// 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.
-class X86VectorVTInfo<int NumElts, ValueType EltVT, RegisterClass rc,
+class X86VectorVTInfo<int numelts, ValueType EltVT, RegisterClass rc,
string suffix = ""> {
RegisterClass RC = rc;
+ int NumElts = numelts;
// Corresponding mask register class.
RegisterClass KRC = !cast<RegisterClass>("VK" # NumElts);
// 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));
+
+ // A vector type of the same width with element type i32. This is used to
+ // create the canonical constant zero node ImmAllZerosV.
+ ValueType i32VT = !cast<ValueType>("v" # !srl(Size, 5) # "i32");
+ dag ImmAllZerosV = (VT (bitconvert (i32VT immAllZerosV)));
}
def v64i8_info : X86VectorVTInfo<64, i8, VR512, "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">;
class AVX512VLVectorVTInfo<X86VectorVTInfo i512, X86VectorVTInfo i256,
X86VectorVTInfo i128> {
v4i32x_info>;
def avx512vl_i64_info : AVX512VLVectorVTInfo<v8i64_info, v4i64x_info,
v2i64x_info>;
-
-
-// Common base class of AVX512_masking and AVX512_masking_3src.
-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,
- string MaskingConstraint = "",
- InstrItinClass itin = NoItinerary,
- bit IsCommutable = 0> {
+def avx512vl_f32_info : AVX512VLVectorVTInfo<v16f32_info, v8f32x_info,
+ v4f32x_info>;
+def avx512vl_f64_info : AVX512VLVectorVTInfo<v8f64_info, v4f64x_info,
+ v2f64x_info>;
+
+// 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 = "",
+ InstrItinClass itin = NoItinerary,
+ bit IsCommutable = 0> {
let isCommutable = IsCommutable in
def NAME: AVX512<O, F, Outs, Ins,
OpcodeStr#"\t{"#AttSrcAsm#", $dst|"#
"$dst, "#IntelSrcAsm#"}",
- [(set _.RC:$dst, RHS)], itin>;
+ Pattern, itin>;
// 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)], itin>,
+ MaskingPattern, itin>,
EVEX_K {
// In case of the 3src subclass this is overridden with a let.
string Constraints = MaskingConstraint;
def NAME#kz: AVX512<O, F, Outs, ZeroMaskingIns,
OpcodeStr#"\t{"#AttSrcAsm#", $dst {${mask}} {z}|"#
"$dst {${mask}} {z}, "#IntelSrcAsm#"}",
- [(set _.RC:$dst,
- (vselect _.KRCWM:$mask, RHS,
- (_.VT (bitconvert
- (v16i32 immAllZerosV)))))],
+ ZeroMaskingPattern,
itin>,
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,
+ string MaskingConstraint = "",
+ InstrItinClass itin = NoItinerary,
+ bit IsCommutable = 0> :
+ AVX512_maskable_custom<O, F, Outs, Ins, MaskingIns, ZeroMaskingIns, OpcodeStr,
+ AttSrcAsm, IntelSrcAsm,
+ [(set _.RC:$dst, RHS)],
+ [(set _.RC:$dst, MaskingRHS)],
+ [(set _.RC:$dst,
+ (vselect _.KRCWM:$mask, RHS, _.ImmAllZerosV))],
+ MaskingConstraint, NoItinerary, IsCommutable>;
+
// 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 $dst.
-multiclass AVX512_masking<bits<8> O, Format F, X86VectorVTInfo _,
- dag Outs, dag Ins, string OpcodeStr,
- string AttSrcAsm, string IntelSrcAsm,
- dag RHS, InstrItinClass itin = NoItinerary,
- bit IsCommutable = 0> :
- AVX512_masking_common<O, F, _, Outs, Ins,
- !con((ins _.RC:$src0, _.KRCWM:$mask), Ins),
- !con((ins _.KRCWM:$mask), Ins),
- OpcodeStr, AttSrcAsm, IntelSrcAsm, RHS,
- (vselect _.KRCWM:$mask, RHS, _.RC:$src0),
- "$src0 = $dst", itin, IsCommutable>;
-
-// Similar to AVX512_masking but in this case one of the source operands
+multiclass AVX512_maskable<bits<8> O, Format F, X86VectorVTInfo _,
+ dag Outs, dag Ins, string OpcodeStr,
+ string AttSrcAsm, string IntelSrcAsm,
+ dag RHS, InstrItinClass itin = NoItinerary,
+ bit IsCommutable = 0> :
+ AVX512_maskable_common<O, F, _, Outs, Ins,
+ !con((ins _.RC:$src0, _.KRCWM:$mask), Ins),
+ !con((ins _.KRCWM:$mask), Ins),
+ OpcodeStr, AttSrcAsm, IntelSrcAsm, RHS,
+ (vselect _.KRCWM:$mask, RHS, _.RC:$src0),
+ "$src0 = $dst", itin, IsCommutable>;
+
+// Similar to AVX512_maskable but in this case one of the source operands
// ($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
// $src1.
-multiclass AVX512_masking_3src<bits<8> O, Format F, X86VectorVTInfo _,
- dag Outs, dag NonTiedIns, string OpcodeStr,
- string AttSrcAsm, string IntelSrcAsm,
- dag RHS> :
- AVX512_masking_common<O, F, _, Outs,
- !con((ins _.RC:$src1), NonTiedIns),
- !con((ins _.RC:$src1, _.KRCWM:$mask), NonTiedIns),
- !con((ins _.RC:$src1, _.KRCWM:$mask), NonTiedIns),
- OpcodeStr, AttSrcAsm, IntelSrcAsm, RHS,
- (vselect _.KRCWM:$mask, RHS, _.RC:$src1)>;
+multiclass AVX512_maskable_3src<bits<8> O, Format F, X86VectorVTInfo _,
+ dag Outs, dag NonTiedIns, string OpcodeStr,
+ string AttSrcAsm, string IntelSrcAsm,
+ dag RHS> :
+ AVX512_maskable_common<O, F, _, Outs,
+ !con((ins _.RC:$src1), NonTiedIns),
+ !con((ins _.RC:$src1, _.KRCWM:$mask), NonTiedIns),
+ !con((ins _.RC:$src1, _.KRCWM:$mask), NonTiedIns),
+ OpcodeStr, AttSrcAsm, IntelSrcAsm, RHS,
+ (vselect _.KRCWM:$mask, RHS, _.RC:$src1)>;
+
+
+multiclass AVX512_maskable_in_asm<bits<8> O, Format F, X86VectorVTInfo _,
+ dag Outs, dag Ins,
+ string OpcodeStr,
+ string AttSrcAsm, string IntelSrcAsm,
+ list<dag> Pattern> :
+ AVX512_maskable_custom<O, F, Outs, Ins,
+ !con((ins _.RC:$src0, _.KRCWM:$mask), Ins),
+ !con((ins _.KRCWM:$mask), Ins),
+ OpcodeStr, AttSrcAsm, IntelSrcAsm, Pattern, [], [],
+ "$src0 = $dst">;
// Bitcasts between 512-bit vector types. Return the original type since
// no instruction is needed for the conversion
// AVX-512 - VECTOR INSERT
//
-multiclass vinsert_for_size<int Opcode,
- X86VectorVTInfo From, X86VectorVTInfo To,
- X86VectorVTInfo AltFrom, X86VectorVTInfo AltTo,
- PatFrag vinsert_insert,
- SDNodeXForm INSERT_get_vinsert_imm> {
+multiclass vinsert_for_size_no_alt<int Opcode,
+ X86VectorVTInfo From, X86VectorVTInfo To,
+ PatFrag vinsert_insert,
+ SDNodeXForm INSERT_get_vinsert_imm> {
let hasSideEffects = 0, ExeDomain = To.ExeDomain in {
def rr : AVX512AIi8<Opcode, MRMSrcReg, (outs VR512:$dst),
(ins VR512:$src1, From.RC:$src2, i8imm:$src3),
- "vinsert" # From.EltTypeName # "x4\t{$src3, $src2, $src1, $dst|"
+ "vinsert" # From.EltTypeName # "x" # From.NumElts #
+ "\t{$src3, $src2, $src1, $dst|"
"$dst, $src1, $src2, $src3}",
[(set To.RC:$dst, (vinsert_insert:$src3 (To.VT VR512:$src1),
(From.VT From.RC:$src2),
let mayLoad = 1 in
def rm : AVX512AIi8<Opcode, MRMSrcMem, (outs VR512:$dst),
(ins VR512:$src1, From.MemOp:$src2, i8imm:$src3),
- "vinsert" # From.EltTypeName # "x4\t{$src3, $src2, $src1, $dst|"
+ "vinsert" # From.EltTypeName # "x" # From.NumElts #
+ "\t{$src3, $src2, $src1, $dst|"
"$dst, $src1, $src2, $src3}",
- []>, EVEX_4V, EVEX_V512, EVEX_CD8<From.EltSize, CD8VT4>;
+ []>,
+ EVEX_4V, EVEX_V512, EVEX_CD8<From.EltSize, From.CD8TupleForm>;
}
-
- // Codegen pattern with the alternative types, e.g. v2i64 -> v8i64 for
- // vinserti32x4
- def : Pat<(vinsert_insert:$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:$ins)))>;
}
-multiclass vinsert_for_type<ValueType EltVT32, int Opcode32,
- ValueType EltVT64, int Opcode64> {
- defm NAME # "32x4" : vinsert_for_size<Opcode32,
+multiclass vinsert_for_size<int Opcode,
+ X86VectorVTInfo From, X86VectorVTInfo To,
+ X86VectorVTInfo AltFrom, X86VectorVTInfo AltTo,
+ PatFrag vinsert_insert,
+ SDNodeXForm INSERT_get_vinsert_imm> :
+ vinsert_for_size_no_alt<Opcode, From, To,
+ vinsert_insert, INSERT_get_vinsert_imm> {
+ // Codegen pattern with the alternative types, e.g. v2i64 -> v8i64 for
+ // vinserti32x4. Only add this if 64x2 and friends are not supported
+ // natively via AVX512DQ.
