VTName)), VTName));
PatFrag ScalarLdFrag = !cast<PatFrag>("load" # EltVT);
+ // Load patterns used for memory operands. We only have this defined in
+ // case of i64 element types for sub-512 integer vectors. For now, keep
+ // MemOpFrag undefined in these cases.
+ PatFrag MemOpFrag =
+ !if (!eq (TypeVariantName, "f"), !cast<PatFrag>("memop" # VTName),
+ !if (!eq (EltTypeName, "i64"), !cast<PatFrag>("memop" # VTName),
+ !if (!eq (VTName, "v16i32"), !cast<PatFrag>("memop" # VTName), ?)));
+
// The corresponding float type, e.g. v16f32 for v16i32
// Note: For EltSize < 32, FloatVT is illegal and TableGen
// fails to compile, so we choose FloatVT = VT
// The string to specify embedded broadcast in assembly.
string BroadcastStr = "{1to" # 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));
}
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">;
// Common base class of AVX512_masking and AVX512_masking_3src.
-multiclass AVX512_masking_common<bits<8> O, Format F, dag Outs, dag Ins,
- dag MaskingIns, dag ZeroMaskingIns,
+multiclass AVX512_masking_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, ValueType OpVT,
- RegisterClass RC, RegisterClass KRC,
+ dag RHS, dag MaskingRHS,
string MaskingConstraint = ""> {
def NAME: AVX512<O, F, Outs, Ins,
- OpcodeStr#" \t{"#AttSrcAsm#", $dst|"#
- "$dst, "#IntelSrcAsm#"}",
- [(set RC:$dst, RHS)]>;
+ OpcodeStr#"\t{"#AttSrcAsm#", $dst|"#
+ "$dst, "#IntelSrcAsm#"}",
+ [(set _.RC:$dst, RHS)]>;
// Prefer over VMOV*rrk Pat<>
let AddedComplexity = 20 in
def NAME#k: AVX512<O, F, Outs, MaskingIns,
- OpcodeStr#" \t{"#AttSrcAsm#", $dst {${mask}}|"#
- "$dst {${mask}}, "#IntelSrcAsm#"}",
- [(set RC:$dst, MaskingRHS)]>,
+ OpcodeStr#"\t{"#AttSrcAsm#", $dst {${mask}}|"#
+ "$dst {${mask}}, "#IntelSrcAsm#"}",
+ [(set _.RC:$dst, MaskingRHS)]>,
EVEX_K {
// In case of the 3src subclass this is overridden with a let.
string Constraints = MaskingConstraint;
}
let AddedComplexity = 30 in // Prefer over VMOV*rrkz Pat<>
def NAME#kz: AVX512<O, F, Outs, ZeroMaskingIns,
- OpcodeStr#" \t{"#AttSrcAsm#", $dst {${mask}} {z}|"#
- "$dst {${mask}} {z}, "#IntelSrcAsm#"}",
- [(set RC:$dst,
- (vselect KRC:$mask, RHS,
- (OpVT (bitconvert
+ OpcodeStr#"\t{"#AttSrcAsm#", $dst {${mask}} {z}|"#
+ "$dst {${mask}} {z}, "#IntelSrcAsm#"}",
+ [(set _.RC:$dst,
+ (vselect _.KRCWM:$mask, RHS,
+ (_.VT (bitconvert
(v16i32 immAllZerosV)))))]>,
EVEX_KZ;
}
// This multiclass generates the unconditional/non-masking, the masking and
// the zero-masking variant of the instruction. In the masking case, the
// perserved vector elements come from a new dummy input operand tied to $dst.
-multiclass AVX512_masking<bits<8> O, Format F, dag Outs, dag Ins,
- string OpcodeStr,
+multiclass AVX512_masking<bits<8> O, Format F, X86VectorVTInfo _,
+ dag Outs, dag Ins, string OpcodeStr,
string AttSrcAsm, string IntelSrcAsm,
- dag RHS, ValueType OpVT, RegisterClass RC,
- RegisterClass KRC> :
- AVX512_masking_common<O, F, Outs,
- Ins,
- !con((ins RC:$src0, KRC:$mask), Ins),
- !con((ins KRC:$mask), Ins),
+ dag RHS> :
+ AVX512_masking_common<O, F, _, Outs, Ins,
+ !con((ins _.RC:$src0, _.KRCWM:$mask), Ins),
+ !con((ins _.KRCWM:$mask), Ins),
OpcodeStr, AttSrcAsm, IntelSrcAsm, RHS,
- (vselect KRC:$mask, RHS, RC:$src0), OpVT, RC, KRC,
+ (vselect _.KRCWM:$mask, RHS, _.RC:$src0),
"$src0 = $dst">;
// Similar to AVX512_masking but in this case one of the source operands
// ($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
// $src1.
-multiclass AVX512_masking_3src<bits<8> O, Format F, dag Outs, dag NonTiedIns,
- string OpcodeStr,
+multiclass AVX512_masking_3src<bits<8> O, Format F, X86VectorVTInfo _,
+ dag Outs, dag NonTiedIns, string OpcodeStr,
string AttSrcAsm, string IntelSrcAsm,
- dag RHS, ValueType OpVT,
- RegisterClass RC, RegisterClass KRC> :
- AVX512_masking_common<O, F, Outs,
- !con((ins RC:$src1), NonTiedIns),
- !con((ins RC:$src1), !con((ins KRC:$mask),
- NonTiedIns)),
- !con((ins RC:$src1), !con((ins KRC:$mask),
- NonTiedIns)),
+ dag RHS> :
+ AVX512_masking_common<O, F, _, Outs,
+ !con((ins _.RC:$src1), NonTiedIns),
+ !con((ins _.RC:$src1, _.KRCWM:$mask), NonTiedIns),
+ !con((ins _.RC:$src1, _.KRCWM:$mask), NonTiedIns),
OpcodeStr, AttSrcAsm, IntelSrcAsm, RHS,
- (vselect KRC:$mask, RHS, RC:$src1), OpVT, RC, KRC>;
+ (vselect _.KRCWM:$mask, RHS, _.RC:$src1)>;
// Bitcasts between 512-bit vector types. Return the original type since
// no instruction is needed for the conversion
//===----------------------------------------------------------------------===//
// AVX-512 - VECTOR INSERT
//
-// -- 32x8 form --
-let hasSideEffects = 0, ExeDomain = SSEPackedSingle in {
-def VINSERTF32x4rr : AVX512AIi8<0x18, MRMSrcReg, (outs VR512:$dst),
- (ins VR512:$src1, VR128X:$src2, i8imm:$src3),
- "vinsertf32x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
- []>, EVEX_4V, EVEX_V512;
-let mayLoad = 1 in
-def VINSERTF32x4rm : AVX512AIi8<0x18, MRMSrcMem, (outs VR512:$dst),
- (ins VR512:$src1, f128mem:$src2, i8imm:$src3),
- "vinsertf32x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
- []>, EVEX_4V, EVEX_V512, EVEX_CD8<32, CD8VT4>;
-}
-
-// -- 64x4 fp form --
-let hasSideEffects = 0, ExeDomain = SSEPackedDouble in {
-def VINSERTF64x4rr : AVX512AIi8<0x1a, MRMSrcReg, (outs VR512:$dst),
