unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / 128)
* ElemsPerChunk);
+ // If the input is a buildvector just emit a smaller one.
+ if (Vec.getOpcode() == ISD::BUILD_VECTOR)
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
+ Vec->op_begin()+NormalizedIdxVal, ElemsPerChunk);
+
SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
SDValue Result = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec,
VecIdx);
Subtarget = &TM.getSubtarget<X86Subtarget>();
X86ScalarSSEf64 = Subtarget->hasSSE2();
X86ScalarSSEf32 = Subtarget->hasSSE1();
-
RegInfo = TM.getRegisterInfo();
TD = getDataLayout();
+ resetOperationActions();
+}
+
+void X86TargetLowering::resetOperationActions() {
+ const TargetMachine &TM = getTargetMachine();
+ static bool FirstTimeThrough = true;
+
+ // If none of the target options have changed, then we don't need to reset the
+ // operation actions.
+ if (!FirstTimeThrough && TO == TM.Options) return;
+
+ if (!FirstTimeThrough) {
+ // Reinitialize the actions.
+ initActions();
+ FirstTimeThrough = false;
+ }
+
+ TO = TM.Options;
+
// Set up the TargetLowering object.
static const MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
setSchedulingPreference(Sched::RegPressure);
setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
- // Bypass i32 with i8 on Atom when compiling with O2
- if (Subtarget->hasSlowDivide() && TM.getOptLevel() >= CodeGenOpt::Default)
+ // Bypass expensive divides on Atom when compiling with O2
+ if (Subtarget->hasSlowDivide() && TM.getOptLevel() >= CodeGenOpt::Default) {
addBypassSlowDiv(32, 8);
+ if (Subtarget->is64Bit())
+ addBypassSlowDiv(64, 16);
+ }
if (Subtarget->isTargetWindows() && !Subtarget->isTargetCygMing()) {
// Setup Windows compiler runtime calls.
setOperationAction(ISD::BR_JT , MVT::Other, Expand);
setOperationAction(ISD::BRCOND , MVT::Other, Custom);
- setOperationAction(ISD::BR_CC , MVT::Other, Expand);
+ setOperationAction(ISD::BR_CC , MVT::f32, Expand);
+ setOperationAction(ISD::BR_CC , MVT::f64, Expand);
+ setOperationAction(ISD::BR_CC , MVT::f80, Expand);
+ setOperationAction(ISD::BR_CC , MVT::i8, Expand);
+ setOperationAction(ISD::BR_CC , MVT::i16, Expand);
+ setOperationAction(ISD::BR_CC , MVT::i32, Expand);
+ setOperationAction(ISD::BR_CC , MVT::i64, Expand);
setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
if (Subtarget->is64Bit())
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
setOperationAction(ISD::SETCC , MVT::i64 , Custom);
}
setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
- // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intened to support
+ // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
// SjLj exception handling but a light-weight setjmp/longjmp replacement to
// support continuation, user-level threading, and etc.. As a result, no
// other SjLj exception interfaces are implemented and please don't build
setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
// We don't support sin/cos/fmod
- setOperationAction(ISD::FSIN , MVT::f64, Expand);
- setOperationAction(ISD::FCOS , MVT::f64, Expand);
- setOperationAction(ISD::FSIN , MVT::f32, Expand);
- setOperationAction(ISD::FCOS , MVT::f32, Expand);
+ setOperationAction(ISD::FSIN , MVT::f64, Expand);
+ setOperationAction(ISD::FCOS , MVT::f64, Expand);
+ setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
+ setOperationAction(ISD::FSIN , MVT::f32, Expand);
+ setOperationAction(ISD::FCOS , MVT::f32, Expand);
+ setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
// Expand FP immediates into loads from the stack, except for the special
// cases we handle.
setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
// We don't support sin/cos/fmod
- setOperationAction(ISD::FSIN , MVT::f32, Expand);
- setOperationAction(ISD::FCOS , MVT::f32, Expand);
+ setOperationAction(ISD::FSIN , MVT::f32, Expand);
+ setOperationAction(ISD::FCOS , MVT::f32, Expand);
+ setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
// Special cases we handle for FP constants.
addLegalFPImmediate(APFloat(+0.0f)); // xorps
addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
if (!TM.Options.UnsafeFPMath) {
- setOperationAction(ISD::FSIN , MVT::f64 , Expand);
- setOperationAction(ISD::FCOS , MVT::f64 , Expand);
+ setOperationAction(ISD::FSIN , MVT::f64, Expand);
+ setOperationAction(ISD::FCOS , MVT::f64, Expand);
+ setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
}
} else if (!TM.Options.UseSoftFloat) {
// f32 and f64 in x87.
setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
if (!TM.Options.UnsafeFPMath) {
- setOperationAction(ISD::FSIN , MVT::f32 , Expand);
- setOperationAction(ISD::FSIN , MVT::f64 , Expand);
- setOperationAction(ISD::FCOS , MVT::f32 , Expand);
- setOperationAction(ISD::FCOS , MVT::f64 , Expand);
+ setOperationAction(ISD::FSIN , MVT::f64, Expand);
+ setOperationAction(ISD::FSIN , MVT::f32, Expand);
+ setOperationAction(ISD::FCOS , MVT::f64, Expand);
+ setOperationAction(ISD::FCOS , MVT::f32, Expand);
+ setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
+ setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
}
addLegalFPImmediate(APFloat(+0.0)); // FLD0
addLegalFPImmediate(APFloat(+1.0)); // FLD1
}
if (!TM.Options.UnsafeFPMath) {
- setOperationAction(ISD::FSIN , MVT::f80 , Expand);
- setOperationAction(ISD::FCOS , MVT::f80 , Expand);
+ setOperationAction(ISD::FSIN , MVT::f80, Expand);
+ setOperationAction(ISD::FCOS , MVT::f80, Expand);
+ setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
}
setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
setOperationAction(ISD::FABS, VT, Expand);
setOperationAction(ISD::FSIN, VT, Expand);
+ setOperationAction(ISD::FSINCOS, VT, Expand);
setOperationAction(ISD::FCOS, VT, Expand);
+ setOperationAction(ISD::FSINCOS, VT, Expand);
setOperationAction(ISD::FREM, VT, Expand);
setOperationAction(ISD::FMA, VT, Expand);
setOperationAction(ISD::FPOWI, VT, Expand);
setOperationAction(ISD::SRA, MVT::v8i16, Custom);
setOperationAction(ISD::SRA, MVT::v16i8, Custom);
- if (Subtarget->hasInt256()) {
- setOperationAction(ISD::SRL, MVT::v2i64, Legal);
- setOperationAction(ISD::SRL, MVT::v4i32, Legal);
-
- setOperationAction(ISD::SHL, MVT::v2i64, Legal);
- setOperationAction(ISD::SHL, MVT::v4i32, Legal);
+ // In the customized shift lowering, the legal cases in AVX2 will be
+ // recognized.
+ setOperationAction(ISD::SRL, MVT::v2i64, Custom);
+ setOperationAction(ISD::SRL, MVT::v4i32, Custom);
- setOperationAction(ISD::SRA, MVT::v4i32, Legal);
- } else {
- setOperationAction(ISD::SRL, MVT::v2i64, Custom);
- setOperationAction(ISD::SRL, MVT::v4i32, Custom);
+ setOperationAction(ISD::SHL, MVT::v2i64, Custom);
+ setOperationAction(ISD::SHL, MVT::v4i32, Custom);
- setOperationAction(ISD::SHL, MVT::v2i64, Custom);
- setOperationAction(ISD::SHL, MVT::v4i32, Custom);
+ setOperationAction(ISD::SRA, MVT::v4i32, Custom);
- setOperationAction(ISD::SRA, MVT::v4i32, Custom);
- }
setOperationAction(ISD::SDIV, MVT::v8i16, Custom);
setOperationAction(ISD::SDIV, MVT::v4i32, Custom);
}
setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Custom);
setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
+ setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
- setOperationAction(ISD::SRL, MVT::v4i64, Legal);
- setOperationAction(ISD::SRL, MVT::v8i32, Legal);
-
- setOperationAction(ISD::SHL, MVT::v4i64, Legal);
- setOperationAction(ISD::SHL, MVT::v8i32, Legal);
-
- setOperationAction(ISD::SRA, MVT::v8i32, Legal);
-
setOperationAction(ISD::SDIV, MVT::v8i32, Custom);
} else {
setOperationAction(ISD::ADD, MVT::v4i64, Custom);
setOperationAction(ISD::MUL, MVT::v8i32, Custom);
setOperationAction(ISD::MUL, MVT::v16i16, Custom);
// Don't lower v32i8 because there is no 128-bit byte mul
+ }
- setOperationAction(ISD::SRL, MVT::v4i64, Custom);
- setOperationAction(ISD::SRL, MVT::v8i32, Custom);
+ // In the customized shift lowering, the legal cases in AVX2 will be
+ // recognized.
+ setOperationAction(ISD::SRL, MVT::v4i64, Custom);
+ setOperationAction(ISD::SRL, MVT::v8i32, Custom);
- setOperationAction(ISD::SHL, MVT::v4i64, Custom);
- setOperationAction(ISD::SHL, MVT::v8i32, Custom);
+ setOperationAction(ISD::SHL, MVT::v4i64, Custom);
+ setOperationAction(ISD::SHL, MVT::v8i32, Custom);
- setOperationAction(ISD::SRA, MVT::v8i32, Custom);
- }
+ setOperationAction(ISD::SRA, MVT::v8i32, Custom);
// Custom lower several nodes for 256-bit types.
for (int i = MVT::FIRST_VECTOR_VALUETYPE;
setLibcallName(RTLIB::SRA_I128, 0);
}
+ // Combine sin / cos into one node or libcall if possible.
+ if (Subtarget->hasSinCos()) {
+ setLibcallName(RTLIB::SINCOS_F32, "sincosf");
+ setLibcallName(RTLIB::SINCOS_F64, "sincos");
+ if (Subtarget->isTargetDarwin()) {
+ // For MacOSX, we don't want to the normal expansion of a libcall to
+ // sincos. We want to issue a libcall to __sincos_stret to avoid memory
+ // traffic.
+ setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
+ setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
+ }
+ }
+
// We have target-specific dag combine patterns for the following nodes:
setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
setTargetDAGCombine(ISD::ZERO_EXTEND);
setTargetDAGCombine(ISD::ANY_EXTEND);
setTargetDAGCombine(ISD::SIGN_EXTEND);
+ setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
setTargetDAGCombine(ISD::TRUNCATE);
setTargetDAGCombine(ISD::SINT_TO_FP);
setTargetDAGCombine(ISD::SETCC);
// 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
- maxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
- maxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
- maxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
+ MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
+ MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
+ MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
+ MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
+ MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
+ MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
setPrefLoopAlignment(4); // 2^4 bytes.
- benefitFromCodePlacementOpt = true;
// Predictable cmov don't hurt on atom because it's in-order.
- predictableSelectIsExpensive = !Subtarget->isAtom();
+ PredictableSelectIsExpensive = !Subtarget->isAtom();
setPrefFunctionAlignment(4); // 2^4 bytes.
}
RVLocs, *DAG.getContext());
CCInfo.AnalyzeReturn(Outs, RetCC_X86);
- // Add the regs to the liveout set for the function.
- MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
- for (unsigned i = 0; i != RVLocs.size(); ++i)
- if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
- MRI.addLiveOut(RVLocs[i].getLocReg());
-
SDValue Flag;
-
SmallVector<SDValue, 6> RetOps;
RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
// Operand #1 = Bytes To Pop
Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
Flag = Chain.getValue(1);
+ RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
}
- // The x86-64 ABI for returning structs by value requires that we copy
- // the sret argument into %rax for the return. We saved the argument into
- // a virtual register in the entry block, so now we copy the value out
- // and into %rax.
- if (Subtarget->is64Bit() &&
- DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
+ // The x86-64 ABIs require that for returning structs by value we copy
+ // the sret argument into %rax/%eax (depending on ABI) for the return.
+ // Win32 requires us to put the sret argument to %eax as well.
+ // We saved the argument into a virtual register in the entry block,
+ // so now we copy the value out and into %rax/%eax.
+ if (DAG.getMachineFunction().getFunction()->hasStructRetAttr() &&
+ (Subtarget->is64Bit() || Subtarget->isTargetWindows())) {
MachineFunction &MF = DAG.getMachineFunction();
X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
unsigned Reg = FuncInfo->getSRetReturnReg();
"SRetReturnReg should have been set in LowerFormalArguments().");
SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
- Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
+ unsigned RetValReg
+ = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
+ X86::RAX : X86::EAX;
+ Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
Flag = Chain.getValue(1);
- // RAX now acts like a return value.
- MRI.addLiveOut(X86::RAX);
+ // RAX/EAX now acts like a return value.
+ RetOps.push_back(DAG.getRegister(RetValReg, getPointerTy()));
}
RetOps[0] = Chain; // Update chain.
InVals.push_back(ArgValue);
}
- // The x86-64 ABI for returning structs by value requires that we copy
- // the sret argument into %rax for the return. Save the argument into
- // a virtual register so that we can access it from the return points.
