#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Target/TargetOptions.h"
+#include "X86IntrinsicsInfo.h"
#include <bitset>
#include <numeric>
#include <cctype>
else
setSchedulingPreference(Sched::RegPressure);
const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(TM.getRegisterInfo());
+ TM.getSubtarget<X86Subtarget>().getRegisterInfo();
setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
// Bypass expensive divides on Atom when compiling with O2
setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
- setOperationAction(ISD::DYNAMIC_STACKALLOC, Subtarget->is64Bit() ?
- MVT::i64 : MVT::i32, Custom);
+ setOperationAction(ISD::DYNAMIC_STACKALLOC, getPointerTy(), Custom);
if (!TM.Options.UseSoftFloat && X86ScalarSSEf64) {
// f32 and f64 use SSE.
if (!TM.Options.UseSoftFloat && Subtarget->hasSSE2()) {
addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
- // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
+ // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
// registers cannot be used even for integer operations.
addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
}
+ // We support custom legalizing of sext and anyext loads for specific
+ // memory vector types which we can load as a scalar (or sequence of
+ // scalars) and extend in-register to a legal 128-bit vector type. For sext
+ // loads these must work with a single scalar load.
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Custom);
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Custom);
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v8i8, Custom);
+ setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Custom);
+ setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Custom);
+ setLoadExtAction(ISD::EXTLOAD, MVT::v2i32, Custom);
+ setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Custom);
+ setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Custom);
+ setLoadExtAction(ISD::EXTLOAD, MVT::v8i8, Custom);
+
setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
// some vselects for now.
setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
- // i8 and i16 vectors are custom , because the source register and source
+ // SSE41 brings specific instructions for doing vector sign extend even in
+ // cases where we don't have SRA.
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Custom);
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Custom);
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v2i32, Custom);
+
+ // i8 and i16 vectors are custom because the source register and source
// source memory operand types are not the same width. f32 vectors are
// custom since the immediate controlling the insert encodes additional
// information.
setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
- // FIXME: these should be Legal but thats only for the case where
+ // FIXME: these should be Legal, but that's only for the case where
// the index is constant. For now custom expand to deal with that.
if (Subtarget->is64Bit()) {
setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
}// 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
PredictableSelectIsExpensive = !Subtarget->isAtom();
setPrefFunctionAlignment(4); // 2^4 bytes.
+
+ verifyIntrinsicTables();
+}
+
+// This has so far only been implemented for 64-bit MachO.
+bool X86TargetLowering::useLoadStackGuardNode() const {
+ return Subtarget->getTargetTriple().getObjectFormat() == Triple::MachO &&
+ Subtarget->is64Bit();
}
TargetLoweringBase::LegalizeTypeAction
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();
}
bool
-X86TargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
- unsigned,
- bool *Fast) const {
+X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
+ unsigned,
+ unsigned,
+ bool *Fast) const {
if (Fast)
*Fast = Subtarget->isUnalignedMemAccessFast();
return true;
default:
return TargetLowering::findRepresentativeClass(VT);
case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
- RRC = Subtarget->is64Bit() ?
- (const TargetRegisterClass*)&X86::GR64RegClass :
- (const TargetRegisterClass*)&X86::GR32RegClass;
+ RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
break;
case MVT::x86mmx:
RRC = &X86::VR64RegClass;
const SmallVectorImpl<ISD::OutputArg> &Outs,
LLVMContext &Context) const {
SmallVector<CCValAssign, 16> RVLocs;
- CCState CCInfo(CallConv, isVarArg, MF, MF.getTarget(),
- RVLocs, Context);
+ CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
return CCInfo.CheckReturn(Outs, RetCC_X86);
}
X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
SmallVector<CCValAssign, 16> RVLocs;
- CCState CCInfo(CallConv, isVarArg, MF, DAG.getTarget(),
- RVLocs, *DAG.getContext());
+ CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
CCInfo.AnalyzeReturn(Outs, RetCC_X86);
SDValue Flag;
// Returns in ST0/ST1 are handled specially: these are pushed as operands to
// the RET instruction and handled by the FP Stackifier.
- if (VA.getLocReg() == X86::ST0 ||
- VA.getLocReg() == X86::ST1) {
+ if (VA.getLocReg() == X86::FP0 ||
+ VA.getLocReg() == X86::FP1) {
// If this is a copy from an xmm register to ST(0), use an FPExtend to
// change the value to the FP stack register class.
if (isScalarFPTypeInSSEReg(VA.getValVT()))
UI != UE; ++UI) {
if (UI->getOpcode() != X86ISD::RET_FLAG)
return false;
+ // If we are returning more than one value, we can definitely
+ // not make a tail call see PR19530
+ if (UI->getNumOperands() > 4)
+ return false;
+ if (UI->getNumOperands() == 4 &&
+ UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
+ return false;
HasRet = true;
}
return true;
}
-MVT
-X86TargetLowering::getTypeForExtArgOrReturn(MVT VT,
+EVT
+X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
ISD::NodeType ExtendKind) const {
MVT ReturnMVT;
// TODO: Is this also valid on 32-bit?
else
ReturnMVT = MVT::i32;
- MVT MinVT = getRegisterType(ReturnMVT);
+ EVT MinVT = getRegisterType(Context, ReturnMVT);
return VT.bitsLT(MinVT) ? MinVT : VT;
}
// Assign locations to each value returned by this call.
SmallVector<CCValAssign, 16> RVLocs;
bool Is64Bit = Subtarget->is64Bit();
- CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
- DAG.getTarget(), RVLocs, *DAG.getContext());
+ CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
+ *DAG.getContext());
CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
// Copy all of the result registers out of their specified physreg.
report_fatal_error("SSE register return with SSE disabled");
}
- SDValue Val;
-
- // If this is a call to a function that returns an fp value on the floating
- // point stack, we must guarantee the value is popped from the stack, so
- // a CopyFromReg is not good enough - the copy instruction may be eliminated
- // if the return value is not used. We use the FpPOP_RETVAL instruction
- // instead.
- if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
- // If we prefer to use the value in xmm registers, copy it out as f80 and
- // use a truncate to move it from fp stack reg to xmm reg.
- if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
- SDValue Ops[] = { Chain, InFlag };
- Chain = SDValue(DAG.getMachineNode(X86::FpPOP_RETVAL, dl, CopyVT,
- MVT::Other, MVT::Glue, Ops), 1);
- Val = Chain.getValue(0);
-
- // Round the f80 to the right size, which also moves it to the appropriate
- // xmm register.
- if (CopyVT != VA.getValVT())
- Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
- // This truncation won't change the value.
- DAG.getIntPtrConstant(1));
- } else {
- Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
- CopyVT, InFlag).getValue(1);
- Val = Chain.getValue(0);
- }
+ // If we prefer to use the value in xmm registers, copy it out as f80 and
+ // use a truncate to move it from fp stack reg to xmm reg.
+ if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
+ isScalarFPTypeInSSEReg(VA.getValVT()))
+ CopyVT = MVT::f80;
+
+ Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
+ CopyVT, InFlag).getValue(1);
+ SDValue Val = Chain.getValue(0);
+
+ if (CopyVT != VA.getValVT())
+ Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
+ // This truncation won't change the value.
+ DAG.getIntPtrConstant(1));
+
InFlag = Chain.getValue(2);
InVals.push_back(Val);
}
}
}
+// 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,
// Assign locations to all of the incoming arguments.
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(CallConv, isVarArg, MF, DAG.getTarget(),
- ArgLocs, *DAG.getContext());
+ CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
// Allocate shadow area for Win64
if (IsWin64)
StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
// 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.
- if (isVarArg) {
- 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);
+ // 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 (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.getTarget().getFrameLowering();
+ const TargetFrameLowering &TFI = *MF.getSubtarget().getFrameLowering();
// Get to the caller-allocated home save location. Add 8 to account
// for the return address.
int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
// 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, 11> SaveXMMOps;
+ 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;
// Analyze operands of the call, assigning locations to each operand.
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(CallConv, isVarArg, MF, MF.getTarget(),
- ArgLocs, *DAG.getContext());
+ CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
// Allocate shadow area for Win64
if (IsWin64)
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;
// Walk the register/memloc assignments, inserting copies/loads. In the case
// of tail call optimization arguments are handle later.
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ DAG.getSubtarget().getRegisterInfo());
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
// Skip inalloca arguments, they have already been written.
ISD::ArgFlagsTy Flags = Outs[i].Flags;
}
}
- 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.
RegsToPass[i].second.getValueType()));
// Add a register mask operand representing the call-preserved registers.
- const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
+ const TargetRegisterInfo *TRI = DAG.getSubtarget().getRegisterInfo();
const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
assert(Mask && "Missing call preserved mask for calling convention");
Ops.push_back(DAG.getRegisterMask(Mask));
SelectionDAG& DAG) const {
MachineFunction &MF = DAG.getMachineFunction();
const TargetMachine &TM = MF.getTarget();
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(TM.getRegisterInfo());
- const TargetFrameLowering &TFI = *TM.getFrameLowering();
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ TM.getSubtargetImpl()->getRegisterInfo());
+ const TargetFrameLowering &TFI = *TM.getSubtargetImpl()->getFrameLowering();
unsigned StackAlignment = TFI.getStackAlignment();
uint64_t AlignMask = StackAlignment - 1;
int64_t Offset = StackSize;
// Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
// emit a special epilogue.
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ DAG.getSubtarget().getRegisterInfo());
if (RegInfo->needsStackRealignment(MF))
return false;
return false;
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
- DAG.getTarget(), ArgLocs, *DAG.getContext());
+ CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
+ *DAG.getContext());
CCInfo.AnalyzeCallOperands(Outs, CC_X86);
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
}
if (Unused) {
SmallVector<CCValAssign, 16> RVLocs;
- CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(),
- DAG.getTarget(), RVLocs, *DAG.getContext());
+ CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
+ *DAG.getContext());
CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
CCValAssign &VA = RVLocs[i];
- if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
+ if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
return false;
}
}
// results are returned in the same way as what the caller expects.
if (!CCMatch) {
SmallVector<CCValAssign, 16> RVLocs1;
- CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
- DAG.getTarget(), RVLocs1, *DAG.getContext());
+ CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
+ *DAG.getContext());
CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
SmallVector<CCValAssign, 16> RVLocs2;
- CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
- DAG.getTarget(), RVLocs2, *DAG.getContext());
+ CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
+ *DAG.getContext());
CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
if (RVLocs1.size() != RVLocs2.size())
// Check if stack adjustment is needed. For now, do not do this if any
// argument is passed on the stack.
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
- DAG.getTarget(), ArgLocs, *DAG.getContext());
+ CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
+ *DAG.getContext());
// Allocate shadow area for Win64
if (IsCalleeWin64)
MachineFrameInfo *MFI = MF.getFrameInfo();
const MachineRegisterInfo *MRI = &MF.getRegInfo();
const X86InstrInfo *TII =
- static_cast<const X86InstrInfo *>(DAG.getTarget().getInstrInfo());
+ static_cast<const X86InstrInfo *>(DAG.getSubtarget().getInstrInfo());
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
CCValAssign &VA = ArgLocs[i];
SDValue Arg = OutVals[i];
static bool isTargetShuffle(unsigned Opcode) {
switch(Opcode) {
default: return false;
+ case X86ISD::PSHUFB:
case X86ISD::PSHUFD:
case X86ISD::PSHUFHW:
case X86ISD::PSHUFLW:
switch(Opc) {
default: llvm_unreachable("Unknown x86 shuffle node");
case X86ISD::PALIGNR:
+ case X86ISD::VALIGN:
case X86ISD::SHUFP:
case X86ISD::VPERM2X128:
return DAG.getNode(Opc, dl, VT, V1, V2,
SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
MachineFunction &MF = DAG.getMachineFunction();
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ DAG.getSubtarget().getRegisterInfo());
X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
int ReturnAddrIndex = FuncInfo->getRAIndex();
/// 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;
}
}
return true;
}
-/// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
-/// is suitable for input to PALIGNR.
-static bool isPALIGNRMask(ArrayRef<int> Mask, MVT VT,
- const X86Subtarget *Subtarget) {
- if ((VT.is128BitVector() && !Subtarget->hasSSSE3()) ||
- (VT.is256BitVector() && !Subtarget->hasInt256()))
- return false;
-
+/// \brief Return true if the mask specifies a shuffle of elements that is
+/// suitable for input to intralane (palignr) or interlane (valign) vector
+/// right-shift.
+static bool isAlignrMask(ArrayRef<int> Mask, MVT VT, bool InterLane) {
unsigned NumElts = VT.getVectorNumElements();
- unsigned NumLanes = VT.is512BitVector() ? 1: VT.getSizeInBits()/128;
+ unsigned NumLanes = InterLane ? 1: VT.getSizeInBits()/128;
unsigned NumLaneElts = NumElts/NumLanes;
// Do not handle 64-bit element shuffles with palignr.
return true;
}
+/// \brief Return true if the node specifies a shuffle of elements that is
+/// suitable for input to PALIGNR.
+static bool isPALIGNRMask(ArrayRef<int> Mask, MVT VT,
+ const X86Subtarget *Subtarget) {
+ if ((VT.is128BitVector() && !Subtarget->hasSSSE3()) ||
+ (VT.is256BitVector() && !Subtarget->hasInt256()) ||
+ VT.is512BitVector())
+ // FIXME: Add AVX512BW.
+ return false;
+
+ return isAlignrMask(Mask, VT, false);
+}
+
+/// \brief Return true if the node specifies a shuffle of elements that is
+/// suitable for input to VALIGN.
+static bool isVALIGNMask(ArrayRef<int> Mask, MVT VT,
+ const X86Subtarget *Subtarget) {
+ // FIXME: Add AVX512VL.
+ if (!VT.is512BitVector() || !Subtarget->hasAVX512())
+ return false;
+ return isAlignrMask(Mask, VT, true);
+}
+
/// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
/// the two vector operands have swapped position.
static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask,
return Mask;
}
-/// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
-/// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
-static unsigned getShufflePALIGNRImmediate(ShuffleVectorSDNode *SVOp) {
+/// \brief Return the appropriate immediate to shuffle the specified
+/// VECTOR_SHUFFLE mask with the PALIGNR (if InterLane is false) or with
+/// VALIGN (if Interlane is true) instructions.
+static unsigned getShuffleAlignrImmediate(ShuffleVectorSDNode *SVOp,
+ bool InterLane) {
MVT VT = SVOp->getSimpleValueType(0);
- unsigned EltSize = VT.is512BitVector() ? 1 :
+ unsigned EltSize = InterLane ? 1 :
VT.getVectorElementType().getSizeInBits() >> 3;
unsigned NumElts = VT.getVectorNumElements();
return (Val - i) * EltSize;
}
+/// \brief Return the appropriate immediate to shuffle the specified
+/// VECTOR_SHUFFLE mask with the PALIGNR instruction.
+static unsigned getShufflePALIGNRImmediate(ShuffleVectorSDNode *SVOp) {
+ return getShuffleAlignrImmediate(SVOp, false);
+}
+
+/// \brief Return the appropriate immediate to shuffle the specified
+/// VECTOR_SHUFFLE mask with the VALIGN instruction.
+static unsigned getShuffleVALIGNImmediate(ShuffleVectorSDNode *SVOp) {
+ return getShuffleAlignrImmediate(SVOp, true);
+}
+
+
static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
}
/// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
-/// target specific opcode. Returns true if the Mask could be calculated.
-/// Sets IsUnary to true if only uses one source.
+/// target specific opcode. Returns true if the Mask could be calculated. Sets
+/// IsUnary to true if only uses one source. Note that this will set IsUnary for
+/// shuffles which use a single input multiple times, and in those cases it will
+/// adjust the mask to only have indices within that single input.
static bool getTargetShuffleMask(SDNode *N, MVT VT,
SmallVectorImpl<int> &Mask, bool &IsUnary) {
unsigned NumElems = VT.getVectorNumElements();
SDValue ImmN;
IsUnary = false;
+ bool IsFakeUnary = false;
switch(N->getOpcode()) {
case X86ISD::SHUFP:
ImmN = N->getOperand(N->getNumOperands()-1);
DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
+ IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
break;
case X86ISD::UNPCKH:
DecodeUNPCKHMask(VT, Mask);
+ IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
break;
case X86ISD::UNPCKL:
DecodeUNPCKLMask(VT, Mask);
+ IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
break;
case X86ISD::MOVHLPS:
DecodeMOVHLPSMask(NumElems, Mask);
+ IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
break;
case X86ISD::MOVLHPS:
DecodeMOVLHPSMask(NumElems, Mask);
+ IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
break;
case X86ISD::PALIGNR:
ImmN = N->getOperand(N->getNumOperands()-1);
DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
IsUnary = true;
break;
+ case X86ISD::PSHUFB: {
+ IsUnary = true;
+ SDValue MaskNode = N->getOperand(1);
+ while (MaskNode->getOpcode() == ISD::BITCAST)
+ MaskNode = MaskNode->getOperand(0);
+
+ if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
+ // If we have a build-vector, then things are easy.
+ EVT VT = MaskNode.getValueType();
+ assert(VT.isVector() &&
+ "Can't produce a non-vector with a build_vector!");
+ if (!VT.isInteger())
+ return false;
+
+ int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
+
+ SmallVector<uint64_t, 32> RawMask;
+ for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
+ auto *CN = dyn_cast<ConstantSDNode>(MaskNode->getOperand(i));
+ if (!CN)
+ return false;
+ APInt MaskElement = CN->getAPIntValue();
+
+ // We now have to decode the element which could be any integer size and
+ // extract each byte of it.