- (ins VR512:$src1, VR256X:$src2, i8imm:$src3),
- "vinsertf64x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
- []>, EVEX_4V, EVEX_V512, VEX_W;
-let mayLoad = 1 in
-def VINSERTF64x4rm : AVX512AIi8<0x1a, MRMSrcMem, (outs VR512:$dst),
- (ins VR512:$src1, i256mem:$src2, i8imm:$src3),
- "vinsertf64x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
- []>, EVEX_4V, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT4>;
-}
-// -- 32x4 integer form --
-let hasSideEffects = 0 in {
-def VINSERTI32x4rr : AVX512AIi8<0x38, MRMSrcReg, (outs VR512:$dst),
- (ins VR512:$src1, VR128X:$src2, i8imm:$src3),
- "vinserti32x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
- []>, EVEX_4V, EVEX_V512;
-let mayLoad = 1 in
-def VINSERTI32x4rm : AVX512AIi8<0x38, MRMSrcMem, (outs VR512:$dst),
- (ins VR512:$src1, i128mem:$src2, i8imm:$src3),
- "vinserti32x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
- []>, EVEX_4V, EVEX_V512, EVEX_CD8<32, CD8VT4>;
-}
-let hasSideEffects = 0 in {
-// -- 64x4 form --
-def VINSERTI64x4rr : AVX512AIi8<0x3a, MRMSrcReg, (outs VR512:$dst),
- (ins VR512:$src1, VR256X:$src2, i8imm:$src3),
- "vinserti64x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
- []>, EVEX_4V, EVEX_V512, VEX_W;
-let mayLoad = 1 in
-def VINSERTI64x4rm : AVX512AIi8<0x3a, MRMSrcMem, (outs VR512:$dst),
- (ins VR512:$src1, i256mem:$src2, i8imm:$src3),
- "vinserti64x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
- []>, EVEX_4V, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT4>;
-}
-
-def : Pat<(vinsert128_insert:$ins (v16f32 VR512:$src1), (v4f32 VR128X:$src2),
- (iPTR imm)), (VINSERTF32x4rr VR512:$src1, VR128X:$src2,
- (INSERT_get_vinsert128_imm VR512:$ins))>;
-def : Pat<(vinsert128_insert:$ins (v8f64 VR512:$src1), (v2f64 VR128X:$src2),
- (iPTR imm)), (VINSERTF32x4rr VR512:$src1, VR128X:$src2,
- (INSERT_get_vinsert128_imm VR512:$ins))>;
-def : Pat<(vinsert128_insert:$ins (v8i64 VR512:$src1), (v2i64 VR128X:$src2),
- (iPTR imm)), (VINSERTI32x4rr VR512:$src1, VR128X:$src2,
- (INSERT_get_vinsert128_imm VR512:$ins))>;
-def : Pat<(vinsert128_insert:$ins (v16i32 VR512:$src1), (v4i32 VR128X:$src2),
- (iPTR imm)), (VINSERTI32x4rr VR512:$src1, VR128X:$src2,
- (INSERT_get_vinsert128_imm VR512:$ins))>;
-
-def : Pat<(vinsert128_insert:$ins (v16f32 VR512:$src1), (loadv4f32 addr:$src2),
- (iPTR imm)), (VINSERTF32x4rm VR512:$src1, addr:$src2,
- (INSERT_get_vinsert128_imm VR512:$ins))>;
-def : Pat<(vinsert128_insert:$ins (v16i32 VR512:$src1),
- (bc_v4i32 (loadv2i64 addr:$src2)),
- (iPTR imm)), (VINSERTI32x4rm VR512:$src1, addr:$src2,
- (INSERT_get_vinsert128_imm VR512:$ins))>;
-def : Pat<(vinsert128_insert:$ins (v8f64 VR512:$src1), (loadv2f64 addr:$src2),
- (iPTR imm)), (VINSERTF32x4rm VR512:$src1, addr:$src2,
- (INSERT_get_vinsert128_imm VR512:$ins))>;
-def : Pat<(vinsert128_insert:$ins (v8i64 VR512:$src1), (loadv2i64 addr:$src2),
- (iPTR imm)), (VINSERTI32x4rm VR512:$src1, addr:$src2,
- (INSERT_get_vinsert128_imm VR512:$ins))>;
-
-def : Pat<(vinsert256_insert:$ins (v16f32 VR512:$src1), (v8f32 VR256X:$src2),
- (iPTR imm)), (VINSERTF64x4rr VR512:$src1, VR256X:$src2,
- (INSERT_get_vinsert256_imm VR512:$ins))>;
-def : Pat<(vinsert256_insert:$ins (v8f64 VR512:$src1), (v4f64 VR256X:$src2),
- (iPTR imm)), (VINSERTF64x4rr VR512:$src1, VR256X:$src2,
- (INSERT_get_vinsert256_imm VR512:$ins))>;
-def : Pat<(vinsert128_insert:$ins (v8i64 VR512:$src1), (v4i64 VR256X:$src2),
- (iPTR imm)), (VINSERTI64x4rr VR512:$src1, VR256X:$src2,
- (INSERT_get_vinsert256_imm VR512:$ins))>;
-def : Pat<(vinsert128_insert:$ins (v16i32 VR512:$src1), (v8i32 VR256X:$src2),
- (iPTR imm)), (VINSERTI64x4rr VR512:$src1, VR256X:$src2,
- (INSERT_get_vinsert256_imm VR512:$ins))>;
-
-def : Pat<(vinsert256_insert:$ins (v16f32 VR512:$src1), (loadv8f32 addr:$src2),
- (iPTR imm)), (VINSERTF64x4rm VR512:$src1, addr:$src2,
- (INSERT_get_vinsert256_imm VR512:$ins))>;
-def : Pat<(vinsert256_insert:$ins (v8f64 VR512:$src1), (loadv4f64 addr:$src2),
- (iPTR imm)), (VINSERTF64x4rm VR512:$src1, addr:$src2,
- (INSERT_get_vinsert256_imm VR512:$ins))>;
-def : Pat<(vinsert256_insert:$ins (v8i64 VR512:$src1), (loadv4i64 addr:$src2),
- (iPTR imm)), (VINSERTI64x4rm VR512:$src1, addr:$src2,
- (INSERT_get_vinsert256_imm VR512:$ins))>;
-def : Pat<(vinsert256_insert:$ins (v16i32 VR512:$src1),
- (bc_v8i32 (loadv4i64 addr:$src2)),
- (iPTR imm)), (VINSERTI64x4rm VR512:$src1, addr:$src2,
- (INSERT_get_vinsert256_imm VR512:$ins))>;
+multiclass vinsert_for_size<int Opcode,
+ X86VectorVTInfo From, X86VectorVTInfo To,
+ X86VectorVTInfo AltFrom, X86VectorVTInfo AltTo,
+ PatFrag vinsert_insert,
+ SDNodeXForm INSERT_get_vinsert_imm> {
+ let hasSideEffects = 0, ExeDomain = To.ExeDomain in {
+ def rr : AVX512AIi8<Opcode, MRMSrcReg, (outs VR512:$dst),
+ (ins VR512:$src1, From.RC:$src2, i8imm:$src3),
+ "vinsert" # From.EltTypeName # "x4\t{$src3, $src2, $src1, $dst|"
+ "$dst, $src1, $src2, $src3}",
+ [(set To.RC:$dst, (vinsert_insert:$src3 (To.VT VR512:$src1),
+ (From.VT From.RC:$src2),
+ (iPTR imm)))]>,
+ EVEX_4V, EVEX_V512;
+
+ let mayLoad = 1 in
+ def rm : AVX512AIi8<Opcode, MRMSrcMem, (outs VR512:$dst),
+ (ins VR512:$src1, From.MemOp:$src2, i8imm:$src3),
+ "vinsert" # From.EltTypeName # "x4\t{$src3, $src2, $src1, $dst|"
