#include "X86ISelLowering.h"
#include "Utils/X86ShuffleDecode.h"
#include "X86CallingConv.h"
+#include "X86FrameLowering.h"
#include "X86InstrBuilder.h"
#include "X86MachineFunctionInfo.h"
#include "X86TargetMachine.h"
"rather than promotion."),
cl::Hidden);
-static cl::opt<bool> ExperimentalVectorShuffleLowering(
- "x86-experimental-vector-shuffle-lowering", cl::init(true),
- cl::desc("Enable an experimental vector shuffle lowering code path."),
- cl::Hidden);
-
static cl::opt<int> ReciprocalEstimateRefinementSteps(
"x86-recip-refinement-steps", cl::init(1),
cl::desc("Specify the number of Newton-Raphson iterations applied to the "
return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
}
-// FIXME: This should stop caching the target machine as soon as
-// we can remove resetOperationActions et al.
-X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM)
- : TargetLowering(TM) {
- Subtarget = &TM.getSubtarget<X86Subtarget>();
+X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
+ const X86Subtarget &STI)
+ : TargetLowering(TM), Subtarget(&STI) {
X86ScalarSSEf64 = Subtarget->hasSSE2();
X86ScalarSSEf32 = Subtarget->hasSSE1();
TD = getDataLayout();
- resetOperationActions();
-}
-
-void X86TargetLowering::resetOperationActions() {
- const TargetMachine &TM = getTargetMachine();
- static bool FirstTimeThrough = true;
-
- // If none of the target options have changed, then we don't need to reset the
- // operation actions.
- if (!FirstTimeThrough && TO == TM.Options) return;
-
- if (!FirstTimeThrough) {
- // Reinitialize the actions.
- initActions();
- FirstTimeThrough = false;
- }
-
- TO = TM.Options;
-
// Set up the TargetLowering object.
static const MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
setSchedulingPreference(Sched::ILP);
else
setSchedulingPreference(Sched::RegPressure);
- const X86RegisterInfo *RegInfo =
- TM.getSubtarget<X86Subtarget>().getRegisterInfo();
+ const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
// Bypass expensive divides on Atom when compiling with O2.
setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
- // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like types,
- // we have to deal with them whether we ask for Expansion or not. Setting
- // Expand causes its own optimisation problems though, so leave them legal.
+ // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
+ // types, we have to deal with them whether we ask for Expansion or not.
+ // Setting Expand causes its own optimisation problems though, so leave
+ // them legal.
if (VT.getVectorElementType() == MVT::i1)
setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
}
setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
+ setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
continue;
setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
+ setOperationAction(ISD::VSELECT, VT, Custom);
setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
}
setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
+ setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
+ setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
// FIXME: Do we need to handle scalar-to-vector here?
setOperationAction(ISD::MUL, MVT::v4i32, Legal);
- setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
- setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
- setOperationAction(ISD::VSELECT, MVT::v4i32, Custom);
- setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
- setOperationAction(ISD::VSELECT, MVT::v8i16, Custom);
- // There is no BLENDI for byte vectors. We don't need to custom lower
- // some vselects for now.
+ // We directly match byte blends in the backend as they match the VSELECT
+ // condition form.
setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
// SSE41 brings specific instructions for doing vector sign extend even in
setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
}
+ // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
+
+ setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
+ setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
+ setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
+ setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
+ setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
+ setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
+
// 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
setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
- setOperationAction(ISD::VSELECT, MVT::v4f64, Custom);
- setOperationAction(ISD::VSELECT, MVT::v4i64, Custom);
- setOperationAction(ISD::VSELECT, MVT::v8i32, Custom);
- setOperationAction(ISD::VSELECT, MVT::v8f32, Custom);
-
setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
- setOperationAction(ISD::VSELECT, MVT::v16i16, Custom);
- setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
-
// The custom lowering for UINT_TO_FP for v8i32 becomes interesting
// when we have a 256bit-wide blend with immediate.
setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
// Custom CTPOP always performs better on natively supported v8i32
setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
+
+ // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
+
+ setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
+ setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
+ setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
+ setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
+ setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
+ setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
} else {
setOperationAction(ISD::ADD, MVT::v4i64, Custom);
setOperationAction(ISD::ADD, MVT::v8i32, Custom);
setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
+ setOperationAction(ISD::VSELECT, VT, Custom);
setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
}
+ if (Subtarget->hasInt256())
+ setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
+
+
// Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
MVT VT = (MVT::SimpleValueType)i;
setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
+ setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
+ setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
+ setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
+ setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
+ setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
+ setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
+ setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
+ setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
+ setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
+ setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
+
setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
setOperationAction(ISD::XOR, MVT::v4i32, Legal);
}
- // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
- // of this type with custom code.
- for (MVT VT : MVT::vector_valuetypes())
- setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Custom);
-
// We want to custom lower some of our intrinsics.
setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
// We have target-specific dag combine patterns for the following nodes:
setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
+ setTargetDAGCombine(ISD::BITCAST);
setTargetDAGCombine(ISD::VSELECT);
setTargetDAGCombine(ISD::SELECT);
setTargetDAGCombine(ISD::SHL);
setTargetDAGCombine(ISD::FMA);
setTargetDAGCombine(ISD::SUB);
setTargetDAGCombine(ISD::LOAD);
+ setTargetDAGCombine(ISD::MLOAD);
setTargetDAGCombine(ISD::STORE);
+ setTargetDAGCombine(ISD::MSTORE);
setTargetDAGCombine(ISD::ZERO_EXTEND);
setTargetDAGCombine(ISD::ANY_EXTEND);
setTargetDAGCombine(ISD::SIGN_EXTEND);
setTargetDAGCombine(ISD::SETCC);
setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
setTargetDAGCombine(ISD::BUILD_VECTOR);
- if (Subtarget->is64Bit())
- setTargetDAGCombine(ISD::MUL);
+ setTargetDAGCombine(ISD::MUL);
setTargetDAGCombine(ISD::XOR);
- computeRegisterProperties();
+ computeRegisterProperties(Subtarget->getRegisterInfo());
// On Darwin, -Os means optimize for size without hurting performance,
// do not reduce the limit.
MachineFunction &MF) const {
const Function *F = MF.getFunction();
if ((!IsMemset || ZeroMemset) &&
- !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
- Attribute::NoImplicitFloat)) {
+ !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
if (Size >= 16 &&
(Subtarget->isUnalignedMemAccessFast() ||
((DstAlign == 0 || DstAlign >= 16) &&
return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
}
-// FIXME: Why this routine is here? Move to RegInfo!
-std::pair<const TargetRegisterClass*, uint8_t>
-X86TargetLowering::findRepresentativeClass(MVT VT) const{
+std::pair<const TargetRegisterClass *, uint8_t>
+X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
+ MVT VT) const {
const TargetRegisterClass *RRC = nullptr;
uint8_t Cost = 1;
switch (VT.SimpleTy) {
default:
- return TargetLowering::findRepresentativeClass(VT);
+ return TargetLowering::findRepresentativeClass(TRI, VT);
case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
break;
// Win32 requires us to put the sret argument to %eax as well.
// We saved the argument into a virtual register in the entry block,
// so now we copy the value out and into %rax/%eax.
- if (DAG.getMachineFunction().getFunction()->hasStructRetAttr() &&
- (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC())) {
- MachineFunction &MF = DAG.getMachineFunction();
- X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
- unsigned Reg = FuncInfo->getSRetReturnReg();
- assert(Reg &&
- "SRetReturnReg should have been set in LowerFormalArguments().");
- SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
+ //
+ // Checking Function.hasStructRetAttr() here is insufficient because the IR
+ // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
+ // false, then an sret argument may be implicitly inserted in the SelDAG. In
+ // either case FuncInfo->setSRetReturnReg() will have been called.
+ if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
+ assert((Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC()) &&
+ "No need for an sret register");
+ SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg, getPointerTy());
unsigned RetValReg
= (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
}
const Function *Fn = MF.getFunction();
- bool NoImplicitFloatOps = Fn->getAttributes().
- hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
+ bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
assert(!(MF.getTarget().Options.UseSoftFloat && NoImplicitFloatOps) &&
"SSE register cannot be used when SSE is disabled!");
if (MF.getTarget().Options.UseSoftFloat || NoImplicitFloatOps ||
// Figure out if XMM registers are in use.
assert(!(MF.getTarget().Options.UseSoftFloat &&
- Fn->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
- Attribute::NoImplicitFloat)) &&
+ Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
"SSE register cannot be used when SSE is disabled!");
// 64-bit calling conventions support varargs and register parameters, so we
// 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());
+ unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
+ unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
"SSE register cannot be used when SSE is disabled!");
}
if (IsWin64) {
- const TargetFrameLowering &TFI = *MF.getSubtarget().getFrameLowering();
+ const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
// Get to the caller-allocated home save location. Add 8 to account
// for the return address.
int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
// 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.getSubtarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = Subtarget->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;
X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
};
- unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
+ unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
assert((Subtarget->hasSSE1() || !NumXMMRegs)
&& "SSE registers cannot be used when SSE is disabled");
// unless we're building with the leopard linker or later, which
// automatically synthesizes these stubs.
OpFlags = X86II::MO_DARWIN_STUB;
- } else if (Subtarget->isPICStyleRIPRel() &&
- isa<Function>(GV) &&
- cast<Function>(GV)->getAttributes().
- hasAttribute(AttributeSet::FunctionIndex,
- Attribute::NonLazyBind)) {
+ } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
+ cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
// If the function is marked as non-lazy, generate an indirect call
// which loads from the GOT directly. This avoids runtime overhead
// at the cost of eager binding (and one extra byte of encoding).
Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
OpFlags);
- } else if (Subtarget->isTarget64BitILP32() && Callee->getValueType(0) == MVT::i32) {
+ } else if (Subtarget->isTarget64BitILP32() &&
+ Callee->getValueType(0) == MVT::i32) {
// Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
}
RegsToPass[i].second.getValueType()));
// Add a register mask operand representing the call-preserved registers.
- const TargetRegisterInfo *TRI = DAG.getSubtarget().getRegisterInfo();
+ const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
assert(Mask && "Missing call preserved mask for calling convention");
Ops.push_back(DAG.getRegisterMask(Mask));
unsigned
X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
SelectionDAG& DAG) const {
- MachineFunction &MF = DAG.getMachineFunction();
- const TargetMachine &TM = MF.getTarget();
- const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
- TM.getSubtargetImpl()->getRegisterInfo());
- const TargetFrameLowering &TFI = *TM.getSubtargetImpl()->getFrameLowering();
+ const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
+ const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
unsigned StackAlignment = TFI.getStackAlignment();
uint64_t AlignMask = StackAlignment - 1;
int64_t Offset = StackSize;
return false;
} else {
unsigned Opcode = Def->getOpcode();
- if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
+ if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
+ Opcode == X86::LEA64_32r) &&
Def->getOperand(1).isFI()) {
FI = Def->getOperand(1).getIndex();
Bytes = Flags.getByValSize();
bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
+ // Win64 functions have extra shadow space for argument homing. Don't do the
+ // sibcall if the caller and callee have mismatched expectations for this
+ // space.
+ if (IsCalleeWin64 != IsCallerWin64)
+ return false;
+
if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
if (IsTailCallConvention(CalleeCC) && CCMatch)
return true;
// 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.getSubtarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
if (RegInfo->needsStackRealignment(MF))
return false;
// the caller's fixed stack objects.
MachineFrameInfo *MFI = MF.getFrameInfo();
const MachineRegisterInfo *MRI = &MF.getRegInfo();
- const X86InstrInfo *TII =
- static_cast<const X86InstrInfo *>(DAG.getSubtarget().getInstrInfo());
+ const X86InstrInfo *TII = Subtarget->getInstrInfo();
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
CCValAssign &VA = ArgLocs[i];
SDValue Arg = OutVals[i];
}
}
-static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
- SDValue V1, SelectionDAG &DAG) {
- switch(Opc) {
- default: llvm_unreachable("Unknown x86 shuffle node");
- case X86ISD::MOVSHDUP:
- case X86ISD::MOVSLDUP:
- case X86ISD::MOVDDUP:
- return DAG.getNode(Opc, dl, VT, V1);
- }
-}
-
static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
SDValue V1, unsigned TargetMask,
SelectionDAG &DAG) {
}
}
-static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
- SDValue V1, SDValue V2, unsigned TargetMask,
- SelectionDAG &DAG) {
- 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,
- DAG.getConstant(TargetMask, MVT::i8));
- }
-}
-
static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
SDValue V1, SDValue V2, SelectionDAG &DAG) {
switch(Opc) {
SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
MachineFunction &MF = DAG.getMachineFunction();
- const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
- DAG.getSubtarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
int ReturnAddrIndex = FuncInfo->getRAIndex();
return true;
}
-/// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
-/// is suitable for input to PSHUFD. That is, it doesn't reference the other
-/// operand - by default will match for first operand.
-static bool isPSHUFDMask(ArrayRef<int> Mask, MVT VT,
- bool TestSecondOperand = false) {
- if (VT != MVT::v4f32 && VT != MVT::v4i32 &&
- VT != MVT::v2f64 && VT != MVT::v2i64)
- return false;
-
- unsigned NumElems = VT.getVectorNumElements();
- unsigned Lo = TestSecondOperand ? NumElems : 0;
- unsigned Hi = Lo + NumElems;
-
- for (unsigned i = 0; i < NumElems; ++i)
- if (!isUndefOrInRange(Mask[i], (int)Lo, (int)Hi))
- return false;
-
- return true;
-}
-
-/// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
-/// is suitable for input to PSHUFHW.
-static bool isPSHUFHWMask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
- if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16))
- return false;
-
- // Lower quadword copied in order or undef.
- if (!isSequentialOrUndefInRange(Mask, 0, 4, 0))
- return false;
-
- // Upper quadword shuffled.
- for (unsigned i = 4; i != 8; ++i)
- if (!isUndefOrInRange(Mask[i], 4, 8))
- return false;
-
- if (VT == MVT::v16i16) {
- // Lower quadword copied in order or undef.
- if (!isSequentialOrUndefInRange(Mask, 8, 4, 8))
- return false;
-
- // Upper quadword shuffled.
- for (unsigned i = 12; i != 16; ++i)
- if (!isUndefOrInRange(Mask[i], 12, 16))
- return false;
- }
-
- return true;
-}
-
-/// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
-/// is suitable for input to PSHUFLW.
-static bool isPSHUFLWMask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
- if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16))
- return false;
-
- // Upper quadword copied in order.
- if (!isSequentialOrUndefInRange(Mask, 4, 4, 4))
- return false;
-
- // Lower quadword shuffled.
- for (unsigned i = 0; i != 4; ++i)
- if (!isUndefOrInRange(Mask[i], 0, 4))
- return false;
-
- if (VT == MVT::v16i16) {
- // Upper quadword copied in order.
- if (!isSequentialOrUndefInRange(Mask, 12, 4, 12))
- return false;
-
- // Lower quadword shuffled.
- for (unsigned i = 8; i != 12; ++i)
- if (!isUndefOrInRange(Mask[i], 8, 12))
- return false;
- }
-
- return true;
-}
-
-/// \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 = InterLane ? 1: VT.getSizeInBits()/128;
- unsigned NumLaneElts = NumElts/NumLanes;
-
- // Do not handle 64-bit element shuffles with palignr.
- if (NumLaneElts == 2)
- return false;
-
- for (unsigned l = 0; l != NumElts; l+=NumLaneElts) {
- unsigned i;
- for (i = 0; i != NumLaneElts; ++i) {
- if (Mask[i+l] >= 0)
- break;
- }
-
- // Lane is all undef, go to next lane
- if (i == NumLaneElts)
- continue;
-
- int Start = Mask[i+l];
-
- // Make sure its in this lane in one of the sources
- if (!isUndefOrInRange(Start, l, l+NumLaneElts) &&
- !isUndefOrInRange(Start, l+NumElts, l+NumElts+NumLaneElts))
- return false;
-
- // If not lane 0, then we must match lane 0
- if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Start, Mask[i]+l))
- return false;
-
- // Correct second source to be contiguous with first source
- if (Start >= (int)NumElts)
- Start -= NumElts - NumLaneElts;
-
- // Make sure we're shifting in the right direction.
- if (Start <= (int)(i+l))
- return false;
-
- Start -= i;
-
- // Check the rest of the elements to see if they are consecutive.
- for (++i; i != NumLaneElts; ++i) {
- int Idx = Mask[i+l];
-
- // Make sure its in this lane
- if (!isUndefOrInRange(Idx, l, l+NumLaneElts) &&
- !isUndefOrInRange(Idx, l+NumElts, l+NumElts+NumLaneElts))
- return false;
-
- // If not lane 0, then we must match lane 0
- if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Idx, Mask[i]+l))
- return false;
-
- if (Idx >= (int)NumElts)
- Idx -= NumElts - NumLaneElts;
-
- if (!isUndefOrEqual(Idx, Start+i))
- return false;
-
- }
- }
-
- 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,
}
}
-/// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to 128/256-bit
-/// SHUFPS and SHUFPD. If Commuted is true, then it checks for sources to be
-/// reverse of what x86 shuffles want.
-static bool isSHUFPMask(ArrayRef<int> Mask, MVT VT, bool Commuted = false) {
-
- unsigned NumElems = VT.getVectorNumElements();
- unsigned NumLanes = VT.getSizeInBits()/128;
- unsigned NumLaneElems = NumElems/NumLanes;
-
- if (NumLaneElems != 2 && NumLaneElems != 4)
+/// isVEXTRACTIndex - Return true if the specified
+/// EXTRACT_SUBVECTOR operand specifies a vector extract that is
+/// suitable for instruction that extract 128 or 256 bit vectors
+static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
+ assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
+ if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
return false;
- unsigned EltSize = VT.getVectorElementType().getSizeInBits();
- bool symetricMaskRequired =
- (VT.getSizeInBits() >= 256) && (EltSize == 32);
+ // The index should be aligned on a vecWidth-bit boundary.
+ uint64_t Index =
+ cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
- // VSHUFPSY divides the resulting vector into 4 chunks.
- // The sources are also splitted into 4 chunks, and each destination
- // chunk must come from a different source chunk.
- //
- // SRC1 => X7 X6 X5 X4 X3 X2 X1 X0
- // SRC2 => Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y9
- //
- // DST => Y7..Y4, Y7..Y4, X7..X4, X7..X4,
- // Y3..Y0, Y3..Y0, X3..X0, X3..X0
- //
- // VSHUFPDY divides the resulting vector into 4 chunks.
- // The sources are also splitted into 4 chunks, and each destination
- // chunk must come from a different source chunk.
- //
- // SRC1 => X3 X2 X1 X0
- // SRC2 => Y3 Y2 Y1 Y0
- //
- // DST => Y3..Y2, X3..X2, Y1..Y0, X1..X0
- //
- SmallVector<int, 4> MaskVal(NumLaneElems, -1);
- unsigned HalfLaneElems = NumLaneElems/2;
- for (unsigned l = 0; l != NumElems; l += NumLaneElems) {
- for (unsigned i = 0; i != NumLaneElems; ++i) {
- int Idx = Mask[i+l];
- unsigned RngStart = l + ((Commuted == (i<HalfLaneElems)) ? NumElems : 0);
- if (!isUndefOrInRange(Idx, RngStart, RngStart+NumLaneElems))
- return false;
- // For VSHUFPSY, the mask of the second half must be the same as the
- // first but with the appropriate offsets. This works in the same way as
- // VPERMILPS works with masks.
- if (!symetricMaskRequired || Idx < 0)
- continue;
- if (MaskVal[i] < 0) {
- MaskVal[i] = Idx - l;
- continue;
- }
- if ((signed)(Idx - l) != MaskVal[i])
- return false;
- }
- }
+ MVT VT = N->getSimpleValueType(0);
+ unsigned ElSize = VT.getVectorElementType().getSizeInBits();
+ bool Result = (Index * ElSize) % vecWidth == 0;
- return true;
+ return Result;
}
-/// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to MOVHLPS.
-static bool isMOVHLPSMask(ArrayRef<int> Mask, MVT VT) {
- if (!VT.is128BitVector())
+/// isVINSERTIndex - Return true if the specified INSERT_SUBVECTOR
+/// operand specifies a subvector insert that is suitable for input to
+/// insertion of 128 or 256-bit subvectors
+static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
+ assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
+ if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
return false;
+ // The index should be aligned on a vecWidth-bit boundary.
+ uint64_t Index =
+ cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
- unsigned NumElems = VT.getVectorNumElements();
-
- if (NumElems != 4)
- return false;
+ MVT VT = N->getSimpleValueType(0);
+ unsigned ElSize = VT.getVectorElementType().getSizeInBits();
+ bool Result = (Index * ElSize) % vecWidth == 0;
- // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
- return isUndefOrEqual(Mask[0], 6) &&
- isUndefOrEqual(Mask[1], 7) &&
- isUndefOrEqual(Mask[2], 2) &&
- isUndefOrEqual(Mask[3], 3);
+ return Result;
}
-/// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
-/// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
-/// <2, 3, 2, 3>
-static bool isMOVHLPS_v_undef_Mask(ArrayRef<int> Mask, MVT VT) {
- if (!VT.is128BitVector())
- return false;
-
- unsigned NumElems = VT.getVectorNumElements();
-
- if (NumElems != 4)
- return false;
+bool X86::isVINSERT128Index(SDNode *N) {
+ return isVINSERTIndex(N, 128);
+}
- return isUndefOrEqual(Mask[0], 2) &&
- isUndefOrEqual(Mask[1], 3) &&
- isUndefOrEqual(Mask[2], 2) &&
- isUndefOrEqual(Mask[3], 3);
+bool X86::isVINSERT256Index(SDNode *N) {
+ return isVINSERTIndex(N, 256);
}
-/// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
-static bool isMOVLPMask(ArrayRef<int> Mask, MVT VT) {
- if (!VT.is128BitVector())
- return false;
+bool X86::isVEXTRACT128Index(SDNode *N) {
+ return isVEXTRACTIndex(N, 128);
+}
- unsigned NumElems = VT.getVectorNumElements();
+bool X86::isVEXTRACT256Index(SDNode *N) {
+ return isVEXTRACTIndex(N, 256);
+}
- if (NumElems != 2 && NumElems != 4)
- return false;
+static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
+ assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
+ if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
+ llvm_unreachable("Illegal extract subvector for VEXTRACT");
- for (unsigned i = 0, e = NumElems/2; i != e; ++i)
- if (!isUndefOrEqual(Mask[i], i + NumElems))
- return false;
+ uint64_t Index =
+ cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
- for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
- if (!isUndefOrEqual(Mask[i], i))
- return false;
+ MVT VecVT = N->getOperand(0).getSimpleValueType();
+ MVT ElVT = VecVT.getVectorElementType();
- return true;
+ unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
+ return Index / NumElemsPerChunk;
}
-/// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to MOVLHPS.
-static bool isMOVLHPSMask(ArrayRef<int> Mask, MVT VT) {
- if (!VT.is128BitVector())
- return false;
-
- unsigned NumElems = VT.getVectorNumElements();
-
- if (NumElems != 2 && NumElems != 4)
- return false;
+static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
+ assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
+ if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
+ llvm_unreachable("Illegal insert subvector for VINSERT");
- for (unsigned i = 0, e = NumElems/2; i != e; ++i)
- if (!isUndefOrEqual(Mask[i], i))
- return false;
+ uint64_t Index =
+ cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
- for (unsigned i = 0, e = NumElems/2; i != e; ++i)
- if (!isUndefOrEqual(Mask[i + e], i + NumElems))
- return false;
+ MVT VecVT = N->getSimpleValueType(0);
+ MVT ElVT = VecVT.getVectorElementType();
- return true;
+ unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
+ return Index / NumElemsPerChunk;
}
-/// isINSERTPSMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to INSERTPS.
-/// i. e: If all but one element come from the same vector.
-static bool isINSERTPSMask(ArrayRef<int> Mask, MVT VT) {
- // TODO: Deal with AVX's VINSERTPS
- if (!VT.is128BitVector() || (VT != MVT::v4f32 && VT != MVT::v4i32))
- return false;
-
- unsigned CorrectPosV1 = 0;
- unsigned CorrectPosV2 = 0;
- for (int i = 0, e = (int)VT.getVectorNumElements(); i != e; ++i) {
- if (Mask[i] == -1) {
- ++CorrectPosV1;
- ++CorrectPosV2;
- continue;
- }
+/// getExtractVEXTRACT128Immediate - Return the appropriate immediate
+/// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
+/// and VINSERTI128 instructions.
+unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
+ return getExtractVEXTRACTImmediate(N, 128);
+}
- if (Mask[i] == i)
- ++CorrectPosV1;
- else if (Mask[i] == i + 4)
- ++CorrectPosV2;
- }
+/// getExtractVEXTRACT256Immediate - Return the appropriate immediate
+/// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF64x4
+/// and VINSERTI64x4 instructions.
+unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
+ return getExtractVEXTRACTImmediate(N, 256);
+}
- if (CorrectPosV1 == 3 || CorrectPosV2 == 3)
- // We have 3 elements (undefs count as elements from any vector) from one
- // vector, and one from another.
- return true;
+/// getInsertVINSERT128Immediate - Return the appropriate immediate
+/// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
+/// and VINSERTI128 instructions.
+unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
+ return getInsertVINSERTImmediate(N, 128);
+}
- return false;
+/// getInsertVINSERT256Immediate - Return the appropriate immediate
+/// to insert at the specified INSERT_SUBVECTOR index with VINSERTF46x4
+/// and VINSERTI64x4 instructions.
+unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
+ return getInsertVINSERTImmediate(N, 256);
}
-//
-// Some special combinations that can be optimized.
-//
-static
-SDValue Compact8x32ShuffleNode(ShuffleVectorSDNode *SVOp,
- SelectionDAG &DAG) {
- MVT VT = SVOp->getSimpleValueType(0);
- SDLoc dl(SVOp);
-
- if (VT != MVT::v8i32 && VT != MVT::v8f32)
- return SDValue();
-
- ArrayRef<int> Mask = SVOp->getMask();
-
- // These are the special masks that may be optimized.
- static const int MaskToOptimizeEven[] = {0, 8, 2, 10, 4, 12, 6, 14};
- static const int MaskToOptimizeOdd[] = {1, 9, 3, 11, 5, 13, 7, 15};
- bool MatchEvenMask = true;
- bool MatchOddMask = true;
- for (int i=0; i<8; ++i) {
- if (!isUndefOrEqual(Mask[i], MaskToOptimizeEven[i]))
- MatchEvenMask = false;
- if (!isUndefOrEqual(Mask[i], MaskToOptimizeOdd[i]))
- MatchOddMask = false;
- }
-
- if (!MatchEvenMask && !MatchOddMask)
- return SDValue();
-
- SDValue UndefNode = DAG.getNode(ISD::UNDEF, dl, VT);
-
- SDValue Op0 = SVOp->getOperand(0);
- SDValue Op1 = SVOp->getOperand(1);
-
- if (MatchEvenMask) {
- // Shift the second operand right to 32 bits.
- static const int ShiftRightMask[] = {-1, 0, -1, 2, -1, 4, -1, 6 };
- Op1 = DAG.getVectorShuffle(VT, dl, Op1, UndefNode, ShiftRightMask);
- } else {
- // Shift the first operand left to 32 bits.
- static const int ShiftLeftMask[] = {1, -1, 3, -1, 5, -1, 7, -1 };
- Op0 = DAG.getVectorShuffle(VT, dl, Op0, UndefNode, ShiftLeftMask);
- }
- static const int BlendMask[] = {0, 9, 2, 11, 4, 13, 6, 15};
- return DAG.getVectorShuffle(VT, dl, Op0, Op1, BlendMask);
+/// isZero - Returns true if Elt is a constant integer zero
+static bool isZero(SDValue V) {
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
+ return C && C->isNullValue();
}
-/// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to UNPCKL.
-static bool isUNPCKLMask(ArrayRef<int> Mask, MVT VT,
- bool HasInt256, bool V2IsSplat = false) {
-
- assert(VT.getSizeInBits() >= 128 &&
- "Unsupported vector type for unpckl");
-
- unsigned NumElts = VT.getVectorNumElements();
- if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
- (!HasInt256 || (NumElts != 16 && NumElts != 32)))
- return false;
-
- assert((!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) &&
- "Unsupported vector type for unpckh");
-
- // AVX defines UNPCK* to operate independently on 128-bit lanes.
- unsigned NumLanes = VT.getSizeInBits()/128;
- unsigned NumLaneElts = NumElts/NumLanes;
-
- for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
- for (unsigned i = 0, j = l; i != NumLaneElts; i += 2, ++j) {
- int BitI = Mask[l+i];
- int BitI1 = Mask[l+i+1];
- if (!isUndefOrEqual(BitI, j))
- return false;
- if (V2IsSplat) {
- if (!isUndefOrEqual(BitI1, NumElts))
- return false;
- } else {
- if (!isUndefOrEqual(BitI1, j + NumElts))
- return false;
- }
- }
- }
-
- return true;
+/// isZeroNode - Returns true if Elt is a constant zero or a floating point
+/// constant +0.0.
+bool X86::isZeroNode(SDValue Elt) {
+ if (isZero(Elt))
+ return true;
+ if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
+ return CFP->getValueAPF().isPosZero();
+ return false;
}
-/// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to UNPCKH.
-static bool isUNPCKHMask(ArrayRef<int> Mask, MVT VT,
- bool HasInt256, bool V2IsSplat = false) {
- assert(VT.getSizeInBits() >= 128 &&
- "Unsupported vector type for unpckh");
-
- unsigned NumElts = VT.getVectorNumElements();
- if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
- (!HasInt256 || (NumElts != 16 && NumElts != 32)))
- return false;
-
- assert((!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) &&
- "Unsupported vector type for unpckh");
-
- // AVX defines UNPCK* to operate independently on 128-bit lanes.
- unsigned NumLanes = VT.getSizeInBits()/128;
- unsigned NumLaneElts = NumElts/NumLanes;
+/// getZeroVector - Returns a vector of specified type with all zero elements.
+///
+static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
+ SelectionDAG &DAG, SDLoc dl) {
+ assert(VT.isVector() && "Expected a vector type");
- for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
- for (unsigned i = 0, j = l+NumLaneElts/2; i != NumLaneElts; i += 2, ++j) {
- int BitI = Mask[l+i];
- int BitI1 = Mask[l+i+1];
- if (!isUndefOrEqual(BitI, j))
- return false;
- if (V2IsSplat) {
- if (isUndefOrEqual(BitI1, NumElts))
- return false;
- } else {
- if (!isUndefOrEqual(BitI1, j+NumElts))
- return false;
- }
+ // Always build SSE zero vectors as <4 x i32> bitcasted
+ // to their dest type. This ensures they get CSE'd.
+ SDValue Vec;
+ if (VT.is128BitVector()) { // SSE
+ if (Subtarget->hasSSE2()) { // SSE2
+ SDValue Cst = DAG.getConstant(0, MVT::i32);
+ Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
+ } else { // SSE1
+ SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32);
+ Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
}
- }
- return true;
-}
-
-/// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
-/// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
-/// <0, 0, 1, 1>
-static bool isUNPCKL_v_undef_Mask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
- unsigned NumElts = VT.getVectorNumElements();
- bool Is256BitVec = VT.is256BitVector();
-
- if (VT.is512BitVector())
- return false;
- assert((VT.is128BitVector() || VT.is256BitVector()) &&
- "Unsupported vector type for unpckh");
-
- if (Is256BitVec && NumElts != 4 && NumElts != 8 &&
- (!HasInt256 || (NumElts != 16 && NumElts != 32)))
- return false;
-
- // For 256-bit i64/f64, use MOVDDUPY instead, so reject the matching pattern
- // FIXME: Need a better way to get rid of this, there's no latency difference
- // between UNPCKLPD and MOVDDUP, the later should always be checked first and
- // the former later. We should also remove the "_undef" special mask.
- if (NumElts == 4 && Is256BitVec)
- return false;
-
- // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
- // independently on 128-bit lanes.
- unsigned NumLanes = VT.getSizeInBits()/128;
- unsigned NumLaneElts = NumElts/NumLanes;
-
- for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
- for (unsigned i = 0, j = l; i != NumLaneElts; i += 2, ++j) {
- int BitI = Mask[l+i];
- int BitI1 = Mask[l+i+1];
-
- if (!isUndefOrEqual(BitI, j))
- return false;
- if (!isUndefOrEqual(BitI1, j))
- return false;
+ } else if (VT.is256BitVector()) { // AVX
+ if (Subtarget->hasInt256()) { // AVX2
+ SDValue Cst = DAG.getConstant(0, MVT::i32);
+ SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
+ Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
+ } else {
+ // 256-bit logic and arithmetic instructions in AVX are all
+ // floating-point, no support for integer ops. Emit fp zeroed vectors.
+ SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32);
+ SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
+ Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
}
- }
+ } else if (VT.is512BitVector()) { // AVX-512
+ SDValue Cst = DAG.getConstant(0, MVT::i32);
+ SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
+ Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
+ Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
+ } else if (VT.getScalarType() == MVT::i1) {
+ assert(VT.getVectorNumElements() <= 16 && "Unexpected vector type");
+ SDValue Cst = DAG.getConstant(0, MVT::i1);
+ SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
+ } else
+ llvm_unreachable("Unexpected vector type");
- return true;
+ return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
}
-/// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
-/// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
-/// <2, 2, 3, 3>
-static bool isUNPCKH_v_undef_Mask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
- unsigned NumElts = VT.getVectorNumElements();
-
- if (VT.is512BitVector())
- return false;
-
- assert((VT.is128BitVector() || VT.is256BitVector()) &&
- "Unsupported vector type for unpckh");
-
- if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
- (!HasInt256 || (NumElts != 16 && NumElts != 32)))
- return false;
-
- // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
- // independently on 128-bit lanes.
- unsigned NumLanes = VT.getSizeInBits()/128;
- unsigned NumLaneElts = NumElts/NumLanes;
+/// getOnesVector - Returns a vector of specified type with all bits set.
+/// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
+/// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
+/// Then bitcast to their original type, ensuring they get CSE'd.
+static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
+ SDLoc dl) {
+ assert(VT.isVector() && "Expected a vector type");
- for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
- for (unsigned i = 0, j = l+NumLaneElts/2; i != NumLaneElts; i += 2, ++j) {
- int BitI = Mask[l+i];
- int BitI1 = Mask[l+i+1];
- if (!isUndefOrEqual(BitI, j))
- return false;
- if (!isUndefOrEqual(BitI1, j))
- return false;
+ SDValue Cst = DAG.getConstant(~0U, MVT::i32);
+ SDValue Vec;
+ if (VT.is256BitVector()) {
+ if (HasInt256) { // AVX2
+ SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
+ Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
+ } else { // AVX
+ Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
+ Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
}
- }
- return true;
-}
-
-// Match for INSERTI64x4 INSERTF64x4 instructions (src0[0], src1[0]) or
-// (src1[0], src0[1]), manipulation with 256-bit sub-vectors
-static bool isINSERT64x4Mask(ArrayRef<int> Mask, MVT VT, unsigned int *Imm) {
- if (!VT.is512BitVector())
- return false;
+ } else if (VT.is128BitVector()) {
+ Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
+ } else
+ llvm_unreachable("Unexpected vector type");
- unsigned NumElts = VT.getVectorNumElements();
- unsigned HalfSize = NumElts/2;
- if (isSequentialOrUndefInRange(Mask, 0, HalfSize, 0)) {
- if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, NumElts)) {
- *Imm = 1;
- return true;
- }
- }
- if (isSequentialOrUndefInRange(Mask, 0, HalfSize, NumElts)) {
- if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, HalfSize)) {
- *Imm = 0;
- return true;
- }
- }
- return false;
+ return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
}
-/// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to MOVSS,
-/// MOVSD, and MOVD, i.e. setting the lowest element.
-static bool isMOVLMask(ArrayRef<int> Mask, EVT VT) {
- if (VT.getVectorElementType().getSizeInBits() < 32)
- return false;
- if (!VT.is128BitVector())
- return false;
-
- unsigned NumElts = VT.getVectorNumElements();
-
- if (!isUndefOrEqual(Mask[0], NumElts))
- return false;
-
- for (unsigned i = 1; i != NumElts; ++i)
- if (!isUndefOrEqual(Mask[i], i))
- return false;
-
- return true;
+/// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
+/// operation of specified width.
+static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
+ SDValue V2) {
+ unsigned NumElems = VT.getVectorNumElements();
+ SmallVector<int, 8> Mask;
+ Mask.push_back(NumElems);
+ for (unsigned i = 1; i != NumElems; ++i)
+ Mask.push_back(i);
+ return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
}
-/// isVPERM2X128Mask - Match 256-bit shuffles where the elements are considered
-/// as permutations between 128-bit chunks or halves. As an example: this
-/// shuffle bellow:
-/// vector_shuffle <4, 5, 6, 7, 12, 13, 14, 15>
-/// The first half comes from the second half of V1 and the second half from the
-/// the second half of V2.
-static bool isVPERM2X128Mask(ArrayRef<int> Mask, MVT VT, bool HasFp256) {
- if (!HasFp256 || !VT.is256BitVector())
- return false;
-
- // The shuffle result is divided into half A and half B. In total the two
- // sources have 4 halves, namely: C, D, E, F. The final values of A and
- // B must come from C, D, E or F.
- unsigned HalfSize = VT.getVectorNumElements()/2;
- bool MatchA = false, MatchB = false;
-
- // Check if A comes from one of C, D, E, F.
- for (unsigned Half = 0; Half != 4; ++Half) {
- if (isSequentialOrUndefInRange(Mask, 0, HalfSize, Half*HalfSize)) {
- MatchA = true;
- break;
- }
- }
-
- // Check if B comes from one of C, D, E, F.
- for (unsigned Half = 0; Half != 4; ++Half) {
- if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, Half*HalfSize)) {
- MatchB = true;
- break;
- }
+/// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
+static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
+ SDValue V2) {
+ unsigned NumElems = VT.getVectorNumElements();
+ SmallVector<int, 8> Mask;
+ for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
+ Mask.push_back(i);
+ Mask.push_back(i + NumElems);
}
-
- return MatchA && MatchB;
+ return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
}
-/// getShuffleVPERM2X128Immediate - Return the appropriate immediate to shuffle
-/// the specified VECTOR_MASK mask with VPERM2F128/VPERM2I128 instructions.
-static unsigned getShuffleVPERM2X128Immediate(ShuffleVectorSDNode *SVOp) {
- MVT VT = SVOp->getSimpleValueType(0);
-
- unsigned HalfSize = VT.getVectorNumElements()/2;
-
- unsigned FstHalf = 0, SndHalf = 0;
- for (unsigned i = 0; i < HalfSize; ++i) {
- if (SVOp->getMaskElt(i) > 0) {
- FstHalf = SVOp->getMaskElt(i)/HalfSize;
- break;
- }
- }
- for (unsigned i = HalfSize; i < HalfSize*2; ++i) {
- if (SVOp->getMaskElt(i) > 0) {
- SndHalf = SVOp->getMaskElt(i)/HalfSize;
- break;
- }
+/// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
+static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
+ SDValue V2) {
+ unsigned NumElems = VT.getVectorNumElements();
+ SmallVector<int, 8> Mask;
+ for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
+ Mask.push_back(i + Half);
+ Mask.push_back(i + NumElems + Half);
}
-
- return (FstHalf | (SndHalf << 4));
+ return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
}
-// Symetric in-lane mask. Each lane has 4 elements (for imm8)
-static bool isPermImmMask(ArrayRef<int> Mask, MVT VT, unsigned& Imm8) {
- unsigned EltSize = VT.getVectorElementType().getSizeInBits();
- if (EltSize < 32)
- return false;
+/// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
+/// vector of zero or undef vector. This produces a shuffle where the low
+/// element of V2 is swizzled into the zero/undef vector, landing at element
+/// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
+static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
+ bool IsZero,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ MVT VT = V2.getSimpleValueType();
+ SDValue V1 = IsZero
+ ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
+ unsigned NumElems = VT.getVectorNumElements();
+ SmallVector<int, 16> MaskVec;
+ for (unsigned i = 0; i != NumElems; ++i)
+ // If this is the insertion idx, put the low elt of V2 here.
+ MaskVec.push_back(i == Idx ? NumElems : i);
+ return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
+}
- unsigned NumElts = VT.getVectorNumElements();
- Imm8 = 0;
- if (VT.is128BitVector() || (VT.is256BitVector() && EltSize == 64)) {
- for (unsigned i = 0; i != NumElts; ++i) {
- if (Mask[i] < 0)
- continue;
- Imm8 |= Mask[i] << (i*2);
- }
- return true;
- }
+/// 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. 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;
- unsigned LaneSize = 4;
- SmallVector<int, 4> MaskVal(LaneSize, -1);
+ IsUnary = false;
+ bool IsFakeUnary = false;
+ switch(N->getOpcode()) {
+ case X86ISD::BLENDI:
+ ImmN = N->getOperand(N->getNumOperands()-1);
+ DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
+ break;
+ 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);
+ DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
+ break;
+ case X86ISD::PSHUFD:
+ case X86ISD::VPERMILPI:
+ ImmN = N->getOperand(N->getNumOperands()-1);
+ DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
+ IsUnary = true;
+ break;
+ case X86ISD::PSHUFHW:
+ ImmN = N->getOperand(N->getNumOperands()-1);
+ DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
+ IsUnary = true;
+ break;
+ case X86ISD::PSHUFLW:
+ 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);
- for (unsigned l = 0; l != NumElts; l += LaneSize) {
- for (unsigned i = 0; i != LaneSize; ++i) {
- if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize))
- return false;
- if (Mask[i+l] < 0)
- continue;
- if (MaskVal[i] < 0) {
- MaskVal[i] = Mask[i+l] - l;
- Imm8 |= MaskVal[i] << (i*2);
- continue;
- }
- if (Mask[i+l] != (signed)(MaskVal[i]+l))
+ 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;
- }
- }
- return true;
-}
-
-/// isVPERMILPMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to VPERMILPD*.
-/// Note that VPERMIL mask matching is different depending whether theunderlying
-/// type is 32 or 64. In the VPERMILPS the high half of the mask should point
-/// to the same elements of the low, but to the higher half of the source.
-/// In VPERMILPD the two lanes could be shuffled independently of each other
-/// with the same restriction that lanes can't be crossed. Also handles PSHUFDY.
-static bool isVPERMILPMask(ArrayRef<int> Mask, MVT VT) {
- unsigned EltSize = VT.getVectorElementType().getSizeInBits();
- if (VT.getSizeInBits() < 256 || EltSize < 32)
- return false;
- bool symetricMaskRequired = (EltSize == 32);
- unsigned NumElts = VT.getVectorNumElements();
- unsigned NumLanes = VT.getSizeInBits()/128;
- unsigned LaneSize = NumElts/NumLanes;
- // 2 or 4 elements in one lane
+ int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
- SmallVector<int, 4> ExpectedMaskVal(LaneSize, -1);
- for (unsigned l = 0; l != NumElts; l += LaneSize) {
- for (unsigned i = 0; i != LaneSize; ++i) {
- if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize))
- return false;
- if (symetricMaskRequired) {
- if (ExpectedMaskVal[i] < 0 && Mask[i+l] >= 0) {
- ExpectedMaskVal[i] = Mask[i+l] - l;
+ SmallVector<uint64_t, 32> RawMask;
+ for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
+ SDValue Op = MaskNode->getOperand(i);
+ if (Op->getOpcode() == ISD::UNDEF) {
+ RawMask.push_back((uint64_t)SM_SentinelUndef);
continue;
}
- if (!isUndefOrEqual(Mask[i+l], ExpectedMaskVal[i]+l))
+ auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
+ 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;
}
- }
- return true;
-}
-
-/// isCommutedMOVLMask - Returns true if the shuffle mask is except the reverse
-/// of what x86 movss want. X86 movs requires the lowest element to be lowest
-/// element of vector 2 and the other elements to come from vector 1 in order.
-static bool isCommutedMOVLMask(ArrayRef<int> Mask, MVT VT,
- bool V2IsSplat = false, bool V2IsUndef = false) {
- if (!VT.is128BitVector())
- return false;
- unsigned NumOps = VT.getVectorNumElements();
- if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
- return false;
+ auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
+ if (!MaskLoad)
+ return false;
- if (!isUndefOrEqual(Mask[0], 0))
- return false;
+ SDValue Ptr = MaskLoad->getBasePtr();
+ if (Ptr->getOpcode() == X86ISD::Wrapper)
+ Ptr = Ptr->getOperand(0);
- for (unsigned i = 1; i != NumOps; ++i)
- if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
- (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
- (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
+ auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
+ if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
return false;
- return true;
-}
+ if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
+ DecodePSHUFBMask(C, Mask);
+ if (Mask.empty())
+ return false;
+ break;
+ }
-/// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
-/// Masks to match: <1, 1, 3, 3> or <1, 1, 3, 3, 5, 5, 7, 7>
-static bool isMOVSHDUPMask(ArrayRef<int> Mask, MVT VT,
- const X86Subtarget *Subtarget) {
- if (!Subtarget->hasSSE3())
return false;
-
- unsigned NumElems = VT.getVectorNumElements();
-
- if ((VT.is128BitVector() && NumElems != 4) ||
- (VT.is256BitVector() && NumElems != 8) ||
- (VT.is512BitVector() && NumElems != 16))
+ }
+ case X86ISD::VPERMI:
+ ImmN = N->getOperand(N->getNumOperands()-1);
+ DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
+ IsUnary = true;
+ break;
+ case X86ISD::MOVSS:
+ case X86ISD::MOVSD:
+ DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
+ break;
+ case X86ISD::VPERM2X128:
+ ImmN = N->getOperand(N->getNumOperands()-1);
+ DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
+ if (Mask.empty()) return false;
+ break;
+ case X86ISD::MOVSLDUP:
+ DecodeMOVSLDUPMask(VT, Mask);
+ IsUnary = true;
+ break;
+ case X86ISD::MOVSHDUP:
+ DecodeMOVSHDUPMask(VT, Mask);
+ IsUnary = true;
+ break;
+ case X86ISD::MOVDDUP:
+ DecodeMOVDDUPMask(VT, Mask);
+ IsUnary = true;
+ break;
+ case X86ISD::MOVLHPD:
+ case X86ISD::MOVLPD:
+ case X86ISD::MOVLPS:
+ // Not yet implemented
return false;
+ default: llvm_unreachable("unknown target shuffle node");
+ }
- // "i+1" is the value the indexed mask element must have
- for (unsigned i = 0; i != NumElems; i += 2)
- if (!isUndefOrEqual(Mask[i], i+1) ||
- !isUndefOrEqual(Mask[i+1], i+1))
- return false;
+ // 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;
}
-/// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
-/// Masks to match: <0, 0, 2, 2> or <0, 0, 2, 2, 4, 4, 6, 6>
-static bool isMOVSLDUPMask(ArrayRef<int> Mask, MVT VT,
- const X86Subtarget *Subtarget) {
- if (!Subtarget->hasSSE3())
- return false;
+/// getShuffleScalarElt - Returns the scalar element that will make up the ith
+/// element of the result of the vector shuffle.
+static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
+ unsigned Depth) {
+ if (Depth == 6)
+ return SDValue(); // Limit search depth.
- unsigned NumElems = VT.getVectorNumElements();
+ SDValue V = SDValue(N, 0);
+ EVT VT = V.getValueType();
+ unsigned Opcode = V.getOpcode();
- if ((VT.is128BitVector() && NumElems != 4) ||
- (VT.is256BitVector() && NumElems != 8) ||
- (VT.is512BitVector() && NumElems != 16))
- return false;
+ // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
+ if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
+ int Elt = SV->getMaskElt(Index);
- // "i" is the value the indexed mask element must have
- for (unsigned i = 0; i != NumElems; i += 2)
- if (!isUndefOrEqual(Mask[i], i) ||
- !isUndefOrEqual(Mask[i+1], i))
- return false;
+ if (Elt < 0)
+ return DAG.getUNDEF(VT.getVectorElementType());
- return true;
-}
+ unsigned NumElems = VT.getVectorNumElements();
+ SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
+ : SV->getOperand(1);
+ return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
+ }
-/// isMOVDDUPYMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to 256-bit
-/// version of MOVDDUP.
-static bool isMOVDDUPYMask(ArrayRef<int> Mask, MVT VT, bool HasFp256) {
- if (!HasFp256 || !VT.is256BitVector())
- return false;
+ // Recurse into target specific vector shuffles to find scalars.
+ if (isTargetShuffle(Opcode)) {
+ MVT ShufVT = V.getSimpleValueType();
+ unsigned NumElems = ShufVT.getVectorNumElements();
+ SmallVector<int, 16> ShuffleMask;
+ bool IsUnary;
- unsigned NumElts = VT.getVectorNumElements();
- if (NumElts != 4)
- return false;
+ if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
+ return SDValue();
- for (unsigned i = 0; i != NumElts/2; ++i)
- if (!isUndefOrEqual(Mask[i], 0))
- return false;
- for (unsigned i = NumElts/2; i != NumElts; ++i)
- if (!isUndefOrEqual(Mask[i], NumElts/2))
- return false;
- return true;
-}
+ int Elt = ShuffleMask[Index];
+ if (Elt < 0)
+ return DAG.getUNDEF(ShufVT.getVectorElementType());
-/// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to 128-bit
-/// version of MOVDDUP.
-static bool isMOVDDUPMask(ArrayRef<int> Mask, MVT VT) {
- if (!VT.is128BitVector())
- return false;
+ SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
+ : N->getOperand(1);
+ return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
+ Depth+1);
+ }
- unsigned e = VT.getVectorNumElements() / 2;
- for (unsigned i = 0; i != e; ++i)
- if (!isUndefOrEqual(Mask[i], i))
- return false;
- for (unsigned i = 0; i != e; ++i)
- if (!isUndefOrEqual(Mask[e+i], i))
- return false;
- return true;
-}
+ // Actual nodes that may contain scalar elements
+ if (Opcode == ISD::BITCAST) {
+ V = V.getOperand(0);
+ EVT SrcVT = V.getValueType();
+ unsigned NumElems = VT.getVectorNumElements();
-/// isVEXTRACTIndex - Return true if the specified
-/// EXTRACT_SUBVECTOR operand specifies a vector extract that is
-/// suitable for instruction that extract 128 or 256 bit vectors
-static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
- assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
- if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
- return false;
+ if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
+ return SDValue();
+ }
- // The index should be aligned on a vecWidth-bit boundary.
- uint64_t Index =
- cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
+ if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
+ return (Index == 0) ? V.getOperand(0)
+ : DAG.getUNDEF(VT.getVectorElementType());
- MVT VT = N->getSimpleValueType(0);
- unsigned ElSize = VT.getVectorElementType().getSizeInBits();
- bool Result = (Index * ElSize) % vecWidth == 0;
+ if (V.getOpcode() == ISD::BUILD_VECTOR)
+ return V.getOperand(Index);
- return Result;
+ return SDValue();
}
-/// isVINSERTIndex - Return true if the specified INSERT_SUBVECTOR
-/// operand specifies a subvector insert that is suitable for input to
-/// insertion of 128 or 256-bit subvectors
-static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
- assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
- if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
- return false;
- // The index should be aligned on a vecWidth-bit boundary.
- uint64_t Index =
- cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
-
- MVT VT = N->getSimpleValueType(0);
- unsigned ElSize = VT.getVectorElementType().getSizeInBits();
- bool Result = (Index * ElSize) % vecWidth == 0;
-
- return Result;
-}
-
-bool X86::isVINSERT128Index(SDNode *N) {
- return isVINSERTIndex(N, 128);
-}
-
-bool X86::isVINSERT256Index(SDNode *N) {
- return isVINSERTIndex(N, 256);
-}
-
-bool X86::isVEXTRACT128Index(SDNode *N) {
- return isVEXTRACTIndex(N, 128);
-}
-
-bool X86::isVEXTRACT256Index(SDNode *N) {
- return isVEXTRACTIndex(N, 256);
-}
-
-/// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
-/// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
-/// Handles 128-bit and 256-bit.
-static unsigned getShuffleSHUFImmediate(ShuffleVectorSDNode *N) {
- MVT VT = N->getSimpleValueType(0);
-
- assert((VT.getSizeInBits() >= 128) &&
- "Unsupported vector type for PSHUF/SHUFP");
+/// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
+///
+static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
+ unsigned NumNonZero, unsigned NumZero,
+ SelectionDAG &DAG,
+ const X86Subtarget* Subtarget,
+ const TargetLowering &TLI) {
+ if (NumNonZero > 8)
+ return SDValue();
- // Handle 128 and 256-bit vector lengths. AVX defines PSHUF/SHUFP to operate
- // independently on 128-bit lanes.
- unsigned NumElts = VT.getVectorNumElements();
- unsigned NumLanes = VT.getSizeInBits()/128;
- unsigned NumLaneElts = NumElts/NumLanes;
+ SDLoc dl(Op);
+ SDValue V;
+ bool First = true;
+ for (unsigned i = 0; i < 16; ++i) {
+ bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
+ if (ThisIsNonZero && First) {
+ if (NumZero)
+ V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
+ else
+ V = DAG.getUNDEF(MVT::v8i16);
+ First = false;
+ }
- assert((NumLaneElts == 2 || NumLaneElts == 4 || NumLaneElts == 8) &&
- "Only supports 2, 4 or 8 elements per lane");
+ if ((i & 1) != 0) {
+ SDValue ThisElt, LastElt;
+ bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
+ if (LastIsNonZero) {
+ LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
+ MVT::i16, Op.getOperand(i-1));
+ }
+ if (ThisIsNonZero) {
+ ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
+ ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
+ ThisElt, DAG.getConstant(8, MVT::i8));
+ if (LastIsNonZero)
+ ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
+ } else
+ ThisElt = LastElt;
- unsigned Shift = (NumLaneElts >= 4) ? 1 : 0;
- unsigned Mask = 0;
- for (unsigned i = 0; i != NumElts; ++i) {
- int Elt = N->getMaskElt(i);
- if (Elt < 0) continue;
- Elt &= NumLaneElts - 1;
- unsigned ShAmt = (i << Shift) % 8;
- Mask |= Elt << ShAmt;
+ if (ThisElt.getNode())
+ V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
+ DAG.getIntPtrConstant(i/2));
+ }
}
- return Mask;
+ return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
}
-/// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
-/// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
-static unsigned getShufflePSHUFHWImmediate(ShuffleVectorSDNode *N) {
- MVT VT = N->getSimpleValueType(0);
-
- assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
- "Unsupported vector type for PSHUFHW");
-
- unsigned NumElts = VT.getVectorNumElements();
+/// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
+///
+static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
+ unsigned NumNonZero, unsigned NumZero,
+ SelectionDAG &DAG,
+ const X86Subtarget* Subtarget,
+ const TargetLowering &TLI) {
+ if (NumNonZero > 4)
+ return SDValue();
- unsigned Mask = 0;
- for (unsigned l = 0; l != NumElts; l += 8) {
- // 8 nodes per lane, but we only care about the last 4.
- for (unsigned i = 0; i < 4; ++i) {
- int Elt = N->getMaskElt(l+i+4);
- if (Elt < 0) continue;
- Elt &= 0x3; // only 2-bits.
- Mask |= Elt << (i * 2);
+ SDLoc dl(Op);
+ SDValue V;
+ bool First = true;
+ for (unsigned i = 0; i < 8; ++i) {
+ bool isNonZero = (NonZeros & (1 << i)) != 0;
+ if (isNonZero) {
+ if (First) {
+ if (NumZero)
+ V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
+ else
+ V = DAG.getUNDEF(MVT::v8i16);
+ First = false;
+ }
+ V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
+ MVT::v8i16, V, Op.getOperand(i),
+ DAG.getIntPtrConstant(i));
}
}
- return Mask;
+ return V;
}
-/// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
-/// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
-static unsigned getShufflePSHUFLWImmediate(ShuffleVectorSDNode *N) {
- MVT VT = N->getSimpleValueType(0);
-
- assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
- "Unsupported vector type for PSHUFHW");
-
- unsigned NumElts = VT.getVectorNumElements();
+/// LowerBuildVectorv4x32 - Custom lower build_vector of v4i32 or v4f32.
+static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
+ const X86Subtarget *Subtarget,
+ const TargetLowering &TLI) {
+ // Find all zeroable elements.
+ std::bitset<4> Zeroable;
+ for (int i=0; i < 4; ++i) {
+ SDValue Elt = Op->getOperand(i);
+ Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
+ }
+ assert(Zeroable.size() - Zeroable.count() > 1 &&
+ "We expect at least two non-zero elements!");
- unsigned Mask = 0;
- for (unsigned l = 0; l != NumElts; l += 8) {
- // 8 nodes per lane, but we only care about the first 4.
- for (unsigned i = 0; i < 4; ++i) {
- int Elt = N->getMaskElt(l+i);
- if (Elt < 0) continue;
- Elt &= 0x3; // only 2-bits
- Mask |= Elt << (i * 2);
+ // We only know how to deal with build_vector nodes where elements are either
+ // zeroable or extract_vector_elt with constant index.
+ SDValue FirstNonZero;
+ unsigned FirstNonZeroIdx;
+ for (unsigned i=0; i < 4; ++i) {
+ if (Zeroable[i])
+ continue;
+ SDValue Elt = Op->getOperand(i);
+ if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
+ !isa<ConstantSDNode>(Elt.getOperand(1)))
+ return SDValue();
+ // Make sure that this node is extracting from a 128-bit vector.
+ MVT VT = Elt.getOperand(0).getSimpleValueType();
+ if (!VT.is128BitVector())
+ return SDValue();
+ if (!FirstNonZero.getNode()) {
+ FirstNonZero = Elt;
+ FirstNonZeroIdx = i;
}
}
- return Mask;
-}
+ assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
+ SDValue V1 = FirstNonZero.getOperand(0);
+ MVT VT = V1.getSimpleValueType();
-/// \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 = InterLane ? 1 :
- VT.getVectorElementType().getSizeInBits() >> 3;
+ // See if this build_vector can be lowered as a blend with zero.
+ SDValue Elt;
+ unsigned EltMaskIdx, EltIdx;
+ int Mask[4];
+ for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
+ if (Zeroable[EltIdx]) {
+ // The zero vector will be on the right hand side.
+ Mask[EltIdx] = EltIdx+4;
+ continue;
+ }
- unsigned NumElts = VT.getVectorNumElements();
- unsigned NumLanes = VT.is512BitVector() ? 1 : VT.getSizeInBits()/128;
- unsigned NumLaneElts = NumElts/NumLanes;
-
- int Val = 0;
- unsigned i;
- for (i = 0; i != NumElts; ++i) {
- Val = SVOp->getMaskElt(i);
- if (Val >= 0)
+ Elt = Op->getOperand(EltIdx);
+ // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
+ EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
+ if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
break;
+ Mask[EltIdx] = EltIdx;
}
- if (Val >= (int)NumElts)
- Val -= NumElts - NumLaneElts;
-
- assert(Val - i > 0 && "PALIGNR imm should be positive");
- 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);
-}
+ if (EltIdx == 4) {
+ // Let the shuffle legalizer deal with blend operations.
+ SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
+ if (V1.getSimpleValueType() != VT)
+ V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
+ return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
+ }
-static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
- assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
- if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
- llvm_unreachable("Illegal extract subvector for VEXTRACT");
-
- uint64_t Index =
- cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
+ // See if we can lower this build_vector to a INSERTPS.
+ if (!Subtarget->hasSSE41())
+ return SDValue();
- MVT VecVT = N->getOperand(0).getSimpleValueType();
- MVT ElVT = VecVT.getVectorElementType();
+ SDValue V2 = Elt.getOperand(0);
+ if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
+ V1 = SDValue();
- unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
- return Index / NumElemsPerChunk;
-}
+ bool CanFold = true;
+ for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
+ if (Zeroable[i])
+ continue;
-static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
- assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
- if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
- llvm_unreachable("Illegal insert subvector for VINSERT");
+ SDValue Current = Op->getOperand(i);
+ SDValue SrcVector = Current->getOperand(0);
+ if (!V1.getNode())
+ V1 = SrcVector;
+ CanFold = SrcVector == V1 &&
+ cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
+ }
- uint64_t Index =
- cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
+ if (!CanFold)
+ return SDValue();
- MVT VecVT = N->getSimpleValueType(0);
- MVT ElVT = VecVT.getVectorElementType();
+ assert(V1.getNode() && "Expected at least two non-zero elements!");
+ if (V1.getSimpleValueType() != MVT::v4f32)
+ V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
+ if (V2.getSimpleValueType() != MVT::v4f32)
+ V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
- unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
- return Index / NumElemsPerChunk;
-}
+ // Ok, we can emit an INSERTPS instruction.
+ unsigned ZMask = Zeroable.to_ulong();
-/// getExtractVEXTRACT128Immediate - Return the appropriate immediate
-/// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
-/// and VINSERTI128 instructions.
-unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
- return getExtractVEXTRACTImmediate(N, 128);
+ unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
+ assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
+ SDValue Result = DAG.getNode(X86ISD::INSERTPS, SDLoc(Op), MVT::v4f32, V1, V2,
+ DAG.getIntPtrConstant(InsertPSMask));
+ return DAG.getNode(ISD::BITCAST, SDLoc(Op), VT, Result);
}
-/// getExtractVEXTRACT256Immediate - Return the appropriate immediate
-/// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF64x4
-/// and VINSERTI64x4 instructions.
-unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
- return getExtractVEXTRACTImmediate(N, 256);
+/// Return a vector logical shift node.
+static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
+ unsigned NumBits, SelectionDAG &DAG,
+ const TargetLowering &TLI, SDLoc dl) {
+ assert(VT.is128BitVector() && "Unknown type for VShift");
+ MVT ShVT = MVT::v2i64;
+ unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
+ SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
+ MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(SrcOp.getValueType());
+ assert(NumBits % 8 == 0 && "Only support byte sized shifts");
+ SDValue ShiftVal = DAG.getConstant(NumBits/8, ScalarShiftTy);
+ return DAG.getNode(ISD::BITCAST, dl, VT,
+ DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
}
-/// getInsertVINSERT128Immediate - Return the appropriate immediate
-/// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
-/// and VINSERTI128 instructions.
-unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
- return getInsertVINSERTImmediate(N, 128);
-}
+static SDValue
+LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
-/// getInsertVINSERT256Immediate - Return the appropriate immediate
-/// to insert at the specified INSERT_SUBVECTOR index with VINSERTF46x4
-/// and VINSERTI64x4 instructions.
-unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
- return getInsertVINSERTImmediate(N, 256);
-}
+ // Check if the scalar load can be widened into a vector load. And if
+ // the address is "base + cst" see if the cst can be "absorbed" into
+ // the shuffle mask.
+ if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
+ SDValue Ptr = LD->getBasePtr();
+ if (!ISD::isNormalLoad(LD) || LD->isVolatile())
+ return SDValue();
+ EVT PVT = LD->getValueType(0);
+ if (PVT != MVT::i32 && PVT != MVT::f32)
+ return SDValue();
-/// isZero - Returns true if Elt is a constant integer zero
-static bool isZero(SDValue V) {
- ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
- return C && C->isNullValue();
-}
-
-/// isZeroNode - Returns true if Elt is a constant zero or a floating point
-/// constant +0.0.
-bool X86::isZeroNode(SDValue Elt) {
- if (isZero(Elt))
- return true;
- if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
- return CFP->getValueAPF().isPosZero();
- return false;
-}
-
-/// ShouldXformToMOVHLPS - Return true if the node should be transformed to
-/// match movhlps. The lower half elements should come from upper half of
-/// V1 (and in order), and the upper half elements should come from the upper
-/// half of V2 (and in order).
-static bool ShouldXformToMOVHLPS(ArrayRef<int> Mask, MVT VT) {
- if (!VT.is128BitVector())
- return false;
- if (VT.getVectorNumElements() != 4)
- return false;
- for (unsigned i = 0, e = 2; i != e; ++i)
- if (!isUndefOrEqual(Mask[i], i+2))
- return false;
- for (unsigned i = 2; i != 4; ++i)
- if (!isUndefOrEqual(Mask[i], i+4))
- return false;
- return true;
-}
-
-/// isScalarLoadToVector - Returns true if the node is a scalar load that
-/// is promoted to a vector. It also returns the LoadSDNode by reference if
-/// required.
-static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = nullptr) {
- if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
- return false;
- N = N->getOperand(0).getNode();
- if (!ISD::isNON_EXTLoad(N))
- return false;
- if (LD)
- *LD = cast<LoadSDNode>(N);
- return true;
-}
-
-// Test whether the given value is a vector value which will be legalized
-// into a load.
-static bool WillBeConstantPoolLoad(SDNode *N) {
- if (N->getOpcode() != ISD::BUILD_VECTOR)
- return false;
-
- // Check for any non-constant elements.
- for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
- switch (N->getOperand(i).getNode()->getOpcode()) {
- case ISD::UNDEF:
- case ISD::ConstantFP:
- case ISD::Constant:
- break;
- default:
- return false;
+ int FI = -1;
+ int64_t Offset = 0;
+ if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
+ FI = FINode->getIndex();
+ Offset = 0;
+ } else if (DAG.isBaseWithConstantOffset(Ptr) &&
+ isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
+ FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
+ Offset = Ptr.getConstantOperandVal(1);
+ Ptr = Ptr.getOperand(0);
+ } else {
+ return SDValue();
}
- // Vectors of all-zeros and all-ones are materialized with special
- // instructions rather than being loaded.
- return !ISD::isBuildVectorAllZeros(N) &&
- !ISD::isBuildVectorAllOnes(N);
-}
+ // FIXME: 256-bit vector instructions don't require a strict alignment,
+ // improve this code to support it better.
+ unsigned RequiredAlign = VT.getSizeInBits()/8;
+ SDValue Chain = LD->getChain();
+ // Make sure the stack object alignment is at least 16 or 32.
+ MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
+ if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
+ if (MFI->isFixedObjectIndex(FI)) {
+ // Can't change the alignment. FIXME: It's possible to compute
+ // the exact stack offset and reference FI + adjust offset instead.
+ // If someone *really* cares about this. That's the way to implement it.
+ return SDValue();
+ } else {
+ MFI->setObjectAlignment(FI, RequiredAlign);
+ }
+ }
-/// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
-/// match movlp{s|d}. The lower half elements should come from lower half of
-/// V1 (and in order), and the upper half elements should come from the upper
-/// half of V2 (and in order). And since V1 will become the source of the
-/// MOVLP, it must be either a vector load or a scalar load to vector.
-static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
- ArrayRef<int> Mask, MVT VT) {
- if (!VT.is128BitVector())
- return false;
+ // (Offset % 16 or 32) must be multiple of 4. Then address is then
+ // Ptr + (Offset & ~15).
+ if (Offset < 0)
+ return SDValue();
+ if ((Offset % RequiredAlign) & 3)
+ return SDValue();
+ int64_t StartOffset = Offset & ~(RequiredAlign-1);
+ if (StartOffset)
+ Ptr = DAG.getNode(ISD::ADD, SDLoc(Ptr), Ptr.getValueType(),
+ Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
- if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
- return false;
- // Is V2 is a vector load, don't do this transformation. We will try to use
- // load folding shufps op.
- if (ISD::isNON_EXTLoad(V2) || WillBeConstantPoolLoad(V2))
- return false;
+ int EltNo = (Offset - StartOffset) >> 2;
+ unsigned NumElems = VT.getVectorNumElements();
- unsigned NumElems = VT.getVectorNumElements();
+ EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
+ SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
+ LD->getPointerInfo().getWithOffset(StartOffset),
+ false, false, false, 0);
- if (NumElems != 2 && NumElems != 4)
- return false;
- for (unsigned i = 0, e = NumElems/2; i != e; ++i)
- if (!isUndefOrEqual(Mask[i], i))
- return false;
- for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
- if (!isUndefOrEqual(Mask[i], i+NumElems))
- return false;
- return true;
-}
+ SmallVector<int, 8> Mask(NumElems, EltNo);
-/// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
-/// to an zero vector.
-/// FIXME: move to dag combiner / method on ShuffleVectorSDNode
-static bool isZeroShuffle(ShuffleVectorSDNode *N) {
- SDValue V1 = N->getOperand(0);
- SDValue V2 = N->getOperand(1);
- unsigned NumElems = N->getValueType(0).getVectorNumElements();
- for (unsigned i = 0; i != NumElems; ++i) {
- int Idx = N->getMaskElt(i);
- if (Idx >= (int)NumElems) {
- unsigned Opc = V2.getOpcode();
- if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
- continue;
- if (Opc != ISD::BUILD_VECTOR ||
- !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
- return false;
- } else if (Idx >= 0) {
- unsigned Opc = V1.getOpcode();
- if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
- continue;
- if (Opc != ISD::BUILD_VECTOR ||
- !X86::isZeroNode(V1.getOperand(Idx)))
- return false;
- }
+ return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
}
- return true;
+
+ return SDValue();
}
-/// getZeroVector - Returns a vector of specified type with all zero elements.
+/// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
+/// elements can be replaced by a single large load which has the same value as
+/// a build_vector or insert_subvector whose loaded operands are 'Elts'.
///
-static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
- SelectionDAG &DAG, SDLoc dl) {
- assert(VT.isVector() && "Expected a vector type");
-
- // Always build SSE zero vectors as <4 x i32> bitcasted
- // to their dest type. This ensures they get CSE'd.
- SDValue Vec;
- if (VT.is128BitVector()) { // SSE
- if (Subtarget->hasSSE2()) { // SSE2
- SDValue Cst = DAG.getConstant(0, MVT::i32);
- Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
- } else { // SSE1
- SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32);
- Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
- }
- } else if (VT.is256BitVector()) { // AVX
- if (Subtarget->hasInt256()) { // AVX2
- SDValue Cst = DAG.getConstant(0, MVT::i32);
- SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
- Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
- } else {
- // 256-bit logic and arithmetic instructions in AVX are all
- // floating-point, no support for integer ops. Emit fp zeroed vectors.
- SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32);
- SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
- Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
- }
- } else if (VT.is512BitVector()) { // AVX-512
- SDValue Cst = DAG.getConstant(0, MVT::i32);
- SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
- Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
- Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
- } else if (VT.getScalarType() == MVT::i1) {
- assert(VT.getVectorNumElements() <= 16 && "Unexpected vector type");
- SDValue Cst = DAG.getConstant(0, MVT::i1);
- SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
- return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
- } else
- llvm_unreachable("Unexpected vector type");
-
- return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
-}
+/// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
+///
+/// FIXME: we'd also like to handle the case where the last elements are zero
+/// rather than undef via VZEXT_LOAD, but we do not detect that case today.
+/// There's even a handy isZeroNode for that purpose.
+static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
+ SDLoc &DL, SelectionDAG &DAG,
+ bool isAfterLegalize) {
+ unsigned NumElems = Elts.size();
-/// getOnesVector - Returns a vector of specified type with all bits set.
-/// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
-/// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
-/// Then bitcast to their original type, ensuring they get CSE'd.
-static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
- SDLoc dl) {
- assert(VT.isVector() && "Expected a vector type");
+ LoadSDNode *LDBase = nullptr;
+ unsigned LastLoadedElt = -1U;
- SDValue Cst = DAG.getConstant(~0U, MVT::i32);
- SDValue Vec;
- if (VT.is256BitVector()) {
- if (HasInt256) { // AVX2
- SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
- Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
- } else { // AVX
- Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
- Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
+ // For each element in the initializer, see if we've found a load or an undef.
+ // If we don't find an initial load element, or later load elements are
+ // non-consecutive, bail out.
+ for (unsigned i = 0; i < NumElems; ++i) {
+ SDValue Elt = Elts[i];
+ // Look through a bitcast.
+ if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
+ Elt = Elt.getOperand(0);
+ if (!Elt.getNode() ||
+ (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
+ return SDValue();
+ if (!LDBase) {
+ if (Elt.getNode()->getOpcode() == ISD::UNDEF)
+ return SDValue();
+ LDBase = cast<LoadSDNode>(Elt.getNode());
+ LastLoadedElt = i;
+ continue;
}
- } else if (VT.is128BitVector()) {
- Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
- } else
- llvm_unreachable("Unexpected vector type");
-
- return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
-}
+ if (Elt.getOpcode() == ISD::UNDEF)
+ continue;
-/// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
-/// that point to V2 points to its first element.
-static void NormalizeMask(SmallVectorImpl<int> &Mask, unsigned NumElems) {
- for (unsigned i = 0; i != NumElems; ++i) {
- if (Mask[i] > (int)NumElems) {
- Mask[i] = NumElems;
- }
+ LoadSDNode *LD = cast<LoadSDNode>(Elt);
+ EVT LdVT = Elt.getValueType();
+ // Each loaded element must be the correct fractional portion of the
+ // requested vector load.
+ if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
+ return SDValue();
+ if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
+ return SDValue();
+ LastLoadedElt = i;
}
-}
-/// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
-/// operation of specified width.
-static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
- SDValue V2) {
- unsigned NumElems = VT.getVectorNumElements();
- SmallVector<int, 8> Mask;
- Mask.push_back(NumElems);
- for (unsigned i = 1; i != NumElems; ++i)
- Mask.push_back(i);
- return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
-}
+ // If we have found an entire vector of loads and undefs, then return a large
+ // load of the entire vector width starting at the base pointer. If we found
+ // consecutive loads for the low half, generate a vzext_load node.
+ if (LastLoadedElt == NumElems - 1) {
+ assert(LDBase && "Did not find base load for merging consecutive loads");
+ EVT EltVT = LDBase->getValueType(0);
+ // Ensure that the input vector size for the merged loads matches the
+ // cumulative size of the input elements.
+ if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
+ return SDValue();
-/// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
-static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
- SDValue V2) {
- unsigned NumElems = VT.getVectorNumElements();
- SmallVector<int, 8> Mask;
- for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
- Mask.push_back(i);
- Mask.push_back(i + NumElems);
- }
- return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
-}
+ if (isAfterLegalize &&
+ !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
+ return SDValue();
-/// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
-static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
- SDValue V2) {
- unsigned NumElems = VT.getVectorNumElements();
- SmallVector<int, 8> Mask;
- for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
- Mask.push_back(i + Half);
- Mask.push_back(i + NumElems + Half);
- }
- return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
-}
+ SDValue NewLd = SDValue();
-// PromoteSplati8i16 - All i16 and i8 vector types can't be used directly by
-// a generic shuffle instruction because the target has no such instructions.
-// Generate shuffles which repeat i16 and i8 several times until they can be
-// represented by v4f32 and then be manipulated by target suported shuffles.
-static SDValue PromoteSplati8i16(SDValue V, SelectionDAG &DAG, int &EltNo) {
- MVT VT = V.getSimpleValueType();
- int NumElems = VT.getVectorNumElements();
- SDLoc dl(V);
+ NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
+ LDBase->getPointerInfo(), LDBase->isVolatile(),
+ LDBase->isNonTemporal(), LDBase->isInvariant(),
+ LDBase->getAlignment());
- while (NumElems > 4) {
- if (EltNo < NumElems/2) {
- V = getUnpackl(DAG, dl, VT, V, V);
- } else {
- V = getUnpackh(DAG, dl, VT, V, V);
- EltNo -= NumElems/2;
+ if (LDBase->hasAnyUseOfValue(1)) {
+ SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
+ SDValue(LDBase, 1),
+ SDValue(NewLd.getNode(), 1));
+ DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
+ DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
+ SDValue(NewLd.getNode(), 1));
}
- NumElems >>= 1;
+
+ return NewLd;
}
- return V;
-}
-/// getLegalSplat - Generate a legal splat with supported x86 shuffles
-static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) {
- MVT VT = V.getSimpleValueType();
- SDLoc dl(V);
-
- if (VT.is128BitVector()) {
- V = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V);
- int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
- V = DAG.getVectorShuffle(MVT::v4f32, dl, V, DAG.getUNDEF(MVT::v4f32),
- &SplatMask[0]);
- } else if (VT.is256BitVector()) {
- // To use VPERMILPS to splat scalars, the second half of indicies must
- // refer to the higher part, which is a duplication of the lower one,
- // because VPERMILPS can only handle in-lane permutations.
- int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo,
- EltNo+4, EltNo+4, EltNo+4, EltNo+4 };
-
- V = DAG.getNode(ISD::BITCAST, dl, MVT::v8f32, V);
- V = DAG.getVectorShuffle(MVT::v8f32, dl, V, DAG.getUNDEF(MVT::v8f32),
- &SplatMask[0]);
- } else
- llvm_unreachable("Vector size not supported");
+ //TODO: The code below fires only for for loading the low v2i32 / v2f32
+ //of a v4i32 / v4f32. It's probably worth generalizing.
+ EVT EltVT = VT.getVectorElementType();
+ if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
+ DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
+ SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
+ SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
+ SDValue ResNode =
+ DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
+ LDBase->getPointerInfo(),
+ LDBase->getAlignment(),
+ false/*isVolatile*/, true/*ReadMem*/,
+ false/*WriteMem*/);
+
+ // Make sure the newly-created LOAD is in the same position as LDBase in
+ // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
+ // update uses of LDBase's output chain to use the TokenFactor.
+ if (LDBase->hasAnyUseOfValue(1)) {
+ SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
+ SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
+ DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
+ DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
+ SDValue(ResNode.getNode(), 1));
+ }
- return DAG.getNode(ISD::BITCAST, dl, VT, V);
+ return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
+ }
+ return SDValue();
}
-/// PromoteSplat - Splat is promoted to target supported vector shuffles.
-static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
- MVT SrcVT = SV->getSimpleValueType(0);
- SDValue V1 = SV->getOperand(0);
- SDLoc dl(SV);
+/// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
+/// to generate a splat value for the following cases:
+/// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
+/// 2. A splat shuffle which uses a scalar_to_vector node which comes from
+/// a scalar load, or a constant.
+/// The VBROADCAST node is returned when a pattern is found,
+/// or SDValue() otherwise.
+static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
+ SelectionDAG &DAG) {
+ // VBROADCAST requires AVX.
+ // TODO: Splats could be generated for non-AVX CPUs using SSE
+ // instructions, but there's less potential gain for only 128-bit vectors.
+ if (!Subtarget->hasAVX())
+ return SDValue();
- int EltNo = SV->getSplatIndex();
- int NumElems = SrcVT.getVectorNumElements();
- bool Is256BitVec = SrcVT.is256BitVector();
+ MVT VT = Op.getSimpleValueType();
+ SDLoc dl(Op);
- assert(((SrcVT.is128BitVector() && NumElems > 4) || Is256BitVec) &&
- "Unknown how to promote splat for type");
+ assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
+ "Unsupported vector type for broadcast.");
- // Extract the 128-bit part containing the splat element and update
- // the splat element index when it refers to the higher register.
- if (Is256BitVec) {
- V1 = Extract128BitVector(V1, EltNo, DAG, dl);
- if (EltNo >= NumElems/2)
- EltNo -= NumElems/2;
- }
+ SDValue Ld;
+ bool ConstSplatVal;
- // All i16 and i8 vector types can't be used directly by a generic shuffle
- // instruction because the target has no such instruction. Generate shuffles
- // which repeat i16 and i8 several times until they fit in i32, and then can
- // be manipulated by target suported shuffles.
- MVT EltVT = SrcVT.getVectorElementType();
- if (EltVT == MVT::i8 || EltVT == MVT::i16)
- V1 = PromoteSplati8i16(V1, DAG, EltNo);
+ switch (Op.getOpcode()) {
+ default:
+ // Unknown pattern found.
+ return SDValue();
- // Recreate the 256-bit vector and place the same 128-bit vector
- // into the low and high part. This is necessary because we want
- // to use VPERM* to shuffle the vectors
- if (Is256BitVec) {
- V1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, SrcVT, V1, V1);
- }
+ case ISD::BUILD_VECTOR: {
+ auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
+ BitVector UndefElements;
+ SDValue Splat = BVOp->getSplatValue(&UndefElements);
- return getLegalSplat(DAG, V1, EltNo);
-}
+ // We need a splat of a single value to use broadcast, and it doesn't
+ // make any sense if the value is only in one element of the vector.
+ if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
+ return SDValue();
-/// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
-/// vector of zero or undef vector. This produces a shuffle where the low
-/// element of V2 is swizzled into the zero/undef vector, landing at element
-/// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
-static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
- bool IsZero,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- MVT VT = V2.getSimpleValueType();
- SDValue V1 = IsZero
- ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
- unsigned NumElems = VT.getVectorNumElements();
- SmallVector<int, 16> MaskVec;
- for (unsigned i = 0; i != NumElems; ++i)
- // If this is the insertion idx, put the low elt of V2 here.
- MaskVec.push_back(i == Idx ? NumElems : i);
- return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
-}
+ Ld = Splat;
+ ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
+ Ld.getOpcode() == ISD::ConstantFP);
-/// 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. 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;
+ // Make sure that all of the users of a non-constant load are from the
+ // BUILD_VECTOR node.
+ if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
+ return SDValue();
+ break;
+ }
- IsUnary = false;
- bool IsFakeUnary = false;
- switch(N->getOpcode()) {
- case X86ISD::BLENDI:
- ImmN = N->getOperand(N->getNumOperands()-1);
- DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
- break;
- 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);
- DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
- break;
- case X86ISD::PSHUFD:
- case X86ISD::VPERMILPI:
- ImmN = N->getOperand(N->getNumOperands()-1);
- DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
- IsUnary = true;
- break;
- case X86ISD::PSHUFHW:
- ImmN = N->getOperand(N->getNumOperands()-1);
- DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
- IsUnary = true;
- break;
- case X86ISD::PSHUFLW:
- 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);
+ case ISD::VECTOR_SHUFFLE: {
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- 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;
+ // Shuffles must have a splat mask where the first element is
+ // broadcasted.
+ if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
+ return SDValue();
- int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
+ SDValue Sc = Op.getOperand(0);
+ if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
+ Sc.getOpcode() != ISD::BUILD_VECTOR) {
- SmallVector<uint64_t, 32> RawMask;
- for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
- SDValue Op = MaskNode->getOperand(i);
- if (Op->getOpcode() == ISD::UNDEF) {
- RawMask.push_back((uint64_t)SM_SentinelUndef);
- continue;
- }
- auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
- if (!CN)
- return false;
- APInt MaskElement = CN->getAPIntValue();
+ if (!Subtarget->hasInt256())
+ return SDValue();
- // 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);
- }
+ // Use the register form of the broadcast instruction available on AVX2.
+ if (VT.getSizeInBits() >= 256)
+ Sc = Extract128BitVector(Sc, 0, DAG, dl);
+ return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
}
- DecodePSHUFBMask(RawMask, Mask);
+
+ Ld = Sc.getOperand(0);
+ ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
+ Ld.getOpcode() == ISD::ConstantFP);
+
+ // The scalar_to_vector node and the suspected
+ // load node must have exactly one user.
+ // Constants may have multiple users.
+
+ // AVX-512 has register version of the broadcast
+ bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
+ Ld.getValueType().getSizeInBits() >= 32;
+ if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
+ !hasRegVer))
+ return SDValue();
break;
}
+ }
- auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
- if (!MaskLoad)
- return false;
+ unsigned ScalarSize = Ld.getValueType().getSizeInBits();
+ bool IsGE256 = (VT.getSizeInBits() >= 256);
- SDValue Ptr = MaskLoad->getBasePtr();
- if (Ptr->getOpcode() == X86ISD::Wrapper)
- Ptr = Ptr->getOperand(0);
+ // When optimizing for size, generate up to 5 extra bytes for a broadcast
+ // instruction to save 8 or more bytes of constant pool data.
+ // TODO: If multiple splats are generated to load the same constant,
+ // it may be detrimental to overall size. There needs to be a way to detect
+ // that condition to know if this is truly a size win.
+ const Function *F = DAG.getMachineFunction().getFunction();
+ bool OptForSize = F->hasFnAttribute(Attribute::OptimizeForSize);
- auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
- if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
- return false;
+ // Handle broadcasting a single constant scalar from the constant pool
+ // into a vector.
+ // On Sandybridge (no AVX2), it is still better to load a constant vector
+ // from the constant pool and not to broadcast it from a scalar.
+ // But override that restriction when optimizing for size.
+ // TODO: Check if splatting is recommended for other AVX-capable CPUs.
+ if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
+ EVT CVT = Ld.getValueType();
+ assert(!CVT.isVector() && "Must not broadcast a vector type");
- if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
- // FIXME: Support AVX-512 here.
- Type *Ty = C->getType();
- if (!Ty->isVectorTy() || (Ty->getVectorNumElements() != 16 &&
- Ty->getVectorNumElements() != 32))
- return false;
+ // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
+ // For size optimization, also splat v2f64 and v2i64, and for size opt
+ // with AVX2, also splat i8 and i16.
+ // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
+ if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
+ (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
+ const Constant *C = nullptr;
+ if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
+ C = CI->getConstantIntValue();
+ else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
+ C = CF->getConstantFPValue();
- DecodePSHUFBMask(C, Mask);
- break;
- }
+ assert(C && "Invalid constant type");
- return false;
- }
- case X86ISD::VPERMI:
- ImmN = N->getOperand(N->getNumOperands()-1);
- DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
- IsUnary = true;
- break;
- case X86ISD::MOVSS:
- case X86ISD::MOVSD: {
- // The index 0 always comes from the first element of the second source,
- // this is why MOVSS and MOVSD are used in the first place. The other
- // elements come from the other positions of the first source vector
- Mask.push_back(NumElems);
- for (unsigned i = 1; i != NumElems; ++i) {
- Mask.push_back(i);
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+ SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
+ unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
+ Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
+ MachinePointerInfo::getConstantPool(),
+ false, false, false, Alignment);
+
+ return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
}
- break;
- }
- case X86ISD::VPERM2X128:
- ImmN = N->getOperand(N->getNumOperands()-1);
- DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
- if (Mask.empty()) return false;
- break;
- case X86ISD::MOVSLDUP:
- DecodeMOVSLDUPMask(VT, Mask);
- break;
- case X86ISD::MOVSHDUP:
- DecodeMOVSHDUPMask(VT, Mask);
- break;
- case X86ISD::MOVDDUP:
- case X86ISD::MOVLHPD:
- case X86ISD::MOVLPD:
- case X86ISD::MOVLPS:
- // Not yet implemented
- return false;
- 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;
-}
-
-/// getShuffleScalarElt - Returns the scalar element that will make up the ith
-/// element of the result of the vector shuffle.
-static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
- unsigned Depth) {
- if (Depth == 6)
- return SDValue(); // Limit search depth.
+ bool IsLoad = ISD::isNormalLoad(Ld.getNode());
- SDValue V = SDValue(N, 0);
- EVT VT = V.getValueType();
- unsigned Opcode = V.getOpcode();
+ // Handle AVX2 in-register broadcasts.
+ if (!IsLoad && Subtarget->hasInt256() &&
+ (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
+ return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
- // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
- if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
- int Elt = SV->getMaskElt(Index);
+ // The scalar source must be a normal load.
+ if (!IsLoad)
+ return SDValue();
- if (Elt < 0)
- return DAG.getUNDEF(VT.getVectorElementType());
+ if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
+ (Subtarget->hasVLX() && ScalarSize == 64))
+ return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
- unsigned NumElems = VT.getVectorNumElements();
- SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
- : SV->getOperand(1);
- return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
+ // The integer check is needed for the 64-bit into 128-bit so it doesn't match
+ // double since there is no vbroadcastsd xmm
+ if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
+ if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
+ return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
}
- // Recurse into target specific vector shuffles to find scalars.
- if (isTargetShuffle(Opcode)) {
- MVT ShufVT = V.getSimpleValueType();
- unsigned NumElems = ShufVT.getVectorNumElements();
- SmallVector<int, 16> ShuffleMask;
- bool IsUnary;
+ // Unsupported broadcast.
+ return SDValue();
+}
- if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
- return SDValue();
+/// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
+/// underlying vector and index.
+///
+/// Modifies \p ExtractedFromVec to the real vector and returns the real
+/// index.
+static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
+ SDValue ExtIdx) {
+ int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
+ if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
+ return Idx;
- int Elt = ShuffleMask[Index];
- if (Elt < 0)
- return DAG.getUNDEF(ShufVT.getVectorElementType());
+ // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
+ // lowered this:
+ // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
+ // to:
+ // (extract_vector_elt (vector_shuffle<2,u,u,u>
+ // (extract_subvector (v8f32 %vreg0), Constant<4>),
+ // undef)
+ // Constant<0>)
+ // In this case the vector is the extract_subvector expression and the index
+ // is 2, as specified by the shuffle.
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
+ SDValue ShuffleVec = SVOp->getOperand(0);
+ MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
+ assert(ShuffleVecVT.getVectorElementType() ==
+ ExtractedFromVec.getSimpleValueType().getVectorElementType());
- SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
- : N->getOperand(1);
- return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
- Depth+1);
+ int ShuffleIdx = SVOp->getMaskElt(Idx);
+ if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
+ ExtractedFromVec = ShuffleVec;
+ return ShuffleIdx;
}
+ return Idx;
+}
- // Actual nodes that may contain scalar elements
- if (Opcode == ISD::BITCAST) {
- V = V.getOperand(0);
- EVT SrcVT = V.getValueType();
- unsigned NumElems = VT.getVectorNumElements();
-
- if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
- return SDValue();
- }
+static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
+ MVT VT = Op.getSimpleValueType();
- if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
- return (Index == 0) ? V.getOperand(0)
- : DAG.getUNDEF(VT.getVectorElementType());
+ // Skip if insert_vec_elt is not supported.
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+ if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
+ return SDValue();
- if (V.getOpcode() == ISD::BUILD_VECTOR)
- return V.getOperand(Index);
+ SDLoc DL(Op);
+ unsigned NumElems = Op.getNumOperands();
- return SDValue();
-}
+ SDValue VecIn1;
+ SDValue VecIn2;
+ SmallVector<unsigned, 4> InsertIndices;
+ SmallVector<int, 8> Mask(NumElems, -1);
-/// getNumOfConsecutiveZeros - Return the number of elements of a vector
-/// shuffle operation which come from a consecutively from a zero. The
-/// search can start in two different directions, from left or right.
-/// We count undefs as zeros until PreferredNum is reached.
-static unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp,
- unsigned NumElems, bool ZerosFromLeft,
- SelectionDAG &DAG,
- unsigned PreferredNum = -1U) {
- unsigned NumZeros = 0;
for (unsigned i = 0; i != NumElems; ++i) {
- unsigned Index = ZerosFromLeft ? i : NumElems - i - 1;
- SDValue Elt = getShuffleScalarElt(SVOp, Index, DAG, 0);
- if (!Elt.getNode())
- break;
-
- if (X86::isZeroNode(Elt))
- ++NumZeros;
- else if (Elt.getOpcode() == ISD::UNDEF) // Undef as zero up to PreferredNum.
- NumZeros = std::min(NumZeros + 1, PreferredNum);
- else
- break;
- }
-
- return NumZeros;
-}
+ unsigned Opc = Op.getOperand(i).getOpcode();
-/// isShuffleMaskConsecutive - Check if the shuffle mask indicies [MaskI, MaskE)
-/// correspond consecutively to elements from one of the vector operands,
-/// starting from its index OpIdx. Also tell OpNum which source vector operand.
-static
-bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp,
- unsigned MaskI, unsigned MaskE, unsigned OpIdx,
- unsigned NumElems, unsigned &OpNum) {
- bool SeenV1 = false;
- bool SeenV2 = false;
-
- for (unsigned i = MaskI; i != MaskE; ++i, ++OpIdx) {
- int Idx = SVOp->getMaskElt(i);
- // Ignore undef indicies
- if (Idx < 0)
+ if (Opc == ISD::UNDEF)
continue;
- if (Idx < (int)NumElems)
- SeenV1 = true;
- else
- SeenV2 = true;
-
- // Only accept consecutive elements from the same vector
- if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
- return false;
- }
-
- OpNum = SeenV1 ? 0 : 1;
- return true;
-}
+ if (Opc != ISD::EXTRACT_VECTOR_ELT) {
+ // Quit if more than 1 elements need inserting.
+ if (InsertIndices.size() > 1)
+ return SDValue();
-/// isVectorShiftRight - Returns true if the shuffle can be implemented as a
-/// logical left shift of a vector.
-static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
- bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
- unsigned NumElems =
- SVOp->getSimpleValueType(0).getVectorNumElements();
- unsigned NumZeros = getNumOfConsecutiveZeros(
- SVOp, NumElems, false /* check zeros from right */, DAG,
- SVOp->getMaskElt(0));
- unsigned OpSrc;
+ InsertIndices.push_back(i);
+ continue;
+ }
- if (!NumZeros)
- return false;
+ SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
+ SDValue ExtIdx = Op.getOperand(i).getOperand(1);
+ // Quit if non-constant index.
+ if (!isa<ConstantSDNode>(ExtIdx))
+ return SDValue();
+ int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
- // Considering the elements in the mask that are not consecutive zeros,
- // check if they consecutively come from only one of the source vectors.
- //
- // V1 = {X, A, B, C} 0
- // \ \ \ /
- // vector_shuffle V1, V2 <1, 2, 3, X>
- //
- if (!isShuffleMaskConsecutive(SVOp,
- 0, // Mask Start Index
- NumElems-NumZeros, // Mask End Index(exclusive)
- NumZeros, // Where to start looking in the src vector
- NumElems, // Number of elements in vector
- OpSrc)) // Which source operand ?
- return false;
+ // Quit if extracted from vector of different type.
+ if (ExtractedFromVec.getValueType() != VT)
+ return SDValue();
- isLeft = false;
- ShAmt = NumZeros;
- ShVal = SVOp->getOperand(OpSrc);
- return true;
-}
+ if (!VecIn1.getNode())
+ VecIn1 = ExtractedFromVec;
+ else if (VecIn1 != ExtractedFromVec) {
+ if (!VecIn2.getNode())
+ VecIn2 = ExtractedFromVec;
+ else if (VecIn2 != ExtractedFromVec)
+ // Quit if more than 2 vectors to shuffle
+ return SDValue();
+ }
-/// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
-/// logical left shift of a vector.
-static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
- bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
- unsigned NumElems =
- SVOp->getSimpleValueType(0).getVectorNumElements();
- unsigned NumZeros = getNumOfConsecutiveZeros(
- SVOp, NumElems, true /* check zeros from left */, DAG,
- NumElems - SVOp->getMaskElt(NumElems - 1) - 1);
- unsigned OpSrc;
+ if (ExtractedFromVec == VecIn1)
+ Mask[i] = Idx;
+ else if (ExtractedFromVec == VecIn2)
+ Mask[i] = Idx + NumElems;
+ }
- if (!NumZeros)
- return false;
+ if (!VecIn1.getNode())
+ return SDValue();
- // Considering the elements in the mask that are not consecutive zeros,
- // check if they consecutively come from only one of the source vectors.
- //
- // 0 { A, B, X, X } = V2
- // / \ / /
- // vector_shuffle V1, V2 <X, X, 4, 5>
- //
- if (!isShuffleMaskConsecutive(SVOp,
- NumZeros, // Mask Start Index
- NumElems, // Mask End Index(exclusive)
- 0, // Where to start looking in the src vector
- NumElems, // Number of elements in vector
- OpSrc)) // Which source operand ?
- return false;
+ VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
+ SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
+ for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
+ unsigned Idx = InsertIndices[i];
+ NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
+ DAG.getIntPtrConstant(Idx));
+ }
- isLeft = true;
- ShAmt = NumZeros;
- ShVal = SVOp->getOperand(OpSrc);
- return true;
+ return NV;
}
-/// isVectorShift - Returns true if the shuffle can be implemented as a
-/// logical left or right shift of a vector.
-static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
- bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
- // Although the logic below support any bitwidth size, there are no
- // shift instructions which handle more than 128-bit vectors.
- if (!SVOp->getSimpleValueType(0).is128BitVector())
- return false;
+// Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
+SDValue
+X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
- if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
- isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
- return true;
+ MVT VT = Op.getSimpleValueType();
+ assert((VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits() <= 16) &&
+ "Unexpected type in LowerBUILD_VECTORvXi1!");
- return false;
-}
+ SDLoc dl(Op);
+ if (ISD::isBuildVectorAllZeros(Op.getNode())) {
+ SDValue Cst = DAG.getTargetConstant(0, MVT::i1);
+ SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
+ }
-/// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
-///
-static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
- unsigned NumNonZero, unsigned NumZero,
- SelectionDAG &DAG,
- const X86Subtarget* Subtarget,
- const TargetLowering &TLI) {
- if (NumNonZero > 8)
- return SDValue();
+ if (ISD::isBuildVectorAllOnes(Op.getNode())) {
+ SDValue Cst = DAG.getTargetConstant(1, MVT::i1);
+ SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
+ }
- SDLoc dl(Op);
- SDValue V;
- bool First = true;
- for (unsigned i = 0; i < 16; ++i) {
- bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
- if (ThisIsNonZero && First) {
- if (NumZero)
- V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
- else
- V = DAG.getUNDEF(MVT::v8i16);
- First = false;
+ bool AllContants = true;
+ uint64_t Immediate = 0;
+ int NonConstIdx = -1;
+ bool IsSplat = true;
+ unsigned NumNonConsts = 0;
+ unsigned NumConsts = 0;
+ for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
+ SDValue In = Op.getOperand(idx);
+ if (In.getOpcode() == ISD::UNDEF)
+ continue;
+ if (!isa<ConstantSDNode>(In)) {
+ AllContants = false;
+ NonConstIdx = idx;
+ NumNonConsts++;
+ } else {
+ NumConsts++;
+ if (cast<ConstantSDNode>(In)->getZExtValue())
+ Immediate |= (1ULL << idx);
}
+ if (In != Op.getOperand(0))
+ IsSplat = false;
+ }
- if ((i & 1) != 0) {
- SDValue ThisElt, LastElt;
- bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
- if (LastIsNonZero) {
- LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
- MVT::i16, Op.getOperand(i-1));
- }
- if (ThisIsNonZero) {
- ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
- ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
- ThisElt, DAG.getConstant(8, MVT::i8));
- if (LastIsNonZero)
- ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
- } else
- ThisElt = LastElt;
+ if (AllContants) {
+ SDValue FullMask = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1,
+ DAG.getConstant(Immediate, MVT::i16));
+ return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, FullMask,
+ DAG.getIntPtrConstant(0));
+ }
- if (ThisElt.getNode())
- V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
- DAG.getIntPtrConstant(i/2));
+ if (NumNonConsts == 1 && NonConstIdx != 0) {
+ SDValue DstVec;
+ if (NumConsts) {
+ SDValue VecAsImm = DAG.getConstant(Immediate,
+ MVT::getIntegerVT(VT.getSizeInBits()));
+ DstVec = DAG.getNode(ISD::BITCAST, dl, VT, VecAsImm);
}
+ else
+ DstVec = DAG.getUNDEF(VT);
+ return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
+ Op.getOperand(NonConstIdx),
+ DAG.getIntPtrConstant(NonConstIdx));
}
-
- return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
+ if (!IsSplat && (NonConstIdx != 0))
+ llvm_unreachable("Unsupported BUILD_VECTOR operation");
+ MVT SelectVT = (VT == MVT::v16i1)? MVT::i16 : MVT::i8;
+ SDValue Select;
+ if (IsSplat)
+ Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
+ DAG.getConstant(-1, SelectVT),
+ DAG.getConstant(0, SelectVT));
+ else
+ Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
+ DAG.getConstant((Immediate | 1), SelectVT),
+ DAG.getConstant(Immediate, SelectVT));
+ return DAG.getNode(ISD::BITCAST, dl, VT, Select);
}
-/// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
+/// \brief Return true if \p N implements a horizontal binop and return the
+/// operands for the horizontal binop into V0 and V1.
///
-static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
- unsigned NumNonZero, unsigned NumZero,
- SelectionDAG &DAG,
- const X86Subtarget* Subtarget,
- const TargetLowering &TLI) {
- if (NumNonZero > 4)
- return SDValue();
+/// This is a helper function of PerformBUILD_VECTORCombine.
+/// This function checks that the build_vector \p N in input implements a
+/// horizontal operation. Parameter \p Opcode defines the kind of horizontal
+/// operation to match.
+/// For example, if \p Opcode is equal to ISD::ADD, then this function
+/// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
+/// is equal to ISD::SUB, then this function checks if this is a horizontal
+/// arithmetic sub.
+///
+/// This function only analyzes elements of \p N whose indices are
+/// in range [BaseIdx, LastIdx).
+static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
+ SelectionDAG &DAG,
+ unsigned BaseIdx, unsigned LastIdx,
+ SDValue &V0, SDValue &V1) {
+ EVT VT = N->getValueType(0);
- SDLoc dl(Op);
- SDValue V;
- bool First = true;
- for (unsigned i = 0; i < 8; ++i) {
- bool isNonZero = (NonZeros & (1 << i)) != 0;
- if (isNonZero) {
- if (First) {
- if (NumZero)
- V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
- else
- V = DAG.getUNDEF(MVT::v8i16);
- First = false;
- }
- V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
- MVT::v8i16, V, Op.getOperand(i),
- DAG.getIntPtrConstant(i));
- }
- }
+ assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
+ assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
+ "Invalid Vector in input!");
- return V;
-}
+ bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
+ bool CanFold = true;
+ unsigned ExpectedVExtractIdx = BaseIdx;
+ unsigned NumElts = LastIdx - BaseIdx;
+ V0 = DAG.getUNDEF(VT);
+ V1 = DAG.getUNDEF(VT);
-/// LowerBuildVectorv4x32 - Custom lower build_vector of v4i32 or v4f32.
-static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
- const X86Subtarget *Subtarget,
- const TargetLowering &TLI) {
- // Find all zeroable elements.
- bool Zeroable[4];
- for (int i=0; i < 4; ++i) {
- SDValue Elt = Op->getOperand(i);
- Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
- }
- assert(std::count_if(&Zeroable[0], &Zeroable[4],
- [](bool M) { return !M; }) > 1 &&
- "We expect at least two non-zero elements!");
+ // Check if N implements a horizontal binop.
+ for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
+ SDValue Op = N->getOperand(i + BaseIdx);
- // We only know how to deal with build_vector nodes where elements are either
- // zeroable or extract_vector_elt with constant index.
- SDValue FirstNonZero;
- unsigned FirstNonZeroIdx;
- for (unsigned i=0; i < 4; ++i) {
- if (Zeroable[i])
+ // Skip UNDEFs.
+ if (Op->getOpcode() == ISD::UNDEF) {
+ // Update the expected vector extract index.
+ if (i * 2 == NumElts)
+ ExpectedVExtractIdx = BaseIdx;
+ ExpectedVExtractIdx += 2;
continue;
- SDValue Elt = Op->getOperand(i);
- if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
- !isa<ConstantSDNode>(Elt.getOperand(1)))
- return SDValue();
- // Make sure that this node is extracting from a 128-bit vector.
- MVT VT = Elt.getOperand(0).getSimpleValueType();
- if (!VT.is128BitVector())
- return SDValue();
- if (!FirstNonZero.getNode()) {
- FirstNonZero = Elt;
- FirstNonZeroIdx = i;
}
- }
-
- assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
- SDValue V1 = FirstNonZero.getOperand(0);
- MVT VT = V1.getSimpleValueType();
- // See if this build_vector can be lowered as a blend with zero.
- SDValue Elt;
- unsigned EltMaskIdx, EltIdx;
- int Mask[4];
- for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
- if (Zeroable[EltIdx]) {
- // The zero vector will be on the right hand side.
- Mask[EltIdx] = EltIdx+4;
- continue;
- }
+ CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
- Elt = Op->getOperand(EltIdx);
- // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
- EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
- if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
+ if (!CanFold)
break;
- Mask[EltIdx] = EltIdx;
- }
- if (EltIdx == 4) {
- // Let the shuffle legalizer deal with blend operations.
- SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
- if (V1.getSimpleValueType() != VT)
- V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
- return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
- }
+ SDValue Op0 = Op.getOperand(0);
+ SDValue Op1 = Op.getOperand(1);
- // See if we can lower this build_vector to a INSERTPS.
- if (!Subtarget->hasSSE41())
- return SDValue();
+ // Try to match the following pattern:
+ // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
+ CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
+ Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
+ Op0.getOperand(0) == Op1.getOperand(0) &&
+ isa<ConstantSDNode>(Op0.getOperand(1)) &&
+ isa<ConstantSDNode>(Op1.getOperand(1)));
+ if (!CanFold)
+ break;
- SDValue V2 = Elt.getOperand(0);
- if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
- V1 = SDValue();
+ unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
+ unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
- bool CanFold = true;
- for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
- if (Zeroable[i])
- continue;
+ if (i * 2 < NumElts) {
+ if (V0.getOpcode() == ISD::UNDEF)
+ V0 = Op0.getOperand(0);
+ } else {
+ if (V1.getOpcode() == ISD::UNDEF)
+ V1 = Op0.getOperand(0);
+ if (i * 2 == NumElts)
+ ExpectedVExtractIdx = BaseIdx;
+ }
- SDValue Current = Op->getOperand(i);
- SDValue SrcVector = Current->getOperand(0);
- if (!V1.getNode())
- V1 = SrcVector;
- CanFold = SrcVector == V1 &&
- cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
+ SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
+ if (I0 == ExpectedVExtractIdx)
+ CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
+ else if (IsCommutable && I1 == ExpectedVExtractIdx) {
+ // Try to match the following dag sequence:
+ // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
+ CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
+ } else
+ CanFold = false;
+
+ ExpectedVExtractIdx += 2;
}
- if (!CanFold)
- return SDValue();
+ return CanFold;
+}
- assert(V1.getNode() && "Expected at least two non-zero elements!");
- if (V1.getSimpleValueType() != MVT::v4f32)
- V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
- if (V2.getSimpleValueType() != MVT::v4f32)
- V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
+/// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
+/// a concat_vector.
+///
+/// This is a helper function of PerformBUILD_VECTORCombine.
+/// This function expects two 256-bit vectors called V0 and V1.
+/// At first, each vector is split into two separate 128-bit vectors.
+/// Then, the resulting 128-bit vectors are used to implement two
+/// horizontal binary operations.
+///
+/// The kind of horizontal binary operation is defined by \p X86Opcode.
+///
+/// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
+/// the two new horizontal binop.
+/// When Mode is set, the first horizontal binop dag node would take as input
+/// the lower 128-bit of V0 and the upper 128-bit of V0. The second
+/// horizontal binop dag node would take as input the lower 128-bit of V1
+/// and the upper 128-bit of V1.
+/// Example:
+/// HADD V0_LO, V0_HI
+/// HADD V1_LO, V1_HI
+///
+/// Otherwise, the first horizontal binop dag node takes as input the lower
+/// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
+/// dag node takes the the upper 128-bit of V0 and the upper 128-bit of V1.
+/// Example:
+/// HADD V0_LO, V1_LO
+/// HADD V0_HI, V1_HI
+///
+/// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
+/// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
+/// the upper 128-bits of the result.
+static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
+ SDLoc DL, SelectionDAG &DAG,
+ unsigned X86Opcode, bool Mode,
+ bool isUndefLO, bool isUndefHI) {
+ EVT VT = V0.getValueType();
+ assert(VT.is256BitVector() && VT == V1.getValueType() &&
+ "Invalid nodes in input!");
- // Ok, we can emit an INSERTPS instruction.
- unsigned ZMask = 0;
- for (int i = 0; i < 4; ++i)
- if (Zeroable[i])
- ZMask |= 1 << i;
+ unsigned NumElts = VT.getVectorNumElements();
+ SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
+ SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
+ SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
+ SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
+ EVT NewVT = V0_LO.getValueType();
- unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
- assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
- SDValue Result = DAG.getNode(X86ISD::INSERTPS, SDLoc(Op), MVT::v4f32, V1, V2,
- DAG.getIntPtrConstant(InsertPSMask));
- return DAG.getNode(ISD::BITCAST, SDLoc(Op), VT, Result);
-}
+ SDValue LO = DAG.getUNDEF(NewVT);
+ SDValue HI = DAG.getUNDEF(NewVT);
-/// getVShift - Return a vector logical shift node.
-///
-static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
- unsigned NumBits, SelectionDAG &DAG,
- const TargetLowering &TLI, SDLoc dl) {
- assert(VT.is128BitVector() && "Unknown type for VShift");
- EVT ShVT = MVT::v2i64;
- unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
- SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
- return DAG.getNode(ISD::BITCAST, dl, VT,
- DAG.getNode(Opc, dl, ShVT, SrcOp,
- DAG.getConstant(NumBits,
- TLI.getScalarShiftAmountTy(SrcOp.getValueType()))));
+ if (Mode) {
+ // Don't emit a horizontal binop if the result is expected to be UNDEF.
+ if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
+ LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
+ if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
+ HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
+ } else {
+ // Don't emit a horizontal binop if the result is expected to be UNDEF.
+ if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
+ V1_LO->getOpcode() != ISD::UNDEF))
+ LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
+
+ if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
+ V1_HI->getOpcode() != ISD::UNDEF))
+ HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
+ }
+
+ return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
}
-static SDValue
-LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
+/// \brief Try to fold a build_vector that performs an 'addsub' into the
+/// sequence of 'vadd + vsub + blendi'.
+static SDValue matchAddSub(const BuildVectorSDNode *BV, SelectionDAG &DAG,
+ const X86Subtarget *Subtarget) {
+ SDLoc DL(BV);
+ EVT VT = BV->getValueType(0);
+ unsigned NumElts = VT.getVectorNumElements();
+ SDValue InVec0 = DAG.getUNDEF(VT);
+ SDValue InVec1 = DAG.getUNDEF(VT);
- // Check if the scalar load can be widened into a vector load. And if
- // the address is "base + cst" see if the cst can be "absorbed" into
- // the shuffle mask.
- if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
- SDValue Ptr = LD->getBasePtr();
- if (!ISD::isNormalLoad(LD) || LD->isVolatile())
- return SDValue();
- EVT PVT = LD->getValueType(0);
- if (PVT != MVT::i32 && PVT != MVT::f32)
- return SDValue();
+ assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
+ VT == MVT::v2f64) && "build_vector with an invalid type found!");
- int FI = -1;
- int64_t Offset = 0;
- if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
- FI = FINode->getIndex();
- Offset = 0;
- } else if (DAG.isBaseWithConstantOffset(Ptr) &&
- isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
- FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
- Offset = Ptr.getConstantOperandVal(1);
- Ptr = Ptr.getOperand(0);
- } else {
- return SDValue();
- }
+ // Odd-numbered elements in the input build vector are obtained from
+ // adding two integer/float elements.
+ // Even-numbered elements in the input build vector are obtained from
+ // subtracting two integer/float elements.
+ unsigned ExpectedOpcode = ISD::FSUB;
+ unsigned NextExpectedOpcode = ISD::FADD;
+ bool AddFound = false;
+ bool SubFound = false;
- // FIXME: 256-bit vector instructions don't require a strict alignment,
- // improve this code to support it better.
- unsigned RequiredAlign = VT.getSizeInBits()/8;
- SDValue Chain = LD->getChain();
- // Make sure the stack object alignment is at least 16 or 32.
- MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
- if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
- if (MFI->isFixedObjectIndex(FI)) {
- // Can't change the alignment. FIXME: It's possible to compute
- // the exact stack offset and reference FI + adjust offset instead.
- // If someone *really* cares about this. That's the way to implement it.
- return SDValue();
- } else {
- MFI->setObjectAlignment(FI, RequiredAlign);
- }
+ for (unsigned i = 0, e = NumElts; i != e; ++i) {
+ SDValue Op = BV->getOperand(i);
+
+ // Skip 'undef' values.
+ unsigned Opcode = Op.getOpcode();
+ if (Opcode == ISD::UNDEF) {
+ std::swap(ExpectedOpcode, NextExpectedOpcode);
+ continue;
}
- // (Offset % 16 or 32) must be multiple of 4. Then address is then
- // Ptr + (Offset & ~15).
- if (Offset < 0)
- return SDValue();
- if ((Offset % RequiredAlign) & 3)
+ // Early exit if we found an unexpected opcode.
+ if (Opcode != ExpectedOpcode)
return SDValue();
- int64_t StartOffset = Offset & ~(RequiredAlign-1);
- if (StartOffset)
- Ptr = DAG.getNode(ISD::ADD, SDLoc(Ptr), Ptr.getValueType(),
- Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
-
- int EltNo = (Offset - StartOffset) >> 2;
- unsigned NumElems = VT.getVectorNumElements();
-
- EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
- SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
- LD->getPointerInfo().getWithOffset(StartOffset),
- false, false, false, 0);
- SmallVector<int, 8> Mask;
- for (unsigned i = 0; i != NumElems; ++i)
- Mask.push_back(EltNo);
+ SDValue Op0 = Op.getOperand(0);
+ SDValue Op1 = Op.getOperand(1);
- return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
- }
+ // Try to match the following pattern:
+ // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
+ // Early exit if we cannot match that sequence.
+ if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
+ Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
+ !isa<ConstantSDNode>(Op0.getOperand(1)) ||
+ !isa<ConstantSDNode>(Op1.getOperand(1)) ||
+ Op0.getOperand(1) != Op1.getOperand(1))
+ return SDValue();
- return SDValue();
-}
+ unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
+ if (I0 != i)
+ return SDValue();
-/// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
-/// vector of type 'VT', see if the elements can be replaced by a single large
-/// load which has the same value as a build_vector whose operands are 'elts'.
-///
-/// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
-///
-/// FIXME: we'd also like to handle the case where the last elements are zero
-/// rather than undef via VZEXT_LOAD, but we do not detect that case today.
-/// There's even a handy isZeroNode for that purpose.
-static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
- SDLoc &DL, SelectionDAG &DAG,
- bool isAfterLegalize) {
- EVT EltVT = VT.getVectorElementType();
- unsigned NumElems = Elts.size();
+ // We found a valid add/sub node. Update the information accordingly.
+ if (i & 1)
+ AddFound = true;
+ else
+ SubFound = true;
- LoadSDNode *LDBase = nullptr;
- unsigned LastLoadedElt = -1U;
+ // Update InVec0 and InVec1.
+ if (InVec0.getOpcode() == ISD::UNDEF)
+ InVec0 = Op0.getOperand(0);
+ if (InVec1.getOpcode() == ISD::UNDEF)
+ InVec1 = Op1.getOperand(0);
- // For each element in the initializer, see if we've found a load or an undef.
- // If we don't find an initial load element, or later load elements are
- // non-consecutive, bail out.
- for (unsigned i = 0; i < NumElems; ++i) {
- SDValue Elt = Elts[i];
+ // Make sure that operands in input to each add/sub node always
+ // come from a same pair of vectors.
+ if (InVec0 != Op0.getOperand(0)) {
+ if (ExpectedOpcode == ISD::FSUB)
+ return SDValue();
- if (!Elt.getNode() ||
- (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
- return SDValue();
- if (!LDBase) {
- if (Elt.getNode()->getOpcode() == ISD::UNDEF)
+ // FADD is commutable. Try to commute the operands
+ // and then test again.
+ std::swap(Op0, Op1);
+ if (InVec0 != Op0.getOperand(0))
return SDValue();
- LDBase = cast<LoadSDNode>(Elt.getNode());
- LastLoadedElt = i;
- continue;
}
- if (Elt.getOpcode() == ISD::UNDEF)
- continue;
- LoadSDNode *LD = cast<LoadSDNode>(Elt);
- if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
+ if (InVec1 != Op1.getOperand(0))
return SDValue();
- LastLoadedElt = i;
- }
- // If we have found an entire vector of loads and undefs, then return a large
- // load of the entire vector width starting at the base pointer. If we found
- // consecutive loads for the low half, generate a vzext_load node.
- if (LastLoadedElt == NumElems - 1) {
+ // Update the pair of expected opcodes.
+ std::swap(ExpectedOpcode, NextExpectedOpcode);
+ }
- if (isAfterLegalize &&
- !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
- return SDValue();
+ // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
+ if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
+ InVec1.getOpcode() != ISD::UNDEF)
+ return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
- SDValue NewLd = SDValue();
+ return SDValue();
+}
- NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
- LDBase->getPointerInfo(), LDBase->isVolatile(),
- LDBase->isNonTemporal(), LDBase->isInvariant(),
- LDBase->getAlignment());
-
- if (LDBase->hasAnyUseOfValue(1)) {
- SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
- SDValue(LDBase, 1),
- SDValue(NewLd.getNode(), 1));
- DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
- DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
- SDValue(NewLd.getNode(), 1));
- }
+static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG,
+ const X86Subtarget *Subtarget) {
+ SDLoc DL(N);
+ EVT VT = N->getValueType(0);
+ unsigned NumElts = VT.getVectorNumElements();
+ BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
+ SDValue InVec0, InVec1;
- return NewLd;
+ // Try to match an ADDSUB.
+ if ((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
+ (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))) {
+ SDValue Value = matchAddSub(BV, DAG, Subtarget);
+ if (Value.getNode())
+ return Value;
}
- //TODO: The code below fires only for for loading the low v2i32 / v2f32
- //of a v4i32 / v4f32. It's probably worth generalizing.
- if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
- DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
- SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
- SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
- SDValue ResNode =
- DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
- LDBase->getPointerInfo(),
- LDBase->getAlignment(),
- false/*isVolatile*/, true/*ReadMem*/,
- false/*WriteMem*/);
+ // Try to match horizontal ADD/SUB.
+ unsigned NumUndefsLO = 0;
+ unsigned NumUndefsHI = 0;
+ unsigned Half = NumElts/2;
- // Make sure the newly-created LOAD is in the same position as LDBase in
- // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
- // update uses of LDBase's output chain to use the TokenFactor.
- if (LDBase->hasAnyUseOfValue(1)) {
- SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
- SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
- DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
- DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
- SDValue(ResNode.getNode(), 1));
- }
+ // Count the number of UNDEF operands in the build_vector in input.
+ for (unsigned i = 0, e = Half; i != e; ++i)
+ if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
+ NumUndefsLO++;
- return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
- }
- return SDValue();
-}
+ for (unsigned i = Half, e = NumElts; i != e; ++i)
+ if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
+ NumUndefsHI++;
-/// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
-/// to generate a splat value for the following cases:
-/// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
-/// 2. A splat shuffle which uses a scalar_to_vector node which comes from
-/// a scalar load, or a constant.
-/// The VBROADCAST node is returned when a pattern is found,
-/// or SDValue() otherwise.
-static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
- SelectionDAG &DAG) {
- // VBROADCAST requires AVX.
- // TODO: Splats could be generated for non-AVX CPUs using SSE
- // instructions, but there's less potential gain for only 128-bit vectors.
- if (!Subtarget->hasAVX())
+ // Early exit if this is either a build_vector of all UNDEFs or all the
+ // operands but one are UNDEF.
+ if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
return SDValue();
- MVT VT = Op.getSimpleValueType();
- SDLoc dl(Op);
+ if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
+ // Try to match an SSE3 float HADD/HSUB.
+ if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
+ return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
- assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
- "Unsupported vector type for broadcast.");
+ if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
+ return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
+ } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
+ // Try to match an SSSE3 integer HADD/HSUB.
+ if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
+ return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
- SDValue Ld;
- bool ConstSplatVal;
+ if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
+ return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
+ }
- switch (Op.getOpcode()) {
- default:
- // Unknown pattern found.
- return SDValue();
+ if (!Subtarget->hasAVX())
+ return SDValue();
- case ISD::BUILD_VECTOR: {
- auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
- BitVector UndefElements;
- SDValue Splat = BVOp->getSplatValue(&UndefElements);
+ if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
+ // Try to match an AVX horizontal add/sub of packed single/double
+ // precision floating point values from 256-bit vectors.
+ SDValue InVec2, InVec3;
+ if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
+ isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
+ ((InVec0.getOpcode() == ISD::UNDEF ||
+ InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
+ ((InVec1.getOpcode() == ISD::UNDEF ||
+ InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
+ return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
- // We need a splat of a single value to use broadcast, and it doesn't
- // make any sense if the value is only in one element of the vector.
- if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
- return SDValue();
+ if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
+ isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
+ ((InVec0.getOpcode() == ISD::UNDEF ||
+ InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
+ ((InVec1.getOpcode() == ISD::UNDEF ||
+ InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
+ return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
+ } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
+ // Try to match an AVX2 horizontal add/sub of signed integers.
+ SDValue InVec2, InVec3;
+ unsigned X86Opcode;
+ bool CanFold = true;
- Ld = Splat;
- ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
- Ld.getOpcode() == ISD::ConstantFP);
+ if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
+ isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
+ ((InVec0.getOpcode() == ISD::UNDEF ||
+ InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
+ ((InVec1.getOpcode() == ISD::UNDEF ||
+ InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
+ X86Opcode = X86ISD::HADD;
+ else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
+ isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
+ ((InVec0.getOpcode() == ISD::UNDEF ||
+ InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
+ ((InVec1.getOpcode() == ISD::UNDEF ||
+ InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
+ X86Opcode = X86ISD::HSUB;
+ else
+ CanFold = false;
- // Make sure that all of the users of a non-constant load are from the
- // BUILD_VECTOR node.
- if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
+ if (CanFold) {
+ // Fold this build_vector into a single horizontal add/sub.
+ // Do this only if the target has AVX2.
+ if (Subtarget->hasAVX2())
+ return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
+
+ // Do not try to expand this build_vector into a pair of horizontal
+ // add/sub if we can emit a pair of scalar add/sub.
+ if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
return SDValue();
- break;
+
+ // Convert this build_vector into a pair of horizontal binop followed by
+ // a concat vector.
+ bool isUndefLO = NumUndefsLO == Half;
+ bool isUndefHI = NumUndefsHI == Half;
+ return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
+ isUndefLO, isUndefHI);
}
+ }
- case ISD::VECTOR_SHUFFLE: {
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
+ VT == MVT::v16i16) && Subtarget->hasAVX()) {
+ unsigned X86Opcode;
+ if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
+ X86Opcode = X86ISD::HADD;
+ else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
+ X86Opcode = X86ISD::HSUB;
+ else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
+ X86Opcode = X86ISD::FHADD;
+ else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
+ X86Opcode = X86ISD::FHSUB;
+ else
+ return SDValue();
- // Shuffles must have a splat mask where the first element is
- // broadcasted.
- if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
- return SDValue();
+ // Don't try to expand this build_vector into a pair of horizontal add/sub
+ // if we can simply emit a pair of scalar add/sub.
+ if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
+ return SDValue();
- SDValue Sc = Op.getOperand(0);
- if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
- Sc.getOpcode() != ISD::BUILD_VECTOR) {
+ // Convert this build_vector into two horizontal add/sub followed by
+ // a concat vector.
+ bool isUndefLO = NumUndefsLO == Half;
+ bool isUndefHI = NumUndefsHI == Half;
+ return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
+ isUndefLO, isUndefHI);
+ }
- if (!Subtarget->hasInt256())
- return SDValue();
+ return SDValue();
+}
- // Use the register form of the broadcast instruction available on AVX2.
- if (VT.getSizeInBits() >= 256)
- Sc = Extract128BitVector(Sc, 0, DAG, dl);
- return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
- }
+SDValue
+X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
+ SDLoc dl(Op);
- Ld = Sc.getOperand(0);
- ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
- Ld.getOpcode() == ISD::ConstantFP);
+ MVT VT = Op.getSimpleValueType();
+ MVT ExtVT = VT.getVectorElementType();
+ unsigned NumElems = Op.getNumOperands();
- // The scalar_to_vector node and the suspected
- // load node must have exactly one user.
- // Constants may have multiple users.
+ // Generate vectors for predicate vectors.
+ if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
+ return LowerBUILD_VECTORvXi1(Op, DAG);
- // AVX-512 has register version of the broadcast
- bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
- Ld.getValueType().getSizeInBits() >= 32;
- if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
- !hasRegVer))
- return SDValue();
- break;
- }
- }
+ // Vectors containing all zeros can be matched by pxor and xorps later
+ if (ISD::isBuildVectorAllZeros(Op.getNode())) {
+ // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
+ // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
+ if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
+ return Op;
- unsigned ScalarSize = Ld.getValueType().getSizeInBits();
- bool IsGE256 = (VT.getSizeInBits() >= 256);
+ return getZeroVector(VT, Subtarget, DAG, dl);
+ }
- // When optimizing for size, generate up to 5 extra bytes for a broadcast
- // instruction to save 8 or more bytes of constant pool data.
