STATISTIC(NumTailCalls, "Number of tail calls");
-static cl::opt<bool>
-DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
-
// Forward declarations.
static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
SDValue V2);
bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
if (TM.getSubtarget<X86Subtarget>().isTargetDarwin()) {
- if (is64Bit) return new X8664_MachoTargetObjectFile();
+ if (is64Bit)
+ return new X8664_MachoTargetObjectFile();
return new TargetLoweringObjectFileMachO();
- } else if (TM.getSubtarget<X86Subtarget>().isTargetELF() ){
- if (is64Bit) return new X8664_ELFTargetObjectFile(TM);
+ }
+
+ if (TM.getSubtarget<X86Subtarget>().isTargetELF() ){
+ if (is64Bit)
+ return new X8664_ELFTargetObjectFile(TM);
return new X8632_ELFTargetObjectFile(TM);
- } else if (TM.getSubtarget<X86Subtarget>().isTargetCOFF()) {
- return new TargetLoweringObjectFileCOFF();
}
+ if (TM.getSubtarget<X86Subtarget>().isTargetCOFF())
+ return new TargetLoweringObjectFileCOFF();
llvm_unreachable("unknown subtarget type");
}
X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
: TargetLowering(TM, createTLOF(TM)) {
Subtarget = &TM.getSubtarget<X86Subtarget>();
- X86ScalarSSEf64 = Subtarget->hasSSE2();
- X86ScalarSSEf32 = Subtarget->hasSSE1();
+ X86ScalarSSEf64 = Subtarget->hasXMMInt();
+ X86ScalarSSEf32 = Subtarget->hasXMM();
X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
RegInfo = TM.getRegisterInfo();
TD = getTargetData();
// Set up the TargetLowering object.
+ static MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
// X86 is weird, it always uses i8 for shift amounts and setcc results.
setShiftAmountType(MVT::i8);
// (low) operations are left as Legal, as there are single-result
// instructions for this in x86. Using the two-result multiply instructions
// when both high and low results are needed must be arranged by dagcombine.
- setOperationAction(ISD::MULHS , MVT::i8 , Expand);
- setOperationAction(ISD::MULHU , MVT::i8 , Expand);
- setOperationAction(ISD::SDIV , MVT::i8 , Expand);
- setOperationAction(ISD::UDIV , MVT::i8 , Expand);
- setOperationAction(ISD::SREM , MVT::i8 , Expand);
- setOperationAction(ISD::UREM , MVT::i8 , Expand);
- setOperationAction(ISD::MULHS , MVT::i16 , Expand);
- setOperationAction(ISD::MULHU , MVT::i16 , Expand);
- setOperationAction(ISD::SDIV , MVT::i16 , Expand);
- setOperationAction(ISD::UDIV , MVT::i16 , Expand);
- setOperationAction(ISD::SREM , MVT::i16 , Expand);
- setOperationAction(ISD::UREM , MVT::i16 , Expand);
- setOperationAction(ISD::MULHS , MVT::i32 , Expand);
- setOperationAction(ISD::MULHU , MVT::i32 , Expand);
- setOperationAction(ISD::SDIV , MVT::i32 , Expand);
- setOperationAction(ISD::UDIV , MVT::i32 , Expand);
- setOperationAction(ISD::SREM , MVT::i32 , Expand);
- setOperationAction(ISD::UREM , MVT::i32 , Expand);
- setOperationAction(ISD::MULHS , MVT::i64 , Expand);
- setOperationAction(ISD::MULHU , MVT::i64 , Expand);
- setOperationAction(ISD::SDIV , MVT::i64 , Expand);
- setOperationAction(ISD::UDIV , MVT::i64 , Expand);
- setOperationAction(ISD::SREM , MVT::i64 , Expand);
- setOperationAction(ISD::UREM , MVT::i64 , Expand);
+ for (unsigned i = 0, e = 4; i != e; ++i) {
+ MVT VT = IntVTs[i];
+ setOperationAction(ISD::MULHS, VT, Expand);
+ setOperationAction(ISD::MULHU, VT, Expand);
+ setOperationAction(ISD::SDIV, VT, Expand);
+ setOperationAction(ISD::UDIV, VT, Expand);
+ setOperationAction(ISD::SREM, VT, Expand);
+ setOperationAction(ISD::UREM, VT, Expand);
+
+ // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
+ setOperationAction(ISD::ADDC, VT, Custom);
+ setOperationAction(ISD::ADDE, VT, Custom);
+ setOperationAction(ISD::SUBC, VT, Custom);
+ setOperationAction(ISD::SUBE, VT, Custom);
+ }
setOperationAction(ISD::BR_JT , MVT::Other, Expand);
setOperationAction(ISD::BRCOND , MVT::Other, Custom);
setOperationAction(ISD::SELECT , MVT::i1 , Promote);
// X86 wants to expand cmov itself.
