blob: f86d9894df9b4c67cf01df32a09d9d2d45006942 [file] [log] [blame]
//===-- X86InstrAVX512.td - AVX512 Instruction Set ---------*- tablegen -*-===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file describes the X86 AVX512 instruction set, defining the
// instructions, and properties of the instructions which are needed for code
// generation, machine code emission, and analysis.
//
//===----------------------------------------------------------------------===//
// Group template arguments that can be derived from the vector type (EltNum x
// EltVT). These are things like the register class for the writemask, etc.
// The idea is to pass one of these as the template argument rather than the
// individual arguments.
// The template is also used for scalar types, in this case numelts is 1.
class X86VectorVTInfo<int numelts, ValueType eltvt, RegisterClass rc,
string suffix = ""> {
RegisterClass RC = rc;
ValueType EltVT = eltvt;
int NumElts = numelts;
// Corresponding mask register class.
RegisterClass KRC = !cast<RegisterClass>("VK" # NumElts);
// Corresponding write-mask register class.
RegisterClass KRCWM = !cast<RegisterClass>("VK" # NumElts # "WM");
// The mask VT.
ValueType KVT = !cast<ValueType>("v" # NumElts # "i1");
// Suffix used in the instruction mnemonic.
string Suffix = suffix;
// VTName is a string name for vector VT. For vector types it will be
// v # NumElts # EltVT, so for vector of 8 elements of i32 it will be v8i32
// It is a little bit complex for scalar types, where NumElts = 1.
// In this case we build v4f32 or v2f64
string VTName = "v" # !if (!eq (NumElts, 1),
!if (!eq (EltVT.Size, 32), 4,
!if (!eq (EltVT.Size, 64), 2, NumElts)), NumElts) # EltVT;
// The vector VT.
ValueType VT = !cast<ValueType>(VTName);
string EltTypeName = !cast<string>(EltVT);
// Size of the element type in bits, e.g. 32 for v16i32.
string EltSizeName = !subst("i", "", !subst("f", "", EltTypeName));
int EltSize = EltVT.Size;
// "i" for integer types and "f" for floating-point types
string TypeVariantName = !subst(EltSizeName, "", EltTypeName);
// Size of RC in bits, e.g. 512 for VR512.
int Size = VT.Size;
// The corresponding memory operand, e.g. i512mem for VR512.
X86MemOperand MemOp = !cast<X86MemOperand>(TypeVariantName # Size # "mem");
X86MemOperand ScalarMemOp = !cast<X86MemOperand>(EltVT # "mem");
// FP scalar memory operand for intrinsics - ssmem/sdmem.
Operand IntScalarMemOp = !if (!eq (EltTypeName, "f32"), !cast<Operand>("ssmem"),
!if (!eq (EltTypeName, "f64"), !cast<Operand>("sdmem"), ?));
// Load patterns
PatFrag LdFrag = !cast<PatFrag>("load" # VTName);
PatFrag AlignedLdFrag = !cast<PatFrag>("alignedload" # VTName);
PatFrag ScalarLdFrag = !cast<PatFrag>("load" # EltVT);
ComplexPattern ScalarIntMemCPat = !if (!eq (EltTypeName, "f32"),
!cast<ComplexPattern>("sse_load_f32"),
!if (!eq (EltTypeName, "f64"),
!cast<ComplexPattern>("sse_load_f64"),
?));
// The string to specify embedded broadcast in assembly.
string BroadcastStr = "{1to" # NumElts # "}";
// 8-bit compressed displacement tuple/subvector format. This is only
// defined for NumElts <= 8.
CD8VForm CD8TupleForm = !if (!eq (!srl(NumElts, 4), 0),
!cast<CD8VForm>("CD8VT" # NumElts), ?);
SubRegIndex SubRegIdx = !if (!eq (Size, 128), sub_xmm,
!if (!eq (Size, 256), sub_ymm, ?));
Domain ExeDomain = !if (!eq (EltTypeName, "f32"), SSEPackedSingle,
!if (!eq (EltTypeName, "f64"), SSEPackedDouble,
SSEPackedInt));
RegisterClass FRC = !if (!eq (EltTypeName, "f32"), FR32X, FR64X);
dag ImmAllZerosV = (VT immAllZerosV);
string ZSuffix = !if (!eq (Size, 128), "Z128",
!if (!eq (Size, 256), "Z256", "Z"));
}
def v64i8_info : X86VectorVTInfo<64, i8, VR512, "b">;
def v32i16_info : X86VectorVTInfo<32, i16, VR512, "w">;
def v16i32_info : X86VectorVTInfo<16, i32, VR512, "d">;
def v8i64_info : X86VectorVTInfo<8, i64, VR512, "q">;
def v16f32_info : X86VectorVTInfo<16, f32, VR512, "ps">;
def v8f64_info : X86VectorVTInfo<8, f64, VR512, "pd">;
// "x" in v32i8x_info means RC = VR256X
def v32i8x_info : X86VectorVTInfo<32, i8, VR256X, "b">;
def v16i16x_info : X86VectorVTInfo<16, i16, VR256X, "w">;
def v8i32x_info : X86VectorVTInfo<8, i32, VR256X, "d">;
def v4i64x_info : X86VectorVTInfo<4, i64, VR256X, "q">;
def v8f32x_info : X86VectorVTInfo<8, f32, VR256X, "ps">;
def v4f64x_info : X86VectorVTInfo<4, f64, VR256X, "pd">;
def v16i8x_info : X86VectorVTInfo<16, i8, VR128X, "b">;
def v8i16x_info : X86VectorVTInfo<8, i16, VR128X, "w">;
def v4i32x_info : X86VectorVTInfo<4, i32, VR128X, "d">;
def v2i64x_info : X86VectorVTInfo<2, i64, VR128X, "q">;
def v4f32x_info : X86VectorVTInfo<4, f32, VR128X, "ps">;
def v2f64x_info : X86VectorVTInfo<2, f64, VR128X, "pd">;
// We map scalar types to the smallest (128-bit) vector type
// with the appropriate element type. This allows to use the same masking logic.
def i32x_info : X86VectorVTInfo<1, i32, GR32, "si">;
def i64x_info : X86VectorVTInfo<1, i64, GR64, "sq">;
def f32x_info : X86VectorVTInfo<1, f32, VR128X, "ss">;
def f64x_info : X86VectorVTInfo<1, f64, VR128X, "sd">;
class AVX512VLVectorVTInfo<X86VectorVTInfo i512, X86VectorVTInfo i256,
X86VectorVTInfo i128> {
X86VectorVTInfo info512 = i512;
X86VectorVTInfo info256 = i256;
X86VectorVTInfo info128 = i128;
}
def avx512vl_i8_info : AVX512VLVectorVTInfo<v64i8_info, v32i8x_info,
v16i8x_info>;
def avx512vl_i16_info : AVX512VLVectorVTInfo<v32i16_info, v16i16x_info,
v8i16x_info>;
def avx512vl_i32_info : AVX512VLVectorVTInfo<v16i32_info, v8i32x_info,
v4i32x_info>;
def avx512vl_i64_info : AVX512VLVectorVTInfo<v8i64_info, v4i64x_info,
v2i64x_info>;
def avx512vl_f32_info : AVX512VLVectorVTInfo<v16f32_info, v8f32x_info,
v4f32x_info>;
def avx512vl_f64_info : AVX512VLVectorVTInfo<v8f64_info, v4f64x_info,
v2f64x_info>;
class X86KVectorVTInfo<RegisterClass _krc, RegisterClass _krcwm,
ValueType _vt> {
RegisterClass KRC = _krc;
RegisterClass KRCWM = _krcwm;
ValueType KVT = _vt;
}
def v1i1_info : X86KVectorVTInfo<VK1, VK1WM, v1i1>;
def v2i1_info : X86KVectorVTInfo<VK2, VK2WM, v2i1>;
def v4i1_info : X86KVectorVTInfo<VK4, VK4WM, v4i1>;
def v8i1_info : X86KVectorVTInfo<VK8, VK8WM, v8i1>;
def v16i1_info : X86KVectorVTInfo<VK16, VK16WM, v16i1>;
def v32i1_info : X86KVectorVTInfo<VK32, VK32WM, v32i1>;
def v64i1_info : X86KVectorVTInfo<VK64, VK64WM, v64i1>;
// This multiclass generates the masking variants from the non-masking
// variant. It only provides the assembly pieces for the masking variants.
// It assumes custom ISel patterns for masking which can be provided as
// template arguments.
multiclass AVX512_maskable_custom<bits<8> O, Format F,
dag Outs,
dag Ins, dag MaskingIns, dag ZeroMaskingIns,
string OpcodeStr,
string AttSrcAsm, string IntelSrcAsm,
list<dag> Pattern,
list<dag> MaskingPattern,
list<dag> ZeroMaskingPattern,
string MaskingConstraint = "",
bit IsCommutable = 0,
bit IsKCommutable = 0,
bit IsKZCommutable = IsCommutable> {
let isCommutable = IsCommutable in
def NAME: AVX512<O, F, Outs, Ins,
OpcodeStr#"\t{"#AttSrcAsm#", $dst|"#
"$dst, "#IntelSrcAsm#"}",
Pattern>;
// Prefer over VMOV*rrk Pat<>
let isCommutable = IsKCommutable in
def NAME#k: AVX512<O, F, Outs, MaskingIns,
OpcodeStr#"\t{"#AttSrcAsm#", $dst {${mask}}|"#
"$dst {${mask}}, "#IntelSrcAsm#"}",
MaskingPattern>,
EVEX_K {
// In case of the 3src subclass this is overridden with a let.
string Constraints = MaskingConstraint;
}
// Zero mask does not add any restrictions to commute operands transformation.
// So, it is Ok to use IsCommutable instead of IsKCommutable.
let isCommutable = IsKZCommutable in // Prefer over VMOV*rrkz Pat<>
def NAME#kz: AVX512<O, F, Outs, ZeroMaskingIns,
OpcodeStr#"\t{"#AttSrcAsm#", $dst {${mask}} {z}|"#
"$dst {${mask}} {z}, "#IntelSrcAsm#"}",
ZeroMaskingPattern>,
EVEX_KZ;
}
// Common base class of AVX512_maskable and AVX512_maskable_3src.
multiclass AVX512_maskable_common<bits<8> O, Format F, X86VectorVTInfo _,
dag Outs,
dag Ins, dag MaskingIns, dag ZeroMaskingIns,
string OpcodeStr,
string AttSrcAsm, string IntelSrcAsm,
dag RHS, dag MaskingRHS,
SDNode Select = vselect,
string MaskingConstraint = "",
bit IsCommutable = 0,
bit IsKCommutable = 0,
bit IsKZCommutable = IsCommutable> :
AVX512_maskable_custom<O, F, Outs, Ins, MaskingIns, ZeroMaskingIns, OpcodeStr,
AttSrcAsm, IntelSrcAsm,
[(set _.RC:$dst, RHS)],
[(set _.RC:$dst, MaskingRHS)],
[(set _.RC:$dst,
(Select _.KRCWM:$mask, RHS, _.ImmAllZerosV))],
MaskingConstraint, IsCommutable,
IsKCommutable, IsKZCommutable>;
// This multiclass generates the unconditional/non-masking, the masking and
// the zero-masking variant of the vector instruction. In the masking case, the
// perserved vector elements come from a new dummy input operand tied to $dst.
// This version uses a separate dag for non-masking and masking.
multiclass AVX512_maskable_split<bits<8> O, Format F, X86VectorVTInfo _,
dag Outs, dag Ins, string OpcodeStr,
string AttSrcAsm, string IntelSrcAsm,
dag RHS, dag MaskRHS,
bit IsCommutable = 0, bit IsKCommutable = 0,
SDNode Select = vselect> :
AVX512_maskable_custom<O, F, Outs, Ins,
!con((ins _.RC:$src0, _.KRCWM:$mask), Ins),
!con((ins _.KRCWM:$mask), Ins),
OpcodeStr, AttSrcAsm, IntelSrcAsm,
[(set _.RC:$dst, RHS)],
[(set _.RC:$dst,
(Select _.KRCWM:$mask, MaskRHS, _.RC:$src0))],
[(set _.RC:$dst,
(Select _.KRCWM:$mask, MaskRHS, _.ImmAllZerosV))],
"$src0 = $dst", IsCommutable, IsKCommutable>;
// This multiclass generates the unconditional/non-masking, the masking and
// the zero-masking variant of the vector instruction. In the masking case, the
// perserved vector elements come from a new dummy input operand tied to $dst.
multiclass AVX512_maskable<bits<8> O, Format F, X86VectorVTInfo _,
dag Outs, dag Ins, string OpcodeStr,
string AttSrcAsm, string IntelSrcAsm,
dag RHS,
bit IsCommutable = 0, bit IsKCommutable = 0,
bit IsKZCommutable = IsCommutable,
SDNode Select = vselect> :
AVX512_maskable_common<O, F, _, Outs, Ins,
!con((ins _.RC:$src0, _.KRCWM:$mask), Ins),
!con((ins _.KRCWM:$mask), Ins),
OpcodeStr, AttSrcAsm, IntelSrcAsm, RHS,
(Select _.KRCWM:$mask, RHS, _.RC:$src0),
Select, "$src0 = $dst", IsCommutable, IsKCommutable,
IsKZCommutable>;
// This multiclass generates the unconditional/non-masking, the masking and
// the zero-masking variant of the scalar instruction.
multiclass AVX512_maskable_scalar<bits<8> O, Format F, X86VectorVTInfo _,
dag Outs, dag Ins, string OpcodeStr,
string AttSrcAsm, string IntelSrcAsm,
dag RHS,
bit IsCommutable = 0> :
AVX512_maskable<O, F, _, Outs, Ins, OpcodeStr, AttSrcAsm, IntelSrcAsm,
RHS, IsCommutable, 0, IsCommutable, X86selects>;
// Similar to AVX512_maskable but in this case one of the source operands
// ($src1) is already tied to $dst so we just use that for the preserved
// vector elements. NOTE that the NonTiedIns (the ins dag) should exclude
// $src1.
multiclass AVX512_maskable_3src<bits<8> O, Format F, X86VectorVTInfo _,
dag Outs, dag NonTiedIns, string OpcodeStr,
string AttSrcAsm, string IntelSrcAsm,
dag RHS,
bit IsCommutable = 0,
bit IsKCommutable = 0,
SDNode Select = vselect,
bit MaskOnly = 0> :
AVX512_maskable_common<O, F, _, Outs,
!con((ins _.RC:$src1), NonTiedIns),
!con((ins _.RC:$src1, _.KRCWM:$mask), NonTiedIns),
!con((ins _.RC:$src1, _.KRCWM:$mask), NonTiedIns),
OpcodeStr, AttSrcAsm, IntelSrcAsm,
!if(MaskOnly, (null_frag), RHS),
(Select _.KRCWM:$mask, RHS, _.RC:$src1),
Select, "", IsCommutable, IsKCommutable>;
// Similar to AVX512_maskable_3src but in this case the input VT for the tied
// operand differs from the output VT. This requires a bitconvert on
// the preserved vector going into the vselect.