+ let Predicates = [NoDQI] in
+ def : Pat<(vinsert_insert:$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:$ins)))>;
+}
+
+multiclass vinsert_for_type<ValueType EltVT32, int Opcode128,
+ ValueType EltVT64, int Opcode256> {
+ defm NAME # "32x4" : vinsert_for_size<Opcode128,
X86VectorVTInfo< 4, EltVT32, VR128X>,
X86VectorVTInfo<16, EltVT32, VR512>,
X86VectorVTInfo< 2, EltVT64, VR128X>,
X86VectorVTInfo< 8, EltVT64, VR512>,
vinsert128_insert,
INSERT_get_vinsert128_imm>;
- defm NAME # "64x4" : vinsert_for_size<Opcode64,
+ let Predicates = [HasDQI] in
+ defm NAME # "64x2" : vinsert_for_size_no_alt<Opcode128,
+ X86VectorVTInfo< 2, EltVT64, VR128X>,
+ X86VectorVTInfo< 8, EltVT64, VR512>,
+ vinsert128_insert,
+ INSERT_get_vinsert128_imm>, VEX_W;
+ defm NAME # "64x4" : vinsert_for_size<Opcode256,
X86VectorVTInfo< 4, EltVT64, VR256X>,
X86VectorVTInfo< 8, EltVT64, VR512>,
X86VectorVTInfo< 8, EltVT32, VR256>,
X86VectorVTInfo<16, EltVT32, VR512>,
vinsert256_insert,
INSERT_get_vinsert256_imm>, VEX_W;
+ let Predicates = [HasDQI] in
+ defm NAME # "32x8" : vinsert_for_size_no_alt<Opcode256,
+ X86VectorVTInfo< 8, EltVT32, VR256X>,
+ X86VectorVTInfo<16, EltVT32, VR512>,
+ vinsert256_insert,
+ INSERT_get_vinsert256_imm>;
}
defm VINSERTF : vinsert_for_type<f32, 0x18, f64, 0x1a>;
PatFrag vextract_extract,
SDNodeXForm EXTRACT_get_vextract_imm> {
let hasSideEffects = 0, ExeDomain = To.ExeDomain in {
- def rr : AVX512AIi8<Opcode, MRMDestReg, (outs To.RC:$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;
+ defm rr : AVX512_maskable_in_asm<Opcode, MRMDestReg, To, (outs To.RC:$dst),
+ (ins VR512:$src1, i8imm:$idx),
+ "vextract" # To.EltTypeName # "x4",
+ "$idx, $src1", "$src1, $idx",
+ [(set To.RC:$dst, (vextract_extract:$idx (From.VT VR512:$src1),
+ (iPTR imm)))]>,
+ AVX512AIi8Base, EVEX, EVEX_V512;
let mayStore = 1 in
def rm : AVX512AIi8<Opcode, MRMDestMem, (outs),
(ins To.MemOp:$dst, VR512:$src1, i8imm:$src2),
def : Pat<(AltTo.VT (extract_subvector (AltFrom.VT VR512:$src), (iPTR 0))),
(AltTo.VT
(EXTRACT_SUBREG (AltFrom.VT VR512:$src), AltTo.SubRegIdx))>;
+
+ // Intrinsic call with masking.
+ def : Pat<(!cast<Intrinsic>("int_x86_avx512_mask_vextract" # To.EltTypeName #
+ "x4_512")
+ VR512:$src1, (iPTR imm:$idx), To.RC:$src0, GR8:$mask),
+ (!cast<Instruction>(NAME # To.EltSize # "x4rrk") To.RC:$src0,
+ (v4i1 (COPY_TO_REGCLASS GR8:$mask, VK4WM)),
+ VR512:$src1, imm:$idx)>;
+
+ // Intrinsic call with zero-masking.
+ def : Pat<(!cast<Intrinsic>("int_x86_avx512_mask_vextract" # To.EltTypeName #
+ "x4_512")
+ VR512:$src1, (iPTR imm:$idx), To.ImmAllZerosV, GR8:$mask),
+ (!cast<Instruction>(NAME # To.EltSize # "x4rrkz")
+ (v4i1 (COPY_TO_REGCLASS GR8:$mask, VK4WM)),
+ VR512:$src1, imm:$idx)>;
+
+ // Intrinsic call without masking.
+ def : Pat<(!cast<Intrinsic>("int_x86_avx512_mask_vextract" # To.EltTypeName #
+ "x4_512")
+ VR512:$src1, (iPTR imm:$idx), To.ImmAllZerosV, (i8 -1)),
+ (!cast<Instruction>(NAME # To.EltSize # "x4rr")
+ VR512:$src1, imm:$idx)>;
}
multiclass vextract_for_type<ValueType EltVT32, int Opcode32,
//===---------------------------------------------------------------------===//
// AVX-512 BROADCAST
//---
-multiclass avx512_fp_broadcast<bits<8> opc, string OpcodeStr,
- RegisterClass DestRC,
- RegisterClass SrcRC, X86MemOperand x86memop> {
- def rr : AVX5128I<opc, MRMSrcReg, (outs DestRC:$dst), (ins SrcRC:$src),
- !strconcat(OpcodeStr, " \t{$src, $dst|$dst, $src}"),
- []>, EVEX;
- def rm : AVX5128I<opc, MRMSrcMem, (outs DestRC:$dst), (ins x86memop:$src),
- !strconcat(OpcodeStr, " \t{$src, $dst|$dst, $src}"),[]>, EVEX;
+multiclass avx512_fp_broadcast<bits<8> opc, SDNode OpNode, RegisterClass SrcRC,
+ ValueType svt, X86VectorVTInfo _> {
+ defm r : AVX512_maskable<opc, MRMSrcReg, _, (outs _.RC:$dst),
+ (ins SrcRC:$src), "vbroadcast"## !subst("p", "s", _.Suffix),
+ "$src", "$src", (_.VT (OpNode (svt SrcRC:$src)))>,
+ T8PD, EVEX;
+
+ let mayLoad = 1 in {
+ defm m : AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst),
+ (ins _.ScalarMemOp:$src),
+ "vbroadcast"##!subst("p", "s", _.Suffix), "$src", "$src",
+ (_.VT (OpNode (_.ScalarLdFrag addr:$src)))>,
+ T8PD, EVEX;
+ }
+}
+
+multiclass avx512_fp_broadcast_vl<bits<8> opc, SDNode OpNode,
+ AVX512VLVectorVTInfo _> {
+ defm Z : avx512_fp_broadcast<opc, OpNode, VR128X, _.info128.VT, _.info512>,
+ EVEX_V512;
+
+ let Predicates = [HasVLX] in {
+ defm Z256 : avx512_fp_broadcast<opc, OpNode, VR128X, _.info128.VT, _.info256>,
+ EVEX_V256;
+ }
}
+
let ExeDomain = SSEPackedSingle in {
- defm VBROADCASTSSZ : avx512_fp_broadcast<0x18, "vbroadcastss", VR512,
- VR128X, f32mem>,
- EVEX_V512, EVEX_CD8<32, CD8VT1>;
+ defm VBROADCASTSS : avx512_fp_broadcast_vl<0x18, X86VBroadcast,
+ avx512vl_f32_info>, EVEX_CD8<32, CD8VT1>;
+ let Predicates = [HasVLX] in {
+ defm VBROADCASTSSZ128 : avx512_fp_broadcast<0x18, X86VBroadcast, VR128X,
+ v4f32, v4f32x_info>, EVEX_V128,
+ EVEX_CD8<32, CD8VT1>;
+ }
}
let ExeDomain = SSEPackedDouble in {
- defm VBROADCASTSDZ : avx512_fp_broadcast<0x19, "vbroadcastsd", VR512,
- VR128X, f64mem>,
- EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT1>;
+ defm VBROADCASTSD : avx512_fp_broadcast_vl<0x19, X86VBroadcast,
+ avx512vl_f64_info>, VEX_W, EVEX_CD8<64, CD8VT1>;
}
def : Pat<(v16f32 (X86VBroadcast (loadf32 addr:$src))),
- (VBROADCASTSSZrm addr:$src)>;
+ (VBROADCASTSSZm addr:$src)>;
def : Pat<(v8f64 (X86VBroadcast (loadf64 addr:$src))),
- (VBROADCASTSDZrm addr:$src)>;
+ (VBROADCASTSDZm addr:$src)>;
def : Pat<(int_x86_avx512_vbroadcast_ss_512 addr:$src),
- (VBROADCASTSSZrm addr:$src)>;
+ (VBROADCASTSSZm addr:$src)>;
def : Pat<(int_x86_avx512_vbroadcast_sd_512 addr:$src),
- (VBROADCASTSDZrm addr:$src)>;
+ (VBROADCASTSDZm addr:$src)>;
multiclass avx512_int_broadcast_reg<bits<8> opc, string OpcodeStr,
RegisterClass SrcRC, RegisterClass KRC> {
(VPBROADCASTQZrr VR128X:$src)>;
