}// has AVX-512
if (!TM.Options.UseSoftFloat && Subtarget->hasBWI()) {
+ addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
+ addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
+
addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
+
+ setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
+ setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
+ setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
+ setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
+
+ for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
+ const MVT VT = (MVT::SimpleValueType)i;
+
+ const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
+
+ // Do not attempt to promote non-256-bit vectors
+ if (!VT.is512BitVector())
+ continue;
+
+ if ( EltSize < 32) {
+ setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
+ setOperationAction(ISD::VSELECT, VT, Legal);
+ }
+ }
+ }
+
+ if (!TM.Options.UseSoftFloat && Subtarget->hasVLX()) {
+ addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
+ addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
+
+ setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
+ setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
}
// SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
setPrefFunctionAlignment(4); // 2^4 bytes.
- InitIntrinsicTables();
+ verifyIntrinsicTables();
}
// This has so far only been implemented for 64-bit MachO.
if (!VT.isVector())
return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
- if (Subtarget->hasAVX512())
- switch(VT.getVectorNumElements()) {
- case 8: return MVT::v8i1;
- case 16: return MVT::v16i1;
+ const unsigned NumElts = VT.getVectorNumElements();
+ const EVT EltVT = VT.getVectorElementType();
+ if (VT.is512BitVector()) {
+ if (Subtarget->hasAVX512())
+ if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
+ EltVT == MVT::f32 || EltVT == MVT::f64)
+ switch(NumElts) {
+ case 8: return MVT::v8i1;
+ case 16: return MVT::v16i1;
+ }
+ if (Subtarget->hasBWI())
+ if (EltVT == MVT::i8 || EltVT == MVT::i16)
+ switch(NumElts) {
+ case 32: return MVT::v32i1;
+ case 64: return MVT::v64i1;
+ }
+ }
+
+ if (VT.is256BitVector() || VT.is128BitVector()) {
+ if (Subtarget->hasVLX())
+ if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
+ EltVT == MVT::f32 || EltVT == MVT::f64)
+ switch(NumElts) {
+ case 2: return MVT::v2i1;
+ case 4: return MVT::v4i1;
+ case 8: return MVT::v8i1;
+ }
+ if (Subtarget->hasBWI() && Subtarget->hasVLX())
+ if (EltVT == MVT::i8 || EltVT == MVT::i16)
+ switch(NumElts) {
+ case 8: return MVT::v8i1;
+ case 16: return MVT::v16i1;
+ case 32: return MVT::v32i1;
+ }
}
return VT.changeVectorElementTypeToInteger();
}
}
+// FIXME: Get this from tablegen.
+static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
+ const X86Subtarget *Subtarget) {
+ assert(Subtarget->is64Bit());
+
+ if (Subtarget->isCallingConvWin64(CallConv)) {
+ static const MCPhysReg GPR64ArgRegsWin64[] = {
+ X86::RCX, X86::RDX, X86::R8, X86::R9
+ };
+ return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
+ }
+
+ static const MCPhysReg GPR64ArgRegs64Bit[] = {
+ X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
+ };
+ return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
+}
+
+// FIXME: Get this from tablegen.
+static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
+ CallingConv::ID CallConv,
+ const X86Subtarget *Subtarget) {
+ assert(Subtarget->is64Bit());
+ if (Subtarget->isCallingConvWin64(CallConv)) {
+ // The XMM registers which might contain var arg parameters are shadowed
+ // in their paired GPR. So we only need to save the GPR to their home
+ // slots.
+ // TODO: __vectorcall will change this.
+ return None;
+ }
+
+ const Function *Fn = MF.getFunction();
+ bool NoImplicitFloatOps = Fn->getAttributes().
+ hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
+ assert(!(MF.getTarget().Options.UseSoftFloat && NoImplicitFloatOps) &&
+ "SSE register cannot be used when SSE is disabled!");
+ if (MF.getTarget().Options.UseSoftFloat || NoImplicitFloatOps ||
+ !Subtarget->hasSSE1())
+ // Kernel mode asks for SSE to be disabled, so there are no XMM argument
+ // registers.
