1 // Bitcasts between 512-bit vector types. Return the original type since
2 // no instruction is needed for the conversion
3 let Predicates = [HasAVX512] in {
4 def : Pat<(v8f64 (bitconvert (v16f32 VR512:$src))), (v8f64 VR512:$src)>;
5 def : Pat<(v8f64 (bitconvert (v16i32 VR512:$src))), (v8f64 VR512:$src)>;
6 def : Pat<(v8f64 (bitconvert (v8i64 VR512:$src))), (v8f64 VR512:$src)>;
7 def : Pat<(v16f32 (bitconvert (v16i32 VR512:$src))), (v16f32 VR512:$src)>;
8 def : Pat<(v16f32 (bitconvert (v8i64 VR512:$src))), (v16f32 VR512:$src)>;
9 def : Pat<(v16f32 (bitconvert (v8f64 VR512:$src))), (v16f32 VR512:$src)>;
10 def : Pat<(v8i64 (bitconvert (v16f32 VR512:$src))), (v8i64 VR512:$src)>;
11 def : Pat<(v8i64 (bitconvert (v16i32 VR512:$src))), (v8i64 VR512:$src)>;
12 def : Pat<(v8i64 (bitconvert (v8f64 VR512:$src))), (v8i64 VR512:$src)>;
13 def : Pat<(v16i32 (bitconvert (v16f32 VR512:$src))), (v16i32 VR512:$src)>;
14 def : Pat<(v16i32 (bitconvert (v8i64 VR512:$src))), (v16i32 VR512:$src)>;
15 def : Pat<(v16i32 (bitconvert (v8f64 VR512:$src))), (v16i32 VR512:$src)>;
16 def : Pat<(v8f64 (bitconvert (v8i64 VR512:$src))), (v8f64 VR512:$src)>;
18 def : Pat<(v2i64 (bitconvert (v4i32 VR128X:$src))), (v2i64 VR128X:$src)>;
19 def : Pat<(v2i64 (bitconvert (v8i16 VR128X:$src))), (v2i64 VR128X:$src)>;
20 def : Pat<(v2i64 (bitconvert (v16i8 VR128X:$src))), (v2i64 VR128X:$src)>;
21 def : Pat<(v2i64 (bitconvert (v2f64 VR128X:$src))), (v2i64 VR128X:$src)>;
22 def : Pat<(v2i64 (bitconvert (v4f32 VR128X:$src))), (v2i64 VR128X:$src)>;
23 def : Pat<(v4i32 (bitconvert (v2i64 VR128X:$src))), (v4i32 VR128X:$src)>;
24 def : Pat<(v4i32 (bitconvert (v8i16 VR128X:$src))), (v4i32 VR128X:$src)>;
25 def : Pat<(v4i32 (bitconvert (v16i8 VR128X:$src))), (v4i32 VR128X:$src)>;
26 def : Pat<(v4i32 (bitconvert (v2f64 VR128X:$src))), (v4i32 VR128X:$src)>;
27 def : Pat<(v4i32 (bitconvert (v4f32 VR128X:$src))), (v4i32 VR128X:$src)>;
28 def : Pat<(v8i16 (bitconvert (v2i64 VR128X:$src))), (v8i16 VR128X:$src)>;
29 def : Pat<(v8i16 (bitconvert (v4i32 VR128X:$src))), (v8i16 VR128X:$src)>;
30 def : Pat<(v8i16 (bitconvert (v16i8 VR128X:$src))), (v8i16 VR128X:$src)>;
31 def : Pat<(v8i16 (bitconvert (v2f64 VR128X:$src))), (v8i16 VR128X:$src)>;
32 def : Pat<(v8i16 (bitconvert (v4f32 VR128X:$src))), (v8i16 VR128X:$src)>;
33 def : Pat<(v16i8 (bitconvert (v2i64 VR128X:$src))), (v16i8 VR128X:$src)>;
34 def : Pat<(v16i8 (bitconvert (v4i32 VR128X:$src))), (v16i8 VR128X:$src)>;
35 def : Pat<(v16i8 (bitconvert (v8i16 VR128X:$src))), (v16i8 VR128X:$src)>;
36 def : Pat<(v16i8 (bitconvert (v2f64 VR128X:$src))), (v16i8 VR128X:$src)>;
37 def : Pat<(v16i8 (bitconvert (v4f32 VR128X:$src))), (v16i8 VR128X:$src)>;
38 def : Pat<(v4f32 (bitconvert (v2i64 VR128X:$src))), (v4f32 VR128X:$src)>;
39 def : Pat<(v4f32 (bitconvert (v4i32 VR128X:$src))), (v4f32 VR128X:$src)>;
40 def : Pat<(v4f32 (bitconvert (v8i16 VR128X:$src))), (v4f32 VR128X:$src)>;
41 def : Pat<(v4f32 (bitconvert (v16i8 VR128X:$src))), (v4f32 VR128X:$src)>;
42 def : Pat<(v4f32 (bitconvert (v2f64 VR128X:$src))), (v4f32 VR128X:$src)>;
43 def : Pat<(v2f64 (bitconvert (v2i64 VR128X:$src))), (v2f64 VR128X:$src)>;
44 def : Pat<(v2f64 (bitconvert (v4i32 VR128X:$src))), (v2f64 VR128X:$src)>;
45 def : Pat<(v2f64 (bitconvert (v8i16 VR128X:$src))), (v2f64 VR128X:$src)>;
46 def : Pat<(v2f64 (bitconvert (v16i8 VR128X:$src))), (v2f64 VR128X:$src)>;
47 def : Pat<(v2f64 (bitconvert (v4f32 VR128X:$src))), (v2f64 VR128X:$src)>;
49 // Bitcasts between 256-bit vector types. Return the original type since
50 // no instruction is needed for the conversion
51 def : Pat<(v4f64 (bitconvert (v8f32 VR256X:$src))), (v4f64 VR256X:$src)>;
52 def : Pat<(v4f64 (bitconvert (v8i32 VR256X:$src))), (v4f64 VR256X:$src)>;
53 def : Pat<(v4f64 (bitconvert (v4i64 VR256X:$src))), (v4f64 VR256X:$src)>;
54 def : Pat<(v4f64 (bitconvert (v16i16 VR256X:$src))), (v4f64 VR256X:$src)>;
55 def : Pat<(v4f64 (bitconvert (v32i8 VR256X:$src))), (v4f64 VR256X:$src)>;
56 def : Pat<(v8f32 (bitconvert (v8i32 VR256X:$src))), (v8f32 VR256X:$src)>;
57 def : Pat<(v8f32 (bitconvert (v4i64 VR256X:$src))), (v8f32 VR256X:$src)>;
58 def : Pat<(v8f32 (bitconvert (v4f64 VR256X:$src))), (v8f32 VR256X:$src)>;
59 def : Pat<(v8f32 (bitconvert (v32i8 VR256X:$src))), (v8f32 VR256X:$src)>;
60 def : Pat<(v8f32 (bitconvert (v16i16 VR256X:$src))), (v8f32 VR256X:$src)>;
61 def : Pat<(v4i64 (bitconvert (v8f32 VR256X:$src))), (v4i64 VR256X:$src)>;
62 def : Pat<(v4i64 (bitconvert (v8i32 VR256X:$src))), (v4i64 VR256X:$src)>;
63 def : Pat<(v4i64 (bitconvert (v4f64 VR256X:$src))), (v4i64 VR256X:$src)>;
64 def : Pat<(v4i64 (bitconvert (v32i8 VR256X:$src))), (v4i64 VR256X:$src)>;
65 def : Pat<(v4i64 (bitconvert (v16i16 VR256X:$src))), (v4i64 VR256X:$src)>;
66 def : Pat<(v32i8 (bitconvert (v4f64 VR256X:$src))), (v32i8 VR256X:$src)>;
67 def : Pat<(v32i8 (bitconvert (v4i64 VR256X:$src))), (v32i8 VR256X:$src)>;
68 def : Pat<(v32i8 (bitconvert (v8f32 VR256X:$src))), (v32i8 VR256X:$src)>;
69 def : Pat<(v32i8 (bitconvert (v8i32 VR256X:$src))), (v32i8 VR256X:$src)>;
70 def : Pat<(v32i8 (bitconvert (v16i16 VR256X:$src))), (v32i8 VR256X:$src)>;
71 def : Pat<(v8i32 (bitconvert (v32i8 VR256X:$src))), (v8i32 VR256X:$src)>;
72 def : Pat<(v8i32 (bitconvert (v16i16 VR256X:$src))), (v8i32 VR256X:$src)>;
73 def : Pat<(v8i32 (bitconvert (v8f32 VR256X:$src))), (v8i32 VR256X:$src)>;
74 def : Pat<(v8i32 (bitconvert (v4i64 VR256X:$src))), (v8i32 VR256X:$src)>;
75 def : Pat<(v8i32 (bitconvert (v4f64 VR256X:$src))), (v8i32 VR256X:$src)>;
76 def : Pat<(v16i16 (bitconvert (v8f32 VR256X:$src))), (v16i16 VR256X:$src)>;
77 def : Pat<(v16i16 (bitconvert (v8i32 VR256X:$src))), (v16i16 VR256X:$src)>;
78 def : Pat<(v16i16 (bitconvert (v4i64 VR256X:$src))), (v16i16 VR256X:$src)>;
79 def : Pat<(v16i16 (bitconvert (v4f64 VR256X:$src))), (v16i16 VR256X:$src)>;
80 def : Pat<(v16i16 (bitconvert (v32i8 VR256X:$src))), (v16i16 VR256X:$src)>;
84 // AVX-512: VPXOR instruction writes zero to its upper part, it's safe build zeros.
87 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
88 isPseudo = 1, Predicates = [HasAVX512] in {
89 def AVX512_512_SET0 : I<0, Pseudo, (outs VR512:$dst), (ins), "",
90 [(set VR512:$dst, (v16f32 immAllZerosV))]>;
93 def : Pat<(v8i64 immAllZerosV), (AVX512_512_SET0)>;
94 def : Pat<(v16i32 immAllZerosV), (AVX512_512_SET0)>;
95 def : Pat<(v8f64 immAllZerosV), (AVX512_512_SET0)>;
96 def : Pat<(v16f32 immAllZerosV), (AVX512_512_SET0)>;
98 //===----------------------------------------------------------------------===//
99 // AVX-512 - VECTOR INSERT
102 let hasSideEffects = 0, ExeDomain = SSEPackedSingle in {
103 def VINSERTF32x4rr : AVX512AIi8<0x18, MRMSrcReg, (outs VR512:$dst),
104 (ins VR512:$src1, VR128X:$src2, i8imm:$src3),
105 "vinsertf32x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
106 []>, EVEX_4V, EVEX_V512;
108 def VINSERTF32x4rm : AVX512AIi8<0x18, MRMSrcMem, (outs VR512:$dst),
109 (ins VR512:$src1, f128mem:$src2, i8imm:$src3),
110 "vinsertf32x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
111 []>, EVEX_4V, EVEX_V512, EVEX_CD8<32, CD8VT4>;
114 // -- 64x4 fp form --
115 let hasSideEffects = 0, ExeDomain = SSEPackedDouble in {
116 def VINSERTF64x4rr : AVX512AIi8<0x1a, MRMSrcReg, (outs VR512:$dst),
117 (ins VR512:$src1, VR256X:$src2, i8imm:$src3),
118 "vinsertf64x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
119 []>, EVEX_4V, EVEX_V512, VEX_W;
121 def VINSERTF64x4rm : AVX512AIi8<0x1a, MRMSrcMem, (outs VR512:$dst),
122 (ins VR512:$src1, i256mem:$src2, i8imm:$src3),
123 "vinsertf64x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
124 []>, EVEX_4V, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT4>;
126 // -- 32x4 integer form --
127 let hasSideEffects = 0 in {
128 def VINSERTI32x4rr : AVX512AIi8<0x38, MRMSrcReg, (outs VR512:$dst),
129 (ins VR512:$src1, VR128X:$src2, i8imm:$src3),
130 "vinserti32x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
131 []>, EVEX_4V, EVEX_V512;
133 def VINSERTI32x4rm : AVX512AIi8<0x38, MRMSrcMem, (outs VR512:$dst),
134 (ins VR512:$src1, i128mem:$src2, i8imm:$src3),
135 "vinserti32x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
136 []>, EVEX_4V, EVEX_V512, EVEX_CD8<32, CD8VT4>;
140 let hasSideEffects = 0 in {
142 def VINSERTI64x4rr : AVX512AIi8<0x3a, MRMSrcReg, (outs VR512:$dst),
143 (ins VR512:$src1, VR256X:$src2, i8imm:$src3),
144 "vinserti64x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
145 []>, EVEX_4V, EVEX_V512, VEX_W;
147 def VINSERTI64x4rm : AVX512AIi8<0x3a, MRMSrcMem, (outs VR512:$dst),
148 (ins VR512:$src1, i256mem:$src2, i8imm:$src3),
149 "vinserti64x4\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
150 []>, EVEX_4V, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT4>;
153 def : Pat<(vinsert128_insert:$ins (v16f32 VR512:$src1), (v4f32 VR128X:$src2),
154 (iPTR imm)), (VINSERTF32x4rr VR512:$src1, VR128X:$src2,
155 (INSERT_get_vinsert128_imm VR512:$ins))>;
156 def : Pat<(vinsert128_insert:$ins (v8f64 VR512:$src1), (v2f64 VR128X:$src2),
157 (iPTR imm)), (VINSERTF32x4rr VR512:$src1, VR128X:$src2,
158 (INSERT_get_vinsert128_imm VR512:$ins))>;
159 def : Pat<(vinsert128_insert:$ins (v8i64 VR512:$src1), (v2i64 VR128X:$src2),
160 (iPTR imm)), (VINSERTI32x4rr VR512:$src1, VR128X:$src2,
161 (INSERT_get_vinsert128_imm VR512:$ins))>;
162 def : Pat<(vinsert128_insert:$ins (v16i32 VR512:$src1), (v4i32 VR128X:$src2),
163 (iPTR imm)), (VINSERTI32x4rr VR512:$src1, VR128X:$src2,
164 (INSERT_get_vinsert128_imm VR512:$ins))>;
166 def : Pat<(vinsert128_insert:$ins (v16f32 VR512:$src1), (loadv4f32 addr:$src2),
167 (iPTR imm)), (VINSERTF32x4rm VR512:$src1, addr:$src2,
168 (INSERT_get_vinsert128_imm VR512:$ins))>;
169 def : Pat<(vinsert128_insert:$ins (v16i32 VR512:$src1),
170 (bc_v4i32 (loadv2i64 addr:$src2)),
171 (iPTR imm)), (VINSERTI32x4rm VR512:$src1, addr:$src2,
172 (INSERT_get_vinsert128_imm VR512:$ins))>;
173 def : Pat<(vinsert128_insert:$ins (v8f64 VR512:$src1), (loadv2f64 addr:$src2),
174 (iPTR imm)), (VINSERTF32x4rm VR512:$src1, addr:$src2,
175 (INSERT_get_vinsert128_imm VR512:$ins))>;
176 def : Pat<(vinsert128_insert:$ins (v8i64 VR512:$src1), (loadv2i64 addr:$src2),
177 (iPTR imm)), (VINSERTI32x4rm VR512:$src1, addr:$src2,
178 (INSERT_get_vinsert128_imm VR512:$ins))>;
180 def : Pat<(vinsert256_insert:$ins (v16f32 VR512:$src1), (v8f32 VR256X:$src2),
181 (iPTR imm)), (VINSERTF64x4rr VR512:$src1, VR256X:$src2,
182 (INSERT_get_vinsert256_imm VR512:$ins))>;
183 def : Pat<(vinsert256_insert:$ins (v8f64 VR512:$src1), (v4f64 VR256X:$src2),
184 (iPTR imm)), (VINSERTF64x4rr VR512:$src1, VR256X:$src2,
185 (INSERT_get_vinsert256_imm VR512:$ins))>;
186 def : Pat<(vinsert128_insert:$ins (v8i64 VR512:$src1), (v4i64 VR256X:$src2),
187 (iPTR imm)), (VINSERTI64x4rr VR512:$src1, VR256X:$src2,
188 (INSERT_get_vinsert256_imm VR512:$ins))>;
189 def : Pat<(vinsert128_insert:$ins (v16i32 VR512:$src1), (v8i32 VR256X:$src2),
190 (iPTR imm)), (VINSERTI64x4rr VR512:$src1, VR256X:$src2,
191 (INSERT_get_vinsert256_imm VR512:$ins))>;
193 def : Pat<(vinsert256_insert:$ins (v16f32 VR512:$src1), (loadv8f32 addr:$src2),
194 (iPTR imm)), (VINSERTF64x4rm VR512:$src1, addr:$src2,
195 (INSERT_get_vinsert256_imm VR512:$ins))>;
196 def : Pat<(vinsert256_insert:$ins (v8f64 VR512:$src1), (loadv4f64 addr:$src2),
197 (iPTR imm)), (VINSERTF64x4rm VR512:$src1, addr:$src2,
198 (INSERT_get_vinsert256_imm VR512:$ins))>;
199 def : Pat<(vinsert256_insert:$ins (v8i64 VR512:$src1), (loadv4i64 addr:$src2),
200 (iPTR imm)), (VINSERTI64x4rm VR512:$src1, addr:$src2,
201 (INSERT_get_vinsert256_imm VR512:$ins))>;
202 def : Pat<(vinsert256_insert:$ins (v16i32 VR512:$src1),
203 (bc_v8i32 (loadv4i64 addr:$src2)),
204 (iPTR imm)), (VINSERTI64x4rm VR512:$src1, addr:$src2,
205 (INSERT_get_vinsert256_imm VR512:$ins))>;
207 // vinsertps - insert f32 to XMM
208 def VINSERTPSzrr : AVX512AIi8<0x21, MRMSrcReg, (outs VR128X:$dst),
209 (ins VR128X:$src1, VR128X:$src2, u32u8imm:$src3),
210 "vinsertps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
211 [(set VR128X:$dst, (X86insrtps VR128X:$src1, VR128X:$src2, imm:$src3))]>,
213 def VINSERTPSzrm: AVX512AIi8<0x21, MRMSrcMem, (outs VR128X:$dst),
214 (ins VR128X:$src1, f32mem:$src2, u32u8imm:$src3),
215 "vinsertps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
216 [(set VR128X:$dst, (X86insrtps VR128X:$src1,
217 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
218 imm:$src3))]>, EVEX_4V, EVEX_CD8<32, CD8VT1>;
220 //===----------------------------------------------------------------------===//
221 // AVX-512 VECTOR EXTRACT
223 let hasSideEffects = 0, ExeDomain = SSEPackedSingle in {
225 def VEXTRACTF32x4rr : AVX512AIi8<0x19, MRMDestReg, (outs VR128X:$dst),
226 (ins VR512:$src1, i8imm:$src2),
227 "vextractf32x4\t{$src2, $src1, $dst|$dst, $src1, $src2}",
228 []>, EVEX, EVEX_V512;
229 def VEXTRACTF32x4mr : AVX512AIi8<0x19, MRMDestMem, (outs),
230 (ins f128mem:$dst, VR512:$src1, i8imm:$src2),
231 "vextractf32x4\t{$src2, $src1, $dst|$dst, $src1, $src2}",
232 []>, EVEX, EVEX_V512, EVEX_CD8<32, CD8VT4>;
235 def VEXTRACTF64x4rr : AVX512AIi8<0x1b, MRMDestReg, (outs VR256X:$dst),
236 (ins VR512:$src1, i8imm:$src2),
237 "vextractf64x4\t{$src2, $src1, $dst|$dst, $src1, $src2}",
238 []>, EVEX, EVEX_V512, VEX_W;
240 def VEXTRACTF64x4mr : AVX512AIi8<0x1b, MRMDestMem, (outs),
241 (ins f256mem:$dst, VR512:$src1, i8imm:$src2),
242 "vextractf64x4\t{$src2, $src1, $dst|$dst, $src1, $src2}",
243 []>, EVEX, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT4>;
246 let hasSideEffects = 0 in {
248 def VEXTRACTI32x4rr : AVX512AIi8<0x39, MRMDestReg, (outs VR128X:$dst),
249 (ins VR512:$src1, i8imm:$src2),
250 "vextracti32x4\t{$src2, $src1, $dst|$dst, $src1, $src2}",
251 []>, EVEX, EVEX_V512;
252 def VEXTRACTI32x4mr : AVX512AIi8<0x39, MRMDestMem, (outs),
253 (ins i128mem:$dst, VR512:$src1, i8imm:$src2),
254 "vextracti32x4\t{$src2, $src1, $dst|$dst, $src1, $src2}",
255 []>, EVEX, EVEX_V512, EVEX_CD8<32, CD8VT4>;
258 def VEXTRACTI64x4rr : AVX512AIi8<0x3b, MRMDestReg, (outs VR256X:$dst),
259 (ins VR512:$src1, i8imm:$src2),
260 "vextracti64x4\t{$src2, $src1, $dst|$dst, $src1, $src2}",
261 []>, EVEX, EVEX_V512, VEX_W;
263 def VEXTRACTI64x4mr : AVX512AIi8<0x3b, MRMDestMem, (outs),
264 (ins i256mem:$dst, VR512:$src1, i8imm:$src2),
265 "vextracti64x4\t{$src2, $src1, $dst|$dst, $src1, $src2}",
266 []>, EVEX, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT4>;
269 def : Pat<(vextract128_extract:$ext (v16f32 VR512:$src1), (iPTR imm)),
270 (v4f32 (VEXTRACTF32x4rr VR512:$src1,
271 (EXTRACT_get_vextract128_imm VR128X:$ext)))>;
273 def : Pat<(vextract128_extract:$ext VR512:$src1, (iPTR imm)),
274 (v4i32 (VEXTRACTF32x4rr VR512:$src1,
275 (EXTRACT_get_vextract128_imm VR128X:$ext)))>;
277 def : Pat<(vextract128_extract:$ext (v8f64 VR512:$src1), (iPTR imm)),
278 (v2f64 (VEXTRACTF32x4rr VR512:$src1,
279 (EXTRACT_get_vextract128_imm VR128X:$ext)))>;
281 def : Pat<(vextract128_extract:$ext (v8i64 VR512:$src1), (iPTR imm)),
282 (v2i64 (VEXTRACTI32x4rr VR512:$src1,
283 (EXTRACT_get_vextract128_imm VR128X:$ext)))>;
286 def : Pat<(vextract256_extract:$ext (v16f32 VR512:$src1), (iPTR imm)),
287 (v8f32 (VEXTRACTF64x4rr VR512:$src1,
288 (EXTRACT_get_vextract256_imm VR256X:$ext)))>;
290 def : Pat<(vextract256_extract:$ext (v16i32 VR512:$src1), (iPTR imm)),
291 (v8i32 (VEXTRACTI64x4rr VR512:$src1,
292 (EXTRACT_get_vextract256_imm VR256X:$ext)))>;
294 def : Pat<(vextract256_extract:$ext (v8f64 VR512:$src1), (iPTR imm)),
295 (v4f64 (VEXTRACTF64x4rr VR512:$src1,
296 (EXTRACT_get_vextract256_imm VR256X:$ext)))>;
298 def : Pat<(vextract256_extract:$ext (v8i64 VR512:$src1), (iPTR imm)),
299 (v4i64 (VEXTRACTI64x4rr VR512:$src1,
300 (EXTRACT_get_vextract256_imm VR256X:$ext)))>;
302 // A 256-bit subvector extract from the first 512-bit vector position
303 // is a subregister copy that needs no instruction.
304 def : Pat<(v8i32 (extract_subvector (v16i32 VR512:$src), (iPTR 0))),
305 (v8i32 (EXTRACT_SUBREG (v16i32 VR512:$src), sub_ymm))>;
306 def : Pat<(v8f32 (extract_subvector (v16f32 VR512:$src), (iPTR 0))),
307 (v8f32 (EXTRACT_SUBREG (v16f32 VR512:$src), sub_ymm))>;
308 def : Pat<(v4i64 (extract_subvector (v8i64 VR512:$src), (iPTR 0))),
309 (v4i64 (EXTRACT_SUBREG (v8i64 VR512:$src), sub_ymm))>;
310 def : Pat<(v4f64 (extract_subvector (v8f64 VR512:$src), (iPTR 0))),
311 (v4f64 (EXTRACT_SUBREG (v8f64 VR512:$src), sub_ymm))>;
314 def : Pat<(v4i32 (extract_subvector (v16i32 VR512:$src), (iPTR 0))),
315 (v4i32 (EXTRACT_SUBREG (v16i32 VR512:$src), sub_xmm))>;
316 def : Pat<(v2i64 (extract_subvector (v8i64 VR512:$src), (iPTR 0))),
317 (v2i64 (EXTRACT_SUBREG (v8i64 VR512:$src), sub_xmm))>;
318 def : Pat<(v2f64 (extract_subvector (v8f64 VR512:$src), (iPTR 0))),
319 (v2f64 (EXTRACT_SUBREG (v8f64 VR512:$src), sub_xmm))>;
320 def : Pat<(v4f32 (extract_subvector (v16f32 VR512:$src), (iPTR 0))),
321 (v4f32 (EXTRACT_SUBREG (v16f32 VR512:$src), sub_xmm))>;
324 // A 128-bit subvector insert to the first 512-bit vector position
325 // is a subregister copy that needs no instruction.