def : Pat<(v16f32 (X86VBroadcast (v4f32 VR128X:$src))),
- (VBROADCASTSSZrr VR128X:$src)>;
+ (VBROADCASTSSZr VR128X:$src)>;
def : Pat<(v8f64 (X86VBroadcast (v2f64 VR128X:$src))),
- (VBROADCASTSDZrr VR128X:$src)>;
+ (VBROADCASTSDZr VR128X:$src)>;
+
+def : Pat<(v16f32 (X86VBroadcast (v16f32 VR512:$src))),
+ (VBROADCASTSSZr (EXTRACT_SUBREG (v16f32 VR512:$src), sub_xmm))>;
+def : Pat<(v8f64 (X86VBroadcast (v8f64 VR512:$src))),
+ (VBROADCASTSDZr (EXTRACT_SUBREG (v8f64 VR512:$src), sub_xmm))>;
+
+def : Pat<(v16i32 (X86VBroadcast (v16i32 VR512:$src))),
+ (VPBROADCASTDZrr (EXTRACT_SUBREG (v16i32 VR512:$src), sub_xmm))>;
+def : Pat<(v8i64 (X86VBroadcast (v8i64 VR512:$src))),
+ (VPBROADCASTQZrr (EXTRACT_SUBREG (v8i64 VR512:$src), sub_xmm))>;
def : Pat<(v16f32 (int_x86_avx512_vbroadcast_ss_ps_512 (v4f32 VR128X:$src))),
- (VBROADCASTSSZrr VR128X:$src)>;
+ (VBROADCASTSSZr VR128X:$src)>;
def : Pat<(v8f64 (int_x86_avx512_vbroadcast_sd_pd_512 (v2f64 VR128X:$src))),
- (VBROADCASTSDZrr VR128X:$src)>;
+ (VBROADCASTSDZr VR128X:$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<(v16f32 (X86VBroadcast FR32X:$src)),
- (VBROADCASTSSZrr (COPY_TO_REGCLASS FR32X:$src, VR128X))>;
+ (VBROADCASTSSZr (COPY_TO_REGCLASS FR32X:$src, VR128X))>;
def : Pat<(v8f64 (X86VBroadcast FR64X:$src)),
- (VBROADCASTSDZrr (COPY_TO_REGCLASS FR64X:$src, VR128X))>;
+ (VBROADCASTSDZr (COPY_TO_REGCLASS FR64X:$src, VR128X))>;
let Predicates = [HasAVX512] in {
//---
multiclass avx512_mask_broadcast<bits<8> opc, string OpcodeStr,
- RegisterClass DstRC, RegisterClass KRC,
- ValueType OpVT, ValueType SrcVT> {
-def rr : AVX512XS8I<opc, MRMDestReg, (outs DstRC:$dst), (ins KRC:$src),
+ RegisterClass KRC> {
+let Predicates = [HasCDI] in
+def Zrr : AVX512XS8I<opc, MRMSrcReg, (outs VR512:$dst), (ins KRC:$src),
!strconcat(OpcodeStr, " \t{$src, $dst|$dst, $src}"),
- []>, EVEX;
+ []>, EVEX, EVEX_V512;
+
+let Predicates = [HasCDI, HasVLX] in {
+def Z128rr : AVX512XS8I<opc, MRMSrcReg, (outs VR128:$dst), (ins KRC:$src),
+ !strconcat(OpcodeStr, " \t{$src, $dst|$dst, $src}"),
+ []>, EVEX, EVEX_V128;
+def Z256rr : AVX512XS8I<opc, MRMSrcReg, (outs VR256:$dst), (ins KRC:$src),
+ !strconcat(OpcodeStr, " \t{$src, $dst|$dst, $src}"),
+ []>, EVEX, EVEX_V256;
+}
}
let Predicates = [HasCDI] in {
-defm VPBROADCASTMW2D : avx512_mask_broadcast<0x3A, "vpbroadcastmw2d", VR512,
- VK16, v16i32, v16i1>, EVEX_V512;
-defm VPBROADCASTMB2Q : avx512_mask_broadcast<0x2A, "vpbroadcastmb2q", VR512,
- VK8, v8i64, v8i1>, EVEX_V512, VEX_W;
+defm VPBROADCASTMW2D : avx512_mask_broadcast<0x3A, "vpbroadcastmw2d",
+ VK16>;
+defm VPBROADCASTMB2Q : avx512_mask_broadcast<0x2A, "vpbroadcastmb2q",
+ VK8>, VEX_W;
}
//===----------------------------------------------------------------------===//
// AVX-512 - VPERM
//
// -- immediate form --
-multiclass avx512_perm_imm<bits<8> opc, string OpcodeStr, RegisterClass RC,
- SDNode OpNode, PatFrag mem_frag,
- X86MemOperand x86memop, ValueType OpVT> {
- def ri : AVX512AIi8<opc, MRMSrcReg, (outs RC:$dst),
- (ins RC:$src1, i8imm:$src2),
+multiclass avx512_perm_imm<bits<8> opc, string OpcodeStr, SDNode OpNode,
+ X86VectorVTInfo _> {
+ let ExeDomain = _.ExeDomain in {
+ def ri : AVX512AIi8<opc, MRMSrcReg, (outs _.RC:$dst),
+ (ins _.RC:$src1, i8imm:$src2),
!strconcat(OpcodeStr,
" \t{$src2, $src1, $dst|$dst, $src1, $src2}"),
- [(set RC:$dst,
- (OpVT (OpNode RC:$src1, (i8 imm:$src2))))]>,
+ [(set _.RC:$dst,
+ (_.VT (OpNode _.RC:$src1, (i8 imm:$src2))))]>,
EVEX;
- def mi : AVX512AIi8<opc, MRMSrcMem, (outs RC:$dst),
- (ins x86memop:$src1, i8imm:$src2),
+ def mi : AVX512AIi8<opc, MRMSrcMem, (outs _.RC:$dst),
+ (ins _.MemOp:$src1, i8imm:$src2),
!strconcat(OpcodeStr,
" \t{$src2, $src1, $dst|$dst, $src1, $src2}"),
- [(set RC:$dst,
- (OpVT (OpNode (mem_frag addr:$src1),
- (i8 imm:$src2))))]>, EVEX;
+ [(set _.RC:$dst,
+ (_.VT (OpNode (_.MemOpFrag addr:$src1),
+ (i8 imm:$src2))))]>,
+ EVEX, EVEX_CD8<_.EltSize, CD8VF>;
+}
+}
+
+multiclass avx512_permil<bits<8> OpcImm, bits<8> OpcVar, X86VectorVTInfo _,
+ X86VectorVTInfo Ctrl> :
+ avx512_perm_imm<OpcImm, "vpermil" # _.Suffix, X86VPermilpi, _> {
+ let ExeDomain = _.ExeDomain in {
+ def rr : AVX5128I<OpcVar, MRMSrcReg, (outs _.RC:$dst),
+ (ins _.RC:$src1, _.RC:$src2),
+ !strconcat("vpermil" # _.Suffix,
+ " \t{$src2, $src1, $dst|$dst, $src1, $src2}"),
+ [(set _.RC:$dst,
+ (_.VT (X86VPermilpv _.RC:$src1,
+ (Ctrl.VT Ctrl.RC:$src2))))]>,
+ EVEX_4V;
+ def rm : AVX5128I<OpcVar, MRMSrcMem, (outs _.RC:$dst),
+ (ins _.RC:$src1, Ctrl.MemOp:$src2),
+ !strconcat("vpermil" # _.Suffix,
+ " \t{$src2, $src1, $dst|$dst, $src1, $src2}"),
+ [(set _.RC:$dst,
+ (_.VT (X86VPermilpv _.RC:$src1,
+ (Ctrl.VT (Ctrl.MemOpFrag addr:$src2)))))]>,
+ EVEX_4V;
+ }
}
-defm VPERMQZ : avx512_perm_imm<0x00, "vpermq", VR512, X86VPermi, memopv8i64,
- i512mem, v8i64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
-let ExeDomain = SSEPackedDouble in
-defm VPERMPDZ : avx512_perm_imm<0x01, "vpermpd", VR512, X86VPermi, memopv8f64,
- f512mem, v8f64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+defm VPERMQZ : avx512_perm_imm<0x00, "vpermq", X86VPermi, v8i64_info>,
+ EVEX_V512, VEX_W;
+defm VPERMPDZ : avx512_perm_imm<0x01, "vpermpd", X86VPermi, v8f64_info>,
+ EVEX_V512, VEX_W;
+
+defm VPERMILPSZ : avx512_permil<0x04, 0x0C, v16f32_info, v16i32_info>,
+ EVEX_V512;
+defm VPERMILPDZ : avx512_permil<0x05, 0x0D, v8f64_info, v8i64_info>,
+ EVEX_V512, VEX_W;
+
+def : Pat<(v16i32 (X86VPermilpi VR512:$src1, (i8 imm:$imm))),
+ (VPERMILPSZri VR512:$src1, imm:$imm)>;
+def : Pat<(v8i64 (X86VPermilpi VR512:$src1, (i8 imm:$imm))),
+ (VPERMILPDZri VR512:$src1, imm:$imm)>;
// -- VPERM - register form --
multiclass avx512_perm<bits<8> opc, string OpcodeStr, RegisterClass RC,
multiclass avx512_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
X86VectorVTInfo _, OpndItins itins,
bit IsCommutable = 0> {