- if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
+ // The x86-64 ABIs require that for returning structs by value we copy
+ // the sret argument into %rax/%eax (depending on ABI) for the return.
+ // Win32 requires us to put the sret argument to %eax as well.
+ // Save the argument into a virtual register so that we can access it
+ // from the return points.
+ if (MF.getFunction()->hasStructRetAttr() &&
+ (Subtarget->is64Bit() || Subtarget->isTargetWindows())) {
X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
unsigned Reg = FuncInfo->getSRetReturnReg();
if (!Reg) {
- Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
+ MVT PtrTy = getPointerTy();
+ Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
FuncInfo->setSRetReturnReg(Reg);
}
SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
// This isn't right, although it's probably harmless on x86; liveouts
// should be computed from returns not tail calls. Consider a void
// function making a tail call to a function returning int.
- return DAG.getNode(X86ISD::TC_RETURN, dl,
- NodeTys, &Ops[0], Ops.size());
+ return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
}
Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
const SmallVectorImpl<ISD::OutputArg> &Outs,
const SmallVectorImpl<SDValue> &OutVals,
const SmallVectorImpl<ISD::InputArg> &Ins,
- SelectionDAG& DAG) const {
+ SelectionDAG &DAG) const {
if (!IsTailCallConvention(CalleeCC) &&
CalleeCC != CallingConv::C)
return false;
// An stdcall caller is expected to clean up its arguments; the callee
// isn't going to do that.
- if (!CCMatch && CallerCC==CallingConv::X86_StdCall)
+ if (!CCMatch && CallerCC == CallingConv::X86_StdCall)
return false;
// Do not sibcall optimize vararg calls unless all arguments are passed via
// callee-saved registers are restored. These happen to be the same
// registers used to pass 'inreg' arguments so watch out for those.
if (!Subtarget->is64Bit() &&
- !isa<GlobalAddressSDNode>(Callee) &&
- !isa<ExternalSymbolSDNode>(Callee)) {
+ ((!isa<GlobalAddressSDNode>(Callee) &&
+ !isa<ExternalSymbolSDNode>(Callee)) ||
+ getTargetMachine().getRelocationModel() == Reloc::PIC_)) {
unsigned NumInRegs = 0;
+ // In PIC we need an extra register to formulate the address computation
+ // for the callee.
+ unsigned MaxInRegs =
+ (getTargetMachine().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
+
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
CCValAssign &VA = ArgLocs[i];
if (!VA.isRegLoc())
switch (Reg) {
default: break;
case X86::EAX: case X86::EDX: case X86::ECX:
- if (++NumInRegs == 3)
+ if (++NumInRegs == MaxInRegs)
return false;
break;
}
case X86ISD::PSHUFHW:
case X86ISD::PSHUFLW:
case X86ISD::SHUFP:
- case X86ISD::PALIGN:
+ case X86ISD::PALIGNR:
case X86ISD::MOVLHPS:
case X86ISD::MOVLHPD:
case X86ISD::MOVHLPS:
SelectionDAG &DAG) {
switch(Opc) {
default: llvm_unreachable("Unknown x86 shuffle node");
- case X86ISD::PALIGN:
+ case X86ISD::PALIGNR:
case X86ISD::SHUFP:
case X86ISD::VPERM2X128:
return DAG.getNode(Opc, dl, VT, V1, V2,
static
SDValue Compact8x32ShuffleNode(ShuffleVectorSDNode *SVOp,
SelectionDAG &DAG) {
- EVT VT = SVOp->getValueType(0);
+ MVT VT = SVOp->getValueType(0).getSimpleVT();
DebugLoc dl = SVOp->getDebugLoc();
if (VT != MVT::v8i32 && VT != MVT::v8f32)
/// isZeroNode - Returns true if Elt is a constant zero or a floating point
/// constant +0.0.
bool X86::isZeroNode(SDValue Elt) {
- return ((isa<ConstantSDNode>(Elt) &&
- cast<ConstantSDNode>(Elt)->isNullValue()) ||
- (isa<ConstantFPSDNode>(Elt) &&
- cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
+ if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Elt))
+ return CN->isNullValue();
+ if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
+ return CFP->getValueAPF().isPosZero();
+ return false;
}
/// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
/// their permute mask.
static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
SelectionDAG &DAG) {
- EVT VT = SVOp->getValueType(0);
+ MVT VT = SVOp->getValueType(0).getSimpleVT();
unsigned NumElems = VT.getVectorNumElements();
SmallVector<int, 8> MaskVec;
/// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
/// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
/// Then bitcast to their original type, ensuring they get CSE'd.
-static SDValue getOnesVector(EVT VT, bool HasInt256, SelectionDAG &DAG,
+static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
DebugLoc dl) {
assert(VT.isVector() && "Expected a vector type");
case X86ISD::MOVLHPS:
DecodeMOVLHPSMask(NumElems, Mask);
break;
+ case X86ISD::PALIGNR:
+ ImmN = N->getOperand(N->getNumOperands()-1);
+ DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
+ break;
case X86ISD::PSHUFD:
case X86ISD::VPERMILP:
ImmN = N->getOperand(N->getNumOperands()-1);
case X86ISD::MOVLPS:
case X86ISD::MOVSHDUP:
case X86ISD::MOVSLDUP:
- case X86ISD::PALIGN:
// Not yet implemented
return false;
default: llvm_unreachable("unknown target shuffle node");
return DAG.getNode(ISD::BITCAST, dl, VT,
DAG.getNode(Opc, dl, ShVT, SrcOp,
DAG.getConstant(NumBits,
- TLI.getShiftAmountTy(SrcOp.getValueType()))));
+ TLI.getScalarShiftAmountTy(SrcOp.getValueType()))));
}
SDValue
if (!Subtarget->hasFp256())
return SDValue();
- EVT VT = Op.getValueType();
+ MVT VT = Op.getValueType().getSimpleVT();
DebugLoc dl = Op.getDebugLoc();
assert((VT.is128BitVector() || VT.is256BitVector()) &&
X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
DebugLoc dl = Op.getDebugLoc();
- EVT VT = Op.getValueType();
- EVT ExtVT = VT.getVectorElementType();
+ MVT VT = Op.getValueType().getSimpleVT();
+ MVT ExtVT = VT.getVectorElementType();
unsigned NumElems = Op.getNumOperands();
// Vectors containing all zeros can be matched by pxor and xorps later
// Vectors containing all ones can be matched by pcmpeqd on 128-bit width
// vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
// vpcmpeqd on 256-bit vectors.
- if (ISD::isBuildVectorAllOnes(Op.getNode())) {
+ if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
return Op;
// to create 256-bit vectors from two other 128-bit ones.
static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
DebugLoc dl = Op.getDebugLoc();
- EVT ResVT = Op.getValueType();
+ MVT ResVT = Op.getValueType().getSimpleVT();
assert(ResVT.is256BitVector() && "Value type must be 256-bit wide");
SDValue V1 = SVOp->getOperand(0);
SDValue V2 = SVOp->getOperand(1);
DebugLoc dl = SVOp->getDebugLoc();
- EVT VT = SVOp->getValueType(0);
- EVT EltVT = VT.getVectorElementType();
+ MVT VT = SVOp->getValueType(0).getSimpleVT();
+ MVT EltVT = VT.getVectorElementType();
unsigned NumElems = VT.getVectorNumElements();
if (!Subtarget->hasSSE41() || EltVT == MVT::i8)
// Check the mask for BLEND and build the value.
unsigned MaskValue = 0;
// There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
- unsigned NumLanes = (NumElems-1)/8 + 1;
+ unsigned NumLanes = (NumElems-1)/8 + 1;
unsigned NumElemsInLane = NumElems / NumLanes;
// Blend for v16i16 should be symetric for the both lanes.
for (unsigned i = 0; i < NumElemsInLane; ++i) {
- int SndLaneEltIdx = (NumLanes == 2) ?
+ int SndLaneEltIdx = (NumLanes == 2) ?
SVOp->getMaskElt(i + NumElemsInLane) : -1;
int EltIdx = SVOp->getMaskElt(i);
- if ((EltIdx == -1 || EltIdx == (int)i) &&
- (SndLaneEltIdx == -1 || SndLaneEltIdx == (int)(i + NumElemsInLane)))
+ if ((EltIdx < 0 || EltIdx == (int)i) &&
+ (SndLaneEltIdx < 0 || SndLaneEltIdx == (int)(i + NumElemsInLane)))
continue;
- if (((unsigned)EltIdx == (i + NumElems)) &&
- (SndLaneEltIdx == -1 ||
+ if (((unsigned)EltIdx == (i + NumElems)) &&
+ (SndLaneEltIdx < 0 ||
(unsigned)SndLaneEltIdx == i + NumElems + NumElemsInLane))
MaskValue |= (1<<i);
- else
+ else
return SDValue();
}
// Convert i32 vectors to floating point if it is not AVX2.
// AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors.
- EVT BlendVT = VT;
+ MVT BlendVT = VT;
if (EltVT == MVT::i64 || (EltVT == MVT::i32 && !Subtarget->hasInt256())) {
- BlendVT = EVT::getVectorVT(*DAG.getContext(),
- EVT::getFloatingPointVT(EltVT.getSizeInBits()),
- NumElems);
+ BlendVT = MVT::getVectorVT(MVT::getFloatingPointVT(EltVT.getSizeInBits()),
+ NumElems);
V1 = DAG.getNode(ISD::BITCAST, dl, VT, V1);
V2 = DAG.getNode(ISD::BITCAST, dl, VT, V2);
}
-
- SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, V1, V2,
- DAG.getConstant(MaskValue, MVT::i32));
+
+ SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, V1, V2,
+ DAG.getConstant(MaskValue, MVT::i32));
return DAG.getNode(ISD::BITCAST, dl, VT, Ret);
}
}
}
+ // Promote splats to a larger type which usually leads to more efficient code.
+ // FIXME: Is this true if pshufb is available?
+ if (SVOp->isSplat())
+ return PromoteSplat(SVOp, DAG);
+
// If we have SSSE3, and all words of the result are from 1 input vector,
// case 2 is generated, otherwise case 3 is generated. If no SSSE3
// is present, fall back to case 4.
DebugLoc dl = SVOp->getDebugLoc();
ArrayRef<int> MaskVals = SVOp->getMask();
+ // Promote splats to a larger type which usually leads to more efficient code.
+ // FIXME: Is this true if pshufb is available?
+ if (SVOp->isSplat())
+ return PromoteSplat(SVOp, DAG);
+
// If we have SSSE3, case 1 is generated when all result bytes come from
// one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
// present, fall back to case 3.
SDValue LowerVECTOR_SHUFFLEv32i8(ShuffleVectorSDNode *SVOp,
const X86Subtarget *Subtarget,
SelectionDAG &DAG) {
- EVT VT = SVOp->getValueType(0);
+ MVT VT = SVOp->getValueType(0).getSimpleVT();
SDValue V1 = SVOp->getOperand(0);
SDValue V2 = SVOp->getOperand(1);
DebugLoc dl = SVOp->getDebugLoc();
/// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
static
SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
- SelectionDAG &DAG, DebugLoc dl) {
+ SelectionDAG &DAG) {
MVT VT = SVOp->getValueType(0).getSimpleVT();
+ DebugLoc dl = SVOp->getDebugLoc();
unsigned NumElems = VT.getVectorNumElements();
MVT NewVT;
unsigned Scale;
/// getVZextMovL - Return a zero-extending vector move low node.
///
-static SDValue getVZextMovL(EVT VT, EVT OpVT,
+static SDValue getVZextMovL(MVT VT, EVT OpVT,
SDValue SrcOp, SelectionDAG &DAG,
const X86Subtarget *Subtarget, DebugLoc dl) {
if (VT == MVT::v2f64 || VT == MVT::v4f32) {
if (NewOp.getNode())
return NewOp;
- EVT VT = SVOp->getValueType(0);
+ MVT VT = SVOp->getValueType(0).getSimpleVT();
unsigned NumElems = VT.getVectorNumElements();
unsigned NumLaneElems = NumElems / 2;
DebugLoc dl = SVOp->getDebugLoc();
- MVT EltVT = VT.getVectorElementType().getSimpleVT();
- EVT NVT = MVT::getVectorVT(EltVT, NumLaneElems);
+ MVT EltVT = VT.getVectorElementType();
+ MVT NVT = MVT::getVectorVT(EltVT, NumLaneElems);
SDValue Output[2];
SmallVector<int, 16> Mask;
SDValue V1 = SVOp->getOperand(0);
SDValue V2 = SVOp->getOperand(1);
DebugLoc dl = SVOp->getDebugLoc();
- EVT VT = SVOp->getValueType(0);
+ MVT VT = SVOp->getValueType(0).getSimpleVT();
assert(VT.is128BitVector() && "Unsupported vector size");
// Reduce a vector shuffle to zext.