+ for (int j = 0; j < NumBytesPerElement; ++j) {
+ // Note that this is x86 and so always little endian: the low byte is
+ // the first byte of the mask.
+ RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
+ MaskElement = MaskElement.lshr(8);
+ }
+ }
+ DecodePSHUFBMask(RawMask, Mask);
+ break;
+ }
+
+ auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
+ if (!MaskLoad)
+ return false;
+
+ SDValue Ptr = MaskLoad->getBasePtr();
+ if (Ptr->getOpcode() == X86ISD::Wrapper)
+ Ptr = Ptr->getOperand(0);
+
+ auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
+ if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
+ return false;
+
+ if (auto *C = dyn_cast<ConstantDataSequential>(MaskCP->getConstVal())) {
+ // FIXME: Support AVX-512 here.
+ if (!C->getType()->isVectorTy() ||
+ (C->getNumElements() != 16 && C->getNumElements() != 32))
+ return false;
+
+ assert(C->getType()->isVectorTy() && "Expected a vector constant.");
+ DecodePSHUFBMask(C, Mask);
+ break;
+ }
+
+ return false;
+ }
case X86ISD::VPERMI:
ImmN = N->getOperand(N->getNumOperands()-1);
DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
default: llvm_unreachable("unknown target shuffle node");
}
+ // If we have a fake unary shuffle, the shuffle mask is spread across two
+ // inputs that are actually the same node. Re-map the mask to always point
+ // into the first input.
+ if (IsFakeUnary)
+ for (int &M : Mask)
+ if (M >= (int)Mask.size())
+ M -= Mask.size();
+
return true;
}
// convert it to a vector with movd (S2V+shuffle to zero extend).
Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
+
+ // If using the new shuffle lowering, just directly insert this.
+ if (ExperimentalVectorShuffleLowering)
+ return DAG.getNode(
+ ISD::BITCAST, dl, VT,
+ getShuffleVectorZeroOrUndef(Item, Idx * 2, true, Subtarget, DAG));
+
Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
// Now we have our 32-bit value zero extended in the low element of
if (EVTBits == 32) {
Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
+ // If using the new shuffle lowering, just directly insert this.
+ if (ExperimentalVectorShuffleLowering)
+ return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
+
// Turn it into a shuffle of zero and zero-extended scalar to vector.
Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0, Subtarget, DAG);
SmallVector<int, 8> MaskVec;
return true;
}
+// Hide this symbol with an anonymous namespace instead of 'static' so that MSVC
+// 2013 will allow us to use it as a non-type template parameter.
+namespace {
+
+/// \brief Implementation of the \c isShuffleEquivalent variadic functor.
+///
+/// See its documentation for details.
+bool isShuffleEquivalentImpl(ArrayRef<int> Mask, ArrayRef<const int *> Args) {
+ if (Mask.size() != Args.size())
+ return false;
+ for (int i = 0, e = Mask.size(); i < e; ++i) {
+ assert(*Args[i] >= 0 && "Arguments must be positive integers!");
+ assert(*Args[i] < (int)Args.size() * 2 &&
+ "Argument outside the range of possible shuffle inputs!");
+ if (Mask[i] != -1 && Mask[i] != *Args[i])
+ return false;
+ }
+ return true;
+}
+
+} // namespace
+
+/// \brief Checks whether a shuffle mask is equivalent to an explicit list of
+/// arguments.
+///
+/// This is a fast way to test a shuffle mask against a fixed pattern:
+///
+/// if (isShuffleEquivalent(Mask, 3, 2, 1, 0)) { ... }
+///
+/// It returns true if the mask is exactly as wide as the argument list, and
+/// each element of the mask is either -1 (signifying undef) or the value given
+/// in the argument.
+static const VariadicFunction1<
+ bool, ArrayRef<int>, int, isShuffleEquivalentImpl> isShuffleEquivalent = {};
+
/// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
///
/// This helper function produces an 8-bit shuffle immediate corresponding to
assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
assert(Mask[1] >= 2 && "Non-canonicalized blend!");
+ // Use dedicated unpack instructions for masks that match their pattern.
+ if (isShuffleEquivalent(Mask, 0, 2))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
+ if (isShuffleEquivalent(Mask, 1, 3))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
+
unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V2,
DAG.getConstant(SHUFPDMask, MVT::i8));
getV4X86ShuffleImm8ForMask(WidenedMask, DAG)));
}
+ // Use dedicated unpack instructions for masks that match their pattern.
+ if (isShuffleEquivalent(Mask, 0, 2))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
+ if (isShuffleEquivalent(Mask, 1, 3))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
+
// We implement this with SHUFPD which is pretty lame because it will likely
// incur 2 cycles of stall for integer vectors on Nehalem and older chips.
// However, all the alternatives are still more cycles and newer chips don't
return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
getV4X86ShuffleImm8ForMask(Mask, DAG));
+ // Use dedicated unpack instructions for masks that match their pattern.
+ if (isShuffleEquivalent(Mask, 0, 4, 1, 5))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
+ if (isShuffleEquivalent(Mask, 2, 6, 3, 7))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
+
if (NumV2Elements == 1) {
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;
getV4X86ShuffleImm8ForMask(NewMask, DAG));
}
+static SDValue lowerIntegerElementInsertionVectorShuffle(
+ MVT VT, SDLoc DL, SDValue V1, SDValue V2, ArrayRef<int> Mask,
+ const X86Subtarget *Subtarget, SelectionDAG &DAG) {
+ int V2Index = std::find_if(Mask.begin(), Mask.end(),
+ [&Mask](int M) { return M >= (int)Mask.size(); }) -
+ Mask.begin();
+
+ // Check for a single input from a SCALAR_TO_VECTOR node.
+ // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
+ // all the smarts here sunk into that routine. However, the current
+ // lowering of BUILD_VECTOR makes that nearly impossible until the old
+ // vector shuffle lowering is dead.
+ if ((Mask[V2Index] == (int)Mask.size() &&
+ V2.getOpcode() == ISD::SCALAR_TO_VECTOR) ||
+ V2.getOpcode() == ISD::BUILD_VECTOR) {
+ SDValue V2S = V2.getOperand(Mask[V2Index] - Mask.size());
+
+ bool V1IsAllZero = false;
+ if (ISD::isBuildVectorAllZeros(V1.getNode())) {
+ V1IsAllZero = true;
+ } else if (V1.getOpcode() == ISD::BUILD_VECTOR) {
+ V1IsAllZero = true;
+ for (int M : Mask) {
+ if (M < 0 || M >= (int)Mask.size())
+ continue;
+ SDValue Input = V1.getOperand(M);
+ if (Input.getOpcode() != ISD::UNDEF && !X86::isZeroNode(Input)) {
+ // A non-zero input!
+ V1IsAllZero = false;
+ break;
+ }
+ }
+ }
+ if (V1IsAllZero) {
+ // First, we need to zext the scalar if it is smaller than an i32.
+ MVT EltVT = VT.getVectorElementType();
+ assert(EltVT == V2S.getSimpleValueType() &&
+ "Different scalar and element types!");
+ MVT ExtVT = VT;
+ if (EltVT == MVT::i8 || EltVT == MVT::i16) {
+ // Zero-extend directly to i32.
+ ExtVT = MVT::v4i32;
+ V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
+ }
+
+ V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT,
+ DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S));
+ if (ExtVT != VT)
+ V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
+
+ if (V2Index != 0) {
+ // If we have 4 or fewer lanes we can cheaply shuffle the element into
+ // the desired position. Otherwise it is more efficient to do a vector
+ // shift left. We know that we can do a vector shift left because all
+ // the inputs are zero.
+ if (VT.getVectorNumElements() <= 4) {
+ SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
+ V2Shuffle[V2Index] = 0;
+ V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
+ } else {
+ V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V2);
+ V2 = DAG.getNode(
+ X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
+ DAG.getConstant(
+ V2Index * EltVT.getSizeInBits(),
+ DAG.getTargetLoweringInfo().getScalarShiftAmountTy(MVT::v2i64)));
+ V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
+ }
+ }
+ return V2;
+ }
+ }
+ return SDValue();
+}
+
/// \brief Lower 4-lane i32 vector shuffles.
///
/// We try to handle these with integer-domain shuffles where we can, but for
ArrayRef<int> Mask = SVOp->getMask();
assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
- if (isSingleInputShuffleMask(Mask))
+ int NumV2Elements =
+ std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
+
+ if (NumV2Elements == 0)
// Straight shuffle of a single input vector. For everything from SSE2
// onward this has a single fast instruction with no scary immediates.
return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
getV4X86ShuffleImm8ForMask(Mask, DAG));
+ // Use dedicated unpack instructions for masks that match their pattern.
+ if (isShuffleEquivalent(Mask, 0, 4, 1, 5))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
+ if (isShuffleEquivalent(Mask, 2, 6, 3, 7))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
+
+ // There are special ways we can lower some single-element blends.
+ if (NumV2Elements == 1)
+ if (SDValue V = lowerIntegerElementInsertionVectorShuffle(
+ MVT::v4i32, DL, V1, V2, Mask, Subtarget, DAG))
+ return V;
+
// We implement this with SHUFPS because it can blend from two vectors.
// Because we're going to eventually use SHUFPS, we use SHUFPS even to build
// up the inputs, bypassing domain shift penalties that we would encur if we
// Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
// Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
//
- // Before we had 3-1 in the low half and 3-1 in the high half. Afterward, 2-2
- // and 2-2.
- auto balanceSides = [&](ArrayRef<int> ThreeInputs, int OneInput,
- int ThreeInputHalfSum, int OneInputHalfOffset) {
+ // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
+ // and an existing 2-into-2 on the other half. In this case we may have to
+ // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
+ // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
+ // Fortunately, we don't have to handle anything but a 2-into-2 pattern
+ // because any other situation (including a 3-into-1 or 1-into-3 in the other
+ // half than the one we target for fixing) will be fixed when we re-enter this
+ // path. We will also combine away any sequence of PSHUFD instructions that
+ // result into a single instruction. Here is an example of the tricky case:
+ //
+ // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
+ // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
+ //
+ // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
+ //
+ // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
+ // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
+ //
+ // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
+ // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
+ //
+ // The result is fine to be handled by the generic logic.
+ auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
+ ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
+ int AOffset, int BOffset) {
+ assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
+ "Must call this with A having 3 or 1 inputs from the A half.");
+ assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
+ "Must call this with B having 1 or 3 inputs from the B half.");
+ assert(AToAInputs.size() + BToAInputs.size() == 4 &&
+ "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
+
// Compute the index of dword with only one word among the three inputs in
// a half by taking the sum of the half with three inputs and subtracting
// the sum of the actual three inputs. The difference is the remaining
// slot.
- int DWordA = (ThreeInputHalfSum -
- std::accumulate(ThreeInputs.begin(), ThreeInputs.end(), 0)) /
- 2;
- int DWordB = OneInputHalfOffset / 2 + (OneInput / 2 + 1) % 2;
+ int ADWord, BDWord;
+ int &TripleDWord = AToAInputs.size() == 3 ? ADWord : BDWord;
+ int &OneInputDWord = AToAInputs.size() == 3 ? BDWord : ADWord;
+ int TripleInputOffset = AToAInputs.size() == 3 ? AOffset : BOffset;
+ ArrayRef<int> TripleInputs = AToAInputs.size() == 3 ? AToAInputs : BToAInputs;
+ int OneInput = AToAInputs.size() == 3 ? BToAInputs[0] : AToAInputs[0];
+ int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
+ int TripleNonInputIdx =
+ TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
+ TripleDWord = TripleNonInputIdx / 2;
+
+ // We use xor with one to compute the adjacent DWord to whichever one the
+ // OneInput is in.
+ OneInputDWord = (OneInput / 2) ^ 1;
+
+ // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
+ // and BToA inputs. If there is also such a problem with the BToB and AToB
+ // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
+ // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
+ // is essential that we don't *create* a 3<-1 as then we might oscillate.
+ if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
+ // Compute how many inputs will be flipped by swapping these DWords. We
+ // need
+ // to balance this to ensure we don't form a 3-1 shuffle in the other
+ // half.
+ int NumFlippedAToBInputs =
+ std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
+ std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
+ int NumFlippedBToBInputs =
+ std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
+ std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
+ if ((NumFlippedAToBInputs == 1 &&
+ (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
+ (NumFlippedBToBInputs == 1 &&
+ (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
+ // We choose whether to fix the A half or B half based on whether that
+ // half has zero flipped inputs. At zero, we may not be able to fix it
+ // with that half. We also bias towards fixing the B half because that
+ // will more commonly be the high half, and we have to bias one way.
+ auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
+ ArrayRef<int> Inputs) {
+ int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
+ bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
+ PinnedIdx ^ 1) != Inputs.end();
+ // Determine whether the free index is in the flipped dword or the
+ // unflipped dword based on where the pinned index is. We use this bit
+ // in an xor to conditionally select the adjacent dword.
+ int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
+ bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
+ FixFreeIdx) != Inputs.end();
+ if (IsFixIdxInput == IsFixFreeIdxInput)
+ FixFreeIdx += 1;
+ IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
+ FixFreeIdx) != Inputs.end();
+ assert(IsFixIdxInput != IsFixFreeIdxInput &&
+ "We need to be changing the number of flipped inputs!");
+ int PSHUFHalfMask[] = {0, 1, 2, 3};
+ std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
+ V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
+ MVT::v8i16, V,
+ getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DAG));
+
+ for (int &M : Mask)
+ if (M != -1 && M == FixIdx)
+ M = FixFreeIdx;
+ else if (M != -1 && M == FixFreeIdx)
+ M = FixIdx;
+ };
+ if (NumFlippedBToBInputs != 0) {
+ int BPinnedIdx =
+ BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
+ FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
+ } else {
+ assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
+ int APinnedIdx =
+ AToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
+ FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
+ }
+ }
+ }
int PSHUFDMask[] = {0, 1, 2, 3};
- PSHUFDMask[DWordA] = DWordB;
- PSHUFDMask[DWordB] = DWordA;
+ PSHUFDMask[ADWord] = BDWord;
+ PSHUFDMask[BDWord] = ADWord;
V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
// Adjust the mask to match the new locations of A and B.
for (int &M : Mask)
- if (M != -1 && M/2 == DWordA)
- M = 2 * DWordB + M % 2;
- else if (M != -1 && M/2 == DWordB)
- M = 2 * DWordA + M % 2;
+ if (M != -1 && M/2 == ADWord)
+ M = 2 * BDWord + M % 2;
+ else if (M != -1 && M/2 == BDWord)
+ M = 2 * ADWord + M % 2;
// Recurse back into this routine to re-compute state now that this isn't
// a 3 and 1 problem.
return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16),
Mask);
};
- if (NumLToL == 3 && NumHToL == 1)
- return balanceSides(LToLInputs, HToLInputs[0], 0 + 1 + 2 + 3, 4);
- else if (NumLToL == 1 && NumHToL == 3)
- return balanceSides(HToLInputs, LToLInputs[0], 4 + 5 + 6 + 7, 0);
- else if (NumLToH == 1 && NumHToH == 3)
- return balanceSides(HToHInputs, LToHInputs[0], 4 + 5 + 6 + 7, 0);
- else if (NumLToH == 3 && NumHToH == 1)
- return balanceSides(LToHInputs, HToHInputs[0], 0 + 1 + 2 + 3, 4);
+ if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
+ return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
+ else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
+ return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
// At this point there are at most two inputs to the low and high halves from
// each half. That means the inputs can always be grouped into dwords and
// First fix the masks for all the inputs that are staying in their
// original halves. This will then dictate the targets of the cross-half
// shuffles.
- auto fixInPlaceInputs = [&PSHUFDMask](
- ArrayRef<int> InPlaceInputs, MutableArrayRef<int> SourceHalfMask,
- MutableArrayRef<int> HalfMask, int HalfOffset) {
+ auto fixInPlaceInputs =
+ [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
+ MutableArrayRef<int> SourceHalfMask,
+ MutableArrayRef<int> HalfMask, int HalfOffset) {
if (InPlaceInputs.empty())
return;
if (InPlaceInputs.size() == 1) {
PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
return;
}
+ if (IncomingInputs.empty()) {
+ // Just fix all of the in place inputs.
+ for (int Input : InPlaceInputs) {
+ SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
+ PSHUFDMask[Input / 2] = Input / 2;
+ }
+ return;
+ }
assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
};
- if (!HToLInputs.empty())
- fixInPlaceInputs(LToLInputs, PSHUFLMask, LoMask, 0);
- if (!LToHInputs.empty())
- fixInPlaceInputs(HToHInputs, PSHUFHMask, HiMask, 4);
+ fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
+ fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
// Now gather the cross-half inputs and place them into a free dword of
// their target half.
auto moveInputsToRightHalf = [&PSHUFDMask](
MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
- int SourceOffset, int DestOffset) {
+ MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
+ int DestOffset) {
auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
};
Input - SourceOffset;
// We have to swap the uses in our half mask in one sweep.
for (int &M : HalfMask)
- if (M == SourceHalfMask[Input - SourceOffset])
+ if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
M = Input;
else if (M == Input)
M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
} else if (IncomingInputs.size() == 2) {
if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
- int SourceDWordBase = !isDWordClobbered(SourceHalfMask, 0) ? 0 : 2;
- assert(!isDWordClobbered(SourceHalfMask, SourceDWordBase) &&
- "Not all dwords can be clobbered!");
- SourceHalfMask[SourceDWordBase] = IncomingInputs[0] - SourceOffset;
- SourceHalfMask[SourceDWordBase + 1] = IncomingInputs[1] - SourceOffset;
+ // We have two non-adjacent or clobbered inputs we need to extract from
+ // the source half. To do this, we need to map them into some adjacent
+ // dword slot in the source mask.