// NOTE: The unmasked pattern is disabled.
multiclass AVX512_maskable_3src_cast<bits<8> O, Format F, X86VectorVTInfo OutVT,
X86VectorVTInfo InVT,
dag Outs, dag NonTiedIns, string OpcodeStr,
string AttSrcAsm, string IntelSrcAsm,
dag RHS, bit IsCommutable = 0> :
AVX512_maskable_common<O, F, OutVT, Outs,
!con((ins InVT.RC:$src1), NonTiedIns),
!con((ins InVT.RC:$src1, InVT.KRCWM:$mask), NonTiedIns),
!con((ins InVT.RC:$src1, InVT.KRCWM:$mask), NonTiedIns),
OpcodeStr, AttSrcAsm, IntelSrcAsm, (null_frag),
(vselect InVT.KRCWM:$mask, RHS,
(bitconvert InVT.RC:$src1)),
vselect, "", IsCommutable>;
multiclass AVX512_maskable_3src_scalar<bits<8> O, Format F, X86VectorVTInfo _,
dag Outs, dag NonTiedIns, string OpcodeStr,
string AttSrcAsm, string IntelSrcAsm,
dag RHS,
bit IsCommutable = 0,
bit IsKCommutable = 0,
bit MaskOnly = 0> :
AVX512_maskable_3src<O, F, _, Outs, NonTiedIns, OpcodeStr, AttSrcAsm,
IntelSrcAsm, RHS, IsCommutable, IsKCommutable,
X86selects, MaskOnly>;
multiclass AVX512_maskable_in_asm<bits<8> O, Format F, X86VectorVTInfo _,
dag Outs, dag Ins,
string OpcodeStr,
string AttSrcAsm, string IntelSrcAsm,
list<dag> Pattern> :
AVX512_maskable_custom<O, F, Outs, Ins,
!con((ins _.RC:$src0, _.KRCWM:$mask), Ins),
!con((ins _.KRCWM:$mask), Ins),
OpcodeStr, AttSrcAsm, IntelSrcAsm, Pattern, [], [],
"$src0 = $dst">;
multiclass AVX512_maskable_3src_in_asm<bits<8> O, Format F, X86VectorVTInfo _,
dag Outs, dag NonTiedIns,
string OpcodeStr,
string AttSrcAsm, string IntelSrcAsm,
list<dag> Pattern> :
AVX512_maskable_custom<O, F, Outs,
!con((ins _.RC:$src1), NonTiedIns),
!con((ins _.RC:$src1, _.KRCWM:$mask), NonTiedIns),
!con((ins _.RC:$src1, _.KRCWM:$mask), NonTiedIns),
OpcodeStr, AttSrcAsm, IntelSrcAsm, Pattern, [], [],
"">;
// Instruction with mask that puts result in mask register,
// like "compare" and "vptest"
multiclass AVX512_maskable_custom_cmp<bits<8> O, Format F,
dag Outs,
dag Ins, dag MaskingIns,
string OpcodeStr,
string AttSrcAsm, string IntelSrcAsm,
list<dag> Pattern,
list<dag> MaskingPattern,
bit IsCommutable = 0> {
let isCommutable = IsCommutable in
def NAME: AVX512<O, F, Outs, Ins,
OpcodeStr#"\t{"#AttSrcAsm#", $dst|"#
"$dst, "#IntelSrcAsm#"}",
Pattern>;
def NAME#k: AVX512<O, F, Outs, MaskingIns,
OpcodeStr#"\t{"#AttSrcAsm#", $dst {${mask}}|"#
"$dst {${mask}}, "#IntelSrcAsm#"}",
MaskingPattern>, EVEX_K;
}
multiclass AVX512_maskable_common_cmp<bits<8> O, Format F, X86VectorVTInfo _,
dag Outs,
dag Ins, dag MaskingIns,
string OpcodeStr,
string AttSrcAsm, string IntelSrcAsm,
dag RHS, dag MaskingRHS,
bit IsCommutable = 0> :
AVX512_maskable_custom_cmp<O, F, Outs, Ins, MaskingIns, OpcodeStr,
AttSrcAsm, IntelSrcAsm,
[(set _.KRC:$dst, RHS)],
[(set _.KRC:$dst, MaskingRHS)], IsCommutable>;
multiclass AVX512_maskable_cmp<bits<8> O, Format F, X86VectorVTInfo _,
dag Outs, dag Ins, string OpcodeStr,
string AttSrcAsm, string IntelSrcAsm,
dag RHS, dag RHS_su, bit IsCommutable = 0> :
AVX512_maskable_common_cmp<O, F, _, Outs, Ins,
!con((ins _.KRCWM:$mask), Ins),
OpcodeStr, AttSrcAsm, IntelSrcAsm, RHS,
(and _.KRCWM:$mask, RHS_su), IsCommutable>;
// Alias instruction that maps zero vector to pxor / xorp* for AVX-512.
// This is expanded by ExpandPostRAPseudos to an xorps / vxorps, and then
// swizzled by ExecutionDomainFix to pxor.
// We set canFoldAsLoad because this can be converted to a constant-pool
// load of an all-zeros value if folding it would be beneficial.
let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
isPseudo = 1, Predicates = [HasAVX512], SchedRW = [WriteZero] in {
def AVX512_512_SET0 : I<0, Pseudo, (outs VR512:$dst), (ins), "",
[(set VR512:$dst, (v16i32 immAllZerosV))]>;
def AVX512_512_SETALLONES : I<0, Pseudo, (outs VR512:$dst), (ins), "",
[(set VR512:$dst, (v16i32 immAllOnesV))]>;
}
// Alias instructions that allow VPTERNLOG to be used with a mask to create
// a mix of all ones and all zeros elements. This is done this way to force
// the same register to be used as input for all three sources.
let isPseudo = 1, Predicates = [HasAVX512], SchedRW = [WriteVecALU] in {
def AVX512_512_SEXT_MASK_32 : I<0, Pseudo, (outs VR512:$dst),
(ins VK16WM:$mask), "",
[(set VR512:$dst, (vselect (v16i1 VK16WM:$mask),
(v16i32 immAllOnesV),
(v16i32 immAllZerosV)))]>;
def AVX512_512_SEXT_MASK_64 : I<0, Pseudo, (outs VR512:$dst),
(ins VK8WM:$mask), "",
[(set VR512:$dst, (vselect (v8i1 VK8WM:$mask),
(v8i64 immAllOnesV),
(v8i64 immAllZerosV)))]>;
}
let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
isPseudo = 1, Predicates = [HasAVX512], SchedRW = [WriteZero] in {
def AVX512_128_SET0 : I<0, Pseudo, (outs VR128X:$dst), (ins), "",
[(set VR128X:$dst, (v4i32 immAllZerosV))]>;
def AVX512_256_SET0 : I<0, Pseudo, (outs VR256X:$dst), (ins), "",
[(set VR256X:$dst, (v8i32 immAllZerosV))]>;
}
// Alias instructions that map fld0 to xorps for sse or vxorps for avx.
// This is expanded by ExpandPostRAPseudos.
let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
isPseudo = 1, SchedRW = [WriteZero], Predicates = [HasAVX512] in {
def AVX512_FsFLD0SS : I<0, Pseudo, (outs FR32X:$dst), (ins), "",
[(set FR32X:$dst, fp32imm0)]>;
def AVX512_FsFLD0SD : I<0, Pseudo, (outs FR64X:$dst), (ins), "",
[(set FR64X:$dst, fpimm0)]>;
}
//===----------------------------------------------------------------------===//
// AVX-512 - VECTOR INSERT
//
// Supports two different pattern operators for mask and unmasked ops. Allows
// null_frag to be passed for one.
multiclass vinsert_for_size_split<int Opcode, X86VectorVTInfo From,
X86VectorVTInfo To,
SDPatternOperator vinsert_insert,
SDPatternOperator vinsert_for_mask,
X86FoldableSchedWrite sched> {
let hasSideEffects = 0, ExeDomain = To.ExeDomain in {
defm rr : AVX512_maskable_split<Opcode, MRMSrcReg, To, (outs To.RC:$dst),
(ins To.RC:$src1, From.RC:$src2, u8imm:$src3),
"vinsert" # From.EltTypeName # "x" # From.NumElts,
"$src3, $src2, $src1", "$src1, $src2, $src3",
(vinsert_insert:$src3 (To.VT To.RC:$src1),
(From.VT From.RC:$src2),
(iPTR imm)),
(vinsert_for_mask:$src3 (To.VT To.RC:$src1),
(From.VT From.RC:$src2),
(iPTR imm))>,
AVX512AIi8Base, EVEX_4V, Sched<[sched]>;
let mayLoad = 1 in
defm rm : AVX512_maskable_split<Opcode, MRMSrcMem, To, (outs To.RC:$dst),
(ins To.RC:$src1, From.MemOp:$src2, u8imm:$src3),
"vinsert" # From.EltTypeName # "x" # From.NumElts,
"$src3, $src2, $src1", "$src1, $src2, $src3",
(vinsert_insert:$src3 (To.VT To.RC:$src1),
(From.VT (From.LdFrag addr:$src2)),
(iPTR imm)),
(vinsert_for_mask:$src3 (To.VT To.RC:$src1),
(From.VT (From.LdFrag addr:$src2)),
(iPTR imm))>, AVX512AIi8Base, EVEX_4V,
EVEX_CD8<From.EltSize, From.CD8TupleForm>,
Sched<[sched.Folded, sched.ReadAfterFold]>;
}
}
// Passes the same pattern operator for masked and unmasked ops.
multiclass vinsert_for_size<int Opcode, X86VectorVTInfo From,
X86VectorVTInfo To,
SDPatternOperator vinsert_insert,
X86FoldableSchedWrite sched> :
vinsert_for_size_split<Opcode, From, To, vinsert_insert, vinsert_insert, sched>;
multiclass vinsert_for_size_lowering<string InstrStr, X86VectorVTInfo From,
X86VectorVTInfo To, PatFrag vinsert_insert,
SDNodeXForm INSERT_get_vinsert_imm , list<Predicate> p> {
let Predicates = p in {
def : Pat<(vinsert_insert:$ins
(To.VT To.RC:$src1), (From.VT From.RC:$src2), (iPTR imm)),
(To.VT (!cast<Instruction>(InstrStr#"rr")
To.RC:$src1, From.RC:$src2,
(INSERT_get_vinsert_imm To.RC:$ins)))>;
def : Pat<(vinsert_insert:$ins
(To.VT To.RC:$src1),
(From.VT (From.LdFrag addr:$src2)),
(iPTR imm)),
(To.VT (!cast<Instruction>(InstrStr#"rm")
To.RC:$src1, addr:$src2,
(INSERT_get_vinsert_imm To.RC:$ins)))>;
}
}
multiclass vinsert_for_type<ValueType EltVT32, int Opcode128,
ValueType EltVT64, int Opcode256,
X86FoldableSchedWrite sched> {
let Predicates = [HasVLX] in
defm NAME # "32x4Z256" : vinsert_for_size<Opcode128,
X86VectorVTInfo< 4, EltVT32, VR128X>,
X86VectorVTInfo< 8, EltVT32, VR256X>,
vinsert128_insert, sched>, EVEX_V256;
defm NAME # "32x4Z" : vinsert_for_size<Opcode128,
X86VectorVTInfo< 4, EltVT32, VR128X>,
X86VectorVTInfo<16, EltVT32, VR512>,
vinsert128_insert, sched>, EVEX_V512;
defm NAME # "64x4Z" : vinsert_for_size<Opcode256,
X86VectorVTInfo< 4, EltVT64, VR256X>,
X86VectorVTInfo< 8, EltVT64, VR512>,
vinsert256_insert, sched>, VEX_W, EVEX_V512;
// Even with DQI we'd like to only use these instructions for masking.
let Predicates = [HasVLX, HasDQI] in
defm NAME # "64x2Z256" : vinsert_for_size_split<Opcode128,
X86VectorVTInfo< 2, EltVT64, VR128X>,
X86VectorVTInfo< 4, EltVT64, VR256X>,
null_frag, vinsert128_insert, sched>,
VEX_W1X, EVEX_V256;
// Even with DQI we'd like to only use these instructions for masking.
let Predicates = [HasDQI] in {
defm NAME # "64x2Z" : vinsert_for_size_split<Opcode128,
X86VectorVTInfo< 2, EltVT64, VR128X>,
X86VectorVTInfo< 8, EltVT64, VR512>,
null_frag, vinsert128_insert, sched>,
VEX_W, EVEX_V512;
defm NAME # "32x8Z" : vinsert_for_size_split<Opcode256,
X86VectorVTInfo< 8, EltVT32, VR256X>,
X86VectorVTInfo<16, EltVT32, VR512>,
null_frag, vinsert256_insert, sched>,
EVEX_V512;
}
}
// FIXME: Is there a better scheduler class for VINSERTF/VINSERTI?
defm VINSERTF : vinsert_for_type<f32, 0x18, f64, 0x1a, WriteFShuffle256>;
defm VINSERTI : vinsert_for_type<i32, 0x38, i64, 0x3a, WriteShuffle256>;
// Codegen pattern with the alternative types,
// Even with AVX512DQ we'll still use these for unmasked operations.