SDValue
-X86TargetLowering::lowerVectorIntExtend(SDValue Op, SelectionDAG &DAG) const {
+X86TargetLowering::LowerVectorIntExtend(SDValue Op, SelectionDAG &DAG) const {
// PMOVZX is only available from SSE41.
if (!Subtarget->hasSSE41())
return SDValue();
return SDValue();
}
+ LLVMContext *Context = DAG.getContext();
unsigned NBits = VT.getVectorElementType().getSizeInBits() << Shift;
- EVT NeVT = EVT::getIntegerVT(*DAG.getContext(), NBits);
- EVT NVT = EVT::getVectorVT(*DAG.getContext(), NeVT, NumElems >> Shift);
+ EVT NeVT = EVT::getIntegerVT(*Context, NBits);
+ EVT NVT = EVT::getVectorVT(*Context, NeVT, NumElems >> Shift);
if (!isTypeLegal(NVT))
return SDValue();
// If it's foldable, i.e. normal load with single use, we will let code
// selection to fold it. Otherwise, we will short the conversion sequence.
if (CIdx && CIdx->getZExtValue() == 0 &&
- (!ISD::isNormalLoad(V.getNode()) || !V.hasOneUse()))
+ (!ISD::isNormalLoad(V.getNode()) || !V.hasOneUse())) {
+ if (V.getValueSizeInBits() > V1.getValueSizeInBits()) {
+ // The "ext_vec_elt" node is wider than the result node.
+ // In this case we should extract subvector from V.
+ // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast (extract_subvector x)).
+ unsigned Ratio = V.getValueSizeInBits() / V1.getValueSizeInBits();
+ EVT FullVT = V.getValueType();
+ EVT SubVecVT = EVT::getVectorVT(*Context,
+ FullVT.getVectorElementType(),
+ FullVT.getVectorNumElements()/Ratio);
+ V = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVecVT, V,
+ DAG.getIntPtrConstant(0));
+ }
V1 = DAG.getNode(ISD::BITCAST, DL, V1.getValueType(), V);
+ }
}
return DAG.getNode(ISD::BITCAST, DL, VT,
SDValue
X86TargetLowering::NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG) const {
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- EVT VT = Op.getValueType();
+ MVT VT = Op.getValueType().getSimpleVT();
DebugLoc dl = Op.getDebugLoc();
SDValue V1 = Op.getOperand(0);
SDValue V2 = Op.getOperand(1);
// Handle splat operations
if (SVOp->isSplat()) {
- unsigned NumElem = VT.getVectorNumElements();
-
// Use vbroadcast whenever the splat comes from a foldable load
SDValue Broadcast = LowerVectorBroadcast(Op, DAG);
if (Broadcast.getNode())
return Broadcast;
-
- // Handle splats by matching through known shuffle masks
- if ((VT.is128BitVector() && NumElem <= 4) ||
- (VT.is256BitVector() && NumElem <= 8))
- return SDValue();
-
- // All remaning splats are promoted to target supported vector shuffles.
- return PromoteSplat(SVOp, DAG);
}
// Check integer expanding shuffles.
- SDValue NewOp = lowerVectorIntExtend(Op, DAG);
+ SDValue NewOp = LowerVectorIntExtend(Op, DAG);
if (NewOp.getNode())
return NewOp;
// do it!
if (VT == MVT::v8i16 || VT == MVT::v16i8 ||
VT == MVT::v16i16 || VT == MVT::v32i8) {
- SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
+ SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
if (NewOp.getNode())
return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
} else if ((VT == MVT::v4i32 ||
// FIXME: Figure out a cleaner way to do this.
// Try to make use of movq to zero out the top part.
if (ISD::isBuildVectorAllZeros(V2.getNode())) {
- SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
+ SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
if (NewOp.getNode()) {
- EVT NewVT = NewOp.getValueType();
+ MVT NewVT = NewOp.getValueType().getSimpleVT();
if (isCommutedMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(),
NewVT, true, false))
return getVZextMovL(VT, NewVT, NewOp.getOperand(0),
DAG, Subtarget, dl);
}
} else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
- SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
+ SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
if (NewOp.getNode()) {
- EVT NewVT = NewOp.getValueType();
+ MVT NewVT = NewOp.getValueType().getSimpleVT();
if (isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(), NewVT))
return getVZextMovL(VT, NewVT, NewOp.getOperand(1),
DAG, Subtarget, dl);
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
SDValue V1 = Op.getOperand(0);
SDValue V2 = Op.getOperand(1);
- EVT VT = Op.getValueType();
+ MVT VT = Op.getValueType().getSimpleVT();
DebugLoc dl = Op.getDebugLoc();
unsigned NumElems = VT.getVectorNumElements();
bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
if (isShift && ShVal.hasOneUse()) {
// If the shifted value has multiple uses, it may be cheaper to use
// v_set0 + movlhps or movhlps, etc.
- EVT EltVT = VT.getVectorElementType();
+ MVT EltVT = VT.getVectorElementType();
ShAmt *= EltVT.getSizeInBits();
return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
}
if (isShift) {
// No better options. Use a vshldq / vsrldq.
- EVT EltVT = VT.getVectorElementType();
+ MVT EltVT = VT.getVectorElementType();
ShAmt *= EltVT.getSizeInBits();
return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
}
// nodes, and remove one by one until they don't return Op anymore.
if (isPALIGNRMask(M, VT, Subtarget))
- return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
+ return getTargetShuffleNode(X86ISD::PALIGNR, dl, VT, V1, V2,
getShufflePALIGNRImmediate(SVOp),
DAG);
return SDValue();
}
-SDValue
-X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
- SelectionDAG &DAG) const {
- EVT VT = Op.getValueType();
+static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
+ MVT VT = Op.getValueType().getSimpleVT();
DebugLoc dl = Op.getDebugLoc();
- if (!Op.getOperand(0).getValueType().is128BitVector())
+ if (!Op.getOperand(0).getValueType().getSimpleVT().is128BitVector())
return SDValue();
if (VT.getSizeInBits() == 8) {
return SDValue();
SDValue Vec = Op.getOperand(0);
- EVT VecVT = Vec.getValueType();
+ MVT VecVT = Vec.getValueType().getSimpleVT();
// If this is a 256-bit vector result, first extract the 128-bit vector and
// then extract the element from the 128-bit vector.
return Res;
}
- EVT VT = Op.getValueType();
+ MVT VT = Op.getValueType().getSimpleVT();
DebugLoc dl = Op.getDebugLoc();
// TODO: handle v16i8.
if (VT.getSizeInBits() == 16) {
MVT::v4i32, Vec),
Op.getOperand(1)));
// Transform it so it match pextrw which produces a 32-bit result.
- EVT EltVT = MVT::i32;
+ MVT EltVT = MVT::i32;
SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
Op.getOperand(0), Op.getOperand(1));
SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
// SHUFPS the element to the lowest double word, then movss.
int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
- EVT VVT = Op.getOperand(0).getValueType();
+ MVT VVT = Op.getOperand(0).getValueType().getSimpleVT();
SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
DAG.getUNDEF(VVT), Mask);
return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
// Note if the lower 64 bits of the result of the UNPCKHPD is then stored
// to a f64mem, the whole operation is folded into a single MOVHPDmr.
int Mask[2] = { 1, -1 };
- EVT VVT = Op.getOperand(0).getValueType();
+ MVT VVT = Op.getOperand(0).getValueType().getSimpleVT();
SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
DAG.getUNDEF(VVT), Mask);
return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
return SDValue();
}
-SDValue
-X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
- SelectionDAG &DAG) const {
- EVT VT = Op.getValueType();
- EVT EltVT = VT.getVectorElementType();
+static SDValue LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
+ MVT VT = Op.getValueType().getSimpleVT();
+ MVT EltVT = VT.getVectorElementType();
DebugLoc dl = Op.getDebugLoc();
SDValue N0 = Op.getOperand(0);
SDValue
X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
- EVT VT = Op.getValueType();
- EVT EltVT = VT.getVectorElementType();
+ MVT VT = Op.getValueType().getSimpleVT();
+ MVT EltVT = VT.getVectorElementType();
DebugLoc dl = Op.getDebugLoc();
SDValue N0 = Op.getOperand(0);
static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
LLVMContext *Context = DAG.getContext();
DebugLoc dl = Op.getDebugLoc();
- EVT OpVT = Op.getValueType();
+ MVT OpVT = Op.getValueType().getSimpleVT();
// If this is a 256-bit vector result, first insert into a 128-bit
// vector and then insert into the 256-bit vector.
SDValue
X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
- int64_t Offset,
- SelectionDAG &DAG) const {
+ int64_t Offset, SelectionDAG &DAG) const {
// Create the TargetGlobalAddress node, folding in the constant
// offset if it is legal.
unsigned char OpFlags =
case TLSModel::LocalExec:
return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
Subtarget->is64Bit(),
- getTargetMachine().getRelocationModel() == Reloc::PIC_);
+ getTargetMachine().getRelocationModel() == Reloc::PIC_);
}
llvm_unreachable("Unknown TLS model.");
}
Chain.getValue(1));
}
- if (Subtarget->isTargetWindows()) {
+ if (Subtarget->isTargetWindows() || Subtarget->isTargetMingw()) {
// Just use the implicit TLS architecture
// Need to generate someting similar to:
// mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
SDValue Chain = DAG.getEntryNode();
// Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
- // %gs:0x58 (64-bit).
+ // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
+ // use its literal value of 0x2C.
Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
? Type::getInt8PtrTy(*DAG.getContext(),
256)
: Type::getInt32PtrTy(*DAG.getContext(),
257));
- SDValue ThreadPointer = DAG.getLoad(getPointerTy(), dl, Chain,
- Subtarget->is64Bit()
- ? DAG.getIntPtrConstant(0x58)
- : DAG.getExternalSymbol("_tls_array",
- getPointerTy()),
+ SDValue TlsArray = Subtarget->is64Bit() ? DAG.getIntPtrConstant(0x58) :
+ (Subtarget->isTargetMingw() ? DAG.getIntPtrConstant(0x2C) :
+ DAG.getExternalSymbol("_tls_array", getPointerTy()));
+
+ SDValue ThreadPointer = DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
MachinePointerInfo(Ptr),
false, false, false, 0);
SmallVector<Constant*,2> CV1;
CV1.push_back(
- ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
+ ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
+ APInt(64, 0x4330000000000000ULL))));
CV1.push_back(
- ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
+ ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
+ APInt(64, 0x4530000000000000ULL))));
Constant *C1 = ConstantVector::get(CV1);
SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
SVT == MVT::v8i8 || SVT == MVT::v8i16) &&
"Custom UINT_TO_FP is not supported!");
- EVT NVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, SVT.getVectorNumElements());
+ EVT NVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
+ SVT.getVectorNumElements());
return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
}
return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
}
-std::pair<SDValue,SDValue> X86TargetLowering::
-FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned, bool IsReplace) const {
+std::pair<SDValue,SDValue>
+X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
+ bool IsSigned, bool IsReplace) const {
DebugLoc DL = Op.getDebugLoc();
EVT DstTy = Op.getValueType();
static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
const X86Subtarget *Subtarget) {
- EVT VT = Op->getValueType(0);
+ MVT VT = Op->getValueType(0).getSimpleVT();
SDValue In = Op->getOperand(0);
- EVT InVT = In.getValueType();
+ MVT InVT = In.getValueType().getSimpleVT();
DebugLoc dl = Op->getDebugLoc();
// Optimize vectors in AVX mode:
SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
- EVT HVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
+ MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
VT.getVectorNumElements()/2);
OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
SDValue X86TargetLowering::LowerZERO_EXTEND(SDValue Op,
SelectionDAG &DAG) const {
DebugLoc DL = Op.getDebugLoc();
- EVT VT = Op.getValueType();
+ MVT VT = Op.getValueType().getSimpleVT();
SDValue In = Op.getOperand(0);
- EVT SVT = In.getValueType();
+ MVT SVT = In.getValueType().getSimpleVT();
if (Subtarget->hasFp256()) {
SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
return DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v8i32, Lo, Hi);
}
-SDValue X86TargetLowering::lowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
+SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
DebugLoc DL = Op.getDebugLoc();
- EVT VT = Op.getValueType();
+ MVT VT = Op.getValueType().getSimpleVT();
SDValue In = Op.getOperand(0);
- EVT SVT = In.getValueType();
+ MVT SVT = In.getValueType().getSimpleVT();
if ((VT == MVT::v4i32) && (SVT == MVT::v4i64)) {
// On AVX2, v4i64 -> v4i32 becomes VPERMD.
SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
SelectionDAG &DAG) const {
- if (Op.getValueType().isVector()) {
- if (Op.getValueType() == MVT::v8i16)
- return DAG.getNode(ISD::TRUNCATE, Op.getDebugLoc(), Op.getValueType(),
+ MVT VT = Op.getValueType().getSimpleVT();
+ if (VT.isVector()) {
+ if (VT == MVT::v8i16)
+ return DAG.getNode(ISD::TRUNCATE, Op.getDebugLoc(), VT,
DAG.getNode(ISD::FP_TO_SINT, Op.getDebugLoc(),
MVT::v8i32, Op.getOperand(0)));
return SDValue();
return FIST;
}
-SDValue X86TargetLowering::lowerFP_EXTEND(SDValue Op,
- SelectionDAG &DAG) const {
+static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
DebugLoc DL = Op.getDebugLoc();
- EVT VT = Op.getValueType();
+ MVT VT = Op.getValueType().getSimpleVT();
SDValue In = Op.getOperand(0);
- EVT SVT = In.getValueType();
+ MVT SVT = In.getValueType().getSimpleVT();
assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) const {
LLVMContext *Context = DAG.getContext();
DebugLoc dl = Op.getDebugLoc();
- EVT VT = Op.getValueType();
- EVT EltVT = VT;
+ MVT VT = Op.getValueType().getSimpleVT();
+ MVT EltVT = VT;
unsigned NumElts = VT == MVT::f64 ? 2 : 4;
if (VT.isVector()) {
EltVT = VT.getVectorElementType();
}
Constant *C;
if (EltVT == MVT::f64)
- C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
+ C = ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
+ APInt(64, ~(1ULL << 63))));
else
- C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
+ C = ConstantFP::get(*Context, APFloat(APFloat::IEEEsingle,
+ APInt(32, ~(1U << 31))));
C = ConstantVector::getSplat(NumElts, C);
SDValue CPIdx = DAG.getConstantPool(C, getPointerTy());
unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
LLVMContext *Context = DAG.getContext();
DebugLoc dl = Op.getDebugLoc();
- EVT VT = Op.getValueType();
- EVT EltVT = VT;
+ MVT VT = Op.getValueType().getSimpleVT();
+ MVT EltVT = VT;
unsigned NumElts = VT == MVT::f64 ? 2 : 4;
if (VT.isVector()) {
EltVT = VT.getVectorElementType();
}
Constant *C;
if (EltVT == MVT::f64)
- C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
+ C = ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
+ APInt(64, 1ULL << 63)));
else
- C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
+ C = ConstantFP::get(*Context, APFloat(APFloat::IEEEsingle,
+ APInt(32, 1U << 31)));
C = ConstantVector::getSplat(NumElts, C);
SDValue CPIdx = DAG.getConstantPool(C, getPointerTy());
unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
DebugLoc dl = Op.getDebugLoc();
- EVT VT = Op.getValueType();
- EVT SrcVT = Op1.getValueType();
+ MVT VT = Op.getValueType().getSimpleVT();
+ MVT SrcVT = Op1.getValueType().getSimpleVT();
// If second operand is smaller, extend it first.
if (SrcVT.bitsLT(VT)) {
// First get the sign bit of second operand.
SmallVector<Constant*,4> CV;
if (SrcVT == MVT::f64) {
- CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
- CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
+ const fltSemantics &Sem = APFloat::IEEEdouble;
+ CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(64, 1ULL << 63))));
+ CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(64, 0))));
} else {
- CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
- CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
- CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
- CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
+ const fltSemantics &Sem = APFloat::IEEEsingle;
+ CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 1U << 31))));
+ CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
+ CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
+ CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
}
Constant *C = ConstantVector::get(CV);
SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
// Clear first operand sign bit.
CV.clear();
if (VT == MVT::f64) {
- CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
- CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
+ const fltSemantics &Sem = APFloat::IEEEdouble;
+ CV.push_back(ConstantFP::get(*Context, APFloat(Sem,
+ APInt(64, ~(1ULL << 63)))));
+ CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(64, 0))));
} else {
- CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
- CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
- CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
- CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
+ const fltSemantics &Sem = APFloat::IEEEsingle;
+ CV.push_back(ConstantFP::get(*Context, APFloat(Sem,
+ APInt(32, ~(1U << 31)))));
+ CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
+ CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
+ CV.push_back(ConstantFP::get(*Context, APFloat(Sem, APInt(32, 0))));
}
C = ConstantVector::get(CV);
CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
SDValue N0 = Op.getOperand(0);
DebugLoc dl = Op.getDebugLoc();
- EVT VT = Op.getValueType();
+ MVT VT = Op.getValueType().getSimpleVT();
// Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
// LowerVectorAllZeroTest - Check whether an OR'd tree is PTEST-able.
//
-SDValue X86TargetLowering::LowerVectorAllZeroTest(SDValue Op, SelectionDAG &DAG) const {
+SDValue X86TargetLowering::LowerVectorAllZeroTest(SDValue Op,
+ SelectionDAG &DAG) const {
assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
if (!Subtarget->hasSSE41())
return SDValue();
}
-SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
-
- if (Op.getValueType().isVector()) return LowerVSETCC(Op, DAG);
-
- assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
- SDValue Op0 = Op.getOperand(0);
- SDValue Op1 = Op.getOperand(1);
- DebugLoc dl = Op.getDebugLoc();
- ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
-
- // Optimize to BT if possible.
- // 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() &&
- Op1.getOpcode() == ISD::Constant &&
- cast<ConstantSDNode>(Op1)->isNullValue() &&
- (CC == ISD::SETEQ || CC == ISD::SETNE)) {
- SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
- if (NewSetCC.getNode())
- return NewSetCC;
- }
-
- // 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)) {
-
- // 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));
- }
- }
-
- bool isFP = Op1.getValueType().isFloatingPoint();
- unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
- if (X86CC == X86::COND_INVALID)
- return SDValue();
-
- SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
- EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
- return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
- DAG.getConstant(X86CC, MVT::i8), EFLAGS);
-}
-
// Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
// ones, and then concatenate the result back.
static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
- EVT VT = Op.getValueType();
+ MVT VT = Op.getValueType().getSimpleVT();
assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
"Unsupported value type for operation");
SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
// Issue the operation on the smaller types and concatenate the result back
- MVT EltVT = VT.getVectorElementType().getSimpleVT();
- EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
+ MVT EltVT = VT.getVectorElementType();
+ MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
}
-SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
+static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
SDValue Cond;
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
SDValue CC = Op.getOperand(2);
- EVT VT = Op.getValueType();
+ MVT VT = Op.getValueType().getSimpleVT();
ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
- bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
+ bool isFP = Op.getOperand(1).getValueType().getSimpleVT().isFloatingPoint();
DebugLoc dl = Op.getDebugLoc();
if (isFP) {
#ifndef NDEBUG
- EVT EltVT = Op0.getValueType().getVectorElementType();
+ MVT EltVT = Op0.getValueType().getVectorElementType().getSimpleVT();
assert(EltVT == MVT::f32 || EltVT == MVT::f64);
#endif
return Result;
}
+SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
+
+ MVT VT = Op.getValueType().getSimpleVT();
+
+ if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
+
+ assert(VT == MVT::i8 && "SetCC type must be 8-bit integer");
+ SDValue Op0 = Op.getOperand(0);
+ SDValue Op1 = Op.getOperand(1);
+ DebugLoc dl = Op.getDebugLoc();
+ ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
+
+ // Optimize to BT if possible.
+ // 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() &&
+ Op1.getOpcode() == ISD::Constant &&
+ cast<ConstantSDNode>(Op1)->isNullValue() &&
+ (CC == ISD::SETEQ || CC == ISD::SETNE)) {
+ SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
+ if (NewSetCC.getNode())
+ return NewSetCC;
+ }
+
+ // 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)) {
+
+ // 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));
+ }
+ }
+
+ bool isFP = Op1.getValueType().getSimpleVT().isFloatingPoint();
+ unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
+ if (X86CC == X86::COND_INVALID)
+ return SDValue();
+
+ SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
+ EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
+ return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
+ DAG.getConstant(X86CC, MVT::i8), EFLAGS);
+}
+
// isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
static bool isX86LogicalCmp(SDValue Op) {
unsigned Opc = Op.getNode()->getOpcode();
SDValue Cmp = Cond.getOperand(1);
unsigned Opc = Cmp.getOpcode();
- EVT VT = Op.getValueType();
+ MVT VT = Op.getValueType().getSimpleVT();
bool IllegalFPCMov = false;
if (VT.isFloatingPoint() && !VT.isVector() &&
SDValue X86TargetLowering::LowerSIGN_EXTEND(SDValue Op,
SelectionDAG &DAG) const {
- EVT VT = Op->getValueType(0);
+ MVT VT = Op->getValueType(0).getSimpleVT();
SDValue In = Op->getOperand(0);
- EVT InVT = In.getValueType();
+ MVT InVT = In.getValueType().getSimpleVT();
DebugLoc dl = Op->getDebugLoc();
if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
- EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), VT.getScalarType(),
+ MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
VT.getVectorNumElements()/2);
OpLo = DAG.getNode(X86ISD::VSEXT_MOVL, dl, HalfVT, OpLo);
MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
}
-// getTargetVShiftNOde - Handle vector element shifts where the shift amount
+// getTargetVShiftNode - Handle vector element shifts where the shift amount
// may or may not be a constant. Takes immediate version of shift as input.
static SDValue getTargetVShiftNode(unsigned Opc, DebugLoc dl, EVT VT,
SDValue SrcOp, SDValue ShAmt,
switch (IntNo) {
default: return SDValue(); // Don't custom lower most intrinsics.
- // RDRAND intrinsics.
+ // RDRAND/RDSEED intrinsics.
case Intrinsic::x86_rdrand_16:
case Intrinsic::x86_rdrand_32:
- case Intrinsic::x86_rdrand_64: {
+ case Intrinsic::x86_rdrand_64:
+ case Intrinsic::x86_rdseed_16:
+ case Intrinsic::x86_rdseed_32:
+ case Intrinsic::x86_rdseed_64: {
+ unsigned Opcode = (IntNo == Intrinsic::x86_rdseed_16 ||
+ IntNo == Intrinsic::x86_rdseed_32 ||
+ IntNo == Intrinsic::x86_rdseed_64) ? X86ISD::RDSEED :
+ X86ISD::RDRAND;
// Emit the node with the right value type.
SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
- SDValue Result = DAG.getNode(X86ISD::RDRAND, dl, VTs, Op.getOperand(0));
+ SDValue Result = DAG.getNode(Opcode, dl, VTs, Op.getOperand(0));
- // If the value returned by RDRAND was valid (CF=1), return 1. Otherwise
- // return the value from Rand, which is always 0, casted to i32.
+ // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
+ // Otherwise return the value from Rand, which is always 0, casted to i32.
SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
DAG.getConstant(1, Op->getValueType(1)),
DAG.getConstant(X86::COND_B, MVT::i32),
return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
SDValue(Result.getNode(), 2));
}
+
+ // XTEST intrinsics.
+ case Intrinsic::x86_xtest: {
+ SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
+ SDValue InTrans = DAG.getNode(X86ISD::XTEST, dl, VTs, Op.getOperand(0));
+ SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
+ DAG.getConstant(X86::COND_NE, MVT::i8),
+ InTrans);
+ SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
+ return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
+ Ret, SDValue(InTrans.getNode(), 1));
+ }
}
}
return SDValue();
}
-SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
-
+static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
+ const X86Subtarget *Subtarget) {
EVT VT = Op.getValueType();
DebugLoc dl = Op.getDebugLoc();
SDValue R = Op.getOperand(0);
SDValue Amt = Op.getOperand(1);
- LLVMContext *Context = DAG.getContext();
-
- if (!Subtarget->hasSSE2())
- return SDValue();
// Optimize shl/srl/sra with constant shift amount.
if (isSplatVector(Amt.getNode())) {
}
}
+ // Special case in 32-bit mode, where i64 is expanded into high and low parts.
+ if (!Subtarget->is64Bit() &&
+ (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
+ Amt.getOpcode() == ISD::BITCAST &&
+ Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
+ Amt = Amt.getOperand(0);
+ unsigned Ratio = Amt.getValueType().getVectorNumElements() /
+ VT.getVectorNumElements();
+ unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
+ uint64_t ShiftAmt = 0;
+ for (unsigned i = 0; i != Ratio; ++i) {
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
+ if (C == 0)
+ return SDValue();
+ // 6 == Log2(64)
+ ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
+ }
+ // Check remaining shift amounts.