+ return None;
+
+ static const MCPhysReg XMMArgRegs64Bit[] = {
+ X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
+ X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
+ };
+ return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
+}
+
SDValue
X86TargetLowering::LowerFormalArguments(SDValue Chain,
CallingConv::ID CallConv,
// If the function takes variable number of arguments, make a frame index for
// the start of the first vararg value... for expansion of llvm.va_start. We
// can skip this if there are no va_start calls.
- if (isVarArg && MFI->hasVAStart()) {
- if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
- CallConv != CallingConv::X86_ThisCall)) {
- FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
- }
- if (Is64Bit) {
- unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
-
- // FIXME: We should really autogenerate these arrays
- static const MCPhysReg GPR64ArgRegsWin64[] = {
- X86::RCX, X86::RDX, X86::R8, X86::R9
- };
- static const MCPhysReg GPR64ArgRegs64Bit[] = {
- X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
- };
- static const MCPhysReg XMMArgRegs64Bit[] = {
- X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
- X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
- };
- const MCPhysReg *GPR64ArgRegs;
- unsigned NumXMMRegs = 0;
-
- if (IsWin64) {
- // The XMM registers which might contain var arg parameters are shadowed
- // in their paired GPR. So we only need to save the GPR to their home
- // slots.
- TotalNumIntRegs = 4;
- GPR64ArgRegs = GPR64ArgRegsWin64;
- } else {
- TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
- GPR64ArgRegs = GPR64ArgRegs64Bit;
-
- NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit,
- TotalNumXMMRegs);
+ if (MFI->hasVAStart() &&
+ (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
+ CallConv != CallingConv::X86_ThisCall))) {
+ FuncInfo->setVarArgsFrameIndex(
+ MFI->CreateFixedObject(1, StackSize, true));
+ }
+
+ // 64-bit calling conventions support varargs and register parameters, so we
+ // have to do extra work to spill them in the prologue or forward them to
+ // musttail calls.
+ if (Is64Bit && isVarArg &&
+ (MFI->hasVAStart() || MFI->hasMustTailInVarArgFunc())) {
+ // Find the first unallocated argument registers.
+ ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
+ ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
+ unsigned NumIntRegs =
+ CCInfo.getFirstUnallocated(ArgGPRs.data(), ArgGPRs.size());
+ unsigned NumXMMRegs =
+ CCInfo.getFirstUnallocated(ArgXMMs.data(), ArgXMMs.size());
+ assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
+ "SSE register cannot be used when SSE is disabled!");
+
+ // Gather all the live in physical registers.
+ SmallVector<SDValue, 6> LiveGPRs;
+ SmallVector<SDValue, 8> LiveXMMRegs;
+ SDValue ALVal;
+ for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
+ unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
+ LiveGPRs.push_back(
+ DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
+ }
+ if (!ArgXMMs.empty()) {
+ unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
+ ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
+ for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
+ unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
+ LiveXMMRegs.push_back(
+ DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
}
- unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
- TotalNumIntRegs);
-
- bool NoImplicitFloatOps = Fn->getAttributes().
- hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
- assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
- "SSE register cannot be used when SSE is disabled!");
- assert(!(NumXMMRegs && MF.getTarget().Options.UseSoftFloat &&
- NoImplicitFloatOps) &&
- "SSE register cannot be used when SSE is disabled!");
- if (MF.getTarget().Options.UseSoftFloat || NoImplicitFloatOps ||
- !Subtarget->hasSSE1())
- // Kernel mode asks for SSE to be disabled, so don't push them
- // on the stack.
- TotalNumXMMRegs = 0;
+ }
+ // Store them to the va_list returned by va_start.
+ if (MFI->hasVAStart()) {
if (IsWin64) {
const TargetFrameLowering &TFI = *MF.getSubtarget().getFrameLowering();
// Get to the caller-allocated home save location. Add 8 to account
// registers, then we must store them to their spots on the stack so
// they may be loaded by deferencing the result of va_next.
FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
- FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
- FuncInfo->setRegSaveFrameIndex(
- MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
- false));
+ FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
+ FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
+ ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
}
// Store the integer parameter registers.
SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
getPointerTy());
unsigned Offset = FuncInfo->getVarArgsGPOffset();
- for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
+ for (SDValue Val : LiveGPRs) {
SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
DAG.getIntPtrConstant(Offset));
- unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
- &X86::GR64RegClass);
- SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
SDValue Store =
DAG.getStore(Val.getValue(1), dl, Val, FIN,
MachinePointerInfo::getFixedStack(
Offset += 8;
}
- if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
+ if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
// Now store the XMM (fp + vector) parameter registers.
SmallVector<SDValue, 12> SaveXMMOps;
SaveXMMOps.push_back(Chain);
-
- unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
- SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
SaveXMMOps.push_back(ALVal);
-
SaveXMMOps.push_back(DAG.getIntPtrConstant(
FuncInfo->getRegSaveFrameIndex()));
SaveXMMOps.push_back(DAG.getIntPtrConstant(
FuncInfo->getVarArgsFPOffset()));
-
- for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
- unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
- &X86::VR128RegClass);
- SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
- SaveXMMOps.push_back(Val);
- }
+ SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
+ LiveXMMRegs.end());
MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
MVT::Other, SaveXMMOps));
}
if (!MemOps.empty())
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
+ } else {
+ // Add all GPRs, al, and XMMs to the list of forwards. We will add then
+ // to the liveout set on a musttail call.
+ assert(MFI->hasMustTailInVarArgFunc());
+ auto &Forwards = FuncInfo->getForwardedMustTailRegParms();
+ typedef X86MachineFunctionInfo::Forward Forward;
+
+ for (unsigned I = 0, E = LiveGPRs.size(); I != E; ++I) {
+ unsigned VReg =
+ MF.getRegInfo().createVirtualRegister(&X86::GR64RegClass);
+ Chain = DAG.getCopyToReg(Chain, dl, VReg, LiveGPRs[I]);
+ Forwards.push_back(Forward(VReg, ArgGPRs[NumIntRegs + I], MVT::i64));
+ }
+
+ if (!ArgXMMs.empty()) {
+ unsigned ALVReg =
+ MF.getRegInfo().createVirtualRegister(&X86::GR8RegClass);
+ Chain = DAG.getCopyToReg(Chain, dl, ALVReg, ALVal);
+ Forwards.push_back(Forward(ALVReg, X86::AL, MVT::i8));
+
+ for (unsigned I = 0, E = LiveXMMRegs.size(); I != E; ++I) {
+ unsigned VReg =
+ MF.getRegInfo().createVirtualRegister(&X86::VR128RegClass);
+ Chain = DAG.getCopyToReg(Chain, dl, VReg, LiveXMMRegs[I]);
+ Forwards.push_back(
+ Forward(VReg, ArgXMMs[NumXMMRegs + I], MVT::v4f32));
+ }
+ }
}
}
bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
StructReturnType SR = callIsStructReturn(Outs);
bool IsSibcall = false;
+ X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
if (MF.getTarget().Options.DisableTailCalls)
isTailCall = false;
int FPDiff = 0;
if (isTailCall && !IsSibcall && !IsMustTail) {
// Lower arguments at fp - stackoffset + fpdiff.
- X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
FPDiff = NumBytesCallerPushed - NumBytes;
}
}
- if (Is64Bit && isVarArg && !IsWin64) {
+ if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
// From AMD64 ABI document:
// For calls that may call functions that use varargs or stdargs
// (prototype-less calls or calls to functions containing ellipsis (...) in
DAG.getConstant(NumXMMRegs, MVT::i8)));
}
+ if (Is64Bit && isVarArg && IsMustTail) {
+ const auto &Forwards = X86Info->getForwardedMustTailRegParms();
+ for (const auto &F : Forwards) {
+ SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
+ RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
+ }
+ }
+
// For tail calls lower the arguments to the 'real' stack slots. Sibcalls
// don't need this because the eligibility check rejects calls that require
// shuffling arguments passed in memory.