defm : vinsert_for_size_lowering<"VINSERTF32x4Z256", v2f64x_info, v4f64x_info,
vinsert128_insert, INSERT_get_vinsert128_imm, [HasVLX]>;
defm : vinsert_for_size_lowering<"VINSERTI32x4Z256", v2i64x_info, v4i64x_info,
vinsert128_insert, INSERT_get_vinsert128_imm, [HasVLX]>;
defm : vinsert_for_size_lowering<"VINSERTF32x4Z", v2f64x_info, v8f64_info,
vinsert128_insert, INSERT_get_vinsert128_imm, [HasAVX512]>;
defm : vinsert_for_size_lowering<"VINSERTI32x4Z", v2i64x_info, v8i64_info,
vinsert128_insert, INSERT_get_vinsert128_imm, [HasAVX512]>;
defm : vinsert_for_size_lowering<"VINSERTF64x4Z", v8f32x_info, v16f32_info,
vinsert256_insert, INSERT_get_vinsert256_imm, [HasAVX512]>;
defm : vinsert_for_size_lowering<"VINSERTI64x4Z", v8i32x_info, v16i32_info,
vinsert256_insert, INSERT_get_vinsert256_imm, [HasAVX512]>;
// Codegen pattern with the alternative types insert VEC128 into VEC256
defm : vinsert_for_size_lowering<"VINSERTI32x4Z256", v8i16x_info, v16i16x_info,
vinsert128_insert, INSERT_get_vinsert128_imm, [HasVLX]>;
defm : vinsert_for_size_lowering<"VINSERTI32x4Z256", v16i8x_info, v32i8x_info,
vinsert128_insert, INSERT_get_vinsert128_imm, [HasVLX]>;
// Codegen pattern with the alternative types insert VEC128 into VEC512
defm : vinsert_for_size_lowering<"VINSERTI32x4Z", v8i16x_info, v32i16_info,
vinsert128_insert, INSERT_get_vinsert128_imm, [HasAVX512]>;
defm : vinsert_for_size_lowering<"VINSERTI32x4Z", v16i8x_info, v64i8_info,
vinsert128_insert, INSERT_get_vinsert128_imm, [HasAVX512]>;
// Codegen pattern with the alternative types insert VEC256 into VEC512
defm : vinsert_for_size_lowering<"VINSERTI64x4Z", v16i16x_info, v32i16_info,
vinsert256_insert, INSERT_get_vinsert256_imm, [HasAVX512]>;
defm : vinsert_for_size_lowering<"VINSERTI64x4Z", v32i8x_info, v64i8_info,
vinsert256_insert, INSERT_get_vinsert256_imm, [HasAVX512]>;
multiclass vinsert_for_mask_cast<string InstrStr, X86VectorVTInfo From,
X86VectorVTInfo To, X86VectorVTInfo Cast,
PatFrag vinsert_insert,
SDNodeXForm INSERT_get_vinsert_imm,
list<Predicate> p> {
let Predicates = p in {
def : Pat<(Cast.VT
(vselect Cast.KRCWM:$mask,
(bitconvert
(vinsert_insert:$ins (To.VT To.RC:$src1),
(From.VT From.RC:$src2),
(iPTR imm))),
Cast.RC:$src0)),
(!cast<Instruction>(InstrStr#"rrk")
Cast.RC:$src0, Cast.KRCWM:$mask, To.RC:$src1, From.RC:$src2,
(INSERT_get_vinsert_imm To.RC:$ins))>;
def : Pat<(Cast.VT
(vselect Cast.KRCWM:$mask,
(bitconvert
(vinsert_insert:$ins (To.VT To.RC:$src1),
(From.VT
(bitconvert
(From.LdFrag addr:$src2))),
(iPTR imm))),
Cast.RC:$src0)),
(!cast<Instruction>(InstrStr#"rmk")
Cast.RC:$src0, Cast.KRCWM:$mask, To.RC:$src1, addr:$src2,
(INSERT_get_vinsert_imm To.RC:$ins))>;
def : Pat<(Cast.VT
(vselect Cast.KRCWM:$mask,
(bitconvert
(vinsert_insert:$ins (To.VT To.RC:$src1),
(From.VT From.RC:$src2),
(iPTR imm))),
Cast.ImmAllZerosV)),
(!cast<Instruction>(InstrStr#"rrkz")
Cast.KRCWM:$mask, To.RC:$src1, From.RC:$src2,
(INSERT_get_vinsert_imm To.RC:$ins))>;
def : Pat<(Cast.VT
(vselect Cast.KRCWM:$mask,
(bitconvert
(vinsert_insert:$ins (To.VT To.RC:$src1),
(From.VT (From.LdFrag addr:$src2)),
(iPTR imm))),
Cast.ImmAllZerosV)),
(!cast<Instruction>(InstrStr#"rmkz")
Cast.KRCWM:$mask, To.RC:$src1, addr:$src2,
(INSERT_get_vinsert_imm To.RC:$ins))>;
}
}
defm : vinsert_for_mask_cast<"VINSERTF32x4Z256", v2f64x_info, v4f64x_info,
v8f32x_info, vinsert128_insert,
INSERT_get_vinsert128_imm, [HasVLX]>;
defm : vinsert_for_mask_cast<"VINSERTF64x2Z256", v4f32x_info, v8f32x_info,
v4f64x_info, vinsert128_insert,
INSERT_get_vinsert128_imm, [HasDQI, HasVLX]>;
defm : vinsert_for_mask_cast<"VINSERTI32x4Z256", v2i64x_info, v4i64x_info,
v8i32x_info, vinsert128_insert,
INSERT_get_vinsert128_imm, [HasVLX]>;
defm : vinsert_for_mask_cast<"VINSERTI32x4Z256", v8i16x_info, v16i16x_info,
v8i32x_info, vinsert128_insert,
INSERT_get_vinsert128_imm, [HasVLX]>;
defm : vinsert_for_mask_cast<"VINSERTI32x4Z256", v16i8x_info, v32i8x_info,
v8i32x_info, vinsert128_insert,
INSERT_get_vinsert128_imm, [HasVLX]>;
defm : vinsert_for_mask_cast<"VINSERTF64x2Z256", v4i32x_info, v8i32x_info,
v4i64x_info, vinsert128_insert,
INSERT_get_vinsert128_imm, [HasDQI, HasVLX]>;
defm : vinsert_for_mask_cast<"VINSERTF64x2Z256", v8i16x_info, v16i16x_info,
v4i64x_info, vinsert128_insert,
INSERT_get_vinsert128_imm, [HasDQI, HasVLX]>;
defm : vinsert_for_mask_cast<"VINSERTF64x2Z256", v16i8x_info, v32i8x_info,
v4i64x_info, vinsert128_insert,
INSERT_get_vinsert128_imm, [HasDQI, HasVLX]>;
defm : vinsert_for_mask_cast<"VINSERTF32x4Z", v2f64x_info, v8f64_info,
v16f32_info, vinsert128_insert,
INSERT_get_vinsert128_imm, [HasAVX512]>;
defm : vinsert_for_mask_cast<"VINSERTF64x2Z", v4f32x_info, v16f32_info,
v8f64_info, vinsert128_insert,
INSERT_get_vinsert128_imm, [HasDQI]>;
defm : vinsert_for_mask_cast<"VINSERTI32x4Z", v2i64x_info, v8i64_info,
v16i32_info, vinsert128_insert,
INSERT_get_vinsert128_imm, [HasAVX512]>;
defm : vinsert_for_mask_cast<"VINSERTI32x4Z", v8i16x_info, v32i16_info,
v16i32_info, vinsert128_insert,
INSERT_get_vinsert128_imm, [HasAVX512]>;
defm : vinsert_for_mask_cast<"VINSERTI32x4Z", v16i8x_info, v64i8_info,
v16i32_info, vinsert128_insert,
INSERT_get_vinsert128_imm, [HasAVX512]>;
defm : vinsert_for_mask_cast<"VINSERTI64x2Z", v4i32x_info, v16i32_info,
v8i64_info, vinsert128_insert,
INSERT_get_vinsert128_imm, [HasDQI]>;
defm : vinsert_for_mask_cast<"VINSERTI64x2Z", v8i16x_info, v32i16_info,
v8i64_info, vinsert128_insert,
INSERT_get_vinsert128_imm, [HasDQI]>;
defm : vinsert_for_mask_cast<"VINSERTI64x2Z", v16i8x_info, v64i8_info,
v8i64_info, vinsert128_insert,
INSERT_get_vinsert128_imm, [HasDQI]>;
defm : vinsert_for_mask_cast<"VINSERTF32x8Z", v4f64x_info, v8f64_info,
v16f32_info, vinsert256_insert,
INSERT_get_vinsert256_imm, [HasDQI]>;
defm : vinsert_for_mask_cast<"VINSERTF64x4Z", v8f32x_info, v16f32_info,
v8f64_info, vinsert256_insert,
INSERT_get_vinsert256_imm, [HasAVX512]>;
defm : vinsert_for_mask_cast<"VINSERTI32x8Z", v4i64x_info, v8i64_info,
v16i32_info, vinsert256_insert,
INSERT_get_vinsert256_imm, [HasDQI]>;
defm : vinsert_for_mask_cast<"VINSERTI32x8Z", v16i16x_info, v32i16_info,
v16i32_info, vinsert256_insert,
INSERT_get_vinsert256_imm, [HasDQI]>;
defm : vinsert_for_mask_cast<"VINSERTI32x8Z", v32i8x_info, v64i8_info,
v16i32_info, vinsert256_insert,
INSERT_get_vinsert256_imm, [HasDQI]>;
defm : vinsert_for_mask_cast<"VINSERTI64x4Z", v8i32x_info, v16i32_info,
v8i64_info, vinsert256_insert,
INSERT_get_vinsert256_imm, [HasAVX512]>;
defm : vinsert_for_mask_cast<"VINSERTI64x4Z", v16i16x_info, v32i16_info,
v8i64_info, vinsert256_insert,
INSERT_get_vinsert256_imm, [HasAVX512]>;
defm : vinsert_for_mask_cast<"VINSERTI64x4Z", v32i8x_info, v64i8_info,
v8i64_info, vinsert256_insert,
INSERT_get_vinsert256_imm, [HasAVX512]>;
// vinsertps - insert f32 to XMM
let ExeDomain = SSEPackedSingle in {
let isCommutable = 1 in
def VINSERTPSZrr : AVX512AIi8<0x21, MRMSrcReg, (outs VR128X:$dst),
(ins VR128X:$src1, VR128X:$src2, u8imm:$src3),
"vinsertps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
[(set VR128X:$dst, (X86insertps VR128X:$src1, VR128X:$src2, imm:$src3))]>,
EVEX_4V, Sched<[SchedWriteFShuffle.XMM]>;
def VINSERTPSZrm: AVX512AIi8<0x21, MRMSrcMem, (outs VR128X:$dst),
(ins VR128X:$src1, f32mem:$src2, u8imm:$src3),
"vinsertps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
[(set VR128X:$dst, (X86insertps VR128X:$src1,
(v4f32 (scalar_to_vector (loadf32 addr:$src2))),
imm:$src3))]>,
EVEX_4V, EVEX_CD8<32, CD8VT1>,
Sched<[SchedWriteFShuffle.XMM.Folded, SchedWriteFShuffle.XMM.ReadAfterFold]>;
}
//===----------------------------------------------------------------------===//
// AVX-512 VECTOR EXTRACT
//---
// Supports two different pattern operators for mask and unmasked ops. Allows
// null_frag to be passed for one.
multiclass vextract_for_size_split<int Opcode,
X86VectorVTInfo From, X86VectorVTInfo To,
SDPatternOperator vextract_extract,
SDPatternOperator vextract_for_mask,
SchedWrite SchedRR, SchedWrite SchedMR> {
let hasSideEffects = 0, ExeDomain = To.ExeDomain in {
defm rr : AVX512_maskable_split<Opcode, MRMDestReg, To, (outs To.RC:$dst),
(ins From.RC:$src1, u8imm:$idx),
"vextract" # To.EltTypeName # "x" # To.NumElts,
"$idx, $src1", "$src1, $idx",
(vextract_extract:$idx (From.VT From.RC:$src1), (iPTR imm)),
(vextract_for_mask:$idx (From.VT From.RC:$src1), (iPTR imm))>,
AVX512AIi8Base, EVEX, Sched<[SchedRR]>;
def mr : AVX512AIi8<Opcode, MRMDestMem, (outs),
(ins To.MemOp:$dst, From.RC:$src1, u8imm:$idx),
"vextract" # To.EltTypeName # "x" # To.NumElts #
"\t{$idx, $src1, $dst|$dst, $src1, $idx}",
[(store (To.VT (vextract_extract:$idx
(From.VT From.RC:$src1), (iPTR imm))),
addr:$dst)]>, EVEX,
Sched<[SchedMR]>;
let mayStore = 1, hasSideEffects = 0 in
def mrk : AVX512AIi8<Opcode, MRMDestMem, (outs),
(ins To.MemOp:$dst, To.KRCWM:$mask,
From.RC:$src1, u8imm:$idx),
"vextract" # To.EltTypeName # "x" # To.NumElts #
"\t{$idx, $src1, $dst {${mask}}|"
"$dst {${mask}}, $src1, $idx}", []>,
EVEX_K, EVEX, Sched<[SchedMR]>, NotMemoryFoldable;
}
}
// Passes the same pattern operator for masked and unmasked ops.
multiclass vextract_for_size<int Opcode, X86VectorVTInfo From,
X86VectorVTInfo To,
SDPatternOperator vextract_extract,
SchedWrite SchedRR, SchedWrite SchedMR> :
vextract_for_size_split<Opcode, From, To, vextract_extract, vextract_extract, SchedRR, SchedMR>;
// Codegen pattern for the alternative types
multiclass vextract_for_size_lowering<string InstrStr, X86VectorVTInfo From,
X86VectorVTInfo To, PatFrag vextract_extract,
SDNodeXForm EXTRACT_get_vextract_imm, list<Predicate> p> {
let Predicates = p in {
def : Pat<(vextract_extract:$ext (From.VT From.RC:$src1), (iPTR imm)),
(To.VT (!cast<Instruction>(InstrStr#"rr")
From.RC:$src1,
(EXTRACT_get_vextract_imm To.RC:$ext)))>;
def : Pat<(store (To.VT (vextract_extract:$ext (From.VT From.RC:$src1),
(iPTR imm))), addr:$dst),
(!cast<Instruction>(InstrStr#"mr") addr:$dst, From.RC:$src1,
(EXTRACT_get_vextract_imm To.RC:$ext))>;
}
}
multiclass vextract_for_type<ValueType EltVT32, int Opcode128,
ValueType EltVT64, int Opcode256,
SchedWrite SchedRR, SchedWrite SchedMR> {
let Predicates = [HasAVX512] in {
defm NAME # "32x4Z" : vextract_for_size<Opcode128,
X86VectorVTInfo<16, EltVT32, VR512>,
X86VectorVTInfo< 4, EltVT32, VR128X>,
vextract128_extract, SchedRR, SchedMR>,
EVEX_V512, EVEX_CD8<32, CD8VT4>;
defm NAME # "64x4Z" : vextract_for_size<Opcode256,
X86VectorVTInfo< 8, EltVT64, VR512>,
X86VectorVTInfo< 4, EltVT64, VR256X>,
vextract256_extract, SchedRR, SchedMR>,
VEX_W, EVEX_V512, EVEX_CD8<64, CD8VT4>;
}
let Predicates = [HasVLX] in
defm NAME # "32x4Z256" : vextract_for_size<Opcode128,
X86VectorVTInfo< 8, EltVT32, VR256X>,
X86VectorVTInfo< 4, EltVT32, VR128X>,
vextract128_extract, SchedRR, SchedMR>,
EVEX_V256, EVEX_CD8<32, CD8VT4>;
// Even with DQI we'd like to only use these instructions for masking.
let Predicates = [HasVLX, HasDQI] in
defm NAME # "64x2Z256" : vextract_for_size_split<Opcode128,
X86VectorVTInfo< 4, EltVT64, VR256X>,
X86VectorVTInfo< 2, EltVT64, VR128X>,
null_frag, vextract128_extract, SchedRR, SchedMR>,
VEX_W1X, EVEX_V256, EVEX_CD8<64, CD8VT2>;
// Even with DQI we'd like to only use these instructions for masking.
let Predicates = [HasDQI] in {
defm NAME # "64x2Z" : vextract_for_size_split<Opcode128,
X86VectorVTInfo< 8, EltVT64, VR512>,
X86VectorVTInfo< 2, EltVT64, VR128X>,
null_frag, vextract128_extract, SchedRR, SchedMR>,
VEX_W, EVEX_V512, EVEX_CD8<64, CD8VT2>;
defm NAME # "32x8Z" : vextract_for_size_split<Opcode256,
X86VectorVTInfo<16, EltVT32, VR512>,
X86VectorVTInfo< 8, EltVT32, VR256X>,
null_frag, vextract256_extract, SchedRR, SchedMR>,
EVEX_V512, EVEX_CD8<32, CD8VT8>;
}
}
// TODO - replace WriteFStore/WriteVecStore with X86SchedWriteMoveLSWidths types.
defm VEXTRACTF : vextract_for_type<f32, 0x19, f64, 0x1b, WriteFShuffle256, WriteFStore>;
defm VEXTRACTI : vextract_for_type<i32, 0x39, i64, 0x3b, WriteShuffle256, WriteVecStore>;
// extract_subvector codegen patterns with the alternative types.
