#include "llvm/CallingConv.h"
#include "llvm/DerivedTypes.h"
#include "llvm/GlobalVariable.h"
+#include "llvm/GlobalAlias.h"
#include "llvm/Instructions.h"
#include "llvm/IntrinsicInst.h"
#include "llvm/Operator.h"
explicit X86FastISel(FunctionLoweringInfo &funcInfo) : FastISel(funcInfo) {
Subtarget = &TM.getSubtarget<X86Subtarget>();
StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
- X86ScalarSSEf64 = Subtarget->hasSSE2() || Subtarget->hasAVX();
- X86ScalarSSEf32 = Subtarget->hasSSE1() || Subtarget->hasAVX();
+ X86ScalarSSEf64 = Subtarget->hasSSE2();
+ X86ScalarSSEf32 = Subtarget->hasSSE1();
}
virtual bool TargetSelectInstruction(const Instruction *I);
case MVT::i1:
case MVT::i8:
Opc = X86::MOV8rm;
- RC = X86::GR8RegisterClass;
+ RC = &X86::GR8RegClass;
break;
case MVT::i16:
Opc = X86::MOV16rm;
- RC = X86::GR16RegisterClass;
+ RC = &X86::GR16RegClass;
break;
case MVT::i32:
Opc = X86::MOV32rm;
- RC = X86::GR32RegisterClass;
+ RC = &X86::GR32RegClass;
break;
case MVT::i64:
// Must be in x86-64 mode.
Opc = X86::MOV64rm;
- RC = X86::GR64RegisterClass;
+ RC = &X86::GR64RegClass;
break;
case MVT::f32:
- if (Subtarget->hasSSE1()) {
- Opc = X86::MOVSSrm;
- RC = X86::FR32RegisterClass;
+ if (X86ScalarSSEf32) {
+ Opc = Subtarget->hasAVX() ? X86::VMOVSSrm : X86::MOVSSrm;
+ RC = &X86::FR32RegClass;
} else {
Opc = X86::LD_Fp32m;
- RC = X86::RFP32RegisterClass;
+ RC = &X86::RFP32RegClass;
}
break;
case MVT::f64:
- if (Subtarget->hasSSE2()) {
- Opc = X86::MOVSDrm;
- RC = X86::FR64RegisterClass;
+ if (X86ScalarSSEf64) {
+ Opc = Subtarget->hasAVX() ? X86::VMOVSDrm : X86::MOVSDrm;
+ RC = &X86::FR64RegClass;
} else {
Opc = X86::LD_Fp64m;
- RC = X86::RFP64RegisterClass;
+ RC = &X86::RFP64RegClass;
}
break;
case MVT::f80:
default: return false;
case MVT::i1: {
// Mask out all but lowest bit.
- unsigned AndResult = createResultReg(X86::GR8RegisterClass);
+ unsigned AndResult = createResultReg(&X86::GR8RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(X86::AND8ri), AndResult).addReg(Val).addImm(1);
Val = AndResult;
case MVT::i32: Opc = X86::MOV32mr; break;
case MVT::i64: Opc = X86::MOV64mr; break; // Must be in x86-64 mode.
case MVT::f32:
- Opc = Subtarget->hasSSE1() ? X86::MOVSSmr : X86::ST_Fp32m;
+ Opc = X86ScalarSSEf32 ?
+ (Subtarget->hasAVX() ? X86::VMOVSSmr : X86::MOVSSmr) : X86::ST_Fp32m;
break;
case MVT::f64:
- Opc = Subtarget->hasSSE2() ? X86::MOVSDmr : X86::ST_Fp64m;
+ Opc = X86ScalarSSEf64 ?
+ (Subtarget->hasAVX() ? X86::VMOVSDmr : X86::MOVSDmr) : X86::ST_Fp64m;
+ break;
+ case MVT::v4f32:
+ Opc = X86::MOVAPSmr;
+ break;
+ case MVT::v2f64:
+ Opc = X86::MOVAPDmr;
+ break;
+ case MVT::v4i32:
+ case MVT::v2i64:
+ case MVT::v8i16:
+ case MVT::v16i8:
+ Opc = X86::MOVDQAmr;
break;
}
// Handle constant address.
if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
- // Can't handle alternate code models or TLS yet.
+ // Can't handle alternate code models yet.
if (TM.getCodeModel() != CodeModel::Small)
return false;
+ // Can't handle TLS yet.
if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
if (GVar->isThreadLocal())
return false;
+ // Can't handle TLS yet, part 2 (this is slightly crazy, but this is how
+ // it works...).