+ int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
+ IncomingInputs[1] - SourceOffset};
+
+ // If there is a free slot in the source half mask adjacent to one of
+ // the inputs, place the other input in it. We use (Index XOR 1) to
+ // compute an adjacent index.
+ if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
+ SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
+ SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
+ SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
+ InputsFixed[1] = InputsFixed[0] ^ 1;
+ } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
+ SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
+ SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
+ SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
+ InputsFixed[0] = InputsFixed[1] ^ 1;
+ } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
+ SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
+ // The two inputs are in the same DWord but it is clobbered and the
+ // adjacent DWord isn't used at all. Move both inputs to the free
+ // slot.
+ SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
+ SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
+ InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
+ InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
+ } else {
+ // The only way we hit this point is if there is no clobbering
+ // (because there are no off-half inputs to this half) and there is no
+ // free slot adjacent to one of the inputs. In this case, we have to
+ // swap an input with a non-input.
+ for (int i = 0; i < 4; ++i)
+ assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
+ "We can't handle any clobbers here!");
+ assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
+ "Cannot have adjacent inputs here!");
+
+ SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
+ SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
+
+ // We also have to update the final source mask in this case because
+ // it may need to undo the above swap.
+ for (int &M : FinalSourceHalfMask)
+ if (M == (InputsFixed[0] ^ 1) + SourceOffset)
+ M = InputsFixed[1] + SourceOffset;
+ else if (M == InputsFixed[1] + SourceOffset)
+ M = (InputsFixed[0] ^ 1) + SourceOffset;
+
+ InputsFixed[1] = InputsFixed[0] ^ 1;
+ }
+
+ // Point everything at the fixed inputs.
for (int &M : HalfMask)
if (M == IncomingInputs[0])
- M = SourceDWordBase + SourceOffset;
+ M = InputsFixed[0] + SourceOffset;
else if (M == IncomingInputs[1])
- M = SourceDWordBase + 1 + SourceOffset;
- IncomingInputs[0] = SourceDWordBase + SourceOffset;
- IncomingInputs[1] = SourceDWordBase + 1 + SourceOffset;
+ M = InputsFixed[1] + SourceOffset;
+
+ IncomingInputs[0] = InputsFixed[0] + SourceOffset;
+ IncomingInputs[1] = InputsFixed[1] + SourceOffset;
}
} else {
llvm_unreachable("Unhandled input size!");
int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
- for (int Input : IncomingInputs)
- std::replace(HalfMask.begin(), HalfMask.end(), Input,
- FreeDWord * 2 + Input % 2);
+ for (int &M : HalfMask)
+ for (int Input : IncomingInputs)
+ if (M == Input)
+ M = FreeDWord * 2 + Input % 2;
};
- moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask,
+ moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
/*SourceOffset*/ 4, /*DestOffset*/ 0);
- moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask,
+ moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
/*SourceOffset*/ 0, /*DestOffset*/ 4);
// Now enact all the shuffles we've computed to move the inputs into their
if (GoodInputs.size() == 2) {
// If the low inputs are spread across two dwords, pack them into
// a single dword.
- MoveMask[Mask[GoodInputs[0]] % 2 + MoveOffset] =
- Mask[GoodInputs[0]] - MaskOffset;
- MoveMask[Mask[GoodInputs[1]] % 2 + MoveOffset] =
- Mask[GoodInputs[1]] - MaskOffset;
- Mask[GoodInputs[0]] = Mask[GoodInputs[0]] % 2 + MoveOffset + MaskOffset;
- Mask[GoodInputs[1]] = Mask[GoodInputs[0]] % 2 + MoveOffset + MaskOffset;
+ MoveMask[MoveOffset] = Mask[GoodInputs[0]] - MaskOffset;
+ MoveMask[MoveOffset + 1] = Mask[GoodInputs[1]] - MaskOffset;
+ Mask[GoodInputs[0]] = MoveOffset + MaskOffset;
+ Mask[GoodInputs[1]] = MoveOffset + 1 + MaskOffset;
} else {
- // Otherwise pin the low inputs.
+ // Otherwise pin the good inputs.
for (int GoodInput : GoodInputs)
MoveMask[Mask[GoodInput] - MaskOffset] = Mask[GoodInput] - MaskOffset;
}
- int MoveMaskIdx =
- std::find(std::begin(MoveMask) + MoveOffset, std::end(MoveMask), -1) -
- std::begin(MoveMask);
- assert(MoveMaskIdx >= MoveOffset && "Established above");
-
if (BadInputs.size() == 2) {
+ // If we have two bad inputs then there may be either one or two good
+ // inputs fixed in place. Find a fixed input, and then find the *other*
+ // two adjacent indices by using modular arithmetic.
+ int GoodMaskIdx =
+ std::find_if(std::begin(MoveMask) + MoveOffset, std::end(MoveMask),
+ [](int M) { return M >= 0; }) -
+ std::begin(MoveMask);
+ int MoveMaskIdx =
+ ((((GoodMaskIdx - MoveOffset) & ~1) + 2) % 4) + MoveOffset;
assert(MoveMask[MoveMaskIdx] == -1 && "Expected empty slot");
assert(MoveMask[MoveMaskIdx + 1] == -1 && "Expected empty slot");
- MoveMask[MoveMaskIdx + Mask[BadInputs[0]] % 2] =
- Mask[BadInputs[0]] - MaskOffset;
- MoveMask[MoveMaskIdx + Mask[BadInputs[1]] % 2] =
- Mask[BadInputs[1]] - MaskOffset;
- Mask[BadInputs[0]] = MoveMaskIdx + Mask[BadInputs[0]] % 2 + MaskOffset;
- Mask[BadInputs[1]] = MoveMaskIdx + Mask[BadInputs[1]] % 2 + MaskOffset;
+ MoveMask[MoveMaskIdx] = Mask[BadInputs[0]] - MaskOffset;
+ MoveMask[MoveMaskIdx + 1] = Mask[BadInputs[1]] - MaskOffset;
+ Mask[BadInputs[0]] = MoveMaskIdx + MaskOffset;
+ Mask[BadInputs[1]] = MoveMaskIdx + 1 + MaskOffset;
} else {
assert(BadInputs.size() == 1 && "All sizes handled");
+ int MoveMaskIdx = std::find(std::begin(MoveMask) + MoveOffset,
+ std::end(MoveMask), -1) -
+ std::begin(MoveMask);
MoveMask[MoveMaskIdx] = Mask[BadInputs[0]] - MaskOffset;
Mask[BadInputs[0]] = MoveMaskIdx + MaskOffset;
}
assert(NumV1Inputs > 0 && "All single-input shuffles should be canonicalized "
"to be V1-input shuffles.");
+ // There are special ways we can lower some single-element blends.
+ if (NumV2Inputs == 1)
+ if (SDValue V = lowerIntegerElementInsertionVectorShuffle(
+ MVT::v8i16, DL, V1, V2, Mask, Subtarget, DAG))
+ return V;
+
if (NumV1Inputs + NumV2Inputs <= 4)
return lowerV8I16BasicBlendVectorShuffle(DL, V1, V2, Mask, Subtarget, DAG);
DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, LoV, HiV));
}
+/// \brief Check whether a compaction lowering can be done by dropping even
+/// elements and compute how many times even elements must be dropped.
+///
+/// This handles shuffles which take every Nth element where N is a power of
+/// two. Example shuffle masks:
+///
+/// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
+/// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
+/// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
+/// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
+/// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
+/// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
+///
+/// Any of these lanes can of course be undef.
+///
+/// This routine only supports N <= 3.
+/// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
+/// for larger N.
+///
+/// \returns N above, or the number of times even elements must be dropped if
+/// there is such a number. Otherwise returns zero.
+static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
+ // Figure out whether we're looping over two inputs or just one.
+ bool IsSingleInput = isSingleInputShuffleMask(Mask);
+
+ // The modulus for the shuffle vector entries is based on whether this is
+ // a single input or not.
+ int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
+ assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
+ "We should only be called with masks with a power-of-2 size!");
+
+ uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
+
+ // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
+ // and 2^3 simultaneously. This is because we may have ambiguity with
+ // partially undef inputs.
+ bool ViableForN[3] = {true, true, true};
+
+ for (int i = 0, e = Mask.size(); i < e; ++i) {
+ // Ignore undef lanes, we'll optimistically collapse them to the pattern we
+ // want.
+ if (Mask[i] == -1)
+ continue;
+
+ bool IsAnyViable = false;
+ for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
+ if (ViableForN[j]) {
+ uint64_t N = j + 1;
+
+ // The shuffle mask must be equal to (i * 2^N) % M.
+ if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
+ IsAnyViable = true;
+ else
+ ViableForN[j] = false;
+ }
+ // Early exit if we exhaust the possible powers of two.
+ if (!IsAnyViable)
+ break;
+ }
+
+ for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
+ if (ViableForN[j])
+ return j + 1;
+
+ // Return 0 as there is no viable power of two.
+ return 0;
+}
+
/// \brief Generic lowering of v16i8 shuffles.
///
/// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
MutableArrayRef<int> LoMask = Mask.slice(0, 8);
MutableArrayRef<int> HiMask = Mask.slice(8, 8);
+ int NumV2Elements =
+ std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
+
// For single-input shuffles, there are some nicer lowering tricks we can use.
- if (isSingleInputShuffleMask(Mask)) {
+ if (NumV2Elements == 0) {
// Check whether we can widen this to an i16 shuffle by duplicating bytes.
// Notably, this handles splat and partial-splat shuffles more efficiently.
// However, it only makes sense if the pre-duplication shuffle simplifies
return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, Evens, Odds);
}
+ // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
+ // with PSHUFB. It is important to do this before we attempt to generate any
+ // blends but after all of the single-input lowerings. If the single input
+ // lowerings can find an instruction sequence that is faster than a PSHUFB, we
+ // want to preserve that and we can DAG combine any longer sequences into
+ // a PSHUFB in the end. But once we start blending from multiple inputs,
+ // the complexity of DAG combining bad patterns back into PSHUFB is too high,
+ // and there are *very* few patterns that would actually be faster than the
+ // PSHUFB approach because of its ability to zero lanes.
+ //
+ // FIXME: The only exceptions to the above are blends which are exact
+ // interleavings with direct instructions supporting them. We currently don't
+ // handle those well here.
+ if (Subtarget->hasSSSE3()) {
+ SDValue V1Mask[16];
+ SDValue V2Mask[16];
+ for (int i = 0; i < 16; ++i)
+ if (Mask[i] == -1) {
+ V1Mask[i] = V2Mask[i] = DAG.getConstant(0x80, MVT::i8);
+ } else {
+ V1Mask[i] = DAG.getConstant(Mask[i] < 16 ? Mask[i] : 0x80, MVT::i8);
+ V2Mask[i] =
+ DAG.getConstant(Mask[i] < 16 ? 0x80 : Mask[i] - 16, MVT::i8);
+ }
+ V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, V1,
+ DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
+ if (isSingleInputShuffleMask(Mask))
+ return V1; // Single inputs are easy.
+
+ // Otherwise, blend the two.
+ V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, V2,
+ DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
+ return DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
+ }
+
+ // There are special ways we can lower some single-element blends.
+ if (NumV2Elements == 1)
+ if (SDValue V = lowerIntegerElementInsertionVectorShuffle(
+ MVT::v16i8, DL, V1, V2, Mask, Subtarget, DAG))
+ return V;
+
+ // Check whether a compaction lowering can be done. This handles shuffles
+ // which take every Nth element for some even N. See the helper function for
+ // details.
+ //
+ // We special case these as they can be particularly efficiently handled with
+ // the PACKUSB instruction on x86 and they show up in common patterns of
+ // rearranging bytes to truncate wide elements.
+ if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
+ // NumEvenDrops is the power of two stride of the elements. Another way of
+ // thinking about it is that we need to drop the even elements this many
+ // times to get the original input.
+ bool IsSingleInput = isSingleInputShuffleMask(Mask);
+
+ // First we need to zero all the dropped bytes.
+ assert(NumEvenDrops <= 3 &&
+ "No support for dropping even elements more than 3 times.");
+ // We use the mask type to pick which bytes are preserved based on how many
+ // elements are dropped.
+ MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
+ SDValue ByteClearMask =
+ DAG.getNode(ISD::BITCAST, DL, MVT::v16i8,
+ DAG.getConstant(0xFF, MaskVTs[NumEvenDrops - 1]));
+ V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
+ if (!IsSingleInput)
+ V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
+
+ // Now pack things back together.
+ V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
+ V2 = IsSingleInput ? V1 : DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
+ SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
+ for (int i = 1; i < NumEvenDrops; ++i) {
+ Result = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, Result);
+ Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
+ }
+
+ return Result;
+ }
+
int V1LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
int V1HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
int V2LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
}
}
-/// \brief Tiny helper function to test whether adjacent masks are sequential.
-static bool areAdjacentMasksSequential(ArrayRef<int> Mask) {
- for (int i = 0, Size = Mask.size(); i < Size; i += 2)
- if (Mask[i] + 1 != Mask[i+1])
- return false;
-
- return true;
+static bool isHalfCrossingShuffleMask(ArrayRef<int> Mask) {
+ int Size = Mask.size();
+ for (int M : Mask.slice(0, Size / 2))
+ if (M >= 0 && (M % Size) >= Size / 2)
+ return true;
+ for (int M : Mask.slice(Size / 2, Size / 2))
+ if (M >= 0 && (M % Size) < Size / 2)
+ return true;
+ return false;
}
-/// \brief Top-level lowering for x86 vector shuffles.
+/// \brief Generic routine to split a 256-bit vector shuffle into 128-bit
+/// shuffles.
///
-/// This handles decomposition, canonicalization, and lowering of all x86
-/// vector shuffles. Most of the specific lowering strategies are encapsulated
-/// above in helper routines. The canonicalization attempts to widen shuffles
-/// to involve fewer lanes of wider elements, consolidate symmetric patterns
+/// There is a severely limited set of shuffles available in AVX1 for 256-bit
+/// vectors resulting in routinely needing to split the shuffle into two 128-bit
+/// shuffles. This can be done generically for any 256-bit vector shuffle and so
+/// we encode the logic here for specific shuffle lowering routines to bail to
+/// when they exhaust the features avaible to more directly handle the shuffle.
+static SDValue splitAndLower256BitVectorShuffle(SDValue Op, SDValue V1,
+ SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ MVT VT = Op.getSimpleValueType();
+ assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
+ assert(V1.getSimpleValueType() == VT && "Bad operand type!");
+ assert(V2.getSimpleValueType() == VT && "Bad operand type!");
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+
+ ArrayRef<int> LoMask = Mask.slice(0, Mask.size()/2);
+ ArrayRef<int> HiMask = Mask.slice(Mask.size()/2);
+
+ int NumElements = VT.getVectorNumElements();
+ int SplitNumElements = NumElements / 2;
+ MVT ScalarVT = VT.getScalarType();
+ MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
+
+ SDValue LoV1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V1,
+ DAG.getIntPtrConstant(0));
+ SDValue HiV1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V1,
+ DAG.getIntPtrConstant(SplitNumElements));
+ SDValue LoV2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V2,
+ DAG.getIntPtrConstant(0));
+ SDValue HiV2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V2,
+ DAG.getIntPtrConstant(SplitNumElements));
+
+ // Now create two 4-way blends of these half-width vectors.
+ auto HalfBlend = [&](ArrayRef<int> HalfMask) {
+ SmallVector<int, 16> V1BlendMask, V2BlendMask, BlendMask;
+ for (int i = 0; i < SplitNumElements; ++i) {
+ int M = HalfMask[i];
+ if (M >= NumElements) {
+ V2BlendMask.push_back(M - NumElements);
+ V1BlendMask.push_back(-1);
+ BlendMask.push_back(SplitNumElements + i);
+ } else if (M >= 0) {
+ V2BlendMask.push_back(-1);
+ V1BlendMask.push_back(M);
+ BlendMask.push_back(i);
+ } else {
+ V2BlendMask.push_back(-1);
+ V1BlendMask.push_back(-1);
+ BlendMask.push_back(-1);
+ }
+ }
+ SDValue V1Blend = DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
+ SDValue V2Blend = DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
+ return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
+ };
+ SDValue Lo = HalfBlend(LoMask);
+ SDValue Hi = HalfBlend(HiMask);
+ return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
+}
+
+/// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
+///
+/// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
+/// isn't available.
+static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
+
+ // FIXME: If we have AVX2, we should delegate to generic code as crossing
+ // shuffles aren't a problem and FP and int have the same patterns.
+
+ // FIXME: We can handle these more cleverly than splitting for v4f64.
+ if (isHalfCrossingShuffleMask(Mask))
+ return splitAndLower256BitVectorShuffle(Op, V1, V2, Subtarget, DAG);
+
+ if (isSingleInputShuffleMask(Mask)) {
+ // Non-half-crossing single input shuffles can be lowerid with an
+ // interleaved permutation.
+ unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
+ ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
+ return DAG.getNode(X86ISD::VPERMILP, DL, MVT::v4f64, V1,
+ DAG.getConstant(VPERMILPMask, MVT::i8));
+ }
+
+ // X86 has dedicated unpack instructions that can handle specific blend
+ // operations: UNPCKH and UNPCKL.
+ if (isShuffleEquivalent(Mask, 0, 4, 2, 6))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
+ if (isShuffleEquivalent(Mask, 1, 5, 3, 7))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
+ // FIXME: It would be nice to find a way to get canonicalization to commute
+ // these patterns.
+ if (isShuffleEquivalent(Mask, 4, 0, 6, 2))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
+ if (isShuffleEquivalent(Mask, 5, 1, 7, 3))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
+
+ // Check if the blend happens to exactly fit that of SHUFPD.
+ if (Mask[0] < 4 && (Mask[1] == -1 || Mask[1] >= 4) &&
+ Mask[2] < 4 && (Mask[3] == -1 || Mask[3] >= 4)) {
+ unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 5) << 1) |
+ ((Mask[2] == 3) << 2) | ((Mask[3] == 7) << 3);
+ return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V1, V2,
+ DAG.getConstant(SHUFPDMask, MVT::i8));
+ }
+ if ((Mask[0] == -1 || Mask[0] >= 4) && Mask[1] < 4 &&
+ (Mask[2] == -1 || Mask[2] >= 4) && Mask[3] < 4) {
+ unsigned SHUFPDMask = (Mask[0] == 5) | ((Mask[1] == 1) << 1) |
+ ((Mask[2] == 7) << 2) | ((Mask[3] == 3) << 3);
+ return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V2, V1,
+ DAG.getConstant(SHUFPDMask, MVT::i8));
+ }
+
+ // Shuffle the input elements into the desired positions in V1 and V2 and
+ // blend them together.
+ int V1Mask[] = {-1, -1, -1, -1};
+ int V2Mask[] = {-1, -1, -1, -1};
+ for (int i = 0; i < 4; ++i)
+ if (Mask[i] >= 0 && Mask[i] < 4)
+ V1Mask[i] = Mask[i];
+ else if (Mask[i] >= 4)
+ V2Mask[i] = Mask[i] - 4;
+
+ V1 = DAG.getVectorShuffle(MVT::v4f64, DL, V1, DAG.getUNDEF(MVT::v4f64), V1Mask);
+ V2 = DAG.getVectorShuffle(MVT::v4f64, DL, V2, DAG.getUNDEF(MVT::v4f64), V2Mask);
+
+ unsigned BlendMask = 0;
+ for (int i = 0; i < 4; ++i)
+ if (Mask[i] >= 4)
+ BlendMask |= 1 << i;
+
+ return DAG.getNode(X86ISD::BLENDI, DL, MVT::v4f64, V1, V2,
+ DAG.getConstant(BlendMask, MVT::i8));
+}
+
+/// \brief Handle lowering of 4-lane 64-bit integer shuffles.
+///
+/// Largely delegates to common code when we have AVX2 and to the floating-point
+/// code when we only have AVX.
+static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ assert(Op.getSimpleValueType() == MVT::v4i64 && "Bad shuffle type!");
+ assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
+
+ // FIXME: If we have AVX2, we should delegate to generic code as crossing
+ // shuffles aren't a problem and FP and int have the same patterns.
+
+ if (isHalfCrossingShuffleMask(Mask))
+ return splitAndLower256BitVectorShuffle(Op, V1, V2, Subtarget, DAG);
+
+ // AVX1 doesn't provide any facilities for v4i64 shuffles, bitcast and
+ // delegate to floating point code.
+ V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v4f64, V1);
+ V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v4f64, V2);
+ return DAG.getNode(ISD::BITCAST, DL, MVT::v4i64,
+ lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG));
+}
+
+/// \brief High-level routine to lower various 256-bit x86 vector shuffles.
+///
+/// This routine either breaks down the specific type of a 256-bit x86 vector
+/// shuffle or splits it into two 128-bit shuffles and fuses the results back
+/// together based on the available instructions.
+static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ MVT VT, const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ switch (VT.SimpleTy) {
+ case MVT::v4f64:
+ return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
+ case MVT::v4i64:
+ return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
+ case MVT::v8i32:
+ case MVT::v8f32:
+ case MVT::v16i16:
+ case MVT::v32i8:
+ // Fall back to the basic pattern of extracting the high half and forming
+ // a 4-way blend.
+ // FIXME: Add targeted lowering for each type that can document rationale
+ // for delegating to this when necessary.
+ return splitAndLower256BitVectorShuffle(Op, V1, V2, Subtarget, DAG);
+
+ default:
+ llvm_unreachable("Not a valid 256-bit x86 vector type!");
+ }
+}
+
+/// \brief Tiny helper function to test whether a shuffle mask could be
+/// simplified by widening the elements being shuffled.
+static bool canWidenShuffleElements(ArrayRef<int> Mask) {
+ for (int i = 0, Size = Mask.size(); i < Size; i += 2)
+ if (Mask[i] % 2 != 0 || Mask[i] + 1 != Mask[i+1])
+ return false;
+
+ return true;
+}
+
+/// \brief Top-level lowering for x86 vector shuffles.
+///
+/// This handles decomposition, canonicalization, and lowering of all x86
+/// vector shuffles. Most of the specific lowering strategies are encapsulated
+/// above in helper routines. The canonicalization attempts to widen shuffles
+/// to involve fewer lanes of wider elements, consolidate symmetric patterns
/// s.t. only one of the two inputs needs to be tested, etc.
static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
SelectionDAG &DAG) {
// but it might be interesting to form i128 integers to handle flipping the
// low and high halves of AVX 256-bit vectors.
if (VT.isInteger() && VT.getScalarSizeInBits() < 64 &&
- areAdjacentMasksSequential(Mask)) {
+ canWidenShuffleElements(Mask)) {
SmallVector<int, 8> NewMask;
for (int i = 0, Size = Mask.size(); i < Size; i += 2)
NewMask.push_back(Mask[i] / 2);
if (VT.getSizeInBits() == 128)
return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
+ if (VT.getSizeInBits() == 256)
+ return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
+
llvm_unreachable("Unimplemented!");
}
getShufflePALIGNRImmediate(SVOp),
DAG);
+ if (isVALIGNMask(M, VT, Subtarget))
+ return getTargetShuffleNode(X86ISD::VALIGN, dl, VT, V1, V2,
+ getShuffleVALIGNImmediate(SVOp),
+ DAG);
+
// Check if this can be converted into a logical shift.
bool isLeft = false;
unsigned ShAmt = 0;
}
SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
+ // A vselect where all conditions and data are constants can be optimized into
+ // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
+ if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
+ ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
+ ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
+ return SDValue();
+
SDValue BlendOp = LowerVSELECTtoBlend(Op, Subtarget, DAG);
if (BlendOp.getNode())
return BlendOp;
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 if (VT == MVT::v16i8)
- Opc = X86ISD::PINSRB;
- else
- 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();
if (Subtarget->is64Bit())
IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain,
IDX, MachinePointerInfo(), MVT::i32,
- false, false, 0);
+ false, false, false, 0);
else
IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
false, false, false, 0);
// FIXME: Avoid the extend by constructing the right constant pool?
SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
FudgePtr, MachinePointerInfo::getConstantPool(),
- MVT::f32, false, false, 4);
+ MVT::f32, false, false, false, 4);
// Extend everything to 80 bits to force it to be done on x87.
SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
if (VT == MVT::i1) {
assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
"Invalid scalar TRUNCATE operation");
- if (InVT == MVT::i32)
+ if (InVT.getSizeInBits() >= 32)
return SDValue();
- if (InVT.getSizeInBits() == 64)
- In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::i32, In);
- else if (InVT.getSizeInBits() < 32)
- In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
+ In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
}
assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
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 &&
return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
}
+// Lower vector extended loads using a shuffle. If SSSE3 is not available we
+// may emit an illegal shuffle but the expansion is still better than scalar
+// code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
+// we'll emit a shuffle and a arithmetic shift.
+// TODO: It is possible to support ZExt by zeroing the undef values during
+// the shuffle phase or after the shuffle.
+static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ MVT RegVT = Op.getSimpleValueType();
+ assert(RegVT.isVector() && "We only custom lower vector sext loads.");
+ assert(RegVT.isInteger() &&
+ "We only custom lower integer vector sext loads.");
+
+ // Nothing useful we can do without SSE2 shuffles.
+ assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
+
+ LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
+ SDLoc dl(Ld);
+ EVT MemVT = Ld->getMemoryVT();
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+ unsigned RegSz = RegVT.getSizeInBits();
+
+ ISD::LoadExtType Ext = Ld->getExtensionType();
+
+ assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
+ && "Only anyext and sext are currently implemented.");
+ assert(MemVT != RegVT && "Cannot extend to the same type");
+ assert(MemVT.isVector() && "Must load a vector from memory");
+
+ unsigned NumElems = RegVT.getVectorNumElements();
+ unsigned MemSz = MemVT.getSizeInBits();
+ assert(RegSz > MemSz && "Register size must be greater than the mem size");
+
+ if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
+ // The only way in which we have a legal 256-bit vector result but not the
+ // integer 256-bit operations needed to directly lower a sextload is if we
+ // have AVX1 but not AVX2. In that case, we can always emit a sextload to
+ // a 128-bit vector and a normal sign_extend to 256-bits that should get
+ // correctly legalized. We do this late to allow the canonical form of
+ // sextload to persist throughout the rest of the DAG combiner -- it wants
+ // to fold together any extensions it can, and so will fuse a sign_extend
+ // of an sextload into a sextload targeting a wider value.
+ SDValue Load;
+ if (MemSz == 128) {
+ // Just switch this to a normal load.
+ assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
+ "it must be a legal 128-bit vector "
+ "type!");
+ Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
+ Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
+ Ld->isInvariant(), Ld->getAlignment());
+ } else {
+ assert(MemSz < 128 &&
+ "Can't extend a type wider than 128 bits to a 256 bit vector!");
+ // Do an sext load to a 128-bit vector type. We want to use the same
+ // number of elements, but elements half as wide. This will end up being
+ // recursively lowered by this routine, but will succeed as we definitely
+ // have all the necessary features if we're using AVX1.
+ EVT HalfEltVT =
+ EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
+ EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
+ Load =
+ DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
+ Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
+ Ld->isNonTemporal(), Ld->isInvariant(),
+ Ld->getAlignment());
+ }
+
+ // Replace chain users with the new chain.
+ assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
+
+ // Finally, do a normal sign-extend to the desired register.
+ return DAG.getSExtOrTrunc(Load, dl, RegVT);
+ }
+
+ // All sizes must be a power of two.
+ assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
+ "Non-power-of-two elements are not custom lowered!");
+
+ // Attempt to load the original value using scalar loads.
+ // Find the largest scalar type that divides the total loaded size.
+ MVT SclrLoadTy = MVT::i8;
+ for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
+ tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
+ MVT Tp = (MVT::SimpleValueType)tp;
+ if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
+ SclrLoadTy = Tp;
+ }
+ }
+
+ // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
+ if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
+ (64 <= MemSz))
+ SclrLoadTy = MVT::f64;
+
+ // Calculate the number of scalar loads that we need to perform
+ // in order to load our vector from memory.
+ unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
+
+ assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
+ "Can only lower sext loads with a single scalar load!");
+
+ unsigned loadRegZize = RegSz;
+ if (Ext == ISD::SEXTLOAD && RegSz == 256)
+ loadRegZize /= 2;
+
+ // Represent our vector as a sequence of elements which are the
+ // largest scalar that we can load.
+ EVT LoadUnitVecVT = EVT::getVectorVT(
+ *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
+
+ // 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(),
+ loadRegZize / MemVT.getScalarType().getSizeInBits());
+
+ assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
+ "Invalid vector type");
+
+ // We can't shuffle using an illegal type.
+ assert(TLI.isTypeLegal(WideVecVT) &&
+ "We only lower types that form legal widened vector types");
+
+ SmallVector<SDValue, 8> Chains;
+ SDValue Ptr = Ld->getBasePtr();
+ SDValue Increment =
+ DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, TLI.getPointerTy());
+ SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
+
+ for (unsigned i = 0; i < NumLoads; ++i) {
+ // Perform a single load.
+ SDValue ScalarLoad =
+ DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
+ Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
+ Ld->getAlignment());
+ Chains.push_back(ScalarLoad.getValue(1));
+ // Create the first element type using SCALAR_TO_VECTOR in order to avoid
+ // another round of DAGCombining.
+ if (i == 0)
+ Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
+ else
+ Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
+ ScalarLoad, DAG.getIntPtrConstant(i));
+
+ Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
+ }
+
+ SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
+
+ // Bitcast the loaded value to a vector of the original element type, in
+ // the size of the target vector type.
+ SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
+ unsigned SizeRatio = RegSz / MemSz;
+
+ if (Ext == ISD::SEXTLOAD) {
+ // If we have SSE4.1, we can directly emit a VSEXT node.
+ if (Subtarget->hasSSE41()) {
+ SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
+ return Sext;
+ }
+
+ // Otherwise we'll shuffle the small elements in the high bits of the
+ // larger type and perform an arithmetic shift. If the shift is not legal
+ // it's better to scalarize.
+ assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
+ "We can't implement a sext load without an arithmetic right shift!");
+
+ // Redistribute the loaded elements into the different locations.
+ SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
+ for (unsigned i = 0; i != NumElems; ++i)
+ ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
+
+ SDValue Shuff = DAG.getVectorShuffle(
+ WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
+
+ Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
+
+ // Build the arithmetic shift.
+ unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
+ MemVT.getVectorElementType().getSizeInBits();
+ Shuff =
+ DAG.getNode(ISD::SRA, dl, RegVT, Shuff, DAG.getConstant(Amt, RegVT));
+
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
+ return Shuff;
+ }
+
+ // Redistribute the loaded elements into the different locations.
+ SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
+ for (unsigned i = 0; i != NumElems; ++i)
+ ShuffleVec[i * SizeRatio] = i;
+
+ SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
+ DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
+
+ // Bitcast to the requested type.
+ Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
+ return Shuff;
+}
+
// isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
// ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
// from the AND / OR.
}
// Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
-// Calls to _alloca is needed to probe the stack when allocating more than 4k
+// Calls to _alloca are needed to probe the stack when allocating more than 4k
// bytes in one go. Touching the stack at 4K increments is necessary to ensure
// that the guard pages used by the OS virtual memory manager are allocated in
// correct sequence.
SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
Chain = SP.getValue(1);
unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
- const TargetFrameLowering &TFI = *DAG.getTarget().getFrameLowering();
+ const TargetFrameLowering &TFI = *DAG.getSubtarget().getFrameLowering();
unsigned StackAlign = TFI.getStackAlignment();
Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
if (Align > StackAlign)
Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ DAG.getSubtarget().getRegisterInfo());
unsigned SPReg = RegInfo->getStackRegister();
SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
Chain = SP.getValue(1);
return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
}
+/// \brief Return (vselect \p Mask, \p Op, \p PreservedSrc) along with the
+/// necessary casting for \p Mask when lowering masking intrinsics.
+static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
+ SDValue PreservedSrc, SelectionDAG &DAG) {
+ EVT VT = Op.getValueType();
+ EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
+ MVT::i1, VT.getVectorNumElements());
+ SDLoc dl(Op);
+
+ assert(MaskVT.isSimple() && "invalid mask type");
+ return DAG.getNode(ISD::VSELECT, dl, VT,
+ DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask),
+ Op, PreservedSrc);
+}
+
+static unsigned getOpcodeForFMAIntrinsic(unsigned IntNo) {
+ switch (IntNo) {
+ default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
+ case Intrinsic::x86_fma_vfmadd_ps:
+ case Intrinsic::x86_fma_vfmadd_pd:
+ case Intrinsic::x86_fma_vfmadd_ps_256:
+ case Intrinsic::x86_fma_vfmadd_pd_256:
+ case Intrinsic::x86_fma_mask_vfmadd_ps_512:
+ case Intrinsic::x86_fma_mask_vfmadd_pd_512:
+ return X86ISD::FMADD;
+ case Intrinsic::x86_fma_vfmsub_ps:
+ case Intrinsic::x86_fma_vfmsub_pd:
+ case Intrinsic::x86_fma_vfmsub_ps_256:
+ case Intrinsic::x86_fma_vfmsub_pd_256:
+ case Intrinsic::x86_fma_mask_vfmsub_ps_512:
+ case Intrinsic::x86_fma_mask_vfmsub_pd_512:
+ return X86ISD::FMSUB;
+ case Intrinsic::x86_fma_vfnmadd_ps:
+ case Intrinsic::x86_fma_vfnmadd_pd:
+ case Intrinsic::x86_fma_vfnmadd_ps_256:
+ case Intrinsic::x86_fma_vfnmadd_pd_256:
+ case Intrinsic::x86_fma_mask_vfnmadd_ps_512:
+ case Intrinsic::x86_fma_mask_vfnmadd_pd_512:
+ return X86ISD::FNMADD;
+ case Intrinsic::x86_fma_vfnmsub_ps:
+ case Intrinsic::x86_fma_vfnmsub_pd:
+ case Intrinsic::x86_fma_vfnmsub_ps_256:
+ case Intrinsic::x86_fma_vfnmsub_pd_256:
+ case Intrinsic::x86_fma_mask_vfnmsub_ps_512:
+ case Intrinsic::x86_fma_mask_vfnmsub_pd_512:
+ return X86ISD::FNMSUB;
+ case Intrinsic::x86_fma_vfmaddsub_ps:
+ case Intrinsic::x86_fma_vfmaddsub_pd:
+ case Intrinsic::x86_fma_vfmaddsub_ps_256:
+ case Intrinsic::x86_fma_vfmaddsub_pd_256:
+ case Intrinsic::x86_fma_mask_vfmaddsub_ps_512:
+ case Intrinsic::x86_fma_mask_vfmaddsub_pd_512:
+ return X86ISD::FMADDSUB;
+ case Intrinsic::x86_fma_vfmsubadd_ps:
+ case Intrinsic::x86_fma_vfmsubadd_pd:
+ case Intrinsic::x86_fma_vfmsubadd_ps_256:
+ case Intrinsic::x86_fma_vfmsubadd_pd_256:
+ case Intrinsic::x86_fma_mask_vfmsubadd_ps_512:
+ case Intrinsic::x86_fma_mask_vfmsubadd_pd_512:
+ return X86ISD::FMSUBADD;
+ }
+}
+
static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
SDLoc dl(Op);
unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
- switch (IntNo) {
- default: return SDValue(); // Don't custom lower most intrinsics.