- defm rr : AVX512_masking<opc, MRMSrcReg, _, (outs _.RC:$dst),
+ defm rr : AVX512_maskable<opc, MRMSrcReg, _, (outs _.RC:$dst),
(ins _.RC:$src1, _.RC:$src2), OpcodeStr##_.Suffix,
"$src2, $src1", "$src1, $src2",
(_.VT (OpNode _.RC:$src1, _.RC:$src2)),
itins.rr, IsCommutable>,
AVX512BIBase, EVEX_4V;
- let mayLoad = 1 in {
- defm rm : AVX512_masking<opc, MRMSrcMem, _, (outs _.RC:$dst),
+ let mayLoad = 1 in
+ defm rm : AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst),
(ins _.RC:$src1, _.MemOp:$src2), OpcodeStr##_.Suffix,
"$src2, $src1", "$src1, $src2",
(_.VT (OpNode _.RC:$src1,
(bitconvert (_.LdFrag addr:$src2)))),
itins.rm>,
AVX512BIBase, EVEX_4V;
- defm rmb : AVX512_masking<opc, MRMSrcMem, _, (outs _.RC:$dst),
+}
+
+multiclass avx512_binop_rmb<bits<8> opc, string OpcodeStr, SDNode OpNode,
+ X86VectorVTInfo _, OpndItins itins,
+ bit IsCommutable = 0> :
+ avx512_binop_rm<opc, OpcodeStr, OpNode, _, itins, IsCommutable> {
+ let mayLoad = 1 in
+ defm rmb : AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst),
(ins _.RC:$src1, _.ScalarMemOp:$src2), OpcodeStr##_.Suffix,
"${src2}"##_.BroadcastStr##", $src1",
"$src1, ${src2}"##_.BroadcastStr,
(_.ScalarLdFrag addr:$src2)))),
itins.rm>,
AVX512BIBase, EVEX_4V, EVEX_B;
+}
+
+multiclass avx512_binop_rm_vl<bits<8> opc, string OpcodeStr, SDNode OpNode,
+ AVX512VLVectorVTInfo VTInfo, OpndItins itins,
+ Predicate prd, bit IsCommutable = 0> {
+ let Predicates = [prd] in
+ defm Z : avx512_binop_rm<opc, OpcodeStr, OpNode, VTInfo.info512, itins,
+ IsCommutable>, EVEX_V512;
+
+ let Predicates = [prd, HasVLX] in {
+ defm Z256 : avx512_binop_rm<opc, OpcodeStr, OpNode, VTInfo.info256, itins,
+ IsCommutable>, EVEX_V256;
+ defm Z128 : avx512_binop_rm<opc, OpcodeStr, OpNode, VTInfo.info128, itins,
+ IsCommutable>, EVEX_V128;
}
}
+multiclass avx512_binop_rmb_vl<bits<8> opc, string OpcodeStr, SDNode OpNode,
+ AVX512VLVectorVTInfo VTInfo, OpndItins itins,
+ Predicate prd, bit IsCommutable = 0> {
+ let Predicates = [prd] in
+ defm Z : avx512_binop_rmb<opc, OpcodeStr, OpNode, VTInfo.info512, itins,
+ IsCommutable>, EVEX_V512;
+
+ let Predicates = [prd, HasVLX] in {
+ defm Z256 : avx512_binop_rmb<opc, OpcodeStr, OpNode, VTInfo.info256, itins,
+ IsCommutable>, EVEX_V256;
+ defm Z128 : avx512_binop_rmb<opc, OpcodeStr, OpNode, VTInfo.info128, itins,
+ IsCommutable>, EVEX_V128;
+ }
+}
+
+multiclass avx512_binop_rm_vl_q<bits<8> opc, string OpcodeStr, SDNode OpNode,
+ OpndItins itins, Predicate prd,
+ bit IsCommutable = 0> {
+ defm NAME : avx512_binop_rmb_vl<opc, OpcodeStr, OpNode, avx512vl_i64_info,
+ itins, prd, IsCommutable>,
+ VEX_W, EVEX_CD8<64, CD8VF>;
+}
+
+multiclass avx512_binop_rm_vl_d<bits<8> opc, string OpcodeStr, SDNode OpNode,
+ OpndItins itins, Predicate prd,
+ bit IsCommutable = 0> {
+ defm NAME : avx512_binop_rmb_vl<opc, OpcodeStr, OpNode, avx512vl_i32_info,
+ itins, prd, IsCommutable>, EVEX_CD8<32, CD8VF>;
+}
+
+multiclass avx512_binop_rm_vl_w<bits<8> opc, string OpcodeStr, SDNode OpNode,
+ OpndItins itins, Predicate prd,
+ bit IsCommutable = 0> {
+ defm NAME : avx512_binop_rm_vl<opc, OpcodeStr, OpNode, avx512vl_i16_info,
+ itins, prd, IsCommutable>, EVEX_CD8<16, CD8VF>;
+}
+
+multiclass avx512_binop_rm_vl_b<bits<8> opc, string OpcodeStr, SDNode OpNode,
+ OpndItins itins, Predicate prd,
+ bit IsCommutable = 0> {
+ defm NAME : avx512_binop_rm_vl<opc, OpcodeStr, OpNode, avx512vl_i8_info,
+ itins, prd, IsCommutable>, EVEX_CD8<8, CD8VF>;
+}
+
+multiclass avx512_binop_rm_vl_dq<bits<8> opc_d, bits<8> opc_q, string OpcodeStr,
+ SDNode OpNode, OpndItins itins, Predicate prd,
+ bit IsCommutable = 0> {
+ defm Q : avx512_binop_rm_vl_q<opc_q, OpcodeStr, OpNode, itins, prd,
+ IsCommutable>;
+
+ defm D : avx512_binop_rm_vl_d<opc_d, OpcodeStr, OpNode, itins, prd,
+ IsCommutable>;
+}
+
+multiclass avx512_binop_rm_vl_bw<bits<8> opc_b, bits<8> opc_w, string OpcodeStr,
+ SDNode OpNode, OpndItins itins, Predicate prd,
+ bit IsCommutable = 0> {
+ defm W : avx512_binop_rm_vl_w<opc_w, OpcodeStr, OpNode, itins, prd,
+ IsCommutable>;
+
+ defm B : avx512_binop_rm_vl_b<opc_b, OpcodeStr, OpNode, itins, prd,
+ IsCommutable>;
+}
+
+multiclass avx512_binop_rm_vl_all<bits<8> opc_b, bits<8> opc_w,
+ bits<8> opc_d, bits<8> opc_q,
+ string OpcodeStr, SDNode OpNode,
+ OpndItins itins, bit IsCommutable = 0> {
+ defm NAME : avx512_binop_rm_vl_dq<opc_d, opc_q, OpcodeStr, OpNode,
+ itins, HasAVX512, IsCommutable>,
+ avx512_binop_rm_vl_bw<opc_b, opc_w, OpcodeStr, OpNode,
+ itins, HasBWI, IsCommutable>;
+}
+
multiclass avx512_binop_rm2<bits<8> opc, string OpcodeStr, ValueType DstVT,
ValueType SrcVT, RegisterClass KRC, RegisterClass RC,
PatFrag memop_frag, X86MemOperand x86memop,
}
}
-defm VPADDDZ : avx512_binop_rm<0xFE, "vpadd", add, v16i32_info,
- SSE_INTALU_ITINS_P, 1>, EVEX_V512, EVEX_CD8<32, CD8VF>;
-
-defm VPSUBDZ : avx512_binop_rm<0xFA, "vpsub", sub, v16i32_info,
- SSE_INTALU_ITINS_P, 0>, EVEX_V512, EVEX_CD8<32, CD8VF>;
-
-defm VPMULLDZ : avx512_binop_rm<0x40, "vpmull", mul, v16i32_info,
- SSE_INTALU_ITINS_P, 1>, T8PD, EVEX_V512, EVEX_CD8<32, CD8VF>;
-
-defm VPADDQZ : avx512_binop_rm<0xD4, "vpadd", add, v8i64_info,
- SSE_INTALU_ITINS_P, 1>, EVEX_CD8<64, CD8VF>, EVEX_V512, VEX_W;
-
-defm VPSUBQZ : avx512_binop_rm<0xFB, "vpsub", sub, v8i64_info,
- SSE_INTALU_ITINS_P, 0>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+defm VPADD : avx512_binop_rm_vl_all<0xFC, 0xFD, 0xFE, 0xD4, "vpadd", add,
+ SSE_INTALU_ITINS_P, 1>;
+defm VPSUB : avx512_binop_rm_vl_all<0xF8, 0xF9, 0xFA, 0xFB, "vpsub", sub,
+ SSE_INTALU_ITINS_P, 0>;
+defm VPMULLD : avx512_binop_rm_vl_d<0x40, "vpmull", mul,
+ SSE_INTALU_ITINS_P, HasAVX512, 1>, T8PD;
+defm VPMULLW : avx512_binop_rm_vl_w<0xD5, "vpmull", mul,
+ SSE_INTALU_ITINS_P, HasBWI, 1>;
+defm VPMULLQ : avx512_binop_rm_vl_q<0x40, "vpmull", mul,
+ SSE_INTALU_ITINS_P, HasDQI, 1>, T8PD;