setOperationAction(ISD::SELECT , MVT::i8 , Custom);
- setOperationAction(ISD::SELECT , MVT::i16 , Custom);
+ setOperationAction(ISD::SELECT , MVT::i16 , Custom);
setOperationAction(ISD::SELECT , MVT::i32 , Custom);
setOperationAction(ISD::SELECT , MVT::f32 , Custom);
setOperationAction(ISD::SELECT , MVT::f64 , Custom);
setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
}
- if (Subtarget->hasSSE1())
+ if (Subtarget->hasXMM())
setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
// We may not have a libcall for MEMBARRIER so we should lower this.
setShouldFoldAtomicFences(true);
// Expand certain atomics
- setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
- setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
- setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
- setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
-
- setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
- setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
- setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
- setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
-
+ for (unsigned i = 0, e = 4; i != e; ++i) {
+ MVT VT = IntVTs[i];
+ setOperationAction(ISD::ATOMIC_CMP_SWAP, VT, Custom);
+ setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
+ }
+
if (!Subtarget->is64Bit()) {
setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
// FIXME: In order to prevent SSE instructions being expanded to MMX ones
// with -msoft-float, disable use of MMX as well.
- if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
+ if (!UseSoftFloat && Subtarget->hasMMX()) {
addRegisterClass(MVT::x86mmx, X86::VR64RegisterClass);
// No operations on x86mmx supported, everything uses intrinsics.
}
setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
- if (!UseSoftFloat && Subtarget->hasSSE1()) {
+ if (!UseSoftFloat && Subtarget->hasXMM()) {
addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
setOperationAction(ISD::FADD, MVT::v4f32, Legal);
setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
}
- if (!UseSoftFloat && Subtarget->hasSSE2()) {
+ if (!UseSoftFloat && Subtarget->hasXMMInt()) {
addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
// FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
}
}
- if (Subtarget->hasSSE42()) {
+ if (Subtarget->hasSSE42())
setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
- }
if (!UseSoftFloat && Subtarget->hasAVX()) {
addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
// We want to custom lower some of our intrinsics.
setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
- // Add/Sub/Mul with overflow operations are custom lowered.
- setOperationAction(ISD::SADDO, MVT::i32, Custom);
- setOperationAction(ISD::UADDO, MVT::i32, Custom);
- setOperationAction(ISD::SSUBO, MVT::i32, Custom);
- setOperationAction(ISD::USUBO, MVT::i32, Custom);
- setOperationAction(ISD::SMULO, MVT::i32, Custom);
-
+
// Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
// handle type legalization for these operations here.
//
// FIXME: We really should do custom legalization for addition and
// subtraction on x86-32 once PR3203 is fixed. We really can't do much better
// than generic legalization for 64-bit multiplication-with-overflow, though.
- if (Subtarget->is64Bit()) {
- setOperationAction(ISD::SADDO, MVT::i64, Custom);
- setOperationAction(ISD::UADDO, MVT::i64, Custom);
- setOperationAction(ISD::SSUBO, MVT::i64, Custom);
- setOperationAction(ISD::USUBO, MVT::i64, Custom);
- setOperationAction(ISD::SMULO, MVT::i64, Custom);
+ for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
+ // Add/Sub/Mul with overflow operations are custom lowered.