- // TODO: If multiple splats are generated to load the same constant,
- // it may be detrimental to overall size. There needs to be a way to detect
- // that condition to know if this is truly a size win.
- const Function *F = DAG.getMachineFunction().getFunction();
- bool OptForSize = F->getAttributes().
- hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
-
- // Handle broadcasting a single constant scalar from the constant pool
- // into a vector.
- // On Sandybridge (no AVX2), it is still better to load a constant vector
- // from the constant pool and not to broadcast it from a scalar.
- // But override that restriction when optimizing for size.
- // TODO: Check if splatting is recommended for other AVX-capable CPUs.
- if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
- EVT CVT = Ld.getValueType();
- assert(!CVT.isVector() && "Must not broadcast a vector type");
+ // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
+ // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
+ // vpcmpeqd on 256-bit vectors.
+ if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
+ if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
+ return Op;
- // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
- // For size optimization, also splat v2f64 and v2i64, and for size opt
- // with AVX2, also splat i8 and i16.
- // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
- if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
- (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
- const Constant *C = nullptr;
- if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
- C = CI->getConstantIntValue();
- else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
- C = CF->getConstantFPValue();
+ if (!VT.is512BitVector())
+ return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
+ }
- assert(C && "Invalid constant type");
+ SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG);
+ if (Broadcast.getNode())
+ return Broadcast;
- const TargetLowering &TLI = DAG.getTargetLoweringInfo();
- SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
- unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
- Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
- MachinePointerInfo::getConstantPool(),
- false, false, false, Alignment);
+ unsigned EVTBits = ExtVT.getSizeInBits();
- return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
+ unsigned NumZero = 0;
+ unsigned NumNonZero = 0;
+ unsigned NonZeros = 0;
+ bool IsAllConstants = true;
+ SmallSet<SDValue, 8> Values;
+ for (unsigned i = 0; i < NumElems; ++i) {
+ SDValue Elt = Op.getOperand(i);
+ if (Elt.getOpcode() == ISD::UNDEF)
+ continue;
+ Values.insert(Elt);
+ if (Elt.getOpcode() != ISD::Constant &&
+ Elt.getOpcode() != ISD::ConstantFP)
+ IsAllConstants = false;
+ if (X86::isZeroNode(Elt))
+ NumZero++;
+ else {
+ NonZeros |= (1 << i);
+ NumNonZero++;
}
}
- bool IsLoad = ISD::isNormalLoad(Ld.getNode());
+ // All undef vector. Return an UNDEF. All zero vectors were handled above.
+ if (NumNonZero == 0)
+ return DAG.getUNDEF(VT);
- // Handle AVX2 in-register broadcasts.
- if (!IsLoad && Subtarget->hasInt256() &&
- (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
- return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
+ // Special case for single non-zero, non-undef, element.
+ if (NumNonZero == 1) {
+ unsigned Idx = countTrailingZeros(NonZeros);
+ SDValue Item = Op.getOperand(Idx);
- // The scalar source must be a normal load.
- if (!IsLoad)
- return SDValue();
+ // If this is an insertion of an i64 value on x86-32, and if the top bits of
+ // the value are obviously zero, truncate the value to i32 and do the
+ // insertion that way. Only do this if the value is non-constant or if the
+ // value is a constant being inserted into element 0. It is cheaper to do
+ // a constant pool load than it is to do a movd + shuffle.
+ if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
+ (!IsAllConstants || Idx == 0)) {
+ if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
+ // Handle SSE only.
+ assert(VT == MVT::v2i64 && "Expected an SSE value type!");
+ EVT VecVT = MVT::v4i32;
- if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
- (Subtarget->hasVLX() && ScalarSize == 64))
- return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
+ // Truncate the value (which may itself be a constant) to i32, and
+ // 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);
+ return DAG.getNode(
+ ISD::BITCAST, dl, VT,
+ getShuffleVectorZeroOrUndef(Item, Idx * 2, true, Subtarget, DAG));
+ }
+ }
- // The integer check is needed for the 64-bit into 128-bit so it doesn't match
- // double since there is no vbroadcastsd xmm
- if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
- if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
- return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
- }
+ // If we have a constant or non-constant insertion into the low element of
+ // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
+ // the rest of the elements. This will be matched as movd/movq/movss/movsd
+ // depending on what the source datatype is.
+ if (Idx == 0) {
+ if (NumZero == 0)
+ return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
- // Unsupported broadcast.
- return SDValue();
-}
+ if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
+ (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
+ if (VT.is256BitVector() || VT.is512BitVector()) {
+ SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
+ return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
+ Item, DAG.getIntPtrConstant(0));
+ }
+ assert(VT.is128BitVector() && "Expected an SSE value type!");
+ Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
+ // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
+ return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
+ }
-/// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
-/// underlying vector and index.
-///
-/// Modifies \p ExtractedFromVec to the real vector and returns the real
-/// index.
-static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
- SDValue ExtIdx) {
- int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
- if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
- return Idx;
+ if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
+ Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
+ Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
+ if (VT.is256BitVector()) {
+ SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
+ Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
+ } else {
+ assert(VT.is128BitVector() && "Expected an SSE value type!");
+ Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
+ }
+ return DAG.getNode(ISD::BITCAST, dl, VT, Item);
+ }
+ }
- // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
- // lowered this:
- // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
- // to:
- // (extract_vector_elt (vector_shuffle<2,u,u,u>
- // (extract_subvector (v8f32 %vreg0), Constant<4>),
- // undef)
- // Constant<0>)
- // In this case the vector is the extract_subvector expression and the index
- // is 2, as specified by the shuffle.
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
- SDValue ShuffleVec = SVOp->getOperand(0);
- MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
- assert(ShuffleVecVT.getVectorElementType() ==
- ExtractedFromVec.getSimpleValueType().getVectorElementType());
+ // Is it a vector logical left shift?
+ if (NumElems == 2 && Idx == 1 &&
+ X86::isZeroNode(Op.getOperand(0)) &&
+ !X86::isZeroNode(Op.getOperand(1))) {
+ unsigned NumBits = VT.getSizeInBits();
+ return getVShift(true, VT,
+ DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
+ VT, Op.getOperand(1)),
+ NumBits/2, DAG, *this, dl);
+ }
- int ShuffleIdx = SVOp->getMaskElt(Idx);
- if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
- ExtractedFromVec = ShuffleVec;
- return ShuffleIdx;
+ if (IsAllConstants) // Otherwise, it's better to do a constpool load.
+ return SDValue();
+
+ // Otherwise, if this is a vector with i32 or f32 elements, and the element
+ // is a non-constant being inserted into an element other than the low one,
+ // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
+ // movd/movss) to move this into the low element, then shuffle it into
+ // place.
+ if (EVTBits == 32) {
+ Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
+ return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
+ }
}
- return Idx;
-}
-static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
- MVT VT = Op.getSimpleValueType();
+ // Splat is obviously ok. Let legalizer expand it to a shuffle.
+ if (Values.size() == 1) {
+ if (EVTBits == 32) {
+ // Instead of a shuffle like this:
+ // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
+ // Check if it's possible to issue this instead.
+ // shuffle (vload ptr)), undef, <1, 1, 1, 1>
+ unsigned Idx = countTrailingZeros(NonZeros);
+ SDValue Item = Op.getOperand(Idx);
+ if (Op.getNode()->isOnlyUserOf(Item.getNode()))
+ return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
+ }
+ return SDValue();
+ }
- // Skip if insert_vec_elt is not supported.
- const TargetLowering &TLI = DAG.getTargetLoweringInfo();
- if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
+ // A vector full of immediates; various special cases are already
+ // handled, so this is best done with a single constant-pool load.
+ if (IsAllConstants)
return SDValue();
- SDLoc DL(Op);
- unsigned NumElems = Op.getNumOperands();
+ // For AVX-length vectors, see if we can use a vector load to get all of the
+ // elements, otherwise build the individual 128-bit pieces and use
+ // shuffles to put them in place.
+ if (VT.is256BitVector() || VT.is512BitVector()) {
+ SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
- SDValue VecIn1;
- SDValue VecIn2;
- SmallVector<unsigned, 4> InsertIndices;
- SmallVector<int, 8> Mask(NumElems, -1);
+ // Check for a build vector of consecutive loads.
+ if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
+ return LD;
- for (unsigned i = 0; i != NumElems; ++i) {
- unsigned Opc = Op.getOperand(i).getOpcode();
+ EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
- if (Opc == ISD::UNDEF)
- continue;
+ // Build both the lower and upper subvector.
+ SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
+ makeArrayRef(&V[0], NumElems/2));
+ SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
+ makeArrayRef(&V[NumElems / 2], NumElems/2));
- if (Opc != ISD::EXTRACT_VECTOR_ELT) {
- // Quit if more than 1 elements need inserting.
- if (InsertIndices.size() > 1)
- return SDValue();
+ // Recreate the wider vector with the lower and upper part.
+ if (VT.is256BitVector())
+ return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
+ return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
+ }
- InsertIndices.push_back(i);
- continue;
+ // Let legalizer expand 2-wide build_vectors.
+ if (EVTBits == 64) {
+ if (NumNonZero == 1) {
+ // One half is zero or undef.
+ unsigned Idx = countTrailingZeros(NonZeros);
+ SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
+ Op.getOperand(Idx));
+ return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
}
+ return SDValue();
+ }
- SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
- SDValue ExtIdx = Op.getOperand(i).getOperand(1);
- // Quit if non-constant index.
- if (!isa<ConstantSDNode>(ExtIdx))
- return SDValue();
- int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
-
- // Quit if extracted from vector of different type.
- if (ExtractedFromVec.getValueType() != VT)
- return SDValue();
+ // If element VT is < 32 bits, convert it to inserts into a zero vector.
+ if (EVTBits == 8 && NumElems == 16) {
+ SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
+ Subtarget, *this);
+ if (V.getNode()) return V;
+ }
- if (!VecIn1.getNode())
- VecIn1 = ExtractedFromVec;
- else if (VecIn1 != ExtractedFromVec) {
- if (!VecIn2.getNode())
- VecIn2 = ExtractedFromVec;
- else if (VecIn2 != ExtractedFromVec)
- // Quit if more than 2 vectors to shuffle
- return SDValue();
- }
+ if (EVTBits == 16 && NumElems == 8) {
+ SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
+ Subtarget, *this);
+ if (V.getNode()) return V;
+ }
- if (ExtractedFromVec == VecIn1)
- Mask[i] = Idx;
- else if (ExtractedFromVec == VecIn2)
- Mask[i] = Idx + NumElems;
+ // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
+ if (EVTBits == 32 && NumElems == 4) {
+ SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this);
+ if (V.getNode())
+ return V;
}
- if (!VecIn1.getNode())
- return SDValue();
+ // If element VT is == 32 bits, turn it into a number of shuffles.
+ SmallVector<SDValue, 8> V(NumElems);
+ if (NumElems == 4 && NumZero > 0) {
+ for (unsigned i = 0; i < 4; ++i) {
+ bool isZero = !(NonZeros & (1 << i));
+ if (isZero)
+ V[i] = getZeroVector(VT, Subtarget, DAG, dl);
+ else
+ V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
+ }
- VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
- SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
- for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
- unsigned Idx = InsertIndices[i];
- NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
- DAG.getIntPtrConstant(Idx));
+ for (unsigned i = 0; i < 2; ++i) {
+ switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
+ default: break;
+ case 0:
+ V[i] = V[i*2]; // Must be a zero vector.
+ break;
+ case 1:
+ V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
+ break;
+ case 2:
+ V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
+ break;
+ case 3:
+ V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
+ break;
+ }
+ }
+
+ bool Reverse1 = (NonZeros & 0x3) == 2;
+ bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
+ int MaskVec[] = {
+ Reverse1 ? 1 : 0,
+ Reverse1 ? 0 : 1,
+ static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
+ static_cast<int>(Reverse2 ? NumElems : NumElems+1)
+ };
+ return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
}
- return NV;
-}
+ if (Values.size() > 1 && VT.is128BitVector()) {
+ // Check for a build vector of consecutive loads.
+ for (unsigned i = 0; i < NumElems; ++i)
+ V[i] = Op.getOperand(i);
-// Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
-SDValue
-X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
+ // Check for elements which are consecutive loads.
+ SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false);
+ if (LD.getNode())
+ return LD;
- MVT VT = Op.getSimpleValueType();
- assert((VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits() <= 16) &&
- "Unexpected type in LowerBUILD_VECTORvXi1!");
+ // Check for a build vector from mostly shuffle plus few inserting.
+ SDValue Sh = buildFromShuffleMostly(Op, DAG);
+ if (Sh.getNode())
+ return Sh;
- SDLoc dl(Op);
- if (ISD::isBuildVectorAllZeros(Op.getNode())) {
- SDValue Cst = DAG.getTargetConstant(0, MVT::i1);
- SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
- return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
- }
+ // For SSE 4.1, use insertps to put the high elements into the low element.
+ if (Subtarget->hasSSE41()) {
+ SDValue Result;
+ if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
+ Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
+ else
+ Result = DAG.getUNDEF(VT);
- if (ISD::isBuildVectorAllOnes(Op.getNode())) {
- SDValue Cst = DAG.getTargetConstant(1, MVT::i1);
- SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
- return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
- }
+ for (unsigned i = 1; i < NumElems; ++i) {
+ if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
+ Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
+ Op.getOperand(i), DAG.getIntPtrConstant(i));
+ }
+ return Result;
+ }
- bool AllContants = true;
- uint64_t Immediate = 0;
- int NonConstIdx = -1;
- bool IsSplat = true;
- unsigned NumNonConsts = 0;
- unsigned NumConsts = 0;
- for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
- SDValue In = Op.getOperand(idx);
- if (In.getOpcode() == ISD::UNDEF)
- continue;
- if (!isa<ConstantSDNode>(In)) {
- AllContants = false;
- NonConstIdx = idx;
- NumNonConsts++;
- } else {
- NumConsts++;
- if (cast<ConstantSDNode>(In)->getZExtValue())
- Immediate |= (1ULL << idx);
+ // Otherwise, expand into a number of unpckl*, start by extending each of
+ // our (non-undef) elements to the full vector width with the element in the
+ // bottom slot of the vector (which generates no code for SSE).
+ for (unsigned i = 0; i < NumElems; ++i) {
+ if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
+ V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
+ else
+ V[i] = DAG.getUNDEF(VT);
}
- if (In != Op.getOperand(0))
- IsSplat = false;
- }
- if (AllContants) {
- SDValue FullMask = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1,
- DAG.getConstant(Immediate, MVT::i16));
- return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, FullMask,
- DAG.getIntPtrConstant(0));
- }
+ // Next, we iteratively mix elements, e.g. for v4f32:
+ // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
+ // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
+ // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
+ unsigned EltStride = NumElems >> 1;
+ while (EltStride != 0) {
+ for (unsigned i = 0; i < EltStride; ++i) {
+ // If V[i+EltStride] is undef and this is the first round of mixing,
+ // then it is safe to just drop this shuffle: V[i] is already in the
+ // right place, the one element (since it's the first round) being
+ // inserted as undef can be dropped. This isn't safe for successive
+ // rounds because they will permute elements within both vectors.
+ if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
+ EltStride == NumElems/2)
+ continue;
- if (NumNonConsts == 1 && NonConstIdx != 0) {
- SDValue DstVec;
- if (NumConsts) {
- SDValue VecAsImm = DAG.getConstant(Immediate,
- MVT::getIntegerVT(VT.getSizeInBits()));
- DstVec = DAG.getNode(ISD::BITCAST, dl, VT, VecAsImm);
+ V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
+ }
+ EltStride >>= 1;
}
- else
- DstVec = DAG.getUNDEF(VT);
- return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
- Op.getOperand(NonConstIdx),
- DAG.getIntPtrConstant(NonConstIdx));
+ return V[0];
}
- if (!IsSplat && (NonConstIdx != 0))
- llvm_unreachable("Unsupported BUILD_VECTOR operation");
- MVT SelectVT = (VT == MVT::v16i1)? MVT::i16 : MVT::i8;
- SDValue Select;
- if (IsSplat)
- Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
- DAG.getConstant(-1, SelectVT),
- DAG.getConstant(0, SelectVT));
- else
- Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
- DAG.getConstant((Immediate | 1), SelectVT),
- DAG.getConstant(Immediate, SelectVT));
- return DAG.getNode(ISD::BITCAST, dl, VT, Select);
+ return SDValue();
}
-/// \brief Return true if \p N implements a horizontal binop and return the
-/// operands for the horizontal binop into V0 and V1.
-///
-/// This is a helper function of PerformBUILD_VECTORCombine.
-/// This function checks that the build_vector \p N in input implements a
-/// horizontal operation. Parameter \p Opcode defines the kind of horizontal
-/// operation to match.
-/// For example, if \p Opcode is equal to ISD::ADD, then this function
-/// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
-/// is equal to ISD::SUB, then this function checks if this is a horizontal
-/// arithmetic sub.
-///
-/// This function only analyzes elements of \p N whose indices are
-/// in range [BaseIdx, LastIdx).
-static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
- SelectionDAG &DAG,
- unsigned BaseIdx, unsigned LastIdx,
- SDValue &V0, SDValue &V1) {
- EVT VT = N->getValueType(0);
-
- assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
- assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
- "Invalid Vector in input!");
+// LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
+// to create 256-bit vectors from two other 128-bit ones.
+static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
+ SDLoc dl(Op);
+ MVT ResVT = Op.getSimpleValueType();
- bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
- bool CanFold = true;
- unsigned ExpectedVExtractIdx = BaseIdx;
- unsigned NumElts = LastIdx - BaseIdx;
- V0 = DAG.getUNDEF(VT);
- V1 = DAG.getUNDEF(VT);
+ assert((ResVT.is256BitVector() ||
+ ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
- // Check if N implements a horizontal binop.
- for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
- SDValue Op = N->getOperand(i + BaseIdx);
+ SDValue V1 = Op.getOperand(0);
+ SDValue V2 = Op.getOperand(1);
+ unsigned NumElems = ResVT.getVectorNumElements();
+ if(ResVT.is256BitVector())
+ return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
- // Skip UNDEFs.
- if (Op->getOpcode() == ISD::UNDEF) {
- // Update the expected vector extract index.
- if (i * 2 == NumElts)
- ExpectedVExtractIdx = BaseIdx;
- ExpectedVExtractIdx += 2;
- continue;
- }
-
- CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
+ if (Op.getNumOperands() == 4) {
+ MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
+ ResVT.getVectorNumElements()/2);
+ SDValue V3 = Op.getOperand(2);
+ SDValue V4 = Op.getOperand(3);
+ return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
+ Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
+ }
+ return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
+}
- if (!CanFold)
- break;
+static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
+ MVT LLVM_ATTRIBUTE_UNUSED VT = Op.getSimpleValueType();
+ assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
+ (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
+ Op.getNumOperands() == 4)));
- SDValue Op0 = Op.getOperand(0);
- SDValue Op1 = Op.getOperand(1);
+ // AVX can use the vinsertf128 instruction to create 256-bit vectors
+ // from two other 128-bit ones.
- // Try to match the following pattern:
- // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
- CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
- Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
- Op0.getOperand(0) == Op1.getOperand(0) &&
- isa<ConstantSDNode>(Op0.getOperand(1)) &&
- isa<ConstantSDNode>(Op1.getOperand(1)));
- if (!CanFold)
- break;
+ // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
+ return LowerAVXCONCAT_VECTORS(Op, DAG);
+}
- unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
- unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
- if (i * 2 < NumElts) {
- if (V0.getOpcode() == ISD::UNDEF)
- V0 = Op0.getOperand(0);
- } else {
- if (V1.getOpcode() == ISD::UNDEF)
- V1 = Op0.getOperand(0);
- if (i * 2 == NumElts)
- ExpectedVExtractIdx = BaseIdx;
- }
+//===----------------------------------------------------------------------===//
+// Vector shuffle lowering
+//
+// This is an experimental code path for lowering vector shuffles on x86. It is
+// designed to handle arbitrary vector shuffles and blends, gracefully
+// degrading performance as necessary. It works hard to recognize idiomatic
+// shuffles and lower them to optimal instruction patterns without leaving
+// a framework that allows reasonably efficient handling of all vector shuffle
+// patterns.
+//===----------------------------------------------------------------------===//
- SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
- if (I0 == ExpectedVExtractIdx)
- CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
- else if (IsCommutable && I1 == ExpectedVExtractIdx) {
- // Try to match the following dag sequence:
- // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
- CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
- } else
- CanFold = false;
+/// \brief Tiny helper function to identify a no-op mask.
+///
+/// This is a somewhat boring predicate function. It checks whether the mask
+/// array input, which is assumed to be a single-input shuffle mask of the kind
+/// used by the X86 shuffle instructions (not a fully general
+/// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
+/// in-place shuffle are 'no-op's.
+static bool isNoopShuffleMask(ArrayRef<int> Mask) {
+ for (int i = 0, Size = Mask.size(); i < Size; ++i)
+ if (Mask[i] != -1 && Mask[i] != i)
+ return false;
+ return true;
+}
- ExpectedVExtractIdx += 2;
- }
+/// \brief Helper function to classify a mask as a single-input mask.
+///
+/// This isn't a generic single-input test because in the vector shuffle
+/// lowering we canonicalize single inputs to be the first input operand. This
+/// means we can more quickly test for a single input by only checking whether
+/// an input from the second operand exists. We also assume that the size of
+/// mask corresponds to the size of the input vectors which isn't true in the
+/// fully general case.
+static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
+ for (int M : Mask)
+ if (M >= (int)Mask.size())
+ return false;
+ return true;
+}
- return CanFold;
+/// \brief Test whether there are elements crossing 128-bit lanes in this
+/// shuffle mask.
+///
+/// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
+/// and we routinely test for these.
+static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
+ int LaneSize = 128 / VT.getScalarSizeInBits();
+ int Size = Mask.size();
+ for (int i = 0; i < Size; ++i)
+ if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
+ return true;
+ return false;
}
-/// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
-/// a concat_vector.
+/// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
///
-/// This is a helper function of PerformBUILD_VECTORCombine.
-/// This function expects two 256-bit vectors called V0 and V1.
-/// At first, each vector is split into two separate 128-bit vectors.
-/// Then, the resulting 128-bit vectors are used to implement two
-/// horizontal binary operations.
+/// This checks a shuffle mask to see if it is performing the same
+/// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
+/// that it is also not lane-crossing. It may however involve a blend from the
+/// same lane of a second vector.
///
-/// The kind of horizontal binary operation is defined by \p X86Opcode.
+/// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
+/// non-trivial to compute in the face of undef lanes. The representation is
+/// *not* suitable for use with existing 128-bit shuffles as it will contain
+/// entries from both V1 and V2 inputs to the wider mask.
+static bool
+is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
+ SmallVectorImpl<int> &RepeatedMask) {
+ int LaneSize = 128 / VT.getScalarSizeInBits();
+ RepeatedMask.resize(LaneSize, -1);
+ int Size = Mask.size();
+ for (int i = 0; i < Size; ++i) {
+ if (Mask[i] < 0)
+ continue;
+ if ((Mask[i] % Size) / LaneSize != i / LaneSize)
+ // This entry crosses lanes, so there is no way to model this shuffle.
+ return false;
+
+ // Ok, handle the in-lane shuffles by detecting if and when they repeat.
+ if (RepeatedMask[i % LaneSize] == -1)
+ // This is the first non-undef entry in this slot of a 128-bit lane.
+ RepeatedMask[i % LaneSize] =
+ Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
+ else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
+ // Found a mismatch with the repeated mask.
+ return false;
+ }
+ return true;
+}
+
+/// \brief Checks whether a shuffle mask is equivalent to an explicit list of
+/// arguments.
///
-/// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
-/// the two new horizontal binop.
-/// When Mode is set, the first horizontal binop dag node would take as input
-/// the lower 128-bit of V0 and the upper 128-bit of V0. The second
-/// horizontal binop dag node would take as input the lower 128-bit of V1
-/// and the upper 128-bit of V1.
-/// Example:
-/// HADD V0_LO, V0_HI
-/// HADD V1_LO, V1_HI
+/// This is a fast way to test a shuffle mask against a fixed pattern:
///
-/// Otherwise, the first horizontal binop dag node takes as input the lower
-/// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
-/// dag node takes the the upper 128-bit of V0 and the upper 128-bit of V1.
-/// Example:
-/// HADD V0_LO, V1_LO
-/// HADD V0_HI, V1_HI
+/// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
///
-/// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
-/// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
-/// the upper 128-bits of the result.
-static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
- SDLoc DL, SelectionDAG &DAG,
- unsigned X86Opcode, bool Mode,
- bool isUndefLO, bool isUndefHI) {
- EVT VT = V0.getValueType();
- assert(VT.is256BitVector() && VT == V1.getValueType() &&
- "Invalid nodes in input!");
+/// 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 bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
+ ArrayRef<int> ExpectedMask) {
+ if (Mask.size() != ExpectedMask.size())
+ return false;
- unsigned NumElts = VT.getVectorNumElements();
- SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
- SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
- SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
- SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
- EVT NewVT = V0_LO.getValueType();
+ int Size = Mask.size();
- SDValue LO = DAG.getUNDEF(NewVT);
- SDValue HI = DAG.getUNDEF(NewVT);
+ // If the values are build vectors, we can look through them to find
+ // equivalent inputs that make the shuffles equivalent.
+ auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
+ auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
- if (Mode) {
- // Don't emit a horizontal binop if the result is expected to be UNDEF.
- if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
- LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
- if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
- HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
- } else {
- // Don't emit a horizontal binop if the result is expected to be UNDEF.
- if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
- V1_LO->getOpcode() != ISD::UNDEF))
- LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
+ for (int i = 0; i < Size; ++i)
+ if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
+ auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
+ auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
+ if (!MaskBV || !ExpectedBV ||
+ MaskBV->getOperand(Mask[i] % Size) !=
+ ExpectedBV->getOperand(ExpectedMask[i] % Size))
+ return false;
+ }
- if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
- V1_HI->getOpcode() != ISD::UNDEF))
- HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
- }
+ return true;
+}
- return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
+/// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
+///
+/// This helper function produces an 8-bit shuffle immediate corresponding to
+/// the ubiquitous shuffle encoding scheme used in x86 instructions for
+/// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
+/// example.
+///
+/// NB: We rely heavily on "undef" masks preserving the input lane.
+static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask,
+ SelectionDAG &DAG) {
+ assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
+ assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
+ assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
+ assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
+ assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
+
+ unsigned Imm = 0;
+ Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
+ Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
+ Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
+ Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
+ return DAG.getConstant(Imm, MVT::i8);
}
-/// \brief Try to fold a build_vector that performs an 'addsub' into the
-/// sequence of 'vadd + vsub + blendi'.
-static SDValue matchAddSub(const BuildVectorSDNode *BV, SelectionDAG &DAG,
- const X86Subtarget *Subtarget) {
- SDLoc DL(BV);
- EVT VT = BV->getValueType(0);
- unsigned NumElts = VT.getVectorNumElements();
- SDValue InVec0 = DAG.getUNDEF(VT);
- SDValue InVec1 = DAG.getUNDEF(VT);
+/// \brief Try to emit a blend instruction for a shuffle using bit math.
+///
+/// This is used as a fallback approach when first class blend instructions are
+/// unavailable. Currently it is only suitable for integer vectors, but could
+/// be generalized for floating point vectors if desirable.
+static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
+ SDValue V2, ArrayRef<int> Mask,
+ SelectionDAG &DAG) {
+ assert(VT.isInteger() && "Only supports integer vector types!");
+ MVT EltVT = VT.getScalarType();
+ int NumEltBits = EltVT.getSizeInBits();
+ SDValue Zero = DAG.getConstant(0, EltVT);
+ SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), EltVT);
+ SmallVector<SDValue, 16> MaskOps;
+ for (int i = 0, Size = Mask.size(); i < Size; ++i) {
+ if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
+ return SDValue(); // Shuffled input!
+ MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
+ }
- assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
- VT == MVT::v2f64) && "build_vector with an invalid type found!");
+ SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
+ V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
+ // We have to cast V2 around.
+ MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
+ V2 = DAG.getNode(ISD::BITCAST, DL, VT,
+ DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
+ DAG.getNode(ISD::BITCAST, DL, MaskVT, V1Mask),
+ DAG.getNode(ISD::BITCAST, DL, MaskVT, V2)));
+ return DAG.getNode(ISD::OR, DL, VT, V1, V2);
+}
- // Odd-numbered elements in the input build vector are obtained from
- // adding two integer/float elements.
- // Even-numbered elements in the input build vector are obtained from
- // subtracting two integer/float elements.
- unsigned ExpectedOpcode = ISD::FSUB;
- unsigned NextExpectedOpcode = ISD::FADD;
- bool AddFound = false;
- bool SubFound = false;
-
- for (unsigned i = 0, e = NumElts; i != e; i++) {
- SDValue Op = BV->getOperand(i);
-
- // Skip 'undef' values.
- unsigned Opcode = Op.getOpcode();
- if (Opcode == ISD::UNDEF) {
- std::swap(ExpectedOpcode, NextExpectedOpcode);
+/// \brief Try to emit a blend instruction for a shuffle.
+///
+/// This doesn't do any checks for the availability of instructions for blending
+/// these values. It relies on the availability of the X86ISD::BLENDI pattern to
+/// be matched in the backend with the type given. What it does check for is
+/// that the shuffle mask is in fact a blend.
+static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
+ SDValue V2, ArrayRef<int> Mask,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ unsigned BlendMask = 0;
+ for (int i = 0, Size = Mask.size(); i < Size; ++i) {
+ if (Mask[i] >= Size) {
+ if (Mask[i] != i + Size)
+ return SDValue(); // Shuffled V2 input!
+ BlendMask |= 1u << i;
continue;
}
+ if (Mask[i] >= 0 && Mask[i] != i)
+ return SDValue(); // Shuffled V1 input!
+ }
+ switch (VT.SimpleTy) {
+ case MVT::v2f64:
+ case MVT::v4f32:
+ case MVT::v4f64:
+ case MVT::v8f32:
+ return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
+ DAG.getConstant(BlendMask, MVT::i8));
- // Early exit if we found an unexpected opcode.
- if (Opcode != ExpectedOpcode)
- return SDValue();
-
- SDValue Op0 = Op.getOperand(0);
- SDValue Op1 = Op.getOperand(1);
+ case MVT::v4i64:
+ case MVT::v8i32:
+ assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
+ // FALLTHROUGH
+ case MVT::v2i64:
+ case MVT::v4i32:
+ // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
+ // that instruction.
+ if (Subtarget->hasAVX2()) {
+ // Scale the blend by the number of 32-bit dwords per element.
+ int Scale = VT.getScalarSizeInBits() / 32;
+ BlendMask = 0;
+ for (int i = 0, Size = Mask.size(); i < Size; ++i)
+ if (Mask[i] >= Size)
+ for (int j = 0; j < Scale; ++j)
+ BlendMask |= 1u << (i * Scale + j);
- // Try to match the following pattern:
- // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
- // Early exit if we cannot match that sequence.
- if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
- Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
- !isa<ConstantSDNode>(Op0.getOperand(1)) ||
- !isa<ConstantSDNode>(Op1.getOperand(1)) ||
- Op0.getOperand(1) != Op1.getOperand(1))
- return SDValue();
+ MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
+ V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
+ V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
+ return DAG.getNode(ISD::BITCAST, DL, VT,
+ DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
+ DAG.getConstant(BlendMask, MVT::i8)));
+ }
+ // FALLTHROUGH
+ case MVT::v8i16: {
+ // For integer shuffles we need to expand the mask and cast the inputs to
+ // v8i16s prior to blending.
+ int Scale = 8 / VT.getVectorNumElements();
+ BlendMask = 0;
+ for (int i = 0, Size = Mask.size(); i < Size; ++i)
+ if (Mask[i] >= Size)
+ for (int j = 0; j < Scale; ++j)
+ BlendMask |= 1u << (i * Scale + j);
- unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
- if (I0 != i)
- return SDValue();
+ V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
+ V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
+ return DAG.getNode(ISD::BITCAST, DL, VT,
+ DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
+ DAG.getConstant(BlendMask, MVT::i8)));
+ }
- // We found a valid add/sub node. Update the information accordingly.
- if (i & 1)
- AddFound = true;
- else
- SubFound = true;
+ case MVT::v16i16: {
+ assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
+ SmallVector<int, 8> RepeatedMask;
+ if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
+ // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
+ assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
+ BlendMask = 0;
+ for (int i = 0; i < 8; ++i)
+ if (RepeatedMask[i] >= 16)
+ BlendMask |= 1u << i;
+ return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
+ DAG.getConstant(BlendMask, MVT::i8));
+ }
+ }
+ // FALLTHROUGH
+ case MVT::v16i8:
+ case MVT::v32i8: {
+ assert((VT.getSizeInBits() == 128 || Subtarget->hasAVX2()) &&
+ "256-bit byte-blends require AVX2 support!");
- // Update InVec0 and InVec1.
- if (InVec0.getOpcode() == ISD::UNDEF)
- InVec0 = Op0.getOperand(0);
- if (InVec1.getOpcode() == ISD::UNDEF)
- InVec1 = Op1.getOperand(0);
+ // Scale the blend by the number of bytes per element.
+ int Scale = VT.getScalarSizeInBits() / 8;
- // Make sure that operands in input to each add/sub node always
- // come from a same pair of vectors.
- if (InVec0 != Op0.getOperand(0)) {
- if (ExpectedOpcode == ISD::FSUB)
- return SDValue();
+ // This form of blend is always done on bytes. Compute the byte vector
+ // type.
+ MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
- // FADD is commutable. Try to commute the operands
- // and then test again.
- std::swap(Op0, Op1);
- if (InVec0 != Op0.getOperand(0))
- return SDValue();
- }
+ // Compute the VSELECT mask. Note that VSELECT is really confusing in the
+ // mix of LLVM's code generator and the x86 backend. We tell the code
+ // generator that boolean values in the elements of an x86 vector register
+ // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
+ // mapping a select to operand #1, and 'false' mapping to operand #2. The
+ // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
+ // of the element (the remaining are ignored) and 0 in that high bit would
+ // mean operand #1 while 1 in the high bit would mean operand #2. So while
+ // the LLVM model for boolean values in vector elements gets the relevant
+ // bit set, it is set backwards and over constrained relative to x86's
+ // actual model.
+ SmallVector<SDValue, 32> VSELECTMask;
+ for (int i = 0, Size = Mask.size(); i < Size; ++i)
+ for (int j = 0; j < Scale; ++j)
+ VSELECTMask.push_back(
+ Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
+ : DAG.getConstant(Mask[i] < Size ? -1 : 0, MVT::i8));
- if (InVec1 != Op1.getOperand(0))
- return SDValue();
+ V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
+ V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
+ return DAG.getNode(
+ ISD::BITCAST, DL, VT,
+ DAG.getNode(ISD::VSELECT, DL, BlendVT,
+ DAG.getNode(ISD::BUILD_VECTOR, DL, BlendVT, VSELECTMask),
+ V1, V2));
+ }
- // Update the pair of expected opcodes.
- std::swap(ExpectedOpcode, NextExpectedOpcode);
+ default:
+ llvm_unreachable("Not a supported integer vector type!");
}
+}
- // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
- if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
- InVec1.getOpcode() != ISD::UNDEF)
- return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
+/// \brief Try to lower as a blend of elements from two inputs followed by
+/// a single-input permutation.
+///
+/// This matches the pattern where we can blend elements from two inputs and
+/// then reduce the shuffle to a single-input permutation.
+static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
+ SDValue V2,
+ ArrayRef<int> Mask,
+ SelectionDAG &DAG) {
+ // We build up the blend mask while checking whether a blend is a viable way
+ // to reduce the shuffle.
+ SmallVector<int, 32> BlendMask(Mask.size(), -1);
+ SmallVector<int, 32> PermuteMask(Mask.size(), -1);
- return SDValue();
-}
+ for (int i = 0, Size = Mask.size(); i < Size; ++i) {
+ if (Mask[i] < 0)
+ continue;
-static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG,
- const X86Subtarget *Subtarget) {
- SDLoc DL(N);
- EVT VT = N->getValueType(0);
- unsigned NumElts = VT.getVectorNumElements();
- BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
- SDValue InVec0, InVec1;
+ assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
- // Try to match an ADDSUB.
- if ((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
- (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))) {
- SDValue Value = matchAddSub(BV, DAG, Subtarget);
- if (Value.getNode())
- return Value;
+ if (BlendMask[Mask[i] % Size] == -1)
+ BlendMask[Mask[i] % Size] = Mask[i];
+ else if (BlendMask[Mask[i] % Size] != Mask[i])
+ return SDValue(); // Can't blend in the needed input!
+
+ PermuteMask[i] = Mask[i] % Size;
}
- // Try to match horizontal ADD/SUB.
- unsigned NumUndefsLO = 0;
- unsigned NumUndefsHI = 0;
- unsigned Half = NumElts/2;
+ SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
+ return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
+}
- // Count the number of UNDEF operands in the build_vector in input.
- for (unsigned i = 0, e = Half; i != e; ++i)
- if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
- NumUndefsLO++;
+/// \brief Generic routine to decompose a shuffle and blend into indepndent
+/// blends and permutes.
+///
+/// This matches the extremely common pattern for handling combined
+/// shuffle+blend operations on newer X86 ISAs where we have very fast blend
+/// operations. It will try to pick the best arrangement of shuffles and
+/// blends.
+static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
+ SDValue V1,
+ SDValue V2,
+ ArrayRef<int> Mask,
+ SelectionDAG &DAG) {
+ // Shuffle the input elements into the desired positions in V1 and V2 and
+ // blend them together.
+ SmallVector<int, 32> V1Mask(Mask.size(), -1);
+ SmallVector<int, 32> V2Mask(Mask.size(), -1);
+ SmallVector<int, 32> BlendMask(Mask.size(), -1);
+ for (int i = 0, Size = Mask.size(); i < Size; ++i)
+ if (Mask[i] >= 0 && Mask[i] < Size) {
+ V1Mask[i] = Mask[i];
+ BlendMask[i] = i;
+ } else if (Mask[i] >= Size) {
+ V2Mask[i] = Mask[i] - Size;
+ BlendMask[i] = i + Size;
+ }
- for (unsigned i = Half, e = NumElts; i != e; ++i)
- if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
- NumUndefsHI++;
+ // Try to lower with the simpler initial blend strategy unless one of the
+ // input shuffles would be a no-op. We prefer to shuffle inputs as the
+ // shuffle may be able to fold with a load or other benefit. However, when
+ // we'll have to do 2x as many shuffles in order to achieve this, blending
+ // first is a better strategy.
+ if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
+ if (SDValue BlendPerm =
+ lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
+ return BlendPerm;
- // Early exit if this is either a build_vector of all UNDEFs or all the
- // operands but one are UNDEF.
- if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
- return SDValue();
+ V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
+ V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
+ return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
+}
- if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
- // Try to match an SSE3 float HADD/HSUB.
- if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
- return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
+/// \brief Try to lower a vector shuffle as a byte rotation.
+///
+/// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
+/// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
+/// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
+/// try to generically lower a vector shuffle through such an pattern. It
+/// does not check for the profitability of lowering either as PALIGNR or
+/// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
+/// This matches shuffle vectors that look like:
+///
+/// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
+///
+/// Essentially it concatenates V1 and V2, shifts right by some number of
+/// elements, and takes the low elements as the result. Note that while this is
+/// specified as a *right shift* because x86 is little-endian, it is a *left
+/// rotate* of the vector lanes.
+static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
+ SDValue V2,
+ ArrayRef<int> Mask,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
- if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
- return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
- } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
- // Try to match an SSSE3 integer HADD/HSUB.
- if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
- return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
+ int NumElts = Mask.size();
+ int NumLanes = VT.getSizeInBits() / 128;
+ int NumLaneElts = NumElts / NumLanes;
- if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
- return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
- }
-
- if (!Subtarget->hasAVX())
- return SDValue();
-
- if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
- // Try to match an AVX horizontal add/sub of packed single/double
- // precision floating point values from 256-bit vectors.
- SDValue InVec2, InVec3;
- if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
- isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
- ((InVec0.getOpcode() == ISD::UNDEF ||
- InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
- ((InVec1.getOpcode() == ISD::UNDEF ||
- InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
- return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
+ // We need to detect various ways of spelling a rotation:
+ // [11, 12, 13, 14, 15, 0, 1, 2]
+ // [-1, 12, 13, 14, -1, -1, 1, -1]
+ // [-1, -1, -1, -1, -1, -1, 1, 2]
+ // [ 3, 4, 5, 6, 7, 8, 9, 10]
+ // [-1, 4, 5, 6, -1, -1, 9, -1]
+ // [-1, 4, 5, 6, -1, -1, -1, -1]
+ int Rotation = 0;
+ SDValue Lo, Hi;
+ for (int l = 0; l < NumElts; l += NumLaneElts) {
+ for (int i = 0; i < NumLaneElts; ++i) {
+ if (Mask[l + i] == -1)
+ continue;
+ assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
- if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
- isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
- ((InVec0.getOpcode() == ISD::UNDEF ||
- InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
- ((InVec1.getOpcode() == ISD::UNDEF ||
- InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
- return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
- } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
- // Try to match an AVX2 horizontal add/sub of signed integers.
- SDValue InVec2, InVec3;
- unsigned X86Opcode;
- bool CanFold = true;
+ // Get the mod-Size index and lane correct it.
+ int LaneIdx = (Mask[l + i] % NumElts) - l;
+ // Make sure it was in this lane.
+ if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
+ return SDValue();
- if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
- isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
- ((InVec0.getOpcode() == ISD::UNDEF ||
- InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
- ((InVec1.getOpcode() == ISD::UNDEF ||
- InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
- X86Opcode = X86ISD::HADD;
- else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
- isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
- ((InVec0.getOpcode() == ISD::UNDEF ||
- InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
- ((InVec1.getOpcode() == ISD::UNDEF ||
- InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
- X86Opcode = X86ISD::HSUB;
- else
- CanFold = false;
+ // Determine where a rotated vector would have started.
+ int StartIdx = i - LaneIdx;
+ if (StartIdx == 0)
+ // The identity rotation isn't interesting, stop.
+ return SDValue();
- if (CanFold) {
- // Fold this build_vector into a single horizontal add/sub.
- // Do this only if the target has AVX2.
- if (Subtarget->hasAVX2())
- return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
+ // If we found the tail of a vector the rotation must be the missing
+ // front. If we found the head of a vector, it must be how much of the
+ // head.
+ int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
- // Do not try to expand this build_vector into a pair of horizontal
- // add/sub if we can emit a pair of scalar add/sub.
- if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
+ if (Rotation == 0)
+ Rotation = CandidateRotation;
+ else if (Rotation != CandidateRotation)
+ // The rotations don't match, so we can't match this mask.
return SDValue();
- // Convert this build_vector into a pair of horizontal binop followed by
- // a concat vector.
- bool isUndefLO = NumUndefsLO == Half;
- bool isUndefHI = NumUndefsHI == Half;
- return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
- isUndefLO, isUndefHI);
+ // Compute which value this mask is pointing at.
+ SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
+
+ // Compute which of the two target values this index should be assigned
+ // to. This reflects whether the high elements are remaining or the low
+ // elements are remaining.
+ SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
+
+ // Either set up this value if we've not encountered it before, or check
+ // that it remains consistent.
+ if (!TargetV)
+ TargetV = MaskV;
+ else if (TargetV != MaskV)
+ // This may be a rotation, but it pulls from the inputs in some
+ // unsupported interleaving.
+ return SDValue();
}
}
- if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
- VT == MVT::v16i16) && Subtarget->hasAVX()) {
- unsigned X86Opcode;
- if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
- X86Opcode = X86ISD::HADD;
- else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
- X86Opcode = X86ISD::HSUB;
- else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
- X86Opcode = X86ISD::FHADD;
- else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
- X86Opcode = X86ISD::FHSUB;
- else
- return SDValue();
+ // Check that we successfully analyzed the mask, and normalize the results.
+ assert(Rotation != 0 && "Failed to locate a viable rotation!");
+ assert((Lo || Hi) && "Failed to find a rotated input vector!");
+ if (!Lo)
+ Lo = Hi;
+ else if (!Hi)
+ Hi = Lo;
- // Don't try to expand this build_vector into a pair of horizontal add/sub
- // if we can simply emit a pair of scalar add/sub.
- if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
- return SDValue();
+ // The actual rotate instruction rotates bytes, so we need to scale the
+ // rotation based on how many bytes are in the vector lane.
+ int Scale = 16 / NumLaneElts;
- // Convert this build_vector into two horizontal add/sub followed by
- // a concat vector.
- bool isUndefLO = NumUndefsLO == Half;
- bool isUndefHI = NumUndefsHI == Half;
- return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
- isUndefLO, isUndefHI);
+ // SSSE3 targets can use the palignr instruction.
+ if (Subtarget->hasSSSE3()) {
+ // Cast the inputs to i8 vector of correct length to match PALIGNR.
+ MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
+ Lo = DAG.getNode(ISD::BITCAST, DL, AlignVT, Lo);
+ Hi = DAG.getNode(ISD::BITCAST, DL, AlignVT, Hi);
+
+ return DAG.getNode(ISD::BITCAST, DL, VT,
+ DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Hi, Lo,
+ DAG.getConstant(Rotation * Scale, MVT::i8)));
}
- return SDValue();
-}
+ assert(VT.getSizeInBits() == 128 &&
+ "Rotate-based lowering only supports 128-bit lowering!");
+ assert(Mask.size() <= 16 &&
+ "Can shuffle at most 16 bytes in a 128-bit vector!");
-SDValue
-X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
- SDLoc dl(Op);
+ // Default SSE2 implementation
+ int LoByteShift = 16 - Rotation * Scale;
+ int HiByteShift = Rotation * Scale;
- MVT VT = Op.getSimpleValueType();
- MVT ExtVT = VT.getVectorElementType();
- unsigned NumElems = Op.getNumOperands();
+ // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
+ Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Lo);
+ Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Hi);
- // Generate vectors for predicate vectors.
- if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
- return LowerBUILD_VECTORvXi1(Op, DAG);
+ SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
+ DAG.getConstant(LoByteShift, MVT::i8));
+ SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
+ DAG.getConstant(HiByteShift, MVT::i8));
+ return DAG.getNode(ISD::BITCAST, DL, VT,
+ DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
+}
- // Vectors containing all zeros can be matched by pxor and xorps later
- if (ISD::isBuildVectorAllZeros(Op.getNode())) {
- // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
- // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
- if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
- return Op;
+/// \brief Compute whether each element of a shuffle is zeroable.
+///
+/// A "zeroable" vector shuffle element is one which can be lowered to zero.
+/// Either it is an undef element in the shuffle mask, the element of the input
+/// referenced is undef, or the element of the input referenced is known to be
+/// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
+/// as many lanes with this technique as possible to simplify the remaining
+/// shuffle.
+static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
+ SDValue V1, SDValue V2) {
+ SmallBitVector Zeroable(Mask.size(), false);
- return getZeroVector(VT, Subtarget, DAG, dl);
- }
+ while (V1.getOpcode() == ISD::BITCAST)
+ V1 = V1->getOperand(0);
+ while (V2.getOpcode() == ISD::BITCAST)
+ V2 = V2->getOperand(0);
- // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
- // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
- // vpcmpeqd on 256-bit vectors.
- if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
- if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
- return Op;
+ bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
+ bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
- if (!VT.is512BitVector())
- return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
- }
+ for (int i = 0, Size = Mask.size(); i < Size; ++i) {
+ int M = Mask[i];
+ // Handle the easy cases.
+ if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
+ Zeroable[i] = true;
+ continue;
+ }
- SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG);
- if (Broadcast.getNode())
- return Broadcast;
+ // If this is an index into a build_vector node (which has the same number
+ // of elements), dig out the input value and use it.
+ SDValue V = M < Size ? V1 : V2;
+ if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
+ continue;
- unsigned EVTBits = ExtVT.getSizeInBits();
+ SDValue Input = V.getOperand(M % Size);
+ // The UNDEF opcode check really should be dead code here, but not quite
+ // worth asserting on (it isn't invalid, just unexpected).
+ if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
+ Zeroable[i] = true;
+ }
- unsigned NumZero = 0;
- unsigned NumNonZero = 0;
- unsigned NonZeros = 0;
- bool IsAllConstants = true;
- SmallSet<SDValue, 8> Values;
- for (unsigned i = 0; i < NumElems; ++i) {
- SDValue Elt = Op.getOperand(i);
- if (Elt.getOpcode() == ISD::UNDEF)
+ return Zeroable;
+}
+
+/// \brief Try to emit a bitmask instruction for a shuffle.
+///
+/// This handles cases where we can model a blend exactly as a bitmask due to
+/// one of the inputs being zeroable.
+static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
+ SDValue V2, ArrayRef<int> Mask,
+ SelectionDAG &DAG) {
+ MVT EltVT = VT.getScalarType();
+ int NumEltBits = EltVT.getSizeInBits();
+ MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
+ SDValue Zero = DAG.getConstant(0, IntEltVT);
+ SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), IntEltVT);
+ if (EltVT.isFloatingPoint()) {
+ Zero = DAG.getNode(ISD::BITCAST, DL, EltVT, Zero);
+ AllOnes = DAG.getNode(ISD::BITCAST, DL, EltVT, AllOnes);
+ }
+ SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
+ SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
+ SDValue V;
+ for (int i = 0, Size = Mask.size(); i < Size; ++i) {
+ if (Zeroable[i])
continue;
- Values.insert(Elt);
- if (Elt.getOpcode() != ISD::Constant &&
- Elt.getOpcode() != ISD::ConstantFP)
- IsAllConstants = false;
- if (X86::isZeroNode(Elt))
- NumZero++;
- else {
- NonZeros |= (1 << i);
- NumNonZero++;
- }
- }
+ if (Mask[i] % Size != i)
+ return SDValue(); // Not a blend.
+ if (!V)
+ V = Mask[i] < Size ? V1 : V2;
+ else if (V != (Mask[i] < Size ? V1 : V2))
+ return SDValue(); // Can only let one input through the mask.
+
+ VMaskOps[i] = AllOnes;
+ }
+ if (!V)
+ return SDValue(); // No non-zeroable elements!
+
+ SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
+ V = DAG.getNode(VT.isFloatingPoint()
+ ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
+ DL, VT, V, VMask);
+ return V;
+}
- // All undef vector. Return an UNDEF. All zero vectors were handled above.
- if (NumNonZero == 0)
- return DAG.getUNDEF(VT);
+/// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
+///
+/// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
+/// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
+/// matches elements from one of the input vectors shuffled to the left or
+/// right with zeroable elements 'shifted in'. It handles both the strictly
+/// bit-wise element shifts and the byte shift across an entire 128-bit double
+/// quad word lane.
+///
+/// PSHL : (little-endian) left bit shift.
+/// [ zz, 0, zz, 2 ]
+/// [ -1, 4, zz, -1 ]
+/// PSRL : (little-endian) right bit shift.
+/// [ 1, zz, 3, zz]
+/// [ -1, -1, 7, zz]
+/// PSLLDQ : (little-endian) left byte shift
+/// [ zz, 0, 1, 2, 3, 4, 5, 6]
+/// [ zz, zz, -1, -1, 2, 3, 4, -1]
+/// [ zz, zz, zz, zz, zz, zz, -1, 1]
+/// PSRLDQ : (little-endian) right byte shift
+/// [ 5, 6, 7, zz, zz, zz, zz, zz]
+/// [ -1, 5, 6, 7, zz, zz, zz, zz]
+/// [ 1, 2, -1, -1, -1, -1, zz, zz]
+static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
+ SDValue V2, ArrayRef<int> Mask,
+ SelectionDAG &DAG) {
+ SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
- // Special case for single non-zero, non-undef, element.
- if (NumNonZero == 1) {
- unsigned Idx = countTrailingZeros(NonZeros);
- SDValue Item = Op.getOperand(Idx);
+ int Size = Mask.size();
+ assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
- // If this is an insertion of an i64 value on x86-32, and if the top bits of
- // the value are obviously zero, truncate the value to i32 and do the
- // insertion that way. Only do this if the value is non-constant or if the
- // value is a constant being inserted into element 0. It is cheaper to do
- // a constant pool load than it is to do a movd + shuffle.
- if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
- (!IsAllConstants || Idx == 0)) {
- if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
- // Handle SSE only.
- assert(VT == MVT::v2i64 && "Expected an SSE value type!");
- EVT VecVT = MVT::v4i32;
- unsigned VecElts = 4;
+ auto CheckZeros = [&](int Shift, int Scale, bool Left) {
+ for (int i = 0; i < Size; i += Scale)
+ for (int j = 0; j < Shift; ++j)
+ if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
+ return false;
- // Truncate the value (which may itself be a constant) to i32, and
- // 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);
+ return true;
+ };
- // 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
- // a vector. If Idx != 0, swizzle it into place.
- if (Idx != 0) {
- SmallVector<int, 4> Mask;
- Mask.push_back(Idx);
- for (unsigned i = 1; i != VecElts; ++i)
- Mask.push_back(i);
- Item = DAG.getVectorShuffle(VecVT, dl, Item, DAG.getUNDEF(VecVT),
- &Mask[0]);
- }
- return DAG.getNode(ISD::BITCAST, dl, VT, Item);
- }
+ auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
+ for (int i = 0; i != Size; i += Scale) {
+ unsigned Pos = Left ? i + Shift : i;
+ unsigned Low = Left ? i : i + Shift;
+ unsigned Len = Scale - Shift;
+ if (!isSequentialOrUndefInRange(Mask, Pos, Len,
+ Low + (V == V1 ? 0 : Size)))
+ return SDValue();
}
- // If we have a constant or non-constant insertion into the low element of
- // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
- // the rest of the elements. This will be matched as movd/movq/movss/movsd
- // depending on what the source datatype is.
- if (Idx == 0) {
- if (NumZero == 0)
- return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
+ int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
+ bool ByteShift = ShiftEltBits > 64;
+ unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
+ : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
+ int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
- if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
- (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
- if (VT.is256BitVector() || VT.is512BitVector()) {
- SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
- return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
- Item, DAG.getIntPtrConstant(0));
- }
- assert(VT.is128BitVector() && "Expected an SSE value type!");
- Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
- // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
- return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
- }
+ // Normalize the scale for byte shifts to still produce an i64 element
+ // type.
+ Scale = ByteShift ? Scale / 2 : Scale;
- if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
- Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
- Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
- if (VT.is256BitVector()) {
- SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
- Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
- } else {
- assert(VT.is128BitVector() && "Expected an SSE value type!");
- Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
- }
- return DAG.getNode(ISD::BITCAST, dl, VT, Item);
- }
- }
+ // We need to round trip through the appropriate type for the shift.
+ MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
+ MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
+ assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
+ "Illegal integer vector type");
+ V = DAG.getNode(ISD::BITCAST, DL, ShiftVT, V);
- // Is it a vector logical left shift?
- if (NumElems == 2 && Idx == 1 &&
- X86::isZeroNode(Op.getOperand(0)) &&
- !X86::isZeroNode(Op.getOperand(1))) {
- unsigned NumBits = VT.getSizeInBits();
- return getVShift(true, VT,
- DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
- VT, Op.getOperand(1)),
- NumBits/2, DAG, *this, dl);
- }
+ V = DAG.getNode(OpCode, DL, ShiftVT, V, DAG.getConstant(ShiftAmt, MVT::i8));
+ return DAG.getNode(ISD::BITCAST, DL, VT, V);
+ };
- if (IsAllConstants) // Otherwise, it's better to do a constpool load.
- return SDValue();
+ // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
+ // keep doubling the size of the integer elements up to that. We can
+ // then shift the elements of the integer vector by whole multiples of
+ // their width within the elements of the larger integer vector. Test each
+ // multiple to see if we can find a match with the moved element indices
+ // and that the shifted in elements are all zeroable.
+ for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
+ for (int Shift = 1; Shift != Scale; ++Shift)
+ for (bool Left : {true, false})
+ if (CheckZeros(Shift, Scale, Left))
+ for (SDValue V : {V1, V2})
+ if (SDValue Match = MatchShift(Shift, Scale, Left, V))
+ return Match;
+
+ // no match
+ return SDValue();
+}
- // Otherwise, if this is a vector with i32 or f32 elements, and the element
- // is a non-constant being inserted into an element other than the low one,
- // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
- // movd/movss) to move this into the low element, then shuffle it into
- // place.
- if (EVTBits == 32) {
- Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
+/// \brief Lower a vector shuffle as a zero or any extension.
+///
+/// Given a specific number of elements, element bit width, and extension
+/// stride, produce either a zero or any extension based on the available
+/// features of the subtarget.
+static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
+ SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
+ const X86Subtarget *Subtarget, SelectionDAG &DAG) {
+ assert(Scale > 1 && "Need a scale to extend.");
+ int NumElements = VT.getVectorNumElements();
+ int EltBits = VT.getScalarSizeInBits();
+ assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
+ "Only 8, 16, and 32 bit elements can be extended.");
+ assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
- // If using the new shuffle lowering, just directly insert this.
- if (ExperimentalVectorShuffleLowering)
- return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
+ // Found a valid zext mask! Try various lowering strategies based on the
+ // input type and available ISA extensions.
+ if (Subtarget->hasSSE41()) {
+ MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
+ NumElements / Scale);
+ return DAG.getNode(ISD::BITCAST, DL, VT,
+ DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
+ }
- // 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;
- for (unsigned i = 0; i != NumElems; ++i)
- MaskVec.push_back(i == Idx ? 0 : 1);
- return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
- }
+ // For any extends we can cheat for larger element sizes and use shuffle
+ // instructions that can fold with a load and/or copy.
+ if (AnyExt && EltBits == 32) {
+ int PSHUFDMask[4] = {0, -1, 1, -1};
+ return DAG.getNode(
+ ISD::BITCAST, DL, VT,
+ DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
+ DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
+ getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
+ }
+ if (AnyExt && EltBits == 16 && Scale > 2) {
+ int PSHUFDMask[4] = {0, -1, 0, -1};
+ InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
+ DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
+ getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG));
+ int PSHUFHWMask[4] = {1, -1, -1, -1};
+ return DAG.getNode(
+ ISD::BITCAST, DL, VT,
+ DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
+ DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, InputV),
+ getV4X86ShuffleImm8ForMask(PSHUFHWMask, DAG)));
}
- // Splat is obviously ok. Let legalizer expand it to a shuffle.
- if (Values.size() == 1) {
- if (EVTBits == 32) {
- // Instead of a shuffle like this:
- // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
- // Check if it's possible to issue this instead.
- // shuffle (vload ptr)), undef, <1, 1, 1, 1>
- unsigned Idx = countTrailingZeros(NonZeros);
- SDValue Item = Op.getOperand(Idx);
- if (Op.getNode()->isOnlyUserOf(Item.getNode()))
- return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
- }
- return SDValue();
+ // If this would require more than 2 unpack instructions to expand, use
+ // pshufb when available. We can only use more than 2 unpack instructions
+ // when zero extending i8 elements which also makes it easier to use pshufb.
+ if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
+ assert(NumElements == 16 && "Unexpected byte vector width!");
+ SDValue PSHUFBMask[16];
+ for (int i = 0; i < 16; ++i)
+ PSHUFBMask[i] =
+ DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, MVT::i8);
+ InputV = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, InputV);
+ return DAG.getNode(ISD::BITCAST, DL, VT,
+ DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
+ DAG.getNode(ISD::BUILD_VECTOR, DL,
+ MVT::v16i8, PSHUFBMask)));
}
- // A vector full of immediates; various special cases are already
- // handled, so this is best done with a single constant-pool load.
- if (IsAllConstants)
- return SDValue();
+ // Otherwise emit a sequence of unpacks.
+ do {
+ MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
+ SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
+ : getZeroVector(InputVT, Subtarget, DAG, DL);
+ InputV = DAG.getNode(ISD::BITCAST, DL, InputVT, InputV);
+ InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
+ Scale /= 2;
+ EltBits *= 2;
+ NumElements /= 2;
+ } while (Scale > 1);
+ return DAG.getNode(ISD::BITCAST, DL, VT, InputV);
+}
- // For AVX-length vectors, see if we can use a vector load to get all of the
- // elements, otherwise build the individual 128-bit pieces and use
- // shuffles to put them in place.
- if (VT.is256BitVector() || VT.is512BitVector()) {
- SmallVector<SDValue, 64> V;
- for (unsigned i = 0; i != NumElems; ++i)
- V.push_back(Op.getOperand(i));
+/// \brief Try to lower a vector shuffle as a zero extension on any microarch.
+///
+/// This routine will try to do everything in its power to cleverly lower
+/// a shuffle which happens to match the pattern of a zero extend. It doesn't
+/// check for the profitability of this lowering, it tries to aggressively
+/// match this pattern. It will use all of the micro-architectural details it
+/// can to emit an efficient lowering. It handles both blends with all-zero
+/// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
+/// masking out later).
+///
+/// The reason we have dedicated lowering for zext-style shuffles is that they
+/// are both incredibly common and often quite performance sensitive.
+static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
+ SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
+ const X86Subtarget *Subtarget, SelectionDAG &DAG) {
+ SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
- // Check for a build vector of consecutive loads.
- if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
- return LD;
+ int Bits = VT.getSizeInBits();
+ int NumElements = VT.getVectorNumElements();
+ assert(VT.getScalarSizeInBits() <= 32 &&
+ "Exceeds 32-bit integer zero extension limit");
+ assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
- EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
+ // Define a helper function to check a particular ext-scale and lower to it if
+ // valid.