// Even with AVX512DQ we'll still use these for unmasked operations.
defm : vextract_for_size_lowering<"VEXTRACTF32x4Z", v8f64_info, v2f64x_info,
vextract128_extract, EXTRACT_get_vextract128_imm, [HasAVX512]>;
defm : vextract_for_size_lowering<"VEXTRACTI32x4Z", v8i64_info, v2i64x_info,
vextract128_extract, EXTRACT_get_vextract128_imm, [HasAVX512]>;
defm : vextract_for_size_lowering<"VEXTRACTF64x4Z", v16f32_info, v8f32x_info,
vextract256_extract, EXTRACT_get_vextract256_imm, [HasAVX512]>;
defm : vextract_for_size_lowering<"VEXTRACTI64x4Z", v16i32_info, v8i32x_info,
vextract256_extract, EXTRACT_get_vextract256_imm, [HasAVX512]>;
defm : vextract_for_size_lowering<"VEXTRACTF32x4Z256", v4f64x_info, v2f64x_info,
vextract128_extract, EXTRACT_get_vextract128_imm, [HasVLX]>;
defm : vextract_for_size_lowering<"VEXTRACTI32x4Z256", v4i64x_info, v2i64x_info,
vextract128_extract, EXTRACT_get_vextract128_imm, [HasVLX]>;
// Codegen pattern with the alternative types extract VEC128 from VEC256
defm : vextract_for_size_lowering<"VEXTRACTI32x4Z256", v16i16x_info, v8i16x_info,
vextract128_extract, EXTRACT_get_vextract128_imm, [HasVLX]>;
defm : vextract_for_size_lowering<"VEXTRACTI32x4Z256", v32i8x_info, v16i8x_info,
vextract128_extract, EXTRACT_get_vextract128_imm, [HasVLX]>;
// Codegen pattern with the alternative types extract VEC128 from VEC512
defm : vextract_for_size_lowering<"VEXTRACTI32x4Z", v32i16_info, v8i16x_info,
vextract128_extract, EXTRACT_get_vextract128_imm, [HasAVX512]>;
defm : vextract_for_size_lowering<"VEXTRACTI32x4Z", v64i8_info, v16i8x_info,
vextract128_extract, EXTRACT_get_vextract128_imm, [HasAVX512]>;
// Codegen pattern with the alternative types extract VEC256 from VEC512
defm : vextract_for_size_lowering<"VEXTRACTI64x4Z", v32i16_info, v16i16x_info,
vextract256_extract, EXTRACT_get_vextract256_imm, [HasAVX512]>;
defm : vextract_for_size_lowering<"VEXTRACTI64x4Z", v64i8_info, v32i8x_info,
vextract256_extract, EXTRACT_get_vextract256_imm, [HasAVX512]>;
// A 128-bit extract from bits [255:128] of a 512-bit vector should use a
// smaller extract to enable EVEX->VEX.
let Predicates = [NoVLX] in {
def : Pat<(v2i64 (extract_subvector (v8i64 VR512:$src), (iPTR 2))),
(v2i64 (VEXTRACTI128rr
(v4i64 (EXTRACT_SUBREG (v8i64 VR512:$src), sub_ymm)),
(iPTR 1)))>;
def : Pat<(v2f64 (extract_subvector (v8f64 VR512:$src), (iPTR 2))),
(v2f64 (VEXTRACTF128rr
(v4f64 (EXTRACT_SUBREG (v8f64 VR512:$src), sub_ymm)),
(iPTR 1)))>;
def : Pat<(v4i32 (extract_subvector (v16i32 VR512:$src), (iPTR 4))),
(v4i32 (VEXTRACTI128rr
(v8i32 (EXTRACT_SUBREG (v16i32 VR512:$src), sub_ymm)),
(iPTR 1)))>;
def : Pat<(v4f32 (extract_subvector (v16f32 VR512:$src), (iPTR 4))),
(v4f32 (VEXTRACTF128rr
(v8f32 (EXTRACT_SUBREG (v16f32 VR512:$src), sub_ymm)),
(iPTR 1)))>;
def : Pat<(v8i16 (extract_subvector (v32i16 VR512:$src), (iPTR 8))),
(v8i16 (VEXTRACTI128rr
(v16i16 (EXTRACT_SUBREG (v32i16 VR512:$src), sub_ymm)),
(iPTR 1)))>;
def : Pat<(v16i8 (extract_subvector (v64i8 VR512:$src), (iPTR 16))),
(v16i8 (VEXTRACTI128rr
(v32i8 (EXTRACT_SUBREG (v64i8 VR512:$src), sub_ymm)),
(iPTR 1)))>;
}
// A 128-bit extract from bits [255:128] of a 512-bit vector should use a
// smaller extract to enable EVEX->VEX.
let Predicates = [HasVLX] in {
def : Pat<(v2i64 (extract_subvector (v8i64 VR512:$src), (iPTR 2))),
(v2i64 (VEXTRACTI32x4Z256rr
(v4i64 (EXTRACT_SUBREG (v8i64 VR512:$src), sub_ymm)),
(iPTR 1)))>;
def : Pat<(v2f64 (extract_subvector (v8f64 VR512:$src), (iPTR 2))),
(v2f64 (VEXTRACTF32x4Z256rr
(v4f64 (EXTRACT_SUBREG (v8f64 VR512:$src), sub_ymm)),
(iPTR 1)))>;
def : Pat<(v4i32 (extract_subvector (v16i32 VR512:$src), (iPTR 4))),
(v4i32 (VEXTRACTI32x4Z256rr
(v8i32 (EXTRACT_SUBREG (v16i32 VR512:$src), sub_ymm)),
(iPTR 1)))>;
def : Pat<(v4f32 (extract_subvector (v16f32 VR512:$src), (iPTR 4))),
(v4f32 (VEXTRACTF32x4Z256rr
(v8f32 (EXTRACT_SUBREG (v16f32 VR512:$src), sub_ymm)),
(iPTR 1)))>;
def : Pat<(v8i16 (extract_subvector (v32i16 VR512:$src), (iPTR 8))),
(v8i16 (VEXTRACTI32x4Z256rr
(v16i16 (EXTRACT_SUBREG (v32i16 VR512:$src), sub_ymm)),
(iPTR 1)))>;
def : Pat<(v16i8 (extract_subvector (v64i8 VR512:$src), (iPTR 16))),
(v16i8 (VEXTRACTI32x4Z256rr
(v32i8 (EXTRACT_SUBREG (v64i8 VR512:$src), sub_ymm)),
(iPTR 1)))>;
}
// Additional patterns for handling a bitcast between the vselect and the
// extract_subvector.
multiclass vextract_for_mask_cast<string InstrStr, X86VectorVTInfo From,
X86VectorVTInfo To, X86VectorVTInfo Cast,
PatFrag vextract_extract,
SDNodeXForm EXTRACT_get_vextract_imm,
list<Predicate> p> {
let Predicates = p in {
def : Pat<(Cast.VT (vselect Cast.KRCWM:$mask,
(bitconvert
(To.VT (vextract_extract:$ext
(From.VT From.RC:$src), (iPTR imm)))),
To.RC:$src0)),
(Cast.VT (!cast<Instruction>(InstrStr#"rrk")
Cast.RC:$src0, Cast.KRCWM:$mask, From.RC:$src,
(EXTRACT_get_vextract_imm To.RC:$ext)))>;
def : Pat<(Cast.VT (vselect Cast.KRCWM:$mask,
(bitconvert
(To.VT (vextract_extract:$ext
(From.VT From.RC:$src), (iPTR imm)))),
Cast.ImmAllZerosV)),
(Cast.VT (!cast<Instruction>(InstrStr#"rrkz")
Cast.KRCWM:$mask, From.RC:$src,
(EXTRACT_get_vextract_imm To.RC:$ext)))>;
}
}
defm : vextract_for_mask_cast<"VEXTRACTF32x4Z256", v4f64x_info, v2f64x_info,
v4f32x_info, vextract128_extract,
EXTRACT_get_vextract128_imm, [HasVLX]>;
defm : vextract_for_mask_cast<"VEXTRACTF64x2Z256", v8f32x_info, v4f32x_info,
v2f64x_info, vextract128_extract,
EXTRACT_get_vextract128_imm, [HasDQI, HasVLX]>;
defm : vextract_for_mask_cast<"VEXTRACTI32x4Z256", v4i64x_info, v2i64x_info,
v4i32x_info, vextract128_extract,
EXTRACT_get_vextract128_imm, [HasVLX]>;
defm : vextract_for_mask_cast<"VEXTRACTI32x4Z256", v16i16x_info, v8i16x_info,
v4i32x_info, vextract128_extract,
EXTRACT_get_vextract128_imm, [HasVLX]>;
defm : vextract_for_mask_cast<"VEXTRACTI32x4Z256", v32i8x_info, v16i8x_info,
v4i32x_info, vextract128_extract,
EXTRACT_get_vextract128_imm, [HasVLX]>;
defm : vextract_for_mask_cast<"VEXTRACTI64x2Z256", v8i32x_info, v4i32x_info,
v2i64x_info, vextract128_extract,
EXTRACT_get_vextract128_imm, [HasDQI, HasVLX]>;
defm : vextract_for_mask_cast<"VEXTRACTI64x2Z256", v16i16x_info, v8i16x_info,
v2i64x_info, vextract128_extract,
EXTRACT_get_vextract128_imm, [HasDQI, HasVLX]>;
defm : vextract_for_mask_cast<"VEXTRACTI64x2Z256", v32i8x_info, v16i8x_info,
v2i64x_info, vextract128_extract,
EXTRACT_get_vextract128_imm, [HasDQI, HasVLX]>;
defm : vextract_for_mask_cast<"VEXTRACTF32x4Z", v8f64_info, v2f64x_info,
v4f32x_info, vextract128_extract,
EXTRACT_get_vextract128_imm, [HasAVX512]>;
defm : vextract_for_mask_cast<"VEXTRACTF64x2Z", v16f32_info, v4f32x_info,
v2f64x_info, vextract128_extract,
EXTRACT_get_vextract128_imm, [HasDQI]>;
defm : vextract_for_mask_cast<"VEXTRACTI32x4Z", v8i64_info, v2i64x_info,
v4i32x_info, vextract128_extract,
EXTRACT_get_vextract128_imm, [HasAVX512]>;
defm : vextract_for_mask_cast<"VEXTRACTI32x4Z", v32i16_info, v8i16x_info,
v4i32x_info, vextract128_extract,
EXTRACT_get_vextract128_imm, [HasAVX512]>;
defm : vextract_for_mask_cast<"VEXTRACTI32x4Z", v64i8_info, v16i8x_info,
v4i32x_info, vextract128_extract,
EXTRACT_get_vextract128_imm, [HasAVX512]>;
defm : vextract_for_mask_cast<"VEXTRACTI64x2Z", v16i32_info, v4i32x_info,
v2i64x_info, vextract128_extract,
EXTRACT_get_vextract128_imm, [HasDQI]>;
defm : vextract_for_mask_cast<"VEXTRACTI64x2Z", v32i16_info, v8i16x_info,
v2i64x_info, vextract128_extract,
EXTRACT_get_vextract128_imm, [HasDQI]>;
defm : vextract_for_mask_cast<"VEXTRACTI64x2Z", v64i8_info, v16i8x_info,
v2i64x_info, vextract128_extract,
EXTRACT_get_vextract128_imm, [HasDQI]>;
defm : vextract_for_mask_cast<"VEXTRACTF32x8Z", v8f64_info, v4f64x_info,
v8f32x_info, vextract256_extract,
EXTRACT_get_vextract256_imm, [HasDQI]>;
defm : vextract_for_mask_cast<"VEXTRACTF64x4Z", v16f32_info, v8f32x_info,
v4f64x_info, vextract256_extract,
EXTRACT_get_vextract256_imm, [HasAVX512]>;
defm : vextract_for_mask_cast<"VEXTRACTI32x8Z", v8i64_info, v4i64x_info,
v8i32x_info, vextract256_extract,
EXTRACT_get_vextract256_imm, [HasDQI]>;
defm : vextract_for_mask_cast<"VEXTRACTI32x8Z", v32i16_info, v16i16x_info,
v8i32x_info, vextract256_extract,
EXTRACT_get_vextract256_imm, [HasDQI]>;
defm : vextract_for_mask_cast<"VEXTRACTI32x8Z", v64i8_info, v32i8x_info,
v8i32x_info, vextract256_extract,
EXTRACT_get_vextract256_imm, [HasDQI]>;
defm : vextract_for_mask_cast<"VEXTRACTI64x4Z", v16i32_info, v8i32x_info,
v4i64x_info, vextract256_extract,
EXTRACT_get_vextract256_imm, [HasAVX512]>;
defm : vextract_for_mask_cast<"VEXTRACTI64x4Z", v32i16_info, v16i16x_info,
v4i64x_info, vextract256_extract,
EXTRACT_get_vextract256_imm, [HasAVX512]>;
defm : vextract_for_mask_cast<"VEXTRACTI64x4Z", v64i8_info, v32i8x_info,
v4i64x_info, vextract256_extract,
EXTRACT_get_vextract256_imm, [HasAVX512]>;
// vextractps - extract 32 bits from XMM
def VEXTRACTPSZrr : AVX512AIi8<0x17, MRMDestReg, (outs GR32:$dst),
(ins VR128X:$src1, u8imm:$src2),
"vextractps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
[(set GR32:$dst, (extractelt (bc_v4i32 (v4f32 VR128X:$src1)), imm:$src2))]>,
EVEX, VEX_WIG, Sched<[WriteVecExtract]>;
def VEXTRACTPSZmr : AVX512AIi8<0x17, MRMDestMem, (outs),
(ins f32mem:$dst, VR128X:$src1, u8imm:$src2),
"vextractps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
[(store (extractelt (bc_v4i32 (v4f32 VR128X:$src1)), imm:$src2),
addr:$dst)]>,
EVEX, VEX_WIG, EVEX_CD8<32, CD8VT1>, Sched<[WriteVecExtractSt]>;
//===---------------------------------------------------------------------===//
// AVX-512 BROADCAST
//---
// broadcast with a scalar argument.
multiclass avx512_broadcast_scalar<bits<8> opc, string OpcodeStr,
string Name,
X86VectorVTInfo DestInfo, X86VectorVTInfo SrcInfo> {
def : Pat<(DestInfo.VT (X86VBroadcast SrcInfo.FRC:$src)),
(!cast<Instruction>(Name#DestInfo.ZSuffix#r)
(SrcInfo.VT (COPY_TO_REGCLASS SrcInfo.FRC:$src, SrcInfo.RC)))>;
def : Pat<(DestInfo.VT (vselect DestInfo.KRCWM:$mask,
(X86VBroadcast SrcInfo.FRC:$src),
DestInfo.RC:$src0)),
(!cast<Instruction>(Name#DestInfo.ZSuffix#rk)
DestInfo.RC:$src0, DestInfo.KRCWM:$mask,
(SrcInfo.VT (COPY_TO_REGCLASS SrcInfo.FRC:$src, SrcInfo.RC)))>;
def : Pat<(DestInfo.VT (vselect DestInfo.KRCWM:$mask,
(X86VBroadcast SrcInfo.FRC:$src),
DestInfo.ImmAllZerosV)),
(!cast<Instruction>(Name#DestInfo.ZSuffix#rkz)
DestInfo.KRCWM:$mask, (SrcInfo.VT (COPY_TO_REGCLASS SrcInfo.FRC:$src, SrcInfo.RC)))>;
}
// Split version to allow mask and broadcast node to be different types. This
// helps support the 32x2 broadcasts.
multiclass avx512_broadcast_rm_split<bits<8> opc, string OpcodeStr,
string Name,
SchedWrite SchedRR, SchedWrite SchedRM,
X86VectorVTInfo MaskInfo,
X86VectorVTInfo DestInfo,
X86VectorVTInfo SrcInfo,
SDPatternOperator UnmaskedOp = X86VBroadcast> {
let ExeDomain = DestInfo.ExeDomain, hasSideEffects = 0 in {
defm r : AVX512_maskable_split<opc, MRMSrcReg, MaskInfo,
(outs MaskInfo.RC:$dst),
(ins SrcInfo.RC:$src), OpcodeStr, "$src", "$src",
(MaskInfo.VT
(bitconvert
(DestInfo.VT
(UnmaskedOp (SrcInfo.VT SrcInfo.RC:$src))))),
(MaskInfo.VT
(bitconvert
(DestInfo.VT
(X86VBroadcast (SrcInfo.VT SrcInfo.RC:$src)))))>,
T8PD, EVEX, Sched<[SchedRR]>;
let mayLoad = 1 in
defm m : AVX512_maskable_split<opc, MRMSrcMem, MaskInfo,
(outs MaskInfo.RC:$dst),
(ins SrcInfo.ScalarMemOp:$src), OpcodeStr, "$src", "$src",
(MaskInfo.VT
(bitconvert
(DestInfo.VT (UnmaskedOp
(SrcInfo.ScalarLdFrag addr:$src))))),
(MaskInfo.VT
(bitconvert
(DestInfo.VT (X86VBroadcast
(SrcInfo.ScalarLdFrag addr:$src)))))>,
T8PD, EVEX, EVEX_CD8<SrcInfo.EltSize, CD8VT1>,
Sched<[SchedRM]>;
}
def : Pat<(MaskInfo.VT
(bitconvert
(DestInfo.VT (UnmaskedOp
(SrcInfo.VT (scalar_to_vector
(SrcInfo.ScalarLdFrag addr:$src))))))),
(!cast<Instruction>(Name#MaskInfo.ZSuffix#m) addr:$src)>;
def : Pat<(MaskInfo.VT (vselect MaskInfo.KRCWM:$mask,
(bitconvert
(DestInfo.VT
(X86VBroadcast
(SrcInfo.VT (scalar_to_vector
(SrcInfo.ScalarLdFrag addr:$src)))))),
MaskInfo.RC:$src0)),
(!cast<Instruction>(Name#DestInfo.ZSuffix#mk)
MaskInfo.RC:$src0, MaskInfo.KRCWM:$mask, addr:$src)>;
def : Pat<(MaskInfo.VT (vselect MaskInfo.KRCWM:$mask,
(bitconvert
(DestInfo.VT
(X86VBroadcast
(SrcInfo.VT (scalar_to_vector
(SrcInfo.ScalarLdFrag addr:$src)))))),
MaskInfo.ImmAllZerosV)),
(!cast<Instruction>(Name#MaskInfo.ZSuffix#mkz)
MaskInfo.KRCWM:$mask, addr:$src)>;
}
// Helper class to force mask and broadcast result to same type.