+ MVT VT = IntVTs[i];
+ setOperationAction(ISD::SADDO, VT, Custom);
+ setOperationAction(ISD::UADDO, VT, Custom);
+ setOperationAction(ISD::SSUBO, VT, Custom);
+ setOperationAction(ISD::USUBO, VT, Custom);
+ setOperationAction(ISD::SMULO, VT, Custom);
+ setOperationAction(ISD::UMULO, VT, Custom);
}
+
+ // There are no 8-bit 3-address imul/mul instructions
+ setOperationAction(ISD::SMULO, MVT::i8, Expand);
+ setOperationAction(ISD::UMULO, MVT::i8, Expand);
if (!Subtarget->is64Bit()) {
// These libcalls are not available in 32-bit.
setTargetDAGCombine(ISD::SRA);
setTargetDAGCombine(ISD::SRL);
setTargetDAGCombine(ISD::OR);
+ setTargetDAGCombine(ISD::AND);
+ setTargetDAGCombine(ISD::ADD);
+ setTargetDAGCombine(ISD::SUB);
setTargetDAGCombine(ISD::STORE);
setTargetDAGCombine(ISD::ZERO_EXTEND);
if (Subtarget->is64Bit())
computeRegisterProperties();
- // FIXME: These should be based on subtarget info. Plus, the values should
- // be smaller when we are in optimizing for size mode.
+ // On Darwin, -Os means optimize for size without hurting performance,
+ // do not reduce the limit.
maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
+ maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
- maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
+ maxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
+ maxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
+ maxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
setPrefLoopAlignment(16);
benefitFromCodePlacementOpt = true;
}
}
unsigned Align = 4;
- if (Subtarget->hasSSE1())
+ if (Subtarget->hasXMM())
getMaxByValAlign(Ty, Align);
return Align;
}
// linux. This is because the stack realignment code can't handle certain
// cases like PR2962. This should be removed when PR2962 is fixed.
const Function *F = MF.getFunction();
- if (NonScalarIntSafe &&
- !F->hasFnAttr(Attribute::NoImplicitFloat)) {
+ if (NonScalarIntSafe && !F->hasFnAttr(Attribute::NoImplicitFloat)) {
if (Size >= 16 &&
- (Subtarget->isUnalignedMemAccessFast() ||
- ((DstAlign == 0 || DstAlign >= 16) &&
- (SrcAlign == 0 || SrcAlign >= 16))) &&
Subtarget->getStackAlignment() >= 16) {
if (Subtarget->hasSSE2())
return MVT::v4i32;
} else if (!MemcpyStrSrc && Size >= 8 &&
!Subtarget->is64Bit() &&
Subtarget->getStackAlignment() >= 8 &&
- Subtarget->hasSSE2()) {
+ Subtarget->hasXMMInt()) {
// Do not use f64 to lower memcpy if source is string constant. It's
// better to use i32 to avoid the loads.
return MVT::f64;
// or SSE or MMX vectors.
if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
- (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
+ (Subtarget->is64Bit() && !Subtarget->hasXMM())) {
report_fatal_error("SSE register return with SSE disabled");
}
// Likewise we can't return F64 values with SSE1 only. gcc does so, but
// llvm-gcc has never done it right and no one has noticed, so this
// should be OK for now.
if (ValVT == MVT::f64 &&
- (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
+ (Subtarget->is64Bit() && !Subtarget->hasXMMInt()))
report_fatal_error("SSE2 register return with SSE2 disabled");
// Returns in ST0/ST1 are handled specially: these are pushed as operands to
// If this is x86-64, and we disabled SSE, we can't return FP values
if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
- ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
+ ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasXMM())) {
report_fatal_error("SSE register return with SSE disabled");
}
if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
SDValue Ops[] = { Chain, InFlag };
- Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Flag,
+ Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Glue,
Ops, 2), 1);
Val = Chain.getValue(0);
TotalNumIntRegs);
bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
- assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
+ assert(!(NumXMMRegs && !Subtarget->hasXMM()) &&
"SSE register cannot be used when SSE is disabled!");
assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
"SSE register cannot be used when SSE is disabled!");
- if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
+ if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasXMM())
// Kernel mode asks for SSE to be disabled, so don't push them
// on the stack.
TotalNumXMMRegs = 0;
X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
};
unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
- assert((Subtarget->hasSSE1() || !NumXMMRegs)
+ assert((Subtarget->hasXMM() || !NumXMMRegs)
&& "SSE registers cannot be used when SSE is disabled");
Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
}
// Returns a chain & a flag for retval copy to use.
- SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
SmallVector<SDValue, 8> Ops;
if (!IsSibcall && isTailCall) {
SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
unsigned char OperandFlags) {
MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
- SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
DebugLoc dl = GA->getDebugLoc();
SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
GA->getValueType(0),
OpFlag = X86II::MO_TLVP;
DebugLoc DL = Op.getDebugLoc();
SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
- getPointerTy(),
+ GA->getValueType(0),
GA->getOffset(), OpFlag);
SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
// Lowering the machine isd will make sure everything is in the right
// location.
- SDValue Args[] = { Offset };
- SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
+ SDValue Chain = DAG.getEntryNode();
+ SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
+ SDValue Args[] = { Chain, Offset };
+ Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
// TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
MFI->setAdjustsStack(true);
-
+
// And our return value (tls address) is in the standard call return value
// location.
unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
SDVTList Tys;
bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
if (useSSE)
- Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
+ Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
else
Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
// Lower (X & (1 << N)) == 0 to BT(X, N).
// Lower ((X >>u N) & 1) != 0 to BT(X, N).
// Lower ((X >>s N) & 1) != 0 to BT(X, N).
- if (Op0.getOpcode() == ISD::AND &&
- Op0.hasOneUse() &&
+ if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
Op1.getOpcode() == ISD::Constant &&
cast<ConstantSDNode>(Op1)->isNullValue() &&
(CC == ISD::SETEQ || CC == ISD::SETNE)) {
return NewSetCC;
}
- // Look for "(setcc) == / != 1" to avoid unncessary setcc.
- if (Op0.getOpcode() == X86ISD::SETCC &&
- Op1.getOpcode() == ISD::Constant &&
+ // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
+ // these.
+ if (Op1.getOpcode() == ISD::Constant &&
(cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
cast<ConstantSDNode>(Op1)->isNullValue()) &&
(CC == ISD::SETEQ || CC == ISD::SETNE)) {
- X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
- bool Invert = (CC == ISD::SETNE) ^
- cast<ConstantSDNode>(Op1)->isNullValue();
- if (Invert)
+
+ // If the input is a setcc, then reuse the input setcc or use a new one with
+ // the inverted condition.
+ if (Op0.getOpcode() == X86ISD::SETCC) {
+ X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
+ bool Invert = (CC == ISD::SETNE) ^
+ cast<ConstantSDNode>(Op1)->isNullValue();
+ if (!Invert) return Op0;
+
CCode = X86::GetOppositeBranchCondition(CCode);
- return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
- DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
+ return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
+ DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
+ }
}
bool isFP = Op1.getValueType().isFloatingPoint();
if (X86CC == X86::COND_INVALID)
return SDValue();
- SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
-
- // Use sbb x, x to materialize carry bit into a GPR.
- if (X86CC == X86::COND_B)
- return DAG.getNode(ISD::AND, dl, MVT::i8,
- DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
- DAG.getConstant(X86CC, MVT::i8), Cond),
- DAG.getConstant(1, MVT::i8));
-
+ SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
- DAG.getConstant(X86CC, MVT::i8), Cond);
+ DAG.getConstant(X86CC, MVT::i8), EFLAGS);
}
SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
if (Op.getResNo() == 1 &&
(Opc == X86ISD::ADD ||
Opc == X86ISD::SUB ||
+ Opc == X86ISD::ADC ||
+ Opc == X86ISD::SBB ||
Opc == X86ISD::SMUL ||
Opc == X86ISD::UMUL ||
Opc == X86ISD::INC ||
Opc == X86ISD::AND))
return true;
+ if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
+ return true;
+
return false;
}
Cond = EmitTest(Cond, X86::COND_NE, DAG);
}
+ // a < b ? -1 : 0 -> RES = ~setcc_carry
+ // a < b ? 0 : -1 -> RES = setcc_carry
+ // a >= b ? -1 : 0 -> RES = setcc_carry
+ // a >= b ? 0 : -1 -> RES = ~setcc_carry
+ if (Cond.getOpcode() == X86ISD::CMP) {
+ unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
+
+ if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
+ (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
+ SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
+ DAG.getConstant(X86::COND_B, MVT::i8), Cond);
+ if (isAllOnes(Op1) != (CondCode == X86::COND_B))
+ return DAG.getNOT(DL, Res, Res.getValueType());
+ return Res;
+ }
+ }
+
// X86ISD::CMOV means set the result (which is operand 1) to the RHS if
// condition is true.
- SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
+ SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
SDValue Ops[] = { Op2, Op1, CC, Cond };
return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
}
Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
Flag = Chain.getValue(1);
- SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
Flag = Chain.getValue(1);
assert(!UseSoftFloat &&
!(DAG.getMachineFunction()
.getFunction()->hasFnAttr(Attribute::NoImplicitFloat)) &&
- Subtarget->hasSSE1());
+ Subtarget->hasXMM());
}
// Insert VAARG_64 node into the DAG
M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
DAG.getConstant(4, MVT::i32));
- R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
- DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
- R, M, Op);
+ R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
// a += a
Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
DAG.getConstant(2, MVT::i32));
- R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
- DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
- R, M, Op);
+ R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
// a += a
Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
// return pblendv(r, r+r, a);
- R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
- DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
+ R = DAG.getNode(X86ISD::PBLENDVB, dl, VT,
R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
return R;
}
SDValue RHS = N->getOperand(1);
unsigned BaseOp = 0;
unsigned Cond = 0;
- DebugLoc dl = Op.getDebugLoc();
-
+ DebugLoc DL = Op.getDebugLoc();
switch (Op.getOpcode()) {
default: llvm_unreachable("Unknown ovf instruction!");
case ISD::SADDO:
BaseOp = X86ISD::SMUL;
Cond = X86::COND_O;
break;
- case ISD::UMULO:
- BaseOp = X86ISD::UMUL;
- Cond = X86::COND_B;
- break;
+ case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
+ SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
+ MVT::i32);
+ SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
+
+ SDValue SetCC =
+ DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
+ DAG.getConstant(X86::COND_O, MVT::i32),
+ SDValue(Sum.getNode(), 2));
+
+ DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
+ return Sum;
+ }
}
// Also sets EFLAGS.
SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
- SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
+ SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
SDValue SetCC =
- DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
- DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
+ DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
+ DAG.getConstant(Cond, MVT::i32),
+ SDValue(Sum.getNode(), 1));
DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
return Sum;
Op.getOperand(3),
DAG.getTargetConstant(size, MVT::i8),
cpIn.getValue(1) };
- SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
Ops, 5, T, MMO);
SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
SelectionDAG &DAG) const {
assert(Subtarget->is64Bit() && "Result not type legalized?");
- SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue TheChain = Op.getOperand(0);
DebugLoc dl = Op.getDebugLoc();
SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
SelectionDAG &DAG) const {
EVT SrcVT = Op.getOperand(0).getValueType();
EVT DstVT = Op.getValueType();
- assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
- Subtarget->hasMMX() && !DisableMMX) &&
- "Unexpected custom BITCAST");
+ assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
+ Subtarget->hasMMX() && "Unexpected custom BITCAST");
assert((DstVT == MVT::i64 ||
(DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
"Unexpected custom BITCAST");
// All other conversions need to be expanded.
return SDValue();
}
+
SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
SDNode *Node = Op.getNode();
DebugLoc dl = Node->getDebugLoc();
cast<AtomicSDNode>(Node)->getAlignment());
}
+static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
+ EVT VT = Op.getNode()->getValueType(0);
+
+ // Let legalize expand this if it isn't a legal type yet.
+ if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
+ return SDValue();
+
+ SDVTList VTs = DAG.getVTList(VT, MVT::i32);
+
+ unsigned Opc;
+ bool ExtraOp = false;
+ switch (Op.getOpcode()) {
+ default: assert(0 && "Invalid code");
+ case ISD::ADDC: Opc = X86ISD::ADD; break;
+ case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
+ case ISD::SUBC: Opc = X86ISD::SUB; break;
+ case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
+ }
+
+ if (!ExtraOp)
+ return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
+ Op.getOperand(1));
+ return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
+ Op.getOperand(1), Op.getOperand(2));
+}
+
/// LowerOperation - Provide custom lowering hooks for some operations.