326 def : Pat<(insert_subvector undef, (v2i64 VR128X:$src), (iPTR 0)),
327 (INSERT_SUBREG (v8i64 (IMPLICIT_DEF)),
328 (INSERT_SUBREG (v4i64 (IMPLICIT_DEF)), VR128X:$src, sub_xmm),
330 def : Pat<(insert_subvector undef, (v2f64 VR128X:$src), (iPTR 0)),
331 (INSERT_SUBREG (v8f64 (IMPLICIT_DEF)),
332 (INSERT_SUBREG (v4f64 (IMPLICIT_DEF)), VR128X:$src, sub_xmm),
334 def : Pat<(insert_subvector undef, (v4i32 VR128X:$src), (iPTR 0)),
335 (INSERT_SUBREG (v16i32 (IMPLICIT_DEF)),
336 (INSERT_SUBREG (v8i32 (IMPLICIT_DEF)), VR128X:$src, sub_xmm),
338 def : Pat<(insert_subvector undef, (v4f32 VR128X:$src), (iPTR 0)),
339 (INSERT_SUBREG (v16f32 (IMPLICIT_DEF)),
340 (INSERT_SUBREG (v8f32 (IMPLICIT_DEF)), VR128X:$src, sub_xmm),
343 def : Pat<(insert_subvector undef, (v4i64 VR256X:$src), (iPTR 0)),
344 (INSERT_SUBREG (v8i64 (IMPLICIT_DEF)), VR256X:$src, sub_ymm)>;
345 def : Pat<(insert_subvector undef, (v4f64 VR256X:$src), (iPTR 0)),
346 (INSERT_SUBREG (v8f64 (IMPLICIT_DEF)), VR256X:$src, sub_ymm)>;
347 def : Pat<(insert_subvector undef, (v8i32 VR256X:$src), (iPTR 0)),
348 (INSERT_SUBREG (v16i32 (IMPLICIT_DEF)), VR256X:$src, sub_ymm)>;
349 def : Pat<(insert_subvector undef, (v8f32 VR256X:$src), (iPTR 0)),
350 (INSERT_SUBREG (v16f32 (IMPLICIT_DEF)), VR256X:$src, sub_ymm)>;
352 // vextractps - extract 32 bits from XMM
353 def VEXTRACTPSzrr : AVX512AIi8<0x17, MRMDestReg, (outs GR32:$dst),
354 (ins VR128X:$src1, u32u8imm:$src2),
355 "vextractps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
356 [(set GR32:$dst, (extractelt (bc_v4i32 (v4f32 VR128X:$src1)), imm:$src2))]>,
359 def VEXTRACTPSzmr : AVX512AIi8<0x17, MRMDestMem, (outs),
360 (ins f32mem:$dst, VR128X:$src1, u32u8imm:$src2),
361 "vextractps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
362 [(store (extractelt (bc_v4i32 (v4f32 VR128X:$src1)), imm:$src2),
365 //===---------------------------------------------------------------------===//
368 multiclass avx512_fp_broadcast<bits<8> opc, string OpcodeStr,
369 RegisterClass DestRC,
370 RegisterClass SrcRC, X86MemOperand x86memop> {
371 def rr : AVX5128I<opc, MRMSrcReg, (outs DestRC:$dst), (ins SrcRC:$src),
372 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
374 def rm : AVX5128I<opc, MRMSrcMem, (outs DestRC:$dst), (ins x86memop:$src),
375 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),[]>, EVEX;
377 let ExeDomain = SSEPackedSingle in {
378 defm VBROADCASTSSZ : avx512_fp_broadcast<0x18, "vbroadcastss", VR512,
380 EVEX_V512, EVEX_CD8<32, CD8VT1>;
383 let ExeDomain = SSEPackedDouble in {
384 defm VBROADCASTSDZ : avx512_fp_broadcast<0x19, "vbroadcastsd", VR512,
386 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT1>;
389 def : Pat<(v16f32 (X86VBroadcast (loadf32 addr:$src))),
390 (VBROADCASTSSZrm addr:$src)>;
391 def : Pat<(v8f64 (X86VBroadcast (loadf64 addr:$src))),
392 (VBROADCASTSDZrm addr:$src)>;
394 def : Pat<(int_x86_avx512_vbroadcast_ss_512 addr:$src),
395 (VBROADCASTSSZrm addr:$src)>;
396 def : Pat<(int_x86_avx512_vbroadcast_sd_512 addr:$src),
397 (VBROADCASTSDZrm addr:$src)>;
399 multiclass avx512_int_broadcast_reg<bits<8> opc, string OpcodeStr,
400 RegisterClass SrcRC, RegisterClass KRC> {
401 def Zrr : AVX5128I<opc, MRMSrcReg, (outs VR512:$dst), (ins SrcRC:$src),
402 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
403 []>, EVEX, EVEX_V512;
404 def Zkrr : AVX5128I<opc, MRMSrcReg, (outs VR512:$dst),
405 (ins KRC:$mask, SrcRC:$src),
406 !strconcat(OpcodeStr,
407 "\t{$src, $dst {${mask}} {z}|$dst {${mask}} {z}, $src}"),
408 []>, EVEX, EVEX_V512, EVEX_KZ;
411 defm VPBROADCASTDr : avx512_int_broadcast_reg<0x7C, "vpbroadcastd", GR32, VK16WM>;
412 defm VPBROADCASTQr : avx512_int_broadcast_reg<0x7C, "vpbroadcastq", GR64, VK8WM>,
415 def : Pat <(v16i32 (X86vzext VK16WM:$mask)),
416 (VPBROADCASTDrZkrr VK16WM:$mask, (i32 (MOV32ri 0x1)))>;
418 def : Pat <(v8i64 (X86vzext VK8WM:$mask)),
419 (VPBROADCASTQrZkrr VK8WM:$mask, (i64 (MOV64ri 0x1)))>;
421 def : Pat<(v16i32 (X86VBroadcast (i32 GR32:$src))),
422 (VPBROADCASTDrZrr GR32:$src)>;
423 def : Pat<(v16i32 (X86VBroadcastm VK16WM:$mask, (i32 GR32:$src))),
424 (VPBROADCASTDrZkrr VK16WM:$mask, GR32:$src)>;
425 def : Pat<(v8i64 (X86VBroadcast (i64 GR64:$src))),
426 (VPBROADCASTQrZrr GR64:$src)>;
427 def : Pat<(v8i64 (X86VBroadcastm VK8WM:$mask, (i64 GR64:$src))),
428 (VPBROADCASTQrZkrr VK8WM:$mask, GR64:$src)>;
430 def : Pat<(v16i32 (int_x86_avx512_pbroadcastd_i32_512 (i32 GR32:$src))),
431 (VPBROADCASTDrZrr GR32:$src)>;
432 def : Pat<(v8i64 (int_x86_avx512_pbroadcastq_i64_512 (i64 GR64:$src))),
433 (VPBROADCASTQrZrr GR64:$src)>;
435 multiclass avx512_int_broadcast_rm<bits<8> opc, string OpcodeStr,
436 X86MemOperand x86memop, PatFrag ld_frag,
437 RegisterClass DstRC, ValueType OpVT, ValueType SrcVT,
439 def rr : AVX5128I<opc, MRMSrcReg, (outs DstRC:$dst), (ins VR128X:$src),
440 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
442 (OpVT (X86VBroadcast (SrcVT VR128X:$src))))]>, EVEX;
443 def krr : AVX5128I<opc, MRMSrcReg, (outs DstRC:$dst), (ins KRC:$mask,
445 !strconcat(OpcodeStr,
446 "\t{$src, ${dst} {${mask}} {z}|${dst} {${mask}} {z}, $src}"),
448 (OpVT (X86VBroadcastm KRC:$mask, (SrcVT VR128X:$src))))]>,
451 def rm : AVX5128I<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src),
452 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
454 (OpVT (X86VBroadcast (ld_frag addr:$src))))]>, EVEX;
455 def krm : AVX5128I<opc, MRMSrcMem, (outs DstRC:$dst), (ins KRC:$mask,
457 !strconcat(OpcodeStr,
458 "\t{$src, ${dst} {${mask}} {z}|${dst} {${mask}} {z}, $src}"),
459 [(set DstRC:$dst, (OpVT (X86VBroadcastm KRC:$mask,
460 (ld_frag addr:$src))))]>, EVEX, EVEX_KZ;
464 defm VPBROADCASTDZ : avx512_int_broadcast_rm<0x58, "vpbroadcastd", i32mem,
465 loadi32, VR512, v16i32, v4i32, VK16WM>,
466 EVEX_V512, EVEX_CD8<32, CD8VT1>;
467 defm VPBROADCASTQZ : avx512_int_broadcast_rm<0x59, "vpbroadcastq", i64mem,
468 loadi64, VR512, v8i64, v2i64, VK8WM>, EVEX_V512, VEX_W,
469 EVEX_CD8<64, CD8VT1>;
471 def : Pat<(v16i32 (int_x86_avx512_pbroadcastd_512 (v4i32 VR128X:$src))),
472 (VPBROADCASTDZrr VR128X:$src)>;
473 def : Pat<(v8i64 (int_x86_avx512_pbroadcastq_512 (v2i64 VR128X:$src))),
474 (VPBROADCASTQZrr VR128X:$src)>;
476 def : Pat<(v16f32 (X86VBroadcast (v4f32 VR128X:$src))),
477 (VBROADCASTSSZrr VR128X:$src)>;
478 def : Pat<(v8f64 (X86VBroadcast (v2f64 VR128X:$src))),
479 (VBROADCASTSDZrr VR128X:$src)>;
481 def : Pat<(v16f32 (int_x86_avx512_vbroadcast_ss_ps_512 (v4f32 VR128X:$src))),
482 (VBROADCASTSSZrr VR128X:$src)>;
483 def : Pat<(v8f64 (int_x86_avx512_vbroadcast_sd_pd_512 (v2f64 VR128X:$src))),
484 (VBROADCASTSDZrr VR128X:$src)>;
486 // Provide fallback in case the load node that is used in the patterns above
487 // is used by additional users, which prevents the pattern selection.
488 def : Pat<(v16f32 (X86VBroadcast FR32X:$src)),
489 (VBROADCASTSSZrr (COPY_TO_REGCLASS FR32X:$src, VR128X))>;
490 def : Pat<(v8f64 (X86VBroadcast FR64X:$src)),
491 (VBROADCASTSDZrr (COPY_TO_REGCLASS FR64X:$src, VR128X))>;
494 let Predicates = [HasAVX512] in {
495 def : Pat<(v8i32 (X86VBroadcastm (v8i1 VK8WM:$mask), (loadi32 addr:$src))),
497 (v16i32 (VPBROADCASTDZkrm (COPY_TO_REGCLASS VK8WM:$mask, VK16WM),
498 addr:$src)), sub_ymm)>;
500 //===----------------------------------------------------------------------===//
501 // AVX-512 BROADCAST MASK TO VECTOR REGISTER
504 multiclass avx512_mask_broadcast<bits<8> opc, string OpcodeStr,
505 RegisterClass DstRC, RegisterClass KRC,
506 ValueType OpVT, ValueType SrcVT> {
507 def rr : AVX512XS8I<opc, MRMDestReg, (outs DstRC:$dst), (ins KRC:$src),
508 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
512 defm VPBROADCASTMW2D : avx512_mask_broadcast<0x3A, "vpbroadcastmw2d", VR512,
513 VK16, v16i32, v16i1>, EVEX_V512;
514 defm VPBROADCASTMB2Q : avx512_mask_broadcast<0x2A, "vpbroadcastmb2q", VR512,
515 VK8, v8i64, v8i1>, EVEX_V512, VEX_W;
517 //===----------------------------------------------------------------------===//
520 // -- immediate form --
521 multiclass avx512_perm_imm<bits<8> opc, string OpcodeStr, RegisterClass RC,
522 SDNode OpNode, PatFrag mem_frag,
523 X86MemOperand x86memop, ValueType OpVT> {
524 def ri : AVX512AIi8<opc, MRMSrcReg, (outs RC:$dst),
525 (ins RC:$src1, i8imm:$src2),
526 !strconcat(OpcodeStr,
527 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
529 (OpVT (OpNode RC:$src1, (i8 imm:$src2))))]>,
531 def mi : AVX512AIi8<opc, MRMSrcMem, (outs RC:$dst),
532 (ins x86memop:$src1, i8imm:$src2),
533 !strconcat(OpcodeStr,
534 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
536 (OpVT (OpNode (mem_frag addr:$src1),
537 (i8 imm:$src2))))]>, EVEX;
540 defm VPERMQZ : avx512_perm_imm<0x00, "vpermq", VR512, X86VPermi, memopv8i64,
541 i512mem, v8i64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
542 let ExeDomain = SSEPackedDouble in
543 defm VPERMPDZ : avx512_perm_imm<0x01, "vpermpd", VR512, X86VPermi, memopv8f64,
544 f512mem, v8f64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
546 // -- VPERM - register form --
547 multiclass avx512_perm<bits<8> opc, string OpcodeStr, RegisterClass RC,
548 PatFrag mem_frag, X86MemOperand x86memop, ValueType OpVT> {
550 def rr : AVX5128I<opc, MRMSrcReg, (outs RC:$dst),
551 (ins RC:$src1, RC:$src2),
552 !strconcat(OpcodeStr,
553 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
555 (OpVT (X86VPermv RC:$src1, RC:$src2)))]>, EVEX_4V;
557 def rm : AVX5128I<opc, MRMSrcMem, (outs RC:$dst),
558 (ins RC:$src1, x86memop:$src2),
559 !strconcat(OpcodeStr,
560 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
562 (OpVT (X86VPermv RC:$src1, (mem_frag addr:$src2))))]>,
566 defm VPERMDZ : avx512_perm<0x36, "vpermd", VR512, memopv16i32, i512mem,
567 v16i32>, EVEX_V512, EVEX_CD8<32, CD8VF>;
568 defm VPERMQZ : avx512_perm<0x36, "vpermq", VR512, memopv8i64, i512mem,
569 v8i64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
570 let ExeDomain = SSEPackedSingle in
571 defm VPERMPSZ : avx512_perm<0x16, "vpermps", VR512, memopv16f32, f512mem,
572 v16f32>, EVEX_V512, EVEX_CD8<32, CD8VF>;
573 let ExeDomain = SSEPackedDouble in
574 defm VPERMPDZ : avx512_perm<0x16, "vpermpd", VR512, memopv8f64, f512mem,
575 v8f64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
577 // -- VPERM2I - 3 source operands form --
578 multiclass avx512_perm_3src<bits<8> opc, string OpcodeStr, RegisterClass RC,
579 PatFrag mem_frag, X86MemOperand x86memop,
581 let Constraints = "$src1 = $dst" in {
582 def rr : AVX5128I<opc, MRMSrcReg, (outs RC:$dst),
583 (ins RC:$src1, RC:$src2, RC:$src3),
584 !strconcat(OpcodeStr,
585 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
587 (OpVT (X86VPermv3 RC:$src1, RC:$src2, RC:$src3)))]>,
590 def rm : AVX5128I<opc, MRMSrcMem, (outs RC:$dst),
591 (ins RC:$src1, RC:$src2, x86memop:$src3),
592 !strconcat(OpcodeStr,
593 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
595 (OpVT (X86VPermv3 RC:$src1, RC:$src2,
596 (mem_frag addr:$src3))))]>, EVEX_4V;
599 defm VPERMI2D : avx512_perm_3src<0x76, "vpermi2d", VR512, memopv16i32, i512mem,
600 v16i32>, EVEX_V512, EVEX_CD8<32, CD8VF>;
601 defm VPERMI2Q : avx512_perm_3src<0x76, "vpermi2q", VR512, memopv8i64, i512mem,
602 v8i64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
603 defm VPERMI2PS : avx512_perm_3src<0x77, "vpermi2ps", VR512, memopv16f32, i512mem,
604 v16f32>, EVEX_V512, EVEX_CD8<32, CD8VF>;
605 defm VPERMI2PD : avx512_perm_3src<0x77, "vpermi2pd", VR512, memopv8f64, i512mem,
606 v8f64>, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
608 //===----------------------------------------------------------------------===//
609 // AVX-512 - BLEND using mask
611 multiclass avx512_blendmask<bits<8> opc, string OpcodeStr, Intrinsic Int,
612 RegisterClass KRC, RegisterClass RC,
613 X86MemOperand x86memop, PatFrag mem_frag,
614 SDNode OpNode, ValueType vt> {
615 def rr : AVX5128I<opc, MRMSrcReg, (outs RC:$dst),
616 (ins KRC:$mask, RC:$src1, RC:$src2),
617 !strconcat(OpcodeStr,
618 "\t{$src2, $src1, ${dst} {${mask}}|${dst} {${mask}}, $src1, $src2}"),
619 [(set RC:$dst, (OpNode KRC:$mask, (vt RC:$src2),
620 (vt RC:$src1)))]>, EVEX_4V, EVEX_K;
621 let isCodeGenOnly = 1 in
622 def rr_Int : AVX5128I<opc, MRMSrcReg, (outs RC:$dst),
623 (ins KRC:$mask, RC:$src1, RC:$src2),
624 !strconcat(OpcodeStr,
625 "\t{$src2, $src1, ${dst} {${mask}}|${dst} {${mask}}, $src1, $src2}"),
626 [(set RC:$dst, (Int KRC:$mask, (vt RC:$src2),
627 (vt RC:$src1)))]>, EVEX_4V, EVEX_K;
630 def rm : AVX5128I<opc, MRMSrcMem, (outs RC:$dst),
631 (ins KRC:$mask, RC:$src1, x86memop:$src2),
632 !strconcat(OpcodeStr,
633 "\t{$src2, $src1, ${dst} {${mask}}|${dst} {${mask}}, $src1, $src2}"),
637 let isCodeGenOnly = 1 in
638 def rm_Int : AVX5128I<opc, MRMSrcMem, (outs RC:$dst),
639 (ins KRC:$mask, RC:$src1, x86memop:$src2),
640 !strconcat(OpcodeStr,
641 "\t{$src2, $src1, ${dst} {${mask}}|${dst} {${mask}}, $src1, $src2}"),
642 [(set RC:$dst, (Int KRC:$mask, (vt RC:$src1),
643 (mem_frag addr:$src2)))]>,
648 let ExeDomain = SSEPackedSingle in
649 defm VBLENDMPSZ : avx512_blendmask<0x65, "vblendmps",
650 int_x86_avx512_mask_blend_ps_512,
651 VK16WM, VR512, f512mem,
652 memopv16f32, vselect, v16f32>,
653 EVEX_CD8<32, CD8VF>, EVEX_V512;
654 let ExeDomain = SSEPackedDouble in
655 defm VBLENDMPDZ : avx512_blendmask<0x65, "vblendmpd",
656 int_x86_avx512_mask_blend_pd_512,
657 VK8WM, VR512, f512mem,
658 memopv8f64, vselect, v8f64>,
659 VEX_W, EVEX_CD8<64, CD8VF>, EVEX_V512;
661 defm VPBLENDMDZ : avx512_blendmask<0x64, "vpblendmd",
662 int_x86_avx512_mask_blend_d_512,
663 VK16WM, VR512, f512mem,
664 memopv16i32, vselect, v16i32>,
665 EVEX_CD8<32, CD8VF>, EVEX_V512;
667 defm VPBLENDMQZ : avx512_blendmask<0x64, "vpblendmq",
668 int_x86_avx512_mask_blend_q_512,
669 VK8WM, VR512, f512mem,
670 memopv8i64, vselect, v8i64>,
671 VEX_W, EVEX_CD8<64, CD8VF>, EVEX_V512;
673 let Predicates = [HasAVX512] in {
674 def : Pat<(v8f32 (vselect (v8i1 VK8WM:$mask), (v8f32 VR256X:$src1),
675 (v8f32 VR256X:$src2))),
677 (v16f32 (VBLENDMPSZrr (COPY_TO_REGCLASS VK8WM:$mask, VK16WM),
678 (v16f32 (SUBREG_TO_REG (i32 0), VR256X:$src2, sub_ymm)),
679 (v16f32 (SUBREG_TO_REG (i32 0), VR256X:$src1, sub_ymm)))), sub_ymm)>;
681 def : Pat<(v8i32 (vselect (v8i1 VK8WM:$mask), (v8i32 VR256X:$src1),
682 (v8i32 VR256X:$src2))),
684 (v16i32 (VPBLENDMDZrr (COPY_TO_REGCLASS VK8WM:$mask, VK16WM),
685 (v16i32 (SUBREG_TO_REG (i32 0), VR256X:$src2, sub_ymm)),
686 (v16i32 (SUBREG_TO_REG (i32 0), VR256X:$src1, sub_ymm)))), sub_ymm)>;
688 //===----------------------------------------------------------------------===//
689 // Compare Instructions
690 //===----------------------------------------------------------------------===//
692 // avx512_cmp_scalar - AVX512 CMPSS and CMPSD
693 multiclass avx512_cmp_scalar<RegisterClass RC, X86MemOperand x86memop,
694 Operand CC, SDNode OpNode, ValueType VT,
695 PatFrag ld_frag, string asm, string asm_alt> {
696 def rr : AVX512Ii8<0xC2, MRMSrcReg,
697 (outs VK1:$dst), (ins RC:$src1, RC:$src2, CC:$cc), asm,
698 [(set VK1:$dst, (OpNode (VT RC:$src1), RC:$src2, imm:$cc))],
699 IIC_SSE_ALU_F32S_RR>, EVEX_4V;
700 def rm : AVX512Ii8<0xC2, MRMSrcMem,
701 (outs VK1:$dst), (ins RC:$src1, x86memop:$src2, CC:$cc), asm,
702 [(set VK1:$dst, (OpNode (VT RC:$src1),
703 (ld_frag addr:$src2), imm:$cc))], IIC_SSE_ALU_F32P_RM>, EVEX_4V;
704 let isAsmParserOnly = 1, hasSideEffects = 0 in {
705 def rri_alt : AVX512Ii8<0xC2, MRMSrcReg,
706 (outs VK1:$dst), (ins RC:$src1, RC:$src2, i8imm:$cc),
707 asm_alt, [], IIC_SSE_ALU_F32S_RR>, EVEX_4V;
708 def rmi_alt : AVX512Ii8<0xC2, MRMSrcMem,
709 (outs VK1:$dst), (ins RC:$src1, x86memop:$src2, i8imm:$cc),
710 asm_alt, [], IIC_SSE_ALU_F32P_RM>, EVEX_4V;
714 let Predicates = [HasAVX512] in {
715 defm VCMPSSZ : avx512_cmp_scalar<FR32X, f32mem, AVXCC, X86cmpms, f32, loadf32,
716 "vcmp${cc}ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
717 "vcmpss\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}">,
719 defm VCMPSDZ : avx512_cmp_scalar<FR64X, f64mem, AVXCC, X86cmpms, f64, loadf64,
720 "vcmp${cc}sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
721 "vcmpsd\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}">,
725 multiclass avx512_icmp_packed<bits<8> opc, string OpcodeStr, RegisterClass KRC,
726 RegisterClass RC, X86MemOperand x86memop, PatFrag memop_frag,
727 SDNode OpNode, ValueType vt> {
728 def rr : AVX512BI<opc, MRMSrcReg,
729 (outs KRC:$dst), (ins RC:$src1, RC:$src2),
730 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
731 [(set KRC:$dst, (OpNode (vt RC:$src1), (vt RC:$src2)))],
732 IIC_SSE_ALU_F32P_RR>, EVEX_4V;
733 def rm : AVX512BI<opc, MRMSrcMem,
734 (outs KRC:$dst), (ins RC:$src1, x86memop:$src2),
735 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
736 [(set KRC:$dst, (OpNode (vt RC:$src1), (memop_frag addr:$src2)))],
737 IIC_SSE_ALU_F32P_RM>, EVEX_4V;
740 defm VPCMPEQDZ : avx512_icmp_packed<0x76, "vpcmpeqd", VK16, VR512, i512mem,
741 memopv16i32, X86pcmpeqm, v16i32>, EVEX_V512;
742 defm VPCMPEQQZ : avx512_icmp_packed<0x29, "vpcmpeqq", VK8, VR512, i512mem,
743 memopv8i64, X86pcmpeqm, v8i64>, T8, EVEX_V512, VEX_W;
745 defm VPCMPGTDZ : avx512_icmp_packed<0x66, "vpcmpgtd", VK16, VR512, i512mem,
746 memopv16i32, X86pcmpgtm, v16i32>, EVEX_V512;
747 defm VPCMPGTQZ : avx512_icmp_packed<0x37, "vpcmpgtq", VK8, VR512, i512mem,
748 memopv8i64, X86pcmpgtm, v8i64>, T8, EVEX_V512, VEX_W;
750 def : Pat<(v8i1 (X86pcmpgtm (v8i32 VR256X:$src1), (v8i32 VR256X:$src2))),
751 (COPY_TO_REGCLASS (VPCMPGTDZrr
752 (v16i32 (SUBREG_TO_REG (i32 0), VR256X:$src1, sub_ymm)),
753 (v16i32 (SUBREG_TO_REG (i32 0), VR256X:$src2, sub_ymm))), VK8)>;
755 def : Pat<(v8i1 (X86pcmpeqm (v8i32 VR256X:$src1), (v8i32 VR256X:$src2))),
756 (COPY_TO_REGCLASS (VPCMPEQDZrr
757 (v16i32 (SUBREG_TO_REG (i32 0), VR256X:$src1, sub_ymm)),
758 (v16i32 (SUBREG_TO_REG (i32 0), VR256X:$src2, sub_ymm))), VK8)>;
760 multiclass avx512_icmp_cc<bits<8> opc, RegisterClass KRC,
761 RegisterClass RC, X86MemOperand x86memop, PatFrag memop_frag,
762 SDNode OpNode, ValueType vt, Operand CC, string asm,
764 def rri : AVX512AIi8<opc, MRMSrcReg,
765 (outs KRC:$dst), (ins RC:$src1, RC:$src2, CC:$cc), asm,
766 [(set KRC:$dst, (OpNode (vt RC:$src1), (vt RC:$src2), imm:$cc))],
767 IIC_SSE_ALU_F32P_RR>, EVEX_4V;
768 def rmi : AVX512AIi8<opc, MRMSrcMem,
769 (outs KRC:$dst), (ins RC:$src1, x86memop:$src2, CC:$cc), asm,
770 [(set KRC:$dst, (OpNode (vt RC:$src1), (memop_frag addr:$src2),
771 imm:$cc))], IIC_SSE_ALU_F32P_RM>, EVEX_4V;
772 // Accept explicit immediate argument form instead of comparison code.