+ auto Lower = [&](int Scale) -> SDValue {
+ SDValue InputV;
+ bool AnyExt = true;
+ for (int i = 0; i < NumElements; ++i) {
+ if (Mask[i] == -1)
+ continue; // Valid anywhere but doesn't tell us anything.
+ if (i % Scale != 0) {
+ // Each of the extended elements need to be zeroable.
+ if (!Zeroable[i])
+ return SDValue();
- // Build both the lower and upper subvector.
- SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
- makeArrayRef(&V[0], NumElems/2));
- SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
- makeArrayRef(&V[NumElems / 2], NumElems/2));
+ // We no longer are in the anyext case.
+ AnyExt = false;
+ continue;
+ }
- // Recreate the wider vector with the lower and upper part.
- if (VT.is256BitVector())
- return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
- return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
- }
+ // Each of the base elements needs to be consecutive indices into the
+ // same input vector.
+ SDValue V = Mask[i] < NumElements ? V1 : V2;
+ if (!InputV)
+ InputV = V;
+ else if (InputV != V)
+ return SDValue(); // Flip-flopping inputs.
- // Let legalizer expand 2-wide build_vectors.
- if (EVTBits == 64) {
- if (NumNonZero == 1) {
- // One half is zero or undef.
- unsigned Idx = countTrailingZeros(NonZeros);
- SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
- Op.getOperand(Idx));
- return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
+ if (Mask[i] % NumElements != i / Scale)
+ return SDValue(); // Non-consecutive strided elements.
}
- return SDValue();
- }
- // If element VT is < 32 bits, convert it to inserts into a zero vector.
- if (EVTBits == 8 && NumElems == 16) {
- SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
- Subtarget, *this);
- if (V.getNode()) return V;
- }
+ // If we fail to find an input, we have a zero-shuffle which should always
+ // have already been handled.
+ // FIXME: Maybe handle this here in case during blending we end up with one?
+ if (!InputV)
+ return SDValue();
- if (EVTBits == 16 && NumElems == 8) {
- SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
- Subtarget, *this);
- if (V.getNode()) return V;
- }
+ return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
+ DL, VT, Scale, AnyExt, InputV, Subtarget, DAG);
+ };
- // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
- if (EVTBits == 32 && NumElems == 4) {
- SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this);
- if (V.getNode())
+ // The widest scale possible for extending is to a 64-bit integer.
+ assert(Bits % 64 == 0 &&
+ "The number of bits in a vector must be divisible by 64 on x86!");
+ int NumExtElements = Bits / 64;
+
+ // Each iteration, try extending the elements half as much, but into twice as
+ // many elements.
+ for (; NumExtElements < NumElements; NumExtElements *= 2) {
+ assert(NumElements % NumExtElements == 0 &&
+ "The input vector size must be divisible by the extended size.");
+ if (SDValue V = Lower(NumElements / NumExtElements))
return V;
}
- // If element VT is == 32 bits, turn it into a number of shuffles.
- SmallVector<SDValue, 8> V(NumElems);
- if (NumElems == 4 && NumZero > 0) {
- for (unsigned i = 0; i < 4; ++i) {
- bool isZero = !(NonZeros & (1 << i));
- if (isZero)
- V[i] = getZeroVector(VT, Subtarget, DAG, dl);
- else
- V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
- }
+ // General extends failed, but 128-bit vectors may be able to use MOVQ.
+ if (Bits != 128)
+ return SDValue();
- for (unsigned i = 0; i < 2; ++i) {
- switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
- default: break;
- case 0:
- V[i] = V[i*2]; // Must be a zero vector.
- break;
- case 1:
- V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
- break;
- case 2:
- V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
- break;
- case 3:
- V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
- break;
- }
- }
+ // Returns one of the source operands if the shuffle can be reduced to a
+ // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
+ auto CanZExtLowHalf = [&]() {
+ for (int i = NumElements / 2; i != NumElements; ++i)
+ if (!Zeroable[i])
+ return SDValue();
+ if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
+ return V1;
+ if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
+ return V2;
+ return SDValue();
+ };
- bool Reverse1 = (NonZeros & 0x3) == 2;
- bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
- int MaskVec[] = {
- Reverse1 ? 1 : 0,
- Reverse1 ? 0 : 1,
- static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
- static_cast<int>(Reverse2 ? NumElems : NumElems+1)
- };
- return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
+ if (SDValue V = CanZExtLowHalf()) {
+ V = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V);
+ V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
+ return DAG.getNode(ISD::BITCAST, DL, VT, V);
}
- if (Values.size() > 1 && VT.is128BitVector()) {
- // Check for a build vector of consecutive loads.
- for (unsigned i = 0; i < NumElems; ++i)
- V[i] = Op.getOperand(i);
+ // No viable ext lowering found.
+ return SDValue();
+}
- // Check for elements which are consecutive loads.
- SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false);
- if (LD.getNode())
- return LD;
+/// \brief Try to get a scalar value for a specific element of a vector.
+///
+/// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
+static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
+ SelectionDAG &DAG) {
+ MVT VT = V.getSimpleValueType();
+ MVT EltVT = VT.getVectorElementType();
+ while (V.getOpcode() == ISD::BITCAST)
+ V = V.getOperand(0);
+ // If the bitcasts shift the element size, we can't extract an equivalent
+ // element from it.
+ MVT NewVT = V.getSimpleValueType();
+ if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
+ return SDValue();
- // Check for a build vector from mostly shuffle plus few inserting.
- SDValue Sh = buildFromShuffleMostly(Op, DAG);
- if (Sh.getNode())
- return Sh;
+ if (V.getOpcode() == ISD::BUILD_VECTOR ||
+ (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR))
+ return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, V.getOperand(Idx));
- // For SSE 4.1, use insertps to put the high elements into the low element.
- if (getSubtarget()->hasSSE41()) {
- SDValue Result;
- if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
- Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
- else
- Result = DAG.getUNDEF(VT);
+ return SDValue();
+}
- for (unsigned i = 1; i < NumElems; ++i) {
- if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
- Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
- Op.getOperand(i), DAG.getIntPtrConstant(i));
- }
- return Result;
- }
+/// \brief Helper to test for a load that can be folded with x86 shuffles.
+///
+/// This is particularly important because the set of instructions varies
+/// significantly based on whether the operand is a load or not.
+static bool isShuffleFoldableLoad(SDValue V) {
+ while (V.getOpcode() == ISD::BITCAST)
+ V = V.getOperand(0);
- // Otherwise, expand into a number of unpckl*, start by extending each of
- // our (non-undef) elements to the full vector width with the element in the
- // bottom slot of the vector (which generates no code for SSE).
- for (unsigned i = 0; i < NumElems; ++i) {
- if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
- V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
- else
- V[i] = DAG.getUNDEF(VT);
+ return ISD::isNON_EXTLoad(V.getNode());
+}
+
+/// \brief Try to lower insertion of a single element into a zero vector.
+///
+/// This is a common pattern that we have especially efficient patterns to lower
+/// across all subtarget feature sets.
+static SDValue lowerVectorShuffleAsElementInsertion(
+ SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
+ const X86Subtarget *Subtarget, SelectionDAG &DAG) {
+ SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
+ MVT ExtVT = VT;
+ MVT EltVT = VT.getVectorElementType();
+
+ int V2Index = std::find_if(Mask.begin(), Mask.end(),
+ [&Mask](int M) { return M >= (int)Mask.size(); }) -
+ Mask.begin();
+ bool IsV1Zeroable = true;
+ for (int i = 0, Size = Mask.size(); i < Size; ++i)
+ if (i != V2Index && !Zeroable[i]) {
+ IsV1Zeroable = false;
+ break;
}
- // Next, we iteratively mix elements, e.g. for v4f32:
- // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
- // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
- // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
- unsigned EltStride = NumElems >> 1;
- while (EltStride != 0) {
- for (unsigned i = 0; i < EltStride; ++i) {
- // If V[i+EltStride] is undef and this is the first round of mixing,
- // then it is safe to just drop this shuffle: V[i] is already in the
- // right place, the one element (since it's the first round) being
- // inserted as undef can be dropped. This isn't safe for successive
- // rounds because they will permute elements within both vectors.
- if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
- EltStride == NumElems/2)
- continue;
+ // 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 (SDValue V2S = getScalarValueForVectorElement(
+ V2, Mask[V2Index] - Mask.size(), DAG)) {
+ // We need to zext the scalar if it is smaller than an i32.
+ V2S = DAG.getNode(ISD::BITCAST, DL, EltVT, V2S);
+ if (EltVT == MVT::i8 || EltVT == MVT::i16) {
+ // Using zext to expand a narrow element won't work for non-zero
+ // insertions.
+ if (!IsV1Zeroable)
+ return SDValue();
- V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
- }
- EltStride >>= 1;
+ // Zero-extend directly to i32.
+ ExtVT = MVT::v4i32;
+ V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
}
- return V[0];
+ V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
+ } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
+ EltVT == MVT::i16) {
+ // Either not inserting from the low element of the input or the input
+ // element size is too small to use VZEXT_MOVL to clear the high bits.
+ return SDValue();
}
- return SDValue();
-}
-// LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
-// to create 256-bit vectors from two other 128-bit ones.
-static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
- SDLoc dl(Op);
- MVT ResVT = Op.getSimpleValueType();
+ if (!IsV1Zeroable) {
+ // If V1 can't be treated as a zero vector we have fewer options to lower
+ // this. We can't support integer vectors or non-zero targets cheaply, and
+ // the V1 elements can't be permuted in any way.
+ assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
+ if (!VT.isFloatingPoint() || V2Index != 0)
+ return SDValue();
+ SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
+ V1Mask[V2Index] = -1;
+ if (!isNoopShuffleMask(V1Mask))
+ return SDValue();
+ // This is essentially a special case blend operation, but if we have
+ // general purpose blend operations, they are always faster. Bail and let
+ // the rest of the lowering handle these as blends.
+ if (Subtarget->hasSSE41())
+ return SDValue();
- assert((ResVT.is256BitVector() ||
- ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
+ // Otherwise, use MOVSD or MOVSS.
+ assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
+ "Only two types of floating point element types to handle!");
+ return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
+ ExtVT, V1, V2);
+ }
- SDValue V1 = Op.getOperand(0);
- SDValue V2 = Op.getOperand(1);
- unsigned NumElems = ResVT.getVectorNumElements();
- if(ResVT.is256BitVector())
- return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
+ // This lowering only works for the low element with floating point vectors.
+ if (VT.isFloatingPoint() && V2Index != 0)
+ return SDValue();
- if (Op.getNumOperands() == 4) {
- MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
- ResVT.getVectorNumElements()/2);
- SDValue V3 = Op.getOperand(2);
- SDValue V4 = Op.getOperand(3);
- return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
- Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
+ V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
+ 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.isFloatingPoint() || 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()/8,
+ DAG.getTargetLoweringInfo().getScalarShiftAmountTy(MVT::v2i64)));
+ V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
+ }
}
- return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
+ return V2;
}
-static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
- MVT LLVM_ATTRIBUTE_UNUSED VT = Op.getSimpleValueType();
- assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
- (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
- Op.getNumOperands() == 4)));
+/// \brief Try to lower broadcast of a single element.
+///
+/// For convenience, this code also bundles all of the subtarget feature set
+/// filtering. While a little annoying to re-dispatch on type here, there isn't
+/// a convenient way to factor it out.
+static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
+ ArrayRef<int> Mask,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ if (!Subtarget->hasAVX())
+ return SDValue();
+ if (VT.isInteger() && !Subtarget->hasAVX2())
+ return SDValue();
- // AVX can use the vinsertf128 instruction to create 256-bit vectors
- // from two other 128-bit ones.
+ // Check that the mask is a broadcast.
+ int BroadcastIdx = -1;
+ for (int M : Mask)
+ if (M >= 0 && BroadcastIdx == -1)
+ BroadcastIdx = M;
+ else if (M >= 0 && M != BroadcastIdx)
+ return SDValue();
- // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
- return LowerAVXCONCAT_VECTORS(Op, DAG);
-}
+ assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
+ "a sorted mask where the broadcast "
+ "comes from V1.");
+ // Go up the chain of (vector) values to try and find a scalar load that
+ // we can combine with the broadcast.
+ for (;;) {
+ switch (V.getOpcode()) {
+ case ISD::CONCAT_VECTORS: {
+ int OperandSize = Mask.size() / V.getNumOperands();
+ V = V.getOperand(BroadcastIdx / OperandSize);
+ BroadcastIdx %= OperandSize;
+ continue;
+ }
-//===----------------------------------------------------------------------===//
-// Vector shuffle lowering
-//
-// This is an experimental code path for lowering vector shuffles on x86. It is
-// designed to handle arbitrary vector shuffles and blends, gracefully
-// degrading performance as necessary. It works hard to recognize idiomatic
-// shuffles and lower them to optimal instruction patterns without leaving
-// a framework that allows reasonably efficient handling of all vector shuffle
-// patterns.
-//===----------------------------------------------------------------------===//
+ case ISD::INSERT_SUBVECTOR: {
+ SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
+ auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
+ if (!ConstantIdx)
+ break;
-/// \brief Tiny helper function to identify a no-op mask.
-///
-/// This is a somewhat boring predicate function. It checks whether the mask
-/// array input, which is assumed to be a single-input shuffle mask of the kind
-/// used by the X86 shuffle instructions (not a fully general
-/// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
-/// in-place shuffle are 'no-op's.
-static bool isNoopShuffleMask(ArrayRef<int> Mask) {
- for (int i = 0, Size = Mask.size(); i < Size; ++i)
- if (Mask[i] != -1 && Mask[i] != i)
- return false;
- return true;
-}
+ int BeginIdx = (int)ConstantIdx->getZExtValue();
+ int EndIdx =
+ BeginIdx + (int)VInner.getValueType().getVectorNumElements();
+ if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
+ BroadcastIdx -= BeginIdx;
+ V = VInner;
+ } else {
+ V = VOuter;
+ }
+ continue;
+ }
+ }
+ break;
+ }
-/// \brief Helper function to classify a mask as a single-input mask.
-///
-/// This isn't a generic single-input test because in the vector shuffle
-/// lowering we canonicalize single inputs to be the first input operand. This
-/// means we can more quickly test for a single input by only checking whether
-/// an input from the second operand exists. We also assume that the size of
-/// mask corresponds to the size of the input vectors which isn't true in the
-/// fully general case.
-static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
- for (int M : Mask)
- if (M >= (int)Mask.size())
- return false;
- return true;
-}
+ // Check if this is a broadcast of a scalar. We special case lowering
+ // for scalars so that we can more effectively fold with loads.
+ if (V.getOpcode() == ISD::BUILD_VECTOR ||
+ (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
+ V = V.getOperand(BroadcastIdx);
-/// \brief Test whether there are elements crossing 128-bit lanes in this
-/// shuffle mask.
-///
-/// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
-/// and we routinely test for these.
-static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
- int LaneSize = 128 / VT.getScalarSizeInBits();
- int Size = Mask.size();
- for (int i = 0; i < Size; ++i)
- if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
- return true;
- return false;
+ // If the scalar isn't a load we can't broadcast from it in AVX1, only with
+ // AVX2.
+ if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
+ return SDValue();
+ } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
+ // We can't broadcast from a vector register w/o AVX2, and we can only
+ // broadcast from the zero-element of a vector register.
+ return SDValue();
+ }
+
+ return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
}
-/// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
-///
-/// This checks a shuffle mask to see if it is performing the same
-/// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
-/// that it is also not lane-crossing. It may however involve a blend from the
-/// same lane of a second vector.
-///
-/// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
-/// non-trivial to compute in the face of undef lanes. The representation is
-/// *not* suitable for use with existing 128-bit shuffles as it will contain
-/// entries from both V1 and V2 inputs to the wider mask.
-static bool
-is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
- SmallVectorImpl<int> &RepeatedMask) {
- int LaneSize = 128 / VT.getScalarSizeInBits();
- RepeatedMask.resize(LaneSize, -1);
- int Size = Mask.size();
- for (int i = 0; i < Size; ++i) {
- if (Mask[i] < 0)
+// Check for whether we can use INSERTPS to perform the shuffle. 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. We can also
+// perform INSERTPS if a single V1 element is out of place and all V2
+// elements are zeroable.
+static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
+ ArrayRef<int> Mask,
+ SelectionDAG &DAG) {
+ assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
+ assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
+ assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
+
+ SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
+
+ unsigned ZMask = 0;
+ int V1DstIndex = -1;
+ int V2DstIndex = -1;
+ bool V1UsedInPlace = false;
+
+ for (int i = 0; i < 4; ++i) {
+ // Synthesize a zero mask from the zeroable elements (includes undefs).
+ if (Zeroable[i]) {
+ ZMask |= 1 << i;
continue;
- if ((Mask[i] % Size) / LaneSize != i / LaneSize)
- // This entry crosses lanes, so there is no way to model this shuffle.
- return false;
+ }
- // Ok, handle the in-lane shuffles by detecting if and when they repeat.
- if (RepeatedMask[i % LaneSize] == -1)
- // This is the first non-undef entry in this slot of a 128-bit lane.
- RepeatedMask[i % LaneSize] =
- Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
- else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
- // Found a mismatch with the repeated mask.
- return false;
+ // Flag if we use any V1 inputs in place.
+ if (i == Mask[i]) {
+ V1UsedInPlace = true;
+ continue;
+ }
+
+ // We can only insert a single non-zeroable element.
+ if (V1DstIndex != -1 || V2DstIndex != -1)
+ return SDValue();
+
+ if (Mask[i] < 4) {
+ // V1 input out of place for insertion.
+ V1DstIndex = i;
+ } else {
+ // V2 input for insertion.
+ V2DstIndex = i;
+ }
}
- 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 {
+ // Don't bother if we have no (non-zeroable) element for insertion.
+ if (V1DstIndex == -1 && V2DstIndex == -1)
+ return SDValue();
-/// \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!");
- if (Mask[i] != -1 && Mask[i] != *Args[i])
- return false;
+ // Determine element insertion src/dst indices. The src index is from the
+ // start of the inserted vector, not the start of the concatenated vector.
+ unsigned V2SrcIndex = 0;
+ if (V1DstIndex != -1) {
+ // If we have a V1 input out of place, we use V1 as the V2 element insertion
+ // and don't use the original V2 at all.
+ V2SrcIndex = Mask[V1DstIndex];
+ V2DstIndex = V1DstIndex;
+ V2 = V1;
+ } else {
+ V2SrcIndex = Mask[V2DstIndex] - 4;
}
- return true;
-}
-} // namespace
+ // If no V1 inputs are used in place, then the result is created only from
+ // the zero mask and the V2 insertion - so remove V1 dependency.
+ if (!V1UsedInPlace)
+ V1 = DAG.getUNDEF(MVT::v4f32);
-/// \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 = {};
+ unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
+ assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
-/// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
-///
-/// This helper function produces an 8-bit shuffle immediate corresponding to
-/// the ubiquitous shuffle encoding scheme used in x86 instructions for
-/// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
-/// example.
+ // Insert the V2 element into the desired position.
+ SDLoc DL(Op);
+ return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
+ DAG.getConstant(InsertPSMask, MVT::i8));
+}
+
+/// \brief Try to lower a shuffle as a permute of the inputs followed by an
+/// UNPCK instruction.
///
-/// NB: We rely heavily on "undef" masks preserving the input lane.
-static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask,
+/// This specifically targets cases where we end up with alternating between
+/// the two inputs, and so can permute them into something that feeds a single
+/// UNPCK instruction. Note that this routine only targets integer vectors
+/// because for floating point vectors we have a generalized SHUFPS lowering
+/// strategy that handles everything that doesn't *exactly* match an unpack,
+/// making this clever lowering unnecessary.
+static SDValue lowerVectorShuffleAsUnpack(SDLoc DL, MVT VT, SDValue V1,
+ SDValue V2, ArrayRef<int> Mask,
SelectionDAG &DAG) {
- assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
- assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
- assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
- assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
- assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
+ assert(!VT.isFloatingPoint() &&
+ "This routine only supports integer vectors.");
+ assert(!isSingleInputShuffleMask(Mask) &&
+ "This routine should only be used when blending two inputs.");
+ assert(Mask.size() >= 2 && "Single element masks are invalid.");
- unsigned Imm = 0;
- Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
- Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
- Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
- Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
- return DAG.getConstant(Imm, MVT::i8);
-}
+ int Size = Mask.size();
-/// \brief Try to emit a blend instruction for a shuffle.
-///
-/// This doesn't do any checks for the availability of instructions for blending
-/// these values. It relies on the availability of the X86ISD::BLENDI pattern to
-/// be matched in the backend with the type given. What it does check for is
-/// that the shuffle mask is in fact a blend.
-static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
- SDValue V2, ArrayRef<int> Mask,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
+ int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
+ return M >= 0 && M % Size < Size / 2;
+ });
+ int NumHiInputs = std::count_if(
+ Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
- unsigned BlendMask = 0;
- for (int i = 0, Size = Mask.size(); i < Size; ++i) {
- if (Mask[i] >= Size) {
- if (Mask[i] != i + Size)
- return SDValue(); // Shuffled V2 input!
- BlendMask |= 1u << i;
- continue;
- }
- if (Mask[i] >= 0 && Mask[i] != i)
- return SDValue(); // Shuffled V1 input!
- }
- switch (VT.SimpleTy) {
- case MVT::v2f64:
- case MVT::v4f32:
- case MVT::v4f64:
- case MVT::v8f32:
- return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
- DAG.getConstant(BlendMask, MVT::i8));
+ bool UnpackLo = NumLoInputs >= NumHiInputs;
- case MVT::v4i64:
- case MVT::v8i32:
- assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
- // FALLTHROUGH
- case MVT::v2i64:
- case MVT::v4i32:
- // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
- // that instruction.
- if (Subtarget->hasAVX2()) {
- // Scale the blend by the number of 32-bit dwords per element.
- int Scale = VT.getScalarSizeInBits() / 32;
- BlendMask = 0;
- for (int i = 0, Size = Mask.size(); i < Size; ++i)
- if (Mask[i] >= Size)
- for (int j = 0; j < Scale; ++j)
- BlendMask |= 1u << (i * Scale + j);
+ auto TryUnpack = [&](MVT UnpackVT, int Scale) {
+ SmallVector<int, 32> V1Mask(Mask.size(), -1);
+ SmallVector<int, 32> V2Mask(Mask.size(), -1);
- MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
- V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
- V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
- return DAG.getNode(ISD::BITCAST, DL, VT,
- DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
- DAG.getConstant(BlendMask, MVT::i8)));
+ for (int i = 0; i < Size; ++i) {
+ if (Mask[i] < 0)
+ continue;
+
+ // Each element of the unpack contains Scale elements from this mask.
+ int UnpackIdx = i / Scale;
+
+ // We only handle the case where V1 feeds the first slots of the unpack.
+ // We rely on canonicalization to ensure this is the case.
+ if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
+ return SDValue();
+
+ // Setup the mask for this input. The indexing is tricky as we have to
+ // handle the unpack stride.
+ SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
+ VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
+ Mask[i] % Size;
}
- // FALLTHROUGH
- case MVT::v8i16: {
- // For integer shuffles we need to expand the mask and cast the inputs to
- // v8i16s prior to blending.
- int Scale = 8 / VT.getVectorNumElements();
- BlendMask = 0;
- for (int i = 0, Size = Mask.size(); i < Size; ++i)
- if (Mask[i] >= Size)
- for (int j = 0; j < Scale; ++j)
- BlendMask |= 1u << (i * Scale + j);
- V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
- V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
+ // If we will have to shuffle both inputs to use the unpack, check whether
+ // we can just unpack first and shuffle the result. If so, skip this unpack.
+ if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
+ !isNoopShuffleMask(V2Mask))
+ return SDValue();
+
+ // Shuffle the inputs into place.
+ V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
+ V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
+
+ // Cast the inputs to the type we will use to unpack them.
+ V1 = DAG.getNode(ISD::BITCAST, DL, UnpackVT, V1);
+ V2 = DAG.getNode(ISD::BITCAST, DL, UnpackVT, V2);
+
+ // Unpack the inputs and cast the result back to the desired type.
return DAG.getNode(ISD::BITCAST, DL, VT,
- DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
- DAG.getConstant(BlendMask, MVT::i8)));
- }
+ DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH,
+ DL, UnpackVT, V1, V2));
+ };
- case MVT::v16i16: {
- assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
- SmallVector<int, 8> RepeatedMask;
- if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
- // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
- assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
- BlendMask = 0;
- for (int i = 0; i < 8; ++i)
- if (RepeatedMask[i] >= 16)
- BlendMask |= 1u << i;
- return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
- DAG.getConstant(BlendMask, MVT::i8));
- }
- }
- // FALLTHROUGH
- case MVT::v32i8: {
- assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
- // Scale the blend by the number of bytes per element.
- int Scale = VT.getScalarSizeInBits() / 8;
- assert(Mask.size() * Scale == 32 && "Not a 256-bit vector!");
+ // We try each unpack from the largest to the smallest to try and find one
+ // that fits this mask.
+ int OrigNumElements = VT.getVectorNumElements();
+ int OrigScalarSize = VT.getScalarSizeInBits();
+ for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
+ int Scale = ScalarSize / OrigScalarSize;
+ int NumElements = OrigNumElements / Scale;
+ MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
+ if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
+ return Unpack;
+ }
+
+ // If none of the unpack-rooted lowerings worked (or were profitable) try an
+ // initial unpack.
+ if (NumLoInputs == 0 || NumHiInputs == 0) {
+ assert((NumLoInputs > 0 || NumHiInputs > 0) &&
+ "We have to have *some* inputs!");
+ int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
+
+ // FIXME: We could consider the total complexity of the permute of each
+ // possible unpacking. Or at the least we should consider how many
+ // half-crossings are created.
+ // FIXME: We could consider commuting the unpacks.
+
+ SmallVector<int, 32> PermMask;
+ PermMask.assign(Size, -1);
+ for (int i = 0; i < Size; ++i) {
+ if (Mask[i] < 0)
+ continue;
- // Compute the VSELECT mask. Note that VSELECT is really confusing in the
- // mix of LLVM's code generator and the x86 backend. We tell the code
- // generator that boolean values in the elements of an x86 vector register
- // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
- // mapping a select to operand #1, and 'false' mapping to operand #2. The
- // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
- // of the element (the remaining are ignored) and 0 in that high bit would
- // mean operand #1 while 1 in the high bit would mean operand #2. So while
- // the LLVM model for boolean values in vector elements gets the relevant
- // bit set, it is set backwards and over constrained relative to x86's
- // actual model.
- SDValue VSELECTMask[32];
- for (int i = 0, Size = Mask.size(); i < Size; ++i)
- for (int j = 0; j < Scale; ++j)
- VSELECTMask[Scale * i + j] =
- Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
- : DAG.getConstant(Mask[i] < Size ? -1 : 0, MVT::i8);
+ assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
- V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1);
- V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V2);
- return DAG.getNode(
- ISD::BITCAST, DL, VT,
- DAG.getNode(ISD::VSELECT, DL, MVT::v32i8,
- DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, VSELECTMask),
- V1, V2));
+ PermMask[i] =
+ 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
+ }
+ return DAG.getVectorShuffle(
+ VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
+ DL, VT, V1, V2),
+ DAG.getUNDEF(VT), PermMask);
}
- default:
- llvm_unreachable("Not a supported integer vector type!");
- }
+ return SDValue();
}
-/// \brief Generic routine to lower a shuffle and blend as a decomposed set of
-/// unblended shuffles followed by an unshuffled blend.
+/// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
///
-/// This matches the extremely common pattern for handling combined
-/// shuffle+blend operations on newer X86 ISAs where we have very fast blend
-/// operations.
-static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
- SDValue V1,
- SDValue V2,
- ArrayRef<int> Mask,
- SelectionDAG &DAG) {
- // Shuffle the input elements into the desired positions in V1 and V2 and
- // blend them together.
- SmallVector<int, 32> V1Mask(Mask.size(), -1);
- SmallVector<int, 32> V2Mask(Mask.size(), -1);
- SmallVector<int, 32> BlendMask(Mask.size(), -1);
- for (int i = 0, Size = Mask.size(); i < Size; ++i)
- if (Mask[i] >= 0 && Mask[i] < Size) {
- V1Mask[i] = Mask[i];
- BlendMask[i] = i;
- } else if (Mask[i] >= Size) {
- V2Mask[i] = Mask[i] - Size;
- BlendMask[i] = i + Size;
+/// This is the basis function for the 2-lane 64-bit shuffles as we have full
+/// support for floating point shuffles but not integer shuffles. These
+/// instructions will incur a domain crossing penalty on some chips though so
+/// it is better to avoid lowering through this for integer vectors where
+/// possible.
+static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
+ assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
+
+ if (isSingleInputShuffleMask(Mask)) {
+ // Use low duplicate instructions for masks that match their pattern.
+ if (Subtarget->hasSSE3())
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
+ return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
+
+ // Straight shuffle of a single input vector. Simulate this by using the
+ // single input as both of the "inputs" to this instruction..
+ unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
+
+ if (Subtarget->hasAVX()) {
+ // If we have AVX, we can use VPERMILPS which will allow folding a load
+ // into the shuffle.
+ return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
+ DAG.getConstant(SHUFPDMask, MVT::i8));
}
- V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
- V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
- return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
-}
+ return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V1,
+ DAG.getConstant(SHUFPDMask, MVT::i8));
+ }
+ assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
+ assert(Mask[1] >= 2 && "Non-canonicalized blend!");
-/// \brief Try to lower a vector shuffle as a byte rotation.
-///
-/// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
-/// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
-/// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
-/// try to generically lower a vector shuffle through such an pattern. It
-/// does not check for the profitability of lowering either as PALIGNR or
-/// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
-/// This matches shuffle vectors that look like:
-///
-/// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
-///
-/// Essentially it concatenates V1 and V2, shifts right by some number of
-/// elements, and takes the low elements as the result. Note that while this is
-/// specified as a *right shift* because x86 is little-endian, it is a *left
-/// rotate* of the vector lanes.
-///
-/// Note that this only handles 128-bit vector widths currently.
-static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
- SDValue V2,
- ArrayRef<int> Mask,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
+ // If we have a single input, insert that into V1 if we can do so cheaply.
+ if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
+ if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
+ DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
+ return Insertion;
+ // Try inverting the insertion since for v2 masks it is easy to do and we
+ // can't reliably sort the mask one way or the other.
+ int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
+ Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
+ if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
+ DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
+ return Insertion;
+ }
- // We need to detect various ways of spelling a rotation:
- // [11, 12, 13, 14, 15, 0, 1, 2]
- // [-1, 12, 13, 14, -1, -1, 1, -1]
- // [-1, -1, -1, -1, -1, -1, 1, 2]
- // [ 3, 4, 5, 6, 7, 8, 9, 10]
- // [-1, 4, 5, 6, -1, -1, 9, -1]
- // [-1, 4, 5, 6, -1, -1, -1, -1]
- int Rotation = 0;
- SDValue Lo, Hi;
- for (int i = 0, Size = Mask.size(); i < Size; ++i) {
- if (Mask[i] == -1)
- continue;
- assert(Mask[i] >= 0 && "Only -1 is a valid negative mask element!");
+ // Try to use one of the special instruction patterns to handle two common
+ // blend patterns if a zero-blend above didn't work.
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
+ isShuffleEquivalent(V1, V2, Mask, {1, 3}))
+ if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
+ // We can either use a special instruction to load over the low double or
+ // to move just the low double.
+ return DAG.getNode(
+ isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
+ DL, MVT::v2f64, V2,
+ DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
- // Based on the mod-Size value of this mask element determine where
- // a rotated vector would have started.
- int StartIdx = i - (Mask[i] % Size);
- if (StartIdx == 0)
- // The identity rotation isn't interesting, stop.
- return SDValue();
+ if (Subtarget->hasSSE41())
+ if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
+ Subtarget, DAG))
+ return Blend;
- // If we found the tail of a vector the rotation must be the missing
- // front. If we found the head of a vector, it must be how much of the head.
- int CandidateRotation = StartIdx < 0 ? -StartIdx : Size - StartIdx;
+ // Use dedicated unpack instructions for masks that match their pattern.
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
- if (Rotation == 0)
- Rotation = CandidateRotation;
- else if (Rotation != CandidateRotation)
- // The rotations don't match, so we can't match this mask.
- return SDValue();
+ 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));
+}
- // Compute which value this mask is pointing at.
- SDValue MaskV = Mask[i] < Size ? V1 : V2;
-
- // Compute which of the two target values this index should be assigned to.
- // This reflects whether the high elements are remaining or the low elements
- // are remaining.
- SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
-
- // Either set up this value if we've not encountered it before, or check
- // that it remains consistent.
- if (!TargetV)
- TargetV = MaskV;
- else if (TargetV != MaskV)
- // This may be a rotation, but it pulls from the inputs in some
- // unsupported interleaving.
- return SDValue();
+/// \brief Handle lowering of 2-lane 64-bit integer shuffles.
+///
+/// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
+/// the integer unit to minimize domain crossing penalties. However, for blends
+/// it falls back to the floating point shuffle operation with appropriate bit
+/// casting.
+static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
+ assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
+
+ if (isSingleInputShuffleMask(Mask)) {
+ // Check for being able to broadcast a single element.
+ if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
+ Mask, Subtarget, DAG))
+ return Broadcast;
+
+ // Straight shuffle of a single input vector. For everything from SSE2
+ // onward this has a single fast instruction with no scary immediates.
+ // We have to map the mask as it is actually a v4i32 shuffle instruction.
+ V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V1);
+ int WidenedMask[4] = {
+ std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
+ std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
+ return DAG.getNode(
+ ISD::BITCAST, DL, MVT::v2i64,
+ DAG.getNode(X86ISD::PSHUFD, SDLoc(Op), MVT::v4i32, V1,
+ getV4X86ShuffleImm8ForMask(WidenedMask, DAG)));
}
+ assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
+ assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
+ assert(Mask[0] < 2 && "We sort V1 to be the first input.");
+ assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
+
+ // If we have a blend of two PACKUS operations an the blend aligns with the
+ // low and half halves, we can just merge the PACKUS operations. This is
+ // particularly important as it lets us merge shuffles that this routine itself
+ // creates.
+ auto GetPackNode = [](SDValue V) {
+ while (V.getOpcode() == ISD::BITCAST)
+ V = V.getOperand(0);
- // Check that we successfully analyzed the mask, and normalize the results.
- assert(Rotation != 0 && "Failed to locate a viable rotation!");
- assert((Lo || Hi) && "Failed to find a rotated input vector!");
- if (!Lo)
- Lo = Hi;
- else if (!Hi)
- Hi = Lo;
+ return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
+ };
+ if (SDValue V1Pack = GetPackNode(V1))
+ if (SDValue V2Pack = GetPackNode(V2))
+ return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
+ DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
+ Mask[0] == 0 ? V1Pack.getOperand(0)
+ : V1Pack.getOperand(1),
+ Mask[1] == 2 ? V2Pack.getOperand(0)
+ : V2Pack.getOperand(1)));
+
+ // Try to use shift instructions.
+ if (SDValue Shift =
+ lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
+ return Shift;
- assert(VT.getSizeInBits() == 128 &&
- "Rotate-based lowering only supports 128-bit lowering!");
- assert(Mask.size() <= 16 &&
- "Can shuffle at most 16 bytes in a 128-bit vector!");
+ // When loading a scalar and then shuffling it into a vector we can often do
+ // the insertion cheaply.
+ if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
+ DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
+ return Insertion;
+ // Try inverting the insertion since for v2 masks it is easy to do and we
+ // can't reliably sort the mask one way or the other.
+ int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
+ if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
+ DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
+ return Insertion;
+
+ // We have different paths for blend lowering, but they all must use the
+ // *exact* same predicate.
+ bool IsBlendSupported = Subtarget->hasSSE41();
+ if (IsBlendSupported)
+ if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
+ Subtarget, DAG))
+ return Blend;
- // The actual rotate instruction rotates bytes, so we need to scale the
- // rotation based on how many bytes are in the vector.
- int Scale = 16 / Mask.size();
+ // Use dedicated unpack instructions for masks that match their pattern.
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
- // SSSE3 targets can use the palignr instruction
- if (Subtarget->hasSSSE3()) {
- // Cast the inputs to v16i8 to match PALIGNR.
- Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Lo);
- Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Hi);
+ // Try to use byte rotation instructions.
+ // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
+ if (Subtarget->hasSSSE3())
+ if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
+ DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
+ return Rotate;
- return DAG.getNode(ISD::BITCAST, DL, VT,
- DAG.getNode(X86ISD::PALIGNR, DL, MVT::v16i8, Hi, Lo,
- DAG.getConstant(Rotation * Scale, MVT::i8)));
- }
+ // If we have direct support for blends, we should lower by decomposing into
+ // a permute. That will be faster than the domain cross.
+ if (IsBlendSupported)
+ return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
+ Mask, DAG);
- // Default SSE2 implementation
- int LoByteShift = 16 - Rotation * Scale;
- int HiByteShift = Rotation * Scale;
+ // 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
+ // have this problem. It would be really nice if x86 had better shuffles here.
+ V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V1);
+ V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V2);
+ return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
+ DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
+}
- // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
- Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Lo);
- Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Hi);
+/// \brief Test whether this can be lowered with a single SHUFPS instruction.
+///
+/// This is used to disable more specialized lowerings when the shufps lowering
+/// will happen to be efficient.
+static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
+ // This routine only handles 128-bit shufps.
+ assert(Mask.size() == 4 && "Unsupported mask size!");
+
+ // To lower with a single SHUFPS we need to have the low half and high half
+ // each requiring a single input.
+ if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
+ return false;
+ if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
+ return false;
- SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
- DAG.getConstant(8 * LoByteShift, MVT::i8));
- SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
- DAG.getConstant(8 * HiByteShift, MVT::i8));
- return DAG.getNode(ISD::BITCAST, DL, VT,
- DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
+ return true;
}
-/// \brief Compute whether each element of a shuffle is zeroable.
+/// \brief Lower a vector shuffle using the SHUFPS instruction.
///
-/// A "zeroable" vector shuffle element is one which can be lowered to zero.
-/// Either it is an undef element in the shuffle mask, the element of the input
-/// referenced is undef, or the element of the input referenced is known to be
-/// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
-/// as many lanes with this technique as possible to simplify the remaining
-/// shuffle.
-static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
- SDValue V1, SDValue V2) {
- SmallBitVector Zeroable(Mask.size(), false);
+/// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
+/// It makes no assumptions about whether this is the *best* lowering, it simply
+/// uses it.
+static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
+ ArrayRef<int> Mask, SDValue V1,
+ SDValue V2, SelectionDAG &DAG) {
+ SDValue LowV = V1, HighV = V2;
+ int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
- bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
- bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
+ int NumV2Elements =
+ std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
- for (int i = 0, Size = Mask.size(); i < Size; ++i) {
- int M = Mask[i];
- // Handle the easy cases.
- if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
- Zeroable[i] = true;
- continue;
+ if (NumV2Elements == 1) {
+ int V2Index =
+ std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
+ Mask.begin();
+
+ // Compute the index adjacent to V2Index and in the same half by toggling
+ // the low bit.
+ int V2AdjIndex = V2Index ^ 1;
+
+ if (Mask[V2AdjIndex] == -1) {
+ // Handles all the cases where we have a single V2 element and an undef.
+ // This will only ever happen in the high lanes because we commute the
+ // vector otherwise.
+ if (V2Index < 2)
+ std::swap(LowV, HighV);
+ NewMask[V2Index] -= 4;
+ } else {
+ // Handle the case where the V2 element ends up adjacent to a V1 element.
+ // To make this work, blend them together as the first step.
+ int V1Index = V2AdjIndex;
+ int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
+ V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
+ getV4X86ShuffleImm8ForMask(BlendMask, DAG));
+
+ // Now proceed to reconstruct the final blend as we have the necessary
+ // high or low half formed.
+ if (V2Index < 2) {
+ LowV = V2;
+ HighV = V1;
+ } else {
+ HighV = V2;
+ }
+ NewMask[V1Index] = 2; // We put the V1 element in V2[2].
+ NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
}
+ } else if (NumV2Elements == 2) {
+ if (Mask[0] < 4 && Mask[1] < 4) {
+ // Handle the easy case where we have V1 in the low lanes and V2 in the
+ // high lanes.
+ NewMask[2] -= 4;
+ NewMask[3] -= 4;
+ } else if (Mask[2] < 4 && Mask[3] < 4) {
+ // We also handle the reversed case because this utility may get called
+ // when we detect a SHUFPS pattern but can't easily commute the shuffle to
+ // arrange things in the right direction.
+ NewMask[0] -= 4;
+ NewMask[1] -= 4;
+ HighV = V1;
+ LowV = V2;
+ } else {
+ // We have a mixture of V1 and V2 in both low and high lanes. Rather than
+ // trying to place elements directly, just blend them and set up the final
+ // shuffle to place them.
- // If this is an index into a build_vector node, dig out the input value and
- // use it.
- SDValue V = M < Size ? V1 : V2;
- if (V.getOpcode() != ISD::BUILD_VECTOR)
- continue;
+ // The first two blend mask elements are for V1, the second two are for
+ // V2.
+ int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
+ Mask[2] < 4 ? Mask[2] : Mask[3],
+ (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
+ (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
+ V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
+ getV4X86ShuffleImm8ForMask(BlendMask, DAG));
- SDValue Input = V.getOperand(M % Size);
- // The UNDEF opcode check really should be dead code here, but not quite
- // worth asserting on (it isn't invalid, just unexpected).
- if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
- Zeroable[i] = true;
+ // Now we do a normal shuffle of V1 by giving V1 as both operands to
+ // a blend.
+ LowV = HighV = V1;
+ NewMask[0] = Mask[0] < 4 ? 0 : 2;
+ NewMask[1] = Mask[0] < 4 ? 2 : 0;
+ NewMask[2] = Mask[2] < 4 ? 1 : 3;
+ NewMask[3] = Mask[2] < 4 ? 3 : 1;
+ }
}
-
- return Zeroable;
+ return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
+ getV4X86ShuffleImm8ForMask(NewMask, DAG));
}
-/// \brief Try to lower a vector shuffle as a byte shift (shifts in zeros).
-///
-/// Attempts to match a shuffle mask against the PSRLDQ and PSLLDQ SSE2
-/// byte-shift instructions. The mask must consist of a shifted sequential
-/// shuffle from one of the input vectors and zeroable elements for the
-/// remaining 'shifted in' elements.
+/// \brief Lower 4-lane 32-bit floating point shuffles.
///
-/// Note that this only handles 128-bit vector widths currently.
-static SDValue lowerVectorShuffleAsByteShift(SDLoc DL, MVT VT, SDValue V1,
- SDValue V2, ArrayRef<int> Mask,
- SelectionDAG &DAG) {
- assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
+/// Uses instructions exclusively from the floating point unit to minimize
+/// domain crossing penalties, as these are sufficient to implement all v4f32
+/// shuffles.
+static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
+ assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
- SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
+ int NumV2Elements =
+ std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
- int Size = Mask.size();
- int Scale = 16 / Size;
+ if (NumV2Elements == 0) {
+ // Check for being able to broadcast a single element.
+ if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
+ Mask, Subtarget, DAG))
+ return Broadcast;
- for (int Shift = 1; Shift < Size; Shift++) {
- int ByteShift = Shift * Scale;
-
- // PSRLDQ : (little-endian) right byte shift
- // [ 5, 6, 7, zz, zz, zz, zz, zz]
- // [ -1, 5, 6, 7, zz, zz, zz, zz]
- // [ 1, 2, -1, -1, -1, -1, zz, zz]
- bool ZeroableRight = true;
- for (int i = Size - Shift; i < Size; i++) {
- ZeroableRight &= Zeroable[i];
- }
-
- if (ZeroableRight) {
- bool ValidShiftRight1 =
- isSequentialOrUndefInRange(Mask, 0, Size - Shift, Shift);
- bool ValidShiftRight2 =
- isSequentialOrUndefInRange(Mask, 0, Size - Shift, Size + Shift);
-
- if (ValidShiftRight1 || ValidShiftRight2) {
- // Cast the inputs to v2i64 to match PSRLDQ.
- SDValue &TargetV = ValidShiftRight1 ? V1 : V2;
- SDValue V = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, TargetV);
- SDValue Shifted = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, V,
- DAG.getConstant(ByteShift * 8, MVT::i8));
- return DAG.getNode(ISD::BITCAST, DL, VT, Shifted);
- }
+ // Use even/odd duplicate instructions for masks that match their pattern.
+ if (Subtarget->hasSSE3()) {
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
+ return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
+ if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
+ return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
}
- // PSLLDQ : (little-endian) left byte shift
- // [ zz, 0, 1, 2, 3, 4, 5, 6]
- // [ zz, zz, -1, -1, 2, 3, 4, -1]
- // [ zz, zz, zz, zz, zz, zz, -1, 1]
- bool ZeroableLeft = true;
- for (int i = 0; i < Shift; i++) {
- ZeroableLeft &= Zeroable[i];
- }
-
- if (ZeroableLeft) {
- bool ValidShiftLeft1 =
- isSequentialOrUndefInRange(Mask, Shift, Size - Shift, 0);
- bool ValidShiftLeft2 =
- isSequentialOrUndefInRange(Mask, Shift, Size - Shift, Size);
-
- if (ValidShiftLeft1 || ValidShiftLeft2) {
- // Cast the inputs to v2i64 to match PSLLDQ.
- SDValue &TargetV = ValidShiftLeft1 ? V1 : V2;
- SDValue V = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, TargetV);
- SDValue Shifted = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, V,
- DAG.getConstant(ByteShift * 8, MVT::i8));
- return DAG.getNode(ISD::BITCAST, DL, VT, Shifted);
- }
+ if (Subtarget->hasAVX()) {
+ // If we have AVX, we can use VPERMILPS which will allow folding a load
+ // into the shuffle.
+ return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
+ getV4X86ShuffleImm8ForMask(Mask, DAG));
}
- }
- return SDValue();
-}
+ // Otherwise, use a straight shuffle of a single input vector. We pass the
+ // input vector to both operands to simulate this with a SHUFPS.
+ return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
+ getV4X86ShuffleImm8ForMask(Mask, DAG));
+ }
-/// \brief Lower a vector shuffle as a zero or any extension.
-///
-/// Given a specific number of elements, element bit width, and extension
-/// stride, produce either a zero or any extension based on the available
-/// features of the subtarget.
-static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
- SDLoc DL, MVT VT, int NumElements, int Scale, bool AnyExt, SDValue InputV,
- const X86Subtarget *Subtarget, SelectionDAG &DAG) {
- assert(Scale > 1 && "Need a scale to extend.");
- int EltBits = VT.getSizeInBits() / NumElements;
- assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
- "Only 8, 16, and 32 bit elements can be extended.");
- assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
+ // There are special ways we can lower some single-element blends. However, we
+ // have custom ways we can lower more complex single-element blends below that
+ // we defer to if both this and BLENDPS fail to match, so restrict this to
+ // when the V2 input is targeting element 0 of the mask -- that is the fast
+ // case here.
+ if (NumV2Elements == 1 && Mask[0] >= 4)
+ if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
+ Mask, Subtarget, DAG))
+ return V;
- // Found a valid zext mask! Try various lowering strategies based on the
- // input type and available ISA extensions.
if (Subtarget->hasSSE41()) {
- MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
- MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
- NumElements / Scale);
- InputV = DAG.getNode(ISD::BITCAST, DL, InputVT, InputV);
- return DAG.getNode(ISD::BITCAST, DL, VT,
- DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
- }
+ if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
+ Subtarget, DAG))
+ return Blend;
- // For any extends we can cheat for larger element sizes and use shuffle
- // instructions that can fold with a load and/or copy.
- if (AnyExt && EltBits == 32) {
- int PSHUFDMask[4] = {0, -1, 1, -1};
- return DAG.getNode(
- ISD::BITCAST, DL, VT,
- DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
- DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
- getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
- }
- if (AnyExt && EltBits == 16 && Scale > 2) {
- int PSHUFDMask[4] = {0, -1, 0, -1};
- InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
- DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
- getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG));
- int PSHUFHWMask[4] = {1, -1, -1, -1};
- return DAG.getNode(
- ISD::BITCAST, DL, VT,
- DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
- DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, InputV),
- getV4X86ShuffleImm8ForMask(PSHUFHWMask, DAG)));
- }
+ // Use INSERTPS if we can complete the shuffle efficiently.
+ if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
+ return V;
- // If this would require more than 2 unpack instructions to expand, use
- // pshufb when available. We can only use more than 2 unpack instructions
- // when zero extending i8 elements which also makes it easier to use pshufb.
- if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
- assert(NumElements == 16 && "Unexpected byte vector width!");
- SDValue PSHUFBMask[16];
- for (int i = 0; i < 16; ++i)
- PSHUFBMask[i] =
- DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, MVT::i8);
- InputV = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, InputV);
- return DAG.getNode(ISD::BITCAST, DL, VT,
- DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
- DAG.getNode(ISD::BUILD_VECTOR, DL,
- MVT::v16i8, PSHUFBMask)));
+ if (!isSingleSHUFPSMask(Mask))
+ if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
+ DL, MVT::v4f32, V1, V2, Mask, DAG))
+ return BlendPerm;
}
- // Otherwise emit a sequence of unpacks.
- do {
- MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
- SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
- : getZeroVector(InputVT, Subtarget, DAG, DL);
- InputV = DAG.getNode(ISD::BITCAST, DL, InputVT, InputV);
- InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
- Scale /= 2;
- EltBits *= 2;
- NumElements /= 2;
- } while (Scale > 1);
- return DAG.getNode(ISD::BITCAST, DL, VT, InputV);
+ // Use dedicated unpack instructions for masks that match their pattern.
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V2, V1);
+ if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V2, V1);
+
+ // Otherwise fall back to a SHUFPS lowering strategy.
+ return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
}
-/// \brief Try to lower a vector shuffle as a zero extension on any micrarch.
-///
-/// This routine will try to do everything in its power to cleverly lower
-/// a shuffle which happens to match the pattern of a zero extend. It doesn't
-/// check for the profitability of this lowering, it tries to aggressively
-/// match this pattern. It will use all of the micro-architectural details it
-/// can to emit an efficient lowering. It handles both blends with all-zero
-/// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
-/// masking out later).
+/// \brief Lower 4-lane i32 vector shuffles.
///
-/// The reason we have dedicated lowering for zext-style shuffles is that they
-/// are both incredibly common and often quite performance sensitive.
-static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
- SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
- const X86Subtarget *Subtarget, SelectionDAG &DAG) {
- SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
-
- int Bits = VT.getSizeInBits();
- int NumElements = Mask.size();
-
- // Define a helper function to check a particular ext-scale and lower to it if
- // valid.
- auto Lower = [&](int Scale) -> SDValue {
- SDValue InputV;
- bool AnyExt = true;
- for (int i = 0; i < NumElements; ++i) {
- if (Mask[i] == -1)
- continue; // Valid anywhere but doesn't tell us anything.
- if (i % Scale != 0) {
- // Each of the extend elements needs to be zeroable.
- if (!Zeroable[i])
- return SDValue();
+/// We try to handle these with integer-domain shuffles where we can, but for
+/// blends we use the floating point domain blend instructions.
+static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
+ assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
- // We no lorger are in the anyext case.
- AnyExt = false;
- continue;
- }
+ // Whenever we can lower this as a zext, that instruction is strictly faster
+ // than any alternative. It also allows us to fold memory operands into the
+ // shuffle in many cases.
+ if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
+ Mask, Subtarget, DAG))
+ return ZExt;
- // Each of the base elements needs to be consecutive indices into the
- // same input vector.
- SDValue V = Mask[i] < NumElements ? V1 : V2;
- if (!InputV)
- InputV = V;
- else if (InputV != V)
- return SDValue(); // Flip-flopping inputs.
+ int NumV2Elements =
+ std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
- if (Mask[i] % NumElements != i / Scale)
- return SDValue(); // Non-consecutive strided elemenst.
- }
+ if (NumV2Elements == 0) {
+ // Check for being able to broadcast a single element.
+ if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
+ Mask, Subtarget, DAG))
+ return Broadcast;
- // If we fail to find an input, we have a zero-shuffle which should always
- // have already been handled.
- // FIXME: Maybe handle this here in case during blending we end up with one?
- if (!InputV)
- return SDValue();
+ // Straight shuffle of a single input vector. For everything from SSE2
+ // onward this has a single fast instruction with no scary immediates.
+ // We coerce the shuffle pattern to be compatible with UNPCK instructions
+ // but we aren't actually going to use the UNPCK instruction because doing
+ // so prevents folding a load into this instruction or making a copy.
+ const int UnpackLoMask[] = {0, 0, 1, 1};
+ const int UnpackHiMask[] = {2, 2, 3, 3};
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
+ Mask = UnpackLoMask;
+ else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
+ Mask = UnpackHiMask;
- return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
- DL, VT, NumElements, Scale, AnyExt, InputV, Subtarget, DAG);
- };
+ return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
+ getV4X86ShuffleImm8ForMask(Mask, DAG));
+ }
- // The widest scale possible for extending is to a 64-bit integer.
- assert(Bits % 64 == 0 &&
- "The number of bits in a vector must be divisible by 64 on x86!");
- int NumExtElements = Bits / 64;
+ // Try to use shift instructions.
+ if (SDValue Shift =
+ lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
+ return Shift;
- // Each iteration, try extending the elements half as much, but into twice as
- // many elements.
- for (; NumExtElements < NumElements; NumExtElements *= 2) {
- assert(NumElements % NumExtElements == 0 &&
- "The input vector size must be divisble by the extended size.");
- if (SDValue V = Lower(NumElements / NumExtElements))
+ // There are special ways we can lower some single-element blends.
+ if (NumV2Elements == 1)
+ if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
+ Mask, Subtarget, DAG))
return V;
- }
- // No viable ext lowering found.
- return SDValue();
-}
+ // We have different paths for blend lowering, but they all must use the
+ // *exact* same predicate.
+ bool IsBlendSupported = Subtarget->hasSSE41();
+ if (IsBlendSupported)
+ if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
+ Subtarget, DAG))
+ return Blend;
-/// \brief Try to get a scalar value for a specific element of a vector.
-///
-/// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
-static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
- SelectionDAG &DAG) {
- MVT VT = V.getSimpleValueType();
- MVT EltVT = VT.getVectorElementType();
- while (V.getOpcode() == ISD::BITCAST)
- V = V.getOperand(0);
- // If the bitcasts shift the element size, we can't extract an equivalent
- // element from it.
- MVT NewVT = V.getSimpleValueType();
- if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
- return SDValue();
+ if (SDValue Masked =
+ lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
+ return Masked;
- if (V.getOpcode() == ISD::BUILD_VECTOR ||
- (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR))
- return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, V.getOperand(Idx));
+ // Use dedicated unpack instructions for masks that match their pattern.
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V2, V1);
+ if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V2, V1);
- return SDValue();
-}
+ // Try to use byte rotation instructions.
+ // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
+ if (Subtarget->hasSSSE3())
+ if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
+ DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
+ return Rotate;
-/// \brief Helper to test for a load that can be folded with x86 shuffles.
-///
-/// This is particularly important because the set of instructions varies
-/// significantly based on whether the operand is a load or not.
-static bool isShuffleFoldableLoad(SDValue V) {
- while (V.getOpcode() == ISD::BITCAST)
- V = V.getOperand(0);
+ // If we have direct support for blends, we should lower by decomposing into
+ // a permute. That will be faster than the domain cross.
+ if (IsBlendSupported)
+ return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
+ Mask, DAG);
- return ISD::isNON_EXTLoad(V.getNode());
+ // Try to lower by permuting the inputs into an unpack instruction.
+ if (SDValue Unpack =
+ lowerVectorShuffleAsUnpack(DL, MVT::v4i32, V1, V2, Mask, DAG))
+ return Unpack;
+
+ // 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
+ // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
+ // relevant.
+ return DAG.getNode(ISD::BITCAST, DL, MVT::v4i32,
+ DAG.getVectorShuffle(
+ MVT::v4f32, DL,
+ DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V1),
+ DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V2), Mask));
}
-/// \brief Try to lower insertion of a single element into a zero vector.
+/// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
+/// shuffle lowering, and the most complex part.
///
-/// This is a common pattern that we have especially efficient patterns to lower
-/// across all subtarget feature sets.
-static SDValue lowerVectorShuffleAsElementInsertion(
- MVT VT, SDLoc DL, SDValue V1, SDValue V2, ArrayRef<int> Mask,
+/// The lowering strategy is to try to form pairs of input lanes which are
+/// targeted at the same half of the final vector, and then use a dword shuffle
+/// to place them onto the right half, and finally unpack the paired lanes into
+/// their final position.
+///
+/// The exact breakdown of how to form these dword pairs and align them on the
+/// correct sides is really tricky. See the comments within the function for
+/// more of the details.
+///
+/// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
+/// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
+/// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
+/// vector, form the analogous 128-bit 8-element Mask.
+static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
+ SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
const X86Subtarget *Subtarget, SelectionDAG &DAG) {
- SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
- MVT ExtVT = VT;
- MVT EltVT = VT.getVectorElementType();
+ assert(VT.getScalarType() == MVT::i16 && "Bad input type!");
+ MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
- int V2Index = std::find_if(Mask.begin(), Mask.end(),
- [&Mask](int M) { return M >= (int)Mask.size(); }) -
- Mask.begin();
- bool IsV1Zeroable = true;
- for (int i = 0, Size = Mask.size(); i < Size; ++i)
- if (i != V2Index && !Zeroable[i]) {
- IsV1Zeroable = false;
- break;
- }
+ assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
+ MutableArrayRef<int> LoMask = Mask.slice(0, 4);
+ MutableArrayRef<int> HiMask = Mask.slice(4, 4);
- // 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 (SDValue V2S = getScalarValueForVectorElement(
- V2, Mask[V2Index] - Mask.size(), DAG)) {
- // We need to zext the scalar if it is smaller than an i32.
- V2S = DAG.getNode(ISD::BITCAST, DL, EltVT, V2S);
- if (EltVT == MVT::i8 || EltVT == MVT::i16) {
- // Using zext to expand a narrow element won't work for non-zero
- // insertions.
- if (!IsV1Zeroable)
- return SDValue();
+ SmallVector<int, 4> LoInputs;
+ std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
+ [](int M) { return M >= 0; });
+ std::sort(LoInputs.begin(), LoInputs.end());
+ LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
+ SmallVector<int, 4> HiInputs;
+ std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
+ [](int M) { return M >= 0; });
+ std::sort(HiInputs.begin(), HiInputs.end());
+ HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
+ int NumLToL =
+ std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
+ int NumHToL = LoInputs.size() - NumLToL;
+ int NumLToH =
+ std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
+ int NumHToH = HiInputs.size() - NumLToH;
+ MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
+ MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
+ MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
+ MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
- // Zero-extend directly to i32.
- ExtVT = MVT::v4i32;
- V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
- }
- V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
- } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
- EltVT == MVT::i16) {
- // Either not inserting from the low element of the input or the input
- // element size is too small to use VZEXT_MOVL to clear the high bits.
- return SDValue();
- }
+ // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
+ // such inputs we can swap two of the dwords across the half mark and end up
+ // with <=2 inputs to each half in each half. Once there, we can fall through
+ // to the generic code below. For example:
+ //
+ // 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]
+ //
+ // 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).");
- if (!IsV1Zeroable) {
- // If V1 can't be treated as a zero vector we have fewer options to lower
- // this. We can't support integer vectors or non-zero targets cheaply, and
- // the V1 elements can't be permuted in any way.
- assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
- if (!VT.isFloatingPoint() || V2Index != 0)
- return SDValue();
- SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
- V1Mask[V2Index] = -1;
- if (!isNoopShuffleMask(V1Mask))
- return SDValue();
- // This is essentially a special case blend operation, but if we have
- // general purpose blend operations, they are always faster. Bail and let
- // the rest of the lowering handle these as blends.
- if (Subtarget->hasSSE41())
- return SDValue();
+ // 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 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;
- // Otherwise, use MOVSD or MOVSS.
- assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
- "Only two types of floating point element types to handle!");
- return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
- ExtVT, V1, V2);
- }
+ // We use xor with one to compute the adjacent DWord to whichever one the
+ // OneInput is in.
+ OneInputDWord = (OneInput / 2) ^ 1;
- V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
- if (ExtVT != VT)
- V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
+ // 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));
- 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.isFloatingPoint() || 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);
+ 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);
+ }
+ }
}
- }
- return V2;
-}
-
-/// \brief Try to lower broadcast of a single element.
-///
-/// For convenience, this code also bundles all of the subtarget feature set
-/// filtering. While a little annoying to re-dispatch on type here, there isn't
-/// a convenient way to factor it out.
-static SDValue lowerVectorShuffleAsBroadcast(MVT VT, SDLoc DL, SDValue V,
- ArrayRef<int> Mask,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- if (!Subtarget->hasAVX())
- return SDValue();
- if (VT.isInteger() && !Subtarget->hasAVX2())
- return SDValue();
- // Check that the mask is a broadcast.
- int BroadcastIdx = -1;
- for (int M : Mask)
- if (M >= 0 && BroadcastIdx == -1)
- BroadcastIdx = M;
- else if (M >= 0 && M != BroadcastIdx)
- return SDValue();
+ int PSHUFDMask[] = {0, 1, 2, 3};
+ PSHUFDMask[ADWord] = BDWord;
+ PSHUFDMask[BDWord] = ADWord;
+ V = DAG.getNode(ISD::BITCAST, DL, VT,
+ DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT,
+ DAG.getNode(ISD::BITCAST, DL, PSHUFDVT, V),
+ getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
- assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
- "a sorted mask where the broadcast "
- "comes from V1.");
+ // Adjust the mask to match the new locations of A and B.
+ for (int &M : Mask)
+ if (M != -1 && M/2 == ADWord)
+ M = 2 * BDWord + M % 2;
+ else if (M != -1 && M/2 == BDWord)
+ M = 2 * ADWord + M % 2;
- // Go up the chain of (vector) values to try and find a scalar load that
- // we can combine with the broadcast.