///
SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
case ISD::UMULO: return LowerXALUO(Op, DAG);
case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
case ISD::BITCAST: return LowerBITCAST(Op, DAG);
+ case ISD::ADDC:
+ case ISD::ADDE:
+ case ISD::SUBC:
+ case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
}
}
default:
assert(false && "Do not know how to custom type legalize this operation!");
return;
+ case ISD::ADDC:
+ case ISD::ADDE:
+ case ISD::SUBC:
+ case ISD::SUBE:
+ // We don't want to expand or promote these.
+ return;
case ISD::FP_TO_SINT: {
std::pair<SDValue,SDValue> Vals =
FP_TO_INTHelper(SDValue(N, 0), DAG, true);
return;
}
case ISD::READCYCLECOUNTER: {
- SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue TheChain = N->getOperand(0);
SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
SDValue Ops[] = { swapInH.getValue(0),
N->getOperand(1),
swapInH.getValue(1) };
- SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG8_DAG, dl, Tys,
Ops, 3, T, MMO);
case X86ISD::PINSRB: return "X86ISD::PINSRB";
case X86ISD::PINSRW: return "X86ISD::PINSRW";
case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
+ case X86ISD::PANDN: return "X86ISD::PANDN";
+ case X86ISD::PSIGNB: return "X86ISD::PSIGNB";
+ case X86ISD::PSIGNW: return "X86ISD::PSIGNW";
+ case X86ISD::PSIGND: return "X86ISD::PSIGND";
+ case X86ISD::PBLENDVB: return "X86ISD::PBLENDVB";
case X86ISD::FMAX: return "X86ISD::FMAX";
case X86ISD::FMIN: return "X86ISD::FMIN";
case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
case X86ISD::ADD: return "X86ISD::ADD";
case X86ISD::SUB: return "X86ISD::SUB";
+ case X86ISD::ADC: return "X86ISD::ADC";
+ case X86ISD::SBB: return "X86ISD::SBB";
case X86ISD::SMUL: return "X86ISD::SMUL";
case X86ISD::UMUL: return "X86ISD::UMUL";
case X86ISD::INC: return "X86ISD::INC";
default: break;
case X86ISD::ADD:
case X86ISD::SUB:
+ case X86ISD::ADC:
+ case X86ISD::SBB:
case X86ISD::SMUL:
case X86ISD::UMUL:
case X86ISD::INC:
/// if the load addresses are consecutive, non-overlapping, and in the right
/// order.
static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
- const TargetLowering &TLI) {
+ TargetLowering::DAGCombinerInfo &DCI) {
DebugLoc dl = N->getDebugLoc();
EVT VT = N->getValueType(0);
if (VT.getSizeInBits() != 128)
return SDValue();
+ // Don't create instructions with illegal types after legalize types has run.
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+ if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
+ return SDValue();
+
SmallVector<SDValue, 16> Elts;
for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
return SDValue();
}
+
+static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
+ TargetLowering::DAGCombinerInfo &DCI,
+ const X86Subtarget *Subtarget) {
+ if (DCI.isBeforeLegalizeOps())
+ return SDValue();
+
+ // Want to form PANDN nodes, in the hopes of then easily combining them with
+ // OR and AND nodes to form PBLEND/PSIGN.