773 let isAsmParserOnly = 1, hasSideEffects = 0 in {
774 def rri_alt : AVX512AIi8<opc, MRMSrcReg,
775 (outs KRC:$dst), (ins RC:$src1, RC:$src2, i8imm:$cc),
776 asm_alt, [], IIC_SSE_ALU_F32P_RR>, EVEX_4V;
777 def rmi_alt : AVX512AIi8<opc, MRMSrcMem,
778 (outs KRC:$dst), (ins RC:$src1, x86memop:$src2, i8imm:$cc),
779 asm_alt, [], IIC_SSE_ALU_F32P_RM>, EVEX_4V;
783 defm VPCMPDZ : avx512_icmp_cc<0x1F, VK16, VR512, i512mem, memopv16i32,
784 X86cmpm, v16i32, AVXCC,
785 "vpcmp${cc}d\t{$src2, $src1, $dst|$dst, $src1, $src2}",
786 "vpcmpd\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}">,
787 EVEX_V512, EVEX_CD8<32, CD8VF>;
788 defm VPCMPUDZ : avx512_icmp_cc<0x1E, VK16, VR512, i512mem, memopv16i32,
789 X86cmpmu, v16i32, AVXCC,
790 "vpcmp${cc}ud\t{$src2, $src1, $dst|$dst, $src1, $src2}",
791 "vpcmpud\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}">,
792 EVEX_V512, EVEX_CD8<32, CD8VF>;
794 defm VPCMPQZ : avx512_icmp_cc<0x1F, VK8, VR512, i512mem, memopv8i64,
795 X86cmpm, v8i64, AVXCC,
796 "vpcmp${cc}q\t{$src2, $src1, $dst|$dst, $src1, $src2}",
797 "vpcmpq\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}">,
798 VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>;
799 defm VPCMPUQZ : avx512_icmp_cc<0x1E, VK8, VR512, i512mem, memopv8i64,
800 X86cmpmu, v8i64, AVXCC,
801 "vpcmp${cc}uq\t{$src2, $src1, $dst|$dst, $src1, $src2}",
802 "vpcmpuq\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}">,
803 VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>;
805 // avx512_cmp_packed - sse 1 & 2 compare packed instructions
806 multiclass avx512_cmp_packed<RegisterClass KRC, RegisterClass RC,
807 X86MemOperand x86memop, ValueType vt,
808 string suffix, Domain d> {
809 def rri : AVX512PIi8<0xC2, MRMSrcReg,
810 (outs KRC:$dst), (ins RC:$src1, RC:$src2, AVXCC:$cc),
811 !strconcat("vcmp${cc}", suffix,
812 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
813 [(set KRC:$dst, (X86cmpm (vt RC:$src1), (vt RC:$src2), imm:$cc))], d>;
814 def rrib: AVX512PIi8<0xC2, MRMSrcReg,
815 (outs KRC:$dst), (ins RC:$src1, RC:$src2, AVXCC:$cc, i32imm:$sae),
816 !strconcat("vcmp${cc}", suffix,
817 "\t{{sae}, $src2, $src1, $dst|$dst, $src1, $src2, {sae}}"),
819 def rmi : AVX512PIi8<0xC2, MRMSrcMem,
820 (outs KRC:$dst), (ins RC:$src1, x86memop:$src2, AVXCC:$cc),
821 !strconcat("vcmp", suffix,
822 "\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}"),
824 (X86cmpm (vt RC:$src1), (memop addr:$src2), imm:$cc))], d>;
826 // Accept explicit immediate argument form instead of comparison code.
827 let isAsmParserOnly = 1, hasSideEffects = 0 in {
828 def rri_alt : AVX512PIi8<0xC2, MRMSrcReg,
829 (outs RC:$dst), (ins RC:$src1, RC:$src2, i8imm:$cc),
830 !strconcat("vcmp", suffix,
831 "\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}"), [], d>;
832 def rmi_alt : AVX512PIi8<0xC2, MRMSrcMem,
833 (outs RC:$dst), (ins RC:$src1, x86memop:$src2, i8imm:$cc),
834 !strconcat("vcmp", suffix,
835 "\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}"), [], d>;
839 defm VCMPPSZ : avx512_cmp_packed<VK16, VR512, f512mem, v16f32,
840 "ps", SSEPackedSingle>, EVEX_4V, EVEX_V512, EVEX_CD8<32, CD8VF>;
841 defm VCMPPDZ : avx512_cmp_packed<VK8, VR512, f512mem, v8f64,
842 "pd", SSEPackedDouble>, OpSize, EVEX_4V, VEX_W, EVEX_V512,
845 def : Pat<(v8i1 (X86cmpm (v8f32 VR256X:$src1), (v8f32 VR256X:$src2), imm:$cc)),
846 (COPY_TO_REGCLASS (VCMPPSZrri
847 (v16f32 (SUBREG_TO_REG (i32 0), VR256X:$src1, sub_ymm)),
848 (v16f32 (SUBREG_TO_REG (i32 0), VR256X:$src2, sub_ymm)),
850 def : Pat<(v8i1 (X86cmpm (v8i32 VR256X:$src1), (v8i32 VR256X:$src2), imm:$cc)),
851 (COPY_TO_REGCLASS (VPCMPDZrri
852 (v16i32 (SUBREG_TO_REG (i32 0), VR256X:$src1, sub_ymm)),
853 (v16i32 (SUBREG_TO_REG (i32 0), VR256X:$src2, sub_ymm)),
855 def : Pat<(v8i1 (X86cmpmu (v8i32 VR256X:$src1), (v8i32 VR256X:$src2), imm:$cc)),
856 (COPY_TO_REGCLASS (VPCMPUDZrri
857 (v16i32 (SUBREG_TO_REG (i32 0), VR256X:$src1, sub_ymm)),
858 (v16i32 (SUBREG_TO_REG (i32 0), VR256X:$src2, sub_ymm)),
861 def : Pat<(i16 (int_x86_avx512_mask_cmp_ps_512 (v16f32 VR512:$src1),
862 (v16f32 VR512:$src2), imm:$cc, (i16 -1),
864 (COPY_TO_REGCLASS (VCMPPSZrrib VR512:$src1, VR512:$src2,
865 (I8Imm imm:$cc), (i32 0)), GR16)>;
867 def : Pat<(i8 (int_x86_avx512_mask_cmp_pd_512 (v8f64 VR512:$src1),
868 (v8f64 VR512:$src2), imm:$cc, (i8 -1),
870 (COPY_TO_REGCLASS (VCMPPDZrrib VR512:$src1, VR512:$src2,
871 (I8Imm imm:$cc), (i32 0)), GR8)>;
873 def : Pat<(i16 (int_x86_avx512_mask_cmp_ps_512 (v16f32 VR512:$src1),
874 (v16f32 VR512:$src2), imm:$cc, (i16 -1),
876 (COPY_TO_REGCLASS (VCMPPSZrri VR512:$src1, VR512:$src2,
877 (I8Imm imm:$cc)), GR16)>;
879 def : Pat<(i8 (int_x86_avx512_mask_cmp_pd_512 (v8f64 VR512:$src1),
880 (v8f64 VR512:$src2), imm:$cc, (i8 -1),
882 (COPY_TO_REGCLASS (VCMPPDZrri VR512:$src1, VR512:$src2,
883 (I8Imm imm:$cc)), GR8)>;
885 // Mask register copy, including
886 // - copy between mask registers
887 // - load/store mask registers
888 // - copy from GPR to mask register and vice versa
890 multiclass avx512_mask_mov<bits<8> opc_kk, bits<8> opc_km, bits<8> opc_mk,
891 string OpcodeStr, RegisterClass KRC,
892 ValueType vt, X86MemOperand x86memop> {
893 let hasSideEffects = 0 in {
894 def kk : I<opc_kk, MRMSrcReg, (outs KRC:$dst), (ins KRC:$src),
895 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"), []>;
897 def km : I<opc_km, MRMSrcMem, (outs KRC:$dst), (ins x86memop:$src),
898 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
899 [(set KRC:$dst, (vt (load addr:$src)))]>;
901 def mk : I<opc_mk, MRMDestMem, (outs), (ins x86memop:$dst, KRC:$src),
902 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"), []>;
906 multiclass avx512_mask_mov_gpr<bits<8> opc_kr, bits<8> opc_rk,
908 RegisterClass KRC, RegisterClass GRC> {
909 let hasSideEffects = 0 in {
910 def kr : I<opc_kr, MRMSrcReg, (outs KRC:$dst), (ins GRC:$src),
911 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"), []>;
912 def rk : I<opc_rk, MRMSrcReg, (outs GRC:$dst), (ins KRC:$src),
913 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"), []>;
917 let Predicates = [HasAVX512] in {
918 defm KMOVW : avx512_mask_mov<0x90, 0x90, 0x91, "kmovw", VK16, v16i1, i16mem>,
920 defm KMOVW : avx512_mask_mov_gpr<0x92, 0x93, "kmovw", VK16, GR32>,
924 let Predicates = [HasAVX512] in {
925 // GR16 from/to 16-bit mask
926 def : Pat<(v16i1 (bitconvert (i16 GR16:$src))),
927 (KMOVWkr (SUBREG_TO_REG (i32 0), GR16:$src, sub_16bit))>;
928 def : Pat<(i16 (bitconvert (v16i1 VK16:$src))),
929 (EXTRACT_SUBREG (KMOVWrk VK16:$src), sub_16bit)>;
931 // Store kreg in memory
932 def : Pat<(store (v16i1 VK16:$src), addr:$dst),
933 (KMOVWmk addr:$dst, VK16:$src)>;
935 def : Pat<(store VK8:$src, addr:$dst),
936 (KMOVWmk addr:$dst, (COPY_TO_REGCLASS VK8:$src, VK16))>;
938 def : Pat<(i1 (load addr:$src)),
939 (COPY_TO_REGCLASS (KMOVWkm addr:$src), VK1)>;
941 def : Pat<(v8i1 (load addr:$src)),
942 (COPY_TO_REGCLASS (KMOVWkm addr:$src), VK8)>;
944 def : Pat<(i1 (trunc (i32 GR32:$src))),
945 (COPY_TO_REGCLASS (KMOVWkr $src), VK1)>;
947 def : Pat<(i1 (trunc (i8 GR8:$src))),
949 (KMOVWkr (SUBREG_TO_REG (i32 0), GR8:$src, sub_8bit)), VK1)>;
951 def : Pat<(i32 (zext VK1:$src)), (KMOVWrk (COPY_TO_REGCLASS VK1:$src, VK16))>;
952 def : Pat<(i8 (zext VK1:$src)),
954 (KMOVWrk (COPY_TO_REGCLASS VK1:$src, VK16)), sub_8bit)>;
956 // With AVX-512 only, 8-bit mask is promoted to 16-bit mask.
957 let Predicates = [HasAVX512] in {
958 // GR from/to 8-bit mask without native support
959 def : Pat<(v8i1 (bitconvert (i8 GR8:$src))),
961 (KMOVWkr (SUBREG_TO_REG (i32 0), GR8:$src, sub_8bit)),
963 def : Pat<(i8 (bitconvert (v8i1 VK8:$src))),
965 (KMOVWrk (COPY_TO_REGCLASS VK8:$src, VK16)),
968 def : Pat<(i1 (extractelt VK16:$src, (iPTR 0))),
969 (COPY_TO_REGCLASS VK16:$src, VK1)>;
970 def : Pat<(i1 (extractelt VK8:$src, (iPTR 0))),
971 (COPY_TO_REGCLASS VK8:$src, VK1)>;
975 // Mask unary operation
977 multiclass avx512_mask_unop<bits<8> opc, string OpcodeStr,
978 RegisterClass KRC, SDPatternOperator OpNode> {
979 let Predicates = [HasAVX512] in
980 def rr : I<opc, MRMSrcReg, (outs KRC:$dst), (ins KRC:$src),
981 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
982 [(set KRC:$dst, (OpNode KRC:$src))]>;
985 multiclass avx512_mask_unop_w<bits<8> opc, string OpcodeStr,
986 SDPatternOperator OpNode> {
987 defm W : avx512_mask_unop<opc, !strconcat(OpcodeStr, "w"), VK16, OpNode>,
991 defm KNOT : avx512_mask_unop_w<0x44, "knot", not>;
993 multiclass avx512_mask_unop_int<string IntName, string InstName> {
994 let Predicates = [HasAVX512] in
995 def : Pat<(!cast<Intrinsic>("int_x86_avx512_"##IntName##"_w")
997 (COPY_TO_REGCLASS (!cast<Instruction>(InstName##"Wrr")
998 (v16i1 (COPY_TO_REGCLASS GR16:$src, VK16))), GR16)>;
1000 defm : avx512_mask_unop_int<"knot", "KNOT">;
1002 def : Pat<(xor VK16:$src1, (v16i1 immAllOnesV)), (KNOTWrr VK16:$src1)>;
1003 def : Pat<(xor VK8:$src1, (v8i1 immAllOnesV)),
1004 (COPY_TO_REGCLASS (KNOTWrr (COPY_TO_REGCLASS VK8:$src1, VK16)), VK8)>;
1006 // With AVX-512, 8-bit mask is promoted to 16-bit mask.
1007 def : Pat<(not VK8:$src),
1009 (KNOTWrr (COPY_TO_REGCLASS VK8:$src, VK16)), VK8)>;
1011 // Mask binary operation
1012 // - KAND, KANDN, KOR, KXNOR, KXOR
1013 multiclass avx512_mask_binop<bits<8> opc, string OpcodeStr,
1014 RegisterClass KRC, SDPatternOperator OpNode> {
1015 let Predicates = [HasAVX512] in
1016 def rr : I<opc, MRMSrcReg, (outs KRC:$dst), (ins KRC:$src1, KRC:$src2),
1017 !strconcat(OpcodeStr,
1018 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1019 [(set KRC:$dst, (OpNode KRC:$src1, KRC:$src2))]>;
1022 multiclass avx512_mask_binop_w<bits<8> opc, string OpcodeStr,
1023 SDPatternOperator OpNode> {
1024 defm W : avx512_mask_binop<opc, !strconcat(OpcodeStr, "w"), VK16, OpNode>,
1028 def andn : PatFrag<(ops node:$i0, node:$i1), (and (not node:$i0), node:$i1)>;
1029 def xnor : PatFrag<(ops node:$i0, node:$i1), (not (xor node:$i0, node:$i1))>;
1031 let isCommutable = 1 in {
1032 defm KAND : avx512_mask_binop_w<0x41, "kand", and>;
1033 let isCommutable = 0 in
1034 defm KANDN : avx512_mask_binop_w<0x42, "kandn", andn>;
1035 defm KOR : avx512_mask_binop_w<0x45, "kor", or>;
1036 defm KXNOR : avx512_mask_binop_w<0x46, "kxnor", xnor>;
1037 defm KXOR : avx512_mask_binop_w<0x47, "kxor", xor>;
1040 def : Pat<(xor VK1:$src1, VK1:$src2),
1041 (COPY_TO_REGCLASS (KXORWrr (COPY_TO_REGCLASS VK1:$src1, VK16),
1042 (COPY_TO_REGCLASS VK1:$src2, VK16)), VK1)>;
1044 def : Pat<(or VK1:$src1, VK1:$src2),
1045 (COPY_TO_REGCLASS (KORWrr (COPY_TO_REGCLASS VK1:$src1, VK16),
1046 (COPY_TO_REGCLASS VK1:$src2, VK16)), VK1)>;
1048 def : Pat<(not VK1:$src),
1049 (COPY_TO_REGCLASS (KXORWrr (COPY_TO_REGCLASS VK1:$src, VK16),
1050 (COPY_TO_REGCLASS (VCMPSSZrr (f32 (IMPLICIT_DEF)),
1051 (f32 (IMPLICIT_DEF)), (i8 0)), VK16)), VK1)>;
1053 def : Pat<(and VK1:$src1, VK1:$src2),
1054 (COPY_TO_REGCLASS (KANDWrr (COPY_TO_REGCLASS VK1:$src1, VK16),
1055 (COPY_TO_REGCLASS VK1:$src2, VK16)), VK1)>;
1057 multiclass avx512_mask_binop_int<string IntName, string InstName> {
1058 let Predicates = [HasAVX512] in
1059 def : Pat<(!cast<Intrinsic>("int_x86_avx512_"##IntName##"_w")
1060 (i16 GR16:$src1), (i16 GR16:$src2)),
1061 (COPY_TO_REGCLASS (!cast<Instruction>(InstName##"Wrr")
1062 (v16i1 (COPY_TO_REGCLASS GR16:$src1, VK16)),
1063 (v16i1 (COPY_TO_REGCLASS GR16:$src2, VK16))), GR16)>;
1066 defm : avx512_mask_binop_int<"kand", "KAND">;
1067 defm : avx512_mask_binop_int<"kandn", "KANDN">;
1068 defm : avx512_mask_binop_int<"kor", "KOR">;
1069 defm : avx512_mask_binop_int<"kxnor", "KXNOR">;
1070 defm : avx512_mask_binop_int<"kxor", "KXOR">;
1072 // With AVX-512, 8-bit mask is promoted to 16-bit mask.
1073 multiclass avx512_binop_pat<SDPatternOperator OpNode, Instruction Inst> {
1074 let Predicates = [HasAVX512] in
1075 def : Pat<(OpNode VK8:$src1, VK8:$src2),
1077 (Inst (COPY_TO_REGCLASS VK8:$src1, VK16),
1078 (COPY_TO_REGCLASS VK8:$src2, VK16)), VK8)>;
1081 defm : avx512_binop_pat<and, KANDWrr>;
1082 defm : avx512_binop_pat<andn, KANDNWrr>;
1083 defm : avx512_binop_pat<or, KORWrr>;
1084 defm : avx512_binop_pat<xnor, KXNORWrr>;
1085 defm : avx512_binop_pat<xor, KXORWrr>;
1088 multiclass avx512_mask_unpck<bits<8> opc, string OpcodeStr,
1089 RegisterClass KRC> {
1090 let Predicates = [HasAVX512] in
1091 def rr : I<opc, MRMSrcReg, (outs KRC:$dst), (ins KRC:$src1, KRC:$src2),
1092 !strconcat(OpcodeStr,
1093 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"), []>;
1096 multiclass avx512_mask_unpck_bw<bits<8> opc, string OpcodeStr> {
1097 defm BW : avx512_mask_unpck<opc, !strconcat(OpcodeStr, "bw"), VK16>,
1098 VEX_4V, VEX_L, OpSize, TB;
1101 defm KUNPCK : avx512_mask_unpck_bw<0x4b, "kunpck">;
1102 def : Pat<(v16i1 (concat_vectors (v8i1 VK8:$src1), (v8i1 VK8:$src2))),
1103 (KUNPCKBWrr (COPY_TO_REGCLASS VK8:$src2, VK16),
1104 (COPY_TO_REGCLASS VK8:$src1, VK16))>;
1107 multiclass avx512_mask_unpck_int<string IntName, string InstName> {
1108 let Predicates = [HasAVX512] in
1109 def : Pat<(!cast<Intrinsic>("int_x86_avx512_"##IntName##"_bw")
1110 (i16 GR16:$src1), (i16 GR16:$src2)),
1111 (COPY_TO_REGCLASS (!cast<Instruction>(InstName##"BWrr")
1112 (v16i1 (COPY_TO_REGCLASS GR16:$src1, VK16)),
1113 (v16i1 (COPY_TO_REGCLASS GR16:$src2, VK16))), GR16)>;
1115 defm : avx512_mask_unpck_int<"kunpck", "KUNPCK">;
1118 multiclass avx512_mask_testop<bits<8> opc, string OpcodeStr, RegisterClass KRC,
1120 let Predicates = [HasAVX512], Defs = [EFLAGS] in
1121 def rr : I<opc, MRMSrcReg, (outs), (ins KRC:$src1, KRC:$src2),
1122 !strconcat(OpcodeStr, "\t{$src2, $src1|$src1, $src2}"),
1123 [(set EFLAGS, (OpNode KRC:$src1, KRC:$src2))]>;
1126 multiclass avx512_mask_testop_w<bits<8> opc, string OpcodeStr, SDNode OpNode> {
1127 defm W : avx512_mask_testop<opc, !strconcat(OpcodeStr, "w"), VK16, OpNode>,
1131 defm KORTEST : avx512_mask_testop_w<0x98, "kortest", X86kortest>;
1133 def : Pat<(X86cmp VK1:$src1, (i1 0)),
1134 (KORTESTWrr (COPY_TO_REGCLASS VK1:$src1, VK16),
1135 (COPY_TO_REGCLASS VK1:$src1, VK16))>;
1138 multiclass avx512_mask_shiftop<bits<8> opc, string OpcodeStr, RegisterClass KRC,
1140 let Predicates = [HasAVX512] in
1141 def ri : Ii8<opc, MRMSrcReg, (outs KRC:$dst), (ins KRC:$src, i8imm:$imm),
1142 !strconcat(OpcodeStr,
1143 "\t{$imm, $src, $dst|$dst, $src, $imm}"),
1144 [(set KRC:$dst, (OpNode KRC:$src, (i8 imm:$imm)))]>;
1147 multiclass avx512_mask_shiftop_w<bits<8> opc1, bits<8> opc2, string OpcodeStr,
1149 defm W : avx512_mask_shiftop<opc1, !strconcat(OpcodeStr, "w"), VK16, OpNode>,
1150 VEX, OpSize, TA, VEX_W;
1153 defm KSHIFTL : avx512_mask_shiftop_w<0x32, 0x33, "kshiftl", X86vshli>;
1154 defm KSHIFTR : avx512_mask_shiftop_w<0x30, 0x31, "kshiftr", X86vsrli>;
1156 // Mask setting all 0s or 1s
1157 multiclass avx512_mask_setop<RegisterClass KRC, ValueType VT, PatFrag Val> {
1158 let Predicates = [HasAVX512] in
1159 let isReMaterializable = 1, isAsCheapAsAMove = 1, isPseudo = 1 in
1160 def #NAME# : I<0, Pseudo, (outs KRC:$dst), (ins), "",
1161 [(set KRC:$dst, (VT Val))]>;
1164 multiclass avx512_mask_setop_w<PatFrag Val> {
1165 defm B : avx512_mask_setop<VK8, v8i1, Val>;
1166 defm W : avx512_mask_setop<VK16, v16i1, Val>;
1169 defm KSET0 : avx512_mask_setop_w<immAllZerosV>;
1170 defm KSET1 : avx512_mask_setop_w<immAllOnesV>;
1172 // With AVX-512 only, 8-bit mask is promoted to 16-bit mask.