+ if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
+ if (const GlobalVariable *GVar =
+ dyn_cast_or_null<GlobalVariable>(GA->resolveAliasedGlobal(false)))
+ if (GVar->isThreadLocal())
+ return false;
+
// RIP-relative addresses can't have additional register operands, so if
// we've already folded stuff into the addressing mode, just force the
// global value into its own register, which we can use as the basereg.
if (TLI.getPointerTy() == MVT::i64) {
Opc = X86::MOV64rm;
- RC = X86::GR64RegisterClass;
+ RC = &X86::GR64RegClass;
if (Subtarget->isPICStyleRIPRel())
StubAM.Base.Reg = X86::RIP;
} else {
Opc = X86::MOV32rm;
- RC = X86::GR32RegisterClass;
+ RC = &X86::GR32RegClass;
}
LoadReg = createResultReg(RC);
/// X86SelectStore - Select and emit code to implement store instructions.
bool X86FastISel::X86SelectStore(const Instruction *I) {
+ // Atomic stores need special handling.
+ const StoreInst *S = cast<StoreInst>(I);
+
+ if (S->isAtomic())
+ return false;
+
+ unsigned SABIAlignment =
+ TD.getABITypeAlignment(S->getValueOperand()->getType());
+ if (S->getAlignment() != 0 && S->getAlignment() < SABIAlignment)
+ return false;
+
MVT VT;
if (!isTypeLegal(I->getOperand(0)->getType(), VT, /*AllowI1=*/true))
return false;
// fastcc with -tailcallopt is intended to provide a guaranteed
// tail call optimization. Fastisel doesn't know how to do that.
- if (CC == CallingConv::Fast && GuaranteedTailCallOpt)
+ if (CC == CallingConv::Fast && TM.Options.GuaranteedTailCallOpt)
return false;
// Let SDISel handle vararg functions.
/// X86SelectLoad - Select and emit code to implement load instructions.
///
bool X86FastISel::X86SelectLoad(const Instruction *I) {
+ // Atomic loads need special handling.
+ if (cast<LoadInst>(I)->isAtomic())
+ return false;
+
MVT VT;
if (!isTypeLegal(I->getType(), VT, /*AllowI1=*/true))
return false;
}
static unsigned X86ChooseCmpOpcode(EVT VT, const X86Subtarget *Subtarget) {
+ bool HasAVX = Subtarget->hasAVX();
+ bool X86ScalarSSEf32 = Subtarget->hasSSE1();
+ bool X86ScalarSSEf64 = Subtarget->hasSSE2();
+
switch (VT.getSimpleVT().SimpleTy) {
default: return 0;
case MVT::i8: return X86::CMP8rr;
case MVT::i16: return X86::CMP16rr;
case MVT::i32: return X86::CMP32rr;
case MVT::i64: return X86::CMP64rr;
- case MVT::f32: return Subtarget->hasSSE1() ? X86::UCOMISSrr : 0;
- case MVT::f64: return Subtarget->hasSSE2() ? X86::UCOMISDrr : 0;
+ case MVT::f32:
+ return X86ScalarSSEf32 ? (HasAVX ? X86::VUCOMISSrr : X86::UCOMISSrr) : 0;
+ case MVT::f64:
+ return X86ScalarSSEf64 ? (HasAVX ? X86::VUCOMISDrr : X86::UCOMISDrr) : 0;
}
}
bool X86FastISel::X86SelectFPExt(const Instruction *I) {
// fpext from float to double.
- if (Subtarget->hasSSE2() &&
+ if (X86ScalarSSEf64 &&
I->getType()->isDoubleTy()) {
const Value *V = I->getOperand(0);
if (V->getType()->isFloatTy()) {
unsigned OpReg = getRegForValue(V);
if (OpReg == 0) return false;
- unsigned ResultReg = createResultReg(X86::FR64RegisterClass);
+ unsigned ResultReg = createResultReg(&X86::FR64RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(X86::CVTSS2SDrr), ResultReg)
.addReg(OpReg);
}
bool X86FastISel::X86SelectFPTrunc(const Instruction *I) {
- if (Subtarget->hasSSE2()) {
+ if (X86ScalarSSEf64) {
if (I->getType()->isFloatTy()) {
const Value *V = I->getOperand(0);
if (V->getType()->isDoubleTy()) {
unsigned OpReg = getRegForValue(V);
if (OpReg == 0) return false;
- unsigned ResultReg = createResultReg(X86::FR32RegisterClass);
+ unsigned ResultReg = createResultReg(&X86::FR32RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(X86::CVTSD2SSrr), ResultReg)
.addReg(OpReg);
if (!Subtarget->is64Bit()) {
// If we're on x86-32; we can't extract an i8 from a general register.
// First issue a copy to GR16_ABCD or GR32_ABCD.