+ "$dst, $src1, $src2, $src3}",
+ []>, EVEX_4V, EVEX_V512, EVEX_CD8<From.EltSize, CD8VT4>;
+ }
+
+ // Codegen pattern with the alternative types, e.g. v2i64 -> v8i64 for
+ // vinserti32x4
+ def : Pat<(vinsert_insert:$ins
+ (AltTo.VT VR512:$src1), (AltFrom.VT From.RC:$src2), (iPTR imm)),
+ (AltTo.VT (!cast<Instruction>(NAME # From.EltSize # "x4rr")
+ VR512:$src1, From.RC:$src2,
+ (INSERT_get_vinsert_imm VR512:$ins)))>;
+}
+
+multiclass vinsert_for_type<ValueType EltVT32, int Opcode32,
+ ValueType EltVT64, int Opcode64> {
+ defm NAME # "32x4" : vinsert_for_size<Opcode32,
+ X86VectorVTInfo< 4, EltVT32, VR128X>,
+ X86VectorVTInfo<16, EltVT32, VR512>,
+ X86VectorVTInfo< 2, EltVT64, VR128X>,
+ X86VectorVTInfo< 8, EltVT64, VR512>,
+ vinsert128_insert,
+ INSERT_get_vinsert128_imm>;
+ defm NAME # "64x4" : vinsert_for_size<Opcode64,
+ X86VectorVTInfo< 4, EltVT64, VR256X>,
+ X86VectorVTInfo< 8, EltVT64, VR512>,
+ X86VectorVTInfo< 8, EltVT32, VR256>,
+ X86VectorVTInfo<16, EltVT32, VR512>,
+ vinsert256_insert,
+ INSERT_get_vinsert256_imm>, VEX_W;
+}
+
+defm VINSERTF : vinsert_for_type<f32, 0x18, f64, 0x1a>;
+defm VINSERTI : vinsert_for_type<i32, 0x38, i64, 0x3a>;
// vinsertps - insert f32 to XMM
def VINSERTPSzrr : AVX512AIi8<0x21, MRMSrcReg, (outs VR128X:$dst),
- (ins VR128X:$src1, VR128X:$src2, u32u8imm:$src3),
+ (ins VR128X:$src1, VR128X:$src2, i8imm:$src3),
"vinsertps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
[(set VR128X:$dst, (X86insertps VR128X:$src1, VR128X:$src2, imm:$src3))]>,
EVEX_4V;
def VINSERTPSzrm: AVX512AIi8<0x21, MRMSrcMem, (outs VR128X:$dst),
- (ins VR128X:$src1, f32mem:$src2, u32u8imm:$src3),
+ (ins VR128X:$src1, f32mem:$src2, i8imm:$src3),
"vinsertps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
[(set VR128X:$dst, (X86insertps VR128X:$src1,
(v4f32 (scalar_to_vector (loadf32 addr:$src2))),
//===----------------------------------------------------------------------===//
// AVX-512 VECTOR EXTRACT
//---
-let hasSideEffects = 0, ExeDomain = SSEPackedSingle in {
-// -- 32x4 form --
-def VEXTRACTF32x4rr : AVX512AIi8<0x19, MRMDestReg, (outs VR128X:$dst),
- (ins VR512:$src1, i8imm:$src2),
- "vextractf32x4\t{$src2, $src1, $dst|$dst, $src1, $src2}",
- []>, EVEX, EVEX_V512;
-def VEXTRACTF32x4mr : AVX512AIi8<0x19, MRMDestMem, (outs),
- (ins f128mem:$dst, VR512:$src1, i8imm:$src2),
- "vextractf32x4\t{$src2, $src1, $dst|$dst, $src1, $src2}",
- []>, EVEX, EVEX_V512, EVEX_CD8<32, CD8VT4>;
-
-// -- 64x4 form --
-def VEXTRACTF64x4rr : AVX512AIi8<0x1b, MRMDestReg, (outs VR256X:$dst),
- (ins VR512:$src1, i8imm:$src2),
- "vextractf64x4\t{$src2, $src1, $dst|$dst, $src1, $src2}",
- []>, EVEX, EVEX_V512, VEX_W;
-let mayStore = 1 in
-def VEXTRACTF64x4mr : AVX512AIi8<0x1b, MRMDestMem, (outs),
- (ins f256mem:$dst, VR512:$src1, i8imm:$src2),
- "vextractf64x4\t{$src2, $src1, $dst|$dst, $src1, $src2}",
- []>, EVEX, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT4>;
-}
-let hasSideEffects = 0 in {
-// -- 32x4 form --
-def VEXTRACTI32x4rr : AVX512AIi8<0x39, MRMDestReg, (outs VR128X:$dst),
- (ins VR512:$src1, i8imm:$src2),
- "vextracti32x4\t{$src2, $src1, $dst|$dst, $src1, $src2}",
- []>, EVEX, EVEX_V512;
-def VEXTRACTI32x4mr : AVX512AIi8<0x39, MRMDestMem, (outs),
- (ins i128mem:$dst, VR512:$src1, i8imm:$src2),
- "vextracti32x4\t{$src2, $src1, $dst|$dst, $src1, $src2}",
- []>, EVEX, EVEX_V512, EVEX_CD8<32, CD8VT4>;
-
-// -- 64x4 form --
-def VEXTRACTI64x4rr : AVX512AIi8<0x3b, MRMDestReg, (outs VR256X:$dst),
- (ins VR512:$src1, i8imm:$src2),
- "vextracti64x4\t{$src2, $src1, $dst|$dst, $src1, $src2}",
- []>, EVEX, EVEX_V512, VEX_W;
-let mayStore = 1 in
-def VEXTRACTI64x4mr : AVX512AIi8<0x3b, MRMDestMem, (outs),
- (ins i256mem:$dst, VR512:$src1, i8imm:$src2),
- "vextracti64x4\t{$src2, $src1, $dst|$dst, $src1, $src2}",
- []>, EVEX, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT4>;
-}
-
-def : Pat<(vextract128_extract:$ext (v16f32 VR512:$src1), (iPTR imm)),
- (v4f32 (VEXTRACTF32x4rr VR512:$src1,
- (EXTRACT_get_vextract128_imm VR128X:$ext)))>;
-
-def : Pat<(vextract128_extract:$ext VR512:$src1, (iPTR imm)),
- (v4i32 (VEXTRACTF32x4rr VR512:$src1,
- (EXTRACT_get_vextract128_imm VR128X:$ext)))>;
-
-def : Pat<(vextract128_extract:$ext (v8f64 VR512:$src1), (iPTR imm)),
- (v2f64 (VEXTRACTF32x4rr VR512:$src1,
- (EXTRACT_get_vextract128_imm VR128X:$ext)))>;
-
-def : Pat<(vextract128_extract:$ext (v8i64 VR512:$src1), (iPTR imm)),
- (v2i64 (VEXTRACTI32x4rr VR512:$src1,
- (EXTRACT_get_vextract128_imm VR128X:$ext)))>;
-
-
-def : Pat<(vextract256_extract:$ext (v16f32 VR512:$src1), (iPTR imm)),
- (v8f32 (VEXTRACTF64x4rr VR512:$src1,
- (EXTRACT_get_vextract256_imm VR256X:$ext)))>;
-
-def : Pat<(vextract256_extract:$ext (v16i32 VR512:$src1), (iPTR imm)),
- (v8i32 (VEXTRACTI64x4rr VR512:$src1,
- (EXTRACT_get_vextract256_imm VR256X:$ext)))>;
-
-def : Pat<(vextract256_extract:$ext (v8f64 VR512:$src1), (iPTR imm)),
- (v4f64 (VEXTRACTF64x4rr VR512:$src1,
- (EXTRACT_get_vextract256_imm VR256X:$ext)))>;
-
-def : Pat<(vextract256_extract:$ext (v8i64 VR512:$src1), (iPTR imm)),
- (v4i64 (VEXTRACTI64x4rr VR512:$src1,
- (EXTRACT_get_vextract256_imm VR256X:$ext)))>;
-
-// A 256-bit subvector extract from the first 512-bit vector position
-// is a subregister copy that needs no instruction.