multiclass avx512_broadcast_rm<bits<8> opc, string OpcodeStr, string Name,
SchedWrite SchedRR, SchedWrite SchedRM,
X86VectorVTInfo DestInfo,
X86VectorVTInfo SrcInfo> :
avx512_broadcast_rm_split<opc, OpcodeStr, Name, SchedRR, SchedRM,
DestInfo, DestInfo, SrcInfo>;
multiclass avx512_fp_broadcast_sd<bits<8> opc, string OpcodeStr,
AVX512VLVectorVTInfo _> {
let Predicates = [HasAVX512] in {
defm Z : avx512_broadcast_rm<opc, OpcodeStr, NAME, WriteFShuffle256,
WriteFShuffle256Ld, _.info512, _.info128>,
avx512_broadcast_scalar<opc, OpcodeStr, NAME, _.info512,
_.info128>,
EVEX_V512;
}
let Predicates = [HasVLX] in {
defm Z256 : avx512_broadcast_rm<opc, OpcodeStr, NAME, WriteFShuffle256,
WriteFShuffle256Ld, _.info256, _.info128>,
avx512_broadcast_scalar<opc, OpcodeStr, NAME, _.info256,
_.info128>,
EVEX_V256;
}
}
multiclass avx512_fp_broadcast_ss<bits<8> opc, string OpcodeStr,
AVX512VLVectorVTInfo _> {
let Predicates = [HasAVX512] in {
defm Z : avx512_broadcast_rm<opc, OpcodeStr, NAME, WriteFShuffle256,
WriteFShuffle256Ld, _.info512, _.info128>,
avx512_broadcast_scalar<opc, OpcodeStr, NAME, _.info512,
_.info128>,
EVEX_V512;
}
let Predicates = [HasVLX] in {
defm Z256 : avx512_broadcast_rm<opc, OpcodeStr, NAME, WriteFShuffle256,
WriteFShuffle256Ld, _.info256, _.info128>,
avx512_broadcast_scalar<opc, OpcodeStr, NAME, _.info256,
_.info128>,
EVEX_V256;
defm Z128 : avx512_broadcast_rm<opc, OpcodeStr, NAME, WriteFShuffle256,
WriteFShuffle256Ld, _.info128, _.info128>,
avx512_broadcast_scalar<opc, OpcodeStr, NAME, _.info128,
_.info128>,
EVEX_V128;
}
}
defm VBROADCASTSS : avx512_fp_broadcast_ss<0x18, "vbroadcastss",
avx512vl_f32_info>;
defm VBROADCASTSD : avx512_fp_broadcast_sd<0x19, "vbroadcastsd",
avx512vl_f64_info>, VEX_W1X;
multiclass avx512_int_broadcast_reg<bits<8> opc, SchedWrite SchedRR,
X86VectorVTInfo _, SDPatternOperator OpNode,
RegisterClass SrcRC> {
let ExeDomain = _.ExeDomain in
defm r : AVX512_maskable<opc, MRMSrcReg, _, (outs _.RC:$dst),
(ins SrcRC:$src),
"vpbroadcast"##_.Suffix, "$src", "$src",
(_.VT (OpNode SrcRC:$src))>, T8PD, EVEX,
Sched<[SchedRR]>;
}
multiclass avx512_int_broadcastbw_reg<bits<8> opc, string Name, SchedWrite SchedRR,
X86VectorVTInfo _, SDPatternOperator OpNode,
RegisterClass SrcRC, SubRegIndex Subreg> {
let hasSideEffects = 0, ExeDomain = _.ExeDomain in
defm r : AVX512_maskable_custom<opc, MRMSrcReg,
(outs _.RC:$dst), (ins GR32:$src),
!con((ins _.RC:$src0, _.KRCWM:$mask), (ins GR32:$src)),
!con((ins _.KRCWM:$mask), (ins GR32:$src)),
"vpbroadcast"##_.Suffix, "$src", "$src", [], [], [],
"$src0 = $dst">, T8PD, EVEX, Sched<[SchedRR]>;
def : Pat <(_.VT (OpNode SrcRC:$src)),
(!cast<Instruction>(Name#r)
(i32 (INSERT_SUBREG (i32 (IMPLICIT_DEF)), SrcRC:$src, Subreg)))>;
def : Pat <(vselect _.KRCWM:$mask, (_.VT (OpNode SrcRC:$src)), _.RC:$src0),
(!cast<Instruction>(Name#rk) _.RC:$src0, _.KRCWM:$mask,
(i32 (INSERT_SUBREG (i32 (IMPLICIT_DEF)), SrcRC:$src, Subreg)))>;
def : Pat <(vselect _.KRCWM:$mask, (_.VT (OpNode SrcRC:$src)), _.ImmAllZerosV),
(!cast<Instruction>(Name#rkz) _.KRCWM:$mask,
(i32 (INSERT_SUBREG (i32 (IMPLICIT_DEF)), SrcRC:$src, Subreg)))>;
}
multiclass avx512_int_broadcastbw_reg_vl<bits<8> opc, string Name,
AVX512VLVectorVTInfo _, SDPatternOperator OpNode,
RegisterClass SrcRC, SubRegIndex Subreg, Predicate prd> {
let Predicates = [prd] in
defm Z : avx512_int_broadcastbw_reg<opc, Name#Z, WriteShuffle256, _.info512,
OpNode, SrcRC, Subreg>, EVEX_V512;
let Predicates = [prd, HasVLX] in {
defm Z256 : avx512_int_broadcastbw_reg<opc, Name#Z256, WriteShuffle256,
_.info256, OpNode, SrcRC, Subreg>, EVEX_V256;
defm Z128 : avx512_int_broadcastbw_reg<opc, Name#Z128, WriteShuffle,
_.info128, OpNode, SrcRC, Subreg>, EVEX_V128;
}
}
multiclass avx512_int_broadcast_reg_vl<bits<8> opc, AVX512VLVectorVTInfo _,
SDPatternOperator OpNode,
RegisterClass SrcRC, Predicate prd> {
let Predicates = [prd] in
defm Z : avx512_int_broadcast_reg<opc, WriteShuffle256, _.info512, OpNode,
SrcRC>, EVEX_V512;
let Predicates = [prd, HasVLX] in {
defm Z256 : avx512_int_broadcast_reg<opc, WriteShuffle256, _.info256, OpNode,
SrcRC>, EVEX_V256;
defm Z128 : avx512_int_broadcast_reg<opc, WriteShuffle, _.info128, OpNode,
SrcRC>, EVEX_V128;
}
}
defm VPBROADCASTBr : avx512_int_broadcastbw_reg_vl<0x7A, "VPBROADCASTBr",
avx512vl_i8_info, X86VBroadcast, GR8, sub_8bit, HasBWI>;
defm VPBROADCASTWr : avx512_int_broadcastbw_reg_vl<0x7B, "VPBROADCASTWr",
avx512vl_i16_info, X86VBroadcast, GR16, sub_16bit,
HasBWI>;
defm VPBROADCASTDr : avx512_int_broadcast_reg_vl<0x7C, avx512vl_i32_info,
X86VBroadcast, GR32, HasAVX512>;
defm VPBROADCASTQr : avx512_int_broadcast_reg_vl<0x7C, avx512vl_i64_info,
X86VBroadcast, GR64, HasAVX512>, VEX_W;
// Provide aliases for broadcast from the same register class that
// automatically does the extract.
multiclass avx512_int_broadcast_rm_lowering<string Name,
X86VectorVTInfo DestInfo,
X86VectorVTInfo SrcInfo,
X86VectorVTInfo ExtInfo> {
def : Pat<(DestInfo.VT (X86VBroadcast (SrcInfo.VT SrcInfo.RC:$src))),
(!cast<Instruction>(Name#DestInfo.ZSuffix#"r")
(ExtInfo.VT (EXTRACT_SUBREG (SrcInfo.VT SrcInfo.RC:$src), sub_xmm)))>;
}
multiclass avx512_int_broadcast_rm_vl<bits<8> opc, string OpcodeStr,
AVX512VLVectorVTInfo _, Predicate prd> {
let Predicates = [prd] in {
defm Z : avx512_broadcast_rm<opc, OpcodeStr, NAME, WriteShuffle256,
WriteShuffle256Ld, _.info512, _.info128>,
avx512_int_broadcast_rm_lowering<NAME, _.info512, _.info256, _.info128>,
EVEX_V512;
// Defined separately to avoid redefinition.
defm Z_Alt : avx512_int_broadcast_rm_lowering<NAME, _.info512, _.info512, _.info128>;
}
let Predicates = [prd, HasVLX] in {
defm Z256 : avx512_broadcast_rm<opc, OpcodeStr, NAME, WriteShuffle256,
WriteShuffle256Ld, _.info256, _.info128>,
avx512_int_broadcast_rm_lowering<NAME, _.info256, _.info256, _.info128>,
EVEX_V256;
defm Z128 : avx512_broadcast_rm<opc, OpcodeStr, NAME, WriteShuffle,
WriteShuffleXLd, _.info128, _.info128>,
EVEX_V128;
}
}
defm VPBROADCASTB : avx512_int_broadcast_rm_vl<0x78, "vpbroadcastb",
avx512vl_i8_info, HasBWI>;
defm VPBROADCASTW : avx512_int_broadcast_rm_vl<0x79, "vpbroadcastw",
avx512vl_i16_info, HasBWI>;
defm VPBROADCASTD : avx512_int_broadcast_rm_vl<0x58, "vpbroadcastd",
avx512vl_i32_info, HasAVX512>;
defm VPBROADCASTQ : avx512_int_broadcast_rm_vl<0x59, "vpbroadcastq",
avx512vl_i64_info, HasAVX512>, VEX_W1X;
multiclass avx512_subvec_broadcast_rm<bits<8> opc, string OpcodeStr,
X86VectorVTInfo _Dst, X86VectorVTInfo _Src> {
defm rm : AVX512_maskable<opc, MRMSrcMem, _Dst, (outs _Dst.RC:$dst),
(ins _Src.MemOp:$src), OpcodeStr, "$src", "$src",
(_Dst.VT (X86SubVBroadcast
(_Src.VT (_Src.LdFrag addr:$src))))>,
Sched<[SchedWriteShuffle.YMM.Folded]>,
AVX5128IBase, EVEX;
}
// This should be used for the AVX512DQ broadcast instructions. It disables
// the unmasked patterns so that we only use the DQ instructions when masking
// is requested.
multiclass avx512_subvec_broadcast_rm_dq<bits<8> opc, string OpcodeStr,
X86VectorVTInfo _Dst, X86VectorVTInfo _Src> {
let hasSideEffects = 0, mayLoad = 1 in
defm rm : AVX512_maskable_split<opc, MRMSrcMem, _Dst, (outs _Dst.RC:$dst),
(ins _Src.MemOp:$src), OpcodeStr, "$src", "$src",
(null_frag),
(_Dst.VT (X86SubVBroadcast
(_Src.VT (_Src.LdFrag addr:$src))))>,
Sched<[SchedWriteShuffle.YMM.Folded]>,
AVX5128IBase, EVEX;
}
let Predicates = [HasAVX512] in {
// 32-bit targets will fail to load a i64 directly but can use ZEXT_LOAD.
def : Pat<(v8i64 (X86VBroadcast (v2i64 (X86vzload addr:$src)))),
(VPBROADCASTQZm addr:$src)>;
}
let Predicates = [HasVLX] in {
// 32-bit targets will fail to load a i64 directly but can use ZEXT_LOAD.
def : Pat<(v2i64 (X86VBroadcast (v2i64 (X86vzload addr:$src)))),
(VPBROADCASTQZ128m addr:$src)>;
def : Pat<(v4i64 (X86VBroadcast (v2i64 (X86vzload addr:$src)))),
(VPBROADCASTQZ256m addr:$src)>;
}
let Predicates = [HasVLX, HasBWI] in {
// loadi16 is tricky to fold, because !isTypeDesirableForOp, justifiably.
// This means we'll encounter truncated i32 loads; match that here.
def : Pat<(v8i16 (X86VBroadcast (i16 (trunc (i32 (load addr:$src)))))),
(VPBROADCASTWZ128m addr:$src)>;
def : Pat<(v16i16 (X86VBroadcast (i16 (trunc (i32 (load addr:$src)))))),
(VPBROADCASTWZ256m addr:$src)>;
def : Pat<(v8i16 (X86VBroadcast
(i16 (trunc (i32 (extloadi16 addr:$src)))))),
(VPBROADCASTWZ128m addr:$src)>;
def : Pat<(v8i16 (X86VBroadcast
(i16 (trunc (i32 (zextloadi16 addr:$src)))))),
(VPBROADCASTWZ128m addr:$src)>;
def : Pat<(v16i16 (X86VBroadcast
(i16 (trunc (i32 (extloadi16 addr:$src)))))),
(VPBROADCASTWZ256m addr:$src)>;
def : Pat<(v16i16 (X86VBroadcast
(i16 (trunc (i32 (zextloadi16 addr:$src)))))),
(VPBROADCASTWZ256m addr:$src)>;
}
let Predicates = [HasBWI] in {
// loadi16 is tricky to fold, because !isTypeDesirableForOp, justifiably.