- const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16)
- ? X86::GR16_ABCDRegisterClass : X86::GR32_ABCDRegisterClass;
+ const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16) ?
+ (const TargetRegisterClass*)&X86::GR16_ABCDRegClass :
+ (const TargetRegisterClass*)&X86::GR32_ABCDRegClass;
unsigned CopyReg = createResultReg(CopyRC);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
CopyReg).addReg(InputReg);
case Intrinsic::memset: {
const MemSetInst &MSI = cast<MemSetInst>(I);
+ if (MSI.isVolatile())
+ return false;
+
unsigned SizeWidth = Subtarget->is64Bit() ? 64 : 32;
if (!MSI.getLength()->getType()->isIntegerTy(SizeWidth))
return false;
return DoSelectCall(&I, "memset");
}
case Intrinsic::stackprotector: {
- // Emit code inline code to store the stack guard onto the stack.
+ // Emit code to store the stack guard onto the stack.
EVT PtrTy = TLI.getPointerTy();
const Value *Op1 = I.getArgOperand(0); // The guard's value.
// fastcc with -tailcallopt is intended to provide a guaranteed
// tail call optimization. Fastisel doesn't know how to do that.
- if (CC == CallingConv::Fast && GuaranteedTailCallOpt)
+ if (CC == CallingConv::Fast && TM.Options.GuaranteedTailCallOpt)
return false;
PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
// Fast-isel doesn't know about callee-pop yet.
if (X86::isCalleePop(CC, Subtarget->is64Bit(), isVarArg,
- GuaranteedTailCallOpt))
+ TM.Options.GuaranteedTailCallOpt))
return false;
// Check whether the function can return without sret-demotion.
SmallVector<ISD::OutputArg, 4> Outs;
- SmallVector<uint64_t, 4> Offsets;
GetReturnInfo(I->getType(), CS.getAttributes().getRetAttributes(),
- Outs, TLI, &Offsets);
+ Outs, TLI);
bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(),
*FuncInfo.MF, FTy->isVarArg(),
Outs, FTy->getContext());
SmallVector<unsigned, 8> Args;
SmallVector<MVT, 8> ArgVTs;
SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
- Args.reserve(CS.arg_size());
- ArgVals.reserve(CS.arg_size());
- ArgVTs.reserve(CS.arg_size());
- ArgFlags.reserve(CS.arg_size());
+ unsigned arg_size = CS.arg_size();
+ Args.reserve(arg_size);
+ ArgVals.reserve(arg_size);
+ ArgVTs.reserve(arg_size);
+ ArgFlags.reserve(arg_size);
for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
i != e; ++i) {
// If we're lowering a mem intrinsic instead of a regular call, skip the
assert(Res && "memcpy length already checked!"); (void)Res;
} else if (isa<ConstantInt>(ArgVal) || isa<ConstantPointerNull>(ArgVal)) {
// If this is a really simple value, emit this with the Value* version
- //of X86FastEmitStore. If it isn't simple, we don't want to do this,
+ // of X86FastEmitStore. If it isn't simple, we don't want to do this,
// as it can cause us to reevaluate the argument.
- X86FastEmitStore(ArgVT, ArgVal, AM);
+ if (!X86FastEmitStore(ArgVT, ArgVal, AM))
+ return false;
} else {
- X86FastEmitStore(ArgVT, Arg, AM);
+ if (!X86FastEmitStore(ArgVT, Arg, AM))
+ return false;
}
}
}
if (Subtarget->is64Bit() && isVarArg && !Subtarget->isTargetWin64()) {
// Count the number of XMM registers allocated.
- static const unsigned XMMArgRegs[] = {
+ static const uint16_t XMMArgRegs[] = {
X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
};
if (CalleeOp) {
// Register-indirect call.
unsigned CallOpc;
- if (Subtarget->isTargetWin64())
- CallOpc = X86::WINCALL64r;
- else if (Subtarget->is64Bit())
+ if (Subtarget->is64Bit())
CallOpc = X86::CALL64r;
else
CallOpc = X86::CALL32r;
// Direct call.
assert(GV && "Not a direct call");
unsigned CallOpc;
- if (Subtarget->isTargetWin64())
- CallOpc = X86::WINCALL64pcrel32;
- else if (Subtarget->is64Bit())
+ if (Subtarget->is64Bit())
CallOpc = X86::CALL64pcrel32;
else
CallOpc = X86::CALLpcrel32;
MIB.addGlobalAddress(GV, 0, OpFlags);
}
+ // Add a register mask with the call-preserved registers.
+ // Proper defs for return values will be added by setPhysRegsDeadExcept().