defm VPMULDQZ : avx512_binop_rm2<0x28, "vpmuldq", v8i64, v16i32, VK8WM, VR512,
memopv8i64, i512mem, loadi64, i64mem, "{1to8}",
(v16i32 VR512:$src2), (bc_v8i64 (v16i32 immAllZerosV)), (i8 -1))),
(VPMULDQZrr VR512:$src1, VR512:$src2)>;
-defm VPMAXUDZ : avx512_binop_rm<0x3F, "vpmaxu", X86umax, v16i32_info,
- SSE_INTALU_ITINS_P, 1>,
- T8PD, EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VPMAXUQZ : avx512_binop_rm<0x3F, "vpmaxu", X86umax, v8i64_info,
- SSE_INTALU_ITINS_P, 0>,
- T8PD, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
-
-defm VPMAXSDZ : avx512_binop_rm<0x3D, "vpmaxs", X86smax, v16i32_info,
- SSE_INTALU_ITINS_P, 1>,
- T8PD, EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VPMAXSQZ : avx512_binop_rm<0x3D, "vpmaxs", X86smax, v8i64_info,
- SSE_INTALU_ITINS_P, 0>,
- T8PD, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
-
-defm VPMINUDZ : avx512_binop_rm<0x3B, "vpminu", X86umin, v16i32_info,
- SSE_INTALU_ITINS_P, 1>,
- T8PD, EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VPMINUQZ : avx512_binop_rm<0x3B, "vpminu", X86umin, v8i64_info,
- SSE_INTALU_ITINS_P, 0>,
- T8PD, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
-
-defm VPMINSDZ : avx512_binop_rm<0x39, "vpmins", X86smin, v16i32_info,
- SSE_INTALU_ITINS_P, 1>,
- T8PD, EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VPMINSQZ : avx512_binop_rm<0x39, "vpmins", X86smin, v8i64_info,
- SSE_INTALU_ITINS_P, 0>,
- T8PD, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+defm VPMAXSB : avx512_binop_rm_vl_b<0x3C, "vpmaxs", X86smax,
+ SSE_INTALU_ITINS_P, HasBWI, 1>, T8PD;
+defm VPMAXSW : avx512_binop_rm_vl_w<0xEE, "vpmaxs", X86smax,
+ SSE_INTALU_ITINS_P, HasBWI, 1>;
+defm VPMAXS : avx512_binop_rm_vl_dq<0x3D, 0x3D, "vpmaxs", X86smax,
+ SSE_INTALU_ITINS_P, HasAVX512, 1>, T8PD;
+
+defm VPMAXUB : avx512_binop_rm_vl_b<0xDE, "vpmaxu", X86umax,
+ SSE_INTALU_ITINS_P, HasBWI, 1>;
+defm VPMAXUW : avx512_binop_rm_vl_w<0x3E, "vpmaxu", X86umax,
+ SSE_INTALU_ITINS_P, HasBWI, 1>, T8PD;
+defm VPMAXU : avx512_binop_rm_vl_dq<0x3F, 0x3F, "vpmaxu", X86umax,
+ SSE_INTALU_ITINS_P, HasAVX512, 1>, T8PD;
+
+defm VPMINSB : avx512_binop_rm_vl_b<0x38, "vpmins", X86smin,
+ SSE_INTALU_ITINS_P, HasBWI, 1>, T8PD;
+defm VPMINSW : avx512_binop_rm_vl_w<0xEA, "vpmins", X86smin,
+ SSE_INTALU_ITINS_P, HasBWI, 1>;
+defm VPMINS : avx512_binop_rm_vl_dq<0x39, 0x39, "vpmins", X86smin,
+ SSE_INTALU_ITINS_P, HasAVX512, 1>, T8PD;
+
+defm VPMINUB : avx512_binop_rm_vl_b<0xDA, "vpminu", X86umin,
+ SSE_INTALU_ITINS_P, HasBWI, 1>;
+defm VPMINUW : avx512_binop_rm_vl_w<0x3A, "vpminu", X86umin,
+ SSE_INTALU_ITINS_P, HasBWI, 1>, T8PD;
+defm VPMINU : avx512_binop_rm_vl_dq<0x3B, 0x3B, "vpminu", X86umin,
+ SSE_INTALU_ITINS_P, HasAVX512, 1>, T8PD;
def : Pat <(v16i32 (int_x86_avx512_mask_pmaxs_d_512 (v16i32 VR512:$src1),
(v16i32 VR512:$src2), (v16i32 immAllZerosV), (i16 -1))),
defm VPSHUFDZ : avx512_pshuf_imm<0x70, "vpshufd", VR512, X86PShufd, memopv16i32,
i512mem, v16i32>, PD, EVEX_V512, EVEX_CD8<32, CD8VF>;
-let ExeDomain = SSEPackedSingle in
-defm VPERMILPSZ : avx512_pshuf_imm<0x04, "vpermilps", VR512, X86VPermilpi,
- memopv16f32, i512mem, v16f32>, TAPD, EVEX_V512,
- EVEX_CD8<32, CD8VF>;
-let ExeDomain = SSEPackedDouble in
-defm VPERMILPDZ : avx512_pshuf_imm<0x05, "vpermilpd", VR512, X86VPermilpi,
- memopv8f64, i512mem, v8f64>, TAPD, EVEX_V512,
- VEX_W, EVEX_CD8<32, CD8VF>;
-
-def : Pat<(v16i32 (X86VPermilpi VR512:$src1, (i8 imm:$imm))),
- (VPERMILPSZri VR512:$src1, imm:$imm)>;
-def : Pat<(v8i64 (X86VPermilpi VR512:$src1, (i8 imm:$imm))),
- (VPERMILPDZri VR512:$src1, imm:$imm)>;
-
//===----------------------------------------------------------------------===//
// AVX-512 Logical Instructions
//===----------------------------------------------------------------------===//
-defm VPANDDZ : avx512_binop_rm<0xDB, "vpand", and, v16i32_info, SSE_BIT_ITINS_P, 1>,
- EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VPANDQZ : avx512_binop_rm<0xDB, "vpand", and, v8i64_info, SSE_BIT_ITINS_P, 1>,
- EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
-defm VPORDZ : avx512_binop_rm<0xEB, "vpor", or, v16i32_info, SSE_BIT_ITINS_P, 1>,
- EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VPORQZ : avx512_binop_rm<0xEB, "vpor", or, v8i64_info, SSE_BIT_ITINS_P, 1>,
- EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
-defm VPXORDZ : avx512_binop_rm<0xEF, "vpxor", xor, v16i32_info, SSE_BIT_ITINS_P, 1>,
- EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VPXORQZ : avx512_binop_rm<0xEF, "vpxor", xor, v8i64_info, SSE_BIT_ITINS_P, 1>,
- EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
-defm VPANDNDZ : avx512_binop_rm<0xDF, "vpandn", X86andnp, v16i32_info,
- SSE_BIT_ITINS_P, 0>, EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VPANDNQZ : avx512_binop_rm<0xDF, "vpandn", X86andnp, v8i64_info,
- SSE_BIT_ITINS_P, 0>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+defm VPAND : avx512_binop_rm_vl_dq<0xDB, 0xDB, "vpand", and,
+ SSE_INTALU_ITINS_P, HasAVX512, 1>;
+defm VPOR : avx512_binop_rm_vl_dq<0xEB, 0xEB, "vpor", or,
+ SSE_INTALU_ITINS_P, HasAVX512, 1>;
+defm VPXOR : avx512_binop_rm_vl_dq<0xEF, 0xEF, "vpxor", xor,
+ SSE_INTALU_ITINS_P, HasAVX512, 1>;
+defm VPANDN : avx512_binop_rm_vl_dq<0xDF, 0xDF, "vpandn", X86andnp,
+ SSE_INTALU_ITINS_P, HasAVX512, 1>;
//===----------------------------------------------------------------------===//
// AVX-512 FP arithmetic
}
multiclass avx512_fp_packed<bits<8> opc, string OpcodeStr, SDNode OpNode,
- RegisterClass KRC,
- RegisterClass RC, ValueType vt,
- X86MemOperand x86memop, PatFrag mem_frag,
- X86MemOperand x86scalar_mop, PatFrag scalar_mfrag,
- string BrdcstStr,
- Domain d, OpndItins itins, bit commutable> {
- let isCommutable = commutable in {
- def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
- !strconcat(OpcodeStr, " \t{$src2, $src1, $dst|$dst, $src1, $src2}"),