1173 let Predicates = [HasAVX512] in {
1174 def : Pat<(v8i1 immAllZerosV), (COPY_TO_REGCLASS (KSET0W), VK8)>;
1175 def : Pat<(v8i1 immAllOnesV), (COPY_TO_REGCLASS (KSET1W), VK8)>;
1177 def : Pat<(v8i1 (extract_subvector (v16i1 VK16:$src), (iPTR 0))),
1178 (v8i1 (COPY_TO_REGCLASS VK16:$src, VK8))>;
1180 def : Pat<(v16i1 (insert_subvector undef, (v8i1 VK8:$src), (iPTR 0))),
1181 (v16i1 (COPY_TO_REGCLASS VK8:$src, VK16))>;
1183 def : Pat<(v8i1 (extract_subvector (v16i1 VK16:$src), (iPTR 8))),
1184 (v8i1 (COPY_TO_REGCLASS (KSHIFTRWri VK16:$src, (i8 8)), VK8))>;
1186 //===----------------------------------------------------------------------===//
1187 // AVX-512 - Aligned and unaligned load and store
1190 multiclass avx512_mov_packed<bits<8> opc, RegisterClass RC, RegisterClass KRC,
1191 X86MemOperand x86memop, PatFrag ld_frag,
1192 string asm, Domain d> {
1193 let hasSideEffects = 0 in
1194 def rr : AVX512PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
1195 !strconcat(asm, "\t{$src, $dst|$dst, $src}"), [], d>,
1197 let canFoldAsLoad = 1 in
1198 def rm : AVX512PI<opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
1199 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
1200 [(set RC:$dst, (ld_frag addr:$src))], d>, EVEX;
1201 let Constraints = "$src1 = $dst" in {
1202 def rrk : AVX512PI<opc, MRMSrcReg, (outs RC:$dst),
1203 (ins RC:$src1, KRC:$mask, RC:$src2),
1205 "\t{$src2, ${dst} {${mask}}|${dst} {${mask}}, $src2}"), [], d>,
1207 def rmk : AVX512PI<opc, MRMSrcMem, (outs RC:$dst),
1208 (ins RC:$src1, KRC:$mask, x86memop:$src2),
1210 "\t{$src2, ${dst} {${mask}}|${dst} {${mask}}, $src2}"),
1211 [], d>, EVEX, EVEX_K;
1215 defm VMOVAPSZ : avx512_mov_packed<0x28, VR512, VK16WM, f512mem, alignedloadv16f32,
1216 "vmovaps", SSEPackedSingle>,
1217 EVEX_V512, EVEX_CD8<32, CD8VF>;
1218 defm VMOVAPDZ : avx512_mov_packed<0x28, VR512, VK8WM, f512mem, alignedloadv8f64,
1219 "vmovapd", SSEPackedDouble>,
1220 OpSize, EVEX_V512, VEX_W,
1221 EVEX_CD8<64, CD8VF>;
1222 defm VMOVUPSZ : avx512_mov_packed<0x10, VR512, VK16WM, f512mem, loadv16f32,
1223 "vmovups", SSEPackedSingle>,
1224 EVEX_V512, EVEX_CD8<32, CD8VF>;
1225 defm VMOVUPDZ : avx512_mov_packed<0x10, VR512, VK8WM, f512mem, loadv8f64,
1226 "vmovupd", SSEPackedDouble>,
1227 OpSize, EVEX_V512, VEX_W,
1228 EVEX_CD8<64, CD8VF>;
1229 def VMOVAPSZmr : AVX512PI<0x29, MRMDestMem, (outs), (ins f512mem:$dst, VR512:$src),
1230 "vmovaps\t{$src, $dst|$dst, $src}",
1231 [(alignedstore512 (v16f32 VR512:$src), addr:$dst)],
1232 SSEPackedSingle>, EVEX, EVEX_V512, EVEX_CD8<32, CD8VF>;
1233 def VMOVAPDZmr : AVX512PI<0x29, MRMDestMem, (outs), (ins f512mem:$dst, VR512:$src),
1234 "vmovapd\t{$src, $dst|$dst, $src}",
1235 [(alignedstore512 (v8f64 VR512:$src), addr:$dst)],
1236 SSEPackedDouble>, EVEX, EVEX_V512,
1237 OpSize, VEX_W, EVEX_CD8<64, CD8VF>;
1238 def VMOVUPSZmr : AVX512PI<0x11, MRMDestMem, (outs), (ins f512mem:$dst, VR512:$src),
1239 "vmovups\t{$src, $dst|$dst, $src}",
1240 [(store (v16f32 VR512:$src), addr:$dst)],
1241 SSEPackedSingle>, EVEX, EVEX_V512, EVEX_CD8<32, CD8VF>;
1242 def VMOVUPDZmr : AVX512PI<0x11, MRMDestMem, (outs), (ins f512mem:$dst, VR512:$src),
1243 "vmovupd\t{$src, $dst|$dst, $src}",
1244 [(store (v8f64 VR512:$src), addr:$dst)],
1245 SSEPackedDouble>, EVEX, EVEX_V512,
1246 OpSize, VEX_W, EVEX_CD8<64, CD8VF>;
1248 let hasSideEffects = 0 in {
1249 def VMOVDQA32rr : AVX512BI<0x6F, MRMSrcReg, (outs VR512:$dst),
1251 "vmovdqa32\t{$src, $dst|$dst, $src}", []>,
1253 def VMOVDQA64rr : AVX512BI<0x6F, MRMSrcReg, (outs VR512:$dst),
1255 "vmovdqa64\t{$src, $dst|$dst, $src}", []>,
1256 EVEX, EVEX_V512, VEX_W;
1257 let mayStore = 1 in {
1258 def VMOVDQA32mr : AVX512BI<0x7F, MRMDestMem, (outs),
1259 (ins i512mem:$dst, VR512:$src),
1260 "vmovdqa32\t{$src, $dst|$dst, $src}", []>,
1261 EVEX, EVEX_V512, EVEX_CD8<32, CD8VF>;
1262 def VMOVDQA64mr : AVX512BI<0x7F, MRMDestMem, (outs),
1263 (ins i512mem:$dst, VR512:$src),
1264 "vmovdqa64\t{$src, $dst|$dst, $src}", []>,
1265 EVEX, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
1267 let mayLoad = 1 in {
1268 def VMOVDQA32rm : AVX512BI<0x6F, MRMSrcMem, (outs VR512:$dst),
1270 "vmovdqa32\t{$src, $dst|$dst, $src}", []>,
1271 EVEX, EVEX_V512, EVEX_CD8<32, CD8VF>;
1272 def VMOVDQA64rm : AVX512BI<0x6F, MRMSrcMem, (outs VR512:$dst),
1274 "vmovdqa64\t{$src, $dst|$dst, $src}", []>,
1275 EVEX, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
1279 // 512-bit aligned load/store
1280 def : Pat<(alignedloadv8i64 addr:$src), (VMOVDQA64rm addr:$src)>;
1281 def : Pat<(alignedloadv16i32 addr:$src), (VMOVDQA32rm addr:$src)>;
1283 def : Pat<(alignedstore512 (v8i64 VR512:$src), addr:$dst),
1284 (VMOVDQA64mr addr:$dst, VR512:$src)>;
1285 def : Pat<(alignedstore512 (v16i32 VR512:$src), addr:$dst),
1286 (VMOVDQA32mr addr:$dst, VR512:$src)>;
1288 multiclass avx512_mov_int<bits<8> load_opc, bits<8> store_opc, string asm,
1289 RegisterClass RC, RegisterClass KRC,
1290 PatFrag ld_frag, X86MemOperand x86memop> {
1291 let hasSideEffects = 0 in
1292 def rr : AVX512XSI<load_opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
1293 !strconcat(asm, "\t{$src, $dst|$dst, $src}"), []>, EVEX;
1294 let canFoldAsLoad = 1 in
1295 def rm : AVX512XSI<load_opc, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
1296 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
1297 [(set RC:$dst, (ld_frag addr:$src))]>, EVEX;
1299 def mr : AVX512XSI<store_opc, MRMDestMem, (outs),
1300 (ins x86memop:$dst, VR512:$src),
1301 !strconcat(asm, "\t{$src, $dst|$dst, $src}"), []>, EVEX;
1302 let Constraints = "$src1 = $dst" in {
1303 def rrk : AVX512XSI<load_opc, MRMSrcReg, (outs RC:$dst),
1304 (ins RC:$src1, KRC:$mask, RC:$src2),
1306 "\t{$src2, ${dst} {${mask}}|${dst} {${mask}}, $src2}"), []>,
1308 def rmk : AVX512XSI<load_opc, MRMSrcMem, (outs RC:$dst),
1309 (ins RC:$src1, KRC:$mask, x86memop:$src2),
1311 "\t{$src2, ${dst} {${mask}}|${dst} {${mask}}, $src2}"),
1316 defm VMOVDQU32 : avx512_mov_int<0x6F, 0x7F, "vmovdqu32", VR512, VK16WM,
1317 memopv16i32, i512mem>,
1318 EVEX_V512, EVEX_CD8<32, CD8VF>;
1319 defm VMOVDQU64 : avx512_mov_int<0x6F, 0x7F, "vmovdqu64", VR512, VK8WM,
1320 memopv8i64, i512mem>,
1321 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
1323 // 512-bit unaligned load/store
1324 def : Pat<(loadv8i64 addr:$src), (VMOVDQU64rm addr:$src)>;
1325 def : Pat<(loadv16i32 addr:$src), (VMOVDQU32rm addr:$src)>;
1327 def : Pat<(store (v8i64 VR512:$src), addr:$dst),
1328 (VMOVDQU64mr addr:$dst, VR512:$src)>;
1329 def : Pat<(store (v16i32 VR512:$src), addr:$dst),
1330 (VMOVDQU32mr addr:$dst, VR512:$src)>;
1332 let AddedComplexity = 20 in {
1333 def : Pat<(v16f32 (vselect VK16WM:$mask, (v16f32 VR512:$src1),
1334 (v16f32 VR512:$src2))),
1335 (VMOVUPSZrrk VR512:$src2, VK16WM:$mask, VR512:$src1)>;
1336 def : Pat<(v8f64 (vselect VK8WM:$mask, (v8f64 VR512:$src1),
1337 (v8f64 VR512:$src2))),
1338 (VMOVUPDZrrk VR512:$src2, VK8WM:$mask, VR512:$src1)>;
1339 def : Pat<(v16i32 (vselect VK16WM:$mask, (v16i32 VR512:$src1),
1340 (v16i32 VR512:$src2))),
1341 (VMOVDQU32rrk VR512:$src2, VK16WM:$mask, VR512:$src1)>;
1342 def : Pat<(v8i64 (vselect VK8WM:$mask, (v8i64 VR512:$src1),
1343 (v8i64 VR512:$src2))),
1344 (VMOVDQU64rrk VR512:$src2, VK8WM:$mask, VR512:$src1)>;
1346 // Move Int Doubleword to Packed Double Int
1348 def VMOVDI2PDIZrr : AVX512SI<0x6E, MRMSrcReg, (outs VR128X:$dst), (ins GR32:$src),
1349 "vmovd\t{$src, $dst|$dst, $src}",
1351 (v4i32 (scalar_to_vector GR32:$src)))], IIC_SSE_MOVDQ>,
1353 def VMOVDI2PDIZrm : AVX512SI<0x6E, MRMSrcMem, (outs VR128X:$dst), (ins i32mem:$src),
1354 "vmovd\t{$src, $dst|$dst, $src}",
1356 (v4i32 (scalar_to_vector (loadi32 addr:$src))))],
1357 IIC_SSE_MOVDQ>, EVEX, VEX_LIG, EVEX_CD8<32, CD8VT1>;
1358 def VMOV64toPQIZrr : AVX512SI<0x6E, MRMSrcReg, (outs VR128X:$dst), (ins GR64:$src),
1359 "vmovq\t{$src, $dst|$dst, $src}",
1361 (v2i64 (scalar_to_vector GR64:$src)))],
1362 IIC_SSE_MOVDQ>, EVEX, VEX_W, VEX_LIG;
1363 let isCodeGenOnly = 1 in {
1364 def VMOV64toSDZrr : AVX512SI<0x6E, MRMSrcReg, (outs FR64:$dst), (ins GR64:$src),
1365 "vmovq\t{$src, $dst|$dst, $src}",
1366 [(set FR64:$dst, (bitconvert GR64:$src))],
1367 IIC_SSE_MOVDQ>, EVEX, VEX_W, Sched<[WriteMove]>;
1368 def VMOVSDto64Zrr : AVX512SI<0x7E, MRMDestReg, (outs GR64:$dst), (ins FR64:$src),
1369 "vmovq\t{$src, $dst|$dst, $src}",
1370 [(set GR64:$dst, (bitconvert FR64:$src))],
1371 IIC_SSE_MOVDQ>, EVEX, VEX_W, Sched<[WriteMove]>;
1373 def VMOVSDto64Zmr : AVX512SI<0x7E, MRMDestMem, (outs), (ins i64mem:$dst, FR64:$src),
1374 "vmovq\t{$src, $dst|$dst, $src}",
1375 [(store (i64 (bitconvert FR64:$src)), addr:$dst)],
1376 IIC_SSE_MOVDQ>, EVEX, VEX_W, Sched<[WriteStore]>,
1377 EVEX_CD8<64, CD8VT1>;
1379 // Move Int Doubleword to Single Scalar
1381 let isCodeGenOnly = 1 in {
1382 def VMOVDI2SSZrr : AVX512SI<0x6E, MRMSrcReg, (outs FR32X:$dst), (ins GR32:$src),
1383 "vmovd\t{$src, $dst|$dst, $src}",
1384 [(set FR32X:$dst, (bitconvert GR32:$src))],
1385 IIC_SSE_MOVDQ>, EVEX, VEX_LIG;
1387 def VMOVDI2SSZrm : AVX512SI<0x6E, MRMSrcMem, (outs FR32X:$dst), (ins i32mem:$src),
1388 "vmovd\t{$src, $dst|$dst, $src}",
1389 [(set FR32X:$dst, (bitconvert (loadi32 addr:$src)))],
1390 IIC_SSE_MOVDQ>, EVEX, VEX_LIG, EVEX_CD8<32, CD8VT1>;
1393 // Move Packed Doubleword Int to Packed Double Int
1395 def VMOVPDI2DIZrr : AVX512SI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128X:$src),
1396 "vmovd\t{$src, $dst|$dst, $src}",
1397 [(set GR32:$dst, (vector_extract (v4i32 VR128X:$src),
1398 (iPTR 0)))], IIC_SSE_MOVD_ToGP>,
1400 def VMOVPDI2DIZmr : AVX512SI<0x7E, MRMDestMem, (outs),
1401 (ins i32mem:$dst, VR128X:$src),
1402 "vmovd\t{$src, $dst|$dst, $src}",
1403 [(store (i32 (vector_extract (v4i32 VR128X:$src),
1404 (iPTR 0))), addr:$dst)], IIC_SSE_MOVDQ>,
1405 EVEX, VEX_LIG, EVEX_CD8<32, CD8VT1>;
1407 // Move Packed Doubleword Int first element to Doubleword Int
1409 def VMOVPQIto64Zrr : I<0x7E, MRMDestReg, (outs GR64:$dst), (ins VR128X:$src),
1410 "vmovq\t{$src, $dst|$dst, $src}",
1411 [(set GR64:$dst, (extractelt (v2i64 VR128X:$src),
1413 IIC_SSE_MOVD_ToGP>, TB, OpSize, EVEX, VEX_LIG, VEX_W,
1414 Requires<[HasAVX512, In64BitMode]>;
1416 def VMOVPQIto64Zmr : I<0xD6, MRMDestMem, (outs),
1417 (ins i64mem:$dst, VR128X:$src),
1418 "vmovq\t{$src, $dst|$dst, $src}",
1419 [(store (extractelt (v2i64 VR128X:$src), (iPTR 0)),
1420 addr:$dst)], IIC_SSE_MOVDQ>,
1421 EVEX, OpSize, VEX_LIG, VEX_W, TB, EVEX_CD8<64, CD8VT1>,
1422 Sched<[WriteStore]>, Requires<[HasAVX512, In64BitMode]>;
1424 // Move Scalar Single to Double Int
1426 let isCodeGenOnly = 1 in {
1427 def VMOVSS2DIZrr : AVX512SI<0x7E, MRMDestReg, (outs GR32:$dst),
1429 "vmovd\t{$src, $dst|$dst, $src}",
1430 [(set GR32:$dst, (bitconvert FR32X:$src))],
1431 IIC_SSE_MOVD_ToGP>, EVEX, VEX_LIG;
1432 def VMOVSS2DIZmr : AVX512SI<0x7E, MRMDestMem, (outs),
1433 (ins i32mem:$dst, FR32X:$src),
1434 "vmovd\t{$src, $dst|$dst, $src}",
1435 [(store (i32 (bitconvert FR32X:$src)), addr:$dst)],
1436 IIC_SSE_MOVDQ>, EVEX, VEX_LIG, EVEX_CD8<32, CD8VT1>;
1439 // Move Quadword Int to Packed Quadword Int
1441 def VMOVQI2PQIZrm : AVX512SI<0x6E, MRMSrcMem, (outs VR128X:$dst),
1443 "vmovq\t{$src, $dst|$dst, $src}",
1445 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>,
1446 EVEX, VEX_LIG, VEX_W, EVEX_CD8<64, CD8VT1>;
1448 //===----------------------------------------------------------------------===//
1449 // AVX-512 MOVSS, MOVSD
1450 //===----------------------------------------------------------------------===//
1452 multiclass avx512_move_scalar <string asm, RegisterClass RC,
1453 SDNode OpNode, ValueType vt,
1454 X86MemOperand x86memop, PatFrag mem_pat> {
1455 let hasSideEffects = 0 in {
1456 def rr : SI<0x10, MRMSrcReg, (outs VR128X:$dst), (ins VR128X:$src1, RC:$src2),
1457 !strconcat(asm, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1458 [(set VR128X:$dst, (vt (OpNode VR128X:$src1,
1459 (scalar_to_vector RC:$src2))))],
1460 IIC_SSE_MOV_S_RR>, EVEX_4V, VEX_LIG;
1461 let Constraints = "$src1 = $dst" in
1462 def rrk : SI<0x10, MRMSrcReg, (outs VR128X:$dst),
1463 (ins VR128X:$src1, VK1WM:$mask, RC:$src2, RC:$src3),
1465 "\t{$src3, $src2, $dst {${mask}}|$dst {${mask}}, $src2, $src3}"),
1466 [], IIC_SSE_MOV_S_RR>, EVEX_4V, VEX_LIG, EVEX_K;
1467 def rm : SI<0x10, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
1468 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
1469 [(set RC:$dst, (mem_pat addr:$src))], IIC_SSE_MOV_S_RM>,
1471 def mr: SI<0x11, MRMDestMem, (outs), (ins x86memop:$dst, RC:$src),
1472 !strconcat(asm, "\t{$src, $dst|$dst, $src}"),
1473 [(store RC:$src, addr:$dst)], IIC_SSE_MOV_S_MR>,
1475 } //hasSideEffects = 0
1478 let ExeDomain = SSEPackedSingle in
1479 defm VMOVSSZ : avx512_move_scalar<"movss", FR32X, X86Movss, v4f32, f32mem,
1480 loadf32>, XS, EVEX_CD8<32, CD8VT1>;
1482 let ExeDomain = SSEPackedDouble in
1483 defm VMOVSDZ : avx512_move_scalar<"movsd", FR64X, X86Movsd, v2f64, f64mem,
1484 loadf64>, XD, VEX_W, EVEX_CD8<64, CD8VT1>;
1486 def : Pat<(f32 (X86select VK1WM:$mask, (f32 FR32X:$src1), (f32 FR32X:$src2))),
1487 (COPY_TO_REGCLASS (VMOVSSZrrk (COPY_TO_REGCLASS FR32X:$src2, VR128X),
1488 VK1WM:$mask, (f32 (IMPLICIT_DEF)), FR32X:$src1), FR32X)>;
1490 def : Pat<(f64 (X86select VK1WM:$mask, (f64 FR64X:$src1), (f64 FR64X:$src2))),
1491 (COPY_TO_REGCLASS (VMOVSDZrrk (COPY_TO_REGCLASS FR64X:$src2, VR128X),
1492 VK1WM:$mask, (f64 (IMPLICIT_DEF)), FR64X:$src1), FR64X)>;
1494 // For the disassembler
1495 let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in {
1496 def VMOVSSZrr_REV : SI<0x11, MRMDestReg, (outs VR128X:$dst),
1497 (ins VR128X:$src1, FR32X:$src2),
1498 "movss\t{$src2, $src1, $dst|$dst, $src1, $src2}", [],
1500 XS, EVEX_4V, VEX_LIG;
1501 def VMOVSDZrr_REV : SI<0x11, MRMDestReg, (outs VR128X:$dst),
1502 (ins VR128X:$src1, FR64X:$src2),
1503 "movsd\t{$src2, $src1, $dst|$dst, $src1, $src2}", [],
1505 XD, EVEX_4V, VEX_LIG, VEX_W;
1508 let Predicates = [HasAVX512] in {
1509 let AddedComplexity = 15 in {
1510 // Move scalar to XMM zero-extended, zeroing a VR128X then do a
1511 // MOVS{S,D} to the lower bits.
1512 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32X:$src)))),
1513 (VMOVSSZrr (v4f32 (V_SET0)), FR32X:$src)>;
1514 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128X:$src))),
1515 (VMOVSSZrr (v4f32 (V_SET0)), (COPY_TO_REGCLASS VR128X:$src, FR32X))>;
1516 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128X:$src))),
1517 (VMOVSSZrr (v4i32 (V_SET0)), (COPY_TO_REGCLASS VR128X:$src, FR32X))>;
1518 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64X:$src)))),
1519 (VMOVSDZrr (v2f64 (V_SET0)), FR64X:$src)>;
1521 // Move low f32 and clear high bits.
1522 def : Pat<(v8f32 (X86vzmovl (v8f32 VR256X:$src))),
1523 (SUBREG_TO_REG (i32 0),
1524 (VMOVSSZrr (v4f32 (V_SET0)),
1525 (EXTRACT_SUBREG (v8f32 VR256X:$src), sub_xmm)), sub_xmm)>;
1526 def : Pat<(v8i32 (X86vzmovl (v8i32 VR256X:$src))),
1527 (SUBREG_TO_REG (i32 0),
1528 (VMOVSSZrr (v4i32 (V_SET0)),
1529 (EXTRACT_SUBREG (v8i32 VR256X:$src), sub_xmm)), sub_xmm)>;
1532 let AddedComplexity = 20 in {
1533 // MOVSSrm zeros the high parts of the register; represent this
1534 // with SUBREG_TO_REG. The AVX versions also write: DST[255:128] <- 0
1535 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
1536 (COPY_TO_REGCLASS (VMOVSSZrm addr:$src), VR128X)>;
1537 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
1538 (COPY_TO_REGCLASS (VMOVSSZrm addr:$src), VR128X)>;
1539 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
1540 (COPY_TO_REGCLASS (VMOVSSZrm addr:$src), VR128X)>;
1542 // MOVSDrm zeros the high parts of the register; represent this
1543 // with SUBREG_TO_REG. The AVX versions also write: DST[255:128] <- 0
1544 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
1545 (COPY_TO_REGCLASS (VMOVSDZrm addr:$src), VR128X)>;
1546 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
1547 (COPY_TO_REGCLASS (VMOVSDZrm addr:$src), VR128X)>;
1548 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
1549 (COPY_TO_REGCLASS (VMOVSDZrm addr:$src), VR128X)>;
1550 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
1551 (COPY_TO_REGCLASS (VMOVSDZrm addr:$src), VR128X)>;
1552 def : Pat<(v2f64 (X86vzload addr:$src)),
1553 (COPY_TO_REGCLASS (VMOVSDZrm addr:$src), VR128X)>;
1555 // Represent the same patterns above but in the form they appear for
1557 def : Pat<(v8i32 (X86vzmovl (insert_subvector undef,
1558 (v4i32 (scalar_to_vector (loadi32 addr:$src))), (iPTR 0)))),
1559 (SUBREG_TO_REG (i32 0), (VMOVDI2PDIZrm addr:$src), sub_xmm)>;
1560 def : Pat<(v8f32 (X86vzmovl (insert_subvector undef,
1561 (v4f32 (scalar_to_vector (loadf32 addr:$src))), (iPTR 0)))),
1562 (SUBREG_TO_REG (i32 0), (VMOVSSZrm addr:$src), sub_xmm)>;
1563 def : Pat<(v4f64 (X86vzmovl (insert_subvector undef,
1564 (v2f64 (scalar_to_vector (loadf64 addr:$src))), (iPTR 0)))),
1565 (SUBREG_TO_REG (i32 0), (VMOVSDZrm addr:$src), sub_xmm)>;
1567 def : Pat<(v8f32 (X86vzmovl (insert_subvector undef,
1568 (v4f32 (scalar_to_vector FR32X:$src)), (iPTR 0)))),
1569 (SUBREG_TO_REG (i32 0), (v4f32 (VMOVSSZrr (v4f32 (V_SET0)),
1570 FR32X:$src)), sub_xmm)>;
1571 def : Pat<(v4f64 (X86vzmovl (insert_subvector undef,
1572 (v2f64 (scalar_to_vector FR64X:$src)), (iPTR 0)))),
1573 (SUBREG_TO_REG (i64 0), (v2f64 (VMOVSDZrr (v2f64 (V_SET0)),
1574 FR64X:$src)), sub_xmm)>;
1575 def : Pat<(v4i64 (X86vzmovl (insert_subvector undef,
1576 (v2i64 (scalar_to_vector (loadi64 addr:$src))), (iPTR 0)))),
1577 (SUBREG_TO_REG (i64 0), (VMOVQI2PQIZrm addr:$src), sub_xmm)>;
1579 // Move low f64 and clear high bits.