-def : Pat<(v8i32 (extract_subvector (v16i32 VR512:$src), (iPTR 0))),
- (v8i32 (EXTRACT_SUBREG (v16i32 VR512:$src), sub_ymm))>;
-def : Pat<(v8f32 (extract_subvector (v16f32 VR512:$src), (iPTR 0))),
- (v8f32 (EXTRACT_SUBREG (v16f32 VR512:$src), sub_ymm))>;
-def : Pat<(v4i64 (extract_subvector (v8i64 VR512:$src), (iPTR 0))),
- (v4i64 (EXTRACT_SUBREG (v8i64 VR512:$src), sub_ymm))>;
-def : Pat<(v4f64 (extract_subvector (v8f64 VR512:$src), (iPTR 0))),
- (v4f64 (EXTRACT_SUBREG (v8f64 VR512:$src), sub_ymm))>;
-
-// zmm -> xmm
-def : Pat<(v4i32 (extract_subvector (v16i32 VR512:$src), (iPTR 0))),
- (v4i32 (EXTRACT_SUBREG (v16i32 VR512:$src), sub_xmm))>;
-def : Pat<(v2i64 (extract_subvector (v8i64 VR512:$src), (iPTR 0))),
- (v2i64 (EXTRACT_SUBREG (v8i64 VR512:$src), sub_xmm))>;
-def : Pat<(v2f64 (extract_subvector (v8f64 VR512:$src), (iPTR 0))),
- (v2f64 (EXTRACT_SUBREG (v8f64 VR512:$src), sub_xmm))>;
-def : Pat<(v4f32 (extract_subvector (v16f32 VR512:$src), (iPTR 0))),
- (v4f32 (EXTRACT_SUBREG (v16f32 VR512:$src), sub_xmm))>;
+multiclass vextract_for_size<int Opcode,
+ X86VectorVTInfo From, X86VectorVTInfo To,
+ X86VectorVTInfo AltFrom, X86VectorVTInfo AltTo,
+ PatFrag vextract_extract,
+ SDNodeXForm EXTRACT_get_vextract_imm> {
+ let hasSideEffects = 0, ExeDomain = To.ExeDomain in {
+ def rr : AVX512AIi8<Opcode, MRMDestReg, (outs To.RC:$dst),
+ (ins VR512:$src1, i8imm:$idx),
+ "vextract" # To.EltTypeName # "x4\t{$idx, $src1, $dst|"
+ "$dst, $src1, $idx}",
+ [(set To.RC:$dst, (vextract_extract:$idx (From.VT VR512:$src1),
+ (iPTR imm)))]>,
+ EVEX, EVEX_V512;
+ let mayStore = 1 in
+ def rm : AVX512AIi8<Opcode, MRMDestMem, (outs),
+ (ins To.MemOp:$dst, VR512:$src1, i8imm:$src2),
+ "vextract" # To.EltTypeName # "x4\t{$src2, $src1, $dst|"
+ "$dst, $src1, $src2}",
+ []>, EVEX, EVEX_V512, EVEX_CD8<To.EltSize, CD8VT4>;
+ }
+ // Codegen pattern with the alternative types, e.g. v8i64 -> v2i64 for
+ // vextracti32x4
+ def : Pat<(vextract_extract:$ext (AltFrom.VT VR512:$src1), (iPTR imm)),
+ (AltTo.VT (!cast<Instruction>(NAME # To.EltSize # "x4rr")
+ VR512:$src1,
+ (EXTRACT_get_vextract_imm To.RC:$ext)))>;
+
+ // A 128/256-bit subvector extract from the first 512-bit vector position is
+ // a subregister copy that needs no instruction.
+ def : Pat<(To.VT (extract_subvector (From.VT VR512:$src), (iPTR 0))),
+ (To.VT
+ (EXTRACT_SUBREG (From.VT VR512:$src), To.SubRegIdx))>;
+
+ // And for the alternative types.
+ def : Pat<(AltTo.VT (extract_subvector (AltFrom.VT VR512:$src), (iPTR 0))),
+ (AltTo.VT
+ (EXTRACT_SUBREG (AltFrom.VT VR512:$src), AltTo.SubRegIdx))>;
+}
+
+multiclass vextract_for_type<ValueType EltVT32, int Opcode32,
+ ValueType EltVT64, int Opcode64> {
+ defm NAME # "32x4" : vextract_for_size<Opcode32,
+ X86VectorVTInfo<16, EltVT32, VR512>,
+ X86VectorVTInfo< 4, EltVT32, VR128X>,
+ X86VectorVTInfo< 8, EltVT64, VR512>,
+ X86VectorVTInfo< 2, EltVT64, VR128X>,
+ vextract128_extract,
+ EXTRACT_get_vextract128_imm>;
+ defm NAME # "64x4" : vextract_for_size<Opcode64,
+ X86VectorVTInfo< 8, EltVT64, VR512>,
+ X86VectorVTInfo< 4, EltVT64, VR256X>,
+ X86VectorVTInfo<16, EltVT32, VR512>,
+ X86VectorVTInfo< 8, EltVT32, VR256>,
+ vextract256_extract,
+ EXTRACT_get_vextract256_imm>, VEX_W;
+}
+
+defm VEXTRACTF : vextract_for_type<f32, 0x19, f64, 0x1b>;
+defm VEXTRACTI : vextract_for_type<i32, 0x39, i64, 0x3b>;
// A 128-bit subvector insert to the first 512-bit vector position
// is a subregister copy that needs no instruction.