- // Comparison intrinsics.
- case Intrinsic::x86_sse_comieq_ss:
- case Intrinsic::x86_sse_comilt_ss:
- case Intrinsic::x86_sse_comile_ss:
- case Intrinsic::x86_sse_comigt_ss:
- case Intrinsic::x86_sse_comige_ss:
- case Intrinsic::x86_sse_comineq_ss:
- case Intrinsic::x86_sse_ucomieq_ss:
- case Intrinsic::x86_sse_ucomilt_ss:
- case Intrinsic::x86_sse_ucomile_ss:
- case Intrinsic::x86_sse_ucomigt_ss:
- case Intrinsic::x86_sse_ucomige_ss:
- case Intrinsic::x86_sse_ucomineq_ss:
- case Intrinsic::x86_sse2_comieq_sd:
- case Intrinsic::x86_sse2_comilt_sd:
- case Intrinsic::x86_sse2_comile_sd:
- case Intrinsic::x86_sse2_comigt_sd:
- case Intrinsic::x86_sse2_comige_sd:
- case Intrinsic::x86_sse2_comineq_sd:
- case Intrinsic::x86_sse2_ucomieq_sd:
- case Intrinsic::x86_sse2_ucomilt_sd:
- case Intrinsic::x86_sse2_ucomile_sd:
- case Intrinsic::x86_sse2_ucomigt_sd:
- case Intrinsic::x86_sse2_ucomige_sd:
- case Intrinsic::x86_sse2_ucomineq_sd: {
- unsigned Opc;
- ISD::CondCode CC;
- switch (IntNo) {
- default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
- case Intrinsic::x86_sse_comieq_ss:
- case Intrinsic::x86_sse2_comieq_sd:
- Opc = X86ISD::COMI;
- CC = ISD::SETEQ;
- break;
- case Intrinsic::x86_sse_comilt_ss:
- case Intrinsic::x86_sse2_comilt_sd:
- Opc = X86ISD::COMI;
- CC = ISD::SETLT;
- break;
- case Intrinsic::x86_sse_comile_ss:
- case Intrinsic::x86_sse2_comile_sd:
- Opc = X86ISD::COMI;
- CC = ISD::SETLE;
- break;
- case Intrinsic::x86_sse_comigt_ss:
- case Intrinsic::x86_sse2_comigt_sd:
- Opc = X86ISD::COMI;
- CC = ISD::SETGT;
- break;
- case Intrinsic::x86_sse_comige_ss:
- case Intrinsic::x86_sse2_comige_sd:
- Opc = X86ISD::COMI;
- CC = ISD::SETGE;
- break;
- case Intrinsic::x86_sse_comineq_ss:
- case Intrinsic::x86_sse2_comineq_sd:
- Opc = X86ISD::COMI;
- CC = ISD::SETNE;
- break;
- case Intrinsic::x86_sse_ucomieq_ss:
- case Intrinsic::x86_sse2_ucomieq_sd:
- Opc = X86ISD::UCOMI;
- CC = ISD::SETEQ;
- break;
- case Intrinsic::x86_sse_ucomilt_ss:
- case Intrinsic::x86_sse2_ucomilt_sd:
- Opc = X86ISD::UCOMI;
- CC = ISD::SETLT;
- break;
- case Intrinsic::x86_sse_ucomile_ss:
- case Intrinsic::x86_sse2_ucomile_sd:
- Opc = X86ISD::UCOMI;
- CC = ISD::SETLE;
- break;
- case Intrinsic::x86_sse_ucomigt_ss:
- case Intrinsic::x86_sse2_ucomigt_sd:
- Opc = X86ISD::UCOMI;
- CC = ISD::SETGT;
- break;
- case Intrinsic::x86_sse_ucomige_ss:
- case Intrinsic::x86_sse2_ucomige_sd:
- Opc = X86ISD::UCOMI;
- CC = ISD::SETGE;
- break;
- case Intrinsic::x86_sse_ucomineq_ss:
- case Intrinsic::x86_sse2_ucomineq_sd:
- Opc = X86ISD::UCOMI;
- CC = ISD::SETNE;
+
+ const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
+ if (IntrData) {
+ switch(IntrData->Type) {
+ case INTR_TYPE_1OP:
+ return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
+ case INTR_TYPE_2OP:
+ return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
+ Op.getOperand(2));
+ case INTR_TYPE_3OP:
+ return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
+ Op.getOperand(2), Op.getOperand(3));
+ case COMI: { // Comparison intrinsics
+ ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
+ SDValue LHS = Op.getOperand(1);
+ SDValue RHS = Op.getOperand(2);
+ unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
+ assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
+ SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
+ SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
+ DAG.getConstant(X86CC, MVT::i8), Cond);
+ return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
+ }
+ case VSHIFT:
+ return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
+ Op.getOperand(1), Op.getOperand(2), DAG);
+ default:
break;
}
-
- SDValue LHS = Op.getOperand(1);
- SDValue RHS = Op.getOperand(2);
- unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
- assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
- SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
- SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
- DAG.getConstant(X86CC, MVT::i8), Cond);
- return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
}
+ switch (IntNo) {
+ default: return SDValue(); // Don't custom lower most intrinsics.
+
// Arithmetic intrinsics.
case Intrinsic::x86_sse2_pmulu_dq:
case Intrinsic::x86_avx2_pmulu_dq:
return DAG.getNode(ISD::MULHS, dl, Op.getValueType(),
Op.getOperand(1), Op.getOperand(2));
- // SSE2/AVX2 sub with unsigned saturation intrinsics
- case Intrinsic::x86_sse2_psubus_b:
- case Intrinsic::x86_sse2_psubus_w:
- case Intrinsic::x86_avx2_psubus_b:
- case Intrinsic::x86_avx2_psubus_w:
- return DAG.getNode(X86ISD::SUBUS, dl, Op.getValueType(),
- Op.getOperand(1), Op.getOperand(2));
-
- // SSE3/AVX horizontal add/sub intrinsics
- case Intrinsic::x86_sse3_hadd_ps:
- case Intrinsic::x86_sse3_hadd_pd:
- case Intrinsic::x86_avx_hadd_ps_256:
- case Intrinsic::x86_avx_hadd_pd_256:
- case Intrinsic::x86_sse3_hsub_ps:
- case Intrinsic::x86_sse3_hsub_pd:
- case Intrinsic::x86_avx_hsub_ps_256:
- case Intrinsic::x86_avx_hsub_pd_256:
- case Intrinsic::x86_ssse3_phadd_w_128:
- case Intrinsic::x86_ssse3_phadd_d_128:
- case Intrinsic::x86_avx2_phadd_w:
- case Intrinsic::x86_avx2_phadd_d:
- case Intrinsic::x86_ssse3_phsub_w_128:
- case Intrinsic::x86_ssse3_phsub_d_128:
- case Intrinsic::x86_avx2_phsub_w:
- case Intrinsic::x86_avx2_phsub_d: {
- unsigned Opcode;
- switch (IntNo) {
- default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
- case Intrinsic::x86_sse3_hadd_ps:
- case Intrinsic::x86_sse3_hadd_pd:
- case Intrinsic::x86_avx_hadd_ps_256:
- case Intrinsic::x86_avx_hadd_pd_256:
- Opcode = X86ISD::FHADD;
- break;
- case Intrinsic::x86_sse3_hsub_ps:
- case Intrinsic::x86_sse3_hsub_pd:
- case Intrinsic::x86_avx_hsub_ps_256:
- case Intrinsic::x86_avx_hsub_pd_256:
- Opcode = X86ISD::FHSUB;
- break;
- case Intrinsic::x86_ssse3_phadd_w_128:
- case Intrinsic::x86_ssse3_phadd_d_128:
- case Intrinsic::x86_avx2_phadd_w:
- case Intrinsic::x86_avx2_phadd_d:
- Opcode = X86ISD::HADD;
- break;
- case Intrinsic::x86_ssse3_phsub_w_128:
- case Intrinsic::x86_ssse3_phsub_d_128:
- case Intrinsic::x86_avx2_phsub_w:
- case Intrinsic::x86_avx2_phsub_d:
- Opcode = X86ISD::HSUB;
- break;
- }
- return DAG.getNode(Opcode, dl, Op.getValueType(),
- Op.getOperand(1), Op.getOperand(2));
- }
-
- // SSE2/SSE41/AVX2 integer max/min intrinsics.
- case Intrinsic::x86_sse2_pmaxu_b:
- case Intrinsic::x86_sse41_pmaxuw:
- case Intrinsic::x86_sse41_pmaxud:
- case Intrinsic::x86_avx2_pmaxu_b:
- case Intrinsic::x86_avx2_pmaxu_w:
- case Intrinsic::x86_avx2_pmaxu_d:
- case Intrinsic::x86_sse2_pminu_b:
- case Intrinsic::x86_sse41_pminuw:
- case Intrinsic::x86_sse41_pminud:
- case Intrinsic::x86_avx2_pminu_b:
- case Intrinsic::x86_avx2_pminu_w:
- case Intrinsic::x86_avx2_pminu_d:
- case Intrinsic::x86_sse41_pmaxsb:
- case Intrinsic::x86_sse2_pmaxs_w:
- case Intrinsic::x86_sse41_pmaxsd:
- case Intrinsic::x86_avx2_pmaxs_b:
- case Intrinsic::x86_avx2_pmaxs_w:
- case Intrinsic::x86_avx2_pmaxs_d:
- case Intrinsic::x86_sse41_pminsb:
- case Intrinsic::x86_sse2_pmins_w:
- case Intrinsic::x86_sse41_pminsd:
- case Intrinsic::x86_avx2_pmins_b:
- case Intrinsic::x86_avx2_pmins_w:
- case Intrinsic::x86_avx2_pmins_d: {
- unsigned Opcode;
- switch (IntNo) {
- default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
- case Intrinsic::x86_sse2_pmaxu_b:
- case Intrinsic::x86_sse41_pmaxuw:
- case Intrinsic::x86_sse41_pmaxud:
- case Intrinsic::x86_avx2_pmaxu_b:
- case Intrinsic::x86_avx2_pmaxu_w:
- case Intrinsic::x86_avx2_pmaxu_d:
- Opcode = X86ISD::UMAX;
- break;
- case Intrinsic::x86_sse2_pminu_b:
- case Intrinsic::x86_sse41_pminuw:
- case Intrinsic::x86_sse41_pminud:
- case Intrinsic::x86_avx2_pminu_b:
- case Intrinsic::x86_avx2_pminu_w:
- case Intrinsic::x86_avx2_pminu_d:
- Opcode = X86ISD::UMIN;
- break;
- case Intrinsic::x86_sse41_pmaxsb:
- case Intrinsic::x86_sse2_pmaxs_w:
- case Intrinsic::x86_sse41_pmaxsd:
- case Intrinsic::x86_avx2_pmaxs_b:
- case Intrinsic::x86_avx2_pmaxs_w:
- case Intrinsic::x86_avx2_pmaxs_d:
- Opcode = X86ISD::SMAX;
- break;
- case Intrinsic::x86_sse41_pminsb:
- case Intrinsic::x86_sse2_pmins_w:
- case Intrinsic::x86_sse41_pminsd:
- case Intrinsic::x86_avx2_pmins_b:
- case Intrinsic::x86_avx2_pmins_w:
- case Intrinsic::x86_avx2_pmins_d:
- Opcode = X86ISD::SMIN;
- break;
- }
- return DAG.getNode(Opcode, dl, Op.getValueType(),
- Op.getOperand(1), Op.getOperand(2));
- }
-
// SSE/SSE2/AVX floating point max/min intrinsics.
case Intrinsic::x86_sse_max_ps:
case Intrinsic::x86_sse2_max_pd:
return DAG.getNode(X86ISD::PSIGN, dl, Op.getValueType(),
Op.getOperand(1), Op.getOperand(2));
- case Intrinsic::x86_sse41_insertps:
- return DAG.getNode(X86ISD::INSERTPS, dl, Op.getValueType(),
- Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
-
- case Intrinsic::x86_avx_vperm2f128_ps_256:
- case Intrinsic::x86_avx_vperm2f128_pd_256:
- case Intrinsic::x86_avx_vperm2f128_si_256:
- case Intrinsic::x86_avx2_vperm2i128:
- return DAG.getNode(X86ISD::VPERM2X128, dl, Op.getValueType(),
- Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
-
case Intrinsic::x86_avx2_permd:
case Intrinsic::x86_avx2_permps:
// Operands intentionally swapped. Mask is last operand to intrinsic,
return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
Op.getOperand(2), Op.getOperand(1));
- case Intrinsic::x86_sse_sqrt_ps:
- case Intrinsic::x86_sse2_sqrt_pd:
- case Intrinsic::x86_avx_sqrt_ps_256:
- case Intrinsic::x86_avx_sqrt_pd_256:
- return DAG.getNode(ISD::FSQRT, dl, Op.getValueType(), Op.getOperand(1));
+ case Intrinsic::x86_avx512_mask_valign_q_512:
+ case Intrinsic::x86_avx512_mask_valign_d_512:
+ // Vector source operands are swapped.