+ for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
+ uint64_t ShAmt = 0;
+ for (unsigned j = 0; j != Ratio; ++j) {
+ ConstantSDNode *C =
+ dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
+ if (C == 0)
+ return SDValue();
+ // 6 == Log2(64)
+ ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
+ }
+ if (ShAmt != ShiftAmt)
+ return SDValue();
+ }
+ switch (Op.getOpcode()) {
+ default:
+ llvm_unreachable("Unknown shift opcode!");
+ case ISD::SHL:
+ return DAG.getNode(X86ISD::VSHLI, dl, VT, R,
+ DAG.getConstant(ShiftAmt, MVT::i32));
+ case ISD::SRL:
+ return DAG.getNode(X86ISD::VSRLI, dl, VT, R,
+ DAG.getConstant(ShiftAmt, MVT::i32));
+ case ISD::SRA:
+ return DAG.getNode(X86ISD::VSRAI, dl, VT, R,
+ DAG.getConstant(ShiftAmt, MVT::i32));
+ }
+ }
+
+ return SDValue();
+}
+
+static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
+ const X86Subtarget* Subtarget) {
+ EVT VT = Op.getValueType();
+ DebugLoc dl = Op.getDebugLoc();
+ SDValue R = Op.getOperand(0);
+ SDValue Amt = Op.getOperand(1);
+
+ if ((VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) ||
+ VT == MVT::v4i32 || VT == MVT::v8i16 ||
+ (Subtarget->hasInt256() &&
+ ((VT == MVT::v4i64 && Op.getOpcode() != ISD::SRA) ||
+ VT == MVT::v8i32 || VT == MVT::v16i16))) {
+ SDValue BaseShAmt;
+ EVT EltVT = VT.getVectorElementType();
+
+ if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
+ unsigned NumElts = VT.getVectorNumElements();
+ unsigned i, j;
+ for (i = 0; i != NumElts; ++i) {
+ if (Amt.getOperand(i).getOpcode() == ISD::UNDEF)
+ continue;
+ break;
+ }
+ for (j = i; j != NumElts; ++j) {
+ SDValue Arg = Amt.getOperand(j);
+ if (Arg.getOpcode() == ISD::UNDEF) continue;
+ if (Arg != Amt.getOperand(i))
+ break;
+ }
+ if (i != NumElts && j == NumElts)
+ BaseShAmt = Amt.getOperand(i);
+ } else {
+ if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
+ Amt = Amt.getOperand(0);
+ if (Amt.getOpcode() == ISD::VECTOR_SHUFFLE &&
+ cast<ShuffleVectorSDNode>(Amt)->isSplat()) {
+ SDValue InVec = Amt.getOperand(0);
+ if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
+ unsigned NumElts = InVec.getValueType().getVectorNumElements();
+ unsigned i = 0;
+ for (; i != NumElts; ++i) {
+ SDValue Arg = InVec.getOperand(i);
+ if (Arg.getOpcode() == ISD::UNDEF) continue;
+ BaseShAmt = Arg;
+ break;
+ }
+ } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
+ if (ConstantSDNode *C =
+ dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
+ unsigned SplatIdx =
+ cast<ShuffleVectorSDNode>(Amt)->getSplatIndex();
+ if (C->getZExtValue() == SplatIdx)
+ BaseShAmt = InVec.getOperand(1);
+ }
+ }
+ if (BaseShAmt.getNode() == 0)
+ BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Amt,
+ DAG.getIntPtrConstant(0));
+ }
+ }
+
+ if (BaseShAmt.getNode()) {
+ if (EltVT.bitsGT(MVT::i32))
+ BaseShAmt = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BaseShAmt);
+ else if (EltVT.bitsLT(MVT::i32))
+ BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
+
+ switch (Op.getOpcode()) {
+ default:
+ llvm_unreachable("Unknown shift opcode!");
+ case ISD::SHL:
+ switch (VT.getSimpleVT().SimpleTy) {
+ default: return SDValue();
+ case MVT::v2i64:
+ case MVT::v4i32:
+ case MVT::v8i16:
+ case MVT::v4i64:
+ case MVT::v8i32:
+ case MVT::v16i16:
+ return getTargetVShiftNode(X86ISD::VSHLI, dl, VT, R, BaseShAmt, DAG);
+ }
+ case ISD::SRA:
+ switch (VT.getSimpleVT().SimpleTy) {
+ default: return SDValue();
+ case MVT::v4i32:
+ case MVT::v8i16:
+ case MVT::v8i32:
+ case MVT::v16i16:
+ return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, R, BaseShAmt, DAG);
+ }
+ case ISD::SRL:
+ switch (VT.getSimpleVT().SimpleTy) {
+ default: return SDValue();
+ case MVT::v2i64:
+ case MVT::v4i32:
+ case MVT::v8i16:
+ case MVT::v4i64:
+ case MVT::v8i32:
+ case MVT::v16i16:
+ return getTargetVShiftNode(X86ISD::VSRLI, dl, VT, R, BaseShAmt, DAG);
+ }
+ }
+ }
+ }
+
+ // Special case in 32-bit mode, where i64 is expanded into high and low parts.
+ if (!Subtarget->is64Bit() &&
+ (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
+ Amt.getOpcode() == ISD::BITCAST &&
+ Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
+ Amt = Amt.getOperand(0);
+ unsigned Ratio = Amt.getValueType().getVectorNumElements() /
+ VT.getVectorNumElements();
+ std::vector<SDValue> Vals(Ratio);
+ for (unsigned i = 0; i != Ratio; ++i)
+ Vals[i] = Amt.getOperand(i);
+ for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
+ for (unsigned j = 0; j != Ratio; ++j)
+ if (Vals[j] != Amt.getOperand(i + j))
+ return SDValue();
+ }
+ switch (Op.getOpcode()) {
+ default:
+ llvm_unreachable("Unknown shift opcode!");
+ case ISD::SHL:
+ return DAG.getNode(X86ISD::VSHL, dl, VT, R, Op.getOperand(1));
+ case ISD::SRL:
+ return DAG.getNode(X86ISD::VSRL, dl, VT, R, Op.getOperand(1));
+ case ISD::SRA:
+ return DAG.getNode(X86ISD::VSRA, dl, VT, R, Op.getOperand(1));
+ }
+ }
+
+ return SDValue();
+}
+
+SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
+
+ EVT VT = Op.getValueType();
+ DebugLoc dl = Op.getDebugLoc();
+ SDValue R = Op.getOperand(0);
+ SDValue Amt = Op.getOperand(1);
+ SDValue V;
+
+ if (!Subtarget->hasSSE2())
+ return SDValue();
+
+ V = LowerScalarImmediateShift(Op, DAG, Subtarget);
+ if (V.getNode())
+ return V;
+
+ V = LowerScalarVariableShift(Op, DAG, Subtarget);
+ if (V.getNode())
+ return V;
+
+ // AVX2 has VPSLLV/VPSRAV/VPSRLV.
+ if (Subtarget->hasInt256()) {
+ if (Op.getOpcode() == ISD::SRL &&
+ (VT == MVT::v2i64 || VT == MVT::v4i32 ||
+ VT == MVT::v4i64 || VT == MVT::v8i32))
+ return Op;
+ if (Op.getOpcode() == ISD::SHL &&
+ (VT == MVT::v2i64 || VT == MVT::v4i32 ||
+ VT == MVT::v4i64 || VT == MVT::v8i32))
+ return Op;
+ if (Op.getOpcode() == ISD::SRA && (VT == MVT::v4i32 || VT == MVT::v8i32))
+ return Op;
+ }
+
// Lower SHL with variable shift amount.
if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
- Op = DAG.getNode(X86ISD::VSHLI, dl, VT, Op.getOperand(1),
- DAG.getConstant(23, MVT::i32));
-
- const uint32_t CV[] = { 0x3f800000U, 0x3f800000U, 0x3f800000U, 0x3f800000U};
- Constant *C = ConstantDataVector::get(*Context, CV);
- SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
- SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
- MachinePointerInfo::getConstantPool(),
- false, false, false, 16);
+ Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, VT));
- Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
+ Op = DAG.getNode(ISD::ADD, dl, VT, Op, DAG.getConstant(0x3f800000U, VT));
Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
return DAG.getNode(ISD::MUL, dl, VT, Op, R);
assert(Subtarget->hasSSE2() && "Need SSE2 for pslli/pcmpeq.");
// a = a << 5;
- Op = DAG.getNode(X86ISD::VSHLI, dl, MVT::v8i16, Op.getOperand(1),
- DAG.getConstant(5, MVT::i32));
+ Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(5, VT));
Op = DAG.getNode(ISD::BITCAST, dl, VT, Op);
// Turn 'a' into a mask suitable for VSELECT
// fall through
case MVT::v4i32:
case MVT::v8i16: {
- SDValue Tmp1 = getTargetVShiftNode(X86ISD::VSHLI, dl, VT,
- Op.getOperand(0), ShAmt, DAG);
+ // (sext (vzext x)) -> (vsext x)
+ SDValue Op0 = Op.getOperand(0);
+ SDValue Op00 = Op0.getOperand(0);
+ SDValue Tmp1;
+ // Hopefully, this VECTOR_SHUFFLE is just a VZEXT.
+ if (Op0.getOpcode() == ISD::BITCAST &&
+ Op00.getOpcode() == ISD::VECTOR_SHUFFLE)
+ Tmp1 = LowerVectorIntExtend(Op00, DAG);
+ if (Tmp1.getNode()) {
+ SDValue Tmp1Op0 = Tmp1.getOperand(0);
+ assert(Tmp1Op0.getOpcode() == X86ISD::VZEXT &&
+ "This optimization is invalid without a VZEXT.");
+ return DAG.getNode(X86ISD::VSEXT, dl, VT, Tmp1Op0.getOperand(0));
+ }
+
+ // If the above didn't work, then just use Shift-Left + Shift-Right.
+ Tmp1 = getTargetVShiftNode(X86ISD::VSHLI, dl, VT, Op0, ShAmt, DAG);
return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, Tmp1, ShAmt, DAG);
}
}
Op.getOperand(1), Op.getOperand(2));
}
+SDValue X86TargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
+ assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
+
+ // For MacOSX, we want to call an alternative entry point: __sincos_stret,
+ // which returns the values in two XMM registers.
+ DebugLoc dl = Op.getDebugLoc();
+ SDValue Arg = Op.getOperand(0);
+ EVT ArgVT = Arg.getValueType();
+ Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
+
+ ArgListTy Args;
+ ArgListEntry Entry;
+
+ Entry.Node = Arg;
+ Entry.Ty = ArgTy;
+ Entry.isSExt = false;
+ Entry.isZExt = false;
+ Args.push_back(Entry);
+
+ // Only optimize x86_64 for now. i386 is a bit messy. For f32,
+ // the small struct {f32, f32} is returned in (eax, edx). For f64,
+ // the results are returned via SRet in memory.
+ const char *LibcallName = (ArgVT == MVT::f64)
+ ? "__sincos_stret" : "__sincosf_stret";
+ SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
+
+ StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
+ TargetLowering::
+ CallLoweringInfo CLI(DAG.getEntryNode(), RetTy,
+ false, false, false, false, 0,
+ CallingConv::C, /*isTaillCall=*/false,
+ /*doesNotRet=*/false, /*isReturnValueUsed*/true,
+ Callee, Args, DAG, dl);
+ std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
+ return CallResult.first;
+}
+
/// LowerOperation - Provide custom lowering hooks for some operations.
///
SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
- case ISD::TRUNCATE: return lowerTRUNCATE(Op, DAG);
+ case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, DAG);
case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, DAG);
case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, DAG);
case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
- case ISD::FP_EXTEND: return lowerFP_EXTEND(Op, DAG);
+ case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
case ISD::FABS: return LowerFABS(Op, DAG);
case ISD::FNEG: return LowerFNEG(Op, DAG);
case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
case ISD::ADD: return LowerADD(Op, DAG);
case ISD::SUB: return LowerSUB(Op, DAG);
case ISD::SDIV: return LowerSDIV(Op, DAG);
+ case ISD::FSINCOS: return LowerFSINCOS(Op, DAG);
}
}
return;
}
case ISD::UINT_TO_FP: {
- if (N->getOperand(0).getValueType() != MVT::v2i32 &&
+ assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
+ if (N->getOperand(0).getValueType() != MVT::v2i32 ||
N->getValueType(0) != MVT::v2f32)
return;
SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
case X86ISD::PTEST: return "X86ISD::PTEST";
case X86ISD::TESTP: return "X86ISD::TESTP";
- case X86ISD::PALIGN: return "X86ISD::PALIGN";
+ case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
case X86ISD::SAHF: return "X86ISD::SAHF";
case X86ISD::RDRAND: return "X86ISD::RDRAND";
+ case X86ISD::RDSEED: return "X86ISD::RDSEED";
case X86ISD::FMADD: return "X86ISD::FMADD";
case X86ISD::FMSUB: return "X86ISD::FMSUB";
case X86ISD::FNMADD: return "X86ISD::FNMADD";
case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
+ case X86ISD::XTEST: return "X86ISD::XTEST";
}
}
// to
//
// ...
-// EAX = LOAD MI.addr
+// t1 = LOAD MI.addr
// loop:
-// t1 = OP MI.val, EAX
-// LCMPXCHG [MI.addr], t1, [EAX is implicitly used & defined]
+// t4 = phi(t1, t3 / loop)
+// t2 = OP MI.val, t4
+// EAX = t4
+// LCMPXCHG [MI.addr], t2, [EAX is implicitly used & defined]
+// t3 = EAX
// JNE loop
// sink:
-// dst = EAX
+// dst = t3
// ...
