ARM::LDRSB;
+ useAM3 = true;
+ }
}
RC = ARM::GPRRegisterClass;
break;
case MVT::i16:
- if (isZExt)
- Opc = isThumb2 ? ARM::t2LDRHi12 : ARM::LDRH;
- else
- Opc = isThumb2 ? ARM::t2LDRSHi12 : ARM::LDRSH;
- if (!isThumb2) useAM3 = true;
+ 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);
+ // 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 (Alignment && Alignment < 4) {
+ // FIXME: Unaligned loads need special handling. Doublewords require
+ // word-alignment.
+ return false;
+ }
Opc = ARM::VLDRD;
RC = TLI.getRegClassFor(VT);
break;
MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(Opc), ResultReg);
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) {
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) useAM3 = true;
+ 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;
+ }
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;
}
// Create the base instruction, then add the operands.
MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(StrOpc))
- .addReg(SrcReg, getKillRegState(true));
+ .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;
}
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) {
unsigned SrcReg1 = getRegForValue(Src1Value);
if (SrcReg1 == 0) return false;
- unsigned SrcReg2;
+ unsigned SrcReg2 = 0;
if (!UseImm) {
SrcReg2 = getRegForValue(Src2Value);
if (SrcReg2 == 0) return false;
(ARM_AM::getSOImmVal(Imm) != -1);
}
- unsigned Op2Reg;
+ unsigned Op2Reg = 0;
if (!UseImm) {
Op2Reg = getRegForValue(I->getOperand(2));
if (Op2Reg == 0) return false;
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();
+ MVT DestVT = VA.getLocVT();
unsigned ResultReg = ARMEmitIntExt(ArgVT, Arg, DestVT,
/*isZExt*/false);
assert (ResultReg != 0 && "Failed to emit a sext");
Arg = ResultReg;
+ ArgVT = DestVT;
break;
}
case CCValAssign::AExt:
// Intentional fall-through. Handle AExt and ZExt.
case CCValAssign::ZExt: {
- EVT DestVT = VA.getLocVT();
+ MVT DestVT = VA.getLocVT();
unsigned ResultReg = ARMEmitIntExt(ArgVT, Arg, DestVT,
/*isZExt*/true);
assert (ResultReg != 0 && "Failed to emit a sext");
Arg = ResultReg;
+ ArgVT = DestVT;
break;
}
case CCValAssign::BCvt: {
if (IntrMemName && e-i <= 2)
break;
- unsigned Arg = getRegForValue(*i);
- 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);
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.");
+
+ 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: {
- // FIXME: Small memcpy/memmove's are common enough that we want to do them
- // without a call if possible.
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 (!ARMComputeAddress(LI->getOperand(0), Addr)) return false;
unsigned ResultReg = MI->getOperand(0).getReg();
- if (!ARMEmitLoad(VT, ResultReg, Addr, isZExt, false))
+ if (!ARMEmitLoad(VT, ResultReg, Addr, LI->getAlignment(), isZExt, false))
return false;
MI->eraseFromParent();
return true;
projects / oota-llvm.git / blobdiff