+ return getVectorMaskingNode(DAG.getNode(X86ISD::VALIGN, dl,
+ Op.getValueType(), Op.getOperand(2),
+ Op.getOperand(1),
+ Op.getOperand(3)),
+ Op.getOperand(5), Op.getOperand(4), DAG);
// ptest and testp intrinsics. The intrinsic these come from are designed to
// return an integer value, not just an instruction so lower it to the ptest
break;
}
- SDValue LHS = Op.getOperand(1);
- SDValue RHS = Op.getOperand(2);
- unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
- SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
- SDValue CC = DAG.getConstant(X86CC, MVT::i8);
- SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
- return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
- }
- case Intrinsic::x86_avx512_kortestz_w:
- case Intrinsic::x86_avx512_kortestc_w: {
- unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
- SDValue LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(1));
- SDValue RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(2));
- SDValue CC = DAG.getConstant(X86CC, MVT::i8);
- SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
- SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
- return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
- }
-
- // SSE/AVX shift intrinsics
- case Intrinsic::x86_sse2_psll_w:
- case Intrinsic::x86_sse2_psll_d:
- case Intrinsic::x86_sse2_psll_q:
- case Intrinsic::x86_avx2_psll_w:
- case Intrinsic::x86_avx2_psll_d:
- case Intrinsic::x86_avx2_psll_q:
- case Intrinsic::x86_sse2_psrl_w:
- case Intrinsic::x86_sse2_psrl_d:
- case Intrinsic::x86_sse2_psrl_q:
- case Intrinsic::x86_avx2_psrl_w:
- case Intrinsic::x86_avx2_psrl_d:
- case Intrinsic::x86_avx2_psrl_q:
- case Intrinsic::x86_sse2_psra_w:
- case Intrinsic::x86_sse2_psra_d:
- case Intrinsic::x86_avx2_psra_w:
- case Intrinsic::x86_avx2_psra_d: {
- unsigned Opcode;
- switch (IntNo) {
- default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
- case Intrinsic::x86_sse2_psll_w:
- case Intrinsic::x86_sse2_psll_d:
- case Intrinsic::x86_sse2_psll_q:
- case Intrinsic::x86_avx2_psll_w:
- case Intrinsic::x86_avx2_psll_d:
- case Intrinsic::x86_avx2_psll_q:
- Opcode = X86ISD::VSHL;
- break;
- case Intrinsic::x86_sse2_psrl_w:
- case Intrinsic::x86_sse2_psrl_d:
- case Intrinsic::x86_sse2_psrl_q:
- case Intrinsic::x86_avx2_psrl_w:
- case Intrinsic::x86_avx2_psrl_d:
- case Intrinsic::x86_avx2_psrl_q:
- Opcode = X86ISD::VSRL;
- break;
- case Intrinsic::x86_sse2_psra_w:
- case Intrinsic::x86_sse2_psra_d:
- case Intrinsic::x86_avx2_psra_w:
- case Intrinsic::x86_avx2_psra_d:
- Opcode = X86ISD::VSRA;
- break;
- }
- return DAG.getNode(Opcode, dl, Op.getValueType(),
- Op.getOperand(1), Op.getOperand(2));
- }
-
- // SSE/AVX immediate shift intrinsics
- case Intrinsic::x86_sse2_pslli_w:
- case Intrinsic::x86_sse2_pslli_d:
- case Intrinsic::x86_sse2_pslli_q:
- case Intrinsic::x86_avx2_pslli_w:
- case Intrinsic::x86_avx2_pslli_d:
- case Intrinsic::x86_avx2_pslli_q:
- case Intrinsic::x86_sse2_psrli_w:
- case Intrinsic::x86_sse2_psrli_d:
- case Intrinsic::x86_sse2_psrli_q:
- case Intrinsic::x86_avx2_psrli_w:
- case Intrinsic::x86_avx2_psrli_d:
- case Intrinsic::x86_avx2_psrli_q:
- case Intrinsic::x86_sse2_psrai_w:
- case Intrinsic::x86_sse2_psrai_d:
- case Intrinsic::x86_avx2_psrai_w:
- case Intrinsic::x86_avx2_psrai_d: {
- unsigned Opcode;
- switch (IntNo) {
- default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
- case Intrinsic::x86_sse2_pslli_w:
- case Intrinsic::x86_sse2_pslli_d:
- case Intrinsic::x86_sse2_pslli_q:
- case Intrinsic::x86_avx2_pslli_w:
- case Intrinsic::x86_avx2_pslli_d:
- case Intrinsic::x86_avx2_pslli_q:
- Opcode = X86ISD::VSHLI;
- break;
- case Intrinsic::x86_sse2_psrli_w:
- case Intrinsic::x86_sse2_psrli_d:
- case Intrinsic::x86_sse2_psrli_q:
- case Intrinsic::x86_avx2_psrli_w:
- case Intrinsic::x86_avx2_psrli_d:
- case Intrinsic::x86_avx2_psrli_q:
- Opcode = X86ISD::VSRLI;
- break;
- case Intrinsic::x86_sse2_psrai_w:
- case Intrinsic::x86_sse2_psrai_d:
- case Intrinsic::x86_avx2_psrai_w:
- case Intrinsic::x86_avx2_psrai_d:
- Opcode = X86ISD::VSRAI;
- break;
- }
- return getTargetVShiftNode(Opcode, dl, Op.getSimpleValueType(),
- Op.getOperand(1), Op.getOperand(2), DAG);
+ SDValue LHS = Op.getOperand(1);
+ SDValue RHS = Op.getOperand(2);
+ unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
+ SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
+ SDValue CC = DAG.getConstant(X86CC, MVT::i8);
+ SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
+ return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
+ }
+ case Intrinsic::x86_avx512_kortestz_w:
+ case Intrinsic::x86_avx512_kortestc_w: {
+ unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
+ SDValue LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(1));
+ SDValue RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(2));
+ SDValue CC = DAG.getConstant(X86CC, MVT::i8);
+ SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
+ SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
+ return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
}
case Intrinsic::x86_sse42_pcmpistria128:
SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
return DAG.getNode(Opcode, dl, VTs, NewOps);
}
+
+ case Intrinsic::x86_fma_mask_vfmadd_ps_512:
+ case Intrinsic::x86_fma_mask_vfmadd_pd_512:
+ case Intrinsic::x86_fma_mask_vfmsub_ps_512:
+ case Intrinsic::x86_fma_mask_vfmsub_pd_512:
+ case Intrinsic::x86_fma_mask_vfnmadd_ps_512:
+ case Intrinsic::x86_fma_mask_vfnmadd_pd_512:
+ case Intrinsic::x86_fma_mask_vfnmsub_ps_512:
+ case Intrinsic::x86_fma_mask_vfnmsub_pd_512:
+ case Intrinsic::x86_fma_mask_vfmaddsub_ps_512:
+ case Intrinsic::x86_fma_mask_vfmaddsub_pd_512:
+ case Intrinsic::x86_fma_mask_vfmsubadd_ps_512:
+ case Intrinsic::x86_fma_mask_vfmsubadd_pd_512: {
+ auto *SAE = cast<ConstantSDNode>(Op.getOperand(5));
+ if (SAE->getZExtValue() == X86::STATIC_ROUNDING::CUR_DIRECTION)
+ return getVectorMaskingNode(DAG.getNode(getOpcodeForFMAIntrinsic(IntNo),
+ dl, Op.getValueType(),
+ Op.getOperand(1),
+ Op.getOperand(2),
+ Op.getOperand(3)),
+ Op.getOperand(4), Op.getOperand(1), DAG);
+ else
+ return SDValue();
+ }
+
case Intrinsic::x86_fma_vfmadd_ps:
case Intrinsic::x86_fma_vfmadd_pd:
case Intrinsic::x86_fma_vfmsub_ps:
case Intrinsic::x86_fma_vfmaddsub_pd_256:
case Intrinsic::x86_fma_vfmsubadd_ps_256:
case Intrinsic::x86_fma_vfmsubadd_pd_256:
- case Intrinsic::x86_fma_vfmadd_ps_512:
- case Intrinsic::x86_fma_vfmadd_pd_512:
- case Intrinsic::x86_fma_vfmsub_ps_512:
- case Intrinsic::x86_fma_vfmsub_pd_512:
- case Intrinsic::x86_fma_vfnmadd_ps_512:
- case Intrinsic::x86_fma_vfnmadd_pd_512:
- case Intrinsic::x86_fma_vfnmsub_ps_512:
- case Intrinsic::x86_fma_vfnmsub_pd_512:
- case Intrinsic::x86_fma_vfmaddsub_ps_512:
- case Intrinsic::x86_fma_vfmaddsub_pd_512:
- case Intrinsic::x86_fma_vfmsubadd_ps_512:
- case Intrinsic::x86_fma_vfmsubadd_pd_512: {
- unsigned Opc;
- switch (IntNo) {
- default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
- case Intrinsic::x86_fma_vfmadd_ps:
- case Intrinsic::x86_fma_vfmadd_pd:
- case Intrinsic::x86_fma_vfmadd_ps_256:
- case Intrinsic::x86_fma_vfmadd_pd_256:
- case Intrinsic::x86_fma_vfmadd_ps_512:
- case Intrinsic::x86_fma_vfmadd_pd_512:
- Opc = X86ISD::FMADD;
- break;
- case Intrinsic::x86_fma_vfmsub_ps:
- case Intrinsic::x86_fma_vfmsub_pd:
- case Intrinsic::x86_fma_vfmsub_ps_256:
- case Intrinsic::x86_fma_vfmsub_pd_256:
- case Intrinsic::x86_fma_vfmsub_ps_512:
- case Intrinsic::x86_fma_vfmsub_pd_512:
- Opc = X86ISD::FMSUB;
- break;
- case Intrinsic::x86_fma_vfnmadd_ps:
- case Intrinsic::x86_fma_vfnmadd_pd:
- case Intrinsic::x86_fma_vfnmadd_ps_256:
- case Intrinsic::x86_fma_vfnmadd_pd_256:
- case Intrinsic::x86_fma_vfnmadd_ps_512:
- case Intrinsic::x86_fma_vfnmadd_pd_512:
- Opc = X86ISD::FNMADD;
- break;
- case Intrinsic::x86_fma_vfnmsub_ps:
- case Intrinsic::x86_fma_vfnmsub_pd:
- case Intrinsic::x86_fma_vfnmsub_ps_256:
- case Intrinsic::x86_fma_vfnmsub_pd_256:
- case Intrinsic::x86_fma_vfnmsub_ps_512:
- case Intrinsic::x86_fma_vfnmsub_pd_512:
- Opc = X86ISD::FNMSUB;
- break;
- case Intrinsic::x86_fma_vfmaddsub_ps:
- case Intrinsic::x86_fma_vfmaddsub_pd:
- case Intrinsic::x86_fma_vfmaddsub_ps_256:
- case Intrinsic::x86_fma_vfmaddsub_pd_256:
- case Intrinsic::x86_fma_vfmaddsub_ps_512:
- case Intrinsic::x86_fma_vfmaddsub_pd_512:
- Opc = X86ISD::FMADDSUB;
- break;
- case Intrinsic::x86_fma_vfmsubadd_ps:
- case Intrinsic::x86_fma_vfmsubadd_pd:
- case Intrinsic::x86_fma_vfmsubadd_ps_256:
- case Intrinsic::x86_fma_vfmsubadd_pd_256:
- case Intrinsic::x86_fma_vfmsubadd_ps_512:
- case Intrinsic::x86_fma_vfmsubadd_pd_512:
- Opc = X86ISD::FMSUBADD;
- break;
- }
-
- return DAG.getNode(Opc, dl, Op.getValueType(), Op.getOperand(1),
- Op.getOperand(2), Op.getOperand(3));
- }
+ return DAG.getNode(getOpcodeForFMAIntrinsic(IntNo), dl, Op.getValueType(),
+ Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
}
}
return DAG.getMergeValues(Results, DL);
}
-enum IntrinsicType {
- GATHER, SCATTER, PREFETCH, RDSEED, RDRAND, RDPMC, RDTSC, XTEST
-};
-
-struct IntrinsicData {
- IntrinsicData(IntrinsicType IType, unsigned IOpc0, unsigned IOpc1)
- :Type(IType), Opc0(IOpc0), Opc1(IOpc1) {}
- IntrinsicType Type;
- unsigned Opc0;
- unsigned Opc1;
-};
-
-std::map < unsigned, IntrinsicData> IntrMap;
-static void InitIntinsicsMap() {
- static bool Initialized = false;
- if (Initialized)
- return;
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gather_qps_512,
- IntrinsicData(GATHER, X86::VGATHERQPSZrm, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gather_qps_512,
- IntrinsicData(GATHER, X86::VGATHERQPSZrm, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gather_qpd_512,
- IntrinsicData(GATHER, X86::VGATHERQPDZrm, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gather_dpd_512,
- IntrinsicData(GATHER, X86::VGATHERDPDZrm, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gather_dps_512,
- IntrinsicData(GATHER, X86::VGATHERDPSZrm, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gather_qpi_512,
- IntrinsicData(GATHER, X86::VPGATHERQDZrm, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gather_qpq_512,
- IntrinsicData(GATHER, X86::VPGATHERQQZrm, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gather_dpi_512,
- IntrinsicData(GATHER, X86::VPGATHERDDZrm, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gather_dpq_512,
- IntrinsicData(GATHER, X86::VPGATHERDQZrm, 0)));
-
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatter_qps_512,
- IntrinsicData(SCATTER, X86::VSCATTERQPSZmr, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatter_qpd_512,
- IntrinsicData(SCATTER, X86::VSCATTERQPDZmr, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatter_dpd_512,
- IntrinsicData(SCATTER, X86::VSCATTERDPDZmr, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatter_dps_512,
- IntrinsicData(SCATTER, X86::VSCATTERDPSZmr, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatter_qpi_512,
- IntrinsicData(SCATTER, X86::VPSCATTERQDZmr, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatter_qpq_512,
- IntrinsicData(SCATTER, X86::VPSCATTERQQZmr, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatter_dpi_512,
- IntrinsicData(SCATTER, X86::VPSCATTERDDZmr, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatter_dpq_512,
- IntrinsicData(SCATTER, X86::VPSCATTERDQZmr, 0)));
-
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gatherpf_qps_512,
- IntrinsicData(PREFETCH, X86::VGATHERPF0QPSm,
- X86::VGATHERPF1QPSm)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gatherpf_qpd_512,
- IntrinsicData(PREFETCH, X86::VGATHERPF0QPDm,
- X86::VGATHERPF1QPDm)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gatherpf_dpd_512,
- IntrinsicData(PREFETCH, X86::VGATHERPF0DPDm,
- X86::VGATHERPF1DPDm)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_gatherpf_dps_512,
- IntrinsicData(PREFETCH, X86::VGATHERPF0DPSm,
- X86::VGATHERPF1DPSm)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatterpf_qps_512,
- IntrinsicData(PREFETCH, X86::VSCATTERPF0QPSm,
- X86::VSCATTERPF1QPSm)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatterpf_qpd_512,
- IntrinsicData(PREFETCH, X86::VSCATTERPF0QPDm,
- X86::VSCATTERPF1QPDm)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatterpf_dpd_512,
- IntrinsicData(PREFETCH, X86::VSCATTERPF0DPDm,
- X86::VSCATTERPF1DPDm)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_avx512_scatterpf_dps_512,
- IntrinsicData(PREFETCH, X86::VSCATTERPF0DPSm,
- X86::VSCATTERPF1DPSm)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_rdrand_16,
- IntrinsicData(RDRAND, X86ISD::RDRAND, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_rdrand_32,
- IntrinsicData(RDRAND, X86ISD::RDRAND, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_rdrand_64,
- IntrinsicData(RDRAND, X86ISD::RDRAND, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_rdseed_16,
- IntrinsicData(RDSEED, X86ISD::RDSEED, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_rdseed_32,
- IntrinsicData(RDSEED, X86ISD::RDSEED, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_rdseed_64,
- IntrinsicData(RDSEED, X86ISD::RDSEED, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_xtest,
- IntrinsicData(XTEST, X86ISD::XTEST, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_rdtsc,
- IntrinsicData(RDTSC, X86ISD::RDTSC_DAG, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_rdtscp,
- IntrinsicData(RDTSC, X86ISD::RDTSCP_DAG, 0)));
- IntrMap.insert(std::make_pair(Intrinsic::x86_rdpmc,
- IntrinsicData(RDPMC, X86ISD::RDPMC_DAG, 0)));
- Initialized = true;
-}
static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
SelectionDAG &DAG) {
- InitIntinsicsMap();
unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
- std::map < unsigned, IntrinsicData>::const_iterator itr = IntrMap.find(IntNo);
- if (itr == IntrMap.end())
+
+ const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
+ if (!IntrData)
return SDValue();
SDLoc dl(Op);
- IntrinsicData Intr = itr->second;
- switch(Intr.Type) {
+ switch(IntrData->Type) {
+ default:
+ llvm_unreachable("Unknown Intrinsic Type");
+ break;
case RDSEED:
case RDRAND: {
// Emit the node with the right value type.
SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
- SDValue Result = DAG.getNode(Intr.Opc0, dl, VTs, Op.getOperand(0));
+ SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
// 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 Index = Op.getOperand(4);
SDValue Mask = Op.getOperand(5);
SDValue Scale = Op.getOperand(6);
- return getGatherNode(Intr.Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain,
+ return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain,
Subtarget);
}
case SCATTER: {
SDValue Index = Op.getOperand(4);
SDValue Src = Op.getOperand(5);
SDValue Scale = Op.getOperand(6);
- return getScatterNode(Intr.Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain);
+ return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain);
}
case PREFETCH: {
SDValue Hint = Op.getOperand(6);
if (dyn_cast<ConstantSDNode> (Hint) == nullptr ||
(HintVal = dyn_cast<ConstantSDNode> (Hint)->getZExtValue()) > 1)
llvm_unreachable("Wrong prefetch hint in intrinsic: should be 0 or 1");
- unsigned Opcode = (HintVal ? Intr.Opc1 : Intr.Opc0);
+ unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
SDValue Chain = Op.getOperand(0);
SDValue Mask = Op.getOperand(2);
SDValue Index = Op.getOperand(3);
// Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
case RDTSC: {
SmallVector<SDValue, 2> Results;
- getReadTimeStampCounter(Op.getNode(), dl, Intr.Opc0, DAG, Subtarget, Results);
+ getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget, Results);
return DAG.getMergeValues(Results, dl);
}
// Read Performance Monitoring Counters.
// XTEST intrinsics.
case XTEST: {
SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
- SDValue InTrans = DAG.getNode(X86ISD::XTEST, dl, VTs, Op.getOperand(0));
+ SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
DAG.getConstant(X86::COND_NE, MVT::i8),
InTrans);
return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
Ret, SDValue(InTrans.getNode(), 1));
}
+ // ADC/ADCX/SBB
+ case ADX: {
+ SmallVector<SDValue, 2> Results;
+ SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
+ SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
+ SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
+ DAG.getConstant(-1, MVT::i8));
+ SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
+ Op.getOperand(4), GenCF.getValue(1));
+ SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
+ Op.getOperand(5), MachinePointerInfo(),
+ false, false, 0);
+ SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
+ DAG.getConstant(X86::COND_B, MVT::i8),
+ Res.getValue(1));
+ Results.push_back(SetCC);
+ Results.push_back(Store);
+ return DAG.getMergeValues(Results, dl);
+ }
}
- llvm_unreachable("Unknown Intrinsic Type");
}
SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
if (Depth > 0) {
SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ DAG.getSubtarget().getRegisterInfo());
SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), PtrVT);
return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
DAG.getNode(ISD::ADD, dl, PtrVT,
EVT VT = Op.getValueType();
SDLoc dl(Op); // FIXME probably not meaningful
unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ DAG.getSubtarget().getRegisterInfo());
unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
(FrameReg == X86::EBP && VT == MVT::i32)) &&
SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
SelectionDAG &DAG) const {
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ DAG.getSubtarget().getRegisterInfo());
return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize());
}
SDLoc dl (Op);
EVT PtrVT = getPointerTy();
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ DAG.getSubtarget().getRegisterInfo());
unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
(FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
SDLoc dl (Op);
const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
- const TargetRegisterInfo* TRI = DAG.getTarget().getRegisterInfo();
+ const TargetRegisterInfo *TRI = DAG.getSubtarget().getRegisterInfo();
if (Subtarget->is64Bit()) {
SDValue OutChains[6];
MachineFunction &MF = DAG.getMachineFunction();
const TargetMachine &TM = MF.getTarget();
- const TargetFrameLowering &TFI = *TM.getFrameLowering();
+ const TargetFrameLowering &TFI = *TM.getSubtargetImpl()->getFrameLowering();
unsigned StackAlignment = TFI.getStackAlignment();
MVT VT = Op.getSimpleValueType();
SDLoc DL(Op);
DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
// Shuffle it back into the right order.