- [(set RC:$dst, (vt (OpNode RC:$src1, RC:$src2)))], itins.rr, d>,
- EVEX_4V;
-
- def rrk: PI<opc, MRMSrcReg, (outs RC:$dst), (ins KRC:$mask, RC:$src1, RC:$src2),
- !strconcat(OpcodeStr,
- " \t{$src2, $src1, $dst {${mask}} |$dst {${mask}}, $src1, $src2}"),
- [], itins.rr, d>, EVEX_4V, EVEX_K;
-
- def rrkz: PI<opc, MRMSrcReg, (outs RC:$dst), (ins KRC:$mask, RC:$src1, RC:$src2),
- !strconcat(OpcodeStr,
- " \t{$src2, $src1, $dst {${mask}} {z}|$dst {${mask}} {z}, $src1, $src2}"),
- [], itins.rr, d>, EVEX_4V, EVEX_KZ;
- }
-
+ X86VectorVTInfo _, bit IsCommutable> {
+ defm rr: AVX512_maskable<opc, MRMSrcReg, _, (outs _.RC:$dst),
+ (ins _.RC:$src1, _.RC:$src2), OpcodeStr##_.Suffix,
+ "$src2, $src1", "$src1, $src2",
+ (_.VT (OpNode _.RC:$src1, _.RC:$src2))>, EVEX_4V;
let mayLoad = 1 in {
- def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
- !strconcat(OpcodeStr, " \t{$src2, $src1, $dst|$dst, $src1, $src2}"),
- [(set RC:$dst, (OpNode RC:$src1, (mem_frag addr:$src2)))],
- itins.rm, d>, EVEX_4V;
+ defm rm: AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst),
+ (ins _.RC:$src1, _.MemOp:$src2), OpcodeStr##_.Suffix,
+ "$src2, $src1", "$src1, $src2",
+ (OpNode _.RC:$src1, (_.LdFrag addr:$src2))>, EVEX_4V;
+ defm rmb: AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst),
+ (ins _.RC:$src1, _.ScalarMemOp:$src2), OpcodeStr##_.Suffix,
+ "${src2}"##_.BroadcastStr##", $src1",
+ "$src1, ${src2}"##_.BroadcastStr,
+ (OpNode _.RC:$src1, (_.VT (X86VBroadcast
+ (_.ScalarLdFrag addr:$src2))))>,
+ EVEX_4V, EVEX_B;
+ }//let mayLoad = 1
+}
+
+multiclass avx512_fp_binop_p<bits<8> opc, string OpcodeStr, SDNode OpNode,
+ bit IsCommutable = 0> {
+ defm PSZ : avx512_fp_packed<opc, OpcodeStr, OpNode, v16f32_info,
+ IsCommutable>, EVEX_V512, PS,
+ EVEX_CD8<32, CD8VF>;
+ defm PDZ : avx512_fp_packed<opc, OpcodeStr, OpNode, v8f64_info,
+ IsCommutable>, EVEX_V512, PD, VEX_W,
+ EVEX_CD8<64, CD8VF>;
- def rmb : PI<opc, MRMSrcMem, (outs RC:$dst),
- (ins RC:$src1, x86scalar_mop:$src2),
- !strconcat(OpcodeStr, " \t{${src2}", BrdcstStr,
- ", $src1, $dst|$dst, $src1, ${src2}", BrdcstStr, "}"),
- [(set RC:$dst, (OpNode RC:$src1,
- (vt (X86VBroadcast (scalar_mfrag addr:$src2)))))],
- itins.rm, d>, EVEX_4V, EVEX_B;
-
- def rmk : PI<opc, MRMSrcMem, (outs RC:$dst),
- (ins KRC:$mask, RC:$src1, x86memop:$src2), !strconcat(OpcodeStr,
- "\t{$src2, $src1, $dst {${mask}}|$dst {${mask}}, $src1, $src2}"),
- [], itins.rm, d>, EVEX_4V, EVEX_K;
-
- def rmkz : PI<opc, MRMSrcMem, (outs RC:$dst),
- (ins KRC:$mask, RC:$src1, x86memop:$src2), !strconcat(OpcodeStr,
- "\t{$src2, $src1, $dst {${mask}} {z}|$dst {${mask}} {z}, $src1, $src2}"),
- [], itins.rm, d>, EVEX_4V, EVEX_KZ;
-
- def rmbk : PI<opc, MRMSrcMem, (outs RC:$dst),
- (ins KRC:$mask, RC:$src1, x86scalar_mop:$src2), !strconcat(OpcodeStr,
- " \t{${src2}", BrdcstStr,
- ", $src1, $dst {${mask}}|$dst {${mask}}, $src1, ${src2}", BrdcstStr, "}"),
- [], itins.rm, d>, EVEX_4V, EVEX_B, EVEX_K;
-
- def rmbkz : PI<opc, MRMSrcMem, (outs RC:$dst),
- (ins KRC:$mask, RC:$src1, x86scalar_mop:$src2), !strconcat(OpcodeStr,
- " \t{${src2}", BrdcstStr,
- ", $src1, $dst {${mask}} {z}|$dst {${mask}} {z}, $src1, ${src2}",
- BrdcstStr, "}"),
- [], itins.rm, d>, EVEX_4V, EVEX_B, EVEX_KZ;
+ // Define only if AVX512VL feature is present.
+ let Predicates = [HasVLX] in {
+ defm PSZ128 : avx512_fp_packed<opc, OpcodeStr, OpNode, v4f32x_info,
+ IsCommutable>, EVEX_V128, PS,
+ EVEX_CD8<32, CD8VF>;
+ defm PSZ256 : avx512_fp_packed<opc, OpcodeStr, OpNode, v8f32x_info,
+ IsCommutable>, EVEX_V256, PS,
+ EVEX_CD8<32, CD8VF>;
+ defm PDZ128 : avx512_fp_packed<opc, OpcodeStr, OpNode, v2f64x_info,
+ IsCommutable>, EVEX_V128, PD, VEX_W,
+ EVEX_CD8<64, CD8VF>;
+ defm PDZ256 : avx512_fp_packed<opc, OpcodeStr, OpNode, v4f64x_info,
+ IsCommutable>, EVEX_V256, PD, VEX_W,
+ EVEX_CD8<64, CD8VF>;
}
}
-defm VADDPSZ : avx512_fp_packed<0x58, "addps", fadd, VK16WM, VR512, v16f32, f512mem,
- memopv16f32, f32mem, loadf32, "{1to16}", SSEPackedSingle,
- SSE_ALU_ITINS_P.s, 1>, EVEX_V512, PS, EVEX_CD8<32, CD8VF>;
-
-defm VADDPDZ : avx512_fp_packed<0x58, "addpd", fadd, VK8WM, VR512, v8f64, f512mem,
- memopv8f64, f64mem, loadf64, "{1to8}", SSEPackedDouble,
- SSE_ALU_ITINS_P.d, 1>,
- EVEX_V512, PD, VEX_W, EVEX_CD8<64, CD8VF>;
-
-defm VMULPSZ : avx512_fp_packed<0x59, "mulps", fmul, VK16WM, VR512, v16f32, f512mem,
- memopv16f32, f32mem, loadf32, "{1to16}", SSEPackedSingle,
- SSE_ALU_ITINS_P.s, 1>, EVEX_V512, PS, EVEX_CD8<32, CD8VF>;
-defm VMULPDZ : avx512_fp_packed<0x59, "mulpd", fmul, VK8WM, VR512, v8f64, f512mem,
- memopv8f64, f64mem, loadf64, "{1to8}", SSEPackedDouble,
- SSE_ALU_ITINS_P.d, 1>,
- EVEX_V512, PD, VEX_W, EVEX_CD8<64, CD8VF>;
-
-defm VMINPSZ : avx512_fp_packed<0x5D, "minps", X86fmin, VK16WM, VR512, v16f32, f512mem,
- memopv16f32, f32mem, loadf32, "{1to16}", SSEPackedSingle,
- SSE_ALU_ITINS_P.s, 1>,
- EVEX_V512, PS, EVEX_CD8<32, CD8VF>;
-defm VMAXPSZ : avx512_fp_packed<0x5F, "maxps", X86fmax, VK16WM, VR512, v16f32, f512mem,
- memopv16f32, f32mem, loadf32, "{1to16}", SSEPackedSingle,
- SSE_ALU_ITINS_P.s, 1>,
- EVEX_V512, PS, EVEX_CD8<32, CD8VF>;
-
-defm VMINPDZ : avx512_fp_packed<0x5D, "minpd", X86fmin, VK8WM, VR512, v8f64, f512mem,
- memopv8f64, f64mem, loadf64, "{1to8}", SSEPackedDouble,
- SSE_ALU_ITINS_P.d, 1>,
- EVEX_V512, PD, VEX_W, EVEX_CD8<64, CD8VF>;
-defm VMAXPDZ : avx512_fp_packed<0x5F, "maxpd", X86fmax, VK8WM, VR512, v8f64, f512mem,
- memopv8f64, f64mem, loadf64, "{1to8}", SSEPackedDouble,
- SSE_ALU_ITINS_P.d, 1>,
- EVEX_V512, PD, VEX_W, EVEX_CD8<64, CD8VF>;
-
-defm VSUBPSZ : avx512_fp_packed<0x5C, "subps", fsub, VK16WM, VR512, v16f32, f512mem,
- memopv16f32, f32mem, loadf32, "{1to16}", SSEPackedSingle,
- SSE_ALU_ITINS_P.s, 0>, EVEX_V512, PS, EVEX_CD8<32, CD8VF>;
-defm VDIVPSZ : avx512_fp_packed<0x5E, "divps", fdiv, VK16WM, VR512, v16f32, f512mem,