- for (;;) {
- switch (V.getOpcode()) {
- case ISD::CONCAT_VECTORS: {
- int OperandSize = Mask.size() / V.getNumOperands();
- V = V.getOperand(BroadcastIdx / OperandSize);
- BroadcastIdx %= OperandSize;
- continue;
- }
+ // Recurse back into this routine to re-compute state now that this isn't
+ // a 3 and 1 problem.
+ return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
+ DAG);
+ };
+ 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);
- case ISD::INSERT_SUBVECTOR: {
- SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
- auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
- if (!ConstantIdx)
- break;
+ // 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
+ // those dwords can then be moved to the correct half with a dword shuffle.
+ // We use at most one low and one high word shuffle to collect these paired
+ // inputs into dwords, and finally a dword shuffle to place them.
+ int PSHUFLMask[4] = {-1, -1, -1, -1};
+ int PSHUFHMask[4] = {-1, -1, -1, -1};
+ int PSHUFDMask[4] = {-1, -1, -1, -1};
- int BeginIdx = (int)ConstantIdx->getZExtValue();
- int EndIdx =
- BeginIdx + (int)VInner.getValueType().getVectorNumElements();
- if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
- BroadcastIdx -= BeginIdx;
- V = VInner;
- } else {
- V = VOuter;
- }
- continue;
+ // 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, ArrayRef<int> IncomingInputs,
+ MutableArrayRef<int> SourceHalfMask,
+ MutableArrayRef<int> HalfMask, int HalfOffset) {
+ if (InPlaceInputs.empty())
+ return;
+ if (InPlaceInputs.size() == 1) {
+ SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
+ InPlaceInputs[0] - HalfOffset;
+ 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;
}
- break;
- }
- // Check if this is a broadcast of a scalar. We special case lowering
- // for scalars so that we can more effectively fold with loads.
- if (V.getOpcode() == ISD::BUILD_VECTOR ||
- (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
- V = V.getOperand(BroadcastIdx);
+ assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
+ SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
+ InPlaceInputs[0] - HalfOffset;
+ // Put the second input next to the first so that they are packed into
+ // a dword. We find the adjacent index by toggling the low bit.
+ int AdjIndex = InPlaceInputs[0] ^ 1;
+ SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
+ std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
+ PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
+ };
+ fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
+ fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
- // If the scalar isn't a load we can't broadcast from it in AVX1, only with
- // AVX2.
- if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
- return SDValue();
- } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
- // We can't broadcast from a vector register w/o AVX2, and we can only
- // broadcast from the zero-element of a vector register.
- return SDValue();
- }
+ // Now gather the cross-half inputs and place them into a free dword of
+ // their target half.
+ // FIXME: This operation could almost certainly be simplified dramatically to
+ // look more like the 3-1 fixing operation.
+ auto moveInputsToRightHalf = [&PSHUFDMask](
+ MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
+ MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
+ MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
+ int DestOffset) {
+ auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
+ return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
+ };
+ auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
+ int Word) {
+ int LowWord = Word & ~1;
+ int HighWord = Word | 1;
+ return isWordClobbered(SourceHalfMask, LowWord) ||
+ isWordClobbered(SourceHalfMask, HighWord);
+ };
- return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
-}
+ if (IncomingInputs.empty())
+ return;
-/// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
-///
-/// This is the basis function for the 2-lane 64-bit shuffles as we have full
-/// support for floating point shuffles but not integer shuffles. These
-/// instructions will incur a domain crossing penalty on some chips though so
-/// it is better to avoid lowering through this for integer vectors where
-/// possible.
-static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- SDLoc DL(Op);
- assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
- assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
- assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> Mask = SVOp->getMask();
- assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
+ if (ExistingInputs.empty()) {
+ // Map any dwords with inputs from them into the right half.
+ for (int Input : IncomingInputs) {
+ // If the source half mask maps over the inputs, turn those into
+ // swaps and use the swapped lane.
+ if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
+ if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
+ SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
+ Input - SourceOffset;
+ // We have to swap the uses in our half mask in one sweep.
+ for (int &M : HalfMask)
+ if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
+ M = Input;
+ else if (M == Input)
+ M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
+ } else {
+ assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
+ Input - SourceOffset &&
+ "Previous placement doesn't match!");
+ }
+ // Note that this correctly re-maps both when we do a swap and when
+ // we observe the other side of the swap above. We rely on that to
+ // avoid swapping the members of the input list directly.
+ Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
+ }
- if (isSingleInputShuffleMask(Mask)) {
- // Straight shuffle of a single input vector. Simulate this by using the
- // single input as both of the "inputs" to this instruction..
- unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
+ // Map the input's dword into the correct half.
+ if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
+ PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
+ else
+ assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
+ Input / 2 &&
+ "Previous placement doesn't match!");
+ }
- if (Subtarget->hasAVX()) {
- // If we have AVX, we can use VPERMILPS which will allow folding a load
- // into the shuffle.
- return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
- DAG.getConstant(SHUFPDMask, MVT::i8));
+ // And just directly shift any other-half mask elements to be same-half
+ // as we will have mirrored the dword containing the element into the
+ // same position within that half.
+ for (int &M : HalfMask)
+ if (M >= SourceOffset && M < SourceOffset + 4) {
+ M = M - SourceOffset + DestOffset;
+ assert(M >= 0 && "This should never wrap below zero!");
+ }
+ return;
}
- return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V1,
- DAG.getConstant(SHUFPDMask, MVT::i8));
- }
- assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
- assert(Mask[1] >= 2 && "Non-canonicalized blend!");
+ // Ensure we have the input in a viable dword of its current half. This
+ // is particularly tricky because the original position may be clobbered
+ // by inputs being moved and *staying* in that half.
+ if (IncomingInputs.size() == 1) {
+ if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
+ int InputFixed = std::find(std::begin(SourceHalfMask),
+ std::end(SourceHalfMask), -1) -
+ std::begin(SourceHalfMask) + SourceOffset;
+ SourceHalfMask[InputFixed - SourceOffset] =
+ IncomingInputs[0] - SourceOffset;
+ std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
+ InputFixed);
+ IncomingInputs[0] = InputFixed;
+ }
+ } else if (IncomingInputs.size() == 2) {
+ if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
+ isDWordClobbered(SourceHalfMask, IncomingInputs[0] - 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};
- // 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);
+ // 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!");
- // If we have a single input, insert that into V1 if we can do so cheaply.
- if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
- if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
- MVT::v2f64, DL, V1, V2, Mask, Subtarget, DAG))
- return Insertion;
- // Try inverting the insertion since for v2 masks it is easy to do and we
- // can't reliably sort the mask one way or the other.
- int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
- Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
- if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
- MVT::v2f64, DL, V2, V1, InverseMask, Subtarget, DAG))
- return Insertion;
- }
+ SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
+ SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
- // Try to use one of the special instruction patterns to handle two common
- // blend patterns if a zero-blend above didn't work.
- if (isShuffleEquivalent(Mask, 0, 3) || isShuffleEquivalent(Mask, 1, 3))
- if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
- // We can either use a special instruction to load over the low double or
- // to move just the low double.
- return DAG.getNode(
- isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
- DL, MVT::v2f64, V2,
- DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
+ // 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;
- if (Subtarget->hasSSE41())
- if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
- Subtarget, DAG))
- return Blend;
+ InputsFixed[1] = InputsFixed[0] ^ 1;
+ }
- 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));
-}
+ // Point everything at the fixed inputs.
+ for (int &M : HalfMask)
+ if (M == IncomingInputs[0])
+ M = InputsFixed[0] + SourceOffset;
+ else if (M == IncomingInputs[1])
+ M = InputsFixed[1] + SourceOffset;
-/// \brief Handle lowering of 2-lane 64-bit integer shuffles.
-///
-/// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
-/// the integer unit to minimize domain crossing penalties. However, for blends
-/// it falls back to the floating point shuffle operation with appropriate bit
-/// casting.
-static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- SDLoc DL(Op);
- assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
- assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
- assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> Mask = SVOp->getMask();
- assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
+ IncomingInputs[0] = InputsFixed[0] + SourceOffset;
+ IncomingInputs[1] = InputsFixed[1] + SourceOffset;
+ }
+ } else {
+ llvm_unreachable("Unhandled input size!");
+ }
- if (isSingleInputShuffleMask(Mask)) {
- // Check for being able to broadcast a single element.
- if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v2i64, DL, V1,
- Mask, Subtarget, DAG))
- return Broadcast;
+ // Now hoist the DWord down to the right half.
+ int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
+ assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
+ PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
+ for (int &M : HalfMask)
+ for (int Input : IncomingInputs)
+ if (M == Input)
+ M = FreeDWord * 2 + Input % 2;
+ };
+ moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
+ /*SourceOffset*/ 4, /*DestOffset*/ 0);
+ moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
+ /*SourceOffset*/ 0, /*DestOffset*/ 4);
- // Straight shuffle of a single input vector. For everything from SSE2
- // onward this has a single fast instruction with no scary immediates.
- // We have to map the mask as it is actually a v4i32 shuffle instruction.
- V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V1);
- int WidenedMask[4] = {
- std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
- std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
- return DAG.getNode(
- ISD::BITCAST, DL, MVT::v2i64,
- DAG.getNode(X86ISD::PSHUFD, SDLoc(Op), MVT::v4i32, V1,
- getV4X86ShuffleImm8ForMask(WidenedMask, DAG)));
- }
+ // Now enact all the shuffles we've computed to move the inputs into their
+ // target half.
+ if (!isNoopShuffleMask(PSHUFLMask))
+ V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
+ getV4X86ShuffleImm8ForMask(PSHUFLMask, DAG));
+ if (!isNoopShuffleMask(PSHUFHMask))
+ V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
+ getV4X86ShuffleImm8ForMask(PSHUFHMask, DAG));
+ if (!isNoopShuffleMask(PSHUFDMask))
+ V = DAG.getNode(ISD::BITCAST, DL, VT,
+ DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT,
+ DAG.getNode(ISD::BITCAST, DL, PSHUFDVT, V),
+ getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
- // Try to use byte shift instructions.
- if (SDValue Shift = lowerVectorShuffleAsByteShift(
- DL, MVT::v2i64, V1, V2, Mask, DAG))
- return Shift;
+ // At this point, each half should contain all its inputs, and we can then
+ // just shuffle them into their final position.
+ assert(std::count_if(LoMask.begin(), LoMask.end(),
+ [](int M) { return M >= 4; }) == 0 &&
+ "Failed to lift all the high half inputs to the low mask!");
+ assert(std::count_if(HiMask.begin(), HiMask.end(),
+ [](int M) { return M >= 0 && M < 4; }) == 0 &&
+ "Failed to lift all the low half inputs to the high mask!");
- // If we have a single input from V2 insert that into V1 if we can do so
- // cheaply.
- if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
- if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
- MVT::v2i64, DL, V1, V2, Mask, Subtarget, DAG))
- return Insertion;
- // Try inverting the insertion since for v2 masks it is easy to do and we
- // can't reliably sort the mask one way or the other.
- int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
- Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
- if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
- MVT::v2i64, DL, V2, V1, InverseMask, Subtarget, DAG))
- return Insertion;
- }
+ // Do a half shuffle for the low mask.
+ if (!isNoopShuffleMask(LoMask))
+ V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
+ getV4X86ShuffleImm8ForMask(LoMask, 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);
-
- if (Subtarget->hasSSE41())
- if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
- Subtarget, DAG))
- return Blend;
-
- // Try to use byte rotation instructions.
- // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
- if (Subtarget->hasSSSE3())
- if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
- DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
- return Rotate;
+ // Do a half shuffle with the high mask after shifting its values down.
+ for (int &M : HiMask)
+ if (M >= 0)
+ M -= 4;
+ if (!isNoopShuffleMask(HiMask))
+ V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
+ getV4X86ShuffleImm8ForMask(HiMask, DAG));
- // 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
- // have this problem. It would be really nice if x86 had better shuffles here.
- V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V1);
- V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V2);
- return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
- DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
+ return V;
}
-/// \brief Lower a vector shuffle using the SHUFPS instruction.
-///
-/// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
-/// It makes no assumptions about whether this is the *best* lowering, it simply
-/// uses it.
-static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
- ArrayRef<int> Mask, SDValue V1,
- SDValue V2, SelectionDAG &DAG) {
- SDValue LowV = V1, HighV = V2;
- int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
-
- int NumV2Elements =
- std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
-
- if (NumV2Elements == 1) {
- int V2Index =
- std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
- Mask.begin();
-
- // Compute the index adjacent to V2Index and in the same half by toggling
- // the low bit.
- int V2AdjIndex = V2Index ^ 1;
-
- if (Mask[V2AdjIndex] == -1) {
- // Handles all the cases where we have a single V2 element and an undef.
- // This will only ever happen in the high lanes because we commute the
- // vector otherwise.
- if (V2Index < 2)
- std::swap(LowV, HighV);
- NewMask[V2Index] -= 4;
- } else {
- // Handle the case where the V2 element ends up adjacent to a V1 element.
- // To make this work, blend them together as the first step.
- int V1Index = V2AdjIndex;
- int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
- V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
- getV4X86ShuffleImm8ForMask(BlendMask, DAG));
+/// \brief Helper to form a PSHUFB-based shuffle+blend.
+static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
+ SDValue V2, ArrayRef<int> Mask,
+ SelectionDAG &DAG, bool &V1InUse,
+ bool &V2InUse) {
+ SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
+ SDValue V1Mask[16];
+ SDValue V2Mask[16];
+ V1InUse = false;
+ V2InUse = false;
- // Now proceed to reconstruct the final blend as we have the necessary
- // high or low half formed.
- if (V2Index < 2) {
- LowV = V2;
- HighV = V1;
- } else {
- HighV = V2;
- }
- NewMask[V1Index] = 2; // We put the V1 element in V2[2].
- NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
- }
- } else if (NumV2Elements == 2) {
- if (Mask[0] < 4 && Mask[1] < 4) {
- // Handle the easy case where we have V1 in the low lanes and V2 in the
- // high lanes.
- NewMask[2] -= 4;
- NewMask[3] -= 4;
- } else if (Mask[2] < 4 && Mask[3] < 4) {
- // We also handle the reversed case because this utility may get called
- // when we detect a SHUFPS pattern but can't easily commute the shuffle to
- // arrange things in the right direction.
- NewMask[0] -= 4;
- NewMask[1] -= 4;
- HighV = V1;
- LowV = V2;
+ int Size = Mask.size();
+ int Scale = 16 / Size;
+ for (int i = 0; i < 16; ++i) {
+ if (Mask[i / Scale] == -1) {
+ V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
} else {
- // We have a mixture of V1 and V2 in both low and high lanes. Rather than
- // trying to place elements directly, just blend them and set up the final
- // shuffle to place them.
+ const int ZeroMask = 0x80;
+ int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
+ : ZeroMask;
+ int V2Idx = Mask[i / Scale] < Size
+ ? ZeroMask
+ : (Mask[i / Scale] - Size) * Scale + i % Scale;
+ if (Zeroable[i / Scale])
+ V1Idx = V2Idx = ZeroMask;
+ V1Mask[i] = DAG.getConstant(V1Idx, MVT::i8);
+ V2Mask[i] = DAG.getConstant(V2Idx, MVT::i8);
+ V1InUse |= (ZeroMask != V1Idx);
+ V2InUse |= (ZeroMask != V2Idx);
+ }
+ }
+
+ if (V1InUse)
+ V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
+ DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, V1),
+ DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
+ if (V2InUse)
+ V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
+ DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, V2),
+ DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
- // The first two blend mask elements are for V1, the second two are for
- // V2.
- int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
- Mask[2] < 4 ? Mask[2] : Mask[3],
- (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
- (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
- V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
- getV4X86ShuffleImm8ForMask(BlendMask, DAG));
+ // If we need shuffled inputs from both, blend the two.
+ SDValue V;
+ if (V1InUse && V2InUse)
+ V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
+ else
+ V = V1InUse ? V1 : V2;
- // Now we do a normal shuffle of V1 by giving V1 as both operands to
- // a blend.
- LowV = HighV = V1;
- NewMask[0] = Mask[0] < 4 ? 0 : 2;
- NewMask[1] = Mask[0] < 4 ? 2 : 0;
- NewMask[2] = Mask[2] < 4 ? 1 : 3;
- NewMask[3] = Mask[2] < 4 ? 3 : 1;
- }
- }
- return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
- getV4X86ShuffleImm8ForMask(NewMask, DAG));
+ // Cast the result back to the correct type.
+ return DAG.getNode(ISD::BITCAST, DL, VT, V);
}
-/// \brief Lower 4-lane 32-bit floating point shuffles.
+/// \brief Generic lowering of 8-lane i16 shuffles.
///
-/// Uses instructions exclusively from the floating point unit to minimize
-/// domain crossing penalties, as these are sufficient to implement all v4f32
-/// shuffles.
-static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+/// This handles both single-input shuffles and combined shuffle/blends with
+/// two inputs. The single input shuffles are immediately delegated to
+/// a dedicated lowering routine.
+///
+/// The blends are lowered in one of three fundamental ways. If there are few
+/// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
+/// of the input is significantly cheaper when lowered as an interleaving of
+/// the two inputs, try to interleave them. Otherwise, blend the low and high
+/// halves of the inputs separately (making them have relatively few inputs)
+/// and then concatenate them.
+static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
const X86Subtarget *Subtarget,
SelectionDAG &DAG) {
SDLoc DL(Op);
- assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
- assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
- assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
+ assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
+ assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> Mask = SVOp->getMask();
- assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
+ ArrayRef<int> OrigMask = SVOp->getMask();
+ int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
+ OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
+ MutableArrayRef<int> Mask(MaskStorage);
- int NumV2Elements =
- std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
+ assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
- if (NumV2Elements == 0) {
+ // Whenever we can lower this as a zext, that instruction is strictly faster
+ // than any alternative.
+ if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
+ DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
+ return ZExt;
+
+ auto isV1 = [](int M) { return M >= 0 && M < 8; };
+ (void)isV1;
+ auto isV2 = [](int M) { return M >= 8; };
+
+ int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
+
+ if (NumV2Inputs == 0) {
// Check for being able to broadcast a single element.
- if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4f32, DL, V1,
+ if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
Mask, Subtarget, DAG))
return Broadcast;
- if (Subtarget->hasAVX()) {
- // If we have AVX, we can use VPERMILPS which will allow folding a load
- // into the shuffle.
- return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
- getV4X86ShuffleImm8ForMask(Mask, DAG));
- }
+ // Try to use shift instructions.
+ if (SDValue Shift =
+ lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
+ return Shift;
- // Otherwise, use a straight shuffle of a single input vector. We pass the
- // input vector to both operands to simulate this with a SHUFPS.
- 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(V1, V1, Mask, {0, 0, 1, 1, 2, 2, 3, 3}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V1);
+ if (isShuffleEquivalent(V1, V1, Mask, {4, 4, 5, 5, 6, 6, 7, 7}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V1);
+
+ // Try to use byte rotation instructions.
+ if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
+ Mask, Subtarget, DAG))
+ return Rotate;
+
+ return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
+ Subtarget, 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);
+ assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
+ "All single-input shuffles should be canonicalized to be V1-input "
+ "shuffles.");
- // There are special ways we can lower some single-element blends. However, we
- // have custom ways we can lower more complex single-element blends below that
- // we defer to if both this and BLENDPS fail to match, so restrict this to
- // when the V2 input is targeting element 0 of the mask -- that is the fast
- // case here.
- if (NumV2Elements == 1 && Mask[0] >= 4)
- if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v4f32, DL, V1, V2,
+ // Try to use shift instructions.
+ if (SDValue Shift =
+ lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
+ return Shift;
+
+ // There are special ways we can lower some single-element blends.
+ if (NumV2Inputs == 1)
+ if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
Mask, Subtarget, DAG))
return V;
- if (Subtarget->hasSSE41())
- if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
+ // We have different paths for blend lowering, but they all must use the
+ // *exact* same predicate.
+ bool IsBlendSupported = Subtarget->hasSSE41();
+ if (IsBlendSupported)
+ if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
Subtarget, DAG))
return Blend;
- // 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 (NumV2Elements == 1 && Subtarget->hasSSE41()) {
- int V2Index =
- std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
- Mask.begin();
+ if (SDValue Masked =
+ lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
+ return Masked;
- // 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.
- SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
+ // Use dedicated unpack instructions for masks that match their pattern.
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 2, 10, 3, 11}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask, {4, 12, 5, 13, 6, 14, 7, 15}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
- // Synthesize a shuffle mask for the non-zero and non-v2 inputs.
- bool InsertNeedsShuffle = false;
- unsigned ZMask = 0;
- for (int i = 0; i < 4; ++i)
- if (i != V2Index) {
- if (Zeroable[i]) {
- ZMask |= 1 << i;
- } else if (Mask[i] != i) {
- InsertNeedsShuffle = true;
- break;
- }
- }
+ // Try to use byte rotation instructions.
+ if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
+ DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
+ return Rotate;
- // We don't want to use INSERTPS or other insertion techniques if it will
- // require shuffling anyways.
- if (!InsertNeedsShuffle) {
- // If all of V1 is zeroable, replace it with undef.
- if ((ZMask | 1 << V2Index) == 0xF)
- V1 = DAG.getUNDEF(MVT::v4f32);
+ if (SDValue BitBlend =
+ lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
+ return BitBlend;
- unsigned InsertPSMask = (Mask[V2Index] - 4) << 6 | V2Index << 4 | ZMask;
- assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
+ if (SDValue Unpack =
+ lowerVectorShuffleAsUnpack(DL, MVT::v8i16, V1, V2, Mask, DAG))
+ return Unpack;
- // Insert the V2 element into the desired position.
- return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
- DAG.getConstant(InsertPSMask, MVT::i8));
- }
+ // If we can't directly blend but can use PSHUFB, that will be better as it
+ // can both shuffle and set up the inefficient blend.
+ if (!IsBlendSupported && Subtarget->hasSSSE3()) {
+ bool V1InUse, V2InUse;
+ return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
+ V1InUse, V2InUse);
}
- // Otherwise fall back to a SHUFPS lowering strategy.
- return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
+ // We can always bit-blend if we have to so the fallback strategy is to
+ // decompose into single-input permutes and blends.
+ return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
+ Mask, DAG);
}
-/// \brief Lower 4-lane i32 vector shuffles.
+/// \brief Check whether a compaction lowering can be done by dropping even
+/// elements and compute how many times even elements must be dropped.
///
-/// We try to handle these with integer-domain shuffles where we can, but for
-/// blends we use the floating point domain blend instructions.
-static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- SDLoc DL(Op);
- assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
- assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
- assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> Mask = SVOp->getMask();
- assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
-
- // Whenever we can lower this as a zext, that instruction is strictly faster
- // than any alternative. It also allows us to fold memory operands into the
- // shuffle in many cases.
- if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
- Mask, Subtarget, DAG))
- return ZExt;
+/// 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);
- int NumV2Elements =
- std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
-
- if (NumV2Elements == 0) {
- // Check for being able to broadcast a single element.
- if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4i32, DL, V1,
- Mask, Subtarget, DAG))
- return Broadcast;
-
- // Straight shuffle of a single input vector. For everything from SSE2
- // onward this has a single fast instruction with no scary immediates.
- // We coerce the shuffle pattern to be compatible with UNPCK instructions
- // but we aren't actually going to use the UNPCK instruction because doing
- // so prevents folding a load into this instruction or making a copy.
- const int UnpackLoMask[] = {0, 0, 1, 1};
- const int UnpackHiMask[] = {2, 2, 3, 3};
- if (isShuffleEquivalent(Mask, 0, 0, 1, 1))
- Mask = UnpackLoMask;
- else if (isShuffleEquivalent(Mask, 2, 2, 3, 3))
- Mask = UnpackHiMask;
+ // 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!");
- return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
- getV4X86ShuffleImm8ForMask(Mask, DAG));
- }
+ uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
- // Try to use byte shift instructions.
- if (SDValue Shift = lowerVectorShuffleAsByteShift(
- DL, MVT::v4i32, V1, V2, Mask, DAG))
- return Shift;
+ // 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};
- // There are special ways we can lower some single-element blends.
- if (NumV2Elements == 1)
- if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v4i32, DL, V1, V2,
- Mask, Subtarget, DAG))
- return V;
+ 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;
- // 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);
+ bool IsAnyViable = false;
+ for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
+ if (ViableForN[j]) {
+ uint64_t N = j + 1;
- if (Subtarget->hasSSE41())
- if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
- Subtarget, DAG))
- return Blend;
+ // 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;
+ }
- // Try to use byte rotation instructions.
- // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
- if (Subtarget->hasSSSE3())
- if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
- DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
- return Rotate;
+ for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
+ if (ViableForN[j])
+ return j + 1;
- // 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
- // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
- // relevant.
- return DAG.getNode(ISD::BITCAST, DL, MVT::v4i32,
- DAG.getVectorShuffle(
- MVT::v4f32, DL,
- DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V1),
- DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V2), Mask));
+ // Return 0 as there is no viable power of two.
+ return 0;
}
-/// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
-/// shuffle lowering, and the most complex part.
-///
-/// The lowering strategy is to try to form pairs of input lanes which are
-/// targeted at the same half of the final vector, and then use a dword shuffle
-/// to place them onto the right half, and finally unpack the paired lanes into
-/// their final position.
+/// \brief Generic lowering of v16i8 shuffles.
///
-/// The exact breakdown of how to form these dword pairs and align them on the
-/// correct sides is really tricky. See the comments within the function for
-/// more of the details.
-static SDValue lowerV8I16SingleInputVectorShuffle(
- SDLoc DL, SDValue V, MutableArrayRef<int> Mask,
- const X86Subtarget *Subtarget, SelectionDAG &DAG) {
- assert(V.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
- MutableArrayRef<int> LoMask = Mask.slice(0, 4);
- MutableArrayRef<int> HiMask = Mask.slice(4, 4);
-
- SmallVector<int, 4> LoInputs;
- std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
- [](int M) { return M >= 0; });
- std::sort(LoInputs.begin(), LoInputs.end());
- LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
- SmallVector<int, 4> HiInputs;
- std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
- [](int M) { return M >= 0; });
- std::sort(HiInputs.begin(), HiInputs.end());
- HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
- int NumLToL =
- std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
- int NumHToL = LoInputs.size() - NumLToL;
- int NumLToH =
- std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
- int NumHToH = HiInputs.size() - NumLToH;
- MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
- MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
- MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
- MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
-
- // Check for being able to broadcast a single element.
- if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v8i16, DL, V,
- Mask, Subtarget, DAG))
- return Broadcast;
+/// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
+/// detect any complexity reducing interleaving. If that doesn't help, it uses
+/// UNPCK to spread the i8 elements across two i16-element vectors, and uses
+/// the existing lowering for v8i16 blends on each half, finally PACK-ing them
+/// back together.
+static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
+ assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
- // Try to use byte shift instructions.
- if (SDValue Shift = lowerVectorShuffleAsByteShift(
- DL, MVT::v8i16, V, V, Mask, DAG))
+ // Try to use shift instructions.
+ if (SDValue Shift =
+ lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
return Shift;
- // Use dedicated unpack instructions for masks that match their pattern.
- if (isShuffleEquivalent(Mask, 0, 0, 1, 1, 2, 2, 3, 3))
- return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V, V);
- if (isShuffleEquivalent(Mask, 4, 4, 5, 5, 6, 6, 7, 7))
- return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V, V);
-
// Try to use byte rotation instructions.
if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
- DL, MVT::v8i16, V, V, Mask, Subtarget, DAG))
+ DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
return Rotate;
- // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
- // such inputs we can swap two of the dwords across the half mark and end up
- // with <=2 inputs to each half in each half. Once there, we can fall through
- // to the generic code below. For example:
- //
- // 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]
- //
- // 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 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;
+ // Try to use a zext lowering.
+ if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
+ DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
+ return ZExt;
- // We use xor with one to compute the adjacent DWord to whichever one the
- // OneInput is in.
- OneInputDWord = (OneInput / 2) ^ 1;
+ int NumV2Elements =
+ std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
- // 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 single-input shuffles, there are some nicer lowering tricks we can use.
+ if (NumV2Elements == 0) {
+ // Check for being able to broadcast a single element.
+ if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
+ Mask, Subtarget, DAG))
+ return Broadcast;
- 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);
- }
- }
- }
+ // 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
+ // things significantly. Currently, this means we need to be able to
+ // express the pre-duplication shuffle as an i16 shuffle.
+ //
+ // FIXME: We should check for other patterns which can be widened into an
+ // i16 shuffle as well.
+ auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
+ for (int i = 0; i < 16; i += 2)
+ if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
+ return false;
- int PSHUFDMask[] = {0, 1, 2, 3};
- 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),
- getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
+ return true;
+ };
+ auto tryToWidenViaDuplication = [&]() -> SDValue {
+ if (!canWidenViaDuplication(Mask))
+ return SDValue();
+ SmallVector<int, 4> LoInputs;
+ std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
+ [](int M) { return M >= 0 && M < 8; });
+ std::sort(LoInputs.begin(), LoInputs.end());
+ LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
+ LoInputs.end());
+ SmallVector<int, 4> HiInputs;
+ std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
+ [](int M) { return M >= 8; });
+ std::sort(HiInputs.begin(), HiInputs.end());
+ HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
+ HiInputs.end());
- // Adjust the mask to match the new locations of A and B.
- for (int &M : Mask)
- if (M != -1 && M/2 == ADWord)
- M = 2 * BDWord + M % 2;
- else if (M != -1 && M/2 == BDWord)
- M = 2 * ADWord + M % 2;
+ bool TargetLo = LoInputs.size() >= HiInputs.size();
+ ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
+ ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
- // 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) || (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);
+ int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
+ SmallDenseMap<int, int, 8> LaneMap;
+ for (int I : InPlaceInputs) {
+ PreDupI16Shuffle[I/2] = I/2;
+ LaneMap[I] = I;
+ }
+ int j = TargetLo ? 0 : 4, je = j + 4;
+ for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
+ // Check if j is already a shuffle of this input. This happens when
+ // there are two adjacent bytes after we move the low one.
+ if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
+ // If we haven't yet mapped the input, search for a slot into which
+ // we can map it.
+ while (j < je && PreDupI16Shuffle[j] != -1)
+ ++j;
- // 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
- // those dwords can then be moved to the correct half with a dword shuffle.
- // We use at most one low and one high word shuffle to collect these paired
- // inputs into dwords, and finally a dword shuffle to place them.
- int PSHUFLMask[4] = {-1, -1, -1, -1};
- int PSHUFHMask[4] = {-1, -1, -1, -1};
- int PSHUFDMask[4] = {-1, -1, -1, -1};
+ if (j == je)
+ // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
+ return SDValue();
- // 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, ArrayRef<int> IncomingInputs,
- MutableArrayRef<int> SourceHalfMask,
- MutableArrayRef<int> HalfMask, int HalfOffset) {
- if (InPlaceInputs.empty())
- return;
- if (InPlaceInputs.size() == 1) {
- SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
- InPlaceInputs[0] - HalfOffset;
- 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;
+ // Map this input with the i16 shuffle.
+ PreDupI16Shuffle[j] = MovingInputs[i] / 2;
+ }
+
+ // Update the lane map based on the mapping we ended up with.
+ LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
}
- return;
- }
+ V1 = DAG.getNode(
+ ISD::BITCAST, DL, MVT::v16i8,
+ DAG.getVectorShuffle(MVT::v8i16, DL,
+ DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
+ DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
- assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
- SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
- InPlaceInputs[0] - HalfOffset;
- // Put the second input next to the first so that they are packed into
- // a dword. We find the adjacent index by toggling the low bit.
- int AdjIndex = InPlaceInputs[0] ^ 1;
- SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
- std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
- PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
- };
- fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
- fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
+ // Unpack the bytes to form the i16s that will be shuffled into place.
+ V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
+ MVT::v16i8, V1, V1);
- // Now gather the cross-half inputs and place them into a free dword of
- // their target half.
- // FIXME: This operation could almost certainly be simplified dramatically to
- // look more like the 3-1 fixing operation.
- auto moveInputsToRightHalf = [&PSHUFDMask](
- MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
- MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
- MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
- int DestOffset) {
- auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
- return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
- };
- auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
- int Word) {
- int LowWord = Word & ~1;
- int HighWord = Word | 1;
- return isWordClobbered(SourceHalfMask, LowWord) ||
- isWordClobbered(SourceHalfMask, HighWord);
+ int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
+ for (int i = 0; i < 16; ++i)
+ if (Mask[i] != -1) {
+ int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
+ assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
+ if (PostDupI16Shuffle[i / 2] == -1)
+ PostDupI16Shuffle[i / 2] = MappedMask;
+ else
+ assert(PostDupI16Shuffle[i / 2] == MappedMask &&
+ "Conflicting entrties in the original shuffle!");
+ }
+ return DAG.getNode(
+ ISD::BITCAST, DL, MVT::v16i8,
+ DAG.getVectorShuffle(MVT::v8i16, DL,
+ DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
+ DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
};
+ if (SDValue V = tryToWidenViaDuplication())
+ return V;
+ }
- if (IncomingInputs.empty())
- return;
+ // Use dedicated unpack instructions for masks that match their pattern.
+ if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
+ 0, 16, 1, 17, 2, 18, 3, 19,
+ // High half.
+ 4, 20, 5, 21, 6, 22, 7, 23}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
+ 8, 24, 9, 25, 10, 26, 11, 27,
+ // High half.
+ 12, 28, 13, 29, 14, 30, 15, 31}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, V2);
- if (ExistingInputs.empty()) {
- // Map any dwords with inputs from them into the right half.
- for (int Input : IncomingInputs) {
- // If the source half mask maps over the inputs, turn those into
- // swaps and use the swapped lane.
- if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
- if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
- SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
- Input - SourceOffset;
- // We have to swap the uses in our half mask in one sweep.
- for (int &M : HalfMask)
- if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
- M = Input;
- else if (M == Input)
- M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
- } else {
- assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
- Input - SourceOffset &&
- "Previous placement doesn't match!");
- }
- // Note that this correctly re-maps both when we do a swap and when
- // we observe the other side of the swap above. We rely on that to
- // avoid swapping the members of the input list directly.
- Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
- }
+ // 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()) {
+ bool V1InUse = false;
+ bool V2InUse = false;
- // Map the input's dword into the correct half.
- if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
- PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
- else
- assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
- Input / 2 &&
- "Previous placement doesn't match!");
- }
+ SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
+ DAG, V1InUse, V2InUse);
- // And just directly shift any other-half mask elements to be same-half
- // as we will have mirrored the dword containing the element into the
- // same position within that half.
- for (int &M : HalfMask)
- if (M >= SourceOffset && M < SourceOffset + 4) {
- M = M - SourceOffset + DestOffset;
- assert(M >= 0 && "This should never wrap below zero!");
- }
- return;
+ // If both V1 and V2 are in use and we can use a direct blend or an unpack,
+ // do so. This avoids using them to handle blends-with-zero which is
+ // important as a single pshufb is significantly faster for that.
+ if (V1InUse && V2InUse) {
+ if (Subtarget->hasSSE41())
+ if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
+ Mask, Subtarget, DAG))
+ return Blend;
+
+ // We can use an unpack to do the blending rather than an or in some
+ // cases. Even though the or may be (very minorly) more efficient, we
+ // preference this lowering because there are common cases where part of
+ // the complexity of the shuffles goes away when we do the final blend as
+ // an unpack.
+ // FIXME: It might be worth trying to detect if the unpack-feeding
+ // shuffles will both be pshufb, in which case we shouldn't bother with
+ // this.
+ if (SDValue Unpack =
+ lowerVectorShuffleAsUnpack(DL, MVT::v16i8, V1, V2, Mask, DAG))
+ return Unpack;
}
- // Ensure we have the input in a viable dword of its current half. This
- // is particularly tricky because the original position may be clobbered
- // by inputs being moved and *staying* in that half.
- if (IncomingInputs.size() == 1) {
- if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
- int InputFixed = std::find(std::begin(SourceHalfMask),
- std::end(SourceHalfMask), -1) -
- std::begin(SourceHalfMask) + SourceOffset;
- SourceHalfMask[InputFixed - SourceOffset] =
- IncomingInputs[0] - SourceOffset;
- std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
- InputFixed);
- IncomingInputs[0] = InputFixed;
- }
- } else if (IncomingInputs.size() == 2) {
- if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
- isDWordClobbered(SourceHalfMask, IncomingInputs[0] - 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};
+ return PSHUFB;
+ }
- // 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;
+ // There are special ways we can lower some single-element blends.
+ if (NumV2Elements == 1)
+ if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
+ Mask, Subtarget, DAG))
+ return V;
- // 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;
+ if (SDValue BitBlend =
+ lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
+ return BitBlend;
- InputsFixed[1] = InputsFixed[0] ^ 1;
- }
+ // 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);
- // Point everything at the fixed inputs.
- for (int &M : HalfMask)
- if (M == IncomingInputs[0])
- M = InputsFixed[0] + SourceOffset;
- else if (M == IncomingInputs[1])
- M = InputsFixed[1] + SourceOffset;
+ // 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);
- IncomingInputs[0] = InputsFixed[0] + SourceOffset;
- IncomingInputs[1] = InputsFixed[1] + SourceOffset;
- }
- } else {
- llvm_unreachable("Unhandled input size!");
+ // 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);
}
- // Now hoist the DWord down to the right half.
- int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
- assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
- PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
- for (int &M : HalfMask)
- for (int Input : IncomingInputs)
- if (M == Input)
- M = FreeDWord * 2 + Input % 2;
- };
- moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
- /*SourceOffset*/ 4, /*DestOffset*/ 0);
- moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
- /*SourceOffset*/ 0, /*DestOffset*/ 4);
+ return Result;
+ }
- // Now enact all the shuffles we've computed to move the inputs into their
- // target half.
- if (!isNoopShuffleMask(PSHUFLMask))
- V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
- getV4X86ShuffleImm8ForMask(PSHUFLMask, DAG));
- if (!isNoopShuffleMask(PSHUFHMask))
- V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
- getV4X86ShuffleImm8ForMask(PSHUFHMask, DAG));
- if (!isNoopShuffleMask(PSHUFDMask))
- V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
- DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
- DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
- getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
+ // Handle multi-input cases by blending single-input shuffles.
+ if (NumV2Elements > 0)
+ return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
+ Mask, DAG);
- // At this point, each half should contain all its inputs, and we can then
- // just shuffle them into their final position.
- assert(std::count_if(LoMask.begin(), LoMask.end(),
- [](int M) { return M >= 4; }) == 0 &&
- "Failed to lift all the high half inputs to the low mask!");
- assert(std::count_if(HiMask.begin(), HiMask.end(),
- [](int M) { return M >= 0 && M < 4; }) == 0 &&
- "Failed to lift all the low half inputs to the high mask!");
+ // The fallback path for single-input shuffles widens this into two v8i16
+ // vectors with unpacks, shuffles those, and then pulls them back together
+ // with a pack.
+ SDValue V = V1;
- // Do a half shuffle for the low mask.
- if (!isNoopShuffleMask(LoMask))
- V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
- getV4X86ShuffleImm8ForMask(LoMask, DAG));
+ int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
+ int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
+ for (int i = 0; i < 16; ++i)
+ if (Mask[i] >= 0)
+ (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
- // Do a half shuffle with the high mask after shifting its values down.
- for (int &M : HiMask)
- if (M >= 0)
- M -= 4;
- if (!isNoopShuffleMask(HiMask))
- V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
- getV4X86ShuffleImm8ForMask(HiMask, DAG));
+ SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
- return V;
+ SDValue VLoHalf, VHiHalf;
+ // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
+ // them out and avoid using UNPCK{L,H} to extract the elements of V as
+ // i16s.
+ if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
+ [](int M) { return M >= 0 && M % 2 == 1; }) &&
+ std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
+ [](int M) { return M >= 0 && M % 2 == 1; })) {
+ // Use a mask to drop the high bytes.
+ VLoHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
+ VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
+ DAG.getConstant(0x00FF, MVT::v8i16));
+
+ // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
+ VHiHalf = DAG.getUNDEF(MVT::v8i16);
+
+ // Squash the masks to point directly into VLoHalf.
+ for (int &M : LoBlendMask)
+ if (M >= 0)
+ M /= 2;
+ for (int &M : HiBlendMask)
+ if (M >= 0)
+ M /= 2;
+ } else {
+ // Otherwise just unpack the low half of V into VLoHalf and the high half into
+ // VHiHalf so that we can blend them as i16s.
+ VLoHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
+ DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
+ VHiHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
+ DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
+ }
+
+ SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
+ SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
+
+ return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
}
-/// \brief Detect whether the mask pattern should be lowered through
-/// interleaving.
+/// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
///
-/// This essentially tests whether viewing the mask as an interleaving of two
-/// sub-sequences reduces the cross-input traffic of a blend operation. If so,
-/// lowering it through interleaving is a significantly better strategy.
-static bool shouldLowerAsInterleaving(ArrayRef<int> Mask) {
- int NumEvenInputs[2] = {0, 0};
- int NumOddInputs[2] = {0, 0};
- int NumLoInputs[2] = {0, 0};
- int NumHiInputs[2] = {0, 0};
- for (int i = 0, Size = Mask.size(); i < Size; ++i) {
- if (Mask[i] < 0)
+/// This routine breaks down the specific type of 128-bit shuffle and
+/// dispatches to the lowering routines accordingly.
+static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ MVT VT, const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ switch (VT.SimpleTy) {
+ case MVT::v2i64:
+ return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
+ case MVT::v2f64:
+ return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
+ case MVT::v4i32:
+ return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
+ case MVT::v4f32:
+ return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
+ case MVT::v8i16:
+ return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
+ case MVT::v16i8:
+ return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
+
+ default:
+ llvm_unreachable("Unimplemented!");
+ }
+}
+
+/// \brief Helper function to test whether a shuffle mask could be
+/// simplified by widening the elements being shuffled.
+///
+/// Appends the mask for wider elements in WidenedMask if valid. Otherwise
+/// leaves it in an unspecified state.
+///
+/// NOTE: This must handle normal vector shuffle masks and *target* vector
+/// shuffle masks. The latter have the special property of a '-2' representing
+/// a zero-ed lane of a vector.
+static bool canWidenShuffleElements(ArrayRef<int> Mask,
+ SmallVectorImpl<int> &WidenedMask) {
+ for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
+ // If both elements are undef, its trivial.
+ if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
+ WidenedMask.push_back(SM_SentinelUndef);
+ continue;
+ }
+
+ // Check for an undef mask and a mask value properly aligned to fit with
+ // a pair of values. If we find such a case, use the non-undef mask's value.
+ if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
+ WidenedMask.push_back(Mask[i + 1] / 2);
+ continue;
+ }
+ if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
+ WidenedMask.push_back(Mask[i] / 2);
continue;
+ }
- int InputIdx = Mask[i] >= Size;
+ // When zeroing, we need to spread the zeroing across both lanes to widen.
+ if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
+ if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
+ (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
+ WidenedMask.push_back(SM_SentinelZero);
+ continue;
+ }
+ return false;
+ }
- if (i < Size / 2)
- ++NumLoInputs[InputIdx];
- else
- ++NumHiInputs[InputIdx];
+ // Finally check if the two mask values are adjacent and aligned with
+ // a pair.
+ if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
+ WidenedMask.push_back(Mask[i] / 2);
+ continue;
+ }
- if ((i % 2) == 0)
- ++NumEvenInputs[InputIdx];
- else
- ++NumOddInputs[InputIdx];
+ // Otherwise we can't safely widen the elements used in this shuffle.
+ return false;
}
+ assert(WidenedMask.size() == Mask.size() / 2 &&
+ "Incorrect size of mask after widening the elements!");
- // The minimum number of cross-input results for both the interleaved and
- // split cases. If interleaving results in fewer cross-input results, return
- // true.
- int InterleavedCrosses = std::min(NumEvenInputs[1] + NumOddInputs[0],
- NumEvenInputs[0] + NumOddInputs[1]);
- int SplitCrosses = std::min(NumLoInputs[1] + NumHiInputs[0],
- NumLoInputs[0] + NumHiInputs[1]);
- return InterleavedCrosses < SplitCrosses;
+ return true;
}
-/// \brief Blend two v8i16 vectors using a naive unpack strategy.
+/// \brief Generic routine to split vector shuffle into half-sized shuffles.
///
-/// This strategy only works when the inputs from each vector fit into a single
-/// half of that vector, and generally there are not so many inputs as to leave
-/// the in-place shuffles required highly constrained (and thus expensive). It
-/// shifts all the inputs into a single side of both input vectors and then
-/// uses an unpack to interleave these inputs in a single vector. At that
-/// point, we will fall back on the generic single input shuffle lowering.
-static SDValue lowerV8I16BasicBlendVectorShuffle(SDLoc DL, SDValue V1,
- SDValue V2,
- MutableArrayRef<int> Mask,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
- assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
- SmallVector<int, 3> LoV1Inputs, HiV1Inputs, LoV2Inputs, HiV2Inputs;
- for (int i = 0; i < 8; ++i)
- if (Mask[i] >= 0 && Mask[i] < 4)
- LoV1Inputs.push_back(i);
- else if (Mask[i] >= 4 && Mask[i] < 8)
- HiV1Inputs.push_back(i);
- else if (Mask[i] >= 8 && Mask[i] < 12)
- LoV2Inputs.push_back(i);
- else if (Mask[i] >= 12)
- HiV2Inputs.push_back(i);
-
- int NumV1Inputs = LoV1Inputs.size() + HiV1Inputs.size();
- int NumV2Inputs = LoV2Inputs.size() + HiV2Inputs.size();
- (void)NumV1Inputs;
- (void)NumV2Inputs;
- assert(NumV1Inputs > 0 && NumV1Inputs <= 3 && "At most 3 inputs supported");
- assert(NumV2Inputs > 0 && NumV2Inputs <= 3 && "At most 3 inputs supported");
- assert(NumV1Inputs + NumV2Inputs <= 4 && "At most 4 combined inputs");
-
- bool MergeFromLo = LoV1Inputs.size() + LoV2Inputs.size() >=
- HiV1Inputs.size() + HiV2Inputs.size();
-
- auto moveInputsToHalf = [&](SDValue V, ArrayRef<int> LoInputs,
- ArrayRef<int> HiInputs, bool MoveToLo,
- int MaskOffset) {
- ArrayRef<int> GoodInputs = MoveToLo ? LoInputs : HiInputs;
- ArrayRef<int> BadInputs = MoveToLo ? HiInputs : LoInputs;
- if (BadInputs.empty())
- return V;
+/// This routine just extracts two subvectors, shuffles them independently, and
+/// then concatenates them back together. This should work effectively with all
+/// AVX vector shuffle types.
+static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
+ SDValue V2, ArrayRef<int> Mask,
+ SelectionDAG &DAG) {
+ assert(VT.getSizeInBits() >= 256 &&
+ "Only for 256-bit or wider vector shuffles!");
+ assert(V1.getSimpleValueType() == VT && "Bad operand type!");
+ assert(V2.getSimpleValueType() == VT && "Bad operand type!");
- int MoveMask[] = {-1, -1, -1, -1, -1, -1, -1, -1};
- int MoveOffset = MoveToLo ? 0 : 4;
+ ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
+ ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
- if (GoodInputs.empty()) {
- for (int BadInput : BadInputs) {
- MoveMask[Mask[BadInput] % 4 + MoveOffset] = Mask[BadInput] - MaskOffset;
- Mask[BadInput] = Mask[BadInput] % 4 + MoveOffset + MaskOffset;
- }
+ int NumElements = VT.getVectorNumElements();
+ int SplitNumElements = NumElements / 2;
+ MVT ScalarVT = VT.getScalarType();
+ MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
+
+ // Rather than splitting build-vectors, just build two narrower build
+ // vectors. This helps shuffling with splats and zeros.
+ auto SplitVector = [&](SDValue V) {
+ while (V.getOpcode() == ISD::BITCAST)
+ V = V->getOperand(0);
+
+ MVT OrigVT = V.getSimpleValueType();
+ int OrigNumElements = OrigVT.getVectorNumElements();
+ int OrigSplitNumElements = OrigNumElements / 2;
+ MVT OrigScalarVT = OrigVT.getScalarType();
+ MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
+
+ SDValue LoV, HiV;
+
+ auto *BV = dyn_cast<BuildVectorSDNode>(V);
+ if (!BV) {
+ LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
+ DAG.getIntPtrConstant(0));
+ HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
+ DAG.getIntPtrConstant(OrigSplitNumElements));
} else {
- if (GoodInputs.size() == 2) {
- // If the low inputs are spread across two dwords, pack them into
- // a single dword.
- 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 good inputs.
- for (int GoodInput : GoodInputs)
- MoveMask[Mask[GoodInput] - MaskOffset] = Mask[GoodInput] - MaskOffset;
- }
- 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]] - 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;
+ SmallVector<SDValue, 16> LoOps, HiOps;
+ for (int i = 0; i < OrigSplitNumElements; ++i) {
+ LoOps.push_back(BV->getOperand(i));
+ HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
}
+ LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
+ HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
}
-
- return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16),
- MoveMask);
+ return std::make_pair(DAG.getNode(ISD::BITCAST, DL, SplitVT, LoV),
+ DAG.getNode(ISD::BITCAST, DL, SplitVT, HiV));
};
- V1 = moveInputsToHalf(V1, LoV1Inputs, HiV1Inputs, MergeFromLo,
- /*MaskOffset*/ 0);
- V2 = moveInputsToHalf(V2, LoV2Inputs, HiV2Inputs, MergeFromLo,
- /*MaskOffset*/ 8);
-
- // FIXME: Select an interleaving of the merge of V1 and V2 that minimizes
- // cross-half traffic in the final shuffle.
- // Munge the mask to be a single-input mask after the unpack merges the
- // results.
- for (int &M : Mask)
- if (M != -1)
- M = 2 * (M % 4) + (M / 8);
-
- return DAG.getVectorShuffle(
- MVT::v8i16, DL, DAG.getNode(MergeFromLo ? X86ISD::UNPCKL : X86ISD::UNPCKH,
- DL, MVT::v8i16, V1, V2),
- DAG.getUNDEF(MVT::v8i16), Mask);
-}
+ SDValue LoV1, HiV1, LoV2, HiV2;
+ std::tie(LoV1, HiV1) = SplitVector(V1);
+ std::tie(LoV2, HiV2) = SplitVector(V2);
-/// \brief Generic lowering of 8-lane i16 shuffles.
-///
-/// This handles both single-input shuffles and combined shuffle/blends with
-/// two inputs. The single input shuffles are immediately delegated to
-/// a dedicated lowering routine.
-///
-/// The blends are lowered in one of three fundamental ways. If there are few
-/// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
-/// of the input is significantly cheaper when lowered as an interleaving of
-/// the two inputs, try to interleave them. Otherwise, blend the low and high
-/// halves of the inputs separately (making them have relatively few inputs)
-/// and then concatenate them.
-static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- SDLoc DL(Op);
- assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
- assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
- assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> OrigMask = SVOp->getMask();
- int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
- OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
- MutableArrayRef<int> Mask(MaskStorage);
-
- assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
-
- // Whenever we can lower this as a zext, that instruction is strictly faster
- // than any alternative.
- if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
- DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
- return ZExt;
-
- auto isV1 = [](int M) { return M >= 0 && M < 8; };
- auto isV2 = [](int M) { return M >= 8; };
-
- int NumV1Inputs = std::count_if(Mask.begin(), Mask.end(), isV1);
- int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
-
- if (NumV2Inputs == 0)
- return lowerV8I16SingleInputVectorShuffle(DL, V1, Mask, Subtarget, DAG);
-
- assert(NumV1Inputs > 0 && "All single-input shuffles should be canonicalized "
- "to be V1-input shuffles.");
-
- // Try to use byte shift instructions.
- if (SDValue Shift = lowerVectorShuffleAsByteShift(
- DL, MVT::v8i16, V1, V2, Mask, DAG))
- return Shift;
-
- // There are special ways we can lower some single-element blends.
- if (NumV2Inputs == 1)
- if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v8i16, DL, V1, V2,
- Mask, Subtarget, DAG))
- return V;
-
- // Use dedicated unpack instructions for masks that match their pattern.
- if (isShuffleEquivalent(Mask, 0, 8, 1, 9, 2, 10, 3, 11))
- return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
- if (isShuffleEquivalent(Mask, 4, 12, 5, 13, 6, 14, 7, 15))
- return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
-
- if (Subtarget->hasSSE41())
- if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
- Subtarget, DAG))
- return Blend;
-
- // Try to use byte rotation instructions.
- if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
- DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
- return Rotate;
-
- if (NumV1Inputs + NumV2Inputs <= 4)
- return lowerV8I16BasicBlendVectorShuffle(DL, V1, V2, Mask, Subtarget, DAG);
-
- // Check whether an interleaving lowering is likely to be more efficient.
- // This isn't perfect but it is a strong heuristic that tends to work well on
- // the kinds of shuffles that show up in practice.
- //
- // FIXME: Handle 1x, 2x, and 4x interleaving.
- if (shouldLowerAsInterleaving(Mask)) {
- // FIXME: Figure out whether we should pack these into the low or high
- // halves.
-
- int EMask[8], OMask[8];
- for (int i = 0; i < 4; ++i) {
- EMask[i] = Mask[2*i];
- OMask[i] = Mask[2*i + 1];
- EMask[i + 4] = -1;
- OMask[i + 4] = -1;
+ // Now create two 4-way blends of these half-width vectors.
+ auto HalfBlend = [&](ArrayRef<int> HalfMask) {
+ bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
+ SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
+ for (int i = 0; i < SplitNumElements; ++i) {
+ int M = HalfMask[i];
+ if (M >= NumElements) {
+ if (M >= NumElements + SplitNumElements)
+ UseHiV2 = true;
+ else
+ UseLoV2 = true;
+ V2BlendMask.push_back(M - NumElements);
+ V1BlendMask.push_back(-1);
+ BlendMask.push_back(SplitNumElements + i);
+ } else if (M >= 0) {
+ if (M >= SplitNumElements)
+ UseHiV1 = true;
+ else
+ UseLoV1 = true;
+ 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 Evens = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, EMask);
- SDValue Odds = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, OMask);
-
- return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, Evens, Odds);
- }
-
- int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
- int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
-
- for (int i = 0; i < 4; ++i) {
- LoBlendMask[i] = Mask[i];
- HiBlendMask[i] = Mask[i + 4];
- }
+ // Because the lowering happens after all combining takes place, we need to
+ // manually combine these blend masks as much as possible so that we create
+ // a minimal number of high-level vector shuffle nodes.
- SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, LoBlendMask);
- SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, HiBlendMask);
- LoV = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, LoV);
- HiV = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, HiV);
+ // First try just blending the halves of V1 or V2.
+ if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
+ return DAG.getUNDEF(SplitVT);
+ if (!UseLoV2 && !UseHiV2)
+ return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
+ if (!UseLoV1 && !UseHiV1)
+ return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
- return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
- DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, LoV, HiV));
+ SDValue V1Blend, V2Blend;
+ if (UseLoV1 && UseHiV1) {
+ V1Blend =
+ DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
+ } else {
+ // We only use half of V1 so map the usage down into the final blend mask.
+ V1Blend = UseLoV1 ? LoV1 : HiV1;
+ for (int i = 0; i < SplitNumElements; ++i)
+ if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
+ BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
+ }
+ if (UseLoV2 && UseHiV2) {
+ V2Blend =
+ DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
+ } else {
+ // We only use half of V2 so map the usage down into the final blend mask.
+ V2Blend = UseLoV2 ? LoV2 : HiV2;
+ for (int i = 0; i < SplitNumElements; ++i)
+ if (BlendMask[i] >= SplitNumElements)
+ BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
+ }
+ 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 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.
+/// \brief Either split a vector in halves or decompose the shuffles and the
+/// blend.
///
-/// \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;
+/// This is provided as a good fallback for many lowerings of non-single-input
+/// shuffles with more than one 128-bit lane. In those cases, we want to select
+/// between splitting the shuffle into 128-bit components and stitching those
+/// back together vs. extracting the single-input shuffles and blending those
+/// results.
+static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
+ SDValue V2, ArrayRef<int> Mask,
+ SelectionDAG &DAG) {
+ assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
+ "lower single-input shuffles as it "
+ "could then recurse on itself.");
+ int Size = Mask.size();
- // 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;
+ // If this can be modeled as a broadcast of two elements followed by a blend,
+ // prefer that lowering. This is especially important because broadcasts can
+ // often fold with memory operands.
+ auto DoBothBroadcast = [&] {
+ int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
+ for (int M : Mask)
+ if (M >= Size) {
+ if (V2BroadcastIdx == -1)
+ V2BroadcastIdx = M - Size;
+ else if (M - Size != V2BroadcastIdx)
+ return false;
+ } else if (M >= 0) {
+ if (V1BroadcastIdx == -1)
+ V1BroadcastIdx = M;
+ else if (M != V1BroadcastIdx)
+ return false;
}
- // Early exit if we exhaust the possible powers of two.
- if (!IsAnyViable)
- break;
- }
+ return true;
+ };
+ if (DoBothBroadcast())
+ return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
+ DAG);
- for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
- if (ViableForN[j])
- return j + 1;
+ // If the inputs all stem from a single 128-bit lane of each input, then we
+ // split them rather than blending because the split will decompose to
+ // unusually few instructions.
+ int LaneCount = VT.getSizeInBits() / 128;
+ int LaneSize = Size / LaneCount;
+ SmallBitVector LaneInputs[2];
+ LaneInputs[0].resize(LaneCount, false);
+ LaneInputs[1].resize(LaneCount, false);
+ for (int i = 0; i < Size; ++i)
+ if (Mask[i] >= 0)
+ LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
+ if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
+ return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
- // Return 0 as there is no viable power of two.
- return 0;
+ // Otherwise, just fall back to decomposed shuffles and a blend. This requires
+ // that the decomposed single-input shuffles don't end up here.
+ return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
}
-/// \brief Generic lowering of v16i8 shuffles.
+/// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
+/// a permutation and blend of those lanes.
///
-/// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
-/// detect any complexity reducing interleaving. If that doesn't help, it uses
-/// UNPCK to spread the i8 elements across two i16-element vectors, and uses
-/// the existing lowering for v8i16 blends on each half, finally PACK-ing them
-/// back together.
-static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- SDLoc DL(Op);
- assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
- assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
- assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> OrigMask = SVOp->getMask();
- assert(OrigMask.size() == 16 && "Unexpected mask size for v16 shuffle!");
-
- // Try to use byte shift instructions.
- if (SDValue Shift = lowerVectorShuffleAsByteShift(
- DL, MVT::v16i8, V1, V2, OrigMask, DAG))
- return Shift;
-
- // Try to use byte rotation instructions.
- if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
- DL, MVT::v16i8, V1, V2, OrigMask, Subtarget, DAG))
- return Rotate;
-
- // Try to use a zext lowering.
- if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
- DL, MVT::v16i8, V1, V2, OrigMask, Subtarget, DAG))
- return ZExt;
-
- int MaskStorage[16] = {
- OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
- OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7],
- OrigMask[8], OrigMask[9], OrigMask[10], OrigMask[11],
- OrigMask[12], OrigMask[13], OrigMask[14], OrigMask[15]};
- MutableArrayRef<int> Mask(MaskStorage);
- 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 (NumV2Elements == 0) {
- // Check for being able to broadcast a single element.
- if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v16i8, DL, V1,
- Mask, Subtarget, DAG))
- return Broadcast;
-
- // 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
- // things significantly. Currently, this means we need to be able to
- // express the pre-duplication shuffle as an i16 shuffle.
- //
- // FIXME: We should check for other patterns which can be widened into an
- // i16 shuffle as well.
- auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
- for (int i = 0; i < 16; i += 2)
- if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
- return false;
-
- return true;
- };
- auto tryToWidenViaDuplication = [&]() -> SDValue {
- if (!canWidenViaDuplication(Mask))
- return SDValue();
- SmallVector<int, 4> LoInputs;
- std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
- [](int M) { return M >= 0 && M < 8; });
- std::sort(LoInputs.begin(), LoInputs.end());
- LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
- LoInputs.end());
- SmallVector<int, 4> HiInputs;
- std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
- [](int M) { return M >= 8; });
- std::sort(HiInputs.begin(), HiInputs.end());
- HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
- HiInputs.end());
-
- bool TargetLo = LoInputs.size() >= HiInputs.size();
- ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
- ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
-
- int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
- SmallDenseMap<int, int, 8> LaneMap;
- for (int I : InPlaceInputs) {
- PreDupI16Shuffle[I/2] = I/2;
- LaneMap[I] = I;
- }
- int j = TargetLo ? 0 : 4, je = j + 4;
- for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
- // Check if j is already a shuffle of this input. This happens when
- // there are two adjacent bytes after we move the low one.
- if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
- // If we haven't yet mapped the input, search for a slot into which
- // we can map it.
- while (j < je && PreDupI16Shuffle[j] != -1)
- ++j;
-
- if (j == je)
- // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
- return SDValue();
-
- // Map this input with the i16 shuffle.
- PreDupI16Shuffle[j] = MovingInputs[i] / 2;
- }
-
- // Update the lane map based on the mapping we ended up with.
- LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
- }
- V1 = DAG.getNode(
- ISD::BITCAST, DL, MVT::v16i8,
- DAG.getVectorShuffle(MVT::v8i16, DL,
- DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
- DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
-
- // Unpack the bytes to form the i16s that will be shuffled into place.