/// own arguments. Callee pop is necessary to support tail calls.
bool X86::isCalleePop(CallingConv::ID CallingConv,
bool is64Bit, bool IsVarArg, bool TailCallOpt) {
- if (IsVarArg)
- return false;
-
switch (CallingConv) {
default:
return false;
case CallingConv::X86_StdCall:
- return !is64Bit;
case CallingConv::X86_FastCall:
- return !is64Bit;
case CallingConv::X86_ThisCall:
return !is64Bit;
case CallingConv::Fast:
- return TailCallOpt;
case CallingConv::GHC:
- return TailCallOpt;
case CallingConv::HiPE:
+ if (IsVarArg)
+ return false;
return TailCallOpt;
}
}
int V2Index =
std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
Mask.begin();
+
+ // Check for whether we can use INSERTPS to perform the blend. We only use
+ // INSERTPS when the V1 elements are already in the correct locations
+ // because otherwise we can just always use two SHUFPS instructions which
+ // are much smaller to encode than a SHUFPS and an INSERTPS.
+ if (Subtarget->hasSSE41()) {
+ // When using INSERTPS we can zero any lane of the destination. Collect
+ // the zero inputs into a mask and drop them from the lanes of V1 which
+ // actually need to be present as inputs to the INSERTPS.
+ unsigned ZMask = 0;
+ if (ISD::isBuildVectorAllZeros(V1.getNode())) {
+ ZMask = 0xF ^ (1 << V2Index);
+ } else if (V1.getOpcode() == ISD::BUILD_VECTOR) {
+ for (int i = 0; i < 4; ++i) {
+ int M = Mask[i];
+ if (M >= 4)
+ continue;
+ if (M > -1) {
+ SDValue Input = V1.getOperand(M);
+ if (Input.getOpcode() != ISD::UNDEF &&
+ !X86::isZeroNode(Input)) {
+ // A non-zero input!
+ ZMask = 0;
+ break;
+ }
+ }
+ ZMask |= 1 << i;
+ }
+ }
+
+ // Synthesize a shuffle mask for the non-zero and non-v2 inputs.
+ int InsertShuffleMask[4] = {-1, -1, -1, -1};
+ for (int i = 0; i < 4; ++i)
+ if (i != V2Index && (ZMask & (1 << i)) == 0)
+ InsertShuffleMask[i] = Mask[i];
+
+ if (isNoopShuffleMask(InsertShuffleMask)) {
+ // Replace V1 with undef if nothing from V1 survives the INSERTPS.
+ if ((ZMask | 1 << V2Index) == 0xF)
+ V1 = DAG.getUNDEF(MVT::v4f32);
+
+ // Insert the V2 element into the desired position.
+ SDValue InsertPSMask =
+ DAG.getIntPtrConstant(Mask[V2Index] << 6 | V2Index << 4 | ZMask);
+ return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
+ InsertPSMask);
+ }
+ }
+
// Compute the index adjacent to V2Index and in the same half by toggling
// the low bit.
int V2AdjIndex = V2Index ^ 1;
break;
case MVT::v8i16:
case MVT::v16i16:
+ if (Subtarget->hasBWI() && Subtarget->hasVLX())
+ break;
return SDValue();
}
return SDValue();
}
-static SDValue LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
- MVT VT = Op.getSimpleValueType();
- MVT EltVT = VT.getVectorElementType();
- SDLoc dl(Op);
-
- SDValue N0 = Op.getOperand(0);
- SDValue N1 = Op.getOperand(1);
- SDValue N2 = Op.getOperand(2);
-
- if (!VT.is128BitVector())
- return SDValue();
-
- if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
- isa<ConstantSDNode>(N2)) {
- unsigned Opc;
- if (VT == MVT::v8i16) {
- Opc = X86ISD::PINSRW;
- } else {
- assert(VT == MVT::v16i8);
- Opc = X86ISD::PINSRB;
- }
-
- // Transform it so it match pinsr{b,w} which expects a GR32 as its second
- // argument.
- if (N1.getValueType() != MVT::i32)
- N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
- if (N2.getValueType() != MVT::i32)
- N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
- return DAG.getNode(Opc, dl, VT, N0, N1, N2);
- }
-
- if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
- // Bits [7:6] of the constant are the source select. This will always be
- // zero here. The DAG Combiner may combine an extract_elt index into these
- // bits. For example (insert (extract, 3), 2) could be matched by putting
- // the '3' into bits [7:6] of X86ISD::INSERTPS.