1580 def : Pat<(v4f64 (X86vzmovl (v4f64 VR256X:$src))),
1581 (SUBREG_TO_REG (i32 0),
1582 (VMOVSDZrr (v2f64 (V_SET0)),
1583 (EXTRACT_SUBREG (v4f64 VR256X:$src), sub_xmm)), sub_xmm)>;
1585 def : Pat<(v4i64 (X86vzmovl (v4i64 VR256X:$src))),
1586 (SUBREG_TO_REG (i32 0), (VMOVSDZrr (v2i64 (V_SET0)),
1587 (EXTRACT_SUBREG (v4i64 VR256X:$src), sub_xmm)), sub_xmm)>;
1589 // Extract and store.
1590 def : Pat<(store (f32 (vector_extract (v4f32 VR128X:$src), (iPTR 0))),
1592 (VMOVSSZmr addr:$dst, (COPY_TO_REGCLASS (v4f32 VR128X:$src), FR32X))>;
1593 def : Pat<(store (f64 (vector_extract (v2f64 VR128X:$src), (iPTR 0))),
1595 (VMOVSDZmr addr:$dst, (COPY_TO_REGCLASS (v2f64 VR128X:$src), FR64X))>;
1597 // Shuffle with VMOVSS
1598 def : Pat<(v4i32 (X86Movss VR128X:$src1, VR128X:$src2)),
1599 (VMOVSSZrr (v4i32 VR128X:$src1),
1600 (COPY_TO_REGCLASS (v4i32 VR128X:$src2), FR32X))>;
1601 def : Pat<(v4f32 (X86Movss VR128X:$src1, VR128X:$src2)),
1602 (VMOVSSZrr (v4f32 VR128X:$src1),
1603 (COPY_TO_REGCLASS (v4f32 VR128X:$src2), FR32X))>;
1606 def : Pat<(v8i32 (X86Movss VR256X:$src1, VR256X:$src2)),
1607 (SUBREG_TO_REG (i32 0),
1608 (VMOVSSZrr (EXTRACT_SUBREG (v8i32 VR256X:$src1), sub_xmm),
1609 (EXTRACT_SUBREG (v8i32 VR256X:$src2), sub_xmm)),
1611 def : Pat<(v8f32 (X86Movss VR256X:$src1, VR256X:$src2)),
1612 (SUBREG_TO_REG (i32 0),
1613 (VMOVSSZrr (EXTRACT_SUBREG (v8f32 VR256X:$src1), sub_xmm),
1614 (EXTRACT_SUBREG (v8f32 VR256X:$src2), sub_xmm)),
1617 // Shuffle with VMOVSD
1618 def : Pat<(v2i64 (X86Movsd VR128X:$src1, VR128X:$src2)),
1619 (VMOVSDZrr VR128X:$src1, (COPY_TO_REGCLASS VR128X:$src2, FR64X))>;
1620 def : Pat<(v2f64 (X86Movsd VR128X:$src1, VR128X:$src2)),
1621 (VMOVSDZrr VR128X:$src1, (COPY_TO_REGCLASS VR128X:$src2, FR64X))>;
1622 def : Pat<(v4f32 (X86Movsd VR128X:$src1, VR128X:$src2)),
1623 (VMOVSDZrr VR128X:$src1, (COPY_TO_REGCLASS VR128X:$src2, FR64X))>;
1624 def : Pat<(v4i32 (X86Movsd VR128X:$src1, VR128X:$src2)),
1625 (VMOVSDZrr VR128X:$src1, (COPY_TO_REGCLASS VR128X:$src2, FR64X))>;
1628 def : Pat<(v4i64 (X86Movsd VR256X:$src1, VR256X:$src2)),
1629 (SUBREG_TO_REG (i32 0),
1630 (VMOVSDZrr (EXTRACT_SUBREG (v4i64 VR256X:$src1), sub_xmm),
1631 (EXTRACT_SUBREG (v4i64 VR256X:$src2), sub_xmm)),
1633 def : Pat<(v4f64 (X86Movsd VR256X:$src1, VR256X:$src2)),
1634 (SUBREG_TO_REG (i32 0),
1635 (VMOVSDZrr (EXTRACT_SUBREG (v4f64 VR256X:$src1), sub_xmm),
1636 (EXTRACT_SUBREG (v4f64 VR256X:$src2), sub_xmm)),
1639 def : Pat<(v2f64 (X86Movlpd VR128X:$src1, VR128X:$src2)),
1640 (VMOVSDZrr VR128X:$src1, (COPY_TO_REGCLASS VR128X:$src2, FR64X))>;
1641 def : Pat<(v2i64 (X86Movlpd VR128X:$src1, VR128X:$src2)),
1642 (VMOVSDZrr VR128X:$src1, (COPY_TO_REGCLASS VR128X:$src2, FR64X))>;
1643 def : Pat<(v4f32 (X86Movlps VR128X:$src1, VR128X:$src2)),
1644 (VMOVSDZrr VR128X:$src1, (COPY_TO_REGCLASS VR128X:$src2, FR64X))>;
1645 def : Pat<(v4i32 (X86Movlps VR128X:$src1, VR128X:$src2)),
1646 (VMOVSDZrr VR128X:$src1, (COPY_TO_REGCLASS VR128X:$src2, FR64X))>;
1649 let AddedComplexity = 15 in
1650 def VMOVZPQILo2PQIZrr : AVX512XSI<0x7E, MRMSrcReg, (outs VR128X:$dst),
1652 "vmovq\t{$src, $dst|$dst, $src}",
1653 [(set VR128X:$dst, (v2i64 (X86vzmovl
1654 (v2i64 VR128X:$src))))],
1655 IIC_SSE_MOVQ_RR>, EVEX, VEX_W;
1657 let AddedComplexity = 20 in
1658 def VMOVZPQILo2PQIZrm : AVX512XSI<0x7E, MRMSrcMem, (outs VR128X:$dst),
1660 "vmovq\t{$src, $dst|$dst, $src}",
1661 [(set VR128X:$dst, (v2i64 (X86vzmovl
1662 (loadv2i64 addr:$src))))],
1663 IIC_SSE_MOVDQ>, EVEX, VEX_W,
1664 EVEX_CD8<8, CD8VT8>;
1666 let Predicates = [HasAVX512] in {
1667 // AVX 128-bit movd/movq instruction write zeros in the high 128-bit part.
1668 let AddedComplexity = 20 in {
1669 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector (loadi32 addr:$src))))),
1670 (VMOVDI2PDIZrm addr:$src)>;
1671 def : Pat<(v2i64 (X86vzmovl (v2i64 (scalar_to_vector GR64:$src)))),
1672 (VMOV64toPQIZrr GR64:$src)>;
1673 def : Pat<(v4i32 (X86vzmovl (v4i32 (scalar_to_vector GR32:$src)))),
1674 (VMOVDI2PDIZrr GR32:$src)>;
1676 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
1677 (VMOVDI2PDIZrm addr:$src)>;
1678 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
1679 (VMOVDI2PDIZrm addr:$src)>;
1680 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
1681 (VMOVZPQILo2PQIZrm addr:$src)>;
1682 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128X:$src))),
1683 (VMOVZPQILo2PQIZrr VR128X:$src)>;
1684 def : Pat<(v2i64 (X86vzload addr:$src)),
1685 (VMOVZPQILo2PQIZrm addr:$src)>;
1688 // Use regular 128-bit instructions to match 256-bit scalar_to_vec+zext.
1689 def : Pat<(v8i32 (X86vzmovl (insert_subvector undef,
1690 (v4i32 (scalar_to_vector GR32:$src)),(iPTR 0)))),
1691 (SUBREG_TO_REG (i32 0), (VMOVDI2PDIZrr GR32:$src), sub_xmm)>;
1692 def : Pat<(v4i64 (X86vzmovl (insert_subvector undef,
1693 (v2i64 (scalar_to_vector GR64:$src)),(iPTR 0)))),
1694 (SUBREG_TO_REG (i64 0), (VMOV64toPQIZrr GR64:$src), sub_xmm)>;
1697 def : Pat<(v16i32 (X86Vinsert (v16i32 immAllZerosV), GR32:$src2, (iPTR 0))),
1698 (SUBREG_TO_REG (i32 0), (VMOVDI2PDIZrr GR32:$src2), sub_xmm)>;
1700 def : Pat<(v8i64 (X86Vinsert (bc_v8i64 (v16i32 immAllZerosV)), GR64:$src2, (iPTR 0))),
1701 (SUBREG_TO_REG (i32 0), (VMOV64toPQIZrr GR64:$src2), sub_xmm)>;
1703 def : Pat<(v16i32 (X86Vinsert undef, GR32:$src2, (iPTR 0))),
1704 (SUBREG_TO_REG (i32 0), (VMOVDI2PDIZrr GR32:$src2), sub_xmm)>;
1706 def : Pat<(v8i64 (X86Vinsert undef, GR64:$src2, (iPTR 0))),
1707 (SUBREG_TO_REG (i32 0), (VMOV64toPQIZrr GR64:$src2), sub_xmm)>;
1709 //===----------------------------------------------------------------------===//
1710 // AVX-512 - Integer arithmetic
1712 multiclass avx512_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1713 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
1714 X86MemOperand x86memop, PatFrag scalar_mfrag,
1715 X86MemOperand x86scalar_mop, string BrdcstStr,
1716 OpndItins itins, bit IsCommutable = 0> {
1717 let isCommutable = IsCommutable in
1718 def rr : AVX512BI<opc, MRMSrcReg, (outs RC:$dst),
1719 (ins RC:$src1, RC:$src2),
1720 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1721 [(set RC:$dst, (OpVT (OpNode (OpVT RC:$src1), (OpVT RC:$src2))))],
1723 def rm : AVX512BI<opc, MRMSrcMem, (outs RC:$dst),
1724 (ins RC:$src1, x86memop:$src2),
1725 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1726 [(set RC:$dst, (OpVT (OpNode (OpVT RC:$src1), (memop_frag addr:$src2))))],
1728 def rmb : AVX512BI<opc, MRMSrcMem, (outs RC:$dst),
1729 (ins RC:$src1, x86scalar_mop:$src2),
1730 !strconcat(OpcodeStr, "\t{${src2}", BrdcstStr,
1731 ", $src1, $dst|$dst, $src1, ${src2}", BrdcstStr, "}"),
1732 [(set RC:$dst, (OpNode RC:$src1,
1733 (OpVT (X86VBroadcast (scalar_mfrag addr:$src2)))))],
1734 itins.rm>, EVEX_4V, EVEX_B;
1736 multiclass avx512_binop_rm2<bits<8> opc, string OpcodeStr,
1737 ValueType DstVT, ValueType SrcVT, RegisterClass RC,
1738 PatFrag memop_frag, X86MemOperand x86memop,
1740 bit IsCommutable = 0> {
1741 let isCommutable = IsCommutable in
1742 def rr : AVX512BI<opc, MRMSrcReg, (outs RC:$dst),
1743 (ins RC:$src1, RC:$src2),
1744 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1745 []>, EVEX_4V, VEX_W;
1746 def rm : AVX512BI<opc, MRMSrcMem, (outs RC:$dst),
1747 (ins RC:$src1, x86memop:$src2),
1748 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1749 []>, EVEX_4V, VEX_W;
1752 defm VPADDDZ : avx512_binop_rm<0xFE, "vpaddd", add, v16i32, VR512, memopv16i32,
1753 i512mem, loadi32, i32mem, "{1to16}", SSE_INTALU_ITINS_P, 1>,
1754 EVEX_V512, EVEX_CD8<32, CD8VF>;
1756 defm VPSUBDZ : avx512_binop_rm<0xFA, "vpsubd", sub, v16i32, VR512, memopv16i32,
1757 i512mem, loadi32, i32mem, "{1to16}", SSE_INTALU_ITINS_P, 0>,
1758 EVEX_V512, EVEX_CD8<32, CD8VF>;
1760 defm VPMULLDZ : avx512_binop_rm<0x40, "vpmulld", mul, v16i32, VR512, memopv16i32,
1761 i512mem, loadi32, i32mem, "{1to16}", SSE_INTALU_ITINS_P, 1>,
1762 T8, EVEX_V512, EVEX_CD8<32, CD8VF>;
1764 defm VPADDQZ : avx512_binop_rm<0xD4, "vpaddq", add, v8i64, VR512, memopv8i64,
1765 i512mem, loadi64, i64mem, "{1to8}", SSE_INTALU_ITINS_P, 1>,
1766 EVEX_CD8<64, CD8VF>, EVEX_V512, VEX_W;
1768 defm VPSUBQZ : avx512_binop_rm<0xFB, "vpsubq", sub, v8i64, VR512, memopv8i64,
1769 i512mem, loadi64, i64mem, "{1to8}", SSE_INTALU_ITINS_P, 0>,
1770 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
1772 defm VPMULDQZ : avx512_binop_rm2<0x28, "vpmuldq", v8i64, v16i32,
1773 VR512, memopv8i64, i512mem, SSE_INTALU_ITINS_P, 1>, T8,
1774 EVEX_V512, EVEX_CD8<64, CD8VF>;
1776 defm VPMULUDQZ : avx512_binop_rm2<0xF4, "vpmuludq", v8i64, v16i32,
1777 VR512, memopv8i64, i512mem, SSE_INTMUL_ITINS_P, 1>, EVEX_V512,
1778 EVEX_CD8<64, CD8VF>;
1780 def : Pat<(v8i64 (X86pmuludq (v16i32 VR512:$src1), (v16i32 VR512:$src2))),
1781 (VPMULUDQZrr VR512:$src1, VR512:$src2)>;
1783 defm VPMAXUDZ : avx512_binop_rm<0x3F, "vpmaxud", X86umax, v16i32, VR512, memopv16i32,
1784 i512mem, loadi32, i32mem, "{1to16}", SSE_INTALU_ITINS_P, 1>,
1785 T8, EVEX_V512, EVEX_CD8<32, CD8VF>;
1786 defm VPMAXUQZ : avx512_binop_rm<0x3F, "vpmaxuq", X86umax, v8i64, VR512, memopv8i64,
1787 i512mem, loadi64, i64mem, "{1to8}", SSE_INTALU_ITINS_P, 0>,
1788 T8, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
1790 defm VPMAXSDZ : avx512_binop_rm<0x3D, "vpmaxsd", X86smax, v16i32, VR512, memopv16i32,
1791 i512mem, loadi32, i32mem, "{1to16}", SSE_INTALU_ITINS_P, 1>,
1792 EVEX_V512, EVEX_CD8<32, CD8VF>;
1793 defm VPMAXSQZ : avx512_binop_rm<0x3D, "vpmaxsq", X86smax, v8i64, VR512, memopv8i64,
1794 i512mem, loadi64, i64mem, "{1to8}", SSE_INTALU_ITINS_P, 0>,
1795 T8, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
1797 defm VPMINUDZ : avx512_binop_rm<0x3B, "vpminud", X86umin, v16i32, VR512, memopv16i32,
1798 i512mem, loadi32, i32mem, "{1to16}", SSE_INTALU_ITINS_P, 1>,
1799 T8, EVEX_V512, EVEX_CD8<32, CD8VF>;
1800 defm VPMINUQZ : avx512_binop_rm<0x3B, "vpminuq", X86umin, v8i64, VR512, memopv8i64,
1801 i512mem, loadi64, i64mem, "{1to8}", SSE_INTALU_ITINS_P, 0>,
1802 T8, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
1804 defm VPMINSDZ : avx512_binop_rm<0x39, "vpminsd", X86smin, v16i32, VR512, memopv16i32,
1805 i512mem, loadi32, i32mem, "{1to16}", SSE_INTALU_ITINS_P, 1>,
1806 T8, EVEX_V512, EVEX_CD8<32, CD8VF>;
1807 defm VPMINSQZ : avx512_binop_rm<0x39, "vpminsq", X86smin, v8i64, VR512, memopv8i64,
1808 i512mem, loadi64, i64mem, "{1to8}", SSE_INTALU_ITINS_P, 0>,
1809 T8, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
1811 //===----------------------------------------------------------------------===//
1812 // AVX-512 - Unpack Instructions
1813 //===----------------------------------------------------------------------===//
1815 multiclass avx512_unpack_fp<bits<8> opc, SDNode OpNode, ValueType vt,
1816 PatFrag mem_frag, RegisterClass RC,
1817 X86MemOperand x86memop, string asm,
1819 def rr : AVX512PI<opc, MRMSrcReg,
1820 (outs RC:$dst), (ins RC:$src1, RC:$src2),
1822 (vt (OpNode RC:$src1, RC:$src2)))],
1824 def rm : AVX512PI<opc, MRMSrcMem,
1825 (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
1827 (vt (OpNode RC:$src1,
1828 (bitconvert (mem_frag addr:$src2)))))],
1832 defm VUNPCKHPSZ: avx512_unpack_fp<0x15, X86Unpckh, v16f32, memopv8f64,
1833 VR512, f512mem, "vunpckhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1834 SSEPackedSingle>, EVEX_V512, EVEX_CD8<32, CD8VF>;
1835 defm VUNPCKHPDZ: avx512_unpack_fp<0x15, X86Unpckh, v8f64, memopv8f64,
1836 VR512, f512mem, "vunpckhpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1837 SSEPackedDouble>, OpSize, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
1838 defm VUNPCKLPSZ: avx512_unpack_fp<0x14, X86Unpckl, v16f32, memopv8f64,
1839 VR512, f512mem, "vunpcklps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1840 SSEPackedSingle>, EVEX_V512, EVEX_CD8<32, CD8VF>;
1841 defm VUNPCKLPDZ: avx512_unpack_fp<0x14, X86Unpckl, v8f64, memopv8f64,
1842 VR512, f512mem, "vunpcklpd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
1843 SSEPackedDouble>, OpSize, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
1845 multiclass avx512_unpack_int<bits<8> opc, string OpcodeStr, SDNode OpNode,
1846 ValueType OpVT, RegisterClass RC, PatFrag memop_frag,
1847 X86MemOperand x86memop> {
1848 def rr : AVX512BI<opc, MRMSrcReg, (outs RC:$dst),
1849 (ins RC:$src1, RC:$src2),
1850 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1851 [(set RC:$dst, (OpVT (OpNode (OpVT RC:$src1), (OpVT RC:$src2))))],
1852 IIC_SSE_UNPCK>, EVEX_4V;
1853 def rm : AVX512BI<opc, MRMSrcMem, (outs RC:$dst),
1854 (ins RC:$src1, x86memop:$src2),
1855 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1856 [(set RC:$dst, (OpVT (OpNode (OpVT RC:$src1),
1857 (bitconvert (memop_frag addr:$src2)))))],
1858 IIC_SSE_UNPCK>, EVEX_4V;
1860 defm VPUNPCKLDQZ : avx512_unpack_int<0x62, "vpunpckldq", X86Unpckl, v16i32,
1861 VR512, memopv16i32, i512mem>, EVEX_V512,
1862 EVEX_CD8<32, CD8VF>;
1863 defm VPUNPCKLQDQZ : avx512_unpack_int<0x6C, "vpunpcklqdq", X86Unpckl, v8i64,
1864 VR512, memopv8i64, i512mem>, EVEX_V512,
1865 VEX_W, EVEX_CD8<64, CD8VF>;
1866 defm VPUNPCKHDQZ : avx512_unpack_int<0x6A, "vpunpckhdq", X86Unpckh, v16i32,
1867 VR512, memopv16i32, i512mem>, EVEX_V512,
1868 EVEX_CD8<32, CD8VF>;
1869 defm VPUNPCKHQDQZ : avx512_unpack_int<0x6D, "vpunpckhqdq", X86Unpckh, v8i64,
1870 VR512, memopv8i64, i512mem>, EVEX_V512,
1871 VEX_W, EVEX_CD8<64, CD8VF>;
1872 //===----------------------------------------------------------------------===//
1876 multiclass avx512_pshuf_imm<bits<8> opc, string OpcodeStr, RegisterClass RC,
1877 SDNode OpNode, PatFrag mem_frag,
1878 X86MemOperand x86memop, ValueType OpVT> {
1879 def ri : AVX512Ii8<opc, MRMSrcReg, (outs RC:$dst),
1880 (ins RC:$src1, i8imm:$src2),
1881 !strconcat(OpcodeStr,
1882 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1884 (OpVT (OpNode RC:$src1, (i8 imm:$src2))))]>,
1886 def mi : AVX512Ii8<opc, MRMSrcMem, (outs RC:$dst),
1887 (ins x86memop:$src1, i8imm:$src2),
1888 !strconcat(OpcodeStr,
1889 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1891 (OpVT (OpNode (mem_frag addr:$src1),
1892 (i8 imm:$src2))))]>, EVEX;
1895 defm VPSHUFDZ : avx512_pshuf_imm<0x70, "vpshufd", VR512, X86PShufd, memopv16i32,
1896 i512mem, v16i32>, OpSize, EVEX_V512, EVEX_CD8<32, CD8VF>;
1898 let ExeDomain = SSEPackedSingle in
1899 defm VPERMILPSZ : avx512_pshuf_imm<0x04, "vpermilps", VR512, X86VPermilp,
1900 memopv16f32, i512mem, v16f32>, OpSize, TA, EVEX_V512,
1901 EVEX_CD8<32, CD8VF>;
1902 let ExeDomain = SSEPackedDouble in
1903 defm VPERMILPDZ : avx512_pshuf_imm<0x05, "vpermilpd", VR512, X86VPermilp,
1904 memopv8f64, i512mem, v8f64>, OpSize, TA, EVEX_V512,
1905 VEX_W, EVEX_CD8<32, CD8VF>;
1907 def : Pat<(v16i32 (X86VPermilp VR512:$src1, (i8 imm:$imm))),
1908 (VPERMILPSZri VR512:$src1, imm:$imm)>;
1909 def : Pat<(v8i64 (X86VPermilp VR512:$src1, (i8 imm:$imm))),
1910 (VPERMILPDZri VR512:$src1, imm:$imm)>;
1912 //===----------------------------------------------------------------------===//
1913 // AVX-512 Logical Instructions
1914 //===----------------------------------------------------------------------===//
1916 defm VPANDDZ : avx512_binop_rm<0xDB, "vpandd", and, v16i32, VR512, memopv16i32,
1917 i512mem, loadi32, i32mem, "{1to16}", SSE_BIT_ITINS_P, 1>,
1918 EVEX_V512, EVEX_CD8<32, CD8VF>;
1919 defm VPANDQZ : avx512_binop_rm<0xDB, "vpandq", and, v8i64, VR512, memopv8i64,
1920 i512mem, loadi64, i64mem, "{1to8}", SSE_BIT_ITINS_P, 1>,
1921 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
1922 defm VPORDZ : avx512_binop_rm<0xEB, "vpord", or, v16i32, VR512, memopv16i32,
1923 i512mem, loadi32, i32mem, "{1to16}", SSE_BIT_ITINS_P, 1>,
1924 EVEX_V512, EVEX_CD8<32, CD8VF>;
1925 defm VPORQZ : avx512_binop_rm<0xEB, "vporq", or, v8i64, VR512, memopv8i64,
1926 i512mem, loadi64, i64mem, "{1to8}", SSE_BIT_ITINS_P, 1>,
1927 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
1928 defm VPXORDZ : avx512_binop_rm<0xEF, "vpxord", xor, v16i32, VR512, memopv16i32,
1929 i512mem, loadi32, i32mem, "{1to16}", SSE_BIT_ITINS_P, 1>,
1930 EVEX_V512, EVEX_CD8<32, CD8VF>;
1931 defm VPXORQZ : avx512_binop_rm<0xEF, "vpxorq", xor, v8i64, VR512, memopv8i64,
1932 i512mem, loadi64, i64mem, "{1to8}", SSE_BIT_ITINS_P, 1>,
1933 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
1934 defm VPANDNDZ : avx512_binop_rm<0xDF, "vpandnd", X86andnp, v16i32, VR512,
1935 memopv16i32, i512mem, loadi32, i32mem, "{1to16}",