- V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
- MVT::v16i8, V1, V1);
-
- int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
- for (int i = 0; i < 16; ++i)
- if (Mask[i] != -1) {
- int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
- assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
- if (PostDupI16Shuffle[i / 2] == -1)
- PostDupI16Shuffle[i / 2] = MappedMask;
- else
- assert(PostDupI16Shuffle[i / 2] == MappedMask &&
- "Conflicting entrties in the original shuffle!");
- }
- return DAG.getNode(
- ISD::BITCAST, DL, MVT::v16i8,
- DAG.getVectorShuffle(MVT::v8i16, DL,
- DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
- DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
- };
- if (SDValue V = tryToWidenViaDuplication())
- return V;
- }
-
- // Check whether an interleaving lowering is likely to be more efficient.
- // This isn't perfect but it is a strong heuristic that tends to work well on
- // the kinds of shuffles that show up in practice.
- //
- // FIXME: We need to handle other interleaving widths (i16, i32, ...).
- if (shouldLowerAsInterleaving(Mask)) {
- int NumLoHalf = std::count_if(Mask.begin(), Mask.end(), [](int M) {
- return (M >= 0 && M < 8) || (M >= 16 && M < 24);
- });
- int NumHiHalf = std::count_if(Mask.begin(), Mask.end(), [](int M) {
- return (M >= 8 && M < 16) || M >= 24;
- });
- int EMask[16] = {-1, -1, -1, -1, -1, -1, -1, -1,
- -1, -1, -1, -1, -1, -1, -1, -1};
- int OMask[16] = {-1, -1, -1, -1, -1, -1, -1, -1,
- -1, -1, -1, -1, -1, -1, -1, -1};
- bool UnpackLo = NumLoHalf >= NumHiHalf;
- MutableArrayRef<int> TargetEMask(UnpackLo ? EMask : EMask + 8, 8);
- MutableArrayRef<int> TargetOMask(UnpackLo ? OMask : OMask + 8, 8);
- for (int i = 0; i < 8; ++i) {
- TargetEMask[i] = Mask[2 * i];
- TargetOMask[i] = Mask[2 * i + 1];
- }
-
- SDValue Evens = DAG.getVectorShuffle(MVT::v16i8, DL, V1, V2, EMask);
- SDValue Odds = DAG.getVectorShuffle(MVT::v16i8, DL, V1, V2, OMask);
-
- return DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, 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.getUNDEF(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 = lowerVectorShuffleAsElementInsertion(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};
- int V2HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
-
- auto buildBlendMasks = [](MutableArrayRef<int> HalfMask,
- MutableArrayRef<int> V1HalfBlendMask,
- MutableArrayRef<int> V2HalfBlendMask) {
- for (int i = 0; i < 8; ++i)
- if (HalfMask[i] >= 0 && HalfMask[i] < 16) {
- V1HalfBlendMask[i] = HalfMask[i];
- HalfMask[i] = i;
- } else if (HalfMask[i] >= 16) {
- V2HalfBlendMask[i] = HalfMask[i] - 16;
- HalfMask[i] = i + 8;
- }
- };
- buildBlendMasks(LoMask, V1LoBlendMask, V2LoBlendMask);
- buildBlendMasks(HiMask, V1HiBlendMask, V2HiBlendMask);
-
- SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
-
- auto buildLoAndHiV8s = [&](SDValue V, MutableArrayRef<int> LoBlendMask,
- MutableArrayRef<int> HiBlendMask) {
- SDValue V1, V2;
- // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
- // them out and avoid using UNPCK{L,H} to extract the elements of V as
- // i16s.
- if (std::none_of(LoBlendMask.begin(), LoBlendMask.end(),
- [](int M) { return M >= 0 && M % 2 == 1; }) &&
- std::none_of(HiBlendMask.begin(), HiBlendMask.end(),
- [](int M) { return M >= 0 && M % 2 == 1; })) {
- // Use a mask to drop the high bytes.
- V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
- V1 = DAG.getNode(ISD::AND, DL, MVT::v8i16, V1,
- DAG.getConstant(0x00FF, MVT::v8i16));
-
- // This will be a single vector shuffle instead of a blend so nuke V2.
- V2 = DAG.getUNDEF(MVT::v8i16);
-
- // Squash the masks to point directly into V1.
- for (int &M : LoBlendMask)
- if (M >= 0)
- M /= 2;
- for (int &M : HiBlendMask)
- if (M >= 0)
- M /= 2;
- } else {
- // Otherwise just unpack the low half of V into V1 and the high half into
- // V2 so that we can blend them as i16s.
- V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
- DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
- V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
- DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
- }
-
- SDValue BlendedLo = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, LoBlendMask);
- SDValue BlendedHi = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, HiBlendMask);
- return std::make_pair(BlendedLo, BlendedHi);
- };
- SDValue V1Lo, V1Hi, V2Lo, V2Hi;
- std::tie(V1Lo, V1Hi) = buildLoAndHiV8s(V1, V1LoBlendMask, V1HiBlendMask);
- std::tie(V2Lo, V2Hi) = buildLoAndHiV8s(V2, V2LoBlendMask, V2HiBlendMask);
-
- SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, V1Lo, V2Lo, LoMask);
- SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, V1Hi, V2Hi, HiMask);
-
- return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
-}
-
-/// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
-///
-/// This routine breaks down the specific type of 128-bit shuffle and
-/// dispatches to the lowering routines accordingly.
-static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- MVT VT, const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- switch (VT.SimpleTy) {
- case MVT::v2i64:
- return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
- case MVT::v2f64:
- return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
- case MVT::v4i32:
- return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
- case MVT::v4f32:
- return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
- case MVT::v8i16:
- return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
- case MVT::v16i8:
- return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
-
- default:
- llvm_unreachable("Unimplemented!");
- }
-}
-
-/// \brief Helper function to test whether a shuffle mask could be
-/// simplified by widening the elements being shuffled.
-///
-/// Appends the mask for wider elements in WidenedMask if valid. Otherwise
-/// leaves it in an unspecified state.
-///
-/// NOTE: This must handle normal vector shuffle masks and *target* vector
-/// shuffle masks. The latter have the special property of a '-2' representing
-/// a zero-ed lane of a vector.
-static bool canWidenShuffleElements(ArrayRef<int> Mask,
- SmallVectorImpl<int> &WidenedMask) {
- for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
- // If both elements are undef, its trivial.
- if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
- WidenedMask.push_back(SM_SentinelUndef);
- continue;
- }
-
- // Check for an undef mask and a mask value properly aligned to fit with
- // a pair of values. If we find such a case, use the non-undef mask's value.
- if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
- WidenedMask.push_back(Mask[i + 1] / 2);
- continue;
- }
- if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
- WidenedMask.push_back(Mask[i] / 2);
- continue;
- }
-
- // When zeroing, we need to spread the zeroing across both lanes to widen.
- if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
- if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
- (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
- WidenedMask.push_back(SM_SentinelZero);
- continue;
- }
- return false;
- }
-
- // Finally check if the two mask values are adjacent and aligned with
- // a pair.
- if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
- WidenedMask.push_back(Mask[i] / 2);
- continue;
- }
-
- // Otherwise we can't safely widen the elements used in this shuffle.
- return false;
- }
- assert(WidenedMask.size() == Mask.size() / 2 &&
- "Incorrect size of mask after widening the elements!");
-
- return true;
-}
-
-/// \brief Generic routine to split ector shuffle into half-sized shuffles.
-///
-/// This routine just extracts two subvectors, shuffles them independently, and
-/// then concatenates them back together. This should work effectively with all
-/// AVX vector shuffle types.
-static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
- SDValue V2, ArrayRef<int> Mask,
- SelectionDAG &DAG) {
- assert(VT.getSizeInBits() >= 256 &&
- "Only for 256-bit or wider vector shuffles!");
- assert(V1.getSimpleValueType() == VT && "Bad operand type!");
- assert(V2.getSimpleValueType() == VT && "Bad operand type!");
-
- 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) {
- bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
- SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
- for (int i = 0; i < SplitNumElements; ++i) {
- int M = HalfMask[i];
- if (M >= NumElements) {
- if (M >= NumElements + SplitNumElements)
- UseHiV2 = true;
- else
- UseLoV2 = true;
- V2BlendMask.push_back(M - NumElements);
- V1BlendMask.push_back(-1);
- BlendMask.push_back(SplitNumElements + i);
- } else if (M >= 0) {
- if (M >= SplitNumElements)
- UseHiV1 = true;
- else
- UseLoV1 = true;
- 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);
- }
- }
-
- // Because the lowering happens after all combining takes place, we need to
- // manually combine these blend masks as much as possible so that we create
- // a minimal number of high-level vector shuffle nodes.
-
- // First try just blending the halves of V1 or V2.
- if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
- return DAG.getUNDEF(SplitVT);
- if (!UseLoV2 && !UseHiV2)
- return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
- if (!UseLoV1 && !UseHiV1)
- return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
-
- SDValue V1Blend, V2Blend;
- if (UseLoV1 && UseHiV1) {
- V1Blend =
- DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
- } else {
- // We only use half of V1 so map the usage down into the final blend mask.
- V1Blend = UseLoV1 ? LoV1 : HiV1;
- for (int i = 0; i < SplitNumElements; ++i)
- if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
- BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
- }
- if (UseLoV2 && UseHiV2) {
- V2Blend =
- DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
- } else {
- // We only use half of V2 so map the usage down into the final blend mask.
- V2Blend = UseLoV2 ? LoV2 : HiV2;
- for (int i = 0; i < SplitNumElements; ++i)
- if (BlendMask[i] >= SplitNumElements)
- BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
- }
- 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 Either split a vector in halves or decompose the shuffles and the
-/// blend.
-///
-/// This is provided as a good fallback for many lowerings of non-single-input
-/// shuffles with more than one 128-bit lane. In those cases, we want to select
-/// between splitting the shuffle into 128-bit components and stitching those
-/// back together vs. extracting the single-input shuffles and blending those
-/// results.
-static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
- SDValue V2, ArrayRef<int> Mask,
- SelectionDAG &DAG) {
- assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
- "lower single-input shuffles as it "
- "could then recurse on itself.");
- int Size = Mask.size();
-
- // If this can be modeled as a broadcast of two elements followed by a blend,
- // prefer that lowering. This is especially important because broadcasts can
- // often fold with memory operands.
- auto DoBothBroadcast = [&] {
- int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
- for (int M : Mask)
- if (M >= Size) {
- if (V2BroadcastIdx == -1)
- V2BroadcastIdx = M - Size;
- else if (M - Size != V2BroadcastIdx)
- return false;
- } else if (M >= 0) {
- if (V1BroadcastIdx == -1)
- V1BroadcastIdx = M;
- else if (M != V1BroadcastIdx)
- return false;
- }
- return true;
- };
- if (DoBothBroadcast())
- return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
- DAG);
-
- // If the inputs all stem from a single 128-bit lane of each input, then we
- // split them rather than blending because the split will decompose to
- // unusually few instructions.
- int LaneCount = VT.getSizeInBits() / 128;
- int LaneSize = Size / LaneCount;
- SmallBitVector LaneInputs[2];
- LaneInputs[0].resize(LaneCount, false);
- LaneInputs[1].resize(LaneCount, false);
- for (int i = 0; i < Size; ++i)
- if (Mask[i] >= 0)
- LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
- if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
- return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
-
- // Otherwise, just fall back to decomposed shuffles and a blend. This requires
- // that the decomposed single-input shuffles don't end up here.
- return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
-}
-
-/// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
-/// a permutation and blend of those lanes.
-///
-/// This essentially blends the out-of-lane inputs to each lane into the lane
-/// from a permuted copy of the vector. This lowering strategy results in four
-/// instructions in the worst case for a single-input cross lane shuffle which
-/// is lower than any other fully general cross-lane shuffle strategy I'm aware
-/// of. Special cases for each particular shuffle pattern should be handled
-/// prior to trying this lowering.
-static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
- SDValue V1, SDValue V2,
- ArrayRef<int> Mask,
- SelectionDAG &DAG) {
- // FIXME: This should probably be generalized for 512-bit vectors as well.
- assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
- int LaneSize = Mask.size() / 2;
-
- // If there are only inputs from one 128-bit lane, splitting will in fact be
- // less expensive. The flags track wether the given lane contains an element
- // that crosses to another lane.
- bool LaneCrossing[2] = {false, false};
- for (int i = 0, Size = Mask.size(); i < Size; ++i)
- if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
- LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
- if (!LaneCrossing[0] || !LaneCrossing[1])
- return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
-
- if (isSingleInputShuffleMask(Mask)) {
- SmallVector<int, 32> FlippedBlendMask;
- for (int i = 0, Size = Mask.size(); i < Size; ++i)
- FlippedBlendMask.push_back(
- Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
- ? Mask[i]
- : Mask[i] % LaneSize +
- (i / LaneSize) * LaneSize + Size));
-
- // Flip the vector, and blend the results which should now be in-lane. The
- // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
- // 5 for the high source. The value 3 selects the high half of source 2 and
- // the value 2 selects the low half of source 2. We only use source 2 to
- // allow folding it into a memory operand.
- unsigned PERMMask = 3 | 2 << 4;
- SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
- V1, DAG.getConstant(PERMMask, MVT::i8));
- return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
- }
-
- // This now reduces to two single-input shuffles of V1 and V2 which at worst
- // will be handled by the above logic and a blend of the results, much like
- // other patterns in AVX.
- return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
-}
-
-/// \brief Handle lowering 2-lane 128-bit shuffles.
-static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
- SDValue V2, ArrayRef<int> Mask,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- // Blends are faster and handle all the non-lane-crossing cases.
- if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
- Subtarget, DAG))
- return Blend;
-
- MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
- VT.getVectorNumElements() / 2);
- // Check for patterns which can be matched with a single insert of a 128-bit
- // subvector.
- if (isShuffleEquivalent(Mask, 0, 1, 0, 1) ||
- isShuffleEquivalent(Mask, 0, 1, 4, 5)) {
- SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
- DAG.getIntPtrConstant(0));
- SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
- Mask[2] < 4 ? V1 : V2, DAG.getIntPtrConstant(0));
- return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
- }
- if (isShuffleEquivalent(Mask, 0, 1, 6, 7)) {
- SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
- DAG.getIntPtrConstant(0));
- SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V2,
- DAG.getIntPtrConstant(2));
- return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
- }
-
- // Otherwise form a 128-bit permutation.
- // FIXME: Detect zero-vector inputs and use the VPERM2X128 to zero that half.
- unsigned PermMask = Mask[0] / 2 | (Mask[2] / 2) << 4;
- return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
- DAG.getConstant(PermMask, MVT::i8));
-}
-
-/// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
-/// shuffling each lane.
-///
-/// This will only succeed when the result of fixing the 128-bit lanes results
-/// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
-/// each 128-bit lanes. This handles many cases where we can quickly blend away
-/// the lane crosses early and then use simpler shuffles within each lane.
-///
-/// FIXME: It might be worthwhile at some point to support this without
-/// requiring the 128-bit lane-relative shuffles to be repeating, but currently
-/// in x86 only floating point has interesting non-repeating shuffles, and even
-/// those are still *marginally* more expensive.
-static SDValue lowerVectorShuffleByMerging128BitLanes(
- SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
- const X86Subtarget *Subtarget, SelectionDAG &DAG) {
- assert(!isSingleInputShuffleMask(Mask) &&
- "This is only useful with multiple inputs.");
-
- int Size = Mask.size();
- int LaneSize = 128 / VT.getScalarSizeInBits();
- int NumLanes = Size / LaneSize;
- assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
-
- // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
- // check whether the in-128-bit lane shuffles share a repeating pattern.
- SmallVector<int, 4> Lanes;
- Lanes.resize(NumLanes, -1);
- SmallVector<int, 4> InLaneMask;
- InLaneMask.resize(LaneSize, -1);
- for (int i = 0; i < Size; ++i) {
- if (Mask[i] < 0)
- continue;
-
- int j = i / LaneSize;
-
- if (Lanes[j] < 0) {
- // First entry we've seen for this lane.
- Lanes[j] = Mask[i] / LaneSize;
- } else if (Lanes[j] != Mask[i] / LaneSize) {
- // This doesn't match the lane selected previously!
- return SDValue();
- }
-
- // Check that within each lane we have a consistent shuffle mask.
- int k = i % LaneSize;
- if (InLaneMask[k] < 0) {
- InLaneMask[k] = Mask[i] % LaneSize;
- } else if (InLaneMask[k] != Mask[i] % LaneSize) {
- // This doesn't fit a repeating in-lane mask.
- return SDValue();
- }
- }
-
- // First shuffle the lanes into place.
- MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
- VT.getSizeInBits() / 64);
- SmallVector<int, 8> LaneMask;
- LaneMask.resize(NumLanes * 2, -1);
- for (int i = 0; i < NumLanes; ++i)
- if (Lanes[i] >= 0) {
- LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
- LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
- }
-
- V1 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V1);
- V2 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V2);
- SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
-
- // Cast it back to the type we actually want.
- LaneShuffle = DAG.getNode(ISD::BITCAST, DL, VT, LaneShuffle);
-
- // Now do a simple shuffle that isn't lane crossing.
- SmallVector<int, 8> NewMask;
- NewMask.resize(Size, -1);
- for (int i = 0; i < Size; ++i)
- if (Mask[i] >= 0)
- NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
- assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
- "Must not introduce lane crosses at this point!");
-
- return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
-}
-
-/// \brief Test whether the specified input (0 or 1) is in-place blended by the
-/// given mask.
-///
-/// This returns true if the elements from a particular input are already in the
-/// slot required by the given mask and require no permutation.
-static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
- assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
- int Size = Mask.size();
- for (int i = 0; i < Size; ++i)
- if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
- return false;
-
- return true;
-}
-
-/// \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!");
-
- SmallVector<int, 4> WidenedMask;
- if (canWidenShuffleElements(Mask, WidenedMask))
- return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
- DAG);
-
- if (isSingleInputShuffleMask(Mask)) {
- // Check for being able to broadcast a single element.
- if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4f64, DL, V1,
- Mask, Subtarget, DAG))
- return Broadcast;
-
- if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, 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::VPERMILPI, DL, MVT::v4f64, V1,
- DAG.getConstant(VPERMILPMask, MVT::i8));
- }
-
- // With AVX2 we have direct support for this permutation.
- if (Subtarget->hasAVX2())
- return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
- getV4X86ShuffleImm8ForMask(Mask, DAG));
-
- // Otherwise, fall back.
- return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
- DAG);
- }
-
- // 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);
-
- // If we have a single input to the zero element, insert that into V1 if we
- // can do so cheaply.
- int NumV2Elements =
- std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
- if (NumV2Elements == 1 && Mask[0] >= 4)
- if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
- MVT::v4f64, DL, V1, V2, Mask, Subtarget, DAG))
- return Insertion;
-
- if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
- Subtarget, DAG))
- return Blend;
-
- // Check if the blend happens to exactly fit that of SHUFPD.
- if ((Mask[0] == -1 || Mask[0] < 2) &&
- (Mask[1] == -1 || (Mask[1] >= 4 && Mask[1] < 6)) &&
- (Mask[2] == -1 || (Mask[2] >= 2 && Mask[2] < 4)) &&
- (Mask[3] == -1 || Mask[3] >= 6)) {
- 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[0] < 6)) &&
- (Mask[1] == -1 || Mask[1] < 2) &&
- (Mask[2] == -1 || Mask[2] >= 6) &&
- (Mask[3] == -1 || (Mask[3] >= 2 && 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));
- }
-
- // Try to simplify this by merging 128-bit lanes to enable a lane-based
- // shuffle. However, if we have AVX2 and either inputs are already in place,
- // we will be able to shuffle even across lanes the other input in a single
- // instruction so skip this pattern.
- if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
- isShuffleMaskInputInPlace(1, Mask))))
- if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
- DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
- return Result;
-
- // If we have AVX2 then we always want to lower with a blend because an v4 we
- // can fully permute the elements.
- if (Subtarget->hasAVX2())
- return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
- Mask, DAG);
-
- // Otherwise fall back on generic lowering.
- return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
-}
-
-/// \brief Handle lowering of 4-lane 64-bit integer shuffles.
-///
-/// This routine is only called when we have AVX2 and thus a reasonable
-/// instruction set for v4i64 shuffling..
-static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- SDLoc DL(Op);
- 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!");
- assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
-
- SmallVector<int, 4> WidenedMask;
- if (canWidenShuffleElements(Mask, WidenedMask))
- return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
- DAG);
-
- if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
- Subtarget, DAG))
- return Blend;
-
- // Check for being able to broadcast a single element.
- if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4i64, DL, V1,
- Mask, Subtarget, DAG))
- return Broadcast;
-
- // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
- // use lower latency instructions that will operate on both 128-bit lanes.
- SmallVector<int, 2> RepeatedMask;
- if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
- if (isSingleInputShuffleMask(Mask)) {
- int PSHUFDMask[] = {-1, -1, -1, -1};
- for (int i = 0; i < 2; ++i)
- if (RepeatedMask[i] >= 0) {
- PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
- PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
- }
- return DAG.getNode(
- ISD::BITCAST, DL, MVT::v4i64,
- DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
- DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, V1),
- getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
- }
-
- // Use dedicated unpack instructions for masks that match their pattern.
- if (isShuffleEquivalent(Mask, 0, 4, 2, 6))
- return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
- if (isShuffleEquivalent(Mask, 1, 5, 3, 7))
- return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
- }
-
- // AVX2 provides a direct instruction for permuting a single input across
- // lanes.
- if (isSingleInputShuffleMask(Mask))
- return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
- getV4X86ShuffleImm8ForMask(Mask, DAG));
-
- // Try to simplify this by merging 128-bit lanes to enable a lane-based
- // shuffle. However, if we have AVX2 and either inputs are already in place,
- // we will be able to shuffle even across lanes the other input in a single
- // instruction so skip this pattern.
- if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
- isShuffleMaskInputInPlace(1, Mask))))
- if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
- DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
- return Result;
-
- // Otherwise fall back on generic blend lowering.
- return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
- Mask, DAG);
-}
-
-/// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
-///
-/// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
-/// isn't available.
-static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- SDLoc DL(Op);
- assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
- assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> Mask = SVOp->getMask();
- assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
-
- if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
- Subtarget, DAG))
- return Blend;
-
- // Check for being able to broadcast a single element.
- if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v8f32, DL, V1,
- Mask, Subtarget, DAG))
- return Broadcast;
-
- // If the shuffle mask is repeated in each 128-bit lane, we have many more
- // options to efficiently lower the shuffle.
- SmallVector<int, 4> RepeatedMask;
- if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
- assert(RepeatedMask.size() == 4 &&
- "Repeated masks must be half the mask width!");
- if (isSingleInputShuffleMask(Mask))
- return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
- getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
-
- // Use dedicated unpack instructions for masks that match their pattern.
- if (isShuffleEquivalent(Mask, 0, 8, 1, 9, 4, 12, 5, 13))
- return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
- if (isShuffleEquivalent(Mask, 2, 10, 3, 11, 6, 14, 7, 15))
- return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
-
- // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
- // have already handled any direct blends. We also need to squash the
- // repeated mask into a simulated v4f32 mask.
- for (int i = 0; i < 4; ++i)
- if (RepeatedMask[i] >= 8)
- RepeatedMask[i] -= 4;
- return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
- }
-
- // If we have a single input shuffle with different shuffle patterns in the
- // two 128-bit lanes use the variable mask to VPERMILPS.
- if (isSingleInputShuffleMask(Mask)) {
- SDValue VPermMask[8];
- for (int i = 0; i < 8; ++i)
- VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
- : DAG.getConstant(Mask[i], MVT::i32);
- if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
- return DAG.getNode(
- X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
- DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
-
- if (Subtarget->hasAVX2())
- return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8f32,
- DAG.getNode(ISD::BITCAST, DL, MVT::v8f32,
- DAG.getNode(ISD::BUILD_VECTOR, DL,
- MVT::v8i32, VPermMask)),
- V1);
-
- // Otherwise, fall back.
- return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
- DAG);
- }
-
- // Try to simplify this by merging 128-bit lanes to enable a lane-based
- // shuffle.
- if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
- DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
- return Result;
-
- // If we have AVX2 then we always want to lower with a blend because at v8 we
- // can fully permute the elements.
- if (Subtarget->hasAVX2())
- return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
- Mask, DAG);
-
- // Otherwise fall back on generic lowering.
- return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
-}
-
-/// \brief Handle lowering of 8-lane 32-bit integer shuffles.
-///
-/// This routine is only called when we have AVX2 and thus a reasonable
-/// instruction set for v8i32 shuffling..
-static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- SDLoc DL(Op);
- assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
- assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> Mask = SVOp->getMask();
- assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
- assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
-
- if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
- Subtarget, DAG))
- return Blend;
-
- // Check for being able to broadcast a single element.
- if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v8i32, DL, V1,
- Mask, Subtarget, DAG))
- return Broadcast;
-
- // If the shuffle mask is repeated in each 128-bit lane we can use more
- // efficient instructions that mirror the shuffles across the two 128-bit
- // lanes.
- SmallVector<int, 4> RepeatedMask;
- if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
- assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
- if (isSingleInputShuffleMask(Mask))
- return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
- getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
-
- // Use dedicated unpack instructions for masks that match their pattern.
- if (isShuffleEquivalent(Mask, 0, 8, 1, 9, 4, 12, 5, 13))
- return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
- if (isShuffleEquivalent(Mask, 2, 10, 3, 11, 6, 14, 7, 15))
- return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
- }
-
- // If the shuffle patterns aren't repeated but it is a single input, directly
- // generate a cross-lane VPERMD instruction.
- if (isSingleInputShuffleMask(Mask)) {
- SDValue VPermMask[8];
- for (int i = 0; i < 8; ++i)
- VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
- : DAG.getConstant(Mask[i], MVT::i32);
- return DAG.getNode(
- X86ISD::VPERMV, DL, MVT::v8i32,
- DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
- }
-
- // Try to simplify this by merging 128-bit lanes to enable a lane-based
- // shuffle.
- if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
- DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
- return Result;
-
- // Otherwise fall back on generic blend lowering.
- return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
- Mask, DAG);
-}
-
-/// \brief Handle lowering of 16-lane 16-bit integer shuffles.
-///
-/// This routine is only called when we have AVX2 and thus a reasonable
-/// instruction set for v16i16 shuffling..
-static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- SDLoc DL(Op);
- assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
- assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> Mask = SVOp->getMask();
- assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
- assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
-
- // Check for being able to broadcast a single element.
- if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v16i16, DL, V1,
- Mask, Subtarget, DAG))
- return Broadcast;
-
- if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
- Subtarget, DAG))
- return Blend;
-
- // Use dedicated unpack instructions for masks that match their pattern.
- if (isShuffleEquivalent(Mask,
- // First 128-bit lane:
- 0, 16, 1, 17, 2, 18, 3, 19,
- // Second 128-bit lane:
- 8, 24, 9, 25, 10, 26, 11, 27))
- return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
- if (isShuffleEquivalent(Mask,
- // First 128-bit lane:
- 4, 20, 5, 21, 6, 22, 7, 23,
- // Second 128-bit lane:
- 12, 28, 13, 29, 14, 30, 15, 31))
- return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
-
- if (isSingleInputShuffleMask(Mask)) {
- // There are no generalized cross-lane shuffle operations available on i16
- // element types.
- if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
- return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
- Mask, DAG);
-
- SDValue PSHUFBMask[32];
- for (int i = 0; i < 16; ++i) {
- if (Mask[i] == -1) {
- PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
- continue;
- }
-
- int M = i < 8 ? Mask[i] : Mask[i] - 8;
- assert(M >= 0 && M < 8 && "Invalid single-input mask!");
- PSHUFBMask[2 * i] = DAG.getConstant(2 * M, MVT::i8);
- PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, MVT::i8);
- }
- return DAG.getNode(
- ISD::BITCAST, DL, MVT::v16i16,
- DAG.getNode(
- X86ISD::PSHUFB, DL, MVT::v32i8,
- DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1),
- DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask)));
- }
-
- // Try to simplify this by merging 128-bit lanes to enable a lane-based
- // shuffle.
- if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
- DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
- return Result;
-
- // Otherwise fall back on generic lowering.
- return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
-}
-
-/// \brief Handle lowering of 32-lane 8-bit integer shuffles.
-///
-/// This routine is only called when we have AVX2 and thus a reasonable
-/// instruction set for v32i8 shuffling..
-static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- SDLoc DL(Op);
- assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
- assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> Mask = SVOp->getMask();
- assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
- assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
-
- // Check for being able to broadcast a single element.
- if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v32i8, DL, V1,
- Mask, Subtarget, DAG))
- return Broadcast;
-
- if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
- Subtarget, DAG))
- return Blend;
-
- // Use dedicated unpack instructions for masks that match their pattern.
- // Note that these are repeated 128-bit lane unpacks, not unpacks across all
- // 256-bit lanes.
- if (isShuffleEquivalent(
- Mask,
- // First 128-bit lane:
- 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
- // Second 128-bit lane:
- 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55))
- return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
- if (isShuffleEquivalent(
- Mask,
- // First 128-bit lane:
- 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
- // Second 128-bit lane:
- 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63))
- return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
-
- if (isSingleInputShuffleMask(Mask)) {
- // There are no generalized cross-lane shuffle operations available on i8
- // element types.
- if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
- return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
- Mask, DAG);
-
- SDValue PSHUFBMask[32];
- for (int i = 0; i < 32; ++i)
- PSHUFBMask[i] =
- Mask[i] < 0
- ? DAG.getUNDEF(MVT::i8)
- : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, MVT::i8);
-
- return DAG.getNode(
- X86ISD::PSHUFB, DL, MVT::v32i8, V1,
- DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
- }
-
- // Try to simplify this by merging 128-bit lanes to enable a lane-based
- // shuffle.
- if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
- DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
- return Result;
-
- // Otherwise fall back on generic lowering.
- return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, 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) {
- SDLoc DL(Op);
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> Mask = SVOp->getMask();
-
- // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
- // check for those subtargets here and avoid much of the subtarget querying in
- // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
- // ability to manipulate a 256-bit vector with integer types. Since we'll use
- // floating point types there eventually, just immediately cast everything to
- // a float and operate entirely in that domain.
- if (VT.isInteger() && !Subtarget->hasAVX2()) {
- int ElementBits = VT.getScalarSizeInBits();
- if (ElementBits < 32)
- // No floating point type available, decompose into 128-bit vectors.
- return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
-
- MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
- VT.getVectorNumElements());
- V1 = DAG.getNode(ISD::BITCAST, DL, FpVT, V1);
- V2 = DAG.getNode(ISD::BITCAST, DL, FpVT, V2);
- return DAG.getNode(ISD::BITCAST, DL, VT,
- DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
- }
-
- 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::v8f32:
- return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
- case MVT::v8i32:
- return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
- case MVT::v16i16:
- return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
- case MVT::v32i8:
- return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
-
- default:
- llvm_unreachable("Not a valid 256-bit x86 vector type!");
- }
-}
-
-/// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
-static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- SDLoc DL(Op);
- assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
- assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> Mask = SVOp->getMask();
- assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
-
- // FIXME: Implement direct support for this type!
- return splitAndLowerVectorShuffle(DL, MVT::v8f64, V1, V2, Mask, DAG);
-}
-
-/// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
-static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- SDLoc DL(Op);
- assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
- assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> Mask = SVOp->getMask();
- assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
-
- // FIXME: Implement direct support for this type!
- return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG);
-}
-
-/// \brief Handle lowering of 8-lane 64-bit integer shuffles.
-static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- SDLoc DL(Op);
- assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
- assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> Mask = SVOp->getMask();
- assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
-
- // FIXME: Implement direct support for this type!
- return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG);
-}
-
-/// \brief Handle lowering of 16-lane 32-bit integer shuffles.
-static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- SDLoc DL(Op);
- assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
- assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> Mask = SVOp->getMask();
- assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
-
- // FIXME: Implement direct support for this type!
- return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
-}
-
-/// \brief Handle lowering of 32-lane 16-bit integer shuffles.
-static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- SDLoc DL(Op);
- assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
- assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> Mask = SVOp->getMask();
- assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
- assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
-
- // FIXME: Implement direct support for this type!
- return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG);
-}
-
-/// \brief Handle lowering of 64-lane 8-bit integer shuffles.
-static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- SDLoc DL(Op);
- assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
- assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> Mask = SVOp->getMask();
- assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
- assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
-
- // FIXME: Implement direct support for this type!
- return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
-}
-
-/// \brief High-level routine to lower various 512-bit x86 vector shuffles.
-///
-/// This routine either breaks down the specific type of a 512-bit x86 vector
-/// shuffle or splits it into two 256-bit shuffles and fuses the results back
-/// together based on the available instructions.
-static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
- MVT VT, const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- SDLoc DL(Op);
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> Mask = SVOp->getMask();
- assert(Subtarget->hasAVX512() &&
- "Cannot lower 512-bit vectors w/ basic ISA!");
-
- // Check for being able to broadcast a single element.
- if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(VT.SimpleTy, DL, V1,
- Mask, Subtarget, DAG))
- return Broadcast;
-
- // Dispatch to each element type for lowering. If we don't have supprot for
- // specific element type shuffles at 512 bits, immediately split them and
- // lower them. Each lowering routine of a given type is allowed to assume that
- // the requisite ISA extensions for that element type are available.
- switch (VT.SimpleTy) {
- case MVT::v8f64:
- return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
- case MVT::v16f32:
- return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
- case MVT::v8i64:
- return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
- case MVT::v16i32:
- return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
- case MVT::v32i16:
- if (Subtarget->hasBWI())
- return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
- break;
- case MVT::v64i8:
- if (Subtarget->hasBWI())
- return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
- break;
-
- default:
- llvm_unreachable("Not a valid 512-bit x86 vector type!");
- }
-
- // Otherwise fall back on splitting.
- return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
-}
-
-/// \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) {
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- ArrayRef<int> Mask = SVOp->getMask();
- SDValue V1 = Op.getOperand(0);
- SDValue V2 = Op.getOperand(1);
- MVT VT = Op.getSimpleValueType();
- int NumElements = VT.getVectorNumElements();
- SDLoc dl(Op);
-
- assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
-
- bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
- bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
- if (V1IsUndef && V2IsUndef)
- return DAG.getUNDEF(VT);
-
- // When we create a shuffle node we put the UNDEF node to second operand,
- // but in some cases the first operand may be transformed to UNDEF.
- // In this case we should just commute the node.
- if (V1IsUndef)
- return DAG.getCommutedVectorShuffle(*SVOp);
-
- // Check for non-undef masks pointing at an undef vector and make the masks
- // undef as well. This makes it easier to match the shuffle based solely on
- // the mask.
- if (V2IsUndef)
- for (int M : Mask)
- if (M >= NumElements) {
- SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
- for (int &M : NewMask)
- if (M >= NumElements)
- M = -1;
- return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
- }
-
- // Try to collapse shuffles into using a vector type with fewer elements but
- // wider element types. We cap this to not form integers or floating point
- // elements wider than 64 bits, but it might be interesting to form i128
- // integers to handle flipping the low and high halves of AVX 256-bit vectors.
- SmallVector<int, 16> WidenedMask;
- if (VT.getScalarSizeInBits() < 64 &&
- canWidenShuffleElements(Mask, WidenedMask)) {
- MVT NewEltVT = VT.isFloatingPoint()
- ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
- : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
- MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
- // Make sure that the new vector type is legal. For example, v2f64 isn't
- // legal on SSE1.
- if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
- V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
- V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
- return DAG.getNode(ISD::BITCAST, dl, VT,
- DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
- }
- }
-
- int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
- for (int M : SVOp->getMask())
- if (M < 0)
- ++NumUndefElements;
- else if (M < NumElements)
- ++NumV1Elements;
- else
- ++NumV2Elements;
-
- // Commute the shuffle as needed such that more elements come from V1 than
- // V2. This allows us to match the shuffle pattern strictly on how many
- // elements come from V1 without handling the symmetric cases.
- if (NumV2Elements > NumV1Elements)
- return DAG.getCommutedVectorShuffle(*SVOp);
-
- // When the number of V1 and V2 elements are the same, try to minimize the
- // number of uses of V2 in the low half of the vector. When that is tied,
- // ensure that the sum of indices for V1 is equal to or lower than the sum
- // indices for V2. When those are equal, try to ensure that the number of odd
- // indices for V1 is lower than the number of odd indices for V2.
- if (NumV1Elements == NumV2Elements) {
- int LowV1Elements = 0, LowV2Elements = 0;
- for (int M : SVOp->getMask().slice(0, NumElements / 2))
- if (M >= NumElements)
- ++LowV2Elements;
- else if (M >= 0)
- ++LowV1Elements;
- if (LowV2Elements > LowV1Elements) {
- return DAG.getCommutedVectorShuffle(*SVOp);
- } else if (LowV2Elements == LowV1Elements) {
- int SumV1Indices = 0, SumV2Indices = 0;
- for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
- if (SVOp->getMask()[i] >= NumElements)
- SumV2Indices += i;
- else if (SVOp->getMask()[i] >= 0)
- SumV1Indices += i;
- if (SumV2Indices < SumV1Indices) {
- return DAG.getCommutedVectorShuffle(*SVOp);
- } else if (SumV2Indices == SumV1Indices) {
- int NumV1OddIndices = 0, NumV2OddIndices = 0;
- for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
- if (SVOp->getMask()[i] >= NumElements)
- NumV2OddIndices += i % 2;
- else if (SVOp->getMask()[i] >= 0)
- NumV1OddIndices += i % 2;
- if (NumV2OddIndices < NumV1OddIndices)
- return DAG.getCommutedVectorShuffle(*SVOp);
- }
- }
- }
+/// This essentially blends the out-of-lane inputs to each lane into the lane
+/// from a permuted copy of the vector. This lowering strategy results in four
+/// instructions in the worst case for a single-input cross lane shuffle which
+/// is lower than any other fully general cross-lane shuffle strategy I'm aware
+/// of. Special cases for each particular shuffle pattern should be handled
+/// prior to trying this lowering.
+static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
+ SDValue V1, SDValue V2,
+ ArrayRef<int> Mask,
+ SelectionDAG &DAG) {
+ // FIXME: This should probably be generalized for 512-bit vectors as well.
+ assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
+ int LaneSize = Mask.size() / 2;
- // For each vector width, delegate to a specialized lowering routine.
- if (VT.getSizeInBits() == 128)
- return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
+ // If there are only inputs from one 128-bit lane, splitting will in fact be
+ // less expensive. The flags track wether the given lane contains an element
+ // that crosses to another lane.
+ bool LaneCrossing[2] = {false, false};
+ for (int i = 0, Size = Mask.size(); i < Size; ++i)
+ if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
+ LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
+ if (!LaneCrossing[0] || !LaneCrossing[1])
+ return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
- if (VT.getSizeInBits() == 256)
- return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
+ if (isSingleInputShuffleMask(Mask)) {
+ SmallVector<int, 32> FlippedBlendMask;
+ for (int i = 0, Size = Mask.size(); i < Size; ++i)
+ FlippedBlendMask.push_back(
+ Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
+ ? Mask[i]
+ : Mask[i] % LaneSize +
+ (i / LaneSize) * LaneSize + Size));
- // Force AVX-512 vectors to be scalarized for now.
- // FIXME: Implement AVX-512 support!
- if (VT.getSizeInBits() == 512)
- return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
+ // Flip the vector, and blend the results which should now be in-lane. The
+ // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
+ // 5 for the high source. The value 3 selects the high half of source 2 and
+ // the value 2 selects the low half of source 2. We only use source 2 to
+ // allow folding it into a memory operand.
+ unsigned PERMMask = 3 | 2 << 4;
+ SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
+ V1, DAG.getConstant(PERMMask, MVT::i8));
+ return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
+ }
- llvm_unreachable("Unimplemented!");
+ // This now reduces to two single-input shuffles of V1 and V2 which at worst
+ // will be handled by the above logic and a blend of the results, much like
+ // other patterns in AVX.
+ return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
}
+/// \brief Handle lowering 2-lane 128-bit shuffles.
+static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
+ SDValue V2, ArrayRef<int> Mask,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ // Blends are faster and handle all the non-lane-crossing cases.
+ if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
+ Subtarget, DAG))
+ return Blend;
-//===----------------------------------------------------------------------===//
-// Legacy vector shuffle lowering
-//
-// This code is the legacy code handling vector shuffles until the above
-// replaces its functionality and performance.
-//===----------------------------------------------------------------------===//
-
-static bool isBlendMask(ArrayRef<int> MaskVals, MVT VT, bool hasSSE41,
- bool hasInt256, unsigned *MaskOut = nullptr) {
- MVT EltVT = VT.getVectorElementType();
+ MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
+ VT.getVectorNumElements() / 2);
+ // Check for patterns which can be matched with a single insert of a 128-bit
+ // subvector.
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1}) ||
+ isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
+ SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
+ DAG.getIntPtrConstant(0));
+ SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
+ Mask[2] < 4 ? V1 : V2, DAG.getIntPtrConstant(0));
+ return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
+ }
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 1, 6, 7})) {
+ SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
+ DAG.getIntPtrConstant(0));
+ SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V2,
+ DAG.getIntPtrConstant(2));
+ return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
+ }
- // There is no blend with immediate in AVX-512.
- if (VT.is512BitVector())
- return false;
+ // Otherwise form a 128-bit permutation.
+ // FIXME: Detect zero-vector inputs and use the VPERM2X128 to zero that half.
+ unsigned PermMask = Mask[0] / 2 | (Mask[2] / 2) << 4;
+ return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
+ DAG.getConstant(PermMask, MVT::i8));
+}
- if (!hasSSE41 || EltVT == MVT::i8)
- return false;
- if (!hasInt256 && VT == MVT::v16i16)
- return false;
+/// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
+/// shuffling each lane.
+///
+/// This will only succeed when the result of fixing the 128-bit lanes results
+/// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
+/// each 128-bit lanes. This handles many cases where we can quickly blend away
+/// the lane crosses early and then use simpler shuffles within each lane.
+///
+/// FIXME: It might be worthwhile at some point to support this without
+/// requiring the 128-bit lane-relative shuffles to be repeating, but currently
+/// in x86 only floating point has interesting non-repeating shuffles, and even
+/// those are still *marginally* more expensive.
+static SDValue lowerVectorShuffleByMerging128BitLanes(
+ SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
+ const X86Subtarget *Subtarget, SelectionDAG &DAG) {
+ assert(!isSingleInputShuffleMask(Mask) &&
+ "This is only useful with multiple inputs.");
- unsigned MaskValue = 0;
- unsigned NumElems = VT.getVectorNumElements();
- // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
- unsigned NumLanes = (NumElems - 1) / 8 + 1;
- unsigned NumElemsInLane = NumElems / NumLanes;
+ int Size = Mask.size();
+ int LaneSize = 128 / VT.getScalarSizeInBits();
+ int NumLanes = Size / LaneSize;
+ assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
- // Blend for v16i16 should be symetric for the both lanes.
- for (unsigned i = 0; i < NumElemsInLane; ++i) {
+ // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
+ // check whether the in-128-bit lane shuffles share a repeating pattern.
+ SmallVector<int, 4> Lanes;
+ Lanes.resize(NumLanes, -1);
+ SmallVector<int, 4> InLaneMask;
+ InLaneMask.resize(LaneSize, -1);
+ for (int i = 0; i < Size; ++i) {
+ if (Mask[i] < 0)
+ continue;
- int SndLaneEltIdx = (NumLanes == 2) ? MaskVals[i + NumElemsInLane] : -1;
- int EltIdx = MaskVals[i];
+ int j = i / LaneSize;
- if ((EltIdx < 0 || EltIdx == (int)i) &&
- (SndLaneEltIdx < 0 || SndLaneEltIdx == (int)(i + NumElemsInLane)))
- continue;
+ if (Lanes[j] < 0) {
+ // First entry we've seen for this lane.
+ Lanes[j] = Mask[i] / LaneSize;
+ } else if (Lanes[j] != Mask[i] / LaneSize) {
+ // This doesn't match the lane selected previously!
+ return SDValue();
+ }
- if (((unsigned)EltIdx == (i + NumElems)) &&
- (SndLaneEltIdx < 0 ||
- (unsigned)SndLaneEltIdx == i + NumElems + NumElemsInLane))
- MaskValue |= (1 << i);
- else
- return false;
+ // Check that within each lane we have a consistent shuffle mask.
+ int k = i % LaneSize;
+ if (InLaneMask[k] < 0) {
+ InLaneMask[k] = Mask[i] % LaneSize;
+ } else if (InLaneMask[k] != Mask[i] % LaneSize) {
+ // This doesn't fit a repeating in-lane mask.
+ return SDValue();
+ }
}
- if (MaskOut)
- *MaskOut = MaskValue;
- return true;
-}
+ // First shuffle the lanes into place.
+ MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
+ VT.getSizeInBits() / 64);
+ SmallVector<int, 8> LaneMask;
+ LaneMask.resize(NumLanes * 2, -1);
+ for (int i = 0; i < NumLanes; ++i)
+ if (Lanes[i] >= 0) {
+ LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
+ LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
+ }
-// Try to lower a shuffle node into a simple blend instruction.
-// This function assumes isBlendMask returns true for this
-// SuffleVectorSDNode
-static SDValue LowerVECTOR_SHUFFLEtoBlend(ShuffleVectorSDNode *SVOp,
- unsigned MaskValue,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- MVT VT = SVOp->getSimpleValueType(0);
- MVT EltVT = VT.getVectorElementType();
- assert(isBlendMask(SVOp->getMask(), VT, Subtarget->hasSSE41(),
- Subtarget->hasInt256() && "Trying to lower a "
- "VECTOR_SHUFFLE to a Blend but "
- "with the wrong mask"));
- SDValue V1 = SVOp->getOperand(0);
- SDValue V2 = SVOp->getOperand(1);
- SDLoc dl(SVOp);
- unsigned NumElems = VT.getVectorNumElements();
+ V1 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V1);
+ V2 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V2);
+ SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
- // Convert i32 vectors to floating point if it is not AVX2.
- // AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors.
- MVT BlendVT = VT;
- if (EltVT == MVT::i64 || (EltVT == MVT::i32 && !Subtarget->hasInt256())) {
- BlendVT = MVT::getVectorVT(MVT::getFloatingPointVT(EltVT.getSizeInBits()),
- NumElems);
- V1 = DAG.getNode(ISD::BITCAST, dl, VT, V1);
- V2 = DAG.getNode(ISD::BITCAST, dl, VT, V2);
- }
+ // Cast it back to the type we actually want.
+ LaneShuffle = DAG.getNode(ISD::BITCAST, DL, VT, LaneShuffle);
- SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, V1, V2,
- DAG.getConstant(MaskValue, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Ret);
-}
+ // Now do a simple shuffle that isn't lane crossing.
+ SmallVector<int, 8> NewMask;
+ NewMask.resize(Size, -1);
+ for (int i = 0; i < Size; ++i)
+ if (Mask[i] >= 0)
+ NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
+ assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
+ "Must not introduce lane crosses at this point!");
-/// In vector type \p VT, return true if the element at index \p InputIdx
-/// falls on a different 128-bit lane than \p OutputIdx.
-static bool ShuffleCrosses128bitLane(MVT VT, unsigned InputIdx,
- unsigned OutputIdx) {
- unsigned EltSize = VT.getVectorElementType().getSizeInBits();
- return InputIdx * EltSize / 128 != OutputIdx * EltSize / 128;
+ return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
}
-/// Generate a PSHUFB if possible. Selects elements from \p V1 according to
-/// \p MaskVals. MaskVals[OutputIdx] = InputIdx specifies that we want to
-/// shuffle the element at InputIdx in V1 to OutputIdx in the result. If \p
-/// MaskVals refers to elements outside of \p V1 or is undef (-1), insert a
-/// zero.
-static SDValue getPSHUFB(ArrayRef<int> MaskVals, SDValue V1, SDLoc &dl,
- SelectionDAG &DAG) {
- MVT VT = V1.getSimpleValueType();
- assert(VT.is128BitVector() || VT.is256BitVector());
-
- MVT EltVT = VT.getVectorElementType();
- unsigned EltSizeInBytes = EltVT.getSizeInBits() / 8;
- unsigned NumElts = VT.getVectorNumElements();
+/// \brief Test whether the specified input (0 or 1) is in-place blended by the
+/// given mask.
+///
+/// This returns true if the elements from a particular input are already in the
+/// slot required by the given mask and require no permutation.
+static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
+ assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
+ int Size = Mask.size();
+ for (int i = 0; i < Size; ++i)
+ if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
+ return false;
- SmallVector<SDValue, 32> PshufbMask;
- for (unsigned OutputIdx = 0; OutputIdx < NumElts; ++OutputIdx) {
- int InputIdx = MaskVals[OutputIdx];
- unsigned InputByteIdx;
+ return true;
+}
- if (InputIdx < 0 || NumElts <= (unsigned)InputIdx)
- InputByteIdx = 0x80;
- else {
- // Cross lane is not allowed.
- if (ShuffleCrosses128bitLane(VT, InputIdx, OutputIdx))
- return SDValue();
- InputByteIdx = InputIdx * EltSizeInBytes;
- // Index is an byte offset within the 128-bit lane.
- InputByteIdx &= 0xf;
- }
+/// \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!");
- for (unsigned j = 0; j < EltSizeInBytes; ++j) {
- PshufbMask.push_back(DAG.getConstant(InputByteIdx, MVT::i8));
- if (InputByteIdx != 0x80)
- ++InputByteIdx;
- }
- }
+ SmallVector<int, 4> WidenedMask;
+ if (canWidenShuffleElements(Mask, WidenedMask))
+ return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
+ DAG);
- MVT ShufVT = MVT::getVectorVT(MVT::i8, PshufbMask.size());
- if (ShufVT != VT)
- V1 = DAG.getNode(ISD::BITCAST, dl, ShufVT, V1);
- return DAG.getNode(X86ISD::PSHUFB, dl, ShufVT, V1,
- DAG.getNode(ISD::BUILD_VECTOR, dl, ShufVT, PshufbMask));
-}
+ if (isSingleInputShuffleMask(Mask)) {
+ // Check for being able to broadcast a single element.
+ if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
+ Mask, Subtarget, DAG))
+ return Broadcast;
-// v8i16 shuffles - Prefer shuffles in the following order:
-// 1. [all] pshuflw, pshufhw, optional move
-// 2. [ssse3] 1 x pshufb
-// 3. [ssse3] 2 x pshufb + 1 x por
-// 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
-static SDValue
-LowerVECTOR_SHUFFLEv8i16(SDValue Op, const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- SDValue V1 = SVOp->getOperand(0);
- SDValue V2 = SVOp->getOperand(1);
- SDLoc dl(SVOp);
- SmallVector<int, 8> MaskVals;
+ // Use low duplicate instructions for masks that match their pattern.
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
+ return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
- // Determine if more than 1 of the words in each of the low and high quadwords
- // of the result come from the same quadword of one of the two inputs. Undef
- // mask values count as coming from any quadword, for better codegen.
- //
- // Lo/HiQuad[i] = j indicates how many words from the ith quad of the input
- // feeds this quad. For i, 0 and 1 refer to V1, 2 and 3 refer to V2.
- unsigned LoQuad[] = { 0, 0, 0, 0 };
- unsigned HiQuad[] = { 0, 0, 0, 0 };
- // Indices of quads used.
- std::bitset<4> InputQuads;
- for (unsigned i = 0; i < 8; ++i) {
- unsigned *Quad = i < 4 ? LoQuad : HiQuad;
- int EltIdx = SVOp->getMaskElt(i);
- MaskVals.push_back(EltIdx);
- if (EltIdx < 0) {
- ++Quad[0];
- ++Quad[1];
- ++Quad[2];
- ++Quad[3];
- continue;
+ if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, 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::VPERMILPI, DL, MVT::v4f64, V1,
+ DAG.getConstant(VPERMILPMask, MVT::i8));
}
- ++Quad[EltIdx / 4];
- InputQuads.set(EltIdx / 4);
- }
- int BestLoQuad = -1;
- unsigned MaxQuad = 1;
- for (unsigned i = 0; i < 4; ++i) {
- if (LoQuad[i] > MaxQuad) {
- BestLoQuad = i;
- MaxQuad = LoQuad[i];
- }
- }
+ // With AVX2 we have direct support for this permutation.
+ if (Subtarget->hasAVX2())
+ return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
+ getV4X86ShuffleImm8ForMask(Mask, DAG));
- int BestHiQuad = -1;
- MaxQuad = 1;
- for (unsigned i = 0; i < 4; ++i) {
- if (HiQuad[i] > MaxQuad) {
- BestHiQuad = i;
- MaxQuad = HiQuad[i];
- }
+ // Otherwise, fall back.
+ return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
+ DAG);
}
- // For SSSE3, If all 8 words of the result come from only 1 quadword of each
- // of the two input vectors, shuffle them into one input vector so only a
- // single pshufb instruction is necessary. If there are more than 2 input
- // quads, disable the next transformation since it does not help SSSE3.
- bool V1Used = InputQuads[0] || InputQuads[1];
- bool V2Used = InputQuads[2] || InputQuads[3];
- if (Subtarget->hasSSSE3()) {
- if (InputQuads.count() == 2 && V1Used && V2Used) {
- BestLoQuad = InputQuads[0] ? 0 : 1;
- BestHiQuad = InputQuads[2] ? 2 : 3;
- }
- if (InputQuads.count() > 2) {
- BestLoQuad = -1;
- BestHiQuad = -1;
- }
- }
+ // X86 has dedicated unpack instructions that can handle specific blend
+ // operations: UNPCKH and UNPCKL.
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
+ if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
- // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
- // the shuffle mask. If a quad is scored as -1, that means that it contains
- // words from all 4 input quadwords.
- SDValue NewV;
- if (BestLoQuad >= 0 || BestHiQuad >= 0) {
- int MaskV[] = {
- BestLoQuad < 0 ? 0 : BestLoQuad,
- BestHiQuad < 0 ? 1 : BestHiQuad
- };
- NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
- DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
- DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
- NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
-
- // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
- // source words for the shuffle, to aid later transformations.
- bool AllWordsInNewV = true;
- bool InOrder[2] = { true, true };
- for (unsigned i = 0; i != 8; ++i) {
- int idx = MaskVals[i];
- if (idx != (int)i)
- InOrder[i/4] = false;
- if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
- continue;
- AllWordsInNewV = false;
- break;
- }
+ // If we have a single input to the zero element, insert that into V1 if we
+ // can do so cheaply.
+ int NumV2Elements =
+ std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
+ if (NumV2Elements == 1 && Mask[0] >= 4)
+ if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
+ DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
+ return Insertion;
- bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
- if (AllWordsInNewV) {
- for (int i = 0; i != 8; ++i) {
- int idx = MaskVals[i];
- if (idx < 0)
- continue;
- idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
- if ((idx != i) && idx < 4)
- pshufhw = false;
- if ((idx != i) && idx > 3)
- pshuflw = false;
- }
- V1 = NewV;
- V2Used = false;
- BestLoQuad = 0;
- BestHiQuad = 1;
- }
-
- // If we've eliminated the use of V2, and the new mask is a pshuflw or
- // pshufhw, that's as cheap as it gets. Return the new shuffle.
- if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
- unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
- unsigned TargetMask = 0;
- NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
- DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
- TargetMask = pshufhw ? getShufflePSHUFHWImmediate(SVOp):
- getShufflePSHUFLWImmediate(SVOp);
- V1 = NewV.getOperand(0);
- return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
- }
- }
-
- // Promote splats to a larger type which usually leads to more efficient code.
- // FIXME: Is this true if pshufb is available?
- if (SVOp->isSplat())
- return PromoteSplat(SVOp, DAG);
-
- // If we have SSSE3, and all words of the result are from 1 input vector,
- // case 2 is generated, otherwise case 3 is generated. If no SSSE3
- // is present, fall back to case 4.
- if (Subtarget->hasSSSE3()) {
- SmallVector<SDValue,16> pshufbMask;
-
- // If we have elements from both input vectors, set the high bit of the
- // shuffle mask element to zero out elements that come from V2 in the V1
- // mask, and elements that come from V1 in the V2 mask, so that the two
- // results can be OR'd together.
- bool TwoInputs = V1Used && V2Used;
- V1 = getPSHUFB(MaskVals, V1, dl, DAG);
- if (!TwoInputs)
- return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
-
- // Calculate the shuffle mask for the second input, shuffle it, and
- // OR it with the first shuffled input.
- CommuteVectorShuffleMask(MaskVals, 8);
- V2 = getPSHUFB(MaskVals, V2, dl, DAG);
- V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
- return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
- }
-
- // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
- // and update MaskVals with new element order.
- std::bitset<8> InOrder;
- if (BestLoQuad >= 0) {
- int MaskV[] = { -1, -1, -1, -1, 4, 5, 6, 7 };
- for (int i = 0; i != 4; ++i) {
- int idx = MaskVals[i];
- if (idx < 0) {
- InOrder.set(i);
- } else if ((idx / 4) == BestLoQuad) {
- MaskV[i] = idx & 3;
- InOrder.set(i);
- }
- }
- NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
- &MaskV[0]);
+ if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
+ Subtarget, DAG))
+ return Blend;
- if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSE2()) {
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
- NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
- NewV.getOperand(0),
- getShufflePSHUFLWImmediate(SVOp), DAG);
- }
+ // Check if the blend happens to exactly fit that of SHUFPD.
+ if ((Mask[0] == -1 || Mask[0] < 2) &&
+ (Mask[1] == -1 || (Mask[1] >= 4 && Mask[1] < 6)) &&
+ (Mask[2] == -1 || (Mask[2] >= 2 && Mask[2] < 4)) &&
+ (Mask[3] == -1 || Mask[3] >= 6)) {
+ 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[0] < 6)) &&
+ (Mask[1] == -1 || Mask[1] < 2) &&
+ (Mask[2] == -1 || Mask[2] >= 6) &&
+ (Mask[3] == -1 || (Mask[3] >= 2 && 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));
}
- // If BestHi >= 0, generate a pshufhw to put the high elements in order,
- // and update MaskVals with the new element order.
- if (BestHiQuad >= 0) {
- int MaskV[] = { 0, 1, 2, 3, -1, -1, -1, -1 };
- for (unsigned i = 4; i != 8; ++i) {
- int idx = MaskVals[i];
- if (idx < 0) {
- InOrder.set(i);
- } else if ((idx / 4) == BestHiQuad) {
- MaskV[i] = (idx & 3) + 4;
- InOrder.set(i);
- }
- }
- NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
- &MaskV[0]);
-
- if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSE2()) {
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
- NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
- NewV.getOperand(0),
- getShufflePSHUFHWImmediate(SVOp), DAG);
- }
- }
+ // Try to simplify this by merging 128-bit lanes to enable a lane-based
+ // shuffle. However, if we have AVX2 and either inputs are already in place,
+ // we will be able to shuffle even across lanes the other input in a single
+ // instruction so skip this pattern.
+ if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
+ isShuffleMaskInputInPlace(1, Mask))))
+ if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
+ DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
+ return Result;
- // In case BestHi & BestLo were both -1, which means each quadword has a word
- // from each of the four input quadwords, calculate the InOrder bitvector now
- // before falling through to the insert/extract cleanup.
- if (BestLoQuad == -1 && BestHiQuad == -1) {
- NewV = V1;
- for (int i = 0; i != 8; ++i)
- if (MaskVals[i] < 0 || MaskVals[i] == i)
- InOrder.set(i);
- }
+ // If we have AVX2 then we always want to lower with a blend because an v4 we
+ // can fully permute the elements.
+ if (Subtarget->hasAVX2())
+ return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
+ Mask, DAG);
- // The other elements are put in the right place using pextrw and pinsrw.
- for (unsigned i = 0; i != 8; ++i) {
- if (InOrder[i])
- continue;
- int EltIdx = MaskVals[i];
- if (EltIdx < 0)
- continue;
- SDValue ExtOp = (EltIdx < 8) ?
- DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
- DAG.getIntPtrConstant(EltIdx)) :
- DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
- DAG.getIntPtrConstant(EltIdx - 8));
- NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
- DAG.getIntPtrConstant(i));
- }
- return NewV;
+ // Otherwise fall back on generic lowering.
+ return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
}
-/// \brief v16i16 shuffles
+/// \brief Handle lowering of 4-lane 64-bit integer shuffles.
///
-/// FIXME: We only support generation of a single pshufb currently. We can
-/// generalize the other applicable cases from LowerVECTOR_SHUFFLEv8i16 as
-/// well (e.g 2 x pshufb + 1 x por).
-static SDValue
-LowerVECTOR_SHUFFLEv16i16(SDValue Op, SelectionDAG &DAG) {
+/// This routine is only called when we have AVX2 and thus a reasonable
+/// instruction set for v4i64 shuffling..
+static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- SDValue V1 = SVOp->getOperand(0);
- SDValue V2 = SVOp->getOperand(1);
- SDLoc dl(SVOp);
-
- if (V2.getOpcode() != ISD::UNDEF)
- return SDValue();
-
- SmallVector<int, 16> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end());
- return getPSHUFB(MaskVals, V1, dl, DAG);
-}
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
+ assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
-// v16i8 shuffles - Prefer shuffles in the following order:
-// 1. [ssse3] 1 x pshufb
-// 2. [ssse3] 2 x pshufb + 1 x por
-// 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
-static SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
- const X86Subtarget* Subtarget,
- SelectionDAG &DAG) {
- const TargetLowering &TLI = DAG.getTargetLoweringInfo();
- SDValue V1 = SVOp->getOperand(0);
- SDValue V2 = SVOp->getOperand(1);
- SDLoc dl(SVOp);
- ArrayRef<int> MaskVals = SVOp->getMask();
+ SmallVector<int, 4> WidenedMask;
+ if (canWidenShuffleElements(Mask, WidenedMask))
+ return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
+ DAG);
- // Promote splats to a larger type which usually leads to more efficient code.
- // FIXME: Is this true if pshufb is available?
- if (SVOp->isSplat())
- return PromoteSplat(SVOp, DAG);
+ if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
+ Subtarget, DAG))
+ return Blend;
- // If we have SSSE3, case 1 is generated when all result bytes come from
- // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
- // present, fall back to case 3.