// vextractps - extract 32 bits from XMM
def VEXTRACTPSzrr : AVX512AIi8<0x17, MRMDestReg, (outs GR32:$dst),
- (ins VR128X:$src1, u32u8imm:$src2),
+ (ins VR128X:$src1, i32i8imm:$src2),
"vextractps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
[(set GR32:$dst, (extractelt (bc_v4i32 (v4f32 VR128X:$src1)), imm:$src2))]>,
EVEX;
def VEXTRACTPSzmr : AVX512AIi8<0x17, MRMDestMem, (outs),
- (ins f32mem:$dst, VR128X:$src1, u32u8imm:$src2),
+ (ins f32mem:$dst, VR128X:$src1, i32i8imm:$src2),
"vextractps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
[(store (extractelt (bc_v4i32 (v4f32 VR128X:$src1)), imm:$src2),
addr:$dst)]>, EVEX, EVEX_CD8<32, CD8VT1>;
}
multiclass avx512_icmp_packed_rmb<bits<8> opc, string OpcodeStr, SDNode OpNode,
- X86VectorVTInfo _> {
+ X86VectorVTInfo _> :
+ avx512_icmp_packed<opc, OpcodeStr, OpNode, _> {
let mayLoad = 1 in {
def rmb : AVX512BI<opc, MRMSrcMem,
(outs _.KRC:$dst), (ins _.RC:$src1, _.ScalarMemOp:$src2),
avx512vl_i16_info, HasBWI>,
EVEX_CD8<16, CD8VF>;
-defm VPCMPEQD : avx512_icmp_packed_vl<0x76, "vpcmpeqd", X86pcmpeqm,
- avx512vl_i32_info, HasAVX512>,
- avx512_icmp_packed_rmb_vl<0x76, "vpcmpeqd", X86pcmpeqm,
+defm VPCMPEQD : avx512_icmp_packed_rmb_vl<0x76, "vpcmpeqd", X86pcmpeqm,
avx512vl_i32_info, HasAVX512>,
EVEX_CD8<32, CD8VF>;
-defm VPCMPEQQ : avx512_icmp_packed_vl<0x29, "vpcmpeqq", X86pcmpeqm,
- avx512vl_i64_info, HasAVX512>,
- avx512_icmp_packed_rmb_vl<0x29, "vpcmpeqq", X86pcmpeqm,
+defm VPCMPEQQ : avx512_icmp_packed_rmb_vl<0x29, "vpcmpeqq", X86pcmpeqm,
avx512vl_i64_info, HasAVX512>,
T8PD, VEX_W, EVEX_CD8<64, CD8VF>;
avx512vl_i16_info, HasBWI>,
EVEX_CD8<16, CD8VF>;
-defm VPCMPGTD : avx512_icmp_packed_vl<0x66, "vpcmpgtd", X86pcmpgtm,
- avx512vl_i32_info, HasAVX512>,
- avx512_icmp_packed_rmb_vl<0x66, "vpcmpgtd", X86pcmpgtm,
+defm VPCMPGTD : avx512_icmp_packed_rmb_vl<0x66, "vpcmpgtd", X86pcmpgtm,
avx512vl_i32_info, HasAVX512>,
EVEX_CD8<32, CD8VF>;
-defm VPCMPGTQ : avx512_icmp_packed_vl<0x37, "vpcmpgtq", X86pcmpgtm,
- avx512vl_i64_info, HasAVX512>,
- avx512_icmp_packed_rmb_vl<0x37, "vpcmpgtq", X86pcmpgtm,
+defm VPCMPGTQ : avx512_icmp_packed_rmb_vl<0x37, "vpcmpgtq", X86pcmpgtm,
avx512vl_i64_info, HasAVX512>,
T8PD, VEX_W, EVEX_CD8<64, CD8VF>;
}
multiclass avx512_icmp_cc_rmb<bits<8> opc, string Suffix, SDNode OpNode,
- X86VectorVTInfo _> {
+ X86VectorVTInfo _> :
+ avx512_icmp_cc<opc, Suffix, OpNode, _> {
let mayLoad = 1 in {
def rmib : AVX512AIi8<opc, MRMSrcMem,
(outs _.KRC:$dst), (ins _.RC:$src1, _.ScalarMemOp:$src2,
defm VPCMPUW : avx512_icmp_cc_vl<0x3E, "uw", X86cmpmu, avx512vl_i16_info,
HasBWI>, VEX_W, EVEX_CD8<16, CD8VF>;
-defm VPCMPD : avx512_icmp_cc_vl<0x1F, "d", X86cmpm, avx512vl_i32_info,
- HasAVX512>,
- avx512_icmp_cc_rmb_vl<0x1F, "d", X86cmpm, avx512vl_i32_info,
+defm VPCMPD : avx512_icmp_cc_rmb_vl<0x1F, "d", X86cmpm, avx512vl_i32_info,
HasAVX512>, EVEX_CD8<32, CD8VF>;
-defm VPCMPUD : avx512_icmp_cc_vl<0x1E, "ud", X86cmpmu, avx512vl_i32_info,
- HasAVX512>,
- avx512_icmp_cc_rmb_vl<0x1E, "ud", X86cmpmu, avx512vl_i32_info,
+defm VPCMPUD : avx512_icmp_cc_rmb_vl<0x1E, "ud", X86cmpmu, avx512vl_i32_info,
HasAVX512>, EVEX_CD8<32, CD8VF>;
-defm VPCMPQ : avx512_icmp_cc_vl<0x1F, "q", X86cmpm, avx512vl_i64_info,
- HasAVX512>,
- avx512_icmp_cc_rmb_vl<0x1F, "q", X86cmpm, avx512vl_i64_info,
+defm VPCMPQ : avx512_icmp_cc_rmb_vl<0x1F, "q", X86cmpm, avx512vl_i64_info,
HasAVX512>, VEX_W, EVEX_CD8<64, CD8VF>;
-defm VPCMPUQ : avx512_icmp_cc_vl<0x1E, "uq", X86cmpmu, avx512vl_i64_info,
- HasAVX512>,
- avx512_icmp_cc_rmb_vl<0x1E, "uq", X86cmpmu, avx512vl_i64_info,
+defm VPCMPUQ : avx512_icmp_cc_rmb_vl<0x1E, "uq", X86cmpmu, avx512vl_i64_info,
HasAVX512>, VEX_W, EVEX_CD8<64, CD8VF>;
// avx512_cmp_packed - compare packed instructions
def : Pat<(v8i1 (extract_subvector (v16i1 VK16:$src), (iPTR 8))),
(v8i1 (COPY_TO_REGCLASS (KSHIFTRWri VK16:$src, (i8 8)), VK8))>;
+let Predicates = [HasVLX] in {
+ def : Pat<(v8i1 (insert_subvector undef, (v4i1 VK4:$src), (iPTR 0))),
+ (v8i1 (COPY_TO_REGCLASS VK4:$src, VK8))>;
+ def : Pat<(v8i1 (insert_subvector undef, (v2i1 VK2:$src), (iPTR 0))),
+ (v8i1 (COPY_TO_REGCLASS VK2:$src, VK8))>;
+ def : Pat<(v4i1 (extract_subvector (v8i1 VK8:$src), (iPTR 0))),
+ (v4i1 (COPY_TO_REGCLASS VK8:$src, VK4))>;
+ def : Pat<(v2i1 (extract_subvector (v8i1 VK8:$src), (iPTR 0))),