// This means we'll encounter truncated i32 loads; match that here.
def : Pat<(v32i16 (X86VBroadcast (i16 (trunc (i32 (load addr:$src)))))),
(VPBROADCASTWZm addr:$src)>;
def : Pat<(v32i16 (X86VBroadcast
(i16 (trunc (i32 (extloadi16 addr:$src)))))),
(VPBROADCASTWZm addr:$src)>;
def : Pat<(v32i16 (X86VBroadcast
(i16 (trunc (i32 (zextloadi16 addr:$src)))))),
(VPBROADCASTWZm addr:$src)>;
}
//===----------------------------------------------------------------------===//
// AVX-512 BROADCAST SUBVECTORS
//
defm VBROADCASTI32X4 : avx512_subvec_broadcast_rm<0x5a, "vbroadcasti32x4",
v16i32_info, v4i32x_info>,
EVEX_V512, EVEX_CD8<32, CD8VT4>;
defm VBROADCASTF32X4 : avx512_subvec_broadcast_rm<0x1a, "vbroadcastf32x4",
v16f32_info, v4f32x_info>,
EVEX_V512, EVEX_CD8<32, CD8VT4>;
defm VBROADCASTI64X4 : avx512_subvec_broadcast_rm<0x5b, "vbroadcasti64x4",
v8i64_info, v4i64x_info>, VEX_W,
EVEX_V512, EVEX_CD8<64, CD8VT4>;
defm VBROADCASTF64X4 : avx512_subvec_broadcast_rm<0x1b, "vbroadcastf64x4",
v8f64_info, v4f64x_info>, VEX_W,
EVEX_V512, EVEX_CD8<64, CD8VT4>;
let Predicates = [HasAVX512] in {
def : Pat<(v16f32 (X86SubVBroadcast (loadv8f32 addr:$src))),
(VBROADCASTF64X4rm addr:$src)>;
def : Pat<(v16i32 (X86SubVBroadcast (loadv8i32 addr:$src))),
(VBROADCASTI64X4rm addr:$src)>;
def : Pat<(v32i16 (X86SubVBroadcast (loadv16i16 addr:$src))),
(VBROADCASTI64X4rm addr:$src)>;
def : Pat<(v64i8 (X86SubVBroadcast (loadv32i8 addr:$src))),
(VBROADCASTI64X4rm addr:$src)>;
// Provide fallback in case the load node that is used in the patterns above
// is used by additional users, which prevents the pattern selection.
def : Pat<(v8f64 (X86SubVBroadcast (v4f64 VR256X:$src))),
(VINSERTF64x4Zrr (INSERT_SUBREG (v8f64 (IMPLICIT_DEF)), VR256X:$src, sub_ymm),
(v4f64 VR256X:$src), 1)>;
def : Pat<(v16f32 (X86SubVBroadcast (v8f32 VR256X:$src))),
(VINSERTF64x4Zrr (INSERT_SUBREG (v16f32 (IMPLICIT_DEF)), VR256X:$src, sub_ymm),
(v8f32 VR256X:$src), 1)>;
def : Pat<(v8i64 (X86SubVBroadcast (v4i64 VR256X:$src))),
(VINSERTI64x4Zrr (INSERT_SUBREG (v8i64 (IMPLICIT_DEF)), VR256X:$src, sub_ymm),
(v4i64 VR256X:$src), 1)>;
def : Pat<(v16i32 (X86SubVBroadcast (v8i32 VR256X:$src))),
(VINSERTI64x4Zrr (INSERT_SUBREG (v16i32 (IMPLICIT_DEF)), VR256X:$src, sub_ymm),
(v8i32 VR256X:$src), 1)>;
def : Pat<(v32i16 (X86SubVBroadcast (v16i16 VR256X:$src))),
(VINSERTI64x4Zrr (INSERT_SUBREG (v32i16 (IMPLICIT_DEF)), VR256X:$src, sub_ymm),
(v16i16 VR256X:$src), 1)>;
def : Pat<(v64i8 (X86SubVBroadcast (v32i8 VR256X:$src))),
(VINSERTI64x4Zrr (INSERT_SUBREG (v64i8 (IMPLICIT_DEF)), VR256X:$src, sub_ymm),
(v32i8 VR256X:$src), 1)>;
def : Pat<(v8f64 (X86SubVBroadcast (loadv2f64 addr:$src))),
(VBROADCASTF32X4rm addr:$src)>;
def : Pat<(v8i64 (X86SubVBroadcast (loadv2i64 addr:$src))),
(VBROADCASTI32X4rm addr:$src)>;
def : Pat<(v32i16 (X86SubVBroadcast (loadv8i16 addr:$src))),
(VBROADCASTI32X4rm addr:$src)>;
def : Pat<(v64i8 (X86SubVBroadcast (loadv16i8 addr:$src))),
(VBROADCASTI32X4rm addr:$src)>;
// Patterns for selects of bitcasted operations.
def : Pat<(vselect VK16WM:$mask,
(bc_v16f32 (v8f64 (X86SubVBroadcast (loadv2f64 addr:$src)))),
(v16f32 immAllZerosV)),
(VBROADCASTF32X4rmkz VK16WM:$mask, addr:$src)>;
def : Pat<(vselect VK16WM:$mask,
(bc_v16f32 (v8f64 (X86SubVBroadcast (loadv2f64 addr:$src)))),
VR512:$src0),
(VBROADCASTF32X4rmk VR512:$src0, VK16WM:$mask, addr:$src)>;
def : Pat<(vselect VK16WM:$mask,
(bc_v16i32 (v8i64 (X86SubVBroadcast (loadv2i64 addr:$src)))),
(v16i32 immAllZerosV)),
(VBROADCASTI32X4rmkz VK16WM:$mask, addr:$src)>;
def : Pat<(vselect VK16WM:$mask,
(bc_v16i32 (v8i64 (X86SubVBroadcast (loadv2i64 addr:$src)))),
VR512:$src0),
(VBROADCASTI32X4rmk VR512:$src0, VK16WM:$mask, addr:$src)>;
def : Pat<(vselect VK8WM:$mask,
(bc_v8f64 (v16f32 (X86SubVBroadcast (loadv8f32 addr:$src)))),
(v8f64 immAllZerosV)),
(VBROADCASTF64X4rmkz VK8WM:$mask, addr:$src)>;
def : Pat<(vselect VK8WM:$mask,
(bc_v8f64 (v16f32 (X86SubVBroadcast (loadv8f32 addr:$src)))),
VR512:$src0),
(VBROADCASTF64X4rmk VR512:$src0, VK8WM:$mask, addr:$src)>;
def : Pat<(vselect VK8WM:$mask,
(bc_v8i64 (v16i32 (X86SubVBroadcast (loadv8i32 addr:$src)))),
(v8i64 immAllZerosV)),
(VBROADCASTI64X4rmkz VK8WM:$mask, addr:$src)>;
def : Pat<(vselect VK8WM:$mask,
(bc_v8i64 (v16i32 (X86SubVBroadcast (loadv8i32 addr:$src)))),
VR512:$src0),
(VBROADCASTI64X4rmk VR512:$src0, VK8WM:$mask, addr:$src)>;
}
let Predicates = [HasVLX] in {
defm VBROADCASTI32X4Z256 : avx512_subvec_broadcast_rm<0x5a, "vbroadcasti32x4",
v8i32x_info, v4i32x_info>,
EVEX_V256, EVEX_CD8<32, CD8VT4>;
defm VBROADCASTF32X4Z256 : avx512_subvec_broadcast_rm<0x1a, "vbroadcastf32x4",
v8f32x_info, v4f32x_info>,
EVEX_V256, EVEX_CD8<32, CD8VT4>;
def : Pat<(v4f64 (X86SubVBroadcast (loadv2f64 addr:$src))),
(VBROADCASTF32X4Z256rm addr:$src)>;
def : Pat<(v4i64 (X86SubVBroadcast (loadv2i64 addr:$src))),
(VBROADCASTI32X4Z256rm addr:$src)>;
def : Pat<(v16i16 (X86SubVBroadcast (loadv8i16 addr:$src))),
(VBROADCASTI32X4Z256rm addr:$src)>;
def : Pat<(v32i8 (X86SubVBroadcast (loadv16i8 addr:$src))),
(VBROADCASTI32X4Z256rm addr:$src)>;
// Patterns for selects of bitcasted operations.
def : Pat<(vselect VK8WM:$mask,
(bc_v8f32 (v4f64 (X86SubVBroadcast (loadv2f64 addr:$src)))),
(v8f32 immAllZerosV)),
(VBROADCASTF32X4Z256rmkz VK8WM:$mask, addr:$src)>;
def : Pat<(vselect VK8WM:$mask,
(bc_v8f32 (v4f64 (X86SubVBroadcast (loadv2f64 addr:$src)))),
VR256X:$src0),
(VBROADCASTF32X4Z256rmk VR256X:$src0, VK8WM:$mask, addr:$src)>;
def : Pat<(vselect VK8WM:$mask,
(bc_v8i32 (v4i64 (X86SubVBroadcast (loadv2i64 addr:$src)))),
(v8i32 immAllZerosV)),
(VBROADCASTI32X4Z256rmkz VK8WM:$mask, addr:$src)>;
def : Pat<(vselect VK8WM:$mask,
(bc_v8i32 (v4i64 (X86SubVBroadcast (loadv2i64 addr:$src)))),
VR256X:$src0),
(VBROADCASTI32X4Z256rmk VR256X:$src0, VK8WM:$mask, addr:$src)>;
// Provide fallback in case the load node that is used in the patterns above
// is used by additional users, which prevents the pattern selection.
def : Pat<(v4f64 (X86SubVBroadcast (v2f64 VR128X:$src))),
(VINSERTF32x4Z256rr (INSERT_SUBREG (v4f64 (IMPLICIT_DEF)), VR128X:$src, sub_xmm),
(v2f64 VR128X:$src), 1)>;
def : Pat<(v8f32 (X86SubVBroadcast (v4f32 VR128X:$src))),
(VINSERTF32x4Z256rr (INSERT_SUBREG (v8f32 (IMPLICIT_DEF)), VR128X:$src, sub_xmm),
(v4f32 VR128X:$src), 1)>;
def : Pat<(v4i64 (X86SubVBroadcast (v2i64 VR128X:$src))),
(VINSERTI32x4Z256rr (INSERT_SUBREG (v4i64 (IMPLICIT_DEF)), VR128X:$src, sub_xmm),
(v2i64 VR128X:$src), 1)>;
def : Pat<(v8i32 (X86SubVBroadcast (v4i32 VR128X:$src))),
(VINSERTI32x4Z256rr (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)), VR128X:$src, sub_xmm),
(v4i32 VR128X:$src), 1)>;
def : Pat<(v16i16 (X86SubVBroadcast (v8i16 VR128X:$src))),
(VINSERTI32x4Z256rr (INSERT_SUBREG (v16i16 (IMPLICIT_DEF)), VR128X:$src, sub_xmm),
(v8i16 VR128X:$src), 1)>;
def : Pat<(v32i8 (X86SubVBroadcast (v16i8 VR128X:$src))),
(VINSERTI32x4Z256rr (INSERT_SUBREG (v32i8 (IMPLICIT_DEF)), VR128X:$src, sub_xmm),
(v16i8 VR128X:$src), 1)>;
}
let Predicates = [HasVLX, HasDQI] in {
defm VBROADCASTI64X2Z128 : avx512_subvec_broadcast_rm_dq<0x5a, "vbroadcasti64x2",
v4i64x_info, v2i64x_info>, VEX_W1X,
EVEX_V256, EVEX_CD8<64, CD8VT2>;
defm VBROADCASTF64X2Z128 : avx512_subvec_broadcast_rm_dq<0x1a, "vbroadcastf64x2",
v4f64x_info, v2f64x_info>, VEX_W1X,
EVEX_V256, EVEX_CD8<64, CD8VT2>;
// Patterns for selects of bitcasted operations.
def : Pat<(vselect VK4WM:$mask,
(bc_v4f64 (v8f32 (X86SubVBroadcast (loadv4f32 addr:$src)))),
(v4f64 immAllZerosV)),
(VBROADCASTF64X2Z128rmkz VK4WM:$mask, addr:$src)>;
def : Pat<(vselect VK4WM:$mask,
(bc_v4f64 (v8f32 (X86SubVBroadcast (loadv4f32 addr:$src)))),
VR256X:$src0),
(VBROADCASTF64X2Z128rmk VR256X:$src0, VK4WM:$mask, addr:$src)>;
def : Pat<(vselect VK4WM:$mask,
(bc_v4i64 (v8i32 (X86SubVBroadcast (loadv4i32 addr:$src)))),
(v4i64 immAllZerosV)),
(VBROADCASTI64X2Z128rmkz VK4WM:$mask, addr:$src)>;
def : Pat<(vselect VK4WM:$mask,
(bc_v4i64 (v8i32 (X86SubVBroadcast (loadv4i32 addr:$src)))),
VR256X:$src0),
(VBROADCASTI64X2Z128rmk VR256X:$src0, VK4WM:$mask, addr:$src)>;
}
let Predicates = [HasDQI] in {
defm VBROADCASTI64X2 : avx512_subvec_broadcast_rm_dq<0x5a, "vbroadcasti64x2",
v8i64_info, v2i64x_info>, VEX_W,
EVEX_V512, EVEX_CD8<64, CD8VT2>;
defm VBROADCASTI32X8 : avx512_subvec_broadcast_rm_dq<0x5b, "vbroadcasti32x8",
v16i32_info, v8i32x_info>,
EVEX_V512, EVEX_CD8<32, CD8VT8>;
defm VBROADCASTF64X2 : avx512_subvec_broadcast_rm_dq<0x1a, "vbroadcastf64x2",
v8f64_info, v2f64x_info>, VEX_W,
EVEX_V512, EVEX_CD8<64, CD8VT2>;
defm VBROADCASTF32X8 : avx512_subvec_broadcast_rm_dq<0x1b, "vbroadcastf32x8",
v16f32_info, v8f32x_info>,
EVEX_V512, EVEX_CD8<32, CD8VT8>;
// Patterns for selects of bitcasted operations.