- // Bits [5:4] of the constant are the destination select. This is the
- // value of the incoming immediate.
- // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
- // combine either bitwise AND or insert of float 0.0 to set these bits.
- N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
- // Create this as a scalar to vector..
- N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
- return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
- }
-
- if ((EltVT == MVT::i32 || EltVT == MVT::i64) && isa<ConstantSDNode>(N2)) {
- // PINSR* works with constant index.
- return Op;
- }
- return SDValue();
-}
-
/// Insert one bit to mask vector, like v16i1 or v8i1.
/// AVX-512 feature.
SDValue
DAG.getConstant(MaxSift - IdxVal, MVT::i8));
return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
}
-SDValue
-X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
+
+SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
+ SelectionDAG &DAG) const {
MVT VT = Op.getSimpleValueType();
MVT EltVT = VT.getVectorElementType();
-
+
if (EltVT == MVT::i1)
return InsertBitToMaskVector(Op, DAG);
SDValue N0 = Op.getOperand(0);
SDValue N1 = Op.getOperand(1);
SDValue N2 = Op.getOperand(2);
+ if (!isa<ConstantSDNode>(N2))
+ return SDValue();
+ auto *N2C = cast<ConstantSDNode>(N2);
+ unsigned IdxVal = N2C->getZExtValue();
- // If this is a 256-bit vector result, first extract the 128-bit vector,
- // insert the element into the extracted half and then place it back.
+ // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
+ // into that, and then insert the subvector back into the result.
if (VT.is256BitVector() || VT.is512BitVector()) {
- if (!isa<ConstantSDNode>(N2))
- return SDValue();
-
// Get the desired 128-bit vector half.
- unsigned IdxVal = cast<ConstantSDNode>(N2)->getZExtValue();
SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
// Insert the element into the desired half.
- unsigned NumEltsIn128 = 128/EltVT.getSizeInBits();
- unsigned IdxIn128 = IdxVal - (IdxVal/NumEltsIn128) * NumEltsIn128;
+ unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
+ unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
DAG.getConstant(IdxIn128, MVT::i32));
// Insert the changed part back to the 256-bit vector
return Insert128BitVector(N0, V, IdxVal, DAG, dl);
}
+ assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
- if (Subtarget->hasSSE41())
- return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
+ if (Subtarget->hasSSE41()) {
+ if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
+ unsigned Opc;
+ if (VT == MVT::v8i16) {
+ Opc = X86ISD::PINSRW;
+ } else {
+ assert(VT == MVT::v16i8);
+ Opc = X86ISD::PINSRB;
+ }
+
+ // Transform it so it match pinsr{b,w} which expects a GR32 as its second
+ // argument.
+ if (N1.getValueType() != MVT::i32)
+ N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
+ if (N2.getValueType() != MVT::i32)
+ N2 = DAG.getIntPtrConstant(IdxVal);
+ return DAG.getNode(Opc, dl, VT, N0, N1, N2);
+ }
+
+ if (EltVT == MVT::f32) {
+ // Bits [7:6] of the constant are the source select. This will always be
+ // zero here. The DAG Combiner may combine an extract_elt index into
+ // these
+ // bits. For example (insert (extract, 3), 2) could be matched by
+ // putting
+ // the '3' into bits [7:6] of X86ISD::INSERTPS.
+ // Bits [5:4] of the constant are the destination select. This is the
+ // value of the incoming immediate.
+ // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
+ // combine either bitwise AND or insert of float 0.0 to set these bits.
+ N2 = DAG.getIntPtrConstant(IdxVal << 4);
+ // Create this as a scalar to vector..
+ N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
+ return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
+ }
+
+ if (EltVT == MVT::i32 || EltVT == MVT::i64) {
+ // PINSR* works with constant index.