MachineBasicBlock *
X86TargetLowering::EmitAtomicLoadArith(MachineInstr *MI,
MachineFunction::iterator I = MBB;
++I;
- assert(MI->getNumOperands() <= X86::AddrNumOperands + 2 &&
+ assert(MI->getNumOperands() <= X86::AddrNumOperands + 4 &&
"Unexpected number of operands");
assert(MI->hasOneMemOperand() &&
const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
MVT::SimpleValueType VT = *RC->vt_begin();
- unsigned AccPhyReg = getX86SubSuperRegister(X86::EAX, VT);
+ unsigned t1 = MRI.createVirtualRegister(RC);
+ unsigned t2 = MRI.createVirtualRegister(RC);
+ unsigned t3 = MRI.createVirtualRegister(RC);
+ unsigned t4 = MRI.createVirtualRegister(RC);
+ unsigned PhyReg = getX86SubSuperRegister(X86::EAX, VT);
unsigned LCMPXCHGOpc = getCmpXChgOpcode(VT);
unsigned LOADOpc = getLoadOpcode(VT);
// For the atomic load-arith operator, we generate
//
// thisMBB:
- // EAX = LOAD [MI.addr]
+ // t1 = LOAD [MI.addr]
// mainMBB:
+ // t4 = phi(t1 / thisMBB, t3 / mainMBB)
// t1 = OP MI.val, EAX
+ // EAX = t4
// LCMPXCHG [MI.addr], t1, [EAX is implicitly used & defined]
+ // t3 = EAX
// JNE mainMBB
// sinkMBB:
+ // dst = t3
MachineBasicBlock *thisMBB = MBB;
MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
// thisMBB:
- MIB = BuildMI(thisMBB, DL, TII->get(LOADOpc), AccPhyReg);
- for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
- MIB.addOperand(MI->getOperand(MemOpndSlot + i));
- MIB.setMemRefs(MMOBegin, MMOEnd);
+ MIB = BuildMI(thisMBB, DL, TII->get(LOADOpc), t1);
+ for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
+ MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
+ if (NewMO.isReg())
+ NewMO.setIsKill(false);
+ MIB.addOperand(NewMO);
+ }
+ for (MachineInstr::mmo_iterator MMOI = MMOBegin; MMOI != MMOEnd; ++MMOI) {
+ unsigned flags = (*MMOI)->getFlags();
+ flags = (flags & ~MachineMemOperand::MOStore) | MachineMemOperand::MOLoad;
+ MachineMemOperand *MMO =
+ MF->getMachineMemOperand((*MMOI)->getPointerInfo(), flags,
+ (*MMOI)->getSize(),
+ (*MMOI)->getBaseAlignment(),
+ (*MMOI)->getTBAAInfo(),
+ (*MMOI)->getRanges());
+ MIB.addMemOperand(MMO);
+ }
thisMBB->addSuccessor(mainMBB);
// mainMBB:
MachineBasicBlock *origMainMBB = mainMBB;
- mainMBB->addLiveIn(AccPhyReg);
- // Copy AccPhyReg as it is used more than once.
- unsigned AccReg = MRI.createVirtualRegister(RC);
- BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), AccReg)
- .addReg(AccPhyReg);
+ // Add a PHI.
+ MachineInstr *Phi = BuildMI(mainMBB, DL, TII->get(X86::PHI), t4)
+ .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(mainMBB);
- unsigned t1 = MRI.createVirtualRegister(RC);
unsigned Opc = MI->getOpcode();
switch (Opc) {
default:
case X86::ATOMXOR32:
case X86::ATOMXOR64: {
unsigned ARITHOpc = getNonAtomicOpcode(Opc);
- BuildMI(mainMBB, DL, TII->get(ARITHOpc), t1).addReg(SrcReg)
- .addReg(AccReg);
+ BuildMI(mainMBB, DL, TII->get(ARITHOpc), t2).addReg(SrcReg)
+ .addReg(t4);
break;
}
case X86::ATOMNAND8:
case X86::ATOMNAND16:
case X86::ATOMNAND32:
case X86::ATOMNAND64: {
- unsigned t2 = MRI.createVirtualRegister(RC);
+ unsigned Tmp = MRI.createVirtualRegister(RC);
unsigned NOTOpc;
unsigned ANDOpc = getNonAtomicOpcodeWithExtraOpc(Opc, NOTOpc);
- BuildMI(mainMBB, DL, TII->get(ANDOpc), t2).addReg(SrcReg)
- .addReg(AccReg);
- BuildMI(mainMBB, DL, TII->get(NOTOpc), t1).addReg(t2);
+ BuildMI(mainMBB, DL, TII->get(ANDOpc), Tmp).addReg(SrcReg)
+ .addReg(t4);
+ BuildMI(mainMBB, DL, TII->get(NOTOpc), t2).addReg(Tmp);
break;
}
case X86::ATOMMAX8:
BuildMI(mainMBB, DL, TII->get(CMPOpc))
.addReg(SrcReg)
- .addReg(AccReg);
+ .addReg(t4);
if (Subtarget->hasCMov()) {
if (VT != MVT::i8) {
// Native support
- BuildMI(mainMBB, DL, TII->get(CMOVOpc), t1)
+ BuildMI(mainMBB, DL, TII->get(CMOVOpc), t2)
.addReg(SrcReg)
- .addReg(AccReg);
+ .addReg(t4);
} else {
// Promote i8 to i32 to use CMOV32
- const TargetRegisterClass *RC32 = getRegClassFor(MVT::i32);
+ const TargetRegisterInfo* TRI = getTargetMachine().getRegisterInfo();
+ const TargetRegisterClass *RC32 =
+ TRI->getSubClassWithSubReg(getRegClassFor(MVT::i32), X86::sub_8bit);
unsigned SrcReg32 = MRI.createVirtualRegister(RC32);
unsigned AccReg32 = MRI.createVirtualRegister(RC32);
- unsigned t2 = MRI.createVirtualRegister(RC32);
+ unsigned Tmp = MRI.createVirtualRegister(RC32);
unsigned Undef = MRI.createVirtualRegister(RC32);
BuildMI(mainMBB, DL, TII->get(TargetOpcode::IMPLICIT_DEF), Undef);
.addImm(X86::sub_8bit);
BuildMI(mainMBB, DL, TII->get(TargetOpcode::INSERT_SUBREG), AccReg32)
.addReg(Undef)
- .addReg(AccReg)
+ .addReg(t4)
.addImm(X86::sub_8bit);
- BuildMI(mainMBB, DL, TII->get(CMOVOpc), t2)
+ BuildMI(mainMBB, DL, TII->get(CMOVOpc), Tmp)
.addReg(SrcReg32)
.addReg(AccReg32);
- BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t1)
- .addReg(t2, 0, X86::sub_8bit);
+ BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t2)
+ .addReg(Tmp, 0, X86::sub_8bit);
}
} else {
// Use pseudo select and lower them.
unsigned SelOpc = getPseudoCMOVOpc(VT);
X86::CondCode CC = X86::getCondFromCMovOpc(CMOVOpc);
assert(CC != X86::COND_INVALID && "Invalid atomic-load-op transformation!");
- MIB = BuildMI(mainMBB, DL, TII->get(SelOpc), t1)
- .addReg(SrcReg).addReg(AccReg)
+ MIB = BuildMI(mainMBB, DL, TII->get(SelOpc), t2)
+ .addReg(SrcReg).addReg(t4)
.addImm(CC);
mainMBB = EmitLoweredSelect(MIB, mainMBB);
+ // Replace the original PHI node as mainMBB is changed after CMOV
+ // lowering.
+ BuildMI(*origMainMBB, Phi, DL, TII->get(X86::PHI), t4)
+ .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(mainMBB);
+ Phi->eraseFromParent();
}
break;
}
}
- // Copy AccPhyReg back from virtual register.
- BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), AccPhyReg)
- .addReg(AccReg);
+ // Copy PhyReg back from virtual register.
+ BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), PhyReg)
+ .addReg(t4);
MIB = BuildMI(mainMBB, DL, TII->get(LCMPXCHGOpc));
- for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
- MIB.addOperand(MI->getOperand(MemOpndSlot + i));
- MIB.addReg(t1);
+ for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
+ MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
+ if (NewMO.isReg())
+ NewMO.setIsKill(false);
+ MIB.addOperand(NewMO);
+ }
+ MIB.addReg(t2);
MIB.setMemRefs(MMOBegin, MMOEnd);
+ // Copy PhyReg back to virtual register.
+ BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t3)
+ .addReg(PhyReg);
+
BuildMI(mainMBB, DL, TII->get(X86::JNE_4)).addMBB(origMainMBB);
mainMBB->addSuccessor(origMainMBB);
mainMBB->addSuccessor(sinkMBB);
// sinkMBB:
- sinkMBB->addLiveIn(AccPhyReg);
-
BuildMI(*sinkMBB, sinkMBB->begin(), DL,
TII->get(TargetOpcode::COPY), DstReg)
- .addReg(AccPhyReg);
+ .addReg(t3);
MI->eraseFromParent();
return sinkMBB;
// to
//
// ...
-// EAX = LOAD [MI.addr + 0]
-// EDX = LOAD [MI.addr + 4]
+// t1L = LOAD [MI.addr + 0]
+// t1H = LOAD [MI.addr + 4]
// loop:
-// EBX = OP MI.val.lo, EAX
-// ECX = OP MI.val.hi, EDX
+// t4L = phi(t1L, t3L / loop)
+// t4H = phi(t1H, t3H / loop)
+// t2L = OP MI.val.lo, t4L
+// t2H = OP MI.val.hi, t4H
+// EAX = t4L
+// EDX = t4H
+// EBX = t2L
+// ECX = t2H
// LCMPXCHG8B [MI.addr], [ECX:EBX & EDX:EAX are implicitly used and EDX:EAX is implicitly defined]
+// t3L = EAX
+// t3H = EDX
// JNE loop
// sink:
-// dst = EDX:EAX
+// dstL = t3L
+// dstH = t3H
// ...
MachineBasicBlock *
X86TargetLowering::EmitAtomicLoadArith6432(MachineInstr *MI,
MachineFunction::iterator I = MBB;
++I;
- assert(MI->getNumOperands() <= X86::AddrNumOperands + 4 &&
+ assert(MI->getNumOperands() <= X86::AddrNumOperands + 7 &&
"Unexpected number of operands");
assert(MI->hasOneMemOperand() &&
const TargetRegisterClass *RC = &X86::GR32RegClass;
const TargetRegisterClass *RC8 = &X86::GR8RegClass;
+ unsigned t1L = MRI.createVirtualRegister(RC);
+ unsigned t1H = MRI.createVirtualRegister(RC);
+ unsigned t2L = MRI.createVirtualRegister(RC);
+ unsigned t2H = MRI.createVirtualRegister(RC);
+ unsigned t3L = MRI.createVirtualRegister(RC);
+ unsigned t3H = MRI.createVirtualRegister(RC);
+ unsigned t4L = MRI.createVirtualRegister(RC);
+ unsigned t4H = MRI.createVirtualRegister(RC);
+
unsigned LCMPXCHGOpc = X86::LCMPXCHG8B;
unsigned LOADOpc = X86::MOV32rm;
// For the atomic load-arith operator, we generate
//
// thisMBB:
- // EAX = LOAD [MI.addr + 0]
- // EDX = LOAD [MI.addr + 4]
+ // t1L = LOAD [MI.addr + 0]
+ // t1H = LOAD [MI.addr + 4]
// mainMBB:
- // EBX = OP MI.vallo, EAX
- // ECX = OP MI.valhi, EDX
+ // t4L = phi(t1L / thisMBB, t3L / mainMBB)
+ // t4H = phi(t1H / thisMBB, t3H / mainMBB)
+ // t2L = OP MI.val.lo, t4L
+ // t2H = OP MI.val.hi, t4H
+ // EBX = t2L
+ // ECX = t2H
// LCMPXCHG8B [MI.addr], [ECX:EBX & EDX:EAX are implicitly used and EDX:EAX is implicitly defined]
- // JNE mainMBB
+ // t3L = EAX
+ // t3H = EDX
+ // JNE loop
// sinkMBB:
+ // dstL = t3L
+ // dstH = t3H
MachineBasicBlock *thisMBB = MBB;
MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
// thisMBB:
// Lo
- MIB = BuildMI(thisMBB, DL, TII->get(LOADOpc), X86::EAX);
- for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
- MIB.addOperand(MI->getOperand(MemOpndSlot + i));
- MIB.setMemRefs(MMOBegin, MMOEnd);
+ MIB = BuildMI(thisMBB, DL, TII->get(LOADOpc), t1L);
+ for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
+ MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
+ if (NewMO.isReg())
+ NewMO.setIsKill(false);
+ MIB.addOperand(NewMO);
+ }
+ for (MachineInstr::mmo_iterator MMOI = MMOBegin; MMOI != MMOEnd; ++MMOI) {
+ unsigned flags = (*MMOI)->getFlags();
+ flags = (flags & ~MachineMemOperand::MOStore) | MachineMemOperand::MOLoad;
+ MachineMemOperand *MMO =
+ MF->getMachineMemOperand((*MMOI)->getPointerInfo(), flags,
+ (*MMOI)->getSize(),
+ (*MMOI)->getBaseAlignment(),
+ (*MMOI)->getTBAAInfo(),
+ (*MMOI)->getRanges());
+ MIB.addMemOperand(MMO);
+ };
+ MachineInstr *LowMI = MIB;
+
// Hi
- MIB = BuildMI(thisMBB, DL, TII->get(LOADOpc), X86::EDX);
+ MIB = BuildMI(thisMBB, DL, TII->get(LOADOpc), t1H);
for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
- if (i == X86::AddrDisp)
+ if (i == X86::AddrDisp) {
MIB.addDisp(MI->getOperand(MemOpndSlot + i), 4); // 4 == sizeof(i32)
- else
- MIB.addOperand(MI->getOperand(MemOpndSlot + i));
+ } else {
+ MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
+ if (NewMO.isReg())
+ NewMO.setIsKill(false);
+ MIB.addOperand(NewMO);
+ }
}
- MIB.setMemRefs(MMOBegin, MMOEnd);
+ MIB.setMemRefs(LowMI->memoperands_begin(), LowMI->memoperands_end());
thisMBB->addSuccessor(mainMBB);
// mainMBB:
MachineBasicBlock *origMainMBB = mainMBB;
- mainMBB->addLiveIn(X86::EAX);
- mainMBB->addLiveIn(X86::EDX);
-
- // Copy EDX:EAX as they are used more than once.