+ EVT VT = N->getValueType(0);
+ if (VT != MVT::v2i64)
+ return SDValue();
+
+ SDValue N0 = N->getOperand(0);
+ SDValue N1 = N->getOperand(1);
+ DebugLoc DL = N->getDebugLoc();
+
+ // Check LHS for vnot
+ if (N0.getOpcode() == ISD::XOR &&
+ ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
+ return DAG.getNode(X86ISD::PANDN, DL, VT, N0.getOperand(0), N1);
+
+ // Check RHS for vnot
+ if (N1.getOpcode() == ISD::XOR &&
+ ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
+ return DAG.getNode(X86ISD::PANDN, DL, VT, N1.getOperand(0), N0);
+
+ return SDValue();
+}
+
static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const X86Subtarget *Subtarget) {
return SDValue();
EVT VT = N->getValueType(0);
- if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
+ if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64 && VT != MVT::v2i64)
return SDValue();
- // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
+
+ // look for psign/blend
+ if (Subtarget->hasSSSE3()) {
+ if (VT == MVT::v2i64) {
+ // Canonicalize pandn to RHS
+ if (N0.getOpcode() == X86ISD::PANDN)
+ std::swap(N0, N1);
+ // or (and (m, x), (pandn m, y))
+ if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::PANDN) {
+ SDValue Mask = N1.getOperand(0);
+ SDValue X = N1.getOperand(1);
+ SDValue Y;
+ if (N0.getOperand(0) == Mask)
+ Y = N0.getOperand(1);
+ if (N0.getOperand(1) == Mask)
+ Y = N0.getOperand(0);
+
+ // Check to see if the mask appeared in both the AND and PANDN and
+ if (!Y.getNode())
+ return SDValue();
+
+ // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
+ if (Mask.getOpcode() != ISD::BITCAST ||
+ X.getOpcode() != ISD::BITCAST ||
+ Y.getOpcode() != ISD::BITCAST)
+ return SDValue();
+
+ // Look through mask bitcast.
+ Mask = Mask.getOperand(0);
+ EVT MaskVT = Mask.getValueType();
+
+ // Validate that the Mask operand is a vector sra node. The sra node
+ // will be an intrinsic.
+ if (Mask.getOpcode() != ISD::INTRINSIC_WO_CHAIN)
+ return SDValue();
+
+ // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
+ // there is no psrai.b
+ switch (cast<ConstantSDNode>(Mask.getOperand(0))->getZExtValue()) {
+ case Intrinsic::x86_sse2_psrai_w:
+ case Intrinsic::x86_sse2_psrai_d:
+ break;
+ default: return SDValue();
+ }
+
+ // Check that the SRA is all signbits.
+ SDValue SraC = Mask.getOperand(2);
+ unsigned SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
+ unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
+ if ((SraAmt + 1) != EltBits)
+ return SDValue();
+
+ DebugLoc DL = N->getDebugLoc();
+
+ // Now we know we at least have a plendvb with the mask val. See if
+ // we can form a psignb/w/d.
+ // psign = x.type == y.type == mask.type && y = sub(0, x);
+ X = X.getOperand(0);
+ Y = Y.getOperand(0);
+ if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
+ ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
+ X.getValueType() == MaskVT && X.getValueType() == Y.getValueType()){
+ unsigned Opc = 0;
+ switch (EltBits) {
+ case 8: Opc = X86ISD::PSIGNB; break;
+ case 16: Opc = X86ISD::PSIGNW; break;
+ case 32: Opc = X86ISD::PSIGND; break;
+ default: break;
+ }
+ if (Opc) {
+ SDValue Sign = DAG.getNode(Opc, DL, MaskVT, X, Mask.getOperand(1));
+ return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Sign);
+ }
+ }
+ // PBLENDVB only available on SSE 4.1
+ if (!Subtarget->hasSSE41())
+ return SDValue();
+
+ X = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, X);
+ Y = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Y);
+ Mask = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Mask);
+ Mask = DAG.getNode(X86ISD::PBLENDVB, DL, MVT::v16i8, X, Y, Mask);
+ return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Mask);
+ }
+ }
+ }
+
+ // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
std::swap(N0, N1);
if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
DAG.getNode(ISD::TRUNCATE, DL,
MVT::i8, ShAmt0));
}
-
+
return SDValue();
}
return SDValue();
}
+// Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
+static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG) {
+ unsigned X86CC = N->getConstantOperandVal(0);
+ SDValue EFLAG = N->getOperand(1);
+ DebugLoc DL = N->getDebugLoc();
+
+ // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
+ // a zext and produces an all-ones bit which is more useful than 0/1 in some
+ // cases.
+ if (X86CC == X86::COND_B)
+ return DAG.getNode(ISD::AND, DL, MVT::i8,
+ DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
+ DAG.getConstant(X86CC, MVT::i8), EFLAG),
+ DAG.getConstant(1, MVT::i8));
+
+ return SDValue();
+}
+
+// Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
+static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
+ X86TargetLowering::DAGCombinerInfo &DCI) {
+ // If the LHS and RHS of the ADC node are zero, then it can't overflow and
+ // the result is either zero or one (depending on the input carry bit).
+ // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
+ if (X86::isZeroNode(N->getOperand(0)) &&
+ X86::isZeroNode(N->getOperand(1)) &&
+ // We don't have a good way to replace an EFLAGS use, so only do this when
+ // dead right now.
+ SDValue(N, 1).use_empty()) {
+ DebugLoc DL = N->getDebugLoc();
+ EVT VT = N->getValueType(0);
+ SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
+ SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
+ DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
+ DAG.getConstant(X86::COND_B,MVT::i8),
+ N->getOperand(2)),
+ DAG.getConstant(1, VT));
+ return DCI.CombineTo(N, Res1, CarryOut);
+ }
+
+ return SDValue();
+}
+
+// fold (add Y, (sete X, 0)) -> adc 0, Y
+// (add Y, (setne X, 0)) -> sbb -1, Y
+// (sub (sete X, 0), Y) -> sbb 0, Y
+// (sub (setne X, 0), Y) -> adc -1, Y
+static SDValue OptimizeConditonalInDecrement(SDNode *N, SelectionDAG &DAG) {
+ DebugLoc DL = N->getDebugLoc();
+
+ // Look through ZExts.
+ SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
+ if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
+ return SDValue();
+
+ SDValue SetCC = Ext.getOperand(0);
+ if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
+ return SDValue();
+
+ X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
+ if (CC != X86::COND_E && CC != X86::COND_NE)
+ return SDValue();
+
+ SDValue Cmp = SetCC.getOperand(1);
+ if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
+ !X86::isZeroNode(Cmp.getOperand(1)))
+ return SDValue();
+
+ SDValue CmpOp0 = Cmp.getOperand(0);
+ SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
+ DAG.getConstant(1, CmpOp0.getValueType()));
+
+ SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
+ if (CC == X86::COND_NE)
+ return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
+ DL, OtherVal.getValueType(), OtherVal,
+ DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
+ return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
+ DL, OtherVal.getValueType(), OtherVal,
+ DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
+}
+
SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
DAGCombinerInfo &DCI) const {
SelectionDAG &DAG = DCI.DAG;
switch (N->getOpcode()) {
default: break;
case ISD::EXTRACT_VECTOR_ELT:
- return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
+ return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
+ case ISD::ADD:
+ case ISD::SUB: return OptimizeConditonalInDecrement(N, DAG);
+ case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
case ISD::SHL:
case ISD::SRA:
case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
+ case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
case X86ISD::FXOR:
case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
+ case X86ISD::SETCC: return PerformSETCCCombine(N, DAG);
case X86ISD::SHUFPS: // Handle all target specific shuffles
case X86ISD::SHUFPD:
case X86ISD::PALIGN:
case X86ISD::PSHUFLW:
case X86ISD::MOVSS:
case X86ISD::MOVSD:
- case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
+ case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI);
}
return SDValue();
weight = CW_SpecificReg;
break;
case 'y':
- if (type->isX86_MMXTy() && !DisableMMX && Subtarget->hasMMX())
+ if (type->isX86_MMXTy() && Subtarget->hasMMX())
weight = CW_SpecificReg;
break;
case 'x':
case 'Y':
- if ((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1())
+ if ((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasXMM())
weight = CW_Register;
break;
case 'I':
// FP X constraints get lowered to SSE1/2 registers if available, otherwise
// 'f' like normal targets.
if (ConstraintVT.isFloatingPoint()) {
- if (Subtarget->hasSSE2())
+ if (Subtarget->hasXMMInt())
return "Y";
- if (Subtarget->hasSSE1())
+ if (Subtarget->hasXMM())
return "x";
}
if (!Subtarget->hasMMX()) break;
return std::make_pair(0U, X86::VR64RegisterClass);
case 'Y': // SSE_REGS if SSE2 allowed
- if (!Subtarget->hasSSE2()) break;
+ if (!Subtarget->hasXMMInt()) break;
// FALL THROUGH.
case 'x': // SSE_REGS if SSE1 allowed
- if (!Subtarget->hasSSE1()) break;
+ if (!Subtarget->hasXMM()) break;
switch (VT.getSimpleVT().SimpleTy) {
default: break;
projects / oota-llvm.git / blobdiff