+ (v2i1 (COPY_TO_REGCLASS VK8:$src, VK2))>;
+}
+
def : Pat<(v8i1 (X86vshli VK8:$src, (i8 imm:$imm))),
(v8i1 (COPY_TO_REGCLASS (KSHIFTLWri (COPY_TO_REGCLASS VK8:$src, VK16), (I8Imm $imm)), VK8))>;
i512mem, v16i32>, PD, EVEX_V512, EVEX_CD8<32, CD8VF>;
let ExeDomain = SSEPackedSingle in
-defm VPERMILPSZ : avx512_pshuf_imm<0x04, "vpermilps", VR512, X86VPermilp,
+defm VPERMILPSZ : avx512_pshuf_imm<0x04, "vpermilps", VR512, X86VPermilpi,
memopv16f32, i512mem, v16f32>, TAPD, EVEX_V512,
EVEX_CD8<32, CD8VF>;
let ExeDomain = SSEPackedDouble in
-defm VPERMILPDZ : avx512_pshuf_imm<0x05, "vpermilpd", VR512, X86VPermilp,
+defm VPERMILPDZ : avx512_pshuf_imm<0x05, "vpermilpd", VR512, X86VPermilpi,
memopv8f64, i512mem, v8f64>, TAPD, EVEX_V512,
VEX_W, EVEX_CD8<32, CD8VF>;
-def : Pat<(v16i32 (X86VPermilp VR512:$src1, (i8 imm:$imm))),
+def : Pat<(v16i32 (X86VPermilpi VR512:$src1, (i8 imm:$imm))),
(VPERMILPSZri VR512:$src1, imm:$imm)>;
-def : Pat<(v8i64 (X86VPermilp VR512:$src1, (i8 imm:$imm))),
+def : Pat<(v8i64 (X86VPermilpi VR512:$src1, (i8 imm:$imm))),
(VPERMILPDZri VR512:$src1, imm:$imm)>;
//===----------------------------------------------------------------------===//
// FMA - Fused Multiply Operations
//
let Constraints = "$src1 = $dst" in {
-multiclass avx512_fma3p_rm<bits<8> opc, string OpcodeStr,
- RegisterClass RC, X86MemOperand x86memop,
- PatFrag mem_frag, X86MemOperand x86scalar_mop, PatFrag scalar_mfrag,
- string BrdcstStr, SDNode OpNode, ValueType OpVT,
- RegisterClass KRC> {
- defm r: AVX512_masking_3src<opc, MRMSrcReg, (outs RC:$dst),
- (ins RC:$src2, RC:$src3),
+multiclass avx512_fma3p_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
+ X86VectorVTInfo _> {
+ defm r: AVX512_masking_3src<opc, MRMSrcReg, _, (outs _.RC:$dst),
+ (ins _.RC:$src2, _.RC:$src3),
OpcodeStr, "$src3, $src2", "$src2, $src3",
- (OpVT (OpNode RC:$src1, RC:$src2, RC:$src3)), OpVT, RC, KRC>,
+ (_.VT (OpNode _.RC:$src1, _.RC:$src2, _.RC:$src3))>,
AVX512FMA3Base;
let mayLoad = 1 in
- def m: AVX512FMA3<opc, MRMSrcMem, (outs RC:$dst),
- (ins RC:$src1, RC:$src2, x86memop:$src3),
+ def m: AVX512FMA3<opc, MRMSrcMem, (outs _.RC:$dst),
+ (ins _.RC:$src1, _.RC:$src2, _.MemOp:$src3),
!strconcat(OpcodeStr, " \t{$src3, $src2, $dst|$dst, $src2, $src3}"),
- [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2,
- (mem_frag addr:$src3))))]>;
- def mb: AVX512FMA3<opc, MRMSrcMem, (outs RC:$dst),
- (ins RC:$src1, RC:$src2, x86scalar_mop:$src3),
- !strconcat(OpcodeStr, " \t{${src3}", BrdcstStr,
- ", $src2, $dst|$dst, $src2, ${src3}", BrdcstStr, "}"),
- [(set RC:$dst, (OpNode RC:$src1, RC:$src2,
- (OpVT (X86VBroadcast (scalar_mfrag addr:$src3)))))]>, EVEX_B;
+ [(set _.RC:$dst, (_.VT (OpNode _.RC:$src1, _.RC:$src2,
+ (_.MemOpFrag addr:$src3))))]>;
+ def mb: AVX512FMA3<opc, MRMSrcMem, (outs _.RC:$dst),
+ (ins _.RC:$src1, _.RC:$src2, _.ScalarMemOp:$src3),
+ !strconcat(OpcodeStr, " \t{${src3}", _.BroadcastStr,
+ ", $src2, $dst|$dst, $src2, ${src3}", _.BroadcastStr, "}"),
+ [(set _.RC:$dst, (OpNode _.RC:$src1, _.RC:$src2,
+ (_.VT (X86VBroadcast (_.ScalarLdFrag addr:$src3)))))]>, EVEX_B;
}
} // Constraints = "$src1 = $dst"
let ExeDomain = SSEPackedSingle in {
- defm VFMADD213PSZ : avx512_fma3p_rm<0xA8, "vfmadd213ps", VR512, f512mem,
- memopv16f32, f32mem, loadf32, "{1to16}",
- X86Fmadd, v16f32, VK16WM>, EVEX_V512,
- EVEX_CD8<32, CD8VF>;
- defm VFMSUB213PSZ : avx512_fma3p_rm<0xAA, "vfmsub213ps", VR512, f512mem,
- memopv16f32, f32mem, loadf32, "{1to16}",
- X86Fmsub, v16f32, VK16WM>, EVEX_V512,
- EVEX_CD8<32, CD8VF>;
- defm VFMADDSUB213PSZ : avx512_fma3p_rm<0xA6, "vfmaddsub213ps", VR512, f512mem,
- memopv16f32, f32mem, loadf32, "{1to16}",
- X86Fmaddsub, v16f32, VK16WM>,
- EVEX_V512, EVEX_CD8<32, CD8VF>;
- defm VFMSUBADD213PSZ : avx512_fma3p_rm<0xA7, "vfmsubadd213ps", VR512, f512mem,
- memopv16f32, f32mem, loadf32, "{1to16}",
- X86Fmsubadd, v16f32, VK16WM>,
- EVEX_V512, EVEX_CD8<32, CD8VF>;
- defm VFNMADD213PSZ : avx512_fma3p_rm<0xAC, "vfnmadd213ps", VR512, f512mem,
- memopv16f32, f32mem, loadf32, "{1to16}",
- X86Fnmadd, v16f32, VK16WM>, EVEX_V512,
- EVEX_CD8<32, CD8VF>;
- defm VFNMSUB213PSZ : avx512_fma3p_rm<0xAE, "vfnmsub213ps", VR512, f512mem,
- memopv16f32, f32mem, loadf32, "{1to16}",
- X86Fnmsub, v16f32, VK16WM>, EVEX_V512,
- EVEX_CD8<32, CD8VF>;
+ defm VFMADD213PSZ : avx512_fma3p_rm<0xA8, "vfmadd213ps", X86Fmadd,