+ return Op;
+ }
+ }
if (EltVT == MVT::i8)
return SDValue();
- if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
+ if (EltVT.getSizeInBits() == 16) {
// Transform it so it match pinsrw which expects a 16-bit value in a GR32
// as its second argument.
if (N1.getValueType() != MVT::i32)
N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
if (N2.getValueType() != MVT::i32)
- N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
+ N2 = DAG.getIntPtrConstant(IdxVal);
return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
}
return SDValue();
In, DAG.getUNDEF(SVT)));
}
-static SDValue LowerFABS(SDValue Op, SelectionDAG &DAG) {
- LLVMContext *Context = DAG.getContext();
- SDLoc dl(Op);
- MVT VT = Op.getSimpleValueType();
- MVT EltVT = VT;
- unsigned NumElts = VT == MVT::f64 ? 2 : 4;
- if (VT.isVector()) {
- EltVT = VT.getVectorElementType();
- NumElts = VT.getVectorNumElements();
- }
- Constant *C;
- if (EltVT == MVT::f64)
- C = ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
- APInt(64, ~(1ULL << 63))));
- else
- C = ConstantFP::get(*Context, APFloat(APFloat::IEEEsingle,
- APInt(32, ~(1U << 31))));
- C = ConstantVector::getSplat(NumElts, C);
- const TargetLowering &TLI = DAG.getTargetLoweringInfo();
- SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy());
- unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
- SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
- MachinePointerInfo::getConstantPool(),
- false, false, false, Alignment);
- if (VT.isVector()) {
- MVT ANDVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
- return DAG.getNode(ISD::BITCAST, dl, VT,
- DAG.getNode(ISD::AND, dl, ANDVT,
- DAG.getNode(ISD::BITCAST, dl, ANDVT,
- Op.getOperand(0)),
- DAG.getNode(ISD::BITCAST, dl, ANDVT, Mask)));
- }
- return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
-}
+// The only differences between FABS and FNEG are the mask and the logic op.
+static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
+ assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
+ "Wrong opcode for lowering FABS or FNEG.");
-static SDValue LowerFNEG(SDValue Op, SelectionDAG &DAG) {
- LLVMContext *Context = DAG.getContext();
+ bool IsFABS = (Op.getOpcode() == ISD::FABS);
SDLoc dl(Op);
MVT VT = Op.getSimpleValueType();
+ // Assume scalar op for initialization; update for vector if needed.
+ // Note that there are no scalar bitwise logical SSE/AVX instructions, so we
+ // generate a 16-byte vector constant and logic op even for the scalar case.
+ // Using a 16-byte mask allows folding the load of the mask with
+ // the logic op, so it can save (~4 bytes) on code size.
MVT EltVT = VT;
unsigned NumElts = VT == MVT::f64 ? 2 : 4;
+ // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
+ // decide if we should generate a 16-byte constant mask when we only need 4 or
+ // 8 bytes for the scalar case.
if (VT.isVector()) {
EltVT = VT.getVectorElementType();
NumElts = VT.getVectorNumElements();
}
- Constant *C;
- if (EltVT == MVT::f64)
- C = ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
- APInt(64, 1ULL << 63)));
- else
- C = ConstantFP::get(*Context, APFloat(APFloat::IEEEsingle,
- APInt(32, 1U << 31)));
+
+ unsigned EltBits = EltVT.getSizeInBits();
+ LLVMContext *Context = DAG.getContext();
+ // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
+ APInt MaskElt =
+ IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
+ Constant *C = ConstantInt::get(*Context, MaskElt);
C = ConstantVector::getSplat(NumElts, C);
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy());
SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
MachinePointerInfo::getConstantPool(),
false, false, false, Alignment);
+
if (VT.isVector()) {
- MVT XORVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits()/64);
+ // For a vector, cast operands to a vector type, perform the logic op,
+ // and cast the result back to the original value type.
+ MVT VecVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
+ SDValue Op0Casted = DAG.getNode(ISD::BITCAST, dl, VecVT, Op.getOperand(0));
+ SDValue MaskCasted = DAG.getNode(ISD::BITCAST, dl, VecVT, Mask);
+ unsigned LogicOp = IsFABS ? ISD::AND : ISD::XOR;
return DAG.getNode(ISD::BITCAST, dl, VT,
- DAG.getNode(ISD::XOR, dl, XORVT,
- DAG.getNode(ISD::BITCAST, dl, XORVT,
- Op.getOperand(0)),
- DAG.getNode(ISD::BITCAST, dl, XORVT, Mask)));
+ DAG.getNode(LogicOp, dl, VecVT, Op0Casted, MaskCasted));
}
-
- return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
+ // If not vector, then scalar.