+ MIB.addRegMask(TRI.getCallPreservedMask(CS.getCallingConv()));
+
// Add an implicit use GOT pointer in EBX.
if (Subtarget->isPICStyleGOT())
- MIB.addReg(X86::EBX);
+ MIB.addReg(X86::EBX, RegState::Implicit);
if (Subtarget->is64Bit() && isVarArg && !Subtarget->isTargetWin64())
- MIB.addReg(X86::AL);
+ MIB.addReg(X86::AL, RegState::Implicit);
// Add implicit physical register uses to the call.
for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
- MIB.addReg(RegArgs[i]);
+ MIB.addReg(RegArgs[i], RegState::Implicit);
// Issue CALLSEQ_END
unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
unsigned NumBytesCallee = 0;
- if (!Subtarget->is64Bit() && CS.paramHasAttr(1, Attribute::StructRet))
+ if (!Subtarget->is64Bit() && !Subtarget->isTargetWindows() &&
+ !(CS.getCallingConv() == CallingConv::Fast ||
+ CS.getCallingConv() == CallingConv::GHC) &&
+ CS.paramHasAttr(1, Attribute::StructRet))
NumBytesCallee = 4;
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackUp))
.addImm(NumBytes).addImm(NumBytesCallee);
RVLocs[i].getLocReg() == X86::ST1)) {
if (isScalarFPTypeInSSEReg(RVLocs[i].getValVT())) {
CopyVT = MVT::f80;
- CopyReg = createResultReg(X86::RFP80RegisterClass);
+ CopyReg = createResultReg(&X86::RFP80RegClass);
}
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::FpPOP_RETVAL),
CopyReg);
default: return false;
case MVT::i8:
Opc = X86::MOV8rm;
- RC = X86::GR8RegisterClass;
+ RC = &X86::GR8RegClass;
break;
case MVT::i16:
Opc = X86::MOV16rm;
- RC = X86::GR16RegisterClass;
+ RC = &X86::GR16RegClass;
break;
case MVT::i32:
Opc = X86::MOV32rm;
- RC = X86::GR32RegisterClass;
+ RC = &X86::GR32RegClass;
break;
case MVT::i64:
// Must be in x86-64 mode.
Opc = X86::MOV64rm;
- RC = X86::GR64RegisterClass;
+ RC = &X86::GR64RegClass;
break;
case MVT::f32:
- if (Subtarget->hasSSE1()) {
- Opc = X86::MOVSSrm;
- RC = X86::FR32RegisterClass;
+ if (X86ScalarSSEf32) {
+ Opc = Subtarget->hasAVX() ? X86::VMOVSSrm : X86::MOVSSrm;
+ RC = &X86::FR32RegClass;
} else {
Opc = X86::LD_Fp32m;
- RC = X86::RFP32RegisterClass;
+ RC = &X86::RFP32RegClass;
}
break;
case MVT::f64:
- if (Subtarget->hasSSE2()) {
- Opc = X86::MOVSDrm;
- RC = X86::FR64RegisterClass;
+ if (X86ScalarSSEf64) {
+ Opc = Subtarget->hasAVX() ? X86::VMOVSDrm : X86::MOVSDrm;
+ RC = &X86::FR64RegClass;
} else {
Opc = X86::LD_Fp64m;
- RC = X86::RFP64RegisterClass;
+ RC = &X86::RFP64RegClass;
}
break;
case MVT::f80:
if (!X86SelectAddress(C, AM))
return 0;
unsigned Opc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
- TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
+ const TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
unsigned ResultReg = createResultReg(RC);
addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(Opc), ResultReg), AM);
switch (VT.SimpleTy) {
default: return false;
case MVT::f32:
- if (Subtarget->hasSSE1()) {
+ if (X86ScalarSSEf32) {
Opc = X86::FsFLD0SS;
- RC = X86::FR32RegisterClass;
+ RC = &X86::FR32RegClass;
} else {
Opc = X86::LD_Fp032;
- RC = X86::RFP32RegisterClass;
+ RC = &X86::RFP32RegClass;
}
break;
case MVT::f64:
- if (Subtarget->hasSSE2()) {
+ if (X86ScalarSSEf64) {
Opc = X86::FsFLD0SD;
- RC = X86::FR64RegisterClass;
+ RC = &X86::FR64RegClass;
} else {
Opc = X86::LD_Fp064;
- RC = X86::RFP64RegisterClass;
+ RC = &X86::RFP64RegClass;
}
break;
case MVT::f80:
namespace llvm {
- llvm::FastISel *X86::createFastISel(FunctionLoweringInfo &funcInfo) {
+ FastISel *X86::createFastISel(FunctionLoweringInfo &funcInfo) {
return new X86FastISel(funcInfo);
}
}
projects / oota-llvm.git / blobdiff