+ // Check for being able to broadcast a single element.
+ if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
+ Mask, Subtarget, DAG))
+ return Broadcast;
- // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
- if (Subtarget->hasSSSE3()) {
- SmallVector<SDValue,16> pshufbMask;
+ // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
+ // use lower latency instructions that will operate on both 128-bit lanes.
+ SmallVector<int, 2> RepeatedMask;
+ if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
+ if (isSingleInputShuffleMask(Mask)) {
+ int PSHUFDMask[] = {-1, -1, -1, -1};
+ for (int i = 0; i < 2; ++i)
+ if (RepeatedMask[i] >= 0) {
+ PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
+ PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
+ }
+ return DAG.getNode(
+ ISD::BITCAST, DL, MVT::v4i64,
+ DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
+ DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, V1),
+ getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
+ }
+ }
- // If all result elements are from one input vector, then only translate
- // undef mask values to 0x80 (zero out result) in the pshufb mask.
- //
- // Otherwise, we have elements from both input vectors, and must zero out
- // elements that come from V2 in the first mask, and V1 in the second mask
- // so that we can OR them together.
- for (unsigned i = 0; i != 16; ++i) {
- int EltIdx = MaskVals[i];
- if (EltIdx < 0 || EltIdx >= 16)
- EltIdx = 0x80;
- pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
- }
- V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
- DAG.getNode(ISD::BUILD_VECTOR, dl,
- MVT::v16i8, pshufbMask));
-
- // As PSHUFB will zero elements with negative indices, it's safe to ignore
- // the 2nd operand if it's undefined or zero.
- if (V2.getOpcode() == ISD::UNDEF ||
- ISD::isBuildVectorAllZeros(V2.getNode()))
- return V1;
+ // AVX2 provides a direct instruction for permuting a single input across
+ // lanes.
+ if (isSingleInputShuffleMask(Mask))
+ return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
+ getV4X86ShuffleImm8ForMask(Mask, DAG));
- // Calculate the shuffle mask for the second input, shuffle it, and
- // OR it with the first shuffled input.
- pshufbMask.clear();
- for (unsigned i = 0; i != 16; ++i) {
- int EltIdx = MaskVals[i];
- EltIdx = (EltIdx < 16) ? 0x80 : EltIdx - 16;
- pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
- }
- V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
- DAG.getNode(ISD::BUILD_VECTOR, dl,
- MVT::v16i8, pshufbMask));
- return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
- }
-
- // No SSSE3 - Calculate in place words and then fix all out of place words
- // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
- // the 16 different words that comprise the two doublequadword input vectors.
- V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
- V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
- SDValue NewV = V1;
- for (int i = 0; i != 8; ++i) {
- int Elt0 = MaskVals[i*2];
- int Elt1 = MaskVals[i*2+1];
-
- // This word of the result is all undef, skip it.
- if (Elt0 < 0 && Elt1 < 0)
- continue;
+ // Try to use shift instructions.
+ if (SDValue Shift =
+ lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
+ return Shift;
- // This word of the result is already in the correct place, skip it.
- if ((Elt0 == i*2) && (Elt1 == i*2+1))
- continue;
+ // Use dedicated unpack instructions for masks that match their pattern.
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V2, V1);
+ if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V2, V1);
- SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
- SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
- SDValue InsElt;
-
- // If Elt0 and Elt1 are defined, are consecutive, and can be load
- // using a single extract together, load it and store it.
- if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
- InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
- DAG.getIntPtrConstant(Elt1 / 2));
- NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
- DAG.getIntPtrConstant(i));
- continue;
- }
+ // Try to simplify this by merging 128-bit lanes to enable a lane-based
+ // shuffle. However, if we have AVX2 and either inputs are already in place,
+ // we will be able to shuffle even across lanes the other input in a single
+ // instruction so skip this pattern.
+ if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
+ isShuffleMaskInputInPlace(1, Mask))))
+ if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
+ DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
+ return Result;
- // If Elt1 is defined, extract it from the appropriate source. If the
- // source byte is not also odd, shift the extracted word left 8 bits
- // otherwise clear the bottom 8 bits if we need to do an or.
- if (Elt1 >= 0) {
- InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
- DAG.getIntPtrConstant(Elt1 / 2));
- if ((Elt1 & 1) == 0)
- InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
- DAG.getConstant(8,
- TLI.getShiftAmountTy(InsElt.getValueType())));
- else if (Elt0 >= 0)
- InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
- DAG.getConstant(0xFF00, MVT::i16));
- }
- // If Elt0 is defined, extract it from the appropriate source. If the
- // source byte is not also even, shift the extracted word right 8 bits. If
- // Elt1 was also defined, OR the extracted values together before
- // inserting them in the result.
- if (Elt0 >= 0) {
- SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
- Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
- if ((Elt0 & 1) != 0)
- InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
- DAG.getConstant(8,
- TLI.getShiftAmountTy(InsElt0.getValueType())));
- else if (Elt1 >= 0)
- InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
- DAG.getConstant(0x00FF, MVT::i16));
- InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
- : InsElt0;
- }
- NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
- DAG.getIntPtrConstant(i));
- }
- return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
-}
-
-// v32i8 shuffles - Translate to VPSHUFB if possible.
-static
-SDValue LowerVECTOR_SHUFFLEv32i8(ShuffleVectorSDNode *SVOp,
- const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- MVT VT = SVOp->getSimpleValueType(0);
- SDValue V1 = SVOp->getOperand(0);
- SDValue V2 = SVOp->getOperand(1);
- SDLoc dl(SVOp);
- SmallVector<int, 32> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end());
+ // Otherwise fall back on generic blend lowering.
+ return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
+ Mask, DAG);
+}
- bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
- bool V1IsAllZero = ISD::isBuildVectorAllZeros(V1.getNode());
- bool V2IsAllZero = ISD::isBuildVectorAllZeros(V2.getNode());
-
- // VPSHUFB may be generated if
- // (1) one of input vector is undefined or zeroinitializer.
- // The mask value 0x80 puts 0 in the corresponding slot of the vector.
- // And (2) the mask indexes don't cross the 128-bit lane.
- if (VT != MVT::v32i8 || !Subtarget->hasInt256() ||
- (!V2IsUndef && !V2IsAllZero && !V1IsAllZero))
- return SDValue();
+/// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
+///
+/// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
+/// isn't available.
+static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
- if (V1IsAllZero && !V2IsAllZero) {
- CommuteVectorShuffleMask(MaskVals, 32);
- V1 = V2;
- }
- return getPSHUFB(MaskVals, V1, dl, DAG);
-}
+ if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
+ Subtarget, DAG))
+ return Blend;
-/// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
-/// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
-/// done when every pair / quad of shuffle mask elements point to elements in
-/// the right sequence. e.g.
-/// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
-static
-SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
- SelectionDAG &DAG) {
- MVT VT = SVOp->getSimpleValueType(0);
- SDLoc dl(SVOp);
- unsigned NumElems = VT.getVectorNumElements();
- MVT NewVT;
- unsigned Scale;
- switch (VT.SimpleTy) {
- default: llvm_unreachable("Unexpected!");
- case MVT::v2i64:
- case MVT::v2f64:
- return SDValue(SVOp, 0);
- case MVT::v4f32: NewVT = MVT::v2f64; Scale = 2; break;
- case MVT::v4i32: NewVT = MVT::v2i64; Scale = 2; break;
- case MVT::v8i16: NewVT = MVT::v4i32; Scale = 2; break;
- case MVT::v16i8: NewVT = MVT::v4i32; Scale = 4; break;
- case MVT::v16i16: NewVT = MVT::v8i32; Scale = 2; break;
- case MVT::v32i8: NewVT = MVT::v8i32; Scale = 4; break;
- }
-
- SmallVector<int, 8> MaskVec;
- for (unsigned i = 0; i != NumElems; i += Scale) {
- int StartIdx = -1;
- for (unsigned j = 0; j != Scale; ++j) {
- int EltIdx = SVOp->getMaskElt(i+j);
- if (EltIdx < 0)
- continue;
- if (StartIdx < 0)
- StartIdx = (EltIdx / Scale);
- if (EltIdx != (int)(StartIdx*Scale + j))
- return SDValue();
- }
- MaskVec.push_back(StartIdx);
- }
+ // Check for being able to broadcast a single element.
+ if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
+ Mask, Subtarget, DAG))
+ return Broadcast;
- SDValue V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(0));
- SDValue V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(1));
- return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
-}
+ // If the shuffle mask is repeated in each 128-bit lane, we have many more
+ // options to efficiently lower the shuffle.
+ SmallVector<int, 4> RepeatedMask;
+ if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
+ assert(RepeatedMask.size() == 4 &&
+ "Repeated masks must be half the mask width!");
-/// getVZextMovL - Return a zero-extending vector move low node.
-///
-static SDValue getVZextMovL(MVT VT, MVT OpVT,
- SDValue SrcOp, SelectionDAG &DAG,
- const X86Subtarget *Subtarget, SDLoc dl) {
- if (VT == MVT::v2f64 || VT == MVT::v4f32) {
- LoadSDNode *LD = nullptr;
- if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
- LD = dyn_cast<LoadSDNode>(SrcOp);
- if (!LD) {
- // movssrr and movsdrr do not clear top bits. Try to use movd, movq
- // instead.
- MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
- if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
- SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
- SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
- SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
- // PR2108
- OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
- return DAG.getNode(ISD::BITCAST, dl, VT,
- DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
- DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
- OpVT,
- SrcOp.getOperand(0)
- .getOperand(0))));
- }
- }
- }
+ // Use even/odd duplicate instructions for masks that match their pattern.
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
+ return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
+ if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
+ return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
- return DAG.getNode(ISD::BITCAST, dl, VT,
- DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
- DAG.getNode(ISD::BITCAST, dl,
- OpVT, SrcOp)));
-}
+ if (isSingleInputShuffleMask(Mask))
+ return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
+ getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
-/// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles
-/// which could not be matched by any known target speficic shuffle
-static SDValue
-LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
+ // Use dedicated unpack instructions for masks that match their pattern.
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V2, V1);
+ if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V2, V1);
- SDValue NewOp = Compact8x32ShuffleNode(SVOp, DAG);
- if (NewOp.getNode())
- return NewOp;
+ // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
+ // have already handled any direct blends. We also need to squash the
+ // repeated mask into a simulated v4f32 mask.
+ for (int i = 0; i < 4; ++i)
+ if (RepeatedMask[i] >= 8)
+ RepeatedMask[i] -= 4;
+ return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
+ }
- MVT VT = SVOp->getSimpleValueType(0);
+ // If we have a single input shuffle with different shuffle patterns in the
+ // two 128-bit lanes use the variable mask to VPERMILPS.
+ if (isSingleInputShuffleMask(Mask)) {
+ SDValue VPermMask[8];
+ for (int i = 0; i < 8; ++i)
+ VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
+ : DAG.getConstant(Mask[i], MVT::i32);
+ if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
+ return DAG.getNode(
+ X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
+ DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
- unsigned NumElems = VT.getVectorNumElements();
- unsigned NumLaneElems = NumElems / 2;
+ if (Subtarget->hasAVX2())
+ return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8f32,
+ DAG.getNode(ISD::BITCAST, DL, MVT::v8f32,
+ DAG.getNode(ISD::BUILD_VECTOR, DL,
+ MVT::v8i32, VPermMask)),
+ V1);
- SDLoc dl(SVOp);
- MVT EltVT = VT.getVectorElementType();
- MVT NVT = MVT::getVectorVT(EltVT, NumLaneElems);
- SDValue Output[2];
+ // Otherwise, fall back.
+ return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
+ DAG);
+ }
- SmallVector<int, 16> Mask;
- for (unsigned l = 0; l < 2; ++l) {
- // Build a shuffle mask for the output, discovering on the fly which
- // input vectors to use as shuffle operands (recorded in InputUsed).
- // If building a suitable shuffle vector proves too hard, then bail
- // out with UseBuildVector set.
- bool UseBuildVector = false;
- int InputUsed[2] = { -1, -1 }; // Not yet discovered.
- unsigned LaneStart = l * NumLaneElems;
- for (unsigned i = 0; i != NumLaneElems; ++i) {
- // The mask element. This indexes into the input.
- int Idx = SVOp->getMaskElt(i+LaneStart);
- if (Idx < 0) {
- // the mask element does not index into any input vector.
- Mask.push_back(-1);
- continue;
- }
+ // Try to simplify this by merging 128-bit lanes to enable a lane-based
+ // shuffle.
+ if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
+ DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
+ return Result;
- // The input vector this mask element indexes into.
- int Input = Idx / NumLaneElems;
+ // If we have AVX2 then we always want to lower with a blend because at v8 we
+ // can fully permute the elements.
+ if (Subtarget->hasAVX2())
+ return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
+ Mask, DAG);
- // Turn the index into an offset from the start of the input vector.
- Idx -= Input * NumLaneElems;
+ // Otherwise fall back on generic lowering.
+ return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
+}
- // Find or create a shuffle vector operand to hold this input.
- unsigned OpNo;
- for (OpNo = 0; OpNo < array_lengthof(InputUsed); ++OpNo) {
- if (InputUsed[OpNo] == Input)
- // This input vector is already an operand.
- break;
- if (InputUsed[OpNo] < 0) {
- // Create a new operand for this input vector.
- InputUsed[OpNo] = Input;
- break;
- }
- }
+/// \brief Handle lowering of 8-lane 32-bit integer shuffles.
+///
+/// This routine is only called when we have AVX2 and thus a reasonable
+/// instruction set for v8i32 shuffling..
+static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
+ assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
- if (OpNo >= array_lengthof(InputUsed)) {
- // More than two input vectors used! Give up on trying to create a
- // shuffle vector. Insert all elements into a BUILD_VECTOR instead.
- UseBuildVector = true;
- break;
- }
+ // Whenever we can lower this as a zext, that instruction is strictly faster
+ // than any alternative. It also allows us to fold memory operands into the
+ // shuffle in many cases.
+ if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
+ Mask, Subtarget, DAG))
+ return ZExt;
- // Add the mask index for the new shuffle vector.
- Mask.push_back(Idx + OpNo * NumLaneElems);
- }
+ if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
+ Subtarget, DAG))
+ return Blend;
- if (UseBuildVector) {
- SmallVector<SDValue, 16> SVOps;
- for (unsigned i = 0; i != NumLaneElems; ++i) {
- // The mask element. This indexes into the input.
- int Idx = SVOp->getMaskElt(i+LaneStart);
- if (Idx < 0) {
- SVOps.push_back(DAG.getUNDEF(EltVT));
- continue;
- }
+ // Check for being able to broadcast a single element.
+ if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
+ Mask, Subtarget, DAG))
+ return Broadcast;
- // The input vector this mask element indexes into.
- int Input = Idx / NumElems;
+ // If the shuffle mask is repeated in each 128-bit lane we can use more
+ // efficient instructions that mirror the shuffles across the two 128-bit
+ // lanes.
+ SmallVector<int, 4> RepeatedMask;
+ if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
+ assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
+ if (isSingleInputShuffleMask(Mask))
+ return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
+ getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
- // Turn the index into an offset from the start of the input vector.
- Idx -= Input * NumElems;
+ // Use dedicated unpack instructions for masks that match their pattern.
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V2, V1);
+ if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V2, V1);
+ }
- // Extract the vector element by hand.
- SVOps.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
- SVOp->getOperand(Input),
- DAG.getIntPtrConstant(Idx)));
- }
+ // Try to use shift instructions.
+ if (SDValue Shift =
+ lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
+ return Shift;
- // Construct the output using a BUILD_VECTOR.
- Output[l] = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, SVOps);
- } else if (InputUsed[0] < 0) {
- // No input vectors were used! The result is undefined.
- Output[l] = DAG.getUNDEF(NVT);
- } else {
- SDValue Op0 = Extract128BitVector(SVOp->getOperand(InputUsed[0] / 2),
- (InputUsed[0] % 2) * NumLaneElems,
- DAG, dl);
- // If only one input was used, use an undefined vector for the other.
- SDValue Op1 = (InputUsed[1] < 0) ? DAG.getUNDEF(NVT) :
- Extract128BitVector(SVOp->getOperand(InputUsed[1] / 2),
- (InputUsed[1] % 2) * NumLaneElems, DAG, dl);
- // At least one input vector was used. Create a new shuffle vector.
- Output[l] = DAG.getVectorShuffle(NVT, dl, Op0, Op1, &Mask[0]);
- }
+ if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
+ DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
+ return Rotate;
- Mask.clear();
+ // If the shuffle patterns aren't repeated but it is a single input, directly
+ // generate a cross-lane VPERMD instruction.
+ if (isSingleInputShuffleMask(Mask)) {
+ SDValue VPermMask[8];
+ for (int i = 0; i < 8; ++i)
+ VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
+ : DAG.getConstant(Mask[i], MVT::i32);
+ return DAG.getNode(
+ X86ISD::VPERMV, DL, MVT::v8i32,
+ DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
}
- // Concatenate the result back
- return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Output[0], Output[1]);
+ // Try to simplify this by merging 128-bit lanes to enable a lane-based
+ // shuffle.
+ if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
+ DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
+ return Result;
+
+ // Otherwise fall back on generic blend lowering.
+ return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
+ Mask, DAG);
}
-/// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
-/// 4 elements, and match them with several different shuffle types.
-static SDValue
-LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
- SDValue V1 = SVOp->getOperand(0);
- SDValue V2 = SVOp->getOperand(1);
- SDLoc dl(SVOp);
- MVT VT = SVOp->getSimpleValueType(0);
+/// \brief Handle lowering of 16-lane 16-bit integer shuffles.
+///
+/// This routine is only called when we have AVX2 and thus a reasonable
+/// instruction set for v16i16 shuffling..
+static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
+ assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
- assert(VT.is128BitVector() && "Unsupported vector size");
+ // Whenever we can lower this as a zext, that instruction is strictly faster
+ // than any alternative. It also allows us to fold memory operands into the
+ // shuffle in many cases.
+ if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
+ Mask, Subtarget, DAG))
+ return ZExt;
- std::pair<int, int> Locs[4];
- int Mask1[] = { -1, -1, -1, -1 };
- SmallVector<int, 8> PermMask(SVOp->getMask().begin(), SVOp->getMask().end());
+ // Check for being able to broadcast a single element.
+ if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
+ Mask, Subtarget, DAG))
+ return Broadcast;
- unsigned NumHi = 0;
- unsigned NumLo = 0;
- for (unsigned i = 0; i != 4; ++i) {
- int Idx = PermMask[i];
- if (Idx < 0) {
- Locs[i] = std::make_pair(-1, -1);
- } else {
- assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
- if (Idx < 4) {
- Locs[i] = std::make_pair(0, NumLo);
- Mask1[NumLo] = Idx;
- NumLo++;
- } else {
- Locs[i] = std::make_pair(1, NumHi);
- if (2+NumHi < 4)
- Mask1[2+NumHi] = Idx;
- NumHi++;
- }
- }
- }
+ if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
+ Subtarget, DAG))
+ return Blend;
- if (NumLo <= 2 && NumHi <= 2) {
- // If no more than two elements come from either vector. This can be
- // implemented with two shuffles. First shuffle gather the elements.
- // The second shuffle, which takes the first shuffle as both of its
- // vector operands, put the elements into the right order.
- V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
+ // Use dedicated unpack instructions for masks that match their pattern.
+ if (isShuffleEquivalent(V1, V2, Mask,
+ {// First 128-bit lane:
+ 0, 16, 1, 17, 2, 18, 3, 19,
+ // Second 128-bit lane:
+ 8, 24, 9, 25, 10, 26, 11, 27}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask,
+ {// First 128-bit lane:
+ 4, 20, 5, 21, 6, 22, 7, 23,
+ // Second 128-bit lane:
+ 12, 28, 13, 29, 14, 30, 15, 31}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
- int Mask2[] = { -1, -1, -1, -1 };
+ // Try to use shift instructions.
+ if (SDValue Shift =
+ lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
+ return Shift;
- for (unsigned i = 0; i != 4; ++i)
- if (Locs[i].first != -1) {
- unsigned Idx = (i < 2) ? 0 : 4;
- Idx += Locs[i].first * 2 + Locs[i].second;
- Mask2[i] = Idx;
- }
+ // Try to use byte rotation instructions.
+ if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
+ DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
+ return Rotate;
- return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
- }
+ if (isSingleInputShuffleMask(Mask)) {
+ // There are no generalized cross-lane shuffle operations available on i16
+ // element types.
+ if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
+ return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
+ Mask, DAG);
- if (NumLo == 3 || NumHi == 3) {
- // Otherwise, we must have three elements from one vector, call it X, and
- // one element from the other, call it Y. First, use a shufps to build an
- // intermediate vector with the one element from Y and the element from X
- // that will be in the same half in the final destination (the indexes don't
- // matter). Then, use a shufps to build the final vector, taking the half
- // containing the element from Y from the intermediate, and the other half
- // from X.
- if (NumHi == 3) {
- // Normalize it so the 3 elements come from V1.
- CommuteVectorShuffleMask(PermMask, 4);
- std::swap(V1, V2);
+ SmallVector<int, 8> RepeatedMask;
+ if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
+ // As this is a single-input shuffle, the repeated mask should be
+ // a strictly valid v8i16 mask that we can pass through to the v8i16
+ // lowering to handle even the v16 case.
+ return lowerV8I16GeneralSingleInputVectorShuffle(
+ DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
}
- // Find the element from V2.
- unsigned HiIndex;
- for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
- int Val = PermMask[HiIndex];
- if (Val < 0)
+ SDValue PSHUFBMask[32];
+ for (int i = 0; i < 16; ++i) {
+ if (Mask[i] == -1) {
+ PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
continue;
- if (Val >= 4)
- break;
- }
+ }
- Mask1[0] = PermMask[HiIndex];
- Mask1[1] = -1;
- Mask1[2] = PermMask[HiIndex^1];
- Mask1[3] = -1;
- V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
-
- if (HiIndex >= 2) {
- Mask1[0] = PermMask[0];
- Mask1[1] = PermMask[1];
- Mask1[2] = HiIndex & 1 ? 6 : 4;
- Mask1[3] = HiIndex & 1 ? 4 : 6;
- return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
- }
-
- Mask1[0] = HiIndex & 1 ? 2 : 0;
- Mask1[1] = HiIndex & 1 ? 0 : 2;
- Mask1[2] = PermMask[2];
- Mask1[3] = PermMask[3];
- if (Mask1[2] >= 0)
- Mask1[2] += 4;
- if (Mask1[3] >= 0)
- Mask1[3] += 4;
- return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
- }
-
- // Break it into (shuffle shuffle_hi, shuffle_lo).
- int LoMask[] = { -1, -1, -1, -1 };
- int HiMask[] = { -1, -1, -1, -1 };
-
- int *MaskPtr = LoMask;
- unsigned MaskIdx = 0;
- unsigned LoIdx = 0;
- unsigned HiIdx = 2;
- for (unsigned i = 0; i != 4; ++i) {
- if (i == 2) {
- MaskPtr = HiMask;
- MaskIdx = 1;
- LoIdx = 0;
- HiIdx = 2;
- }
- int Idx = PermMask[i];
- if (Idx < 0) {
- Locs[i] = std::make_pair(-1, -1);
- } else if (Idx < 4) {
- Locs[i] = std::make_pair(MaskIdx, LoIdx);
- MaskPtr[LoIdx] = Idx;
- LoIdx++;
- } else {
- Locs[i] = std::make_pair(MaskIdx, HiIdx);
- MaskPtr[HiIdx] = Idx;
- HiIdx++;
+ int M = i < 8 ? Mask[i] : Mask[i] - 8;
+ assert(M >= 0 && M < 8 && "Invalid single-input mask!");
+ PSHUFBMask[2 * i] = DAG.getConstant(2 * M, MVT::i8);
+ PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, MVT::i8);
}
+ return DAG.getNode(
+ ISD::BITCAST, DL, MVT::v16i16,
+ DAG.getNode(
+ X86ISD::PSHUFB, DL, MVT::v32i8,
+ DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1),
+ DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask)));
}
- SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
- SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
- int MaskOps[] = { -1, -1, -1, -1 };
- for (unsigned i = 0; i != 4; ++i)
- if (Locs[i].first != -1)
- MaskOps[i] = Locs[i].first * 4 + Locs[i].second;
- return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
-}
-
-static bool MayFoldVectorLoad(SDValue V) {
- while (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
- V = V.getOperand(0);
-
- if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
- V = V.getOperand(0);
- if (V.hasOneUse() && V.getOpcode() == ISD::BUILD_VECTOR &&
- V.getNumOperands() == 2 && V.getOperand(1).getOpcode() == ISD::UNDEF)
- // BUILD_VECTOR (load), undef
- V = V.getOperand(0);
-
- return MayFoldLoad(V);
-}
-
-static
-SDValue getMOVDDup(SDValue &Op, SDLoc &dl, SDValue V1, SelectionDAG &DAG) {
- MVT VT = Op.getSimpleValueType();
+ // Try to simplify this by merging 128-bit lanes to enable a lane-based
+ // shuffle.
+ if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
+ DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
+ return Result;
- // Canonizalize to v2f64.
- V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
- return DAG.getNode(ISD::BITCAST, dl, VT,
- getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
- V1, DAG));
+ // Otherwise fall back on generic lowering.
+ return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
}
-static
-SDValue getMOVLowToHigh(SDValue &Op, SDLoc &dl, SelectionDAG &DAG,
- bool HasSSE2) {
- SDValue V1 = Op.getOperand(0);
- SDValue V2 = Op.getOperand(1);
- MVT VT = Op.getSimpleValueType();
+/// \brief Handle lowering of 32-lane 8-bit integer shuffles.
+///
+/// This routine is only called when we have AVX2 and thus a reasonable
+/// instruction set for v32i8 shuffling..
+static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
+ assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
- assert(VT != MVT::v2i64 && "unsupported shuffle type");
+ // Whenever we can lower this as a zext, that instruction is strictly faster
+ // than any alternative. It also allows us to fold memory operands into the
+ // shuffle in many cases.
+ if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
+ Mask, Subtarget, DAG))
+ return ZExt;
- if (HasSSE2 && VT == MVT::v2f64)
- return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
+ // Check for being able to broadcast a single element.
+ if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
+ Mask, Subtarget, DAG))
+ return Broadcast;
- // v4f32 or v4i32: canonizalized to v4f32 (which is legal for SSE1)
- return DAG.getNode(ISD::BITCAST, dl, VT,
- getTargetShuffleNode(X86ISD::MOVLHPS, dl, MVT::v4f32,
- DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V1),
- DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V2), DAG));
-}
+ if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
+ Subtarget, DAG))
+ return Blend;
-static
-SDValue getMOVHighToLow(SDValue &Op, SDLoc &dl, SelectionDAG &DAG) {
- SDValue V1 = Op.getOperand(0);
- SDValue V2 = Op.getOperand(1);
- MVT VT = Op.getSimpleValueType();
+ // Use dedicated unpack instructions for masks that match their pattern.
+ // Note that these are repeated 128-bit lane unpacks, not unpacks across all
+ // 256-bit lanes.
+ if (isShuffleEquivalent(
+ V1, V2, Mask,
+ {// First 128-bit lane:
+ 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
+ // Second 128-bit lane:
+ 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
+ if (isShuffleEquivalent(
+ V1, V2, Mask,
+ {// First 128-bit lane:
+ 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
+ // Second 128-bit lane:
+ 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
- assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
- "unsupported shuffle type");
+ // Try to use shift instructions.
+ if (SDValue Shift =
+ lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
+ return Shift;
- if (V2.getOpcode() == ISD::UNDEF)
- V2 = V1;
+ // Try to use byte rotation instructions.
+ if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
+ DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
+ return Rotate;
- // v4i32 or v4f32
- return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
-}
+ if (isSingleInputShuffleMask(Mask)) {
+ // There are no generalized cross-lane shuffle operations available on i8
+ // element types.
+ if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
+ return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
+ Mask, DAG);
-static
-SDValue getMOVLP(SDValue &Op, SDLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
- SDValue V1 = Op.getOperand(0);
- SDValue V2 = Op.getOperand(1);
- MVT VT = Op.getSimpleValueType();
- unsigned NumElems = VT.getVectorNumElements();
+ SDValue PSHUFBMask[32];
+ for (int i = 0; i < 32; ++i)
+ PSHUFBMask[i] =
+ Mask[i] < 0
+ ? DAG.getUNDEF(MVT::i8)
+ : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, MVT::i8);
- // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
- // operand of these instructions is only memory, so check if there's a
- // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
- // same masks.
- bool CanFoldLoad = false;
+ return DAG.getNode(
+ X86ISD::PSHUFB, DL, MVT::v32i8, V1,
+ DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
+ }
- // Trivial case, when V2 comes from a load.
- if (MayFoldVectorLoad(V2))
- CanFoldLoad = true;
+ // Try to simplify this by merging 128-bit lanes to enable a lane-based
+ // shuffle.
+ if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
+ DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
+ return Result;
- // When V1 is a load, it can be folded later into a store in isel, example:
- // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
- // turns into:
- // (MOVLPSmr addr:$src1, VR128:$src2)
- // So, recognize this potential and also use MOVLPS or MOVLPD
- else if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
- CanFoldLoad = true;
+ // Otherwise fall back on generic lowering.
+ return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, 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) {
+ SDLoc DL(Op);
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- if (CanFoldLoad) {
- if (HasSSE2 && NumElems == 2)
- return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
+ ArrayRef<int> Mask = SVOp->getMask();
- if (NumElems == 4)
- // If we don't care about the second element, proceed to use movss.
- if (SVOp->getMaskElt(1) != -1)
- return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
- }
+ // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
+ // check for those subtargets here and avoid much of the subtarget querying in
+ // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
+ // ability to manipulate a 256-bit vector with integer types. Since we'll use
+ // floating point types there eventually, just immediately cast everything to
+ // a float and operate entirely in that domain.
+ if (VT.isInteger() && !Subtarget->hasAVX2()) {
+ int ElementBits = VT.getScalarSizeInBits();
+ if (ElementBits < 32)
+ // No floating point type available, decompose into 128-bit vectors.
+ return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
- // movl and movlp will both match v2i64, but v2i64 is never matched by
- // movl earlier because we make it strict to avoid messing with the movlp load
- // folding logic (see the code above getMOVLP call). Match it here then,
- // this is horrible, but will stay like this until we move all shuffle
- // matching to x86 specific nodes. Note that for the 1st condition all
- // types are matched with movsd.
- if (HasSSE2) {
- // FIXME: isMOVLMask should be checked and matched before getMOVLP,
- // as to remove this logic from here, as much as possible
- if (NumElems == 2 || !isMOVLMask(SVOp->getMask(), VT))
- return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
- return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
+ MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
+ VT.getVectorNumElements());
+ V1 = DAG.getNode(ISD::BITCAST, DL, FpVT, V1);
+ V2 = DAG.getNode(ISD::BITCAST, DL, FpVT, V2);
+ return DAG.getNode(ISD::BITCAST, DL, VT,
+ DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
}
- assert(VT != MVT::v4i32 && "unsupported shuffle type");
+ 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::v8f32:
+ return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
+ case MVT::v8i32:
+ return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
+ case MVT::v16i16:
+ return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
+ case MVT::v32i8:
+ return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
- // Invert the operand order and use SHUFPS to match it.
- return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V2, V1,
- getShuffleSHUFImmediate(SVOp), DAG);
+ default:
+ llvm_unreachable("Not a valid 256-bit x86 vector type!");
+ }
}
-static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
- SelectionDAG &DAG) {
- SDLoc dl(Load);
- MVT VT = Load->getSimpleValueType(0);
- MVT EVT = VT.getVectorElementType();
- SDValue Addr = Load->getOperand(1);
- SDValue NewAddr = DAG.getNode(
- ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
- DAG.getConstant(Index * EVT.getStoreSize(), Addr.getSimpleValueType()));
+/// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
+static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
+
+ // X86 has dedicated unpack instructions that can handle specific blend
+ // operations: UNPCKH and UNPCKL.
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f64, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f64, V1, V2);
- SDValue NewLoad =
- DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
- DAG.getMachineFunction().getMachineMemOperand(
- Load->getMemOperand(), 0, EVT.getStoreSize()));
- return NewLoad;
+ // FIXME: Implement direct support for this type!
+ return splitAndLowerVectorShuffle(DL, MVT::v8f64, V1, V2, Mask, DAG);
}
-// It is only safe to call this function if isINSERTPSMask is true for
-// this shufflevector mask.
-static SDValue getINSERTPS(ShuffleVectorSDNode *SVOp, SDLoc &dl,
- SelectionDAG &DAG) {
- // Generate an insertps instruction when inserting an f32 from memory onto a
- // v4f32 or when copying a member from one v4f32 to another.
- // We also use it for transferring i32 from one register to another,
- // since it simply copies the same bits.
- // If we're transferring an i32 from memory to a specific element in a
- // register, we output a generic DAG that will match the PINSRD
- // instruction.
- MVT VT = SVOp->getSimpleValueType(0);
- MVT EVT = VT.getVectorElementType();
- SDValue V1 = SVOp->getOperand(0);
- SDValue V2 = SVOp->getOperand(1);
- auto Mask = SVOp->getMask();
- assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
- "unsupported vector type for insertps/pinsrd");
-
- auto FromV1Predicate = [](const int &i) { return i < 4 && i > -1; };
- auto FromV2Predicate = [](const int &i) { return i >= 4; };
- int FromV1 = std::count_if(Mask.begin(), Mask.end(), FromV1Predicate);
-
- SDValue From;
- SDValue To;
- unsigned DestIndex;
- if (FromV1 == 1) {
- From = V1;
- To = V2;
- DestIndex = std::find_if(Mask.begin(), Mask.end(), FromV1Predicate) -
- Mask.begin();
-
- // If we have 1 element from each vector, we have to check if we're
- // changing V1's element's place. If so, we're done. Otherwise, we
- // should assume we're changing V2's element's place and behave
- // accordingly.
- int FromV2 = std::count_if(Mask.begin(), Mask.end(), FromV2Predicate);
- assert(DestIndex <= INT32_MAX && "truncated destination index");
- if (FromV1 == FromV2 &&
- static_cast<int>(DestIndex) == Mask[DestIndex] % 4) {
- From = V2;
- To = V1;
- DestIndex =
- std::find_if(Mask.begin(), Mask.end(), FromV2Predicate) - Mask.begin();
- }
- } else {
- assert(std::count_if(Mask.begin(), Mask.end(), FromV2Predicate) == 1 &&
- "More than one element from V1 and from V2, or no elements from one "
- "of the vectors. This case should not have returned true from "
- "isINSERTPSMask");
- From = V2;
- To = V1;
- DestIndex =
- std::find_if(Mask.begin(), Mask.end(), FromV2Predicate) - Mask.begin();
- }
-
- // Get an index into the source vector in the range [0,4) (the mask is
- // in the range [0,8) because it can address V1 and V2)
- unsigned SrcIndex = Mask[DestIndex] % 4;
- if (MayFoldLoad(From)) {
- // Trivial case, when From comes from a load and is only used by the
- // shuffle. Make it use insertps from the vector that we need from that
- // load.
- SDValue NewLoad =
- NarrowVectorLoadToElement(cast<LoadSDNode>(From), SrcIndex, DAG);
- if (!NewLoad.getNode())
- return SDValue();
+/// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
+static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
- if (EVT == MVT::f32) {
- // Create this as a scalar to vector to match the instruction pattern.
- SDValue LoadScalarToVector =
- DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, NewLoad);
- SDValue InsertpsMask = DAG.getIntPtrConstant(DestIndex << 4);
- return DAG.getNode(X86ISD::INSERTPS, dl, VT, To, LoadScalarToVector,
- InsertpsMask);
- } else { // EVT == MVT::i32
- // If we're getting an i32 from memory, use an INSERT_VECTOR_ELT
- // instruction, to match the PINSRD instruction, which loads an i32 to a
- // certain vector element.
- return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, To, NewLoad,
- DAG.getConstant(DestIndex, MVT::i32));
- }
- }
+ // Use dedicated unpack instructions for masks that match their pattern.
+ if (isShuffleEquivalent(V1, V2, Mask,
+ {// First 128-bit lane.
+ 0, 16, 1, 17, 4, 20, 5, 21,
+ // Second 128-bit lane.
+ 8, 24, 9, 25, 12, 28, 13, 29}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16f32, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask,
+ {// First 128-bit lane.
+ 2, 18, 3, 19, 6, 22, 7, 23,
+ // Second 128-bit lane.
+ 10, 26, 11, 27, 14, 30, 15, 31}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16f32, V1, V2);
- // Vector-element-to-vector
- SDValue InsertpsMask = DAG.getIntPtrConstant(DestIndex << 4 | SrcIndex << 6);
- return DAG.getNode(X86ISD::INSERTPS, dl, VT, To, From, InsertpsMask);
+ // FIXME: Implement direct support for this type!
+ return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG);
}
-// Reduce a vector shuffle to zext.
-static SDValue LowerVectorIntExtend(SDValue Op, const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- // PMOVZX is only available from SSE41.
- if (!Subtarget->hasSSE41())
- return SDValue();
+/// \brief Handle lowering of 8-lane 64-bit integer shuffles.
+static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
- MVT VT = Op.getSimpleValueType();
+ // X86 has dedicated unpack instructions that can handle specific blend
+ // operations: UNPCKH and UNPCKL.
+ if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i64, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i64, V1, V2);
- // Only AVX2 support 256-bit vector integer extending.
- if (!Subtarget->hasInt256() && VT.is256BitVector())
- return SDValue();
+ // FIXME: Implement direct support for this type!
+ return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG);
+}
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+/// \brief Handle lowering of 16-lane 32-bit integer shuffles.
+static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
SDLoc DL(Op);
- SDValue V1 = Op.getOperand(0);
- SDValue V2 = Op.getOperand(1);
- unsigned NumElems = VT.getVectorNumElements();
+ assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
- // Extending is an unary operation and the element type of the source vector
- // won't be equal to or larger than i64.
- if (V2.getOpcode() != ISD::UNDEF || !VT.isInteger() ||
- VT.getVectorElementType() == MVT::i64)
- return SDValue();
+ // Use dedicated unpack instructions for masks that match their pattern.
+ if (isShuffleEquivalent(V1, V2, Mask,
+ {// First 128-bit lane.
+ 0, 16, 1, 17, 4, 20, 5, 21,
+ // Second 128-bit lane.
+ 8, 24, 9, 25, 12, 28, 13, 29}))
+ return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i32, V1, V2);
+ if (isShuffleEquivalent(V1, V2, Mask,
+ {// First 128-bit lane.
+ 2, 18, 3, 19, 6, 22, 7, 23,
+ // Second 128-bit lane.
+ 10, 26, 11, 27, 14, 30, 15, 31}))
+ return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i32, V1, V2);
- // Find the expansion ratio, e.g. expanding from i8 to i32 has a ratio of 4.
- unsigned Shift = 1; // Start from 2, i.e. 1 << 1.
- while ((1U << Shift) < NumElems) {
- if (SVOp->getMaskElt(1U << Shift) == 1)
- break;
- Shift += 1;
- // The maximal ratio is 8, i.e. from i8 to i64.
- if (Shift > 3)
- return SDValue();
- }
+ // FIXME: Implement direct support for this type!
+ return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
+}
- // Check the shuffle mask.
- unsigned Mask = (1U << Shift) - 1;
- for (unsigned i = 0; i != NumElems; ++i) {
- int EltIdx = SVOp->getMaskElt(i);
- if ((i & Mask) != 0 && EltIdx != -1)
- return SDValue();
- if ((i & Mask) == 0 && (unsigned)EltIdx != (i >> Shift))
- return SDValue();
- }
+/// \brief Handle lowering of 32-lane 16-bit integer shuffles.
+static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
+ assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
- unsigned NBits = VT.getVectorElementType().getSizeInBits() << Shift;
- MVT NeVT = MVT::getIntegerVT(NBits);
- MVT NVT = MVT::getVectorVT(NeVT, NumElems >> Shift);
+ // FIXME: Implement direct support for this type!
+ return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG);
+}
- if (!DAG.getTargetLoweringInfo().isTypeLegal(NVT))
- return SDValue();
+/// \brief Handle lowering of 64-lane 8-bit integer shuffles.
+static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
+ assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
+ assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
- return DAG.getNode(ISD::BITCAST, DL, VT,
- DAG.getNode(X86ISD::VZEXT, DL, NVT, V1));
+ // FIXME: Implement direct support for this type!
+ return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
}
-static SDValue NormalizeVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
+/// \brief High-level routine to lower various 512-bit x86 vector shuffles.
+///
+/// This routine either breaks down the specific type of a 512-bit x86 vector
+/// shuffle or splits it into two 256-bit shuffles and fuses the results back
+/// together based on the available instructions.
+static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ MVT VT, const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- MVT VT = Op.getSimpleValueType();
- SDLoc dl(Op);
- SDValue V1 = Op.getOperand(0);
- SDValue V2 = Op.getOperand(1);
-
- if (isZeroShuffle(SVOp))
- return getZeroVector(VT, Subtarget, DAG, dl);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Subtarget->hasAVX512() &&
+ "Cannot lower 512-bit vectors w/ basic ISA!");
- // Handle splat operations
- if (SVOp->isSplat()) {
- // Use vbroadcast whenever the splat comes from a foldable load
- SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG);
- if (Broadcast.getNode())
- return Broadcast;
- }
+ // Check for being able to broadcast a single element.
+ if (SDValue Broadcast =
+ lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
+ return Broadcast;
- // Check integer expanding shuffles.
- SDValue NewOp = LowerVectorIntExtend(Op, Subtarget, DAG);
- if (NewOp.getNode())
- return NewOp;
+ // Dispatch to each element type for lowering. If we don't have supprot for
+ // specific element type shuffles at 512 bits, immediately split them and
+ // lower them. Each lowering routine of a given type is allowed to assume that
+ // the requisite ISA extensions for that element type are available.
+ switch (VT.SimpleTy) {
+ case MVT::v8f64:
+ return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
+ case MVT::v16f32:
+ return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
+ case MVT::v8i64:
+ return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
+ case MVT::v16i32:
+ return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
+ case MVT::v32i16:
+ if (Subtarget->hasBWI())
+ return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
+ break;
+ case MVT::v64i8:
+ if (Subtarget->hasBWI())
+ return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
+ break;
- // If the shuffle can be profitably rewritten as a narrower shuffle, then
- // do it!
- if (VT == MVT::v8i16 || VT == MVT::v16i8 || VT == MVT::v16i16 ||
- VT == MVT::v32i8) {
- SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
- if (NewOp.getNode())
- return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
- } else if (VT.is128BitVector() && Subtarget->hasSSE2()) {
- // FIXME: Figure out a cleaner way to do this.
- if (ISD::isBuildVectorAllZeros(V2.getNode())) {
- SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
- if (NewOp.getNode()) {
- MVT NewVT = NewOp.getSimpleValueType();
- if (isCommutedMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(),
- NewVT, true, false))
- return getVZextMovL(VT, NewVT, NewOp.getOperand(0), DAG, Subtarget,
- dl);
- }
- } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
- SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
- if (NewOp.getNode()) {
- MVT NewVT = NewOp.getSimpleValueType();
- if (isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(), NewVT))
- return getVZextMovL(VT, NewVT, NewOp.getOperand(1), DAG, Subtarget,
- dl);
- }
- }
+ default:
+ llvm_unreachable("Not a valid 512-bit x86 vector type!");
}
- return SDValue();
+
+ // Otherwise fall back on splitting.
+ return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
}
-SDValue
-X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
+/// \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) {
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
SDValue V1 = Op.getOperand(0);
SDValue V2 = Op.getOperand(1);
MVT VT = Op.getSimpleValueType();
+ int NumElements = VT.getVectorNumElements();
SDLoc dl(Op);
- unsigned NumElems = VT.getVectorNumElements();
- bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
- bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
- bool V1IsSplat = false;
- bool V2IsSplat = false;
- bool HasSSE2 = Subtarget->hasSSE2();
- bool HasFp256 = Subtarget->hasFp256();
- bool HasInt256 = Subtarget->hasInt256();
- MachineFunction &MF = DAG.getMachineFunction();
- bool OptForSize = MF.getFunction()->getAttributes().
- hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
-
- // Check if we should use the experimental vector shuffle lowering. If so,
- // delegate completely to that code path.
- if (ExperimentalVectorShuffleLowering)
- return lowerVectorShuffle(Op, Subtarget, DAG);
assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
+ bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
+ bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
if (V1IsUndef && V2IsUndef)
return DAG.getUNDEF(VT);
if (V1IsUndef)
return DAG.getCommutedVectorShuffle(*SVOp);
- // Vector shuffle lowering takes 3 steps:
- //
- // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
- // narrowing and commutation of operands should be handled.
- // 2) Matching of shuffles with known shuffle masks to x86 target specific
- // shuffle nodes.
- // 3) Rewriting of unmatched masks into new generic shuffle operations,
- // so the shuffle can be broken into other shuffles and the legalizer can
- // try the lowering again.
- //
- // The general idea is that no vector_shuffle operation should be left to
- // be matched during isel, all of them must be converted to a target specific
- // node here.
-
- // Normalize the input vectors. Here splats, zeroed vectors, profitable
- // narrowing and commutation of operands should be handled. The actual code
- // doesn't include all of those, work in progress...
- SDValue NewOp = NormalizeVectorShuffle(Op, Subtarget, DAG);
- if (NewOp.getNode())
- return NewOp;
-
- SmallVector<int, 8> M(SVOp->getMask().begin(), SVOp->getMask().end());
-
- // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
- // unpckh_undef). Only use pshufd if speed is more important than size.
- if (OptForSize && isUNPCKL_v_undef_Mask(M, VT, HasInt256))
- return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
- if (OptForSize && isUNPCKH_v_undef_Mask(M, VT, HasInt256))
- return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
-
- if (isMOVDDUPMask(M, VT) && Subtarget->hasSSE3() &&
- V2IsUndef && MayFoldVectorLoad(V1))
- return getMOVDDup(Op, dl, V1, DAG);
-
- if (isMOVHLPS_v_undef_Mask(M, VT))
- return getMOVHighToLow(Op, dl, DAG);
-
- // Use to match splats
- if (HasSSE2 && isUNPCKHMask(M, VT, HasInt256) && V2IsUndef &&
- (VT == MVT::v2f64 || VT == MVT::v2i64))
- return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
-
- if (isPSHUFDMask(M, VT)) {
- // The actual implementation will match the mask in the if above and then
- // during isel it can match several different instructions, not only pshufd
- // as its name says, sad but true, emulate the behavior for now...
- if (isMOVDDUPMask(M, VT) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
- return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
-
- unsigned TargetMask = getShuffleSHUFImmediate(SVOp);
-
- if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
- return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
-
- if (HasFp256 && (VT == MVT::v4f32 || VT == MVT::v2f64))
- return getTargetShuffleNode(X86ISD::VPERMILPI, dl, VT, V1, TargetMask,
- DAG);
-
- return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V1,
- TargetMask, DAG);
- }
-
- if (isPALIGNRMask(M, VT, Subtarget))
- return getTargetShuffleNode(X86ISD::PALIGNR, dl, VT, V1, V2,
- 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 ShVal;
- bool isShift = HasSSE2 && isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
- if (isShift && ShVal.hasOneUse()) {
- // If the shifted value has multiple uses, it may be cheaper to use
- // v_set0 + movlhps or movhlps, etc.
- MVT EltVT = VT.getVectorElementType();
- ShAmt *= EltVT.getSizeInBits();
- return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
- }
-
- if (isMOVLMask(M, VT)) {
- if (ISD::isBuildVectorAllZeros(V1.getNode()))
- return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
- if (!isMOVLPMask(M, VT)) {
- if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
- return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
-
- if (VT == MVT::v4i32 || VT == MVT::v4f32)
- return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
- }
- }
-
- // FIXME: fold these into legal mask.
- if (isMOVLHPSMask(M, VT) && !isUNPCKLMask(M, VT, HasInt256))
- return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
-
- if (isMOVHLPSMask(M, VT))
- return getMOVHighToLow(Op, dl, DAG);
-
- if (V2IsUndef && isMOVSHDUPMask(M, VT, Subtarget))
- return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
-
- if (V2IsUndef && isMOVSLDUPMask(M, VT, Subtarget))
- return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
-
- if (isMOVLPMask(M, VT))
- return getMOVLP(Op, dl, DAG, HasSSE2);
-
- if (ShouldXformToMOVHLPS(M, VT) ||
- ShouldXformToMOVLP(V1.getNode(), V2.getNode(), M, VT))
- return DAG.getCommutedVectorShuffle(*SVOp);
-
- if (isShift) {
- // No better options. Use a vshldq / vsrldq.
- MVT EltVT = VT.getVectorElementType();
- ShAmt *= EltVT.getSizeInBits();
- return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
- }
-
- bool Commuted = false;
- // FIXME: This should also accept a bitcast of a splat? Be careful, not
- // 1,1,1,1 -> v8i16 though.
- BitVector UndefElements;
- if (auto *BVOp = dyn_cast<BuildVectorSDNode>(V1.getNode()))
- if (BVOp->getConstantSplatNode(&UndefElements) && UndefElements.none())
- V1IsSplat = true;
- if (auto *BVOp = dyn_cast<BuildVectorSDNode>(V2.getNode()))
- if (BVOp->getConstantSplatNode(&UndefElements) && UndefElements.none())
- V2IsSplat = true;
-
- // Canonicalize the splat or undef, if present, to be on the RHS.
- if (!V2IsUndef && V1IsSplat && !V2IsSplat) {
- CommuteVectorShuffleMask(M, NumElems);
- std::swap(V1, V2);
- std::swap(V1IsSplat, V2IsSplat);
- Commuted = true;
- }
-
- if (isCommutedMOVLMask(M, VT, V2IsSplat, V2IsUndef)) {
- // Shuffling low element of v1 into undef, just return v1.
- if (V2IsUndef)
- return V1;
- // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
- // the instruction selector will not match, so get a canonical MOVL with
- // swapped operands to undo the commute.
- return getMOVL(DAG, dl, VT, V2, V1);
- }
-
- if (isUNPCKLMask(M, VT, HasInt256))
- return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
-
- if (isUNPCKHMask(M, VT, HasInt256))
- return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
-
- if (V2IsSplat) {
- // Normalize mask so all entries that point to V2 points to its first
- // element then try to match unpck{h|l} again. If match, return a
- // new vector_shuffle with the corrected mask.p
- SmallVector<int, 8> NewMask(M.begin(), M.end());
- NormalizeMask(NewMask, NumElems);
- if (isUNPCKLMask(NewMask, VT, HasInt256, true))
- return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
- if (isUNPCKHMask(NewMask, VT, HasInt256, true))
- return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
- }
-
- if (Commuted) {
- // Commute is back and try unpck* again.
- // FIXME: this seems wrong.
- CommuteVectorShuffleMask(M, NumElems);
- std::swap(V1, V2);
- std::swap(V1IsSplat, V2IsSplat);
-
- if (isUNPCKLMask(M, VT, HasInt256))
- return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
-
- if (isUNPCKHMask(M, VT, HasInt256))
- return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
- }
-
- // Normalize the node to match x86 shuffle ops if needed
- if (!V2IsUndef && (isSHUFPMask(M, VT, /* Commuted */ true)))
- return DAG.getCommutedVectorShuffle(*SVOp);
-
- // The checks below are all present in isShuffleMaskLegal, but they are
- // inlined here right now to enable us to directly emit target specific
- // nodes, and remove one by one until they don't return Op anymore.
-
- if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
- SVOp->getSplatIndex() == 0 && V2IsUndef) {
- if (VT == MVT::v2f64 || VT == MVT::v2i64)
- return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
- }
-
- if (isPSHUFHWMask(M, VT, HasInt256))
- return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
- getShufflePSHUFHWImmediate(SVOp),
- DAG);
-
- if (isPSHUFLWMask(M, VT, HasInt256))
- return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
- getShufflePSHUFLWImmediate(SVOp),
- DAG);
-
- unsigned MaskValue;
- if (isBlendMask(M, VT, Subtarget->hasSSE41(), Subtarget->hasInt256(),
- &MaskValue))
- return LowerVECTOR_SHUFFLEtoBlend(SVOp, MaskValue, Subtarget, DAG);
-
- if (isSHUFPMask(M, VT))
- return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V2,
- getShuffleSHUFImmediate(SVOp), DAG);
-
- if (isUNPCKL_v_undef_Mask(M, VT, HasInt256))
- return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
- if (isUNPCKH_v_undef_Mask(M, VT, HasInt256))
- return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
-
- //===--------------------------------------------------------------------===//
- // Generate target specific nodes for 128 or 256-bit shuffles only
- // supported in the AVX instruction set.
- //
-
- // Handle VMOVDDUPY permutations
- if (V2IsUndef && isMOVDDUPYMask(M, VT, HasFp256))
- return getTargetShuffleNode(X86ISD::MOVDDUP, dl, VT, V1, DAG);
-
- // Handle VPERMILPS/D* permutations
- if (isVPERMILPMask(M, VT)) {
- if ((HasInt256 && VT == MVT::v8i32) || VT == MVT::v16i32)
- return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1,
- getShuffleSHUFImmediate(SVOp), DAG);
- return getTargetShuffleNode(X86ISD::VPERMILPI, dl, VT, V1,
- getShuffleSHUFImmediate(SVOp), DAG);
- }
-
- unsigned Idx;
- if (VT.is512BitVector() && isINSERT64x4Mask(M, VT, &Idx))
- return Insert256BitVector(V1, Extract256BitVector(V2, 0, DAG, dl),
- Idx*(NumElems/2), DAG, dl);
-
- // Handle VPERM2F128/VPERM2I128 permutations
- if (isVPERM2X128Mask(M, VT, HasFp256))
- return getTargetShuffleNode(X86ISD::VPERM2X128, dl, VT, V1,
- V2, getShuffleVPERM2X128Immediate(SVOp), DAG);
-
- if (Subtarget->hasSSE41() && isINSERTPSMask(M, VT))
- return getINSERTPS(SVOp, dl, DAG);
+ // Check for non-undef masks pointing at an undef vector and make the masks
+ // undef as well. This makes it easier to match the shuffle based solely on
+ // the mask.
+ if (V2IsUndef)
+ for (int M : Mask)
+ if (M >= NumElements) {
+ SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
+ for (int &M : NewMask)
+ if (M >= NumElements)
+ M = -1;
+ return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
+ }
- unsigned Imm8;
- if (V2IsUndef && HasInt256 && isPermImmMask(M, VT, Imm8))
- return getTargetShuffleNode(X86ISD::VPERMI, dl, VT, V1, Imm8, DAG);
+ // We actually see shuffles that are entirely re-arrangements of a set of
+ // zero inputs. This mostly happens while decomposing complex shuffles into
+ // simple ones. Directly lower these as a buildvector of zeros.
+ SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
+ if (Zeroable.all())
+ return getZeroVector(VT, Subtarget, DAG, dl);
- if ((V2IsUndef && HasInt256 && VT.is256BitVector() && NumElems == 8) ||
- VT.is512BitVector()) {
- MVT MaskEltVT = MVT::getIntegerVT(VT.getVectorElementType().getSizeInBits());
- MVT MaskVectorVT = MVT::getVectorVT(MaskEltVT, NumElems);
- SmallVector<SDValue, 16> permclMask;
- for (unsigned i = 0; i != NumElems; ++i) {
- permclMask.push_back(DAG.getConstant((M[i]>=0) ? M[i] : 0, MaskEltVT));
+ // Try to collapse shuffles into using a vector type with fewer elements but
+ // wider element types. We cap this to not form integers or floating point
+ // elements wider than 64 bits, but it might be interesting to form i128
+ // integers to handle flipping the low and high halves of AVX 256-bit vectors.
+ SmallVector<int, 16> WidenedMask;
+ if (VT.getScalarSizeInBits() < 64 &&
+ canWidenShuffleElements(Mask, WidenedMask)) {
+ MVT NewEltVT = VT.isFloatingPoint()
+ ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
+ : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
+ MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
+ // Make sure that the new vector type is legal. For example, v2f64 isn't
+ // legal on SSE1.
+ if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
+ V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
+ V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
+ return DAG.getNode(ISD::BITCAST, dl, VT,
+ DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
}
-
- SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVectorVT, permclMask);
- if (V2IsUndef)
- // Bitcast is for VPERMPS since mask is v8i32 but node takes v8f32
- return DAG.getNode(X86ISD::VPERMV, dl, VT,
- DAG.getNode(ISD::BITCAST, dl, VT, Mask), V1);
- return DAG.getNode(X86ISD::VPERMV3, dl, VT, V1,
- DAG.getNode(ISD::BITCAST, dl, VT, Mask), V2);
}
- //===--------------------------------------------------------------------===//
- // Since no target specific shuffle was selected for this generic one,
- // lower it into other known shuffles. FIXME: this isn't true yet, but
- // this is the plan.
- //
-
- // Handle v8i16 specifically since SSE can do byte extraction and insertion.
- if (VT == MVT::v8i16) {
- SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, Subtarget, DAG);
- if (NewOp.getNode())
- return NewOp;
- }
+ int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
+ for (int M : SVOp->getMask())
+ if (M < 0)
+ ++NumUndefElements;
+ else if (M < NumElements)
+ ++NumV1Elements;
+ else
+ ++NumV2Elements;
- if (VT == MVT::v16i16 && Subtarget->hasInt256()) {
- SDValue NewOp = LowerVECTOR_SHUFFLEv16i16(Op, DAG);
- if (NewOp.getNode())
- return NewOp;
- }
+ // Commute the shuffle as needed such that more elements come from V1 than
+ // V2. This allows us to match the shuffle pattern strictly on how many
+ // elements come from V1 without handling the symmetric cases.
+ if (NumV2Elements > NumV1Elements)
+ return DAG.getCommutedVectorShuffle(*SVOp);
- if (VT == MVT::v16i8) {
- SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, Subtarget, DAG);
- if (NewOp.getNode())
- return NewOp;
+ // When the number of V1 and V2 elements are the same, try to minimize the
+ // number of uses of V2 in the low half of the vector. When that is tied,
+ // ensure that the sum of indices for V1 is equal to or lower than the sum
+ // indices for V2. When those are equal, try to ensure that the number of odd
+ // indices for V1 is lower than the number of odd indices for V2.
+ if (NumV1Elements == NumV2Elements) {
+ int LowV1Elements = 0, LowV2Elements = 0;
+ for (int M : SVOp->getMask().slice(0, NumElements / 2))
+ if (M >= NumElements)
+ ++LowV2Elements;
+ else if (M >= 0)
+ ++LowV1Elements;
+ if (LowV2Elements > LowV1Elements) {
+ return DAG.getCommutedVectorShuffle(*SVOp);
+ } else if (LowV2Elements == LowV1Elements) {
+ int SumV1Indices = 0, SumV2Indices = 0;
+ for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
+ if (SVOp->getMask()[i] >= NumElements)
+ SumV2Indices += i;
+ else if (SVOp->getMask()[i] >= 0)
+ SumV1Indices += i;
+ if (SumV2Indices < SumV1Indices) {
+ return DAG.getCommutedVectorShuffle(*SVOp);
+ } else if (SumV2Indices == SumV1Indices) {
+ int NumV1OddIndices = 0, NumV2OddIndices = 0;
+ for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
+ if (SVOp->getMask()[i] >= NumElements)
+ NumV2OddIndices += i % 2;
+ else if (SVOp->getMask()[i] >= 0)
+ NumV1OddIndices += i % 2;
+ if (NumV2OddIndices < NumV1OddIndices)
+ return DAG.getCommutedVectorShuffle(*SVOp);
+ }
+ }
}
- if (VT == MVT::v32i8) {
- SDValue NewOp = LowerVECTOR_SHUFFLEv32i8(SVOp, Subtarget, DAG);
- if (NewOp.getNode())
- return NewOp;
- }
+ // For each vector width, delegate to a specialized lowering routine.
+ if (VT.getSizeInBits() == 128)
+ return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
- // Handle all 128-bit wide vectors with 4 elements, and match them with
- // several different shuffle types.
- if (NumElems == 4 && VT.is128BitVector())
- return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
+ if (VT.getSizeInBits() == 256)
+ return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
- // Handle general 256-bit shuffles
- if (VT.is256BitVector())
- return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
+ // Force AVX-512 vectors to be scalarized for now.
+ // FIXME: Implement AVX-512 support!
+ if (VT.getSizeInBits() == 512)
+ return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
- return SDValue();
+ llvm_unreachable("Unimplemented!");
}
// This function assumes its argument is a BUILD_VECTOR of constants or
return true;
}
-/// \brief Try to lower a VSELECT instruction to an immediate-controlled blend
-/// instruction.
-static SDValue lowerVSELECTtoBLENDI(SDValue Op, const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
+/// \brief Try to lower a VSELECT instruction to a vector shuffle.
+static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
+ const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
SDValue Cond = Op.getOperand(0);
SDValue LHS = Op.getOperand(1);
SDValue RHS = Op.getOperand(2);
SDLoc dl(Op);
MVT VT = Op.getSimpleValueType();
- MVT EltVT = VT.getVectorElementType();
- unsigned NumElems = VT.getVectorNumElements();
-
- // There is no blend with immediate in AVX-512.
- if (VT.is512BitVector())
- return SDValue();
-
- if (!Subtarget->hasSSE41() || EltVT == MVT::i8)
- return SDValue();
- if (!Subtarget->hasInt256() && VT == MVT::v16i16)
- return SDValue();
if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
return SDValue();
+ auto *CondBV = cast<BuildVectorSDNode>(Cond);
- // Check the mask for BLEND and build the value.
- unsigned MaskValue = 0;
- if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
- return SDValue();
-
- // Convert i32 vectors to floating point if it is not AVX2.
- // AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors.