+ unsigned LogicOp = IsFABS ? X86ISD::FAND : X86ISD::FXOR;
+ return DAG.getNode(LogicOp, dl, VT, Op.getOperand(0), Mask);
}
static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
MVT VT = Op.getSimpleValueType();
SDLoc dl(Op);
- assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 32 &&
+ assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
Op.getValueType().getScalarType() == MVT::i1 &&
"Cannot set masked compare for this operation");
EVT OpVT = Op1.getValueType();
if (Subtarget->hasAVX512()) {
if (Op1.getValueType().is512BitVector() ||
+ (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
(MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
// In AVX-512 architecture setcc returns mask with i1 elements,
- // But there is no compare instruction for i8 and i16 elements.
+ // But there is no compare instruction for i8 and i16 elements in KNL.
// We are not talking about 512-bit operands in this case, these
// types are illegal.
if (MaskResult &&
SDLoc dl(Op);
unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
- const IntrinsicData* IntrData = GetIntrinsicWithoutChain(IntNo);
+ const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
if (IntrData) {
switch(IntrData->Type) {
case INTR_TYPE_1OP:
SelectionDAG &DAG) {
unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
- const IntrinsicData* IntrData = GetIntrinsicWithChain(IntNo);
+ const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
if (!IntrData)
return SDValue();
return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
}
+// Sign extension of the low part of vector elements. This may be used either
+// when sign extend instructions are not available or if the vector element
+// sizes already match the sign-extended size. If the vector elements are in
+// their pre-extended size and sign extend instructions are available, that will
+// be handled by LowerSIGN_EXTEND.
SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
SelectionDAG &DAG) const {
SDLoc dl(Op);
case MVT::v4i32:
case MVT::v8i16: {
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) {
- // (sext (vzext x)) -> (vsext x)
- Tmp1 = LowerVectorIntExtend(Op00, Subtarget, DAG);
- if (Tmp1.getNode()) {
- EVT ExtraEltVT = ExtraVT.getVectorElementType();
- // This folding is only valid when the in-reg type is a vector of i8,
- // i16, or i32.
- if (ExtraEltVT == MVT::i8 || ExtraEltVT == MVT::i16 ||
- ExtraEltVT == MVT::i32) {
- 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));
- }
- Op0 = Tmp1;
- }
- }
- // If the above didn't work, then just use Shift-Left + Shift-Right.
- Tmp1 = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, Op0, BitsDiff,
- DAG);
- return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, Tmp1, BitsDiff,
+ // This is a sign extension of some low part of vector elements without
+ // changing the size of the vector elements themselves:
+ // Shift-Left + Shift-Right-Algebraic.
+ SDValue Shl = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, Op0,
+ BitsDiff, DAG);
+ return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, Shl, BitsDiff,
DAG);
}
}
case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
- case ISD::FABS: return LowerFABS(Op, DAG);
- case ISD::FNEG: return LowerFNEG(Op, DAG);
+ case ISD::FABS:
+ case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
case ISD::SETCC: return LowerSETCC(Op, DAG);
if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
CondVT.getVectorElementType() == MVT::i1) {
// v16i8 (select v16i1, v16i8, v16i8) does not have a proper
- // lowering on AVX-512. In this case we convert it to
+ // lowering on KNL. In this case we convert it to
// v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
- // The same situation for all 128 and 256-bit vectors of i8 and i16
+ // The same situation for all 128 and 256-bit vectors of i8 and i16.
+ // Since SKX these selects have a proper lowering.
EVT OpVT = LHS.getValueType();
if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
(OpVT.getVectorElementType() == MVT::i8 ||
- OpVT.getVectorElementType() == MVT::i16)) {
+ OpVT.getVectorElementType() == MVT::i16) &&
+ !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
DCI.AddToWorklist(Cond.getNode());
return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);