- // The internal representation is big endian.
- // In other words, a i64 bitcasted to 2 x i32 has its high part at index 0
- // and its low part at index 1.
- // Moreover, we have: Mul1 = <ae|cg> ; Mul2 = <bf|dh>
- // Vector index 0 1 ; 2 3
- // We want <ae|bf|cg|dh>
- // Vector index 0 2 1 3
- // Since each element is seen as 2 x i32, we get:
- // high_mask[i] = 2 x vector_index[i]
- // low_mask[i] = 2 x vector_index[i] + 1
- // where vector_index = {0, Size/2, 1, Size/2 + 1, ...,
- // Size/2 - 1, Size/2 + Size/2 - 1}
- // where Size is the number of element of the final vector.
SDValue Highs, Lows;
if (VT == MVT::v8i32) {
- const int HighMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
+ const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
- const int LowMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
+ const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
} else {
- const int HighMask[] = {0, 4, 2, 6};
+ const int HighMask[] = {1, 5, 3, 7};
Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
- const int LowMask[] = {1, 5, 3, 7};
+ const int LowMask[] = {0, 4, 2, 6};
Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
}
Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
}
- // The low part of a MUL_LOHI is supposed to be the first value and the
- // high part the second value.
- return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getValueType(), Lows, Highs);
+ // The first result of MUL_LOHI is actually the low value, followed by the
+ // high value.
+ SDValue Ops[] = {Lows, Highs};
+ return DAG.getMergeValues(Ops, dl);
}
static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
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_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::FABS: return LowerFABS(Op, DAG);
- case ISD::FNEG: return LowerFNEG(Op, DAG);
+ case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, 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);
case ISD::ATOMIC_LOAD_UMIN:
case ISD::ATOMIC_LOAD_UMAX:
// Delegate to generic TypeLegalization. Situations we can really handle
- // should have already been dealt with by X86AtomicExpand.cpp.
+ // should have already been dealt with by X86AtomicExpandPass.cpp.
break;
case ISD::ATOMIC_LOAD: {
ReplaceATOMIC_LOAD(N, Results, DAG);
case X86ISD::PACKSS: return "X86ISD::PACKSS";
case X86ISD::PACKUS: return "X86ISD::PACKUS";
case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
+ case X86ISD::VALIGN: return "X86ISD::VALIGN";
case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
// Machine Information
- const TargetInstrInfo *TII = MBB->getParent()->getTarget().getInstrInfo();
+ const TargetInstrInfo *TII = MBB->getParent()->getSubtarget().getInstrInfo();
MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
XMMSaveMBB->addSuccessor(EndMBB);
// Now add the instructions.
- const TargetInstrInfo *TII = MBB->getParent()->getTarget().getInstrInfo();
+ const TargetInstrInfo *TII = MBB->getParent()->getSubtarget().getInstrInfo();
DebugLoc DL = MI->getDebugLoc();
unsigned CountReg = MI->getOperand(0).getReg();
MachineBasicBlock *
X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
MachineBasicBlock *BB) const {
- const TargetInstrInfo *TII = BB->getParent()->getTarget().getInstrInfo();
+ const TargetInstrInfo *TII = BB->getParent()->getSubtarget().getInstrInfo();
DebugLoc DL = MI->getDebugLoc();
// To "insert" a SELECT_CC instruction, we actually have to insert the
// If the EFLAGS register isn't dead in the terminator, then claim that it's
// live into the sink and copy blocks.
- const TargetRegisterInfo* TRI = BB->getParent()->getTarget().getRegisterInfo();
+ const TargetRegisterInfo *TRI =
+ BB->getParent()->getSubtarget().getRegisterInfo();
if (!MI->killsRegister(X86::EFLAGS) &&
!checkAndUpdateEFLAGSKill(MI, BB, TRI)) {
copy0MBB->addLiveIn(X86::EFLAGS);
X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI, MachineBasicBlock *BB,
bool Is64Bit) const {
MachineFunction *MF = BB->getParent();
- const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
+ const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
DebugLoc DL = MI->getDebugLoc();
const BasicBlock *LLVM_BB = BB->getBasicBlock();
BuildMI(bumpMBB, DL, TII->get(X86::JMP_4)).addMBB(continueMBB);
// Calls into a routine in libgcc to allocate more space from the heap.
- const uint32_t *RegMask =
- MF->getTarget().getRegisterInfo()->getCallPreservedMask(CallingConv::C);
+ const uint32_t *RegMask = MF->getTarget()
+ .getSubtargetImpl()
+ ->getRegisterInfo()
+ ->getCallPreservedMask(CallingConv::C);
if (Is64Bit) {
BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
.addReg(sizeVReg);
MachineBasicBlock *
X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
MachineBasicBlock *BB) const {
- const TargetInstrInfo *TII = BB->getParent()->getTarget().getInstrInfo();
+ const TargetInstrInfo *TII = BB->getParent()->getSubtarget().getInstrInfo();
DebugLoc DL = MI->getDebugLoc();
assert(!Subtarget->isTargetMacho());
// or EAX and doing an indirect call. The return value will then
// be in the normal return register.
MachineFunction *F = BB->getParent();
- const X86InstrInfo *TII
- = static_cast<const X86InstrInfo*>(F->getTarget().getInstrInfo());
+ const X86InstrInfo *TII =
+ static_cast<const X86InstrInfo *>(F->getSubtarget().getInstrInfo());
DebugLoc DL = MI->getDebugLoc();
assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
// Get a register mask for the lowered call.
// FIXME: The 32-bit calls have non-standard calling conventions. Use a
// proper register mask.
- const uint32_t *RegMask =
- F->getTarget().getRegisterInfo()->getCallPreservedMask(CallingConv::C);
+ const uint32_t *RegMask = F->getTarget()
+ .getSubtargetImpl()
+ ->getRegisterInfo()
+ ->getCallPreservedMask(CallingConv::C);
if (Subtarget->is64Bit()) {
MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
TII->get(X86::MOV64rm), X86::RDI)
MachineBasicBlock *MBB) const {
DebugLoc DL = MI->getDebugLoc();
MachineFunction *MF = MBB->getParent();
- const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
+ const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
MachineRegisterInfo &MRI = MF->getRegInfo();
const BasicBlock *BB = MBB->getBasicBlock();
MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
.addMBB(restoreMBB);
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(MF->getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ MF->getSubtarget().getRegisterInfo());
MIB.addRegMask(RegInfo->getNoPreservedMask());
thisMBB->addSuccessor(mainMBB);
thisMBB->addSuccessor(restoreMBB);
MachineBasicBlock *MBB) const {
DebugLoc DL = MI->getDebugLoc();
MachineFunction *MF = MBB->getParent();
- const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
+ const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
MachineRegisterInfo &MRI = MF->getRegInfo();
// Memory Reference
(PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
unsigned Tmp = MRI.createVirtualRegister(RC);
// Since FP is only updated here but NOT referenced, it's treated as GPR.
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(MF->getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ MF->getSubtarget().getRegisterInfo());
unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
unsigned SP = RegInfo->getStackRegister();
default: llvm_unreachable("Unrecognized FMA variant.");
}
- const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo();
+ const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
MachineInstrBuilder MIB =
BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
.addOperand(MI->getOperand(0))
case X86::FP80_TO_INT32_IN_MEM:
case X86::FP80_TO_INT64_IN_MEM: {
MachineFunction *F = BB->getParent();
- const TargetInstrInfo *TII = F->getTarget().getInstrInfo();
+ const TargetInstrInfo *TII = F->getSubtarget().getInstrInfo();
DebugLoc DL = MI->getDebugLoc();
// Change the floating point control register to use "round towards zero"
case X86::VPCMPESTRM128MEM:
assert(Subtarget->hasSSE42() &&
"Target must have SSE4.2 or AVX features enabled");
- return EmitPCMPSTRM(MI, BB, BB->getParent()->getTarget().getInstrInfo());
+ return EmitPCMPSTRM(MI, BB, BB->getParent()->getSubtarget().getInstrInfo());
// String/text processing lowering.
case X86::PCMPISTRIREG:
case X86::VPCMPESTRIMEM:
assert(Subtarget->hasSSE42() &&
"Target must have SSE4.2 or AVX features enabled");
- return EmitPCMPSTRI(MI, BB, BB->getParent()->getTarget().getInstrInfo());
+ return EmitPCMPSTRI(MI, BB, BB->getParent()->getSubtarget().getInstrInfo());
// Thread synchronization.
case X86::MONITOR:
- return EmitMonitor(MI, BB, BB->getParent()->getTarget().getInstrInfo(), Subtarget);
+ return EmitMonitor(MI, BB, BB->getParent()->getSubtarget().getInstrInfo(),
+ Subtarget);
// xbegin
case X86::XBEGIN:
- return EmitXBegin(MI, BB, BB->getParent()->getTarget().getInstrInfo());
+ return EmitXBegin(MI, BB, BB->getParent()->getSubtarget().getInstrInfo());
case X86::VASTART_SAVE_XMM_REGS:
return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
return SDValue();
}
+/// \brief Combine an arbitrary chain of shuffles into a single instruction if
+/// possible.
+///
+/// This is the leaf of the recursive combinine below. When we have found some
+/// chain of single-use x86 shuffle instructions and accumulated the combined
+/// shuffle mask represented by them, this will try to pattern match that mask
+/// into either a single instruction if there is a special purpose instruction
+/// for this operation, or into a PSHUFB instruction which is a fully general
+/// instruction but should only be used to replace chains over a certain depth.
+static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
+ int Depth, bool HasPSHUFB, SelectionDAG &DAG,
+ TargetLowering::DAGCombinerInfo &DCI,
+ const X86Subtarget *Subtarget) {
+ assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
+
+ // Find the operand that enters the chain. Note that multiple uses are OK
+ // here, we're not going to remove the operand we find.
+ SDValue Input = Op.getOperand(0);
+ while (Input.getOpcode() == ISD::BITCAST)
+ Input = Input.getOperand(0);
+
+ MVT VT = Input.getSimpleValueType();
+ MVT RootVT = Root.getSimpleValueType();
+ SDLoc DL(Root);
+
+ // Just remove no-op shuffle masks.
+ if (Mask.size() == 1) {
+ DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Input),
+ /*AddTo*/ true);
+ return true;
+ }
+
+ // Use the float domain if the operand type is a floating point type.
+ bool FloatDomain = VT.isFloatingPoint();
+
+ // If we don't have access to VEX encodings, the generic PSHUF instructions
+ // are preferable to some of the specialized forms despite requiring one more
+ // byte to encode because they can implicitly copy.
+ //
+ // IF we *do* have VEX encodings, than we can use shorter, more specific
+ // shuffle instructions freely as they can copy due to the extra register
+ // operand.
+ if (Subtarget->hasAVX()) {
+ // We have both floating point and integer variants of shuffles that dup
+ // either the low or high half of the vector.
+ if (Mask.equals(0, 0) || Mask.equals(1, 1)) {
+ bool Lo = Mask.equals(0, 0);
+ unsigned Shuffle = FloatDomain ? (Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS)
+ : (Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH);
+ if (Depth == 1 && Root->getOpcode() == Shuffle)
+ return false; // Nothing to do!
+ MVT ShuffleVT = FloatDomain ? MVT::v4f32 : MVT::v2i64;
+ Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
+ DCI.AddToWorklist(Op.getNode());
+ Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
+ DCI.AddToWorklist(Op.getNode());
+ DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
+ /*AddTo*/ true);
+ return true;
+ }
+
+ // FIXME: We should match UNPCKLPS and UNPCKHPS here.
+
+ // For the integer domain we have specialized instructions for duplicating
+ // any element size from the low or high half.
+ if (!FloatDomain &&
+ (Mask.equals(0, 0, 1, 1) || Mask.equals(2, 2, 3, 3) ||
+ Mask.equals(0, 0, 1, 1, 2, 2, 3, 3) ||
+ Mask.equals(4, 4, 5, 5, 6, 6, 7, 7) ||
+ Mask.equals(0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7) ||
+ Mask.equals(8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15,
+ 15))) {
+ bool Lo = Mask[0] == 0;
+ unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
+ if (Depth == 1 && Root->getOpcode() == Shuffle)
+ return false; // Nothing to do!
+ MVT ShuffleVT;
+ switch (Mask.size()) {
+ case 4: ShuffleVT = MVT::v4i32; break;
+ case 8: ShuffleVT = MVT::v8i16; break;
+ case 16: ShuffleVT = MVT::v16i8; break;
+ };
+ Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
+ DCI.AddToWorklist(Op.getNode());
+ Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
+ DCI.AddToWorklist(Op.getNode());
+ DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
+ /*AddTo*/ true);
+ return true;
+ }
+ }
+
+ // Don't try to re-form single instruction chains under any circumstances now
+ // that we've done encoding canonicalization for them.
+ if (Depth < 2)
+ return false;
+
+ // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
+ // can replace them with a single PSHUFB instruction profitably. Intel's
+ // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
+ // in practice PSHUFB tends to be *very* fast so we're more aggressive.
+ if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
+ SmallVector<SDValue, 16> PSHUFBMask;
+ assert(Mask.size() <= 16 && "Can't shuffle elements smaller than bytes!");
+ int Ratio = 16 / Mask.size();
+ for (unsigned i = 0; i < 16; ++i) {
+ int M = Mask[i / Ratio] != SM_SentinelZero
+ ? Ratio * Mask[i / Ratio] + i % Ratio
+ : 255;
+ PSHUFBMask.push_back(DAG.getConstant(M, MVT::i8));
+ }
+ Op = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Input);
+ DCI.AddToWorklist(Op.getNode());
+ SDValue PSHUFBMaskOp =
+ DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, PSHUFBMask);
+ DCI.AddToWorklist(PSHUFBMaskOp.getNode());
+ Op = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, Op, PSHUFBMaskOp);
+ DCI.AddToWorklist(Op.getNode());
+ DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
+ /*AddTo*/ true);
+ return true;
+ }
+
+ // Failed to find any combines.
+ return false;
+}
+
+/// \brief Fully generic combining of x86 shuffle instructions.
+///
+/// This should be the last combine run over the x86 shuffle instructions. Once
+/// they have been fully optimized, this will recursively consider all chains
+/// of single-use shuffle instructions, build a generic model of the cumulative
+/// shuffle operation, and check for simpler instructions which implement this
+/// operation. We use this primarily for two purposes:
+///
+/// 1) Collapse generic shuffles to specialized single instructions when
+/// equivalent. In most cases, this is just an encoding size win, but
+/// sometimes we will collapse multiple generic shuffles into a single
+/// special-purpose shuffle.
+/// 2) Look for sequences of shuffle instructions with 3 or more total
+/// instructions, and replace them with the slightly more expensive SSSE3
+/// PSHUFB instruction if available. We do this as the last combining step
+/// to ensure we avoid using PSHUFB if we can implement the shuffle with
+/// a suitable short sequence of other instructions. The PHUFB will either
+/// use a register or have to read from memory and so is slightly (but only
+/// slightly) more expensive than the other shuffle instructions.
+///
+/// Because this is inherently a quadratic operation (for each shuffle in
+/// a chain, we recurse up the chain), the depth is limited to 8 instructions.
+/// This should never be an issue in practice as the shuffle lowering doesn't
+/// produce sequences of more than 8 instructions.
+///
+/// FIXME: We will currently miss some cases where the redundant shuffling
+/// would simplify under the threshold for PSHUFB formation because of
+/// combine-ordering. To fix this, we should do the redundant instruction
+/// combining in this recursive walk.
+static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
+ ArrayRef<int> RootMask,
+ int Depth, bool HasPSHUFB,
+ SelectionDAG &DAG,
+ TargetLowering::DAGCombinerInfo &DCI,
+ const X86Subtarget *Subtarget) {
+ // Bound the depth of our recursive combine because this is ultimately
+ // quadratic in nature.
+ if (Depth > 8)
+ return false;
+
+ // Directly rip through bitcasts to find the underlying operand.
+ while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
+ Op = Op.getOperand(0);
+
+ MVT VT = Op.getSimpleValueType();
+ if (!VT.isVector())
+ return false; // Bail if we hit a non-vector.
+ // FIXME: This routine should be taught about 256-bit shuffles, or a 256-bit
+ // version should be added.
+ if (VT.getSizeInBits() != 128)
+ return false;
+
+ assert(Root.getSimpleValueType().isVector() &&
+ "Shuffles operate on vector types!");
+ assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
+ "Can only combine shuffles of the same vector register size.");
+
+ if (!isTargetShuffle(Op.getOpcode()))
+ return false;
+ SmallVector<int, 16> OpMask;
+ bool IsUnary;
+ bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
+ // We only can combine unary shuffles which we can decode the mask for.