def : Pat<(vselect VK16WM:$mask,
(bc_v16f32 (v8f64 (X86SubVBroadcast (loadv4f64 addr:$src)))),
(v16f32 immAllZerosV)),
(VBROADCASTF32X8rmkz VK16WM:$mask, addr:$src)>;
def : Pat<(vselect VK16WM:$mask,
(bc_v16f32 (v8f64 (X86SubVBroadcast (loadv4f64 addr:$src)))),
VR512:$src0),
(VBROADCASTF32X8rmk VR512:$src0, VK16WM:$mask, addr:$src)>;
def : Pat<(vselect VK16WM:$mask,
(bc_v16i32 (v8i64 (X86SubVBroadcast (loadv4i64 addr:$src)))),
(v16i32 immAllZerosV)),
(VBROADCASTI32X8rmkz VK16WM:$mask, addr:$src)>;
def : Pat<(vselect VK16WM:$mask,
(bc_v16i32 (v8i64 (X86SubVBroadcast (loadv4i64 addr:$src)))),
VR512:$src0),
(VBROADCASTI32X8rmk VR512:$src0, VK16WM:$mask, addr:$src)>;
def : Pat<(vselect VK8WM:$mask,
(bc_v8f64 (v16f32 (X86SubVBroadcast (loadv4f32 addr:$src)))),
(v8f64 immAllZerosV)),
(VBROADCASTF64X2rmkz VK8WM:$mask, addr:$src)>;
def : Pat<(vselect VK8WM:$mask,
(bc_v8f64 (v16f32 (X86SubVBroadcast (loadv4f32 addr:$src)))),
VR512:$src0),
(VBROADCASTF64X2rmk VR512:$src0, VK8WM:$mask, addr:$src)>;
def : Pat<(vselect VK8WM:$mask,
(bc_v8i64 (v16i32 (X86SubVBroadcast (loadv4i32 addr:$src)))),
(v8i64 immAllZerosV)),
(VBROADCASTI64X2rmkz VK8WM:$mask, addr:$src)>;
def : Pat<(vselect VK8WM:$mask,
(bc_v8i64 (v16i32 (X86SubVBroadcast (loadv4i32 addr:$src)))),
VR512:$src0),
(VBROADCASTI64X2rmk VR512:$src0, VK8WM:$mask, addr:$src)>;
}
multiclass avx512_common_broadcast_32x2<bits<8> opc, string OpcodeStr,
AVX512VLVectorVTInfo _Dst, AVX512VLVectorVTInfo _Src> {
let Predicates = [HasDQI] in
defm Z : avx512_broadcast_rm_split<opc, OpcodeStr, NAME, WriteShuffle256,
WriteShuffle256Ld, _Dst.info512,
_Src.info512, _Src.info128, null_frag>,
EVEX_V512;
let Predicates = [HasDQI, HasVLX] in
defm Z256 : avx512_broadcast_rm_split<opc, OpcodeStr, NAME, WriteShuffle256,
WriteShuffle256Ld, _Dst.info256,
_Src.info256, _Src.info128, null_frag>,
EVEX_V256;
}
multiclass avx512_common_broadcast_i32x2<bits<8> opc, string OpcodeStr,
AVX512VLVectorVTInfo _Dst, AVX512VLVectorVTInfo _Src> :
avx512_common_broadcast_32x2<opc, OpcodeStr, _Dst, _Src> {
let Predicates = [HasDQI, HasVLX] in
defm Z128 : avx512_broadcast_rm_split<opc, OpcodeStr, NAME, WriteShuffle,
WriteShuffleXLd, _Dst.info128,
_Src.info128, _Src.info128, null_frag>,
EVEX_V128;
}
defm VBROADCASTI32X2 : avx512_common_broadcast_i32x2<0x59, "vbroadcasti32x2",
avx512vl_i32_info, avx512vl_i64_info>;
defm VBROADCASTF32X2 : avx512_common_broadcast_32x2<0x19, "vbroadcastf32x2",
avx512vl_f32_info, avx512vl_f64_info>;
let Predicates = [HasVLX] in {
def : Pat<(v8f32 (X86VBroadcast (v8f32 VR256X:$src))),
(VBROADCASTSSZ256r (v4f32 (EXTRACT_SUBREG (v8f32 VR256X:$src), sub_xmm)))>;
def : Pat<(v4f64 (X86VBroadcast (v4f64 VR256X:$src))),
(VBROADCASTSDZ256r (v2f64 (EXTRACT_SUBREG (v4f64 VR256X:$src), sub_xmm)))>;
}
def : Pat<(v16f32 (X86VBroadcast (v16f32 VR512:$src))),
(VBROADCASTSSZr (v4f32 (EXTRACT_SUBREG (v16f32 VR512:$src), sub_xmm)))>;
def : Pat<(v16f32 (X86VBroadcast (v8f32 VR256X:$src))),
(VBROADCASTSSZr (v4f32 (EXTRACT_SUBREG (v8f32 VR256X:$src), sub_xmm)))>;
def : Pat<(v8f64 (X86VBroadcast (v8f64 VR512:$src))),
(VBROADCASTSDZr (v2f64 (EXTRACT_SUBREG (v8f64 VR512:$src), sub_xmm)))>;
def : Pat<(v8f64 (X86VBroadcast (v4f64 VR256X:$src))),
(VBROADCASTSDZr (v2f64 (EXTRACT_SUBREG (v4f64 VR256X:$src), sub_xmm)))>;
//===----------------------------------------------------------------------===//
// AVX-512 BROADCAST MASK TO VECTOR REGISTER
//---
multiclass avx512_mask_broadcastm<bits<8> opc, string OpcodeStr,
X86VectorVTInfo _, RegisterClass KRC> {
def rr : AVX512XS8I<opc, MRMSrcReg, (outs _.RC:$dst), (ins KRC:$src),
!strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
[(set _.RC:$dst, (_.VT (X86VBroadcastm KRC:$src)))]>,
EVEX, Sched<[WriteShuffle]>;
}
multiclass avx512_mask_broadcast<bits<8> opc, string OpcodeStr,
AVX512VLVectorVTInfo VTInfo, RegisterClass KRC> {
let Predicates = [HasCDI] in
defm Z : avx512_mask_broadcastm<opc, OpcodeStr, VTInfo.info512, KRC>, EVEX_V512;
let Predicates = [HasCDI, HasVLX] in {
defm Z256 : avx512_mask_broadcastm<opc, OpcodeStr, VTInfo.info256, KRC>, EVEX_V256;
defm Z128 : avx512_mask_broadcastm<opc, OpcodeStr, VTInfo.info128, KRC>, EVEX_V128;
}
}
defm VPBROADCASTMW2D : avx512_mask_broadcast<0x3A, "vpbroadcastmw2d",
avx512vl_i32_info, VK16>;
defm VPBROADCASTMB2Q : avx512_mask_broadcast<0x2A, "vpbroadcastmb2q",
avx512vl_i64_info, VK8>, VEX_W;
//===----------------------------------------------------------------------===//
// -- VPERMI2 - 3 source operands form --
multiclass avx512_perm_i<bits<8> opc, string OpcodeStr,
X86FoldableSchedWrite sched,
X86VectorVTInfo _, X86VectorVTInfo IdxVT> {
let Constraints = "$src1 = $dst", ExeDomain = _.ExeDomain,
hasSideEffects = 0 in {
defm rr: AVX512_maskable_3src_cast<opc, MRMSrcReg, _, IdxVT, (outs _.RC:$dst),
(ins _.RC:$src2, _.RC:$src3),
OpcodeStr, "$src3, $src2", "$src2, $src3",
(_.VT (X86VPermt2 _.RC:$src2, IdxVT.RC:$src1, _.RC:$src3)), 1>,
EVEX_4V, AVX5128IBase, Sched<[sched]>;
let mayLoad = 1 in
defm rm: AVX512_maskable_3src_cast<opc, MRMSrcMem, _, IdxVT, (outs _.RC:$dst),
(ins _.RC:$src2, _.MemOp:$src3),
OpcodeStr, "$src3, $src2", "$src2, $src3",
(_.VT (X86VPermt2 _.RC:$src2, IdxVT.RC:$src1,
(_.VT (_.LdFrag addr:$src3)))), 1>,
EVEX_4V, AVX5128IBase, Sched<[sched.Folded, sched.ReadAfterFold]>;
}
}
multiclass avx512_perm_i_mb<bits<8> opc, string OpcodeStr,
X86FoldableSchedWrite sched,
X86VectorVTInfo _, X86VectorVTInfo IdxVT> {
let Constraints = "$src1 = $dst", ExeDomain = _.ExeDomain,
hasSideEffects = 0, mayLoad = 1 in
defm rmb: AVX512_maskable_3src_cast<opc, MRMSrcMem, _, IdxVT, (outs _.RC:$dst),
(ins _.RC:$src2, _.ScalarMemOp:$src3),
OpcodeStr, !strconcat("${src3}", _.BroadcastStr,", $src2"),
!strconcat("$src2, ${src3}", _.BroadcastStr ),
(_.VT (X86VPermt2 _.RC:$src2,
IdxVT.RC:$src1,(_.VT (X86VBroadcast (_.ScalarLdFrag addr:$src3))))), 1>,
AVX5128IBase, EVEX_4V, EVEX_B,
Sched<[sched.Folded, sched.ReadAfterFold]>;
}
multiclass avx512_perm_i_sizes<bits<8> opc, string OpcodeStr,
X86FoldableSchedWrite sched,
AVX512VLVectorVTInfo VTInfo,
AVX512VLVectorVTInfo ShuffleMask> {
defm NAME: avx512_perm_i<opc, OpcodeStr, sched, VTInfo.info512,
ShuffleMask.info512>,
avx512_perm_i_mb<opc, OpcodeStr, sched, VTInfo.info512,
ShuffleMask.info512>, EVEX_V512;
let Predicates = [HasVLX] in {
defm NAME#128: avx512_perm_i<opc, OpcodeStr, sched, VTInfo.info128,
ShuffleMask.info128>,
avx512_perm_i_mb<opc, OpcodeStr, sched, VTInfo.info128,
ShuffleMask.info128>, EVEX_V128;
defm NAME#256: avx512_perm_i<opc, OpcodeStr, sched, VTInfo.info256,
ShuffleMask.info256>,
avx512_perm_i_mb<opc, OpcodeStr, sched, VTInfo.info256,
ShuffleMask.info256>, EVEX_V256;
}
}
multiclass avx512_perm_i_sizes_bw<bits<8> opc, string OpcodeStr,
X86FoldableSchedWrite sched,
AVX512VLVectorVTInfo VTInfo,
AVX512VLVectorVTInfo Idx,
Predicate Prd> {
let Predicates = [Prd] in
defm NAME: avx512_perm_i<opc, OpcodeStr, sched, VTInfo.info512,
Idx.info512>, EVEX_V512;
let Predicates = [Prd, HasVLX] in {
defm NAME#128: avx512_perm_i<opc, OpcodeStr, sched, VTInfo.info128,
Idx.info128>, EVEX_V128;
defm NAME#256: avx512_perm_i<opc, OpcodeStr, sched, VTInfo.info256,
Idx.info256>, EVEX_V256;
}
}
defm VPERMI2D : avx512_perm_i_sizes<0x76, "vpermi2d", WriteVarShuffle256,
avx512vl_i32_info, avx512vl_i32_info>, EVEX_CD8<32, CD8VF>;
defm VPERMI2Q : avx512_perm_i_sizes<0x76, "vpermi2q", WriteVarShuffle256,
avx512vl_i64_info, avx512vl_i64_info>, VEX_W, EVEX_CD8<64, CD8VF>;
defm VPERMI2W : avx512_perm_i_sizes_bw<0x75, "vpermi2w", WriteVarShuffle256,
avx512vl_i16_info, avx512vl_i16_info, HasBWI>,
VEX_W, EVEX_CD8<16, CD8VF>;
defm VPERMI2B : avx512_perm_i_sizes_bw<0x75, "vpermi2b", WriteVarShuffle256,
avx512vl_i8_info, avx512vl_i8_info, HasVBMI>,
EVEX_CD8<8, CD8VF>;
defm VPERMI2PS : avx512_perm_i_sizes<0x77, "vpermi2ps", WriteFVarShuffle256,
avx512vl_f32_info, avx512vl_i32_info>, EVEX_CD8<32, CD8VF>;
defm VPERMI2PD : avx512_perm_i_sizes<0x77, "vpermi2pd", WriteFVarShuffle256,
avx512vl_f64_info, avx512vl_i64_info>, VEX_W, EVEX_CD8<64, CD8VF>;
// Extra patterns to deal with extra bitcasts due to passthru and index being
// different types on the fp versions.
multiclass avx512_perm_i_lowering<string InstrStr, X86VectorVTInfo _,
X86VectorVTInfo IdxVT,
X86VectorVTInfo CastVT> {
def : Pat<(_.VT (vselect _.KRCWM:$mask,
(X86VPermt2 (_.VT _.RC:$src2),
(IdxVT.VT (bitconvert (CastVT.VT _.RC:$src1))), _.RC:$src3),
(_.VT (bitconvert (CastVT.VT _.RC:$src1))))),
(!cast<Instruction>(InstrStr#"rrk") _.RC:$src1, _.KRCWM:$mask,
_.RC:$src2, _.RC:$src3)>;
def : Pat<(_.VT (vselect _.KRCWM:$mask,
(X86VPermt2 _.RC:$src2,
(IdxVT.VT (bitconvert (CastVT.VT _.RC:$src1))),
(_.LdFrag addr:$src3)),
(_.VT (bitconvert (CastVT.VT _.RC:$src1))))),
(!cast<Instruction>(InstrStr#"rmk") _.RC:$src1, _.KRCWM:$mask,
_.RC:$src2, addr:$src3)>;
def : Pat<(_.VT (vselect _.KRCWM:$mask,
(X86VPermt2 _.RC:$src2,
(IdxVT.VT (bitconvert (CastVT.VT _.RC:$src1))),
(X86VBroadcast (_.ScalarLdFrag addr:$src3))),
(_.VT (bitconvert (CastVT.VT _.RC:$src1))))),
(!cast<Instruction>(InstrStr#"rmbk") _.RC:$src1, _.KRCWM:$mask,
_.RC:$src2, addr:$src3)>;
}
// TODO: Should we add more casts? The vXi64 case is common due to ABI.
defm : avx512_perm_i_lowering<"VPERMI2PS", v16f32_info, v16i32_info, v8i64_info>;
defm : avx512_perm_i_lowering<"VPERMI2PS256", v8f32x_info, v8i32x_info, v4i64x_info>;
defm : avx512_perm_i_lowering<"VPERMI2PS128", v4f32x_info, v4i32x_info, v2i64x_info>;
// VPERMT2
multiclass avx512_perm_t<bits<8> opc, string OpcodeStr,
X86FoldableSchedWrite sched,
X86VectorVTInfo _, X86VectorVTInfo IdxVT> {
let Constraints = "$src1 = $dst", ExeDomain = _.ExeDomain in {
defm rr: AVX512_maskable_3src<opc, MRMSrcReg, _, (outs _.RC:$dst),
(ins IdxVT.RC:$src2, _.RC:$src3),
OpcodeStr, "$src3, $src2", "$src2, $src3",
(_.VT (X86VPermt2 _.RC:$src1, IdxVT.RC:$src2, _.RC:$src3)), 1>,
EVEX_4V, AVX5128IBase, Sched<[sched]>;
defm rm: AVX512_maskable_3src<opc, MRMSrcMem, _, (outs _.RC:$dst),
(ins IdxVT.RC:$src2, _.MemOp:$src3),
OpcodeStr, "$src3, $src2", "$src2, $src3",
(_.VT (X86VPermt2 _.RC:$src1, IdxVT.RC:$src2,
(_.LdFrag addr:$src3))), 1>,
EVEX_4V, AVX5128IBase, Sched<[sched.Folded, sched.ReadAfterFold]>;
}
}
multiclass avx512_perm_t_mb<bits<8> opc, string OpcodeStr,
X86FoldableSchedWrite sched,
X86VectorVTInfo _, X86VectorVTInfo IdxVT> {
let Constraints = "$src1 = $dst", ExeDomain = _.ExeDomain in
defm rmb: AVX512_maskable_3src<opc, MRMSrcMem, _, (outs _.RC:$dst),
(ins IdxVT.RC:$src2, _.ScalarMemOp:$src3),
OpcodeStr, !strconcat("${src3}", _.BroadcastStr,", $src2"),
!strconcat("$src2, ${src3}", _.BroadcastStr ),
(_.VT (X86VPermt2 _.RC:$src1,
IdxVT.RC:$src2,(_.VT (X86VBroadcast (_.ScalarLdFrag addr:$src3))))), 1>,
AVX5128IBase, EVEX_4V, EVEX_B,
Sched<[sched.Folded, sched.ReadAfterFold]>;
}
multiclass avx512_perm_t_sizes<bits<8> opc, string OpcodeStr,
X86FoldableSchedWrite sched,
AVX512VLVectorVTInfo VTInfo,
AVX512VLVectorVTInfo ShuffleMask> {
defm NAME: avx512_perm_t<opc, OpcodeStr, sched, VTInfo.info512,
ShuffleMask.info512>,
avx512_perm_t_mb<opc, OpcodeStr, sched, VTInfo.info512,
ShuffleMask.info512>, EVEX_V512;
let Predicates = [HasVLX] in {
defm NAME#128: avx512_perm_t<opc, OpcodeStr, sched, VTInfo.info128,
ShuffleMask.info128>,
avx512_perm_t_mb<opc, OpcodeStr, sched, VTInfo.info128,