+ v16f32_info>,
+ EVEX_V512, EVEX_CD8<32, CD8VF>;
+ defm VFMSUB213PSZ : avx512_fma3p_rm<0xAA, "vfmsub213ps", X86Fmsub,
+ v16f32_info>,
+ EVEX_V512, EVEX_CD8<32, CD8VF>;
+ defm VFMADDSUB213PSZ : avx512_fma3p_rm<0xA6, "vfmaddsub213ps", X86Fmaddsub,
+ v16f32_info>,
+ EVEX_V512, EVEX_CD8<32, CD8VF>;
+ defm VFMSUBADD213PSZ : avx512_fma3p_rm<0xA7, "vfmsubadd213ps", X86Fmsubadd,
+ v16f32_info>,
+ EVEX_V512, EVEX_CD8<32, CD8VF>;
+ defm VFNMADD213PSZ : avx512_fma3p_rm<0xAC, "vfnmadd213ps", X86Fnmadd,
+ v16f32_info>,
+ EVEX_V512, EVEX_CD8<32, CD8VF>;
+ defm VFNMSUB213PSZ : avx512_fma3p_rm<0xAE, "vfnmsub213ps", X86Fnmsub,
+ v16f32_info>,
+ EVEX_V512, EVEX_CD8<32, CD8VF>;
}
let ExeDomain = SSEPackedDouble in {
- defm VFMADD213PDZ : avx512_fma3p_rm<0xA8, "vfmadd213pd", VR512, f512mem,
- memopv8f64, f64mem, loadf64, "{1to8}",
- X86Fmadd, v8f64, VK8WM>, EVEX_V512,
- VEX_W, EVEX_CD8<64, CD8VF>;
- defm VFMSUB213PDZ : avx512_fma3p_rm<0xAA, "vfmsub213pd", VR512, f512mem,
- memopv8f64, f64mem, loadf64, "{1to8}",
- X86Fmsub, v8f64, VK8WM>, EVEX_V512, VEX_W,
- EVEX_CD8<64, CD8VF>;
- defm VFMADDSUB213PDZ : avx512_fma3p_rm<0xA6, "vfmaddsub213pd", VR512, f512mem,
- memopv8f64, f64mem, loadf64, "{1to8}",
- X86Fmaddsub, v8f64, VK8WM>,
+ defm VFMADD213PDZ : avx512_fma3p_rm<0xA8, "vfmadd213pd", X86Fmadd,
+ v8f64_info>,
+ EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+ defm VFMSUB213PDZ : avx512_fma3p_rm<0xAA, "vfmsub213pd", X86Fmsub,
+ v8f64_info>,
EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
- defm VFMSUBADD213PDZ : avx512_fma3p_rm<0xA7, "vfmsubadd213pd", VR512, f512mem,
- memopv8f64, f64mem, loadf64, "{1to8}",
- X86Fmsubadd, v8f64, VK8WM>,
+ defm VFMADDSUB213PDZ : avx512_fma3p_rm<0xA6, "vfmaddsub213pd", X86Fmaddsub,
+ v8f64_info>,
+ EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+ defm VFMSUBADD213PDZ : avx512_fma3p_rm<0xA7, "vfmsubadd213pd", X86Fmsubadd,
+ v8f64_info>,
+ EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+ defm VFNMADD213PDZ : avx512_fma3p_rm<0xAC, "vfnmadd213pd", X86Fnmadd,
+ v8f64_info>,
+ EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+ defm VFNMSUB213PDZ : avx512_fma3p_rm<0xAE, "vfnmsub213pd", X86Fnmsub,
+ v8f64_info>,
EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
- defm VFNMADD213PDZ : avx512_fma3p_rm<0xAC, "vfnmadd213pd", VR512, f512mem,
- memopv8f64, f64mem, loadf64, "{1to8}",
- X86Fnmadd, v8f64, VK8WM>, EVEX_V512, VEX_W,
- EVEX_CD8<64, CD8VF>;
- defm VFNMSUB213PDZ : avx512_fma3p_rm<0xAE, "vfnmsub213pd", VR512, f512mem,
- memopv8f64, f64mem, loadf64, "{1to8}",
- X86Fnmsub, v8f64, VK8WM>, EVEX_V512, VEX_W,
- EVEX_CD8<64, CD8VF>;
}
let Constraints = "$src1 = $dst" in {
-multiclass avx512_fma3p_m132<bits<8> opc, string OpcodeStr,
- RegisterClass RC, X86MemOperand x86memop,
- PatFrag mem_frag, X86MemOperand x86scalar_mop, PatFrag scalar_mfrag,
- string BrdcstStr, SDNode OpNode, ValueType OpVT> {
+multiclass avx512_fma3p_m132<bits<8> opc, string OpcodeStr, SDNode OpNode,
+ X86VectorVTInfo _> {
let mayLoad = 1 in
- def m: AVX512FMA3<opc, MRMSrcMem, (outs RC:$dst),
- (ins RC:$src1, RC:$src3, x86memop:$src2),
+ def m: AVX512FMA3<opc, MRMSrcMem, (outs _.RC:$dst),
+ (ins _.RC:$src1, _.RC:$src3, _.MemOp:$src2),
!strconcat(OpcodeStr, " \t{$src2, $src3, $dst|$dst, $src3, $src2}"),
- [(set RC:$dst, (OpVT (OpNode RC:$src1, (mem_frag addr:$src2), RC:$src3)))]>;
- def mb: AVX512FMA3<opc, MRMSrcMem, (outs RC:$dst),
- (ins RC:$src1, RC:$src3, x86scalar_mop:$src2),
- !strconcat(OpcodeStr, " \t{${src2}", BrdcstStr,
- ", $src3, $dst|$dst, $src3, ${src2}", BrdcstStr, "}"),
- [(set RC:$dst, (OpNode RC:$src1,
- (OpVT (X86VBroadcast (scalar_mfrag addr:$src2))), RC:$src3))]>, EVEX_B;
+ [(set _.RC:$dst, (_.VT (OpNode _.RC:$src1, (_.MemOpFrag addr:$src2),
+ _.RC:$src3)))]>;
+ def mb: AVX512FMA3<opc, MRMSrcMem, (outs _.RC:$dst),
+ (ins _.RC:$src1, _.RC:$src3, _.ScalarMemOp:$src2),
+ !strconcat(OpcodeStr, " \t{${src2}", _.BroadcastStr,
+ ", $src3, $dst|$dst, $src3, ${src2}", _.BroadcastStr, "}"),
+ [(set _.RC:$dst,
+ (OpNode _.RC:$src1, (_.VT (X86VBroadcast
+ (_.ScalarLdFrag addr:$src2))),
+ _.RC:$src3))]>, EVEX_B;
}
} // Constraints = "$src1 = $dst"
let ExeDomain = SSEPackedSingle in {
- defm VFMADD132PSZ : avx512_fma3p_m132<0x98, "vfmadd132ps", VR512, f512mem,