- MVT BlendVT = VT;
- if (EltVT == MVT::i64 || (EltVT == MVT::i32 && !Subtarget->hasInt256())) {
- BlendVT = MVT::getVectorVT(MVT::getFloatingPointVT(EltVT.getSizeInBits()),
- NumElems);
- LHS = DAG.getNode(ISD::BITCAST, dl, VT, LHS);
- RHS = DAG.getNode(ISD::BITCAST, dl, VT, RHS);
+ // Only non-legal VSELECTs reach this lowering, convert those into generic
+ // shuffles and re-use the shuffle lowering path for blends.
+ SmallVector<int, 32> Mask;
+ for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
+ SDValue CondElt = CondBV->getOperand(i);
+ Mask.push_back(
+ isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
}
-
- SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, LHS, RHS,
- DAG.getConstant(MaskValue, MVT::i32));
- return DAG.getNode(ISD::BITCAST, dl, VT, Ret);
+ return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
}
SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
return SDValue();
- SDValue BlendOp = lowerVSELECTtoBLENDI(Op, Subtarget, DAG);
+ // Try to lower this to a blend-style vector shuffle. This can handle all
+ // constant condition cases.
+ SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG);
if (BlendOp.getNode())
return BlendOp;
- // Some types for vselect were previously set to Expand, not Legal or
- // Custom. Return an empty SDValue so we fall-through to Expand, after
- // the Custom lowering phase.
- MVT VT = Op.getSimpleValueType();
- switch (VT.SimpleTy) {
+ // Variable blends are only legal from SSE4.1 onward.
+ if (!Subtarget->hasSSE41())
+ return SDValue();
+
+ // Only some types will be legal on some subtargets. If we can emit a legal
+ // VSELECT-matching blend, return Op, and but if we need to expand, return
+ // a null value.
+ switch (Op.getSimpleValueType().SimpleTy) {
default:
- break;
+ // Most of the vector types have blends past SSE4.1.
+ return Op;
+
+ case MVT::v32i8:
+ // The byte blends for AVX vectors were introduced only in AVX2.
+ if (Subtarget->hasAVX2())
+ return Op;
+
+ return SDValue();
+
case MVT::v8i16:
case MVT::v16i16:
+ // AVX-512 BWI and VLX features support VSELECT with i16 elements.
if (Subtarget->hasBWI() && Subtarget->hasVLX())
- break;
+ return Op;
+
+ // FIXME: We should custom lower this by fixing the condition and using i8
+ // blends.
return SDValue();
}
-
- // We couldn't create a "Blend with immediate" node.
- // This node should still be legal, but we'll have to emit a blendv*
- // instruction.
- return Op;
}
static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
MVT EltVT = Op.getSimpleValueType();
assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
+ assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
+ "Unexpected vector type in ExtractBitFromMaskVector");
// variable index can't be handled in mask registers,
// extend vector to VR512
unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
const TargetRegisterClass* rc = getRegClassFor(VecVT);
+ if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
+ rc = getRegClassFor(MVT::v16i1);
unsigned MaxSift = rc->getSize()*8 - 1;
Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
DAG.getConstant(MaxSift - IdxVal, MVT::i8));
// the upper bits of a vector.
static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
SelectionDAG &DAG) {
- if (Subtarget->hasFp256()) {
- SDLoc dl(Op.getNode());
- SDValue Vec = Op.getNode()->getOperand(0);
- SDValue SubVec = Op.getNode()->getOperand(1);
- SDValue Idx = Op.getNode()->getOperand(2);
-
- if ((Op.getNode()->getSimpleValueType(0).is256BitVector() ||
- Op.getNode()->getSimpleValueType(0).is512BitVector()) &&
- SubVec.getNode()->getSimpleValueType(0).is128BitVector() &&
- isa<ConstantSDNode>(Idx)) {
- unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
- return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
- }
+ if (!Subtarget->hasAVX())
+ return SDValue();
- if (Op.getNode()->getSimpleValueType(0).is512BitVector() &&
- SubVec.getNode()->getSimpleValueType(0).is256BitVector() &&
- isa<ConstantSDNode>(Idx)) {
- unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
- return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
+ SDLoc dl(Op);
+ SDValue Vec = Op.getOperand(0);
+ SDValue SubVec = Op.getOperand(1);
+ SDValue Idx = Op.getOperand(2);
+
+ if (!isa<ConstantSDNode>(Idx))
+ return SDValue();
+
+ unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
+ MVT OpVT = Op.getSimpleValueType();
+ MVT SubVecVT = SubVec.getSimpleValueType();
+
+ // Fold two 16-byte subvector loads into one 32-byte load:
+ // (insert_subvector (insert_subvector undef, (load addr), 0),
+ // (load addr + 16), Elts/2)
+ // --> load32 addr
+ if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
+ Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
+ OpVT.is256BitVector() && SubVecVT.is128BitVector() &&
+ !Subtarget->isUnalignedMem32Slow()) {
+ SDValue SubVec2 = Vec.getOperand(1);
+ if (auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2))) {
+ if (Idx2->getZExtValue() == 0) {
+ SDValue Ops[] = { SubVec2, SubVec };
+ SDValue LD = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false);
+ if (LD.getNode())
+ return LD;
+ }
}
}
+
+ if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
+ SubVecVT.is128BitVector())
+ return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
+
+ if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
+ return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
+
return SDValue();
}
/// equivalent.
SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
SelectionDAG &DAG) const {
- if (Op.getValueType() == MVT::i1)
- // KORTEST instruction should be selected
- return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
- DAG.getConstant(0, Op.getValueType()));
-
+ if (Op.getValueType() == MVT::i1) {
+ SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
+ return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
+ DAG.getConstant(0, MVT::i8));
+ }
// CF and OF aren't always set the way we want. Determine which
// of these we need.
bool NeedCF = false;
DAG.getConstant(0, Op.getValueType()));
SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
- SmallVector<SDValue, 4> Ops;
- for (unsigned i = 0; i != NumOperands; ++i)
- Ops.push_back(Op.getOperand(i));
+ SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
DAG.ReplaceAllUsesWith(Op, New);
// if we're optimizing for size, however, as that'll allow better folding
// of memory operations.
if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
- !DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute(
- AttributeSet::FunctionIndex, Attribute::MinSize) &&
+ !DAG.getMachineFunction().getFunction()->hasFnAttribute(
+ Attribute::MinSize) &&
!Subtarget->isAtom()) {
unsigned ExtendOp =
isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
// 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.
+// FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
// 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,
SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
Chain = SP.getValue(1);
unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
- const TargetFrameLowering &TFI = *DAG.getSubtarget().getFrameLowering();
+ const TargetFrameLowering &TFI = *Subtarget->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.getSubtarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
unsigned SPReg = RegInfo->getStackRegister();
SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
Chain = SP.getValue(1);
if (ArgMode == 2) {
// Sanity Check: Make sure using fp_offset makes sense.
assert(!DAG.getTarget().Options.UseSoftFloat &&
- !(DAG.getMachineFunction()
- .getFunction()->getAttributes()
- .hasAttribute(AttributeSet::FunctionIndex,
- Attribute::NoImplicitFloat)) &&
+ !(DAG.getMachineFunction().getFunction()->hasFnAttribute(
+ Attribute::NoImplicitFloat)) &&
Subtarget->hasSSE1());
}
// Insert VAARG_64 node into the DAG
// VAARG_64 returns two values: Variable Argument Address, Chain
- SmallVector<SDValue, 11> InstOps;
- InstOps.push_back(Chain);
- InstOps.push_back(SrcPtr);
- InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
- InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
- InstOps.push_back(DAG.getConstant(Align, MVT::i32));
+ SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, MVT::i32),
+ DAG.getConstant(ArgMode, MVT::i8),
+ DAG.getConstant(Align, MVT::i32)};
SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
VTs, InstOps, MVT::i64,
}
const X86Subtarget &Subtarget =
- DAG.getTarget().getSubtarget<X86Subtarget>();
+ static_cast<const X86Subtarget &>(DAG.getSubtarget());
if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
// Let the shuffle legalizer expand this shift amount node.
return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, 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, const X86Subtarget *Subtarget,
SelectionDAG &DAG) {
SDLoc dl(Op);
SDValue Src2 = Op.getOperand(2);
SDValue Src0 = Op.getOperand(3);
SDValue Mask = Op.getOperand(4);
- SDValue RoundingMode = Op.getOperand(5);
+ // There are 2 kinds of intrinsics in this group:
+ // (1) With supress-all-exceptions (sae) - 6 operands
+ // (2) With rounding mode and sae - 7 operands.
+ if (Op.getNumOperands() == 6) {
+ SDValue Sae = Op.getOperand(5);
+ return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
+ Sae),
+ Mask, Src0, Subtarget, DAG);
+ }
+ assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
+ SDValue RoundingMode = Op.getOperand(5);
+ SDValue Sae = Op.getOperand(6);
return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
- RoundingMode),
+ RoundingMode, Sae),
Mask, Src0, Subtarget, DAG);
}
case INTR_TYPE_2OP_MASK: {
- return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Op.getOperand(1),
- Op.getOperand(2)),
- Op.getOperand(4), Op.getOperand(3), Subtarget, DAG);
+ SDValue Src1 = Op.getOperand(1);
+ SDValue Src2 = Op.getOperand(2);
+ SDValue PassThru = Op.getOperand(3);
+ SDValue Mask = Op.getOperand(4);
+ // We specify 2 possible opcodes for intrinsics with rounding modes.
+ // First, we check if the intrinsic may have non-default rounding mode,
+ // (IntrData->Opc1 != 0), then we check the rounding mode operand.
+ unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
+ if (IntrWithRoundingModeOpcode != 0) {
+ SDValue Rnd = Op.getOperand(5);
+ unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
+ if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
+ return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
+ dl, Op.getValueType(),
+ Src1, Src2, Rnd),
+ Mask, PassThru, Subtarget, DAG);
+ }
+ }
+ return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
+ Src1,Src2),
+ Mask, PassThru, Subtarget, DAG);
+ }
+ case FMA_OP_MASK: {
+ SDValue Src1 = Op.getOperand(1);
+ SDValue Src2 = Op.getOperand(2);
+ SDValue Src3 = Op.getOperand(3);
+ SDValue Mask = Op.getOperand(4);
+ // We specify 2 possible opcodes for intrinsics with rounding modes.
+ // First, we check if the intrinsic may have non-default rounding mode,
+ // (IntrData->Opc1 != 0), then we check the rounding mode operand.
+ unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
+ if (IntrWithRoundingModeOpcode != 0) {
+ SDValue Rnd = Op.getOperand(5);
+ if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
+ X86::STATIC_ROUNDING::CUR_DIRECTION)
+ return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
+ dl, Op.getValueType(),
+ Src1, Src2, Src3, Rnd),
+ Mask, Src1, Subtarget, DAG);
+ }
+ return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
+ dl, Op.getValueType(),
+ Src1, Src2, Src3),
+ Mask, Src1, Subtarget, DAG);
}
case CMP_MASK:
case CMP_MASK_CC: {
return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
Op.getOperand(2));
}
- case FMA_OP_MASK:
- {
- return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
- dl, Op.getValueType(),
- Op.getOperand(1),
- Op.getOperand(2),
- Op.getOperand(3)),
- Op.getOperand(4), Op.getOperand(1),
- Subtarget, DAG);
- }
default:
break;
}
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),
- Subtarget, 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_vfmsub_pd:
- case Intrinsic::x86_fma_vfnmadd_ps:
- case Intrinsic::x86_fma_vfnmadd_pd:
- case Intrinsic::x86_fma_vfnmsub_ps:
- case Intrinsic::x86_fma_vfnmsub_pd:
- case Intrinsic::x86_fma_vfmaddsub_ps:
- case Intrinsic::x86_fma_vfmaddsub_pd:
- case Intrinsic::x86_fma_vfmsubadd_ps:
- case Intrinsic::x86_fma_vfmsubadd_pd:
- case Intrinsic::x86_fma_vfmadd_ps_256:
- case Intrinsic::x86_fma_vfmadd_pd_256:
- case Intrinsic::x86_fma_vfmsub_ps_256:
- case Intrinsic::x86_fma_vfmsub_pd_256:
- case Intrinsic::x86_fma_vfnmadd_ps_256:
- case Intrinsic::x86_fma_vfnmadd_pd_256:
- case Intrinsic::x86_fma_vfnmsub_ps_256:
- case Intrinsic::x86_fma_vfnmsub_pd_256:
- case Intrinsic::x86_fma_vfmaddsub_ps_256:
- case Intrinsic::x86_fma_vfmaddsub_pd_256:
- case Intrinsic::x86_fma_vfmsubadd_ps_256:
- case Intrinsic::x86_fma_vfmsubadd_pd_256:
- return DAG.getNode(getOpcodeForFMAIntrinsic(IntNo), dl, Op.getValueType(),
- Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
}
}
SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
false, false, false, 0);
- SmallVector<SDValue, 2> Results;
- Results.push_back(DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToExpand,
- PathThru));
- Results.push_back(Chain);
+ SDValue Results[] = {
+ DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToExpand, PathThru),
+ Chain};
return DAG.getMergeValues(Results, dl);
}
}
if (Depth > 0) {
SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
- const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
- DAG.getSubtarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), PtrVT);
return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
DAG.getNode(ISD::ADD, dl, PtrVT,
}
SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
- MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
+ MachineFunction &MF = DAG.getMachineFunction();
+ MachineFrameInfo *MFI = MF.getFrameInfo();
+ X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
+ const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
+ EVT VT = Op.getValueType();
+
MFI->setFrameAddressIsTaken(true);
- EVT VT = Op.getValueType();
+ if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
+ // Depth > 0 makes no sense on targets which use Windows unwind codes. It
+ // is not possible to crawl up the stack without looking at the unwind codes
+ // simultaneously.
+ int FrameAddrIndex = FuncInfo->getFAIndex();
+ if (!FrameAddrIndex) {
+ // Set up a frame object for the return address.
+ unsigned SlotSize = RegInfo->getSlotSize();
+ FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
+ SlotSize, /*Offset=*/INT64_MIN, /*IsImmutable=*/false);
+ FuncInfo->setFAIndex(FrameAddrIndex);
+ }
+ return DAG.getFrameIndex(FrameAddrIndex, VT);
+ }
+
+ unsigned FrameReg =
+ RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
SDLoc dl(Op); // FIXME probably not meaningful
unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
- const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
- DAG.getSubtarget().getRegisterInfo());
- unsigned FrameReg = RegInfo->getPtrSizedFrameRegister(
- DAG.getMachineFunction());
assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
(FrameReg == X86::EBP && VT == MVT::i32)) &&
"Invalid Frame Register!");
SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
SelectionDAG &DAG) const {
- const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
- DAG.getSubtarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize());
}
SDLoc dl (Op);
EVT PtrVT = getPointerTy();
- const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
- DAG.getSubtarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = Subtarget->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.getSubtarget().getRegisterInfo();
+ const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
if (Subtarget->is64Bit()) {
SDValue OutChains[6];
*/
MachineFunction &MF = DAG.getMachineFunction();
- const TargetMachine &TM = MF.getTarget();
- const TargetFrameLowering &TFI = *TM.getSubtargetImpl()->getFrameLowering();
+ const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
unsigned StackAlignment = TFI.getStackAlignment();
MVT VT = Op.getSimpleValueType();
SDLoc DL(Op);
DAG);
}
- if (VT == MVT::v16i8) {
- if (Op.getOpcode() == ISD::SHL) {
- // Make a large shift.
- SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl,
- MVT::v8i16, R, ShiftAmt,
- DAG);
- SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
- // Zero out the rightmost bits.
- SmallVector<SDValue, 16> V(16,
- DAG.getConstant(uint8_t(-1U << ShiftAmt),
- MVT::i8));
- return DAG.getNode(ISD::AND, dl, VT, SHL,
- DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
- }
- if (Op.getOpcode() == ISD::SRL) {
- // Make a large shift.
- SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl,
- MVT::v8i16, R, ShiftAmt,
- DAG);
- SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
- // Zero out the leftmost bits.
- SmallVector<SDValue, 16> V(16,
- DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
- MVT::i8));
- return DAG.getNode(ISD::AND, dl, VT, SRL,
- DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
- }
- if (Op.getOpcode() == ISD::SRA) {
- if (ShiftAmt == 7) {
- // R s>> 7 === R s< 0
- SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
- return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
- }
-
- // R s>> a === ((R u>> a) ^ m) - m
- SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
- SmallVector<SDValue, 16> V(16, DAG.getConstant(128 >> ShiftAmt,
- MVT::i8));
- SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
- Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
- Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
- return Res;
- }
- llvm_unreachable("Unknown shift opcode.");
- }
+ if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
+ unsigned NumElts = VT.getVectorNumElements();
+ MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
- if (Subtarget->hasInt256() && VT == MVT::v32i8) {
if (Op.getOpcode() == ISD::SHL) {
// Make a large shift.
- SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl,
- MVT::v16i16, R, ShiftAmt,
- DAG);
+ SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
+ R, ShiftAmt, DAG);
SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
// Zero out the rightmost bits.
- SmallVector<SDValue, 32> V(32,
- DAG.getConstant(uint8_t(-1U << ShiftAmt),
- MVT::i8));
+ SmallVector<SDValue, 32> V(
+ NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), MVT::i8));
return DAG.getNode(ISD::AND, dl, VT, SHL,
DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
}
if (Op.getOpcode() == ISD::SRL) {
// Make a large shift.
- SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl,
- MVT::v16i16, R, ShiftAmt,
- DAG);
+ SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
+ R, ShiftAmt, DAG);
SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
// Zero out the leftmost bits.
- SmallVector<SDValue, 32> V(32,
- DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
- MVT::i8));
+ SmallVector<SDValue, 32> V(
+ NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, MVT::i8));
return DAG.getNode(ISD::AND, dl, VT, SRL,
DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
}
// R s>> a === ((R u>> a) ^ m) - m
SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
- SmallVector<SDValue, 32> V(32, DAG.getConstant(128 >> ShiftAmt,
- MVT::i8));
+ SmallVector<SDValue, 32> V(NumElts,
+ DAG.getConstant(128 >> ShiftAmt, MVT::i8));
SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
Amt = DAG.getNode(ISD::ANY_EXTEND, dl, NewVT, Amt);
return DAG.getNode(ISD::TRUNCATE, dl, VT,
DAG.getNode(Op.getOpcode(), dl, NewVT, R, Amt));
- }
+ }
// Decompose 256-bit shifts into smaller 128-bit shifts.
if (VT.is256BitVector()) {
SDValue Amt1, Amt2;
if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
// Constant shift amount
- SmallVector<SDValue, 4> Amt1Csts;
- SmallVector<SDValue, 4> Amt2Csts;
- for (unsigned i = 0; i != NumElems/2; ++i)
- Amt1Csts.push_back(Amt->getOperand(i));
- for (unsigned i = NumElems/2; i != NumElems; ++i)
- Amt2Csts.push_back(Amt->getOperand(i));
+ SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
+ ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
+ ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
SDValue SetCC =
- DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
- DAG.getConstant(X86::COND_O, MVT::i32),
- SDValue(Sum.getNode(), 2));
-
- return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
- }
- }
-
- // Also sets EFLAGS.
- SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
- SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
-
- SDValue SetCC =
- DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
- DAG.getConstant(Cond, MVT::i32),
- SDValue(Sum.getNode(), 1));
-
- 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);
- EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
- MVT VT = Op.getSimpleValueType();
-
- if (!Subtarget->hasSSE2() || !VT.isVector())
- return SDValue();
-
- unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
- ExtraVT.getScalarType().getSizeInBits();
-
- switch (VT.SimpleTy) {
- default: return SDValue();
- case MVT::v8i32:
- case MVT::v16i16:
- if (!Subtarget->hasFp256())
- return SDValue();
- if (!Subtarget->hasInt256()) {
- // needs to be split
- unsigned NumElems = VT.getVectorNumElements();
-
- // Extract the LHS vectors
- SDValue LHS = Op.getOperand(0);
- SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
- SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
+ DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
+ DAG.getConstant(X86::COND_O, MVT::i32),
+ SDValue(Sum.getNode(), 2));
- MVT EltVT = VT.getVectorElementType();
- EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
+ return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
+ }
+ }
- EVT ExtraEltVT = ExtraVT.getVectorElementType();
- unsigned ExtraNumElems = ExtraVT.getVectorNumElements();
- ExtraVT = EVT::getVectorVT(*DAG.getContext(), ExtraEltVT,
- ExtraNumElems/2);
- SDValue Extra = DAG.getValueType(ExtraVT);
+ // Also sets EFLAGS.
+ SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
+ SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
- LHS1 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, Extra);
- LHS2 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, Extra);
+ SDValue SetCC =
+ DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
+ DAG.getConstant(Cond, MVT::i32),
+ SDValue(Sum.getNode(), 1));
- return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, LHS1, LHS2);
- }
- // fall through
- case MVT::v4i32:
- case MVT::v8i16: {
- SDValue Op0 = Op.getOperand(0);
-
- // 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);
- }
- }
+ return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
}
/// Returns true if the operand type is exactly twice the native width, and
/// Used to know whether to use cmpxchg8/16b when expanding atomic operations
/// (otherwise we leave them alone to become __sync_fetch_and_... calls).
bool X86TargetLowering::needsCmpXchgNb(const Type *MemType) const {
- const X86Subtarget &Subtarget =
- getTargetMachine().getSubtarget<X86Subtarget>();
unsigned OpWidth = MemType->getPrimitiveSizeInBits();
if (OpWidth == 64)
- return !Subtarget.is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
+ return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
else if (OpWidth == 128)
- return Subtarget.hasCmpxchg16b();
+ return Subtarget->hasCmpxchg16b();
else
return false;
}
}
bool X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
- const X86Subtarget &Subtarget =
- getTargetMachine().getSubtarget<X86Subtarget>();
- unsigned NativeWidth = Subtarget.is64Bit() ? 64 : 32;
+ unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
const Type *MemType = AI->getType();
// If the operand is too big, we must see if cmpxchg8/16b is available
LoadInst *
X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
- const X86Subtarget &Subtarget =
- getTargetMachine().getSubtarget<X86Subtarget>();
- unsigned NativeWidth = Subtarget.is64Bit() ? 64 : 32;
+ unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
const Type *MemType = AI->getType();
// Accesses larger than the native width are turned into cmpxchg/libcalls, so
// there is no benefit in turning such RMWs into loads, and it is actually
// FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
// the IR level, so we must wrap it in an intrinsic.
return nullptr;
- } else if (hasMFENCE(Subtarget)) {
+ } else if (hasMFENCE(*Subtarget)) {
Function *MFence = llvm::Intrinsic::getDeclaration(M,
Intrinsic::x86_sse2_mfence);
Builder.CreateCall(MFence);
DAG.getIntPtrConstant(i)));
// Explicitly mark the extra elements as Undef.
- SDValue Undef = DAG.getUNDEF(SVT);
- for (unsigned i = NumElts, e = NumElts * 2; i != e; ++i)
- Elts.push_back(Undef);
+ Elts.append(NumElts, DAG.getUNDEF(SVT));
EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
switch (Op.getOpcode()) {
default: llvm_unreachable("Should not custom lower this!");
- case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
return LowerCMP_SWAP(Op, Subtarget, DAG);
case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
- case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
+ case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
case ISD::VSELECT: return LowerVSELECT(Op, DAG);
case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
switch (N->getOpcode()) {
default:
llvm_unreachable("Do not know how to custom type legalize this operation!");
+ // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
+ case X86ISD::FMINC:
+ case X86ISD::FMIN:
+ case X86ISD::FMAXC:
+ case X86ISD::FMAX: {
+ EVT VT = N->getValueType(0);
+ if (VT != MVT::v2f32)
+ llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
+ SDValue UNDEF = DAG.getUNDEF(VT);
+ SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
+ N->getOperand(0), UNDEF);
+ SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
+ N->getOperand(1), UNDEF);
+ Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
+ return;
+ }
case ISD::SIGN_EXTEND_INREG:
case ISD::ADDC:
case ISD::ADDE:
case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
case X86ISD::EXPAND: return "X86ISD::EXPAND";
case X86ISD::SELECT: return "X86ISD::SELECT";
+ case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
+ case X86ISD::RCP28: return "X86ISD::RCP28";
+ case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
+ case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
+ case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
+ case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
+ case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
}
}
return false;
}
+bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
+
bool
X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
if (!(Subtarget->hasFMA() || Subtarget->hasFMA4()))
if (!VT.isSimple())
return false;
- MVT SVT = VT.getSimpleVT();
-
// Very little shuffling can be done for 64-bit vectors right now.
if (VT.getSizeInBits() == 64)
return false;
- // If this is a single-input shuffle with no 128 bit lane crossings we can
- // lower it into pshufb.
- if ((SVT.is128BitVector() && Subtarget->hasSSSE3()) ||
- (SVT.is256BitVector() && Subtarget->hasInt256())) {
- bool isLegal = true;
- for (unsigned I = 0, E = M.size(); I != E; ++I) {
- if (M[I] >= (int)SVT.getVectorNumElements() ||
- ShuffleCrosses128bitLane(SVT, I, M[I])) {
- isLegal = false;
- break;
- }
- }
- if (isLegal)
- return true;
- }
-
- // FIXME: blends, shifts.
- return (SVT.getVectorNumElements() == 2 ||
- ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
- isMOVLMask(M, SVT) ||
- isCommutedMOVLMask(M, SVT) ||
- isMOVHLPSMask(M, SVT) ||
- isSHUFPMask(M, SVT) ||
- isSHUFPMask(M, SVT, /* Commuted */ true) ||
- isPSHUFDMask(M, SVT) ||
- isPSHUFDMask(M, SVT, /* SecondOperand */ true) ||
- isPSHUFHWMask(M, SVT, Subtarget->hasInt256()) ||
- isPSHUFLWMask(M, SVT, Subtarget->hasInt256()) ||
- isPALIGNRMask(M, SVT, Subtarget) ||
- isUNPCKLMask(M, SVT, Subtarget->hasInt256()) ||
- isUNPCKHMask(M, SVT, Subtarget->hasInt256()) ||
- isUNPCKL_v_undef_Mask(M, SVT, Subtarget->hasInt256()) ||
- isUNPCKH_v_undef_Mask(M, SVT, Subtarget->hasInt256()) ||
- isBlendMask(M, SVT, Subtarget->hasSSE41(), Subtarget->hasInt256()) ||
- (Subtarget->hasSSE41() && isINSERTPSMask(M, SVT)));
+ // We only care that the types being shuffled are legal. The lowering can
+ // handle any possible shuffle mask that results.
+ return isTypeLegal(VT.getSimpleVT());
}
bool
X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
EVT VT) const {
- if (!VT.isSimple())
- return false;
-
- MVT SVT = VT.getSimpleVT();
- unsigned NumElts = SVT.getVectorNumElements();
- // FIXME: This collection of masks seems suspect.
- if (NumElts == 2)
- return true;
- if (NumElts == 4 && SVT.is128BitVector()) {
- return (isMOVLMask(Mask, SVT) ||
- isCommutedMOVLMask(Mask, SVT, true) ||
- isSHUFPMask(Mask, SVT) ||
- isSHUFPMask(Mask, SVT, /* Commuted */ true) ||
- isBlendMask(Mask, SVT, Subtarget->hasSSE41(),
- Subtarget->hasInt256()));
- }
- return false;
+ // Just delegate to the generic legality, clear masks aren't special.
+ return isShuffleMaskLegal(Mask, VT);
}
//===----------------------------------------------------------------------===//
return BB;
}
-static MachineBasicBlock * EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
- const TargetInstrInfo *TII,
- const X86Subtarget* Subtarget) {
+static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
+ const X86Subtarget *Subtarget) {
DebugLoc dl = MI->getDebugLoc();
-
+ const TargetInstrInfo *TII = Subtarget->getInstrInfo();
// Address into RAX/EAX, other two args into ECX, EDX.
unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
}
MachineBasicBlock *
-X86TargetLowering::EmitVAARG64WithCustomInserter(
- MachineInstr *MI,
- MachineBasicBlock *MBB) const {
+X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
+ MachineBasicBlock *MBB) const {
// Emit va_arg instruction on X86-64.
// Operands to this pseudo-instruction:
MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
// Machine Information
- const TargetInstrInfo *TII = MBB->getParent()->getSubtarget().getInstrInfo();
+ const TargetInstrInfo *TII = Subtarget->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()->getSubtarget().getInstrInfo();
+ const TargetInstrInfo *TII = Subtarget->getInstrInfo();
DebugLoc DL = MI->getDebugLoc();
unsigned CountReg = MI->getOperand(0).getReg();
MachineBasicBlock *
X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
MachineBasicBlock *BB) const {
- const TargetInstrInfo *TII = BB->getParent()->getSubtarget().getInstrInfo();
+ const TargetInstrInfo *TII = Subtarget->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()->getSubtarget().getRegisterInfo();
+ const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
if (!MI->killsRegister(X86::EFLAGS) &&
!checkAndUpdateEFLAGSKill(MI, BB, TRI)) {
copy0MBB->addLiveIn(X86::EFLAGS);
X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
MachineBasicBlock *BB) const {
MachineFunction *MF = BB->getParent();
- const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
+ const TargetInstrInfo *TII = Subtarget->getInstrInfo();
DebugLoc DL = MI->getDebugLoc();
const BasicBlock *LLVM_BB = BB->getBasicBlock();
BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
// Calls into a routine in libgcc to allocate more space from the heap.
- const uint32_t *RegMask = MF->getTarget()
- .getSubtargetImpl()
- ->getRegisterInfo()
- ->getCallPreservedMask(CallingConv::C);
+ const uint32_t *RegMask =
+ Subtarget->getRegisterInfo()->getCallPreservedMask(CallingConv::C);
if (IsLP64) {
BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
.addReg(sizeVReg);
MachineBasicBlock *
X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
MachineBasicBlock *BB) const {
- const TargetInstrInfo *TII = BB->getParent()->getSubtarget().getInstrInfo();
DebugLoc DL = MI->getDebugLoc();
assert(!Subtarget->isTargetMachO());
- // The lowering is pretty easy: we're just emitting the call to _alloca. The
- // non-trivial part is impdef of ESP.
-
- if (Subtarget->isTargetWin64()) {
- if (Subtarget->isTargetCygMing()) {
- // ___chkstk(Mingw64):
- // Clobbers R10, R11, RAX and EFLAGS.
- // Updates RSP.
- BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
- .addExternalSymbol("___chkstk")
- .addReg(X86::RAX, RegState::Implicit)
- .addReg(X86::RSP, RegState::Implicit)
- .addReg(X86::RAX, RegState::Define | RegState::Implicit)
- .addReg(X86::RSP, RegState::Define | RegState::Implicit)
- .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
- } else {
- // __chkstk(MSVCRT): does not update stack pointer.
- // Clobbers R10, R11 and EFLAGS.
- BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
- .addExternalSymbol("__chkstk")
- .addReg(X86::RAX, RegState::Implicit)
- .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
- // RAX has the offset to be subtracted from RSP.
- BuildMI(*BB, MI, DL, TII->get(X86::SUB64rr), X86::RSP)
- .addReg(X86::RSP)
- .addReg(X86::RAX);
- }
- } else {
- const char *StackProbeSymbol = (Subtarget->isTargetKnownWindowsMSVC() ||
- Subtarget->isTargetWindowsItanium())
- ? "_chkstk"
- : "_alloca";
-
- BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
- .addExternalSymbol(StackProbeSymbol)
- .addReg(X86::EAX, RegState::Implicit)
- .addReg(X86::ESP, RegState::Implicit)
- .addReg(X86::EAX, RegState::Define | RegState::Implicit)
- .addReg(X86::ESP, RegState::Define | RegState::Implicit)
- .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
- }
+ X86FrameLowering::emitStackProbeCall(*BB->getParent(), *BB, MI, DL);
MI->eraseFromParent(); // The pseudo instruction is gone now.
return BB;
// 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->getSubtarget().getInstrInfo());
+ const X86InstrInfo *TII = Subtarget->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()
- .getSubtargetImpl()
- ->getRegisterInfo()
- ->getCallPreservedMask(CallingConv::C);
+ const uint32_t *RegMask =
+ Subtarget->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->getSubtarget().getInstrInfo();
+ const TargetInstrInfo *TII = Subtarget->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->getSubtarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
MIB.addRegMask(RegInfo->getNoPreservedMask());
thisMBB->addSuccessor(mainMBB);
thisMBB->addSuccessor(restoreMBB);
// restoreMBB:
if (RegInfo->hasBasePointer(*MF)) {
- const X86Subtarget &STI = MF->getTarget().getSubtarget<X86Subtarget>();
- const bool Uses64BitFramePtr = STI.isTarget64BitLP64() || STI.isTargetNaCl64();
+ const bool Uses64BitFramePtr =
+ Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
X86FI->setRestoreBasePointer(MF);
unsigned FramePtr = RegInfo->getFrameRegister(*MF);
MachineBasicBlock *MBB) const {
DebugLoc DL = MI->getDebugLoc();
MachineFunction *MF = MBB->getParent();
- const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
+ const TargetInstrInfo *TII = Subtarget->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->getSubtarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
unsigned SP = RegInfo->getStackRegister();
default: llvm_unreachable("Unrecognized FMA variant.");
}
- const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
+ const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
MachineInstrBuilder MIB =
BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
.addOperand(MI->getOperand(0))
case X86::TAILJMPd64:
case X86::TAILJMPr64:
case X86::TAILJMPm64:
+ case X86::TAILJMPd64_REX:
+ case X86::TAILJMPr64_REX:
+ case X86::TAILJMPm64_REX:
llvm_unreachable("TAILJMP64 would not be touched here.");
case X86::TCRETURNdi64:
case X86::TCRETURNri64:
case X86::FP80_TO_INT32_IN_MEM:
case X86::FP80_TO_INT64_IN_MEM: {
MachineFunction *F = BB->getParent();
- const TargetInstrInfo *TII = F->getSubtarget().getInstrInfo();
+ const TargetInstrInfo *TII = Subtarget->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()->getSubtarget().getInstrInfo());
+ return EmitPCMPSTRM(MI, BB, Subtarget->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()->getSubtarget().getInstrInfo());
+ return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
// Thread synchronization.
case X86::MONITOR:
- return EmitMonitor(MI, BB, BB->getParent()->getSubtarget().getInstrInfo(),
- Subtarget);
+ return EmitMonitor(MI, BB, Subtarget);
// xbegin
case X86::XBEGIN:
- return EmitXBegin(MI, BB, BB->getParent()->getSubtarget().getInstrInfo());
+ return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
case X86::VASTART_SAVE_XMM_REGS:
return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
// Note that even with AVX we prefer the PSHUFD form of shuffle for integer
// vectors because it can have a load folded into it that UNPCK cannot. This
// doesn't preclude something switching to the shorter encoding post-RA.
- if (FloatDomain) {
+ //
+ // FIXME: Should teach these routines about AVX vector widths.
+ if (FloatDomain && VT.getSizeInBits() == 128) {
if (Mask.equals(0, 0) || Mask.equals(1, 1)) {
bool Lo = Mask.equals(0, 0);
unsigned Shuffle;
// We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
// variants as none of these have single-instruction variants that are
// superior to the UNPCK formulation.
- if (!FloatDomain &&
+ if (!FloatDomain && VT.getSizeInBits() == 128 &&
(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) ||
// 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 NumBytes = VT.getSizeInBits() / 8;
+ int Ratio = NumBytes / Mask.size();
+ for (int i = 0; i < NumBytes; ++i) {
if (Mask[i / Ratio] == SM_SentinelUndef) {
PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
continue;
: 255;
PSHUFBMask.push_back(DAG.getConstant(M, MVT::i8));
}
- Op = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Input);
+ MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
+ Op = DAG.getNode(ISD::BITCAST, DL, ByteVT, Input);
DCI.AddToWorklist(Op.getNode());
SDValue PSHUFBMaskOp =
- DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, PSHUFBMask);
+ DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
DCI.AddToWorklist(PSHUFBMaskOp.getNode());
- Op = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, Op, PSHUFBMaskOp);
+ Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
DCI.AddToWorklist(Op.getNode());
DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
/*AddTo*/ true);
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!");
/// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
/// PSHUF-style masks that can be reused with such instructions.
static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
+ MVT VT = N.getSimpleValueType();
SmallVector<int, 4> Mask;
bool IsUnary;
- bool HaveMask = getTargetShuffleMask(N.getNode(), N.getSimpleValueType(), Mask, IsUnary);
+ bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
(void)HaveMask;
assert(HaveMask);
+ // If we have more than 128-bits, only the low 128-bits of shuffle mask
+ // matter. Check that the upper masks are repeats and remove them.
+ if (VT.getSizeInBits() > 128) {
+ int LaneElts = 128 / VT.getScalarSizeInBits();
+#ifndef NDEBUG
+ for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
+ for (int j = 0; j < LaneElts; ++j)
+ assert(Mask[j] == Mask[i * LaneElts + j] - LaneElts &&
+ "Mask doesn't repeat in high 128-bit lanes!");
+#endif
+ Mask.resize(LaneElts);
+ }
+
switch (N.getOpcode()) {
case X86ISD::PSHUFD:
return Mask;
case X86ISD::UNPCKH:
// 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)
+ if (V.getSimpleValueType().getScalarType() != MVT::i8 &&
+ V.getSimpleValueType().getScalarType() != MVT::i16)
return SDValue();
// Search for a half-shuffle which we can combine with.
break;
case X86ISD::PSHUFLW:
case X86ISD::PSHUFHW:
- assert(VT == MVT::v8i16);
- (void)VT;
+ assert(VT.getScalarType() == MVT::i16 && "Bad word shuffle type!");
if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
return SDValue(); // We combined away this shuffle, so we're done.
int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
DMask[DOffset + 0] = DOffset + 1;
DMask[DOffset + 1] = DOffset + 0;
- V = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V);
+ MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
+ V = DAG.getNode(ISD::BITCAST, DL, DVT, V);
DCI.AddToWorklist(V.getNode());
- V = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V,
+ V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
getV4X86ShuffleImm8ForMask(DMask, DAG));
DCI.AddToWorklist(V.getNode());
- return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
+ return DAG.getNode(ISD::BITCAST, DL, VT, V);
}
// Look for shuffle patterns which can be implemented as a single unpack.
std::equal(std::begin(MappedMask), std::end(MappedMask),
std::begin(UnpackHiMask))) {
// We can replace all three shuffles with an unpack.
- V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, D.getOperand(0));
+ V = DAG.getNode(ISD::BITCAST, DL, VT, D.getOperand(0));
DCI.AddToWorklist(V.getNode());
return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
: X86ISD::UNPCKH,
- DL, MVT::v8i16, V, V);
+ DL, VT, V, V);
}
}
}
// We're looking for blends between FADD and FSUB nodes. We insist on these
// nodes being lined up in a specific expected pattern.
- if (!(isShuffleEquivalent(Mask, 0, 3) ||
- isShuffleEquivalent(Mask, 0, 5, 2, 7) ||
- isShuffleEquivalent(Mask, 0, 9, 2, 11, 4, 13, 6, 15)))
+ if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
+ isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
+ isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
return SDValue();
// Only specific types are legal at this point, assert so we notice if and
}
}
- // Only handle 128 wide vector from here on.
- if (!VT.is128BitVector())
- return SDValue();
-
// Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
// load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
// consecutive, non-overlapping, and in the right order.
: InVec.getOperand(1);
// If inputs to shuffle are the same for both ops, then allow 2 uses
- unsigned AllowedUses = InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
+ unsigned AllowedUses = InVec.getNumOperands() > 1 &&
+ InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
if (LdNode.getOpcode() == ISD::BITCAST) {
// Don't duplicate a load with other uses.
EltNo);
}
+/// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
+/// special and don't usually play with other vector types, it's better to
+/// handle them early to be sure we emit efficient code by avoiding
+/// store-load conversions.
+static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
+ if (N->getValueType(0) != MVT::x86mmx ||
+ N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
+ N->getOperand(0)->getValueType(0) != MVT::v2i32)
+ return SDValue();
+
+ SDValue V = N->getOperand(0);
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
+ if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
+ return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
+ N->getValueType(0), V.getOperand(0));
+
+ return SDValue();
+}
+
/// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
/// generation and convert it from being a bunch of shuffles and extracts
/// into a somewhat faster sequence. For i686, the best sequence is apparently
SDValue InputVector = N->getOperand(0);
- // Detect whether we are trying to convert from mmx to i32 and the bitcast
- // from mmx to v2i32 has a single usage.
- if (InputVector.getNode()->getOpcode() == llvm::ISD::BITCAST &&
- InputVector.getNode()->getOperand(0).getValueType() == MVT::x86mmx &&
- InputVector.hasOneUse() && N->getValueType(0) == MVT::i32)
- return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
- N->getValueType(0),
- InputVector.getNode()->getOperand(0));
+ // Detect mmx to i32 conversion through a v2i32 elt extract.
+ if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
+ N->getValueType(0) == MVT::i32 &&
+ InputVector.getValueType() == MVT::v2i32) {
+
+ // The bitcast source is a direct mmx result.
+ SDValue MMXSrc = InputVector.getNode()->getOperand(0);
+ if (MMXSrc.getValueType() == MVT::x86mmx)
+ return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
+ N->getValueType(0),
+ InputVector.getNode()->getOperand(0));
+
+ // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
+ SDValue MMXSrcOp = MMXSrc.getOperand(0);
+ if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
+ MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
+ MMXSrcOp.getOpcode() == ISD::BITCAST &&
+ MMXSrcOp.getValueType() == MVT::v1i64 &&
+ MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
+ return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
+ N->getValueType(0),
+ MMXSrcOp.getOperand(0));
+ }
// Only operate on vectors of 4 elements, where the alternative shuffling
// gets to be more expensive.
// instructions match the semantics of the common C idiom x<y?x:y but not
// x<=y?x:y, because of how they handle negative zero (which can be
// ignored in unsafe-math mode).
+ // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
- VT != MVT::f80 && TLI.isTypeLegal(VT) &&
+ VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
(Subtarget->hasSSE2() ||
(Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
}
}
- // If we know that this node is legal then we know that it is going to be
- // matched by one of the SSE/AVX BLEND instructions. These instructions only
- // depend on the highest bit in each word. Try to use SimplifyDemandedBits
- // to simplify previous instructions.
+ // We should generate an X86ISD::BLENDI from a vselect if its argument
+ // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
+ // constants. This specific pattern gets generated when we split a
+ // selector for a 512 bit vector in a machine without AVX512 (but with
+ // 256-bit vectors), during legalization:
+ //
+ // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
+ //
+ // Iff we find this pattern and the build_vectors are built from
+ // constants, we translate the vselect into a shuffle_vector that we
+ // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
+ if ((N->getOpcode() == ISD::VSELECT ||
+ N->getOpcode() == X86ISD::SHRUNKBLEND) &&
+ !DCI.isBeforeLegalize()) {
+ SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
+ if (Shuffle.getNode())
+ return Shuffle;
+ }
+
+ // If this is a *dynamic* select (non-constant condition) and we can match
+ // this node with one of the variable blend instructions, restructure the
+ // condition so that the blends can use the high bit of each element and use
+ // SimplifyDemandedBits to simplify the condition operand.
if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
!DCI.isBeforeLegalize() &&
- // We explicitly check against v8i16 and v16i16 because, although
- // they're marked as Custom, they might only be legal when Cond is a
- // build_vector of constants. This will be taken care in a later
- // condition.
- (TLI.isOperationLegalOrCustom(ISD::VSELECT, VT) && VT != MVT::v16i16 &&
- VT != MVT::v8i16) &&
- // Don't optimize vector of constants. Those are handled by
- // the generic code and all the bits must be properly set for
- // the generic optimizer.
!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
if (BitWidth == 1)
return SDValue();
+ // We can only handle the cases where VSELECT is directly legal on the
+ // subtarget. We custom lower VSELECT nodes with constant conditions and
+ // this makes it hard to see whether a dynamic VSELECT will correctly
+ // lower, so we both check the operation's status and explicitly handle the
+ // cases where a *dynamic* blend will fail even though a constant-condition
+ // blend could be custom lowered.
+ // FIXME: We should find a better way to handle this class of problems.
+ // Potentially, we should combine constant-condition vselect nodes
+ // pre-legalization into shuffles and not mark as many types as custom
+ // lowered.
+ if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
+ return SDValue();
+ // FIXME: We don't support i16-element blends currently. We could and
+ // should support them by making *all* the bits in the condition be set
+ // rather than just the high bit and using an i8-element blend.
+ if (VT.getScalarType() == MVT::i16)
+ return SDValue();
+ // Dynamic blending was only available from SSE4.1 onward.
+ if (VT.getSizeInBits() == 128 && !Subtarget->hasSSE41())
+ return SDValue();
+ // Byte blends are only available in AVX2
+ if (VT.getSizeInBits() == 256 && VT.getScalarType() == MVT::i8 &&
+ !Subtarget->hasAVX2())
+ return SDValue();
+
assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
}
}
- // We should generate an X86ISD::BLENDI from a vselect if its argument
- // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
- // constants. This specific pattern gets generated when we split a
- // selector for a 512 bit vector in a machine without AVX512 (but with
- // 256-bit vectors), during legalization:
- //
- // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
- //
- // Iff we find this pattern and the build_vectors are built from
- // constants, we translate the vselect into a shuffle_vector that we
- // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
- if ((N->getOpcode() == ISD::VSELECT ||
- N->getOpcode() == X86ISD::SHRUNKBLEND) &&
- !DCI.isBeforeLegalize()) {
- SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
- if (Shuffle.getNode())
- return Shuffle;
- }
-
return SDValue();
}
return SDValue();
EVT VT = N->getValueType(0);
- if (VT != MVT::i64)
+ if (VT != MVT::i64 && VT != MVT::i32)
return SDValue();
ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
}
}
+static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
+ TargetLowering::DAGCombinerInfo &DCI,
+ const X86Subtarget *Subtarget) {
+ SDValue N0 = N->getOperand(0);
+ SDValue N1 = N->getOperand(1);
+ SDLoc DL(N);
+
+ // A vector zext_in_reg may be represented as a shuffle,
+ // feeding into a bitcast (this represents anyext) feeding into
+ // an and with a mask.
+ // We'd like to try to combine that into a shuffle with zero
+ // plus a bitcast, removing the and.
+ if (N0.getOpcode() != ISD::BITCAST ||
+ N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
+ return SDValue();
+
+ // The other side of the AND should be a splat of 2^C, where C
+ // is the number of bits in the source type.
+ if (N1.getOpcode() == ISD::BITCAST)
+ N1 = N1.getOperand(0);
+ if (N1.getOpcode() != ISD::BUILD_VECTOR)
+ return SDValue();
+ BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
+
+ ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
+ EVT SrcType = Shuffle->getValueType(0);
+
+ // We expect a single-source shuffle
+ if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
+ return SDValue();
+
+ unsigned SrcSize = SrcType.getScalarSizeInBits();
+
+ APInt SplatValue, SplatUndef;
+ unsigned SplatBitSize;
+ bool HasAnyUndefs;
+ if (!Vector->isConstantSplat(SplatValue, SplatUndef,
+ SplatBitSize, HasAnyUndefs))
+ return SDValue();
+
+ unsigned ResSize = N1.getValueType().getScalarSizeInBits();
+ // Make sure the splat matches the mask we expect
+ if (SplatBitSize > ResSize ||
+ (SplatValue + 1).exactLogBase2() != (int)SrcSize)
+ return SDValue();
+
+ // Make sure the input and output size make sense
+ if (SrcSize >= ResSize || ResSize % SrcSize)
+ return SDValue();
+
+ // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
+ // The number of u's between each two values depends on the ratio between
+ // the source and dest type.
+ unsigned ZextRatio = ResSize / SrcSize;
+ bool IsZext = true;
+ for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
+ if (i % ZextRatio) {
+ if (Shuffle->getMaskElt(i) > 0) {
+ // Expected undef
+ IsZext = false;
+ break;
+ }
+ } else {
+ if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
+ // Expected element number
+ IsZext = false;
+ break;
+ }
+ }
+ }
+
+ if (!IsZext)
+ return SDValue();
+
+ // Ok, perform the transformation - replace the shuffle with
+ // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
+ // (instead of undef) where the k elements come from the zero vector.
+ SmallVector<int, 8> Mask;
+ unsigned NumElems = SrcType.getVectorNumElements();
+ for (unsigned i = 0; i < NumElems; ++i)
+ if (i % ZextRatio)
+ Mask.push_back(NumElems);
+ else
+ Mask.push_back(i / ZextRatio);
+
+ SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
+ Shuffle->getOperand(0), DAG.getConstant(0, SrcType), Mask);
+ return DAG.getNode(ISD::BITCAST, DL, N0.getValueType(), NewShuffle);
+}
+
static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const X86Subtarget *Subtarget) {
- EVT VT = N->getValueType(0);
if (DCI.isBeforeLegalizeOps())
return SDValue();
+ SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget);
+ if (Zext.getNode())
+ return Zext;
+
SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
if (R.getNode())
return R;
+ EVT VT = N->getValueType(0);
+ SDValue N0 = N->getOperand(0);
+ SDValue N1 = N->getOperand(1);
+ SDLoc DL(N);
+
// Create BEXTR instructions
// BEXTR is ((X >> imm) & (2**size-1))
if (VT == MVT::i32 || VT == MVT::i64) {
- SDValue N0 = N->getOperand(0);
- SDValue N1 = N->getOperand(1);
- SDLoc DL(N);
-
// Check for BEXTR.
if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
(N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
uint64_t Mask = MaskNode->getZExtValue();
uint64_t Shift = ShiftNode->getZExtValue();
if (isMask_64(Mask)) {
- uint64_t MaskSize = CountPopulation_64(Mask);
+ uint64_t MaskSize = countPopulation(Mask);
if (Shift + MaskSize <= VT.getSizeInBits())
return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
DAG.getConstant(Shift | (MaskSize << 8), VT));
if (VT != MVT::v2i64 && VT != MVT::v4i64)
return SDValue();
- SDValue N0 = N->getOperand(0);
- SDValue N1 = N->getOperand(1);
- SDLoc DL(N);
-
// Check LHS for vnot
if (N0.getOpcode() == ISD::XOR &&
//ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
// fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
MachineFunction &MF = DAG.getMachineFunction();
- bool OptForSize = MF.getFunction()->getAttributes().
- hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
+ bool OptForSize =
+ MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize);
// SHLD/SHRD instructions have lower register pressure, but on some
// platforms they have higher latency than the equivalent
return SDValue();
}
+/// PerformMLOADCombine - Resolve extending loads
+static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
+ TargetLowering::DAGCombinerInfo &DCI,
+ const X86Subtarget *Subtarget) {
+ MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
+ if (Mld->getExtensionType() != ISD::SEXTLOAD)
+ return SDValue();
+
+ EVT VT = Mld->getValueType(0);
+ unsigned NumElems = VT.getVectorNumElements();
+ EVT LdVT = Mld->getMemoryVT();
+ SDLoc dl(Mld);
+
+ assert(LdVT != VT && "Cannot extend to the same type");
+ unsigned ToSz = VT.getVectorElementType().getSizeInBits();
+ unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
+ // From, To sizes and ElemCount must be pow of two
+ assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
+ "Unexpected size for extending masked load");
+
+ unsigned SizeRatio = ToSz / FromSz;
+ assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
+
+ // Create a type on which we perform the shuffle
+ EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
+ LdVT.getScalarType(), NumElems*SizeRatio);
+ assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
+
+ // Convert Src0 value
+ SDValue WideSrc0 = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mld->getSrc0());
+ if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
+ SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
+ for (unsigned i = 0; i != NumElems; ++i)
+ ShuffleVec[i] = i * SizeRatio;
+
+ // Can't shuffle using an illegal type.
+ assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
+ && "WideVecVT should be legal");
+ WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
+ DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
+ }
+ // Prepare the new mask
+ SDValue NewMask;
+ SDValue Mask = Mld->getMask();
+ if (Mask.getValueType() == VT) {
+ // Mask and original value have the same type
+ NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
+ SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
+ for (unsigned i = 0; i != NumElems; ++i)
+ ShuffleVec[i] = i * SizeRatio;
+ for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
+ ShuffleVec[i] = NumElems*SizeRatio;
+ NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
+ DAG.getConstant(0, WideVecVT),
+ &ShuffleVec[0]);
+ }
+ else {
+ assert(Mask.getValueType().getVectorElementType() == MVT::i1);
+ unsigned WidenNumElts = NumElems*SizeRatio;
+ unsigned MaskNumElts = VT.getVectorNumElements();
+ EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
+ WidenNumElts);
+
+ unsigned NumConcat = WidenNumElts / MaskNumElts;
+ SmallVector<SDValue, 16> Ops(NumConcat);
+ SDValue ZeroVal = DAG.getConstant(0, Mask.getValueType());
+ Ops[0] = Mask;
+ for (unsigned i = 1; i != NumConcat; ++i)
+ Ops[i] = ZeroVal;
+
+ NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
+ }
+
+ SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
+ Mld->getBasePtr(), NewMask, WideSrc0,
+ Mld->getMemoryVT(), Mld->getMemOperand(),
+ ISD::NON_EXTLOAD);
+ SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
+ return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
+
+}
+/// PerformMSTORECombine - Resolve truncating stores
+static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
+ const X86Subtarget *Subtarget) {
+ MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
+ if (!Mst->isTruncatingStore())
+ return SDValue();
+
+ EVT VT = Mst->getValue().getValueType();
+ unsigned NumElems = VT.getVectorNumElements();
+ EVT StVT = Mst->getMemoryVT();
+ SDLoc dl(Mst);
+
+ assert(StVT != VT && "Cannot truncate to the same type");
+ unsigned FromSz = VT.getVectorElementType().getSizeInBits();
+ unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
+
+ // From, To sizes and ElemCount must be pow of two
+ assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
+ "Unexpected size for truncating masked store");
+ // We are going to use the original vector elt for storing.
+ // Accumulated smaller vector elements must be a multiple of the store size.
+ assert (((NumElems * FromSz) % ToSz) == 0 &&
+ "Unexpected ratio for truncating masked store");
+
+ unsigned SizeRatio = FromSz / ToSz;
+ assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
+
+ // Create a type on which we perform the shuffle
+ EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
+ StVT.getScalarType(), NumElems*SizeRatio);
+
+ assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
+
+ SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mst->getValue());
+ SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
+ for (unsigned i = 0; i != NumElems; ++i)
+ ShuffleVec[i] = i * SizeRatio;
+
+ // Can't shuffle using an illegal type.
+ assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
+ && "WideVecVT should be legal");
+
+ SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
+ DAG.getUNDEF(WideVecVT),
+ &ShuffleVec[0]);
+
+ SDValue NewMask;
+ SDValue Mask = Mst->getMask();
+ if (Mask.getValueType() == VT) {
+ // Mask and original value have the same type
+ NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
+ for (unsigned i = 0; i != NumElems; ++i)
+ ShuffleVec[i] = i * SizeRatio;
+ for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
+ ShuffleVec[i] = NumElems*SizeRatio;
+ NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
+ DAG.getConstant(0, WideVecVT),
+ &ShuffleVec[0]);
+ }
+ else {
+ assert(Mask.getValueType().getVectorElementType() == MVT::i1);
+ unsigned WidenNumElts = NumElems*SizeRatio;
+ unsigned MaskNumElts = VT.getVectorNumElements();
+ EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
+ WidenNumElts);
+
+ unsigned NumConcat = WidenNumElts / MaskNumElts;
+ SmallVector<SDValue, 16> Ops(NumConcat);
+ SDValue ZeroVal = DAG.getConstant(0, Mask.getValueType());
+ Ops[0] = Mask;
+ for (unsigned i = 1; i != NumConcat; ++i)
+ Ops[i] = ZeroVal;
+
+ NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
+ }
+
+ return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
+ NewMask, StVT, Mst->getMemOperand(), false);
+}
/// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
const X86Subtarget *Subtarget) {
return SDValue();
const Function *F = DAG.getMachineFunction().getFunction();
- bool NoImplicitFloatOps = F->getAttributes().
- hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
+ bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps
&& Subtarget->hasSSE2();
if ((VT.isVector() ||
/// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
+
// F[X]OR(0.0, x) -> x
- // F[X]OR(x, 0.0) -> x
if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
if (C->getValueAPF().isPosZero())
return N->getOperand(1);
+
+ // F[X]OR(x, 0.0) -> x
if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
if (C->getValueAPF().isPosZero())
return N->getOperand(0);
/// Do target-specific dag combines on X86ISD::FAND nodes.
static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
// FAND(0.0, x) -> 0.0
- // FAND(x, 0.0) -> 0.0
if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
if (C->getValueAPF().isPosZero())
return N->getOperand(0);
+
+ // FAND(x, 0.0) -> 0.0
if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
if (C->getValueAPF().isPosZero())
return N->getOperand(1);
+
return SDValue();
}
/// Do target-specific dag combines on X86ISD::FANDN nodes
static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
- // FANDN(x, 0.0) -> 0.0
// FANDN(0.0, x) -> x
if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
if (C->getValueAPF().isPosZero())
return N->getOperand(1);
+
+ // FANDN(x, 0.0) -> 0.0
if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
if (C->getValueAPF().isPosZero())
return N->getOperand(1);
+
return SDValue();
}
return SDValue();
}
+static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
+ SelectionDAG &DAG) {
+ SDLoc dl(Load);
+ MVT VT = Load->getSimpleValueType(0);
+ MVT EVT = VT.getVectorElementType();
+ SDValue Addr = Load->getOperand(1);
+ SDValue NewAddr = DAG.getNode(
+ ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
+ DAG.getConstant(Index * EVT.getStoreSize(), Addr.getSimpleValueType()));
+
+ SDValue NewLoad =
+ DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
+ DAG.getMachineFunction().getMachineMemOperand(
+ Load->getMemOperand(), 0, EVT.getStoreSize()));
+ return NewLoad;
+}
+
static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
const X86Subtarget *Subtarget) {
SDLoc dl(N);
if (MayFoldLoad(Ld)) {
// Extract the countS bits from the immediate so we can get the proper
// address when narrowing the vector load to a specific element.
- // When the second source op is a memory address, interps doesn't use
+ // When the second source op is a memory address, insertps doesn't use
// countS and just gets an f32 from that address.
unsigned DestIndex =
cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
+
Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
- } else
- return SDValue();
- // Create this as a scalar to vector to match the instruction pattern.
- SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
- // countS bits are ignored when loading from memory on insertps, which
- // means we don't need to explicitly set them to 0.
- return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
- LoadScalarToVector, N->getOperand(2));
+ // Create this as a scalar to vector to match the instruction pattern.
+ SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
+ // countS bits are ignored when loading from memory on insertps, which
+ // means we don't need to explicitly set them to 0.
+ return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
+ LoadScalarToVector, N->getOperand(2));
+ }
+ return SDValue();
+}
+
+static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
+ SDValue V0 = N->getOperand(0);
+ SDValue V1 = N->getOperand(1);
+ SDLoc DL(N);
+ EVT VT = N->getValueType(0);
+
+ // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
+ // operands and changing the mask to 1. This saves us a bunch of
+ // pattern-matching possibilities related to scalar math ops in SSE/AVX.
+ // x86InstrInfo knows how to commute this back after instruction selection
+ // if it would help register allocation.
+
+ // TODO: If optimizing for size or a processor that doesn't suffer from
+ // partial register update stalls, this should be transformed into a MOVSD
+ // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
+
+ if (VT == MVT::v2f64)
+ if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
+ if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
+ SDValue NewMask = DAG.getConstant(1, MVT::i8);
+ return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
+ }
+
+ return SDValue();
}
// Helper function of PerformSETCCCombine. It is to materialize "setb reg"
}
static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
- const X86TargetLowering *XTLI) {
+ const X86Subtarget *Subtarget) {
// First try to optimize away the conversion entirely when it's
// conditionally from a constant. Vectors only.
SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
EVT VT = Ld->getValueType(0);
if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
- !XTLI->getSubtarget()->is64Bit() &&
- VT == MVT::i64) {
- SDValue FILDChain = XTLI->BuildFILD(SDValue(N, 0), Ld->getValueType(0),
- Ld->getChain(), Op0, DAG);
+ !Subtarget->is64Bit() && VT == MVT::i64) {
+ SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
+ SDValue(N, 0), Ld->getValueType(0), Ld->getChain(), Op0, DAG);
DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
return FILDChain;
}
case ISD::SELECT:
case X86ISD::SHRUNKBLEND:
return PerformSELECTCombine(N, DAG, DCI, Subtarget);
+ case ISD::BITCAST: return PerformBITCASTCombine(N, DAG);
case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
+ case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
- case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
+ case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
+ case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
case X86ISD::FXOR:
case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
case ISD::INTRINSIC_WO_CHAIN:
return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget);
- case X86ISD::INSERTPS:
- return PerformINSERTPSCombine(N, DAG, Subtarget);
+ case X86ISD::INSERTPS: {
+ if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
+ return PerformINSERTPSCombine(N, DAG, Subtarget);
+ break;
+ }
+ case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DAG, Subtarget);
}
}
}
return;
+ case 'L':
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
+ if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
+ (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
+ Result = DAG.getTargetConstant(C->getSExtValue(), Op.getValueType());
+ break;
+ }
+ }
+ return;
+ case 'M':
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
+ if (C->getZExtValue() <= 3) {
+ Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
+ break;
+ }
+ }
+ return;
case 'N':
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
if (C->getZExtValue() <= 255) {
}
}
return;
+ case 'O':
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
+ if (C->getZExtValue() <= 127) {
+ Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
+ break;
+ }
+ }
+ return;
case 'e': {
// 32-bit signed value
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
}
-std::pair<unsigned, const TargetRegisterClass*>
-X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
+std::pair<unsigned, const TargetRegisterClass *>
+X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
+ const std::string &Constraint,
MVT VT) const {
// First, see if this is a constraint that directly corresponds to an LLVM
// register class.
// Use the default implementation in TargetLowering to convert the register
// constraint into a member of a register class.
std::pair<unsigned, const TargetRegisterClass*> Res;
- Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
+ Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
// Not found as a standard register?
if (!Res.second) {