1936 SSE_BIT_ITINS_P, 0>, EVEX_V512, EVEX_CD8<32, CD8VF>;
1937 defm VPANDNQZ : avx512_binop_rm<0xDF, "vpandnq", X86andnp, v8i64, VR512, memopv8i64,
1938 i512mem, loadi64, i64mem, "{1to8}", SSE_BIT_ITINS_P, 0>,
1939 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
1941 //===----------------------------------------------------------------------===//
1942 // AVX-512 FP arithmetic
1943 //===----------------------------------------------------------------------===//
1945 multiclass avx512_binop_s<bits<8> opc, string OpcodeStr, SDNode OpNode,
1947 defm SSZ : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "ss"), OpNode, FR32X,
1948 f32mem, itins.s, 0>, XS, EVEX_4V, VEX_LIG,
1949 EVEX_CD8<32, CD8VT1>;
1950 defm SDZ : sse12_fp_scalar<opc, !strconcat(OpcodeStr, "sd"), OpNode, FR64X,
1951 f64mem, itins.d, 0>, XD, VEX_W, EVEX_4V, VEX_LIG,
1952 EVEX_CD8<64, CD8VT1>;
1955 let isCommutable = 1 in {
1956 defm VADD : avx512_binop_s<0x58, "add", fadd, SSE_ALU_ITINS_S>;
1957 defm VMUL : avx512_binop_s<0x59, "mul", fmul, SSE_ALU_ITINS_S>;
1958 defm VMIN : avx512_binop_s<0x5D, "min", X86fmin, SSE_ALU_ITINS_S>;
1959 defm VMAX : avx512_binop_s<0x5F, "max", X86fmax, SSE_ALU_ITINS_S>;
1961 let isCommutable = 0 in {
1962 defm VSUB : avx512_binop_s<0x5C, "sub", fsub, SSE_ALU_ITINS_S>;
1963 defm VDIV : avx512_binop_s<0x5E, "div", fdiv, SSE_ALU_ITINS_S>;
1966 multiclass avx512_fp_packed<bits<8> opc, string OpcodeStr, SDNode OpNode,
1967 RegisterClass RC, ValueType vt,
1968 X86MemOperand x86memop, PatFrag mem_frag,
1969 X86MemOperand x86scalar_mop, PatFrag scalar_mfrag,
1971 Domain d, OpndItins itins, bit commutable> {
1972 let isCommutable = commutable in
1973 def rr : PI<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src1, RC:$src2),
1974 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1975 [(set RC:$dst, (vt (OpNode RC:$src1, RC:$src2)))], itins.rr, d>,
1977 let mayLoad = 1 in {
1978 def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
1979 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
1980 [(set RC:$dst, (OpNode RC:$src1, (mem_frag addr:$src2)))],
1981 itins.rm, d>, EVEX_4V, TB;
1982 def rmb : PI<opc, MRMSrcMem, (outs RC:$dst),
1983 (ins RC:$src1, x86scalar_mop:$src2),
1984 !strconcat(OpcodeStr, "\t{${src2}", BrdcstStr,
1985 ", $src1, $dst|$dst, $src1, ${src2}", BrdcstStr, "}"),
1986 [(set RC:$dst, (OpNode RC:$src1,
1987 (vt (X86VBroadcast (scalar_mfrag addr:$src2)))))],
1988 itins.rm, d>, EVEX_4V, EVEX_B, TB;
1992 defm VADDPSZ : avx512_fp_packed<0x58, "addps", fadd, VR512, v16f32, f512mem,
1993 memopv16f32, f32mem, loadf32, "{1to16}", SSEPackedSingle,
1994 SSE_ALU_ITINS_P.s, 1>, EVEX_V512, EVEX_CD8<32, CD8VF>;
1996 defm VADDPDZ : avx512_fp_packed<0x58, "addpd", fadd, VR512, v8f64, f512mem,
1997 memopv8f64, f64mem, loadf64, "{1to8}", SSEPackedDouble,
1998 SSE_ALU_ITINS_P.d, 1>,
1999 EVEX_V512, OpSize, VEX_W, EVEX_CD8<64, CD8VF>;
2001 defm VMULPSZ : avx512_fp_packed<0x59, "mulps", fmul, VR512, v16f32, f512mem,
2002 memopv16f32, f32mem, loadf32, "{1to16}", SSEPackedSingle,
2003 SSE_ALU_ITINS_P.s, 1>, EVEX_V512, EVEX_CD8<32, CD8VF>;
2004 defm VMULPDZ : avx512_fp_packed<0x59, "mulpd", fmul, VR512, v8f64, f512mem,
2005 memopv8f64, f64mem, loadf64, "{1to8}", SSEPackedDouble,
2006 SSE_ALU_ITINS_P.d, 1>,
2007 EVEX_V512, OpSize, VEX_W, EVEX_CD8<64, CD8VF>;
2009 defm VMINPSZ : avx512_fp_packed<0x5D, "minps", X86fmin, VR512, v16f32, f512mem,
2010 memopv16f32, f32mem, loadf32, "{1to16}", SSEPackedSingle,
2011 SSE_ALU_ITINS_P.s, 1>,
2012 EVEX_V512, EVEX_CD8<32, CD8VF>;
2013 defm VMAXPSZ : avx512_fp_packed<0x5F, "maxps", X86fmax, VR512, v16f32, f512mem,
2014 memopv16f32, f32mem, loadf32, "{1to16}", SSEPackedSingle,
2015 SSE_ALU_ITINS_P.s, 1>,
2016 EVEX_V512, EVEX_CD8<32, CD8VF>;
2018 defm VMINPDZ : avx512_fp_packed<0x5D, "minpd", X86fmin, VR512, v8f64, f512mem,
2019 memopv8f64, f64mem, loadf64, "{1to8}", SSEPackedDouble,
2020 SSE_ALU_ITINS_P.d, 1>,
2021 EVEX_V512, OpSize, VEX_W, EVEX_CD8<64, CD8VF>;
2022 defm VMAXPDZ : avx512_fp_packed<0x5F, "maxpd", X86fmax, VR512, v8f64, f512mem,
2023 memopv8f64, f64mem, loadf64, "{1to8}", SSEPackedDouble,
2024 SSE_ALU_ITINS_P.d, 1>,
2025 EVEX_V512, OpSize, VEX_W, EVEX_CD8<64, CD8VF>;
2027 defm VSUBPSZ : avx512_fp_packed<0x5C, "subps", fsub, VR512, v16f32, f512mem,
2028 memopv16f32, f32mem, loadf32, "{1to16}", SSEPackedSingle,
2029 SSE_ALU_ITINS_P.s, 0>, EVEX_V512, EVEX_CD8<32, CD8VF>;
2030 defm VDIVPSZ : avx512_fp_packed<0x5E, "divps", fdiv, VR512, v16f32, f512mem,
2031 memopv16f32, f32mem, loadf32, "{1to16}", SSEPackedSingle,
2032 SSE_ALU_ITINS_P.s, 0>, EVEX_V512, EVEX_CD8<32, CD8VF>;
2034 defm VSUBPDZ : avx512_fp_packed<0x5C, "subpd", fsub, VR512, v8f64, f512mem,
2035 memopv8f64, f64mem, loadf64, "{1to8}", SSEPackedDouble,
2036 SSE_ALU_ITINS_P.d, 0>,
2037 EVEX_V512, OpSize, VEX_W, EVEX_CD8<64, CD8VF>;
2038 defm VDIVPDZ : avx512_fp_packed<0x5E, "divpd", fdiv, VR512, v8f64, f512mem,
2039 memopv8f64, f64mem, loadf64, "{1to8}", SSEPackedDouble,
2040 SSE_ALU_ITINS_P.d, 0>,
2041 EVEX_V512, OpSize, VEX_W, EVEX_CD8<64, CD8VF>;
2043 def : Pat<(v16f32 (int_x86_avx512_mask_max_ps_512 (v16f32 VR512:$src1),
2044 (v16f32 VR512:$src2), (bc_v16f32 (v16i32 immAllZerosV)),
2045 (i16 -1), FROUND_CURRENT)),
2046 (VMAXPSZrr VR512:$src1, VR512:$src2)>;
2048 def : Pat<(v8f64 (int_x86_avx512_mask_max_pd_512 (v8f64 VR512:$src1),
2049 (v8f64 VR512:$src2), (bc_v8f64 (v16i32 immAllZerosV)),
2050 (i8 -1), FROUND_CURRENT)),
2051 (VMAXPDZrr VR512:$src1, VR512:$src2)>;
2053 def : Pat<(v16f32 (int_x86_avx512_mask_min_ps_512 (v16f32 VR512:$src1),
2054 (v16f32 VR512:$src2), (bc_v16f32 (v16i32 immAllZerosV)),
2055 (i16 -1), FROUND_CURRENT)),
2056 (VMINPSZrr VR512:$src1, VR512:$src2)>;
2058 def : Pat<(v8f64 (int_x86_avx512_mask_min_pd_512 (v8f64 VR512:$src1),
2059 (v8f64 VR512:$src2), (bc_v8f64 (v16i32 immAllZerosV)),
2060 (i8 -1), FROUND_CURRENT)),
2061 (VMINPDZrr VR512:$src1, VR512:$src2)>;
2062 //===----------------------------------------------------------------------===//
2063 // AVX-512 VPTESTM instructions
2064 //===----------------------------------------------------------------------===//
2066 multiclass avx512_vptest<bits<8> opc, string OpcodeStr, RegisterClass KRC,
2067 RegisterClass RC, X86MemOperand x86memop, PatFrag memop_frag,
2068 SDNode OpNode, ValueType vt> {
2069 def rr : AVX5128I<opc, MRMSrcReg,
2070 (outs KRC:$dst), (ins RC:$src1, RC:$src2),
2071 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
2072 [(set KRC:$dst, (OpNode (vt RC:$src1), (vt RC:$src2)))]>, EVEX_4V;
2073 def rm : AVX5128I<opc, MRMSrcMem,
2074 (outs KRC:$dst), (ins RC:$src1, x86memop:$src2),
2075 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
2076 [(set KRC:$dst, (OpNode (vt RC:$src1),
2077 (bitconvert (memop_frag addr:$src2))))]>, EVEX_4V;
2080 defm VPTESTMDZ : avx512_vptest<0x27, "vptestmd", VK16, VR512, f512mem,
2081 memopv16i32, X86testm, v16i32>, EVEX_V512,
2082 EVEX_CD8<32, CD8VF>;
2083 defm VPTESTMQZ : avx512_vptest<0x27, "vptestmq", VK8, VR512, f512mem,
2084 memopv8i64, X86testm, v8i64>, EVEX_V512, VEX_W,
2085 EVEX_CD8<64, CD8VF>;
2087 //===----------------------------------------------------------------------===//
2088 // AVX-512 Shift instructions
2089 //===----------------------------------------------------------------------===//
2090 multiclass avx512_shift_rmi<bits<8> opc, Format ImmFormR, Format ImmFormM,
2091 string OpcodeStr, SDNode OpNode, RegisterClass RC,
2092 ValueType vt, X86MemOperand x86memop, PatFrag mem_frag,
2093 RegisterClass KRC> {
2094 def ri : AVX512BIi8<opc, ImmFormR, (outs RC:$dst),
2095 (ins RC:$src1, i8imm:$src2),
2096 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
2097 [(set RC:$dst, (vt (OpNode RC:$src1, (i8 imm:$src2))))],
2098 SSE_INTSHIFT_ITINS_P.rr>, EVEX_4V;
2099 def rik : AVX512BIi8<opc, ImmFormR, (outs RC:$dst),
2100 (ins KRC:$mask, RC:$src1, i8imm:$src2),
2101 !strconcat(OpcodeStr,
2102 "\t{$src2, $src1, $dst {${mask}}|$dst {${mask}}, $src1, $src2}"),
2103 [], SSE_INTSHIFT_ITINS_P.rr>, EVEX_4V, EVEX_K;
2104 def mi: AVX512BIi8<opc, ImmFormM, (outs RC:$dst),
2105 (ins x86memop:$src1, i8imm:$src2),
2106 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
2107 [(set RC:$dst, (OpNode (mem_frag addr:$src1),
2108 (i8 imm:$src2)))], SSE_INTSHIFT_ITINS_P.rm>, EVEX_4V;
2109 def mik: AVX512BIi8<opc, ImmFormM, (outs RC:$dst),
2110 (ins KRC:$mask, x86memop:$src1, i8imm:$src2),
2111 !strconcat(OpcodeStr,
2112 "\t{$src2, $src1, $dst {${mask}}|$dst {${mask}}, $src1, $src2}"),
2113 [], SSE_INTSHIFT_ITINS_P.rm>, EVEX_4V, EVEX_K;
2116 multiclass avx512_shift_rrm<bits<8> opc, string OpcodeStr, SDNode OpNode,
2117 RegisterClass RC, ValueType vt, ValueType SrcVT,
2118 PatFrag bc_frag, RegisterClass KRC> {
2119 // src2 is always 128-bit
2120 def rr : AVX512BI<opc, MRMSrcReg, (outs RC:$dst),
2121 (ins RC:$src1, VR128X:$src2),
2122 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
2123 [(set RC:$dst, (vt (OpNode RC:$src1, (SrcVT VR128X:$src2))))],
2124 SSE_INTSHIFT_ITINS_P.rr>, EVEX_4V;
2125 def rrk : AVX512BI<opc, MRMSrcReg, (outs RC:$dst),
2126 (ins KRC:$mask, RC:$src1, VR128X:$src2),
2127 !strconcat(OpcodeStr,
2128 "\t{$src2, $src1, $dst {${mask}}|$dst {${mask}}, $src1, $src2}"),
2129 [], SSE_INTSHIFT_ITINS_P.rr>, EVEX_4V, EVEX_K;
2130 def rm : AVX512BI<opc, MRMSrcMem, (outs RC:$dst),
2131 (ins RC:$src1, i128mem:$src2),
2132 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
2133 [(set RC:$dst, (vt (OpNode RC:$src1,
2134 (bc_frag (memopv2i64 addr:$src2)))))],
2135 SSE_INTSHIFT_ITINS_P.rm>, EVEX_4V;
2136 def rmk : AVX512BI<opc, MRMSrcMem, (outs RC:$dst),
2137 (ins KRC:$mask, RC:$src1, i128mem:$src2),
2138 !strconcat(OpcodeStr,
2139 "\t{$src2, $src1, $dst {${mask}}|$dst {${mask}}, $src1, $src2}"),
2140 [], SSE_INTSHIFT_ITINS_P.rm>, EVEX_4V, EVEX_K;
2143 defm VPSRLDZ : avx512_shift_rmi<0x72, MRM2r, MRM2m, "vpsrld", X86vsrli,
2144 VR512, v16i32, i512mem, memopv16i32, VK16WM>,
2145 EVEX_V512, EVEX_CD8<32, CD8VF>;
2146 defm VPSRLDZ : avx512_shift_rrm<0xD2, "vpsrld", X86vsrl,
2147 VR512, v16i32, v4i32, bc_v4i32, VK16WM>, EVEX_V512,
2148 EVEX_CD8<32, CD8VQ>;
2150 defm VPSRLQZ : avx512_shift_rmi<0x73, MRM2r, MRM2m, "vpsrlq", X86vsrli,
2151 VR512, v8i64, i512mem, memopv8i64, VK8WM>, EVEX_V512,
2152 EVEX_CD8<64, CD8VF>, VEX_W;
2153 defm VPSRLQZ : avx512_shift_rrm<0xD3, "vpsrlq", X86vsrl,
2154 VR512, v8i64, v2i64, bc_v2i64, VK8WM>, EVEX_V512,
2155 EVEX_CD8<64, CD8VQ>, VEX_W;
2157 defm VPSLLDZ : avx512_shift_rmi<0x72, MRM6r, MRM6m, "vpslld", X86vshli,
2158 VR512, v16i32, i512mem, memopv16i32, VK16WM>, EVEX_V512,
2159 EVEX_CD8<32, CD8VF>;
2160 defm VPSLLDZ : avx512_shift_rrm<0xF2, "vpslld", X86vshl,
2161 VR512, v16i32, v4i32, bc_v4i32, VK16WM>, EVEX_V512,
2162 EVEX_CD8<32, CD8VQ>;
2164 defm VPSLLQZ : avx512_shift_rmi<0x73, MRM6r, MRM6m, "vpsllq", X86vshli,
2165 VR512, v8i64, i512mem, memopv8i64, VK8WM>, EVEX_V512,
2166 EVEX_CD8<64, CD8VF>, VEX_W;
2167 defm VPSLLQZ : avx512_shift_rrm<0xF3, "vpsllq", X86vshl,
2168 VR512, v8i64, v2i64, bc_v2i64, VK8WM>, EVEX_V512,
2169 EVEX_CD8<64, CD8VQ>, VEX_W;
2171 defm VPSRADZ : avx512_shift_rmi<0x72, MRM4r, MRM4m, "vpsrad", X86vsrai,
2172 VR512, v16i32, i512mem, memopv16i32, VK16WM>,
2173 EVEX_V512, EVEX_CD8<32, CD8VF>;
2174 defm VPSRADZ : avx512_shift_rrm<0xE2, "vpsrad", X86vsra,
2175 VR512, v16i32, v4i32, bc_v4i32, VK16WM>, EVEX_V512,
2176 EVEX_CD8<32, CD8VQ>;
2178 defm VPSRAQZ : avx512_shift_rmi<0x72, MRM4r, MRM4m, "vpsraq", X86vsrai,
2179 VR512, v8i64, i512mem, memopv8i64, VK8WM>, EVEX_V512,
2180 EVEX_CD8<64, CD8VF>, VEX_W;
2181 defm VPSRAQZ : avx512_shift_rrm<0xE2, "vpsraq", X86vsra,
2182 VR512, v8i64, v2i64, bc_v2i64, VK8WM>, EVEX_V512,
2183 EVEX_CD8<64, CD8VQ>, VEX_W;
2185 //===-------------------------------------------------------------------===//
2186 // Variable Bit Shifts
2187 //===-------------------------------------------------------------------===//
2188 multiclass avx512_var_shift<bits<8> opc, string OpcodeStr, SDNode OpNode,
2189 RegisterClass RC, ValueType vt,
2190 X86MemOperand x86memop, PatFrag mem_frag> {
2191 def rr : AVX5128I<opc, MRMSrcReg, (outs RC:$dst),
2192 (ins RC:$src1, RC:$src2),
2193 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
2195 (vt (OpNode RC:$src1, (vt RC:$src2))))]>,
2197 def rm : AVX5128I<opc, MRMSrcMem, (outs RC:$dst),
2198 (ins RC:$src1, x86memop:$src2),
2199 !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
2201 (vt (OpNode RC:$src1, (mem_frag addr:$src2))))]>,
2205 defm VPSLLVDZ : avx512_var_shift<0x47, "vpsllvd", shl, VR512, v16i32,
2206 i512mem, memopv16i32>, EVEX_V512,
2207 EVEX_CD8<32, CD8VF>;
2208 defm VPSLLVQZ : avx512_var_shift<0x47, "vpsllvq", shl, VR512, v8i64,
2209 i512mem, memopv8i64>, EVEX_V512, VEX_W,
2210 EVEX_CD8<64, CD8VF>;
2211 defm VPSRLVDZ : avx512_var_shift<0x45, "vpsrlvd", srl, VR512, v16i32,
2212 i512mem, memopv16i32>, EVEX_V512,
2213 EVEX_CD8<32, CD8VF>;
2214 defm VPSRLVQZ : avx512_var_shift<0x45, "vpsrlvq", srl, VR512, v8i64,
2215 i512mem, memopv8i64>, EVEX_V512, VEX_W,
2216 EVEX_CD8<64, CD8VF>;
2217 defm VPSRAVDZ : avx512_var_shift<0x46, "vpsravd", sra, VR512, v16i32,
2218 i512mem, memopv16i32>, EVEX_V512,
2219 EVEX_CD8<32, CD8VF>;
2220 defm VPSRAVQZ : avx512_var_shift<0x46, "vpsravq", sra, VR512, v8i64,
2221 i512mem, memopv8i64>, EVEX_V512, VEX_W,
2222 EVEX_CD8<64, CD8VF>;
2224 //===----------------------------------------------------------------------===//
2225 // AVX-512 - MOVDDUP
2226 //===----------------------------------------------------------------------===//
2228 multiclass avx512_movddup<string OpcodeStr, RegisterClass RC, ValueType VT,
2229 X86MemOperand x86memop, PatFrag memop_frag> {
2230 def rr : AVX512PDI<0x12, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
2231 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2232 [(set RC:$dst, (VT (X86Movddup RC:$src)))]>, EVEX;
2233 def rm : AVX512PDI<0x12, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
2234 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2236 (VT (X86Movddup (memop_frag addr:$src))))]>, EVEX;
2239 defm VMOVDDUPZ : avx512_movddup<"vmovddup", VR512, v8f64, f512mem, memopv8f64>,
2240 VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>;
2241 def : Pat<(X86Movddup (v8f64 (scalar_to_vector (loadf64 addr:$src)))),
2242 (VMOVDDUPZrm addr:$src)>;
2244 //===---------------------------------------------------------------------===//
2245 // Replicate Single FP - MOVSHDUP and MOVSLDUP
2246 //===---------------------------------------------------------------------===//
2247 multiclass avx512_replicate_sfp<bits<8> op, SDNode OpNode, string OpcodeStr,
2248 ValueType vt, RegisterClass RC, PatFrag mem_frag,
2249 X86MemOperand x86memop> {
2250 def rr : AVX512XSI<op, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
2251 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2252 [(set RC:$dst, (vt (OpNode RC:$src)))]>, EVEX;
2254 def rm : AVX512XSI<op, MRMSrcMem, (outs RC:$dst), (ins x86memop:$src),
2255 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2256 [(set RC:$dst, (OpNode (mem_frag addr:$src)))]>, EVEX;
2259 defm VMOVSHDUPZ : avx512_replicate_sfp<0x16, X86Movshdup, "vmovshdup",
2260 v16f32, VR512, memopv16f32, f512mem>, EVEX_V512,
2261 EVEX_CD8<32, CD8VF>;
2262 defm VMOVSLDUPZ : avx512_replicate_sfp<0x12, X86Movsldup, "vmovsldup",
2263 v16f32, VR512, memopv16f32, f512mem>, EVEX_V512,
2264 EVEX_CD8<32, CD8VF>;
2266 def : Pat<(v16i32 (X86Movshdup VR512:$src)), (VMOVSHDUPZrr VR512:$src)>;
2267 def : Pat<(v16i32 (X86Movshdup (memopv16i32 addr:$src))),
2268 (VMOVSHDUPZrm addr:$src)>;
2269 def : Pat<(v16i32 (X86Movsldup VR512:$src)), (VMOVSLDUPZrr VR512:$src)>;
2270 def : Pat<(v16i32 (X86Movsldup (memopv16i32 addr:$src))),
2271 (VMOVSLDUPZrm addr:$src)>;
2273 //===----------------------------------------------------------------------===//
2274 // Move Low to High and High to Low packed FP Instructions
2275 //===----------------------------------------------------------------------===//
2276 def VMOVLHPSZrr : AVX512PSI<0x16, MRMSrcReg, (outs VR128X:$dst),
2277 (ins VR128X:$src1, VR128X:$src2),
2278 "vmovlhps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2279 [(set VR128X:$dst, (v4f32 (X86Movlhps VR128X:$src1, VR128X:$src2)))],
2280 IIC_SSE_MOV_LH>, EVEX_4V;
2281 def VMOVHLPSZrr : AVX512PSI<0x12, MRMSrcReg, (outs VR128X:$dst),