- unsigned LoReg = MRI.createVirtualRegister(RC);
- unsigned HiReg = MRI.createVirtualRegister(RC);
- BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), LoReg).addReg(X86::EAX);
- BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), HiReg).addReg(X86::EDX);
- unsigned t1L = MRI.createVirtualRegister(RC);
- unsigned t1H = MRI.createVirtualRegister(RC);
+ // Add PHIs.
+ MachineInstr *PhiL = BuildMI(mainMBB, DL, TII->get(X86::PHI), t4L)
+ .addReg(t1L).addMBB(thisMBB).addReg(t3L).addMBB(mainMBB);
+ MachineInstr *PhiH = BuildMI(mainMBB, DL, TII->get(X86::PHI), t4H)
+ .addReg(t1H).addMBB(thisMBB).addReg(t3H).addMBB(mainMBB);
unsigned Opc = MI->getOpcode();
switch (Opc) {
case X86::ATOMSUB6432: {
unsigned HiOpc;
unsigned LoOpc = getNonAtomic6432Opcode(Opc, HiOpc);
- BuildMI(mainMBB, DL, TII->get(LoOpc), t1L).addReg(LoReg).addReg(SrcLoReg);
- BuildMI(mainMBB, DL, TII->get(HiOpc), t1H).addReg(HiReg).addReg(SrcHiReg);
+ BuildMI(mainMBB, DL, TII->get(LoOpc), t2L).addReg(t4L)
+ .addReg(SrcLoReg);
+ BuildMI(mainMBB, DL, TII->get(HiOpc), t2H).addReg(t4H)
+ .addReg(SrcHiReg);
break;
}
case X86::ATOMNAND6432: {
unsigned HiOpc, NOTOpc;
unsigned LoOpc = getNonAtomic6432OpcodeWithExtraOpc(Opc, HiOpc, NOTOpc);
- unsigned t2L = MRI.createVirtualRegister(RC);
- unsigned t2H = MRI.createVirtualRegister(RC);
- BuildMI(mainMBB, DL, TII->get(LoOpc), t2L).addReg(SrcLoReg).addReg(LoReg);
- BuildMI(mainMBB, DL, TII->get(HiOpc), t2H).addReg(SrcHiReg).addReg(HiReg);
- BuildMI(mainMBB, DL, TII->get(NOTOpc), t1L).addReg(t2L);
- BuildMI(mainMBB, DL, TII->get(NOTOpc), t1H).addReg(t2H);
+ unsigned TmpL = MRI.createVirtualRegister(RC);
+ unsigned TmpH = MRI.createVirtualRegister(RC);
+ BuildMI(mainMBB, DL, TII->get(LoOpc), TmpL).addReg(SrcLoReg)
+ .addReg(t4L);
+ BuildMI(mainMBB, DL, TII->get(HiOpc), TmpH).addReg(SrcHiReg)
+ .addReg(t4H);
+ BuildMI(mainMBB, DL, TII->get(NOTOpc), t2L).addReg(TmpL);
+ BuildMI(mainMBB, DL, TII->get(NOTOpc), t2H).addReg(TmpH);
break;
}
case X86::ATOMMAX6432:
unsigned cc = MRI.createVirtualRegister(RC);
// cl := cmp src_lo, lo
BuildMI(mainMBB, DL, TII->get(X86::CMP32rr))
- .addReg(SrcLoReg).addReg(LoReg);
+ .addReg(SrcLoReg).addReg(t4L);
BuildMI(mainMBB, DL, TII->get(LoOpc), cL);
BuildMI(mainMBB, DL, TII->get(X86::MOVZX32rr8), cL32).addReg(cL);
// ch := cmp src_hi, hi
BuildMI(mainMBB, DL, TII->get(X86::CMP32rr))
- .addReg(SrcHiReg).addReg(HiReg);
+ .addReg(SrcHiReg).addReg(t4H);
BuildMI(mainMBB, DL, TII->get(HiOpc), cH);
BuildMI(mainMBB, DL, TII->get(X86::MOVZX32rr8), cH32).addReg(cH);
// cc := if (src_hi == hi) ? cl : ch;
}
BuildMI(mainMBB, DL, TII->get(X86::TEST32rr)).addReg(cc).addReg(cc);
if (Subtarget->hasCMov()) {
- BuildMI(mainMBB, DL, TII->get(X86::CMOVNE32rr), t1L)
- .addReg(SrcLoReg).addReg(LoReg);
- BuildMI(mainMBB, DL, TII->get(X86::CMOVNE32rr), t1H)
- .addReg(SrcHiReg).addReg(HiReg);
+ BuildMI(mainMBB, DL, TII->get(X86::CMOVNE32rr), t2L)
+ .addReg(SrcLoReg).addReg(t4L);
+ BuildMI(mainMBB, DL, TII->get(X86::CMOVNE32rr), t2H)
+ .addReg(SrcHiReg).addReg(t4H);
} else {
- MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), t1L)
- .addReg(SrcLoReg).addReg(LoReg)
+ MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), t2L)
+ .addReg(SrcLoReg).addReg(t4L)
.addImm(X86::COND_NE);
mainMBB = EmitLoweredSelect(MIB, mainMBB);
- MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), t1H)
- .addReg(SrcHiReg).addReg(HiReg)
+ // As the lowered CMOV won't clobber EFLAGS, we could reuse it for the
+ // 2nd CMOV lowering.
+ mainMBB->addLiveIn(X86::EFLAGS);
+ MIB = BuildMI(mainMBB, DL, TII->get(X86::CMOV_GR32), t2H)
+ .addReg(SrcHiReg).addReg(t4H)
.addImm(X86::COND_NE);
mainMBB = EmitLoweredSelect(MIB, mainMBB);
+ // Replace the original PHI node as mainMBB is changed after CMOV
+ // lowering.
+ BuildMI(*origMainMBB, PhiL, DL, TII->get(X86::PHI), t4L)
+ .addReg(t1L).addMBB(thisMBB).addReg(t3L).addMBB(mainMBB);
+ BuildMI(*origMainMBB, PhiH, DL, TII->get(X86::PHI), t4H)
+ .addReg(t1H).addMBB(thisMBB).addReg(t3H).addMBB(mainMBB);
+ PhiL->eraseFromParent();
+ PhiH->eraseFromParent();
}
break;
}
case X86::ATOMSWAP6432: {
unsigned HiOpc;
unsigned LoOpc = getNonAtomic6432Opcode(Opc, HiOpc);
- BuildMI(mainMBB, DL, TII->get(LoOpc), t1L).addReg(SrcLoReg);
- BuildMI(mainMBB, DL, TII->get(HiOpc), t1H).addReg(SrcHiReg);
+ BuildMI(mainMBB, DL, TII->get(LoOpc), t2L).addReg(SrcLoReg);
+ BuildMI(mainMBB, DL, TII->get(HiOpc), t2H).addReg(SrcHiReg);
break;
}
}
// Copy EDX:EAX back from HiReg:LoReg
- BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::EAX).addReg(LoReg);
- BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::EDX).addReg(HiReg);
+ BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::EAX).addReg(t4L);
+ BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::EDX).addReg(t4H);
// Copy ECX:EBX from t1H:t1L
- BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::EBX).addReg(t1L);
- BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::ECX).addReg(t1H);
+ BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::EBX).addReg(t2L);
+ BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), X86::ECX).addReg(t2H);
MIB = BuildMI(mainMBB, DL, TII->get(LCMPXCHGOpc));
- for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
- MIB.addOperand(MI->getOperand(MemOpndSlot + i));
+ for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
+ MachineOperand NewMO = MI->getOperand(MemOpndSlot + i);
+ if (NewMO.isReg())
+ NewMO.setIsKill(false);
+ MIB.addOperand(NewMO);
+ }
MIB.setMemRefs(MMOBegin, MMOEnd);
+ // Copy EDX:EAX back to t3H:t3L
+ BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t3L).addReg(X86::EAX);
+ BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t3H).addReg(X86::EDX);
+
BuildMI(mainMBB, DL, TII->get(X86::JNE_4)).addMBB(origMainMBB);
mainMBB->addSuccessor(origMainMBB);
mainMBB->addSuccessor(sinkMBB);
// sinkMBB:
- sinkMBB->addLiveIn(X86::EAX);
- sinkMBB->addLiveIn(X86::EDX);
-
BuildMI(*sinkMBB, sinkMBB->begin(), DL,
TII->get(TargetOpcode::COPY), DstLoReg)
- .addReg(X86::EAX);
+ .addReg(t3L);
BuildMI(*sinkMBB, sinkMBB->begin(), DL,
TII->get(TargetOpcode::COPY), DstHiReg)
- .addReg(X86::EDX);
+ .addReg(t3H);
MI->eraseFromParent();
return sinkMBB;
if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
isSplatVector(CondRHS.getNode()) && isSplatVector(OpRHS.getNode())) {
APInt A = cast<ConstantSDNode>(OpRHS.getOperand(0))->getAPIntValue();
- if (CondRHS.getConstantOperandVal(0) == -A-1) {
- SmallVector<SDValue, 32> V(VT.getVectorNumElements(),
- DAG.getConstant(-A, VT.getScalarType()));
+ if (CondRHS.getConstantOperandVal(0) == -A-1)
return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS,
- DAG.getNode(ISD::BUILD_VECTOR, DL, VT,
- V.data(), V.size()));
- }
+ DAG.getConstant(-A, VT));
}
// Another special case: If C was a sign bit, the sub has been
// Quit if the constant is neither 0 or 1.
return SDValue();
- // Skip 'zext' node.
- if (SetCC.getOpcode() == ISD::ZERO_EXTEND)
+ // Skip 'zext' or 'trunc' node.
+ if (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
+ SetCC.getOpcode() == ISD::TRUNCATE)
SetCC = SetCC.getOperand(0);
switch (SetCC.getOpcode()) {
return SDValue();
// Quit if false value is not a constant.
if (!FVal) {
- // A special case for rdrand, where 0 is set if false cond is found.
SDValue Op = SetCC.getOperand(0);
- if (Op.getOpcode() != X86ISD::RDRAND)
+ // Skip 'zext' or 'trunc' node.
+ if (Op.getOpcode() == ISD::ZERO_EXTEND ||
+ Op.getOpcode() == ISD::TRUNCATE)
+ Op = Op.getOperand(0);
+ // A special case for rdrand/rdseed, where 0 is set if false cond is
+ // found.
+ if ((Op.getOpcode() != X86ISD::RDRAND &&
+ Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
return SDValue();
}
// Quit if false value is not the constant 0 or 1.
ConstantSDNode *CmpAgainst = 0;
if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
(CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
- dyn_cast<ConstantSDNode>(Cond.getOperand(0)) == 0) {
+ !isa<ConstantSDNode>(Cond.getOperand(0))) {
if (CC == X86::COND_NE &&
CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const X86Subtarget *Subtarget) {
- EVT VT = N->getValueType(0);
if (N->getOpcode() == ISD::SHL) {
SDValue V = PerformSHLCombine(N, DAG);
if (V.getNode()) return V;
}
- // On X86 with SSE2 support, we can transform this to a vector shift if
- // all elements are shifted by the same amount. We can't do this in legalize
- // because the a constant vector is typically transformed to a constant pool
- // so we have no knowledge of the shift amount.
- if (!Subtarget->hasSSE2())
- return SDValue();
-
- if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
- (!Subtarget->hasInt256() ||
- (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
- return SDValue();
-
- SDValue ShAmtOp = N->getOperand(1);
- EVT EltVT = VT.getVectorElementType();
- DebugLoc DL = N->getDebugLoc();
- SDValue BaseShAmt = SDValue();
- if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
- unsigned NumElts = VT.getVectorNumElements();
- unsigned i = 0;
- for (; i != NumElts; ++i) {
- SDValue Arg = ShAmtOp.getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) continue;
- BaseShAmt = Arg;
- break;
- }
- // Handle the case where the build_vector is all undef
- // FIXME: Should DAG allow this?