+ if (!HaveMask || !IsUnary)
+ return false;
+
+ assert(VT.getVectorNumElements() == OpMask.size() &&
+ "Different mask size from vector size!");
+ assert(((RootMask.size() > OpMask.size() &&
+ RootMask.size() % OpMask.size() == 0) ||
+ (OpMask.size() > RootMask.size() &&
+ OpMask.size() % RootMask.size() == 0) ||
+ OpMask.size() == RootMask.size()) &&
+ "The smaller number of elements must divide the larger.");
+ int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
+ int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
+ assert(((RootRatio == 1 && OpRatio == 1) ||
+ (RootRatio == 1) != (OpRatio == 1)) &&
+ "Must not have a ratio for both incoming and op masks!");
+
+ SmallVector<int, 16> Mask;
+ Mask.reserve(std::max(OpMask.size(), RootMask.size()));
+
+ // Merge this shuffle operation's mask into our accumulated mask. Note that
+ // this shuffle's mask will be the first applied to the input, followed by the
+ // root mask to get us all the way to the root value arrangement. The reason
+ // for this order is that we are recursing up the operation chain.
+ for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
+ int RootIdx = i / RootRatio;
+ if (RootMask[RootIdx] == SM_SentinelZero) {
+ // This is a zero-ed lane, we're done.
+ Mask.push_back(SM_SentinelZero);
+ continue;
+ }
+
+ int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
+ int OpIdx = RootMaskedIdx / OpRatio;
+ if (OpMask[OpIdx] == SM_SentinelZero) {
+ // The incoming lanes are zero, it doesn't matter which ones we are using.
+ Mask.push_back(SM_SentinelZero);
+ continue;
+ }
+
+ // Ok, we have non-zero lanes, map them through.
+ Mask.push_back(OpMask[OpIdx] * OpRatio +
+ RootMaskedIdx % OpRatio);
+ }
+
+ // See if we can recurse into the operand to combine more things.
+ switch (Op.getOpcode()) {
+ case X86ISD::PSHUFB:
+ HasPSHUFB = true;
+ case X86ISD::PSHUFD:
+ case X86ISD::PSHUFHW:
+ case X86ISD::PSHUFLW:
+ if (Op.getOperand(0).hasOneUse() &&
+ combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
+ HasPSHUFB, DAG, DCI, Subtarget))
+ return true;
+ break;
+
+ case X86ISD::UNPCKL:
+ case X86ISD::UNPCKH:
+ assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
+ // We can't check for single use, we have to check that this shuffle is the only user.
+ if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
+ combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
+ HasPSHUFB, DAG, DCI, Subtarget))
+ return true;
+ break;
+ }
+
+ // Minor canonicalization of the accumulated shuffle mask to make it easier
+ // to match below. All this does is detect masks with squential pairs of
+ // elements, and shrink them to the half-width mask. It does this in a loop
+ // so it will reduce the size of the mask to the minimal width mask which
+ // performs an equivalent shuffle.
+ while (Mask.size() > 1 && canWidenShuffleElements(Mask)) {
+ for (int i = 0, e = Mask.size() / 2; i < e; ++i)
+ Mask[i] = Mask[2 * i] / 2;
+ Mask.resize(Mask.size() / 2);
+ }
+
+ return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
+ Subtarget);
+}
+
/// \brief Get the PSHUF-style mask from PSHUF node.
///
/// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
/// We walk up the chain and look for a combinable shuffle, skipping over
/// shuffles that we could hoist this shuffle's transformation past without
/// altering anything.
-static bool combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
- SelectionDAG &DAG,
- TargetLowering::DAGCombinerInfo &DCI) {
+static SDValue
+combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
+ SelectionDAG &DAG,
+ TargetLowering::DAGCombinerInfo &DCI) {
assert(N.getOpcode() == X86ISD::PSHUFD &&
"Called with something other than an x86 128-bit half shuffle!");
SDLoc DL(N);
- // Walk up a single-use chain looking for a combinable shuffle.
+ // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
+ // of the shuffles in the chain so that we can form a fresh chain to replace
+ // this one.
+ SmallVector<SDValue, 8> Chain;
SDValue V = N.getOperand(0);
for (; V.hasOneUse(); V = V.getOperand(0)) {
switch (V.getOpcode()) {
default:
- return false; // Nothing combined!
+ return SDValue(); // Nothing combined!
case ISD::BITCAST:
// Skip bitcasts as we always know the type for the target specific
// dword shuffle, and the high words are self-contained.
if (Mask[0] != 0 || Mask[1] != 1 ||
!(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
- return false;
+ return SDValue();
+ Chain.push_back(V);
continue;
case X86ISD::PSHUFHW:
// dword shuffle, and the low words are self-contained.
if (Mask[2] != 2 || Mask[3] != 3 ||
!(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
- return false;
+ return SDValue();
+ Chain.push_back(V);
continue;
case X86ISD::UNPCKL:
// For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
// shuffle into a preceding word shuffle.
if (V.getValueType() != MVT::v16i8 && V.getValueType() != MVT::v8i16)
- return false;
+ return SDValue();
// Search for a half-shuffle which we can combine with.
unsigned CombineOp =
V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
if (V.getOperand(0) != V.getOperand(1) ||
!V->isOnlyUserOf(V.getOperand(0).getNode()))
- return false;
+ return SDValue();
+ Chain.push_back(V);
V = V.getOperand(0);
do {
switch (V.getOpcode()) {
default:
- return false; // Nothing to combine.
+ return SDValue(); // Nothing to combine.
case X86ISD::PSHUFLW:
case X86ISD::PSHUFHW:
if (V.getOpcode() == CombineOp)
break;
+ Chain.push_back(V);
+
// Fallthrough!
case ISD::BITCAST:
V = V.getOperand(0);
if (!V.hasOneUse())
// We fell out of the loop without finding a viable combining instruction.
- return false;
-
- // Record the old value to use in RAUW-ing.
- SDValue Old = V;
+ return SDValue();
// Merge this node's mask and our incoming mask.
SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
getV4X86ShuffleImm8ForMask(Mask, DAG));
- // It is possible that one of the combinable shuffles was completely absorbed
- // by the other, just replace it and revisit all users in that case.
- if (Old.getNode() == V.getNode()) {
- DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo=*/true);
- return true;
- }
+ // Rebuild the chain around this new shuffle.
+ while (!Chain.empty()) {
+ SDValue W = Chain.pop_back_val();
- // Replace N with its operand as we're going to combine that shuffle away.
- DAG.ReplaceAllUsesWith(N, N.getOperand(0));
+ if (V.getValueType() != W.getOperand(0).getValueType())
+ V = DAG.getNode(ISD::BITCAST, DL, W.getOperand(0).getValueType(), V);
- // Replace the combinable shuffle with the combined one, updating all users
- // so that we re-evaluate the chain here.
- DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
- return true;
+ switch (W.getOpcode()) {
+ default:
+ llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
+
+ case X86ISD::UNPCKL:
+ case X86ISD::UNPCKH:
+ V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
+ break;
+
+ case X86ISD::PSHUFD:
+ case X86ISD::PSHUFLW:
+ case X86ISD::PSHUFHW:
+ V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
+ break;
+ }
+ }
+ if (V.getValueType() != N.getValueType())
+ V = DAG.getNode(ISD::BITCAST, DL, N.getValueType(), V);
+
+ // Return the new chain to replace N.
+ return V;
}
/// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
// Other-half shuffles are no-ops.
continue;
-
- case X86ISD::PSHUFD: {
- // We can only handle pshufd if the half we are combining either stays in
- // its half, or switches to the other half. Bail if one of these isn't
- // true.
- SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
- int DOffset = CombineOpcode == X86ISD::PSHUFLW ? 0 : 2;
- if (!((VMask[DOffset + 0] < 2 && VMask[DOffset + 1] < 2) ||
- (VMask[DOffset + 0] >= 2 && VMask[DOffset + 1] >= 2)))
- return false;
-
- // Map the mask through the pshufd and keep walking up the chain.
- for (int i = 0; i < 4; ++i)
- Mask[i] = 2 * (VMask[DOffset + Mask[i] / 2] % 2) + Mask[i] % 2;
-
- // Switch halves if the pshufd does.
- CombineOpcode =
- VMask[DOffset + Mask[0] / 2] < 2 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
- continue;
- }
}
// Break out of the loop if we break out of the switch.
break;
// We fell out of the loop without finding a viable combining instruction.
return false;
- // Record the old value to use in RAUW-ing.
+ // Combine away the bottom node as its shuffle will be accumulated into
+ // a preceding shuffle.
+ DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
+
+ // Record the old value.
SDValue Old = V;
// Merge this node's mask and our incoming mask (adjusted to account for all
V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
getV4X86ShuffleImm8ForMask(Mask, DAG));
- // Replace N with its operand as we're going to combine that shuffle away.
- DAG.ReplaceAllUsesWith(N, N.getOperand(0));
+ // Check that the shuffles didn't cancel each other out. If not, we need to
+ // combine to the new one.
+ if (Old != V)
+ // Replace the combinable shuffle with the combined one, updating all users
+ // so that we re-evaluate the chain here.
+ DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
- // Replace the combinable shuffle with the combined one, updating all users
- // so that we re-evaluate the chain here.
- DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
return true;
}
// See if this reduces to a PSHUFD which is no more expensive and can
// combine with more operations.
- if (Mask[0] % 2 == 0 && Mask[2] % 2 == 0 &&
- areAdjacentMasksSequential(Mask)) {
+ if (canWidenShuffleElements(Mask)) {
int DMask[] = {-1, -1, -1, -1};
int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
DMask[DOffset + 0] = DOffset + Mask[0] / 2;
break;
case X86ISD::PSHUFD:
- if (combineRedundantDWordShuffle(N, Mask, DAG, DCI))
- return SDValue(); // We combined away this shuffle.
+ if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
+ return NewN;
break;
}
PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
if (Shuffle.getNode())
return Shuffle;
+
+ // Try recursively combining arbitrary sequences of x86 shuffle
+ // instructions into higher-order shuffles. We do this after combining
+ // specific PSHUF instruction sequences into their minimal form so that we
+ // can evaluate how many specialized shuffle instructions are involved in
+ // a particular chain.
+ SmallVector<int, 1> NonceMask; // Just a placeholder.
+ NonceMask.push_back(0);
+ if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
+ /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
+ DCI, Subtarget))
+ return SDValue(); // This routine will use CombineTo to replace N.
}
return SDValue();
EVT VT = InVec.getValueType();
- bool HasShuffleIntoBitcast = false;
if (InVec.getOpcode() == ISD::BITCAST) {
// Don't duplicate a load with other uses.
if (!InVec.hasOneUse())
if (BCVT.getVectorNumElements() != VT.getVectorNumElements())
return SDValue();
InVec = InVec.getOperand(0);
- HasShuffleIntoBitcast = true;
}
if (!isTargetShuffle(InVec.getOpcode()))
if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
return SDValue();
- if (HasShuffleIntoBitcast) {
- // If there's a bitcast before the shuffle, check if the load type and
- // alignment is valid.
- unsigned Align = LN0->getAlignment();
- const TargetLowering &TLI = DAG.getTargetLoweringInfo();
- unsigned NewAlign = TLI.getDataLayout()->
- getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
+ EVT EltVT = N->getValueType(0);
+ // If there's a bitcast before the shuffle, check if the load type and
+ // alignment is valid.
+ unsigned Align = LN0->getAlignment();
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+ unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
+ EltVT.getTypeForEVT(*DAG.getContext()));
- if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
- return SDValue();
- }
+ if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
+ return SDValue();
// All checks match so transform back to vector_shuffle so that DAG combiner
// can finish the job
if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
return SDValue();
+ // A vselect where all conditions and data are constants can be optimized into
+ // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
+ if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
+ ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
+ return SDValue();
+
unsigned MaskValue = 0;
if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
return SDValue();
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);
EVT MemVT = Ld->getMemoryVT();
SDLoc dl(Ld);
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
- unsigned RegSz = RegVT.getSizeInBits();
// On Sandybridge unaligned 256bit loads are inefficient.
ISD::LoadExtType Ext = Ld->getExtensionType();
return DCI.CombineTo(N, NewVec, TF, true);
}
- // If this is a vector EXT Load then attempt to optimize it using a
- // shuffle. If SSSE3 is not available we may emit an illegal shuffle but the
- // expansion is still better than scalar code.
- // We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise we'll
- // emit a shuffle and a arithmetic shift.
- // TODO: It is possible to support ZExt by zeroing the undef values
- // during the shuffle phase or after the shuffle.
- if (RegVT.isVector() && RegVT.isInteger() && Subtarget->hasSSE2() &&
- (Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)) {
- assert(MemVT != RegVT && "Cannot extend to the same type");
- assert(MemVT.isVector() && "Must load a vector from memory");
-
- unsigned NumElems = RegVT.getVectorNumElements();
- unsigned MemSz = MemVT.getSizeInBits();
- assert(RegSz > MemSz && "Register size must be greater than the mem size");
-
- if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256())
- return SDValue();
-
- // All sizes must be a power of two.
- if (!isPowerOf2_32(RegSz * MemSz * NumElems))
- return SDValue();
-
- // Attempt to load the original value using scalar loads.
- // Find the largest scalar type that divides the total loaded size.
- MVT SclrLoadTy = MVT::i8;
- for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
- tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
- MVT Tp = (MVT::SimpleValueType)tp;
- if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
- SclrLoadTy = Tp;
- }
- }
-
- // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
- if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
- (64 <= MemSz))
- SclrLoadTy = MVT::f64;
-
- // Calculate the number of scalar loads that we need to perform
- // in order to load our vector from memory.
- unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
- if (Ext == ISD::SEXTLOAD && NumLoads > 1)
- return SDValue();
-
- unsigned loadRegZize = RegSz;
- if (Ext == ISD::SEXTLOAD && RegSz == 256)
- loadRegZize /= 2;
-
- // Represent our vector as a sequence of elements which are the
- // largest scalar that we can load.
- EVT LoadUnitVecVT = EVT::getVectorVT(*DAG.getContext(), SclrLoadTy,
- loadRegZize/SclrLoadTy.getSizeInBits());
-
- // 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(),
- loadRegZize/MemVT.getScalarType().getSizeInBits());
-
- assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
- "Invalid vector type");
-
- // We can't shuffle using an illegal type.
- if (!TLI.isTypeLegal(WideVecVT))
- return SDValue();
-
- SmallVector<SDValue, 8> Chains;
- SDValue Ptr = Ld->getBasePtr();
- SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits()/8,
- TLI.getPointerTy());
- SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
-
- for (unsigned i = 0; i < NumLoads; ++i) {
- // Perform a single load.
- SDValue ScalarLoad = DAG.getLoad(SclrLoadTy, dl, Ld->getChain(),
- Ptr, Ld->getPointerInfo(),
- Ld->isVolatile(), Ld->isNonTemporal(),
- Ld->isInvariant(), Ld->getAlignment());
- Chains.push_back(ScalarLoad.getValue(1));
- // Create the first element type using SCALAR_TO_VECTOR in order to avoid
- // another round of DAGCombining.
- if (i == 0)
- Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
- else
- Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
- ScalarLoad, DAG.getIntPtrConstant(i));
-
- Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
- }
-
- SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
-
- // Bitcast the loaded value to a vector of the original element type, in
- // the size of the target vector type.
- SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
- unsigned SizeRatio = RegSz/MemSz;
-
- if (Ext == ISD::SEXTLOAD) {
- // If we have SSE4.1 we can directly emit a VSEXT node.
- if (Subtarget->hasSSE41()) {
- SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
- return DCI.CombineTo(N, Sext, TF, true);
- }
-
- // Otherwise we'll shuffle the small elements in the high bits of the
- // larger type and perform an arithmetic shift. If the shift is not legal
- // it's better to scalarize.
- if (!TLI.isOperationLegalOrCustom(ISD::SRA, RegVT))
- return SDValue();
-
- // Redistribute the loaded elements into the different locations.
- SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
- for (unsigned i = 0; i != NumElems; ++i)
- ShuffleVec[i*SizeRatio + SizeRatio-1] = i;
-
- SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
- DAG.getUNDEF(WideVecVT),
- &ShuffleVec[0]);
-
- Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
-
- // Build the arithmetic shift.
- unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
- MemVT.getVectorElementType().getSizeInBits();
- Shuff = DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
- DAG.getConstant(Amt, RegVT));
-
- return DCI.CombineTo(N, Shuff, TF, true);
- }
-
- // Redistribute the loaded elements into the different locations.
- SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
- for (unsigned i = 0; i != NumElems; ++i)
- ShuffleVec[i*SizeRatio] = i;
-
- SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
- DAG.getUNDEF(WideVecVT),
- &ShuffleVec[0]);
-
- // Bitcast to the requested type.
- Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
- // Replace the original load with the new sequence
- // and return the new chain.
- return DCI.CombineTo(N, Shuff, TF, true);
- }
-
return SDValue();
}
case X86ISD::UNPCKL:
case X86ISD::MOVHLPS:
case X86ISD::MOVLHPS:
+ case X86ISD::PSHUFB:
case X86ISD::PSHUFD:
case X86ISD::PSHUFHW:
case X86ISD::PSHUFLW:
Constraint[5] == ')' &&
Constraint[6] == '}') {
- Res.first = X86::ST0+Constraint[4]-'0';
+ Res.first = X86::FP0+Constraint[4]-'0';
Res.second = &X86::RFP80RegClass;
return Res;
}
// GCC allows "st(0)" to be called just plain "st".
if (StringRef("{st}").equals_lower(Constraint)) {
- Res.first = X86::ST0;
+ Res.first = X86::FP0;
Res.second = &X86::RFP80RegClass;
return Res;
}
projects / oota-llvm.git / blobdiff