2282 (ins VR128X:$src1, VR128X:$src2),
2283 "vmovhlps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2284 [(set VR128X:$dst, (v4f32 (X86Movhlps VR128X:$src1, VR128X:$src2)))],
2285 IIC_SSE_MOV_LH>, EVEX_4V;
2287 let Predicates = [HasAVX512] in {
2289 def : Pat<(v4i32 (X86Movlhps VR128X:$src1, VR128X:$src2)),
2290 (VMOVLHPSZrr VR128X:$src1, VR128X:$src2)>;
2291 def : Pat<(v2i64 (X86Movlhps VR128X:$src1, VR128X:$src2)),
2292 (VMOVLHPSZrr (v2i64 VR128X:$src1), VR128X:$src2)>;
2295 def : Pat<(v4i32 (X86Movhlps VR128X:$src1, VR128X:$src2)),
2296 (VMOVHLPSZrr VR128X:$src1, VR128X:$src2)>;
2299 //===----------------------------------------------------------------------===//
2300 // FMA - Fused Multiply Operations
2302 let Constraints = "$src1 = $dst" in {
2303 multiclass avx512_fma3p_rm<bits<8> opc, string OpcodeStr,
2304 RegisterClass RC, X86MemOperand x86memop,
2305 PatFrag mem_frag, X86MemOperand x86scalar_mop, PatFrag scalar_mfrag,
2306 string BrdcstStr, SDNode OpNode, ValueType OpVT> {
2307 def r: AVX512FMA3<opc, MRMSrcReg, (outs RC:$dst),
2308 (ins RC:$src1, RC:$src2, RC:$src3),
2309 !strconcat(OpcodeStr,"\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
2310 [(set RC:$dst, (OpVT(OpNode RC:$src1, RC:$src2, RC:$src3)))]>;
2313 def m: AVX512FMA3<opc, MRMSrcMem, (outs RC:$dst),
2314 (ins RC:$src1, RC:$src2, x86memop:$src3),
2315 !strconcat(OpcodeStr, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
2316 [(set RC:$dst, (OpVT (OpNode RC:$src1, RC:$src2,
2317 (mem_frag addr:$src3))))]>;
2318 def mb: AVX512FMA3<opc, MRMSrcMem, (outs RC:$dst),
2319 (ins RC:$src1, RC:$src2, x86scalar_mop:$src3),
2320 !strconcat(OpcodeStr, "\t{${src3}", BrdcstStr,
2321 ", $src2, $dst|$dst, $src2, ${src3}", BrdcstStr, "}"),
2322 [(set RC:$dst, (OpNode RC:$src1, RC:$src2,
2323 (OpVT (X86VBroadcast (scalar_mfrag addr:$src3)))))]>, EVEX_B;
2325 } // Constraints = "$src1 = $dst"
2327 let ExeDomain = SSEPackedSingle in {
2328 defm VFMADD213PSZ : avx512_fma3p_rm<0xA8, "vfmadd213ps", VR512, f512mem,
2329 memopv16f32, f32mem, loadf32, "{1to16}",
2330 X86Fmadd, v16f32>, EVEX_V512,
2331 EVEX_CD8<32, CD8VF>;
2332 defm VFMSUB213PSZ : avx512_fma3p_rm<0xAA, "vfmsub213ps", VR512, f512mem,
2333 memopv16f32, f32mem, loadf32, "{1to16}",
2334 X86Fmsub, v16f32>, EVEX_V512,
2335 EVEX_CD8<32, CD8VF>;
2336 defm VFMADDSUB213PSZ : avx512_fma3p_rm<0xA6, "vfmaddsub213ps", VR512, f512mem,
2337 memopv16f32, f32mem, loadf32, "{1to16}",
2338 X86Fmaddsub, v16f32>,
2339 EVEX_V512, EVEX_CD8<32, CD8VF>;
2340 defm VFMSUBADD213PSZ : avx512_fma3p_rm<0xA7, "vfmsubadd213ps", VR512, f512mem,
2341 memopv16f32, f32mem, loadf32, "{1to16}",
2342 X86Fmsubadd, v16f32>,
2343 EVEX_V512, EVEX_CD8<32, CD8VF>;
2344 defm VFNMADD213PSZ : avx512_fma3p_rm<0xAC, "vfnmadd213ps", VR512, f512mem,
2345 memopv16f32, f32mem, loadf32, "{1to16}",
2346 X86Fnmadd, v16f32>, EVEX_V512,
2347 EVEX_CD8<32, CD8VF>;
2348 defm VFNMSUB213PSZ : avx512_fma3p_rm<0xAE, "vfnmsub213ps", VR512, f512mem,
2349 memopv16f32, f32mem, loadf32, "{1to16}",
2350 X86Fnmsub, v16f32>, EVEX_V512,
2351 EVEX_CD8<32, CD8VF>;
2353 let ExeDomain = SSEPackedDouble in {
2354 defm VFMADD213PDZ : avx512_fma3p_rm<0xA8, "vfmadd213pd", VR512, f512mem,
2355 memopv8f64, f64mem, loadf64, "{1to8}",
2356 X86Fmadd, v8f64>, EVEX_V512,
2357 VEX_W, EVEX_CD8<64, CD8VF>;
2358 defm VFMSUB213PDZ : avx512_fma3p_rm<0xAA, "vfmsub213pd", VR512, f512mem,
2359 memopv8f64, f64mem, loadf64, "{1to8}",
2360 X86Fmsub, v8f64>, EVEX_V512, VEX_W,
2361 EVEX_CD8<64, CD8VF>;
2362 defm VFMADDSUB213PDZ : avx512_fma3p_rm<0xA6, "vfmaddsub213pd", VR512, f512mem,
2363 memopv8f64, f64mem, loadf64, "{1to8}",
2364 X86Fmaddsub, v8f64>, EVEX_V512, VEX_W,
2365 EVEX_CD8<64, CD8VF>;
2366 defm VFMSUBADD213PDZ : avx512_fma3p_rm<0xA7, "vfmsubadd213pd", VR512, f512mem,
2367 memopv8f64, f64mem, loadf64, "{1to8}",
2368 X86Fmsubadd, v8f64>, EVEX_V512, VEX_W,
2369 EVEX_CD8<64, CD8VF>;
2370 defm VFNMADD213PDZ : avx512_fma3p_rm<0xAC, "vfnmadd213pd", VR512, f512mem,
2371 memopv8f64, f64mem, loadf64, "{1to8}",
2372 X86Fnmadd, v8f64>, EVEX_V512, VEX_W,
2373 EVEX_CD8<64, CD8VF>;
2374 defm VFNMSUB213PDZ : avx512_fma3p_rm<0xAE, "vfnmsub213pd", VR512, f512mem,
2375 memopv8f64, f64mem, loadf64, "{1to8}",
2376 X86Fnmsub, v8f64>, EVEX_V512, VEX_W,
2377 EVEX_CD8<64, CD8VF>;
2380 let Constraints = "$src1 = $dst" in {
2381 multiclass avx512_fma3p_m132<bits<8> opc, string OpcodeStr,
2382 RegisterClass RC, X86MemOperand x86memop,
2383 PatFrag mem_frag, X86MemOperand x86scalar_mop, PatFrag scalar_mfrag,
2384 string BrdcstStr, SDNode OpNode, ValueType OpVT> {
2386 def m: AVX512FMA3<opc, MRMSrcMem, (outs RC:$dst),
2387 (ins RC:$src1, RC:$src3, x86memop:$src2),
2388 !strconcat(OpcodeStr, "\t{$src2, $src3, $dst|$dst, $src3, $src2}"),
2389 [(set RC:$dst, (OpVT (OpNode RC:$src1, (mem_frag addr:$src2), RC:$src3)))]>;
2390 def mb: AVX512FMA3<opc, MRMSrcMem, (outs RC:$dst),
2391 (ins RC:$src1, RC:$src3, x86scalar_mop:$src2),
2392 !strconcat(OpcodeStr, "\t{${src2}", BrdcstStr,
2393 ", $src3, $dst|$dst, $src3, ${src2}", BrdcstStr, "}"),
2394 [(set RC:$dst, (OpNode RC:$src1,
2395 (OpVT (X86VBroadcast (scalar_mfrag addr:$src2))), RC:$src3))]>, EVEX_B;
2397 } // Constraints = "$src1 = $dst"
2400 let ExeDomain = SSEPackedSingle in {
2401 defm VFMADD132PSZ : avx512_fma3p_m132<0x98, "vfmadd132ps", VR512, f512mem,
2402 memopv16f32, f32mem, loadf32, "{1to16}",
2403 X86Fmadd, v16f32>, EVEX_V512,
2404 EVEX_CD8<32, CD8VF>;
2405 defm VFMSUB132PSZ : avx512_fma3p_m132<0x9A, "vfmsub132ps", VR512, f512mem,
2406 memopv16f32, f32mem, loadf32, "{1to16}",
2407 X86Fmsub, v16f32>, EVEX_V512,
2408 EVEX_CD8<32, CD8VF>;
2409 defm VFMADDSUB132PSZ : avx512_fma3p_m132<0x96, "vfmaddsub132ps", VR512, f512mem,
2410 memopv16f32, f32mem, loadf32, "{1to16}",
2411 X86Fmaddsub, v16f32>,
2412 EVEX_V512, EVEX_CD8<32, CD8VF>;
2413 defm VFMSUBADD132PSZ : avx512_fma3p_m132<0x97, "vfmsubadd132ps", VR512, f512mem,
2414 memopv16f32, f32mem, loadf32, "{1to16}",
2415 X86Fmsubadd, v16f32>,
2416 EVEX_V512, EVEX_CD8<32, CD8VF>;
2417 defm VFNMADD132PSZ : avx512_fma3p_m132<0x9C, "vfnmadd132ps", VR512, f512mem,
2418 memopv16f32, f32mem, loadf32, "{1to16}",
2419 X86Fnmadd, v16f32>, EVEX_V512,
2420 EVEX_CD8<32, CD8VF>;
2421 defm VFNMSUB132PSZ : avx512_fma3p_m132<0x9E, "vfnmsub132ps", VR512, f512mem,
2422 memopv16f32, f32mem, loadf32, "{1to16}",
2423 X86Fnmsub, v16f32>, EVEX_V512,
2424 EVEX_CD8<32, CD8VF>;
2426 let ExeDomain = SSEPackedDouble in {
2427 defm VFMADD132PDZ : avx512_fma3p_m132<0x98, "vfmadd132pd", VR512, f512mem,
2428 memopv8f64, f64mem, loadf64, "{1to8}",
2429 X86Fmadd, v8f64>, EVEX_V512,
2430 VEX_W, EVEX_CD8<64, CD8VF>;
2431 defm VFMSUB132PDZ : avx512_fma3p_m132<0x9A, "vfmsub132pd", VR512, f512mem,
2432 memopv8f64, f64mem, loadf64, "{1to8}",
2433 X86Fmsub, v8f64>, EVEX_V512, VEX_W,
2434 EVEX_CD8<64, CD8VF>;
2435 defm VFMADDSUB132PDZ : avx512_fma3p_m132<0x96, "vfmaddsub132pd", VR512, f512mem,
2436 memopv8f64, f64mem, loadf64, "{1to8}",
2437 X86Fmaddsub, v8f64>, EVEX_V512, VEX_W,
2438 EVEX_CD8<64, CD8VF>;
2439 defm VFMSUBADD132PDZ : avx512_fma3p_m132<0x97, "vfmsubadd132pd", VR512, f512mem,
2440 memopv8f64, f64mem, loadf64, "{1to8}",
2441 X86Fmsubadd, v8f64>, EVEX_V512, VEX_W,
2442 EVEX_CD8<64, CD8VF>;
2443 defm VFNMADD132PDZ : avx512_fma3p_m132<0x9C, "vfnmadd132pd", VR512, f512mem,
2444 memopv8f64, f64mem, loadf64, "{1to8}",
2445 X86Fnmadd, v8f64>, EVEX_V512, VEX_W,
2446 EVEX_CD8<64, CD8VF>;
2447 defm VFNMSUB132PDZ : avx512_fma3p_m132<0x9E, "vfnmsub132pd", VR512, f512mem,
2448 memopv8f64, f64mem, loadf64, "{1to8}",
2449 X86Fnmsub, v8f64>, EVEX_V512, VEX_W,
2450 EVEX_CD8<64, CD8VF>;
2454 let Constraints = "$src1 = $dst" in {
2455 multiclass avx512_fma3s_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
2456 RegisterClass RC, ValueType OpVT,
2457 X86MemOperand x86memop, Operand memop,
2459 let isCommutable = 1 in
2460 def r : AVX512FMA3<opc, MRMSrcReg, (outs RC:$dst),
2461 (ins RC:$src1, RC:$src2, RC:$src3),
2462 !strconcat(OpcodeStr,
2463 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
2465 (OpVT (OpNode RC:$src2, RC:$src1, RC:$src3)))]>;
2467 def m : AVX512FMA3<opc, MRMSrcMem, (outs RC:$dst),
2468 (ins RC:$src1, RC:$src2, f128mem:$src3),
2469 !strconcat(OpcodeStr,
2470 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
2472 (OpVT (OpNode RC:$src2, RC:$src1,
2473 (mem_frag addr:$src3))))]>;
2476 } // Constraints = "$src1 = $dst"
2478 defm VFMADDSSZ : avx512_fma3s_rm<0xA9, "vfmadd213ss", X86Fmadd, FR32X,
2479 f32, f32mem, ssmem, loadf32>, EVEX_CD8<32, CD8VT1>;
2480 defm VFMADDSDZ : avx512_fma3s_rm<0xA9, "vfmadd213sd", X86Fmadd, FR64X,
2481 f64, f64mem, sdmem, loadf64>, VEX_W, EVEX_CD8<64, CD8VT1>;
2482 defm VFMSUBSSZ : avx512_fma3s_rm<0xAB, "vfmsub213ss", X86Fmsub, FR32X,
2483 f32, f32mem, ssmem, loadf32>, EVEX_CD8<32, CD8VT1>;
2484 defm VFMSUBSDZ : avx512_fma3s_rm<0xAB, "vfmsub213sd", X86Fmsub, FR64X,
2485 f64, f64mem, sdmem, loadf64>, VEX_W, EVEX_CD8<64, CD8VT1>;
2486 defm VFNMADDSSZ : avx512_fma3s_rm<0xAD, "vfnmadd213ss", X86Fnmadd, FR32X,
2487 f32, f32mem, ssmem, loadf32>, EVEX_CD8<32, CD8VT1>;
2488 defm VFNMADDSDZ : avx512_fma3s_rm<0xAD, "vfnmadd213sd", X86Fnmadd, FR64X,
2489 f64, f64mem, sdmem, loadf64>, VEX_W, EVEX_CD8<64, CD8VT1>;
2490 defm VFNMSUBSSZ : avx512_fma3s_rm<0xAF, "vfnmsub213ss", X86Fnmsub, FR32X,
2491 f32, f32mem, ssmem, loadf32>, EVEX_CD8<32, CD8VT1>;
2492 defm VFNMSUBSDZ : avx512_fma3s_rm<0xAF, "vfnmsub213sd", X86Fnmsub, FR64X,
2493 f64, f64mem, sdmem, loadf64>, VEX_W, EVEX_CD8<64, CD8VT1>;
2495 //===----------------------------------------------------------------------===//
2496 // AVX-512 Scalar convert from sign integer to float/double
2497 //===----------------------------------------------------------------------===//
2499 multiclass avx512_vcvtsi<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
2500 X86MemOperand x86memop, string asm> {
2501 let hasSideEffects = 0 in {
2502 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins DstRC:$src1, SrcRC:$src),
2503 !strconcat(asm,"\t{$src, $src1, $dst|$dst, $src1, $src}"), []>,
2506 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst),
2507 (ins DstRC:$src1, x86memop:$src),
2508 !strconcat(asm,"\t{$src, $src1, $dst|$dst, $src1, $src}"), []>,
2510 } // hasSideEffects = 0
2512 let Predicates = [HasAVX512] in {
2513 defm VCVTSI2SSZ : avx512_vcvtsi<0x2A, GR32, FR32X, i32mem, "cvtsi2ss{l}">,
2514 XS, VEX_LIG, EVEX_CD8<32, CD8VT1>;
2515 defm VCVTSI642SSZ : avx512_vcvtsi<0x2A, GR64, FR32X, i64mem, "cvtsi2ss{q}">,
2516 XS, VEX_W, VEX_LIG, EVEX_CD8<64, CD8VT1>;
2517 defm VCVTSI2SDZ : avx512_vcvtsi<0x2A, GR32, FR64X, i32mem, "cvtsi2sd{l}">,
2518 XD, VEX_LIG, EVEX_CD8<32, CD8VT1>;
2519 defm VCVTSI642SDZ : avx512_vcvtsi<0x2A, GR64, FR64X, i64mem, "cvtsi2sd{q}">,
2520 XD, VEX_W, VEX_LIG, EVEX_CD8<64, CD8VT1>;
2522 def : Pat<(f32 (sint_to_fp (loadi32 addr:$src))),
2523 (VCVTSI2SSZrm (f32 (IMPLICIT_DEF)), addr:$src)>;
2524 def : Pat<(f32 (sint_to_fp (loadi64 addr:$src))),
2525 (VCVTSI642SSZrm (f32 (IMPLICIT_DEF)), addr:$src)>;
2526 def : Pat<(f64 (sint_to_fp (loadi32 addr:$src))),
2527 (VCVTSI2SDZrm (f64 (IMPLICIT_DEF)), addr:$src)>;
2528 def : Pat<(f64 (sint_to_fp (loadi64 addr:$src))),
2529 (VCVTSI642SDZrm (f64 (IMPLICIT_DEF)), addr:$src)>;
2531 def : Pat<(f32 (sint_to_fp GR32:$src)),
2532 (VCVTSI2SSZrr (f32 (IMPLICIT_DEF)), GR32:$src)>;
2533 def : Pat<(f32 (sint_to_fp GR64:$src)),
2534 (VCVTSI642SSZrr (f32 (IMPLICIT_DEF)), GR64:$src)>;
2535 def : Pat<(f64 (sint_to_fp GR32:$src)),
2536 (VCVTSI2SDZrr (f64 (IMPLICIT_DEF)), GR32:$src)>;
2537 def : Pat<(f64 (sint_to_fp GR64:$src)),
2538 (VCVTSI642SDZrr (f64 (IMPLICIT_DEF)), GR64:$src)>;
2540 defm VCVTUSI2SSZ : avx512_vcvtsi<0x7B, GR32, FR32X, i32mem, "cvtusi2ss{l}">,
2541 XS, VEX_LIG, EVEX_CD8<32, CD8VT1>;
2542 defm VCVTUSI642SSZ : avx512_vcvtsi<0x7B, GR64, FR32X, i64mem, "cvtusi2ss{q}">,
2543 XS, VEX_W, VEX_LIG, EVEX_CD8<64, CD8VT1>;
2544 defm VCVTUSI2SDZ : avx512_vcvtsi<0x7B, GR32, FR64X, i32mem, "cvtusi2sd{l}">,
2545 XD, VEX_LIG, EVEX_CD8<32, CD8VT1>;
2546 defm VCVTUSI642SDZ : avx512_vcvtsi<0x7B, GR64, FR64X, i64mem, "cvtusi2sd{q}">,
2547 XD, VEX_W, VEX_LIG, EVEX_CD8<64, CD8VT1>;
2549 def : Pat<(f32 (uint_to_fp (loadi32 addr:$src))),
2550 (VCVTUSI2SSZrm (f32 (IMPLICIT_DEF)), addr:$src)>;
2551 def : Pat<(f32 (uint_to_fp (loadi64 addr:$src))),
2552 (VCVTUSI642SSZrm (f32 (IMPLICIT_DEF)), addr:$src)>;
2553 def : Pat<(f64 (uint_to_fp (loadi32 addr:$src))),
2554 (VCVTUSI2SDZrm (f64 (IMPLICIT_DEF)), addr:$src)>;
2555 def : Pat<(f64 (uint_to_fp (loadi64 addr:$src))),
2556 (VCVTUSI642SDZrm (f64 (IMPLICIT_DEF)), addr:$src)>;
2558 def : Pat<(f32 (uint_to_fp GR32:$src)),
2559 (VCVTUSI2SSZrr (f32 (IMPLICIT_DEF)), GR32:$src)>;
2560 def : Pat<(f32 (uint_to_fp GR64:$src)),
2561 (VCVTUSI642SSZrr (f32 (IMPLICIT_DEF)), GR64:$src)>;
2562 def : Pat<(f64 (uint_to_fp GR32:$src)),
2563 (VCVTUSI2SDZrr (f64 (IMPLICIT_DEF)), GR32:$src)>;
2564 def : Pat<(f64 (uint_to_fp GR64:$src)),
2565 (VCVTUSI642SDZrr (f64 (IMPLICIT_DEF)), GR64:$src)>;
2568 //===----------------------------------------------------------------------===//
2569 // AVX-512 Scalar convert from float/double to integer
2570 //===----------------------------------------------------------------------===//
2571 multiclass avx512_cvt_s_int<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
2572 Intrinsic Int, Operand memop, ComplexPattern mem_cpat,
2574 let hasSideEffects = 0 in {
2575 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src),
2576 !strconcat(asm,"\t{$src, $dst|$dst, $src}"),
2577 [(set DstRC:$dst, (Int SrcRC:$src))]>, EVEX, VEX_LIG,
2578 Requires<[HasAVX512]>;
2580 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins memop:$src),
2581 !strconcat(asm,"\t{$src, $dst|$dst, $src}"), []>, EVEX, VEX_LIG,
2582 Requires<[HasAVX512]>;
2583 } // hasSideEffects = 0
2585 let Predicates = [HasAVX512] in {
2586 // Convert float/double to signed/unsigned int 32/64
2587 defm VCVTSS2SIZ: avx512_cvt_s_int<0x2D, VR128X, GR32, int_x86_sse_cvtss2si,
2588 ssmem, sse_load_f32, "cvtss2si">,
2589 XS, EVEX_CD8<32, CD8VT1>;
2590 defm VCVTSS2SI64Z: avx512_cvt_s_int<0x2D, VR128X, GR64, int_x86_sse_cvtss2si64,
2591 ssmem, sse_load_f32, "cvtss2si">,
2592 XS, VEX_W, EVEX_CD8<32, CD8VT1>;
2593 defm VCVTSS2USIZ: avx512_cvt_s_int<0x79, VR128X, GR32, int_x86_avx512_cvtss2usi,
2594 ssmem, sse_load_f32, "cvtss2usi">,
2595 XS, EVEX_CD8<32, CD8VT1>;
2596 defm VCVTSS2USI64Z: avx512_cvt_s_int<0x79, VR128X, GR64,
2597 int_x86_avx512_cvtss2usi64, ssmem,
2598 sse_load_f32, "cvtss2usi">, XS, VEX_W,
2599 EVEX_CD8<32, CD8VT1>;
2600 defm VCVTSD2SIZ: avx512_cvt_s_int<0x2D, VR128X, GR32, int_x86_sse2_cvtsd2si,
2601 sdmem, sse_load_f64, "cvtsd2si">,
2602 XD, EVEX_CD8<64, CD8VT1>;
2603 defm VCVTSD2SI64Z: avx512_cvt_s_int<0x2D, VR128X, GR64, int_x86_sse2_cvtsd2si64,
2604 sdmem, sse_load_f64, "cvtsd2si">,
2605 XD, VEX_W, EVEX_CD8<64, CD8VT1>;
2606 defm VCVTSD2USIZ: avx512_cvt_s_int<0x79, VR128X, GR32, int_x86_avx512_cvtsd2usi,
2607 sdmem, sse_load_f64, "cvtsd2usi">,
2608 XD, EVEX_CD8<64, CD8VT1>;
2609 defm VCVTSD2USI64Z: avx512_cvt_s_int<0x79, VR128X, GR64,
2610 int_x86_avx512_cvtsd2usi64, sdmem,
2611 sse_load_f64, "cvtsd2usi">, XD, VEX_W,
2612 EVEX_CD8<64, CD8VT1>;
2614 let isCodeGenOnly = 1 in {
2615 defm Int_VCVTSI2SSZ : sse12_cvt_sint_3addr<0x2A, GR32, VR128X,
2616 int_x86_sse_cvtsi2ss, i32mem, loadi32, "cvtsi2ss{l}",
2617 SSE_CVT_Scalar, 0>, XS, EVEX_4V;
2618 defm Int_VCVTSI2SS64Z : sse12_cvt_sint_3addr<0x2A, GR64, VR128X,
2619 int_x86_sse_cvtsi642ss, i64mem, loadi64, "cvtsi2ss{q}",
2620 SSE_CVT_Scalar, 0>, XS, EVEX_4V, VEX_W;
2621 defm Int_VCVTSI2SDZ : sse12_cvt_sint_3addr<0x2A, GR32, VR128X,
2622 int_x86_sse2_cvtsi2sd, i32mem, loadi32, "cvtsi2sd{l}",
2623 SSE_CVT_Scalar, 0>, XD, EVEX_4V;
2624 defm Int_VCVTSI2SD64Z : sse12_cvt_sint_3addr<0x2A, GR64, VR128X,
2625 int_x86_sse2_cvtsi642sd, i64mem, loadi64, "cvtsi2sd{q}",
2626 SSE_CVT_Scalar, 0>, XD, EVEX_4V, VEX_W;
2628 defm Int_VCVTUSI2SSZ : sse12_cvt_sint_3addr<0x2A, GR32, VR128X,
2629 int_x86_avx512_cvtusi2ss, i32mem, loadi32, "cvtusi2ss{l}",
2630 SSE_CVT_Scalar, 0>, XS, EVEX_4V;
2631 defm Int_VCVTUSI2SS64Z : sse12_cvt_sint_3addr<0x2A, GR64, VR128X,
2632 int_x86_avx512_cvtusi642ss, i64mem, loadi64, "cvtusi2ss{q}",
2633 SSE_CVT_Scalar, 0>, XS, EVEX_4V, VEX_W;
2634 defm Int_VCVTUSI2SDZ : sse12_cvt_sint_3addr<0x2A, GR32, VR128X,