- memopv16f32, f32mem, loadf32, "{1to16}", SSEPackedSingle,
- SSE_ALU_ITINS_P.s, 0>, EVEX_V512, PS, EVEX_CD8<32, CD8VF>;
-
-defm VSUBPDZ : avx512_fp_packed<0x5C, "subpd", fsub, VK8WM, VR512, v8f64, f512mem,
- memopv8f64, f64mem, loadf64, "{1to8}", SSEPackedDouble,
- SSE_ALU_ITINS_P.d, 0>,
- EVEX_V512, PD, VEX_W, EVEX_CD8<64, CD8VF>;
-defm VDIVPDZ : avx512_fp_packed<0x5E, "divpd", fdiv, VK8WM, VR512, v8f64, f512mem,
- memopv8f64, f64mem, loadf64, "{1to8}", SSEPackedDouble,
- SSE_ALU_ITINS_P.d, 0>,
- EVEX_V512, PD, VEX_W, EVEX_CD8<64, CD8VF>;
+defm VADD : avx512_fp_binop_p<0x58, "vadd", fadd, 1>;
+defm VMUL : avx512_fp_binop_p<0x59, "vmul", fmul, 1>;
+defm VMIN : avx512_fp_binop_p<0x5D, "vmin", X86fmin, 1>;
+defm VMAX : avx512_fp_binop_p<0x5F, "vmax", X86fmax, 1>;
+defm VSUB : avx512_fp_binop_p<0x5C, "vsub", fsub>;
+defm VDIV : avx512_fp_binop_p<0x5E, "vdiv", fdiv>;
def : Pat<(v16f32 (int_x86_avx512_mask_max_ps_512 (v16f32 VR512:$src1),
(v16f32 VR512:$src2), (bc_v16f32 (v16i32 immAllZerosV)),
//===----------------------------------------------------------------------===//
// FMA - Fused Multiply Operations
//
+
let Constraints = "$src1 = $dst" in {
-multiclass avx512_fma3p_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
- X86VectorVTInfo _> {
- defm r: AVX512_masking_3src<opc, MRMSrcReg, _, (outs _.RC:$dst),
+// Omitting the parameter OpNode (= null_frag) disables ISel pattern matching.
+multiclass avx512_fma3p_rm<bits<8> opc, string OpcodeStr, X86VectorVTInfo _,
+ SDPatternOperator OpNode = null_frag> {
+ defm r: AVX512_maskable_3src<opc, MRMSrcReg, _, (outs _.RC:$dst),
(ins _.RC:$src2, _.RC:$src3),
OpcodeStr, "$src3, $src2", "$src2, $src3",
(_.VT (OpNode _.RC:$src1, _.RC:$src2, _.RC:$src3))>,
}
} // Constraints = "$src1 = $dst"
+multiclass avx512_fma3p_forms<bits<8> opc213, bits<8> opc231,
+ string OpcodeStr, X86VectorVTInfo VTI,
+ SDPatternOperator OpNode> {
+ defm v213 : avx512_fma3p_rm<opc213, !strconcat(OpcodeStr, "213", VTI.Suffix),
+ VTI, OpNode>,
+ EVEX_V512, EVEX_CD8<VTI.EltSize, CD8VF>;
+
+ defm v231 : avx512_fma3p_rm<opc231, !strconcat(OpcodeStr, "231", VTI.Suffix),
+ VTI>,
+ EVEX_V512, EVEX_CD8<VTI.EltSize, CD8VF>;
+}
+
let ExeDomain = SSEPackedSingle in {
- 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>;
+ defm VFMADDPSZ : avx512_fma3p_forms<0xA8, 0xB8, "vfmadd",
+ v16f32_info, X86Fmadd>;
+ defm VFMSUBPSZ : avx512_fma3p_forms<0xAA, 0xBA, "vfmsub",
+ v16f32_info, X86Fmsub>;
+ defm VFMADDSUBPSZ : avx512_fma3p_forms<0xA6, 0xB6, "vfmaddsub",
+ v16f32_info, X86Fmaddsub>;
+ defm VFMSUBADDPSZ : avx512_fma3p_forms<0xA7, 0xB7, "vfmsubadd",
+ v16f32_info, X86Fmsubadd>;
+ defm VFNMADDPSZ : avx512_fma3p_forms<0xAC, 0xBC, "vfnmadd",
+ v16f32_info, X86Fnmadd>;
+ defm VFNMSUBPSZ : avx512_fma3p_forms<0xAE, 0xBE, "vfnmsub",
+ v16f32_info, X86Fnmsub>;
}
let ExeDomain = SSEPackedDouble in {
- 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 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 VFMADDPDZ : avx512_fma3p_forms<0xA8, 0xB8, "vfmadd",
+ v8f64_info, X86Fmadd>, VEX_W;
+ defm VFMSUBPDZ : avx512_fma3p_forms<0xAA, 0xBA, "vfmsub",
+ v8f64_info, X86Fmsub>, VEX_W;
+ defm VFMADDSUBPDZ : avx512_fma3p_forms<0xA6, 0xB6, "vfmaddsub",
+ v8f64_info, X86Fmaddsub>, VEX_W;
+ defm VFMSUBADDPDZ : avx512_fma3p_forms<0xA7, 0xB7, "vfmsubadd",
+ v8f64_info, X86Fmsubadd>, VEX_W;
+ defm VFNMADDPDZ : avx512_fma3p_forms<0xAC, 0xBC, "vfnmadd",
+ v8f64_info, X86Fnmadd>, VEX_W;
+ defm VFNMSUBPDZ : avx512_fma3p_forms<0xAE, 0xBE, "vfnmsub",
+ v8f64_info, X86Fnmsub>, VEX_W;
}
let Constraints = "$src1 = $dst" in {
/// avx512_fp14_p rcp14ps, rcp14pd, rsqrt14ps, rsqrt14pd
multiclass avx512_fp14_p<bits<8> opc, string OpcodeStr, SDNode OpNode,
- RegisterClass RC, X86MemOperand x86memop,
- PatFrag mem_frag, ValueType OpVt> {
- def r : AVX5128I<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
- !strconcat(OpcodeStr,
- " \t{$src, $dst|$dst, $src}"),
- [(set RC:$dst, (OpVt (OpNode RC:$src)))]>,
- EVEX;
- def m : AVX5128I<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
- !strconcat(OpcodeStr, " \t{$src, $dst|$dst, $src}"),
- [(set RC:$dst, (OpVt (OpNode (mem_frag addr:$src))))]>,
- EVEX;
-}
-defm VRSQRT14PSZ : avx512_fp14_p<0x4E, "vrsqrt14ps", X86frsqrt, VR512, f512mem,
- memopv16f32, v16f32>, EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VRSQRT14PDZ : avx512_fp14_p<0x4E, "vrsqrt14pd", X86frsqrt, VR512, f512mem,
- memopv8f64, v8f64>, VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>;
-defm VRCP14PSZ : avx512_fp14_p<0x4C, "vrcp14ps", X86frcp, VR512, f512mem,
- memopv16f32, v16f32>, EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VRCP14PDZ : avx512_fp14_p<0x4C, "vrcp14pd", X86frcp, VR512, f512mem,
- memopv8f64, v8f64>, VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>;
+ X86VectorVTInfo _> {
+ defm r: AVX512_maskable<opc, MRMSrcReg, _, (outs _.RC:$dst),
+ (ins _.RC:$src), OpcodeStr, "$src", "$src",
+ (_.FloatVT (OpNode _.RC:$src))>, EVEX, T8PD;
+ let mayLoad = 1 in {
+ defm m: AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst),
+ (ins _.MemOp:$src), OpcodeStr, "$src", "$src",
+ (OpNode (_.FloatVT
+ (bitconvert (_.LdFrag addr:$src))))>, EVEX, T8PD;
+ defm mb: AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst),
+ (ins _.ScalarMemOp:$src), OpcodeStr,
+ "${src}"##_.BroadcastStr, "${src}"##_.BroadcastStr,
+ (OpNode (_.FloatVT
+ (X86VBroadcast (_.ScalarLdFrag addr:$src))))>,
+ EVEX, T8PD, EVEX_B;
+ }
+}
+
+multiclass avx512_fp14_p_vl_all<bits<8> opc, string OpcodeStr, SDNode OpNode> {
+ defm PSZ : avx512_fp14_p<opc, !strconcat(OpcodeStr, "ps"), OpNode, v16f32_info>,
+ EVEX_V512, EVEX_CD8<32, CD8VF>;
+ defm PDZ : avx512_fp14_p<opc, !strconcat(OpcodeStr, "pd"), OpNode, v8f64_info>,
+ EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+
+ // Define only if AVX512VL feature is present.