- if (i == NumElts)
- return SDValue();
-
- for (; i != NumElts; ++i) {
- SDValue Arg = ShAmtOp.getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) continue;
- if (Arg != BaseShAmt) {
- return SDValue();
- }
- }
- } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
- cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
- SDValue InVec = ShAmtOp.getOperand(0);
- if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
- unsigned NumElts = InVec.getValueType().getVectorNumElements();
- unsigned i = 0;
- for (; i != NumElts; ++i) {
- SDValue Arg = InVec.getOperand(i);
- if (Arg.getOpcode() == ISD::UNDEF) continue;
- BaseShAmt = Arg;
- break;
- }
- } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
- if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
- unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
- if (C->getZExtValue() == SplatIdx)
- BaseShAmt = InVec.getOperand(1);
- }
- }
- if (BaseShAmt.getNode() == 0) {
- // Don't create instructions with illegal types after legalize
- // types has run.
- if (!DAG.getTargetLoweringInfo().isTypeLegal(EltVT) &&
- !DCI.isBeforeLegalize())
- return SDValue();
-
- BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
- DAG.getIntPtrConstant(0));
- }
- } else
- return SDValue();
-
- // The shift amount is an i32.
- if (EltVT.bitsGT(MVT::i32))
- BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
- else if (EltVT.bitsLT(MVT::i32))
- BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
-
- // The shift amount is identical so we can do a vector shift.
- SDValue ValOp = N->getOperand(0);
- switch (N->getOpcode()) {
- default:
- llvm_unreachable("Unknown shift opcode!");
- case ISD::SHL:
- switch (VT.getSimpleVT().SimpleTy) {
- default: return SDValue();
- case MVT::v2i64:
- case MVT::v4i32:
- case MVT::v8i16:
- case MVT::v4i64:
- case MVT::v8i32:
- case MVT::v16i16:
- return getTargetVShiftNode(X86ISD::VSHLI, DL, VT, ValOp, BaseShAmt, DAG);
- }
- case ISD::SRA:
- switch (VT.getSimpleVT().SimpleTy) {
- default: return SDValue();
- case MVT::v4i32:
- case MVT::v8i16:
- case MVT::v8i32:
- case MVT::v16i16:
- return getTargetVShiftNode(X86ISD::VSRAI, DL, VT, ValOp, BaseShAmt, DAG);
- }
- case ISD::SRL:
- switch (VT.getSimpleVT().SimpleTy) {
- default: return SDValue();
- case MVT::v2i64:
- case MVT::v4i32:
- case MVT::v8i16:
- case MVT::v4i64:
- case MVT::v8i32:
- case MVT::v16i16:
- return getTargetVShiftNode(X86ISD::VSRLI, DL, VT, ValOp, BaseShAmt, DAG);
- }
- }
+ return SDValue();
}
// CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
if (VT == MVT::f32 || VT == MVT::f64) {
bool ExpectingFlags = false;
// Check for any users that want flags:
- for (SDNode::use_iterator UI = N->use_begin(),
- UE = N->use_end();
+ for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
!ExpectingFlags && UI != UE; ++UI)
switch (UI->getOpcode()) {
default:
// Validate that the Mask operand is a vector sra node.
// FIXME: what to do for bytes, since there is a psignb/pblendvb, but
// there is no psrai.b
- if (Mask.getOpcode() != X86ISD::VSRAI)
- return SDValue();
-
- // Check that the SRA is all signbits.
- SDValue SraC = Mask.getOperand(1);
- unsigned SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
+ unsigned SraAmt = ~0;
+ if (Mask.getOpcode() == ISD::SRA) {
+ SDValue Amt = Mask.getOperand(1);
+ if (isSplatVector(Amt.getNode())) {
+ SDValue SclrAmt = Amt->getOperand(0);
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt))
+ SraAmt = C->getZExtValue();
+ }
+ } else if (Mask.getOpcode() == X86ISD::VSRAI) {
+ SDValue SraC = Mask.getOperand(1);
+ SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
+ }
if ((SraAmt + 1) != EltBits)
return SDValue();
DebugLoc DL = N->getDebugLoc();
- // We are going to replace the AND, OR, NAND with either BLEND
- // or PSIGN, which only look at the MSB. The VSRAI instruction
- // does not affect the highest bit, so we can get rid of it.
- Mask = Mask.getOperand(0);
-
// 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.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
"Unsupported VT for PSIGN");
- Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask);
+ Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
}
// PBLENDVB only available on SSE 4.1
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
unsigned RegSz = RegVT.getSizeInBits();
+ // On Sandybridge unaligned 256bit loads are inefficient.
ISD::LoadExtType Ext = Ld->getExtensionType();
unsigned Alignment = Ld->getAlignment();
-
- // On Sandybridge unaligned 256bit loads are inefficient.
+ bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
if (RegVT.is256BitVector() && !Subtarget->hasInt256() &&
- !DCI.isBeforeLegalizeOps() && Alignment < 32 &&
- Ext == ISD::NON_EXTLOAD) {
+ !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
unsigned NumElems = RegVT.getVectorNumElements();
+ if (NumElems < 2)
+ return SDValue();
+
SDValue Ptr = Ld->getBasePtr();
SDValue Increment = DAG.getConstant(16, TLI.getPointerTy());
SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
Ld->getPointerInfo(), Ld->isVolatile(),
Ld->isNonTemporal(), Ld->isInvariant(),
- Alignment);
+ std::min(16U, Alignment));
SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
Load1.getValue(1),
Load2.getValue(1));
// Represent the data using the same element type that is stored in
// memory. In practice, we ''widen'' MemVT.
- EVT WideVecVT =
- EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
+ EVT WideVecVT =
+ EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
loadRegZize/MemVT.getScalarType().getSizeInBits());
assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
// Build the arithmetic shift.
unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
MemVT.getVectorElementType().getSizeInBits();
- SmallVector<SDValue, 8> C(NumElems,
- DAG.getConstant(Amt, RegVT.getScalarType()));
- SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, RegVT, &C[0], C.size());
- Shuff = DAG.getNode(ISD::SRA, dl, RegVT, Shuff, BV);
+ Shuff = DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
+ DAG.getConstant(Amt, RegVT));
return DCI.CombineTo(N, Shuff, TF, true);
}
// On Sandy Bridge, 256-bit memory operations are executed by two
// 128-bit ports. However, on Haswell it is better to issue a single 256-bit
// memory operation.
+ unsigned Alignment = St->getAlignment();
+ bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
if (VT.is256BitVector() && !Subtarget->hasInt256() &&
- StoredVal.getNode()->getOpcode() == ISD::CONCAT_VECTORS &&
- StoredVal.getNumOperands() == 2) {
- SDValue Value0 = StoredVal.getOperand(0);
- SDValue Value1 = StoredVal.getOperand(1);
+ StVT == VT && !IsAligned) {
+ unsigned NumElems = VT.getVectorNumElements();
+ if (NumElems < 2)
+ return SDValue();
+
+ SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
+ SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
SDValue Ptr0 = St->getBasePtr();
SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
St->getPointerInfo(), St->isVolatile(),
- St->isNonTemporal(), St->getAlignment());
+ St->isNonTemporal(), Alignment);
SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
St->getPointerInfo(), St->isVolatile(),
- St->isNonTemporal(), St->getAlignment());
+ St->isNonTemporal(),
+ std::min(16U, Alignment));
return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
}
return SDValue();
}
-static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
- TargetLowering::DAGCombinerInfo &DCI,
- const X86Subtarget *Subtarget) {
+static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
+ const X86Subtarget *Subtarget) {
EVT VT = N->getValueType(0);
-
if (!VT.isVector())
return SDValue();
- SDValue In = N->getOperand(0);
- EVT InVT = In.getValueType();
+ SDValue N0 = N->getOperand(0);
+ SDValue N1 = N->getOperand(1);
+ EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
DebugLoc dl = N->getDebugLoc();
- unsigned ExtendedEltSize = VT.getVectorElementType().getSizeInBits();
- // Split SIGN_EXTEND operation to use vmovsx instruction when possible
- if (InVT == MVT::v8i8) {
- if (ExtendedEltSize > 16 && !Subtarget->hasInt256())
- In = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, In);
- if (ExtendedEltSize > 32)
- In = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i32, In);
- return DAG.getNode(ISD::SIGN_EXTEND, dl, VT, In);
- }
+ // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
+ // both SSE and AVX2 since there is no sign-extended shift right
+ // operation on a vector with 64-bit elements.
+ //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
+ // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
+ if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
+ N0.getOpcode() == ISD::SIGN_EXTEND)) {
+ SDValue N00 = N0.getOperand(0);
+
+ // EXTLOAD has a better solution on AVX2,
+ // it may be replaced with X86ISD::VSEXT node.
+ if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
+ if (!ISD::isNormalLoad(N00.getNode()))
+ return SDValue();
- if ((InVT == MVT::v4i8 || InVT == MVT::v4i16) &&
- ExtendedEltSize > 32 && !Subtarget->hasInt256()) {
- In = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, In);
- return DAG.getNode(ISD::SIGN_EXTEND, dl, VT, In);
+ if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
+ SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
+ N00, N1);
+ return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
+ }
}
+ return SDValue();
+}
+static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
+ TargetLowering::DAGCombinerInfo &DCI,
+ const X86Subtarget *Subtarget) {
if (!DCI.isBeforeLegalizeOps())
return SDValue();
if (!Subtarget->hasFp256())
return SDValue();
- if (VT.is256BitVector()) {
+ EVT VT = N->getValueType(0);
+ if (VT.isVector() && VT.getSizeInBits() == 256) {
SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
if (R.getNode())
return R;
return SDValue();
}
-// Helper function of PerformSETCCCombine. It is to materialize "setb reg"
-// as "sbb reg,reg", since it can be extended without zext and produces
+// Helper function of PerformSETCCCombine. It is to materialize "setb reg"
+// as "sbb reg,reg", since it can be extended without zext and produces
// an all-ones bit which is more useful than 0/1 in some cases.
static SDValue MaterializeSETB(DebugLoc DL, SDValue EFLAGS, SelectionDAG &DAG) {
return DAG.getNode(ISD::AND, DL, MVT::i8,
SDValue EFLAGS = N->getOperand(1);
if (CC == X86::COND_A) {
- // Try to convert COND_A into COND_B in an attempt to facilitate
+ // Try to convert COND_A into COND_B in an attempt to facilitate
// materializing "setb reg".
//
// Do not flip "e > c", where "c" is a constant, because Cmp instruction
// cannot take an immediate as its first operand.
//
- if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
+ if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
EFLAGS.getValueType().isInteger() &&
!isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
SDValue NewSub = DAG.getNode(X86ISD::SUB, EFLAGS.getDebugLoc(),
if (In.getOpcode() != X86ISD::VZEXT)
return SDValue();
- return DAG.getNode(X86ISD::VZEXT, N->getDebugLoc(), N->getValueType(0), In.getOperand(0));
+ return DAG.getNode(X86ISD::VZEXT, N->getDebugLoc(), N->getValueType(0),
+ In.getOperand(0));
}
SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
case ISD::ANY_EXTEND:
case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
+ case ISD::SIGN_EXTEND_INREG: return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG,DCI,Subtarget);
case ISD::SETCC: return PerformISDSETCCCombine(N, DAG);
case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
case X86ISD::SHUFP: // Handle all target specific shuffles
- case X86ISD::PALIGN:
+ case X86ISD::PALIGNR:
case X86ISD::UNPCKH:
case X86ISD::UNPCKL:
case X86ISD::MOVHLPS:
AsmPieces.clear();
const std::string &ConstraintsStr = IA->getConstraintString();
SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
- std::sort(AsmPieces.begin(), AsmPieces.end());
+ array_pod_sort(AsmPieces.begin(), AsmPieces.end());
if (AsmPieces.size() == 4 &&
AsmPieces[0] == "~{cc}" &&
AsmPieces[1] == "~{dirflag}" &&
AsmPieces.clear();
const std::string &ConstraintsStr = IA->getConstraintString();
SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
- std::sort(AsmPieces.begin(), AsmPieces.end());
+ array_pod_sort(AsmPieces.begin(), AsmPieces.end());
if (AsmPieces.size() == 4 &&
AsmPieces[0] == "~{cc}" &&
AsmPieces[1] == "~{dirflag}" &&
// really want an 8-bit or 32-bit register, map to the appropriate register
// class and return the appropriate register.
if (Res.second == &X86::GR16RegClass) {
- if (VT == MVT::i8) {
+ if (VT == MVT::i8 || VT == MVT::i1) {
unsigned DestReg = 0;
switch (Res.first) {
default: break;
Res.first = DestReg;
Res.second = &X86::GR8RegClass;
}
- } else if (VT == MVT::i32) {
+ } else if (VT == MVT::i32 || VT == MVT::f32) {
unsigned DestReg = 0;
switch (Res.first) {
default: break;
Res.first = DestReg;
Res.second = &X86::GR32RegClass;
}
- } else if (VT == MVT::i64) {
+ } else if (VT == MVT::i64 || VT == MVT::f64) {
unsigned DestReg = 0;
switch (Res.first) {
default: break;