ShuffleMask.info128>, EVEX_V128;
defm NAME#256: avx512_perm_t<opc, OpcodeStr, sched, VTInfo.info256,
ShuffleMask.info256>,
avx512_perm_t_mb<opc, OpcodeStr, sched, VTInfo.info256,
ShuffleMask.info256>, EVEX_V256;
}
}
multiclass avx512_perm_t_sizes_bw<bits<8> opc, string OpcodeStr,
X86FoldableSchedWrite sched,
AVX512VLVectorVTInfo VTInfo,
AVX512VLVectorVTInfo Idx, Predicate Prd> {
let Predicates = [Prd] in
defm NAME: avx512_perm_t<opc, OpcodeStr, sched, VTInfo.info512,
Idx.info512>, EVEX_V512;
let Predicates = [Prd, HasVLX] in {
defm NAME#128: avx512_perm_t<opc, OpcodeStr, sched, VTInfo.info128,
Idx.info128>, EVEX_V128;
defm NAME#256: avx512_perm_t<opc, OpcodeStr, sched, VTInfo.info256,
Idx.info256>, EVEX_V256;
}
}
defm VPERMT2D : avx512_perm_t_sizes<0x7E, "vpermt2d", WriteVarShuffle256,
avx512vl_i32_info, avx512vl_i32_info>, EVEX_CD8<32, CD8VF>;
defm VPERMT2Q : avx512_perm_t_sizes<0x7E, "vpermt2q", WriteVarShuffle256,
avx512vl_i64_info, avx512vl_i64_info>, VEX_W, EVEX_CD8<64, CD8VF>;
defm VPERMT2W : avx512_perm_t_sizes_bw<0x7D, "vpermt2w", WriteVarShuffle256,
avx512vl_i16_info, avx512vl_i16_info, HasBWI>,
VEX_W, EVEX_CD8<16, CD8VF>;
defm VPERMT2B : avx512_perm_t_sizes_bw<0x7D, "vpermt2b", WriteVarShuffle256,
avx512vl_i8_info, avx512vl_i8_info, HasVBMI>,
EVEX_CD8<8, CD8VF>;
defm VPERMT2PS : avx512_perm_t_sizes<0x7F, "vpermt2ps", WriteFVarShuffle256,
avx512vl_f32_info, avx512vl_i32_info>, EVEX_CD8<32, CD8VF>;
defm VPERMT2PD : avx512_perm_t_sizes<0x7F, "vpermt2pd", WriteFVarShuffle256,
avx512vl_f64_info, avx512vl_i64_info>, VEX_W, EVEX_CD8<64, CD8VF>;
//===----------------------------------------------------------------------===//
// AVX-512 - BLEND using mask
//
multiclass WriteFVarBlendask<bits<8> opc, string OpcodeStr,
X86FoldableSchedWrite sched, X86VectorVTInfo _> {
let ExeDomain = _.ExeDomain, hasSideEffects = 0 in {
def rr : AVX5128I<opc, MRMSrcReg, (outs _.RC:$dst),
(ins _.RC:$src1, _.RC:$src2),
!strconcat(OpcodeStr,
"\t{$src2, $src1, ${dst}|${dst}, $src1, $src2}"), []>,
EVEX_4V, Sched<[sched]>;
def rrk : AVX5128I<opc, MRMSrcReg, (outs _.RC:$dst),
(ins _.KRCWM:$mask, _.RC:$src1, _.RC:$src2),
!strconcat(OpcodeStr,
"\t{$src2, $src1, ${dst} {${mask}}|${dst} {${mask}}, $src1, $src2}"),
[]>, EVEX_4V, EVEX_K, Sched<[sched]>;
def rrkz : AVX5128I<opc, MRMSrcReg, (outs _.RC:$dst),
(ins _.KRCWM:$mask, _.RC:$src1, _.RC:$src2),
!strconcat(OpcodeStr,
"\t{$src2, $src1, ${dst} {${mask}} {z}|${dst} {${mask}} {z}, $src1, $src2}"),
[]>, EVEX_4V, EVEX_KZ, Sched<[sched]>, NotMemoryFoldable;
let mayLoad = 1 in {
def rm : AVX5128I<opc, MRMSrcMem, (outs _.RC:$dst),
(ins _.RC:$src1, _.MemOp:$src2),
!strconcat(OpcodeStr,
"\t{$src2, $src1, ${dst}|${dst}, $src1, $src2}"),
[]>, EVEX_4V, EVEX_CD8<_.EltSize, CD8VF>,
Sched<[sched.Folded, sched.ReadAfterFold]>;
def rmk : AVX5128I<opc, MRMSrcMem, (outs _.RC:$dst),
(ins _.KRCWM:$mask, _.RC:$src1, _.MemOp:$src2),
!strconcat(OpcodeStr,
"\t{$src2, $src1, ${dst} {${mask}}|${dst} {${mask}}, $src1, $src2}"),
[]>, EVEX_4V, EVEX_K, EVEX_CD8<_.EltSize, CD8VF>,
Sched<[sched.Folded, sched.ReadAfterFold]>;
def rmkz : AVX5128I<opc, MRMSrcMem, (outs _.RC:$dst),
(ins _.KRCWM:$mask, _.RC:$src1, _.MemOp:$src2),
!strconcat(OpcodeStr,
"\t{$src2, $src1, ${dst} {${mask}} {z}|${dst} {${mask}} {z}, $src1, $src2}"),
[]>, EVEX_4V, EVEX_KZ, EVEX_CD8<_.EltSize, CD8VF>,
Sched<[sched.Folded, sched.ReadAfterFold]>, NotMemoryFoldable;
}
}
}
multiclass WriteFVarBlendask_rmb<bits<8> opc, string OpcodeStr,
X86FoldableSchedWrite sched, X86VectorVTInfo _> {
let mayLoad = 1, hasSideEffects = 0 in {
def rmbk : AVX5128I<opc, MRMSrcMem, (outs _.RC:$dst),
(ins _.KRCWM:$mask, _.RC:$src1, _.ScalarMemOp:$src2),
!strconcat(OpcodeStr,
"\t{${src2}", _.BroadcastStr, ", $src1, $dst {${mask}}|",
"$dst {${mask}}, $src1, ${src2}", _.BroadcastStr, "}"), []>,
EVEX_4V, EVEX_K, EVEX_B, EVEX_CD8<_.EltSize, CD8VF>,
Sched<[sched.Folded, sched.ReadAfterFold]>;
def rmbkz : AVX5128I<opc, MRMSrcMem, (outs _.RC:$dst),
(ins _.KRCWM:$mask, _.RC:$src1, _.ScalarMemOp:$src2),
!strconcat(OpcodeStr,
"\t{${src2}", _.BroadcastStr, ", $src1, $dst {${mask}} {z}|",
"$dst {${mask}} {z}, $src1, ${src2}", _.BroadcastStr, "}"), []>,
EVEX_4V, EVEX_KZ, EVEX_B, EVEX_CD8<_.EltSize, CD8VF>,
Sched<[sched.Folded, sched.ReadAfterFold]>, NotMemoryFoldable;
def rmb : AVX5128I<opc, MRMSrcMem, (outs _.RC:$dst),
(ins _.RC:$src1, _.ScalarMemOp:$src2),
!strconcat(OpcodeStr,
"\t{${src2}", _.BroadcastStr, ", $src1, $dst|",
"$dst, $src1, ${src2}", _.BroadcastStr, "}"), []>,
EVEX_4V, EVEX_B, EVEX_CD8<_.EltSize, CD8VF>,
Sched<[sched.Folded, sched.ReadAfterFold]>;
}
}
multiclass blendmask_dq<bits<8> opc, string OpcodeStr, X86SchedWriteWidths sched,
AVX512VLVectorVTInfo VTInfo> {
defm Z : WriteFVarBlendask<opc, OpcodeStr, sched.ZMM, VTInfo.info512>,
WriteFVarBlendask_rmb<opc, OpcodeStr, sched.ZMM, VTInfo.info512>,
EVEX_V512;
let Predicates = [HasVLX] in {
defm Z256 : WriteFVarBlendask<opc, OpcodeStr, sched.YMM, VTInfo.info256>,
WriteFVarBlendask_rmb<opc, OpcodeStr, sched.YMM, VTInfo.info256>,
EVEX_V256;
defm Z128 : WriteFVarBlendask<opc, OpcodeStr, sched.XMM, VTInfo.info128>,
WriteFVarBlendask_rmb<opc, OpcodeStr, sched.XMM, VTInfo.info128>,
EVEX_V128;
}
}
multiclass blendmask_bw<bits<8> opc, string OpcodeStr, X86SchedWriteWidths sched,
AVX512VLVectorVTInfo VTInfo> {
let Predicates = [HasBWI] in
defm Z : WriteFVarBlendask<opc, OpcodeStr, sched.ZMM, VTInfo.info512>,
EVEX_V512;
let Predicates = [HasBWI, HasVLX] in {
defm Z256 : WriteFVarBlendask<opc, OpcodeStr, sched.YMM, VTInfo.info256>,
EVEX_V256;
defm Z128 : WriteFVarBlendask<opc, OpcodeStr, sched.XMM, VTInfo.info128>,
EVEX_V128;
}
}
defm VBLENDMPS : blendmask_dq<0x65, "vblendmps", SchedWriteFVarBlend,
avx512vl_f32_info>;
defm VBLENDMPD : blendmask_dq<0x65, "vblendmpd", SchedWriteFVarBlend,
avx512vl_f64_info>, VEX_W;
defm VPBLENDMD : blendmask_dq<0x64, "vpblendmd", SchedWriteVarBlend,
avx512vl_i32_info>;
defm VPBLENDMQ : blendmask_dq<0x64, "vpblendmq", SchedWriteVarBlend,
avx512vl_i64_info>, VEX_W;
defm VPBLENDMB : blendmask_bw<0x66, "vpblendmb", SchedWriteVarBlend,
avx512vl_i8_info>;
defm VPBLENDMW : blendmask_bw<0x66, "vpblendmw", SchedWriteVarBlend,
avx512vl_i16_info>, VEX_W;
//===----------------------------------------------------------------------===//
// Compare Instructions
//===----------------------------------------------------------------------===//
// avx512_cmp_scalar - AVX512 CMPSS and CMPSD
multiclass avx512_cmp_scalar<X86VectorVTInfo _, SDNode OpNode, SDNode OpNodeSAE,
PatFrag OpNode_su, PatFrag OpNodeSAE_su,
X86FoldableSchedWrite sched> {
defm rr_Int : AVX512_maskable_cmp<0xC2, MRMSrcReg, _,
(outs _.KRC:$dst),
(ins _.RC:$src1, _.RC:$src2, u8imm:$cc),
"vcmp"#_.Suffix,
"$cc, $src2, $src1", "$src1, $src2, $cc",
(OpNode (_.VT _.RC:$src1), (_.VT _.RC:$src2), imm:$cc),
(OpNode_su (_.VT _.RC:$src1), (_.VT _.RC:$src2),
imm:$cc)>, EVEX_4V, VEX_LIG, Sched<[sched]>;
let mayLoad = 1 in
defm rm_Int : AVX512_maskable_cmp<0xC2, MRMSrcMem, _,
(outs _.KRC:$dst),
(ins _.RC:$src1, _.IntScalarMemOp:$src2, u8imm:$cc),
"vcmp"#_.Suffix,
"$cc, $src2, $src1", "$src1, $src2, $cc",
(OpNode (_.VT _.RC:$src1), _.ScalarIntMemCPat:$src2,
imm:$cc),
(OpNode_su (_.VT _.RC:$src1), _.ScalarIntMemCPat:$src2,
imm:$cc)>, EVEX_4V, VEX_LIG, EVEX_CD8<_.EltSize, CD8VT1>,
Sched<[sched.Folded, sched.ReadAfterFold]>;
defm rrb_Int : AVX512_maskable_cmp<0xC2, MRMSrcReg, _,
(outs _.KRC:$dst),
(ins _.RC:$src1, _.RC:$src2, u8imm:$cc),
"vcmp"#_.Suffix,
"$cc, {sae}, $src2, $src1","$src1, $src2, {sae}, $cc",
(OpNodeSAE (_.VT _.RC:$src1), (_.VT _.RC:$src2),
imm:$cc),
(OpNodeSAE_su (_.VT _.RC:$src1), (_.VT _.RC:$src2),
imm:$cc)>,
EVEX_4V, VEX_LIG, EVEX_B, Sched<[sched]>;
let isCodeGenOnly = 1 in {
let isCommutable = 1 in
def rr : AVX512Ii8<0xC2, MRMSrcReg,
(outs _.KRC:$dst), (ins _.FRC:$src1, _.FRC:$src2, u8imm:$cc),
!strconcat("vcmp", _.Suffix,
"\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}"),
[(set _.KRC:$dst, (OpNode _.FRC:$src1,
_.FRC:$src2,
imm:$cc))]>,
EVEX_4V, VEX_LIG, Sched<[sched]>;
def rm : AVX512Ii8<0xC2, MRMSrcMem,
(outs _.KRC:$dst),
(ins _.FRC:$src1, _.ScalarMemOp:$src2, u8imm:$cc),
!strconcat("vcmp", _.Suffix,
"\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}"),
[(set _.KRC:$dst, (OpNode _.FRC:$src1,
(_.ScalarLdFrag addr:$src2),
imm:$cc))]>,
EVEX_4V, VEX_LIG, EVEX_CD8<_.EltSize, CD8VT1>,
Sched<[sched.Folded, sched.ReadAfterFold]>;
}
}
def X86cmpms_su : PatFrag<(ops node:$src1, node:$src2, node:$cc),
(X86cmpms node:$src1, node:$src2, node:$cc), [{
return N->hasOneUse();
}]>;
def X86cmpmsSAE_su : PatFrag<(ops node:$src1, node:$src2, node:$cc),
(X86cmpmsSAE node:$src1, node:$src2, node:$cc), [{
return N->hasOneUse();
}]>;
let Predicates = [HasAVX512] in {
let ExeDomain = SSEPackedSingle in
defm VCMPSSZ : avx512_cmp_scalar<f32x_info, X86cmpms, X86cmpmsSAE,
X86cmpms_su, X86cmpmsSAE_su,
SchedWriteFCmp.Scl>, AVX512XSIi8Base;
let ExeDomain = SSEPackedDouble in
defm VCMPSDZ : avx512_cmp_scalar<f64x_info, X86cmpms, X86cmpmsSAE,
X86cmpms_su, X86cmpmsSAE_su,
SchedWriteFCmp.Scl>, AVX512XDIi8Base, VEX_W;
}
multiclass avx512_icmp_packed<bits<8> opc, string OpcodeStr, PatFrag OpNode,
PatFrag OpNode_su, X86FoldableSchedWrite sched,
X86VectorVTInfo _, bit IsCommutable> {
let isCommutable = IsCommutable in
def rr : AVX512BI<opc, MRMSrcReg,
(outs _.KRC:$dst), (ins _.RC:$src1, _.RC:$src2),
!strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
[(set _.KRC:$dst, (OpNode (_.VT _.RC:$src1), (_.VT _.RC:$src2)))]>,
EVEX_4V, Sched<[sched]>;
def rm : AVX512BI<opc, MRMSrcMem,
(outs _.KRC:$dst), (ins _.RC:$src1, _.MemOp:$src2),
!strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
[(set _.KRC:$dst, (OpNode (_.VT _.RC:$src1),
(_.VT (_.LdFrag addr:$src2))))]>,
EVEX_4V, Sched<[sched.Folded, sched.ReadAfterFold]>;
let isCommutable = IsCommutable in
def rrk : AVX512BI<opc, MRMSrcReg,
(outs _.KRC:$dst), (ins _.KRCWM:$mask, _.RC:$src1, _.RC:$src2),
!strconcat(OpcodeStr, "\t{$src2, $src1, $dst {${mask}}|",
"$dst {${mask}}, $src1, $src2}"),
[(set _.KRC:$dst, (and _.KRCWM:$mask,
(OpNode_su (_.VT _.RC:$src1), (_.VT _.RC:$src2))))]>,
EVEX_4V, EVEX_K, Sched<[sched]>;
def rmk : AVX512BI<opc, MRMSrcMem,
(outs _.KRC:$dst), (ins _.KRCWM:$mask, _.RC:$src1, _.MemOp:$src2),
!strconcat(OpcodeStr, "\t{$src2, $src1, $dst {${mask}}|",
"$dst {${mask}}, $src1, $src2}"),
[(set _.KRC:$dst, (and _.KRCWM:$mask,
(OpNode_su (_.VT _.RC:$src1),
(_.VT (_.LdFrag addr:$src2)))))]>,
EVEX_4V, EVEX_K, Sched<[sched.Folded, sched.ReadAfterFold]>;
}
multiclass avx512_icmp_packed_rmb<bits<8> opc, string OpcodeStr, PatFrag OpNode,
PatFrag OpNode_su,
X86FoldableSchedWrite sched, X86VectorVTInfo _,
bit IsCommutable> :
avx512_icmp_packed<opc, OpcodeStr, OpNode, OpNode_su, sched, _, IsCommutable> {
def rmb : AVX512BI<opc, MRMSrcMem,
(outs _.KRC:$dst), (ins _.RC:$src1, _.ScalarMemOp:$src2),
!strconcat(OpcodeStr, "\t{${src2}", _.BroadcastStr, ", $src1, $dst",
"|$dst, $src1, ${src2}", _.BroadcastStr, "}"),
[(set _.KRC:$dst, (OpNode (_.VT _.RC:$src1),
(X86VBroadcast (_.ScalarLdFrag addr:$src2))))]>,
EVEX_4V, EVEX_B, Sched<[sched.Folded, sched.ReadAfterFold]>;
def rmbk : AVX512BI<opc, MRMSrcMem,
(outs _.KRC:$dst), (ins _.KRCWM:$mask, _.RC:$src1,
_.ScalarMemOp:$src2),