+ let Predicates = [HasVLX] in {
+ defm PSZ128 : avx512_fp14_p<opc, !strconcat(OpcodeStr, "ps"),
+ OpNode, v4f32x_info>,
+ EVEX_V128, EVEX_CD8<32, CD8VF>;
+ defm PSZ256 : avx512_fp14_p<opc, !strconcat(OpcodeStr, "ps"),
+ OpNode, v8f32x_info>,
+ EVEX_V256, EVEX_CD8<32, CD8VF>;
+ defm PDZ128 : avx512_fp14_p<opc, !strconcat(OpcodeStr, "pd"),
+ OpNode, v2f64x_info>,
+ EVEX_V128, VEX_W, EVEX_CD8<64, CD8VF>;
+ defm PDZ256 : avx512_fp14_p<opc, !strconcat(OpcodeStr, "pd"),
+ OpNode, v4f64x_info>,
+ EVEX_V256, VEX_W, EVEX_CD8<64, CD8VF>;
+ }
+}
+
+defm VRSQRT14 : avx512_fp14_p_vl_all<0x4E, "vrsqrt14", X86frsqrt>;
+defm VRCP14 : avx512_fp14_p_vl_all<0x4C, "vrcp14", X86frcp>;
def : Pat <(v16f32 (int_x86_avx512_rsqrt14_ps_512 (v16f32 VR512:$src),
(bc_v16f32 (v16i32 immAllZerosV)), (i16 -1))),
(bc_v8f64 (v16i32 immAllZerosV)), (i8 -1), FROUND_NO_EXC)),
(VRCP28PDZrb VR512:$src)>;
-multiclass avx512_sqrt_packed<bits<8> opc, string OpcodeStr, SDNode OpNode,
- OpndItins itins_s, OpndItins itins_d> {
- def PSZrr :AVX512PSI<opc, MRMSrcReg, (outs VR512:$dst), (ins VR512:$src),
- !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
- [(set VR512:$dst, (v16f32 (OpNode VR512:$src)))], itins_s.rr>,
- EVEX, EVEX_V512;
-
- let mayLoad = 1 in
- def PSZrm : AVX512PSI<opc, MRMSrcMem, (outs VR512:$dst), (ins f512mem:$src),
- !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
- [(set VR512:$dst,
- (OpNode (v16f32 (bitconvert (memopv16f32 addr:$src)))))],
- itins_s.rm>, EVEX, EVEX_V512, EVEX_CD8<32, CD8VF>;
-
- def PDZrr : AVX512PDI<opc, MRMSrcReg, (outs VR512:$dst), (ins VR512:$src),
- !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
- [(set VR512:$dst, (v8f64 (OpNode VR512:$src)))], itins_d.rr>,
- EVEX, EVEX_V512;
-
- let mayLoad = 1 in
- def PDZrm : AVX512PDI<opc, MRMSrcMem, (outs VR512:$dst), (ins f512mem:$src),
- !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
- [(set VR512:$dst, (OpNode
- (v8f64 (bitconvert (memopv16f32 addr:$src)))))],
- itins_d.rm>, EVEX, EVEX_V512, EVEX_CD8<64, CD8VF>;
-
+multiclass avx512_sqrt_packed<bits<8> opc, string OpcodeStr,
+ SDNode OpNode, X86VectorVTInfo _>{
+ defm r: AVX512_maskable<opc, MRMSrcReg, _, (outs _.RC:$dst),
+ (ins _.RC:$src), OpcodeStr, "$src", "$src",
+ (_.FloatVT (OpNode _.RC:$src))>, EVEX;
+ let mayLoad = 1 in {
+ defm m: AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst),
+ (ins _.MemOp:$src), OpcodeStr, "$src", "$src",
+ (OpNode (_.FloatVT
+ (bitconvert (_.LdFrag addr:$src))))>, EVEX;
+
+ defm mb: AVX512_maskable<opc, MRMSrcMem, _, (outs _.RC:$dst),
+ (ins _.ScalarMemOp:$src), OpcodeStr,
+ "${src}"##_.BroadcastStr, "${src}"##_.BroadcastStr,
+ (OpNode (_.FloatVT
+ (X86VBroadcast (_.ScalarLdFrag addr:$src))))>,
+ EVEX, EVEX_B;
+ }
}
multiclass avx512_sqrt_scalar<bits<8> opc, string OpcodeStr,
}
}
+multiclass avx512_sqrt_packed_all<bits<8> opc, string OpcodeStr,
+ SDNode OpNode> {
+ defm PSZ : avx512_sqrt_packed<opc, !strconcat(OpcodeStr, "ps"), OpNode,
+ v16f32_info>,
+ EVEX_V512, PS, EVEX_CD8<32, CD8VF>;
+ defm PDZ : avx512_sqrt_packed<opc, !strconcat(OpcodeStr, "pd"), OpNode,
+ v8f64_info>,
+ EVEX_V512, VEX_W, PD, EVEX_CD8<64, CD8VF>;
+ // Define only if AVX512VL feature is present.
+ let Predicates = [HasVLX] in {
+ defm PSZ128 : avx512_sqrt_packed<opc, !strconcat(OpcodeStr, "ps"),
+ OpNode, v4f32x_info>,
+ EVEX_V128, PS, EVEX_CD8<32, CD8VF>;
+ defm PSZ256 : avx512_sqrt_packed<opc, !strconcat(OpcodeStr, "ps"),
+ OpNode, v8f32x_info>,
+ EVEX_V256, PS, EVEX_CD8<32, CD8VF>;
+ defm PDZ128 : avx512_sqrt_packed<opc, !strconcat(OpcodeStr, "pd"),
+ OpNode, v2f64x_info>,
+ EVEX_V128, VEX_W, PD, EVEX_CD8<64, CD8VF>;
+ defm PDZ256 : avx512_sqrt_packed<opc, !strconcat(OpcodeStr, "pd"),
+ OpNode, v4f64x_info>,
+ EVEX_V256, VEX_W, PD, EVEX_CD8<64, CD8VF>;
+ }
+}
+
+defm VSQRT : avx512_sqrt_packed_all<0x51, "vsqrt", fsqrt>;
defm VSQRT : avx512_sqrt_scalar<0x51, "sqrt",
int_x86_avx512_sqrt_ss, int_x86_avx512_sqrt_sd,
- SSE_SQRTSS, SSE_SQRTSD>,
- avx512_sqrt_packed<0x51, "vsqrt", fsqrt,
- SSE_SQRTPS, SSE_SQRTPD>;
+ SSE_SQRTSS, SSE_SQRTSD>;
let Predicates = [HasAVX512] in {
def : Pat<(v16f32 (int_x86_avx512_sqrt_ps_512 (v16f32 VR512:$src1),
(bc_v16f32 (v16i32 immAllZerosV)), (i16 -1), FROUND_CURRENT)),
- (VSQRTPSZrr VR512:$src1)>;
+ (VSQRTPSZr VR512:$src1)>;
def : Pat<(v8f64 (int_x86_avx512_sqrt_pd_512 (v8f64 VR512:$src1),
(bc_v8f64 (v16i32 immAllZerosV)), (i8 -1), FROUND_CURRENT)),
- (VSQRTPDZrr VR512:$src1)>;
+ (VSQRTPDZr VR512:$src1)>;
def : Pat<(f32 (fsqrt FR32X:$src)),
(VSQRTSSZr (f32 (IMPLICIT_DEF)), FR32X:$src)>;
(VSHUFPDZrmi VR512:$src1, addr:$src2, imm:$imm)>;
multiclass avx512_valign<X86VectorVTInfo _> {
- defm rri : AVX512_masking<0x03, MRMSrcReg, _, (outs _.RC:$dst),
+ defm rri : AVX512_maskable<0x03, MRMSrcReg, _, (outs _.RC:$dst),
(ins _.RC:$src1, _.RC:$src2, i8imm:$src3),
"valign"##_.Suffix,
"$src3, $src2, $src1", "$src1, $src2, $src3",