- memopv16f32, f32mem, loadf32, "{1to16}",
- X86Fmadd, v16f32>, EVEX_V512,
- EVEX_CD8<32, CD8VF>;
- defm VFMSUB132PSZ : avx512_fma3p_m132<0x9A, "vfmsub132ps", VR512, f512mem,
- memopv16f32, f32mem, loadf32, "{1to16}",
- X86Fmsub, v16f32>, EVEX_V512,
- EVEX_CD8<32, CD8VF>;
- defm VFMADDSUB132PSZ : avx512_fma3p_m132<0x96, "vfmaddsub132ps", VR512, f512mem,
- memopv16f32, f32mem, loadf32, "{1to16}",
- X86Fmaddsub, v16f32>,
- EVEX_V512, EVEX_CD8<32, CD8VF>;
- defm VFMSUBADD132PSZ : avx512_fma3p_m132<0x97, "vfmsubadd132ps", VR512, f512mem,
- memopv16f32, f32mem, loadf32, "{1to16}",
- X86Fmsubadd, v16f32>,
- EVEX_V512, EVEX_CD8<32, CD8VF>;
- defm VFNMADD132PSZ : avx512_fma3p_m132<0x9C, "vfnmadd132ps", VR512, f512mem,
- memopv16f32, f32mem, loadf32, "{1to16}",
- X86Fnmadd, v16f32>, EVEX_V512,
- EVEX_CD8<32, CD8VF>;
- defm VFNMSUB132PSZ : avx512_fma3p_m132<0x9E, "vfnmsub132ps", VR512, f512mem,
- memopv16f32, f32mem, loadf32, "{1to16}",
- X86Fnmsub, v16f32>, EVEX_V512,
- EVEX_CD8<32, CD8VF>;
+ defm VFMADD132PSZ : avx512_fma3p_m132<0x98, "vfmadd132ps", X86Fmadd,
+ v16f32_info>,
+ EVEX_V512, EVEX_CD8<32, CD8VF>;
+ defm VFMSUB132PSZ : avx512_fma3p_m132<0x9A, "vfmsub132ps", X86Fmsub,
+ v16f32_info>,
+ EVEX_V512, EVEX_CD8<32, CD8VF>;
+ defm VFMADDSUB132PSZ : avx512_fma3p_m132<0x96, "vfmaddsub132ps", X86Fmaddsub,
+ v16f32_info>,
+ EVEX_V512, EVEX_CD8<32, CD8VF>;
+ defm VFMSUBADD132PSZ : avx512_fma3p_m132<0x97, "vfmsubadd132ps", X86Fmsubadd,
+ v16f32_info>,
+ EVEX_V512, EVEX_CD8<32, CD8VF>;
+ defm VFNMADD132PSZ : avx512_fma3p_m132<0x9C, "vfnmadd132ps", X86Fnmadd,
+ v16f32_info>,
+ EVEX_V512, EVEX_CD8<32, CD8VF>;
+ defm VFNMSUB132PSZ : avx512_fma3p_m132<0x9E, "vfnmsub132ps", X86Fnmsub,
+ v16f32_info>,
+ EVEX_V512, EVEX_CD8<32, CD8VF>;
}
let ExeDomain = SSEPackedDouble in {
- defm VFMADD132PDZ : avx512_fma3p_m132<0x98, "vfmadd132pd", VR512, f512mem,
- memopv8f64, f64mem, loadf64, "{1to8}",
- X86Fmadd, v8f64>, EVEX_V512,
- VEX_W, EVEX_CD8<64, CD8VF>;
- defm VFMSUB132PDZ : avx512_fma3p_m132<0x9A, "vfmsub132pd", VR512, f512mem,
- memopv8f64, f64mem, loadf64, "{1to8}",
- X86Fmsub, v8f64>, EVEX_V512, VEX_W,
- EVEX_CD8<64, CD8VF>;
- defm VFMADDSUB132PDZ : avx512_fma3p_m132<0x96, "vfmaddsub132pd", VR512, f512mem,
- memopv8f64, f64mem, loadf64, "{1to8}",
- X86Fmaddsub, v8f64>, EVEX_V512, VEX_W,
- EVEX_CD8<64, CD8VF>;
- defm VFMSUBADD132PDZ : avx512_fma3p_m132<0x97, "vfmsubadd132pd", VR512, f512mem,
- memopv8f64, f64mem, loadf64, "{1to8}",
- X86Fmsubadd, v8f64>, EVEX_V512, VEX_W,
- EVEX_CD8<64, CD8VF>;
- defm VFNMADD132PDZ : avx512_fma3p_m132<0x9C, "vfnmadd132pd", VR512, f512mem,
- memopv8f64, f64mem, loadf64, "{1to8}",
- X86Fnmadd, v8f64>, EVEX_V512, VEX_W,
- EVEX_CD8<64, CD8VF>;
- defm VFNMSUB132PDZ : avx512_fma3p_m132<0x9E, "vfnmsub132pd", VR512, f512mem,
- memopv8f64, f64mem, loadf64, "{1to8}",
- X86Fnmsub, v8f64>, EVEX_V512, VEX_W,
- EVEX_CD8<64, CD8VF>;
+ defm VFMADD132PDZ : avx512_fma3p_m132<0x98, "vfmadd132pd", X86Fmadd,
+ v8f64_info>,
+ EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+ defm VFMSUB132PDZ : avx512_fma3p_m132<0x9A, "vfmsub132pd", X86Fmsub,
+ v8f64_info>,
+ EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+ defm VFMADDSUB132PDZ : avx512_fma3p_m132<0x96, "vfmaddsub132pd", X86Fmaddsub,
+ v8f64_info>,
+ EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+ defm VFMSUBADD132PDZ : avx512_fma3p_m132<0x97, "vfmsubadd132pd", X86Fmsubadd,
+ v8f64_info>,
+ EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+ defm VFNMADD132PDZ : avx512_fma3p_m132<0x9C, "vfnmadd132pd", X86Fnmadd,
+ v8f64_info>,
+ EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+ defm VFNMSUB132PDZ : avx512_fma3p_m132<0x9E, "vfnmsub132pd", X86Fnmsub,
+ v8f64_info>,
+ EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
}
// Scalar FMA
(VSHUFPDZrmi VR512:$src1, addr:$src2, imm:$imm)>;
multiclass avx512_valign<X86VectorVTInfo _> {
- defm rri : AVX512_masking<0x03, MRMSrcReg, (outs _.RC:$dst),
+ defm rri : AVX512_masking<0x03, MRMSrcReg, _, (outs _.RC:$dst),
(ins _.RC:$src1, _.RC:$src2, i8imm:$src3),
"valign"##_.Suffix,
"$src3, $src2, $src1", "$src1, $src2, $src3",
(_.VT (X86VAlign _.RC:$src2, _.RC:$src1,
- (i8 imm:$src3))),
- _.VT, _.RC, _.KRCWM>,
+ (i8 imm:$src3)))>,
AVX512AIi8Base, EVEX_4V;
// Also match valign of packed floats.
projects / oota-llvm.git / blobdiff