2635 int_x86_avx512_cvtusi2sd, i32mem, loadi32, "cvtusi2sd{l}",
2636 SSE_CVT_Scalar, 0>, XD, EVEX_4V;
2637 defm Int_VCVTUSI2SD64Z : sse12_cvt_sint_3addr<0x2A, GR64, VR128X,
2638 int_x86_avx512_cvtusi642sd, i64mem, loadi64, "cvtusi2sd{q}",
2639 SSE_CVT_Scalar, 0>, XD, EVEX_4V, VEX_W;
2640 } // isCodeGenOnly = 1
2642 // Convert float/double to signed/unsigned int 32/64 with truncation
2643 let isCodeGenOnly = 1 in {
2644 defm Int_VCVTTSS2SIZ : avx512_cvt_s_int<0x2C, VR128X, GR32, int_x86_sse_cvttss2si,
2645 ssmem, sse_load_f32, "cvttss2si">,
2646 XS, EVEX_CD8<32, CD8VT1>;
2647 defm Int_VCVTTSS2SI64Z : avx512_cvt_s_int<0x2C, VR128X, GR64,
2648 int_x86_sse_cvttss2si64, ssmem, sse_load_f32,
2649 "cvttss2si">, XS, VEX_W,
2650 EVEX_CD8<32, CD8VT1>;
2651 defm Int_VCVTTSD2SIZ : avx512_cvt_s_int<0x2C, VR128X, GR32, int_x86_sse2_cvttsd2si,
2652 sdmem, sse_load_f64, "cvttsd2si">, XD,
2653 EVEX_CD8<64, CD8VT1>;
2654 defm Int_VCVTTSD2SI64Z : avx512_cvt_s_int<0x2C, VR128X, GR64,
2655 int_x86_sse2_cvttsd2si64, sdmem, sse_load_f64,
2656 "cvttsd2si">, XD, VEX_W,
2657 EVEX_CD8<64, CD8VT1>;
2658 defm Int_VCVTTSS2USIZ : avx512_cvt_s_int<0x78, VR128X, GR32,
2659 int_x86_avx512_cvttss2usi, ssmem, sse_load_f32,
2660 "cvttss2usi">, XS, EVEX_CD8<32, CD8VT1>;
2661 defm Int_VCVTTSS2USI64Z : avx512_cvt_s_int<0x78, VR128X, GR64,
2662 int_x86_avx512_cvttss2usi64, ssmem,
2663 sse_load_f32, "cvttss2usi">, XS, VEX_W,
2664 EVEX_CD8<32, CD8VT1>;
2665 defm Int_VCVTTSD2USIZ : avx512_cvt_s_int<0x78, VR128X, GR32,
2666 int_x86_avx512_cvttsd2usi,
2667 sdmem, sse_load_f64, "cvttsd2usi">, XD,
2668 EVEX_CD8<64, CD8VT1>;
2669 defm Int_VCVTTSD2USI64Z : avx512_cvt_s_int<0x78, VR128X, GR64,
2670 int_x86_avx512_cvttsd2usi64, sdmem,
2671 sse_load_f64, "cvttsd2usi">, XD, VEX_W,
2672 EVEX_CD8<64, CD8VT1>;
2673 } // isCodeGenOnly = 1
2675 multiclass avx512_cvt_s<bits<8> opc, RegisterClass SrcRC, RegisterClass DstRC,
2676 SDNode OpNode, X86MemOperand x86memop, PatFrag ld_frag,
2678 def rr : SI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src),
2679 !strconcat(asm,"\t{$src, $dst|$dst, $src}"),
2680 [(set DstRC:$dst, (OpNode SrcRC:$src))]>, EVEX;
2681 def rm : SI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src),
2682 !strconcat(asm,"\t{$src, $dst|$dst, $src}"),
2683 [(set DstRC:$dst, (OpNode (ld_frag addr:$src)))]>, EVEX;
2686 defm VCVTTSS2SIZ : avx512_cvt_s<0x2C, FR32X, GR32, fp_to_sint, f32mem,
2687 loadf32, "cvttss2si">, XS,
2688 EVEX_CD8<32, CD8VT1>;
2689 defm VCVTTSS2USIZ : avx512_cvt_s<0x78, FR32X, GR32, fp_to_uint, f32mem,
2690 loadf32, "cvttss2usi">, XS,
2691 EVEX_CD8<32, CD8VT1>;
2692 defm VCVTTSS2SI64Z : avx512_cvt_s<0x2C, FR32X, GR64, fp_to_sint, f32mem,
2693 loadf32, "cvttss2si">, XS, VEX_W,
2694 EVEX_CD8<32, CD8VT1>;
2695 defm VCVTTSS2USI64Z : avx512_cvt_s<0x78, FR32X, GR64, fp_to_uint, f32mem,
2696 loadf32, "cvttss2usi">, XS, VEX_W,
2697 EVEX_CD8<32, CD8VT1>;
2698 defm VCVTTSD2SIZ : avx512_cvt_s<0x2C, FR64X, GR32, fp_to_sint, f64mem,
2699 loadf64, "cvttsd2si">, XD,
2700 EVEX_CD8<64, CD8VT1>;
2701 defm VCVTTSD2USIZ : avx512_cvt_s<0x78, FR64X, GR32, fp_to_uint, f64mem,
2702 loadf64, "cvttsd2usi">, XD,
2703 EVEX_CD8<64, CD8VT1>;
2704 defm VCVTTSD2SI64Z : avx512_cvt_s<0x2C, FR64X, GR64, fp_to_sint, f64mem,
2705 loadf64, "cvttsd2si">, XD, VEX_W,
2706 EVEX_CD8<64, CD8VT1>;
2707 defm VCVTTSD2USI64Z : avx512_cvt_s<0x78, FR64X, GR64, fp_to_uint, f64mem,
2708 loadf64, "cvttsd2usi">, XD, VEX_W,
2709 EVEX_CD8<64, CD8VT1>;
2711 //===----------------------------------------------------------------------===//
2712 // AVX-512 Convert form float to double and back
2713 //===----------------------------------------------------------------------===//
2714 let hasSideEffects = 0 in {
2715 def VCVTSS2SDZrr : AVX512XSI<0x5A, MRMSrcReg, (outs FR64X:$dst),
2716 (ins FR32X:$src1, FR32X:$src2),
2717 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2718 []>, EVEX_4V, VEX_LIG, Sched<[WriteCvtF2F]>;
2720 def VCVTSS2SDZrm : AVX512XSI<0x5A, MRMSrcMem, (outs FR64X:$dst),
2721 (ins FR32X:$src1, f32mem:$src2),
2722 "vcvtss2sd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2723 []>, EVEX_4V, VEX_LIG, Sched<[WriteCvtF2FLd, ReadAfterLd]>,
2724 EVEX_CD8<32, CD8VT1>;
2726 // Convert scalar double to scalar single
2727 def VCVTSD2SSZrr : AVX512XDI<0x5A, MRMSrcReg, (outs FR32X:$dst),
2728 (ins FR64X:$src1, FR64X:$src2),
2729 "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2730 []>, EVEX_4V, VEX_LIG, VEX_W, Sched<[WriteCvtF2F]>;
2732 def VCVTSD2SSZrm : AVX512XDI<0x5A, MRMSrcMem, (outs FR32X:$dst),
2733 (ins FR64X:$src1, f64mem:$src2),
2734 "vcvtsd2ss\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2735 []>, EVEX_4V, VEX_LIG, VEX_W,
2736 Sched<[WriteCvtF2FLd, ReadAfterLd]>, EVEX_CD8<64, CD8VT1>;
2739 def : Pat<(f64 (fextend FR32X:$src)), (VCVTSS2SDZrr FR32X:$src, FR32X:$src)>,
2740 Requires<[HasAVX512]>;
2741 def : Pat<(fextend (loadf32 addr:$src)),
2742 (VCVTSS2SDZrm (f32 (IMPLICIT_DEF)), addr:$src)>, Requires<[HasAVX512]>;
2744 def : Pat<(extloadf32 addr:$src),
2745 (VCVTSS2SDZrm (f32 (IMPLICIT_DEF)), addr:$src)>,
2746 Requires<[HasAVX512, OptForSize]>;
2748 def : Pat<(extloadf32 addr:$src),
2749 (VCVTSS2SDZrr (f32 (IMPLICIT_DEF)), (VMOVSSZrm addr:$src))>,
2750 Requires<[HasAVX512, OptForSpeed]>;
2752 def : Pat<(f32 (fround FR64X:$src)), (VCVTSD2SSZrr FR64X:$src, FR64X:$src)>,
2753 Requires<[HasAVX512]>;
2755 multiclass avx512_vcvt_fp_with_rc<bits<8> opc, string asm, RegisterClass SrcRC,
2756 RegisterClass DstRC, SDNode OpNode, PatFrag mem_frag,
2757 X86MemOperand x86memop, ValueType OpVT, ValueType InVT,
2759 let hasSideEffects = 0 in {
2760 def rr : AVX512PI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src),
2761 !strconcat(asm,"\t{$src, $dst|$dst, $src}"),
2763 (OpVT (OpNode (InVT SrcRC:$src))))], d>, EVEX;
2764 def rrb : AVX512PI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src, AVX512RC:$rc),
2765 !strconcat(asm,"\t{$rc, $src, $dst|$dst, $src, $rc}"),
2766 [], d>, EVEX, EVEX_B;
2768 def rm : AVX512PI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src),
2769 !strconcat(asm,"\t{$src, $dst|$dst, $src}"),
2771 (OpVT (OpNode (InVT (bitconvert (mem_frag addr:$src))))))], d>, EVEX;
2772 } // hasSideEffects = 0
2775 multiclass avx512_vcvt_fp<bits<8> opc, string asm, RegisterClass SrcRC,
2776 RegisterClass DstRC, SDNode OpNode, PatFrag mem_frag,
2777 X86MemOperand x86memop, ValueType OpVT, ValueType InVT,
2779 let hasSideEffects = 0 in {
2780 def rr : AVX512PI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src),
2781 !strconcat(asm,"\t{$src, $dst|$dst, $src}"),
2783 (OpVT (OpNode (InVT SrcRC:$src))))], d>, EVEX;
2785 def rm : AVX512PI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src),
2786 !strconcat(asm,"\t{$src, $dst|$dst, $src}"),
2788 (OpVT (OpNode (InVT (bitconvert (mem_frag addr:$src))))))], d>, EVEX;
2789 } // hasSideEffects = 0
2792 defm VCVTPD2PSZ : avx512_vcvt_fp_with_rc<0x5A, "vcvtpd2ps", VR512, VR256X, fround,
2793 memopv8f64, f512mem, v8f32, v8f64,
2794 SSEPackedSingle>, EVEX_V512, VEX_W, OpSize,
2795 EVEX_CD8<64, CD8VF>;
2797 defm VCVTPS2PDZ : avx512_vcvt_fp<0x5A, "vcvtps2pd", VR256X, VR512, fextend,
2798 memopv4f64, f256mem, v8f64, v8f32,
2799 SSEPackedDouble>, EVEX_V512, EVEX_CD8<32, CD8VH>;
2800 def : Pat<(v8f64 (extloadv8f32 addr:$src)),
2801 (VCVTPS2PDZrm addr:$src)>;
2803 //===----------------------------------------------------------------------===//
2804 // AVX-512 Vector convert from sign integer to float/double
2805 //===----------------------------------------------------------------------===//
2807 defm VCVTDQ2PSZ : avx512_vcvt_fp_with_rc<0x5B, "vcvtdq2ps", VR512, VR512, sint_to_fp,
2808 memopv8i64, i512mem, v16f32, v16i32,
2809 SSEPackedSingle>, EVEX_V512, EVEX_CD8<32, CD8VF>;
2811 defm VCVTDQ2PDZ : avx512_vcvt_fp<0xE6, "vcvtdq2pd", VR256X, VR512, sint_to_fp,
2812 memopv4i64, i256mem, v8f64, v8i32,
2813 SSEPackedDouble>, EVEX_V512, XS,
2814 EVEX_CD8<32, CD8VH>;
2816 defm VCVTTPS2DQZ : avx512_vcvt_fp<0x5B, "vcvttps2dq", VR512, VR512, fp_to_sint,
2817 memopv16f32, f512mem, v16i32, v16f32,
2818 SSEPackedSingle>, EVEX_V512, XS,
2819 EVEX_CD8<32, CD8VF>;
2821 defm VCVTTPD2DQZ : avx512_vcvt_fp<0xE6, "vcvttpd2dq", VR512, VR256X, fp_to_sint,
2822 memopv8f64, f512mem, v8i32, v8f64,
2823 SSEPackedDouble>, EVEX_V512, OpSize, VEX_W,
2824 EVEX_CD8<64, CD8VF>;
2826 defm VCVTTPS2UDQZ : avx512_vcvt_fp<0x78, "vcvttps2udq", VR512, VR512, fp_to_uint,
2827 memopv16f32, f512mem, v16i32, v16f32,
2828 SSEPackedSingle>, EVEX_V512,
2829 EVEX_CD8<32, CD8VF>;
2831 // cvttps2udq (src, 0, mask-all-ones, sae-current)
2832 def : Pat<(v16i32 (int_x86_avx512_mask_cvttps2udq_512 (v16f32 VR512:$src),
2833 (v16i32 immAllZerosV), (i16 -1), FROUND_CURRENT)),
2834 (VCVTTPS2UDQZrr VR512:$src)>;
2836 defm VCVTTPD2UDQZ : avx512_vcvt_fp<0x78, "vcvttpd2udq", VR512, VR256X, fp_to_uint,
2837 memopv8f64, f512mem, v8i32, v8f64,
2838 SSEPackedDouble>, EVEX_V512, VEX_W,
2839 EVEX_CD8<64, CD8VF>;
2841 // cvttpd2udq (src, 0, mask-all-ones, sae-current)
2842 def : Pat<(v8i32 (int_x86_avx512_mask_cvttpd2udq_512 (v8f64 VR512:$src),
2843 (v8i32 immAllZerosV), (i8 -1), FROUND_CURRENT)),
2844 (VCVTTPD2UDQZrr VR512:$src)>;
2846 defm VCVTUDQ2PDZ : avx512_vcvt_fp<0x7A, "vcvtudq2pd", VR256X, VR512, uint_to_fp,
2847 memopv4i64, f256mem, v8f64, v8i32,
2848 SSEPackedDouble>, EVEX_V512, XS,
2849 EVEX_CD8<32, CD8VH>;
2851 defm VCVTUDQ2PSZ : avx512_vcvt_fp_with_rc<0x7A, "vcvtudq2ps", VR512, VR512, uint_to_fp,
2852 memopv16i32, f512mem, v16f32, v16i32,
2853 SSEPackedSingle>, EVEX_V512, XD,
2854 EVEX_CD8<32, CD8VF>;
2856 def : Pat<(v8i32 (fp_to_uint (v8f32 VR256X:$src1))),
2857 (EXTRACT_SUBREG (v16i32 (VCVTTPS2UDQZrr
2858 (v16f32 (SUBREG_TO_REG (i32 0), VR256X:$src1, sub_ymm)))), sub_ymm)>;
2861 def : Pat<(v16f32 (int_x86_avx512_mask_cvtdq2ps_512 (v16i32 VR512:$src),
2862 (bc_v16f32 (v16i32 immAllZerosV)), (i16 -1), imm:$rc)),
2863 (VCVTDQ2PSZrrb VR512:$src, imm:$rc)>;
2864 def : Pat<(v8f64 (int_x86_avx512_mask_cvtdq2pd_512 (v8i32 VR256X:$src),
2865 (bc_v8f64 (v16i32 immAllZerosV)), (i8 -1))),
2866 (VCVTDQ2PDZrr VR256X:$src)>;
2867 def : Pat<(v16f32 (int_x86_avx512_mask_cvtudq2ps_512 (v16i32 VR512:$src),
2868 (bc_v16f32 (v16i32 immAllZerosV)), (i16 -1), imm:$rc)),
2869 (VCVTUDQ2PSZrrb VR512:$src, imm:$rc)>;
2870 def : Pat<(v8f64 (int_x86_avx512_mask_cvtudq2pd_512 (v8i32 VR256X:$src),
2871 (bc_v8f64 (v16i32 immAllZerosV)), (i8 -1))),
2872 (VCVTUDQ2PDZrr VR256X:$src)>;
2874 multiclass avx512_vcvt_fp2int<bits<8> opc, string asm, RegisterClass SrcRC,
2875 RegisterClass DstRC, PatFrag mem_frag,
2876 X86MemOperand x86memop, Domain d> {
2877 let hasSideEffects = 0 in {
2878 def rr : AVX512PI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src),
2879 !strconcat(asm,"\t{$src, $dst|$dst, $src}"),
2881 def rrb : AVX512PI<opc, MRMSrcReg, (outs DstRC:$dst), (ins SrcRC:$src, AVX512RC:$rc),
2882 !strconcat(asm,"\t{$rc, $src, $dst|$dst, $src, $rc}"),
2883 [], d>, EVEX, EVEX_B;
2885 def rm : AVX512PI<opc, MRMSrcMem, (outs DstRC:$dst), (ins x86memop:$src),
2886 !strconcat(asm,"\t{$src, $dst|$dst, $src}"),
2888 } // hasSideEffects = 0
2891 defm VCVTPS2DQZ : avx512_vcvt_fp2int<0x5B, "vcvtps2dq", VR512, VR512,
2892 memopv16f32, f512mem, SSEPackedSingle>, OpSize,
2893 EVEX_V512, EVEX_CD8<32, CD8VF>;
2894 defm VCVTPD2DQZ : avx512_vcvt_fp2int<0xE6, "vcvtpd2dq", VR512, VR256X,
2895 memopv8f64, f512mem, SSEPackedDouble>, XD, VEX_W,
2896 EVEX_V512, EVEX_CD8<64, CD8VF>;
2898 def : Pat <(v16i32 (int_x86_avx512_mask_cvtps2dq_512 (v16f32 VR512:$src),
2899 (v16i32 immAllZerosV), (i16 -1), imm:$rc)),
2900 (VCVTPS2DQZrrb VR512:$src, imm:$rc)>;
2902 def : Pat <(v8i32 (int_x86_avx512_mask_cvtpd2dq_512 (v8f64 VR512:$src),
2903 (v8i32 immAllZerosV), (i8 -1), imm:$rc)),
2904 (VCVTPD2DQZrrb VR512:$src, imm:$rc)>;
2906 defm VCVTPS2UDQZ : avx512_vcvt_fp2int<0x79, "vcvtps2udq", VR512, VR512,
2907 memopv16f32, f512mem, SSEPackedSingle>,
2908 EVEX_V512, EVEX_CD8<32, CD8VF>;
2909 defm VCVTPD2UDQZ : avx512_vcvt_fp2int<0x79, "vcvtpd2udq", VR512, VR256X,
2910 memopv8f64, f512mem, SSEPackedDouble>, VEX_W,
2911 EVEX_V512, EVEX_CD8<64, CD8VF>;
2913 def : Pat <(v16i32 (int_x86_avx512_mask_cvtps2udq_512 (v16f32 VR512:$src),
2914 (v16i32 immAllZerosV), (i16 -1), imm:$rc)),
2915 (VCVTPS2UDQZrrb VR512:$src, imm:$rc)>;
2917 def : Pat <(v8i32 (int_x86_avx512_mask_cvtpd2udq_512 (v8f64 VR512:$src),
2918 (v8i32 immAllZerosV), (i8 -1), imm:$rc)),
2919 (VCVTPD2UDQZrrb VR512:$src, imm:$rc)>;
2921 let Predicates = [HasAVX512] in {
2922 def : Pat<(v8f32 (fround (loadv8f64 addr:$src))),
2923 (VCVTPD2PSZrm addr:$src)>;
2924 def : Pat<(v8f64 (extloadv8f32 addr:$src)),
2925 (VCVTPS2PDZrm addr:$src)>;
2928 //===----------------------------------------------------------------------===//
2929 // Half precision conversion instructions
2930 //===----------------------------------------------------------------------===//
2931 multiclass avx512_f16c_ph2ps<RegisterClass destRC, RegisterClass srcRC,
2932 X86MemOperand x86memop, Intrinsic Int> {
2933 def rr : AVX5128I<0x13, MRMSrcReg, (outs destRC:$dst), (ins srcRC:$src),
2934 "vcvtph2ps\t{$src, $dst|$dst, $src}",
2935 [(set destRC:$dst, (Int srcRC:$src))]>, EVEX;
2936 let hasSideEffects = 0, mayLoad = 1 in
2937 def rm : AVX5128I<0x13, MRMSrcMem, (outs destRC:$dst), (ins x86memop:$src),
2938 "vcvtph2ps\t{$src, $dst|$dst, $src}", []>, EVEX;
2941 multiclass avx512_f16c_ps2ph<RegisterClass destRC, RegisterClass srcRC,
2942 X86MemOperand x86memop, Intrinsic Int> {
2943 def rr : AVX512AIi8<0x1D, MRMDestReg, (outs destRC:$dst),
2944 (ins srcRC:$src1, i32i8imm:$src2),
2945 "vcvtps2ph\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2946 [(set destRC:$dst, (Int srcRC:$src1, imm:$src2))]>, EVEX;
2947 let hasSideEffects = 0, mayStore = 1 in
2948 def mr : AVX512AIi8<0x1D, MRMDestMem, (outs),
2949 (ins x86memop:$dst, srcRC:$src1, i32i8imm:$src2),
2950 "vcvtps2ph\t{$src2, $src1, $dst|$dst, $src1, $src2}", []>, EVEX;
2953 defm VCVTPH2PSZ : avx512_f16c_ph2ps<VR512, VR256X, f256mem,
2954 int_x86_avx512_vcvtph2ps_512>, EVEX_V512,
2955 EVEX_CD8<32, CD8VH>;
2956 defm VCVTPS2PHZ : avx512_f16c_ps2ph<VR256X, VR512, f256mem,
2957 int_x86_avx512_vcvtps2ph_512>, EVEX_V512,
2958 EVEX_CD8<32, CD8VH>;
2960 let Defs = [EFLAGS], Predicates = [HasAVX512] in {
2961 defm VUCOMISSZ : sse12_ord_cmp<0x2E, FR32X, X86cmp, f32, f32mem, loadf32,
2962 "ucomiss">, TB, EVEX, VEX_LIG,
2963 EVEX_CD8<32, CD8VT1>;
2964 defm VUCOMISDZ : sse12_ord_cmp<0x2E, FR64X, X86cmp, f64, f64mem, loadf64,
2965 "ucomisd">, TB, OpSize, EVEX,
2966 VEX_LIG, VEX_W, EVEX_CD8<64, CD8VT1>;
2967 let Pattern = []<dag> in {
2968 defm VCOMISSZ : sse12_ord_cmp<0x2F, VR128X, undef, v4f32, f128mem, load,
2969 "comiss">, TB, EVEX, VEX_LIG,
2970 EVEX_CD8<32, CD8VT1>;
2971 defm VCOMISDZ : sse12_ord_cmp<0x2F, VR128X, undef, v2f64, f128mem, load,
2972 "comisd">, TB, OpSize, EVEX,
2973 VEX_LIG, VEX_W, EVEX_CD8<64, CD8VT1>;
2975 let isCodeGenOnly = 1 in {
2976 defm Int_VUCOMISSZ : sse12_ord_cmp<0x2E, VR128X, X86ucomi, v4f32, f128mem,
2977 load, "ucomiss">, TB, EVEX, VEX_LIG,
2978 EVEX_CD8<32, CD8VT1>;
2979 defm Int_VUCOMISDZ : sse12_ord_cmp<0x2E, VR128X, X86ucomi, v2f64, f128mem,
2980 load, "ucomisd">, TB, OpSize, EVEX,
2981 VEX_LIG, VEX_W, EVEX_CD8<64, CD8VT1>;
2983 defm Int_VCOMISSZ : sse12_ord_cmp<0x2F, VR128X, X86comi, v4f32, f128mem,
2984 load, "comiss">, TB, EVEX, VEX_LIG,
2985 EVEX_CD8<32, CD8VT1>;
2986 defm Int_VCOMISDZ : sse12_ord_cmp<0x2F, VR128X, X86comi, v2f64, f128mem,
2987 load, "comisd">, TB, OpSize, EVEX,
2988 VEX_LIG, VEX_W, EVEX_CD8<64, CD8VT1>;
2992 /// avx512_unop_p - AVX-512 unops in packed form.
2993 multiclass avx512_fp_unop_p<bits<8> opc, string OpcodeStr, SDNode OpNode> {
2994 def PSZr : AVX5128I<opc, MRMSrcReg, (outs VR512:$dst), (ins VR512:$src),
2995 !strconcat(OpcodeStr,
2996 "ps\t{$src, $dst|$dst, $src}"),
2997 [(set VR512:$dst, (v16f32 (OpNode VR512:$src)))]>,
2999 def PSZm : AVX5128I<opc, MRMSrcMem, (outs VR512:$dst), (ins f256mem:$src),
3000 !strconcat(OpcodeStr,
3001 "ps\t{$src, $dst|$dst, $src}"),
3002 [(set VR512:$dst, (OpNode (memopv16f32 addr:$src)))]>,
3003 EVEX, EVEX_V512, EVEX_CD8<32, CD8VF>;
3004 def PDZr : AVX5128I<opc, MRMSrcReg, (outs VR512:$dst), (ins VR512:$src),
3005 !strconcat(OpcodeStr,
3006 "pd\t{$src, $dst|$dst, $src}"),
3007 [(set VR512:$dst, (v8f64 (OpNode VR512:$src)))]>,
3008 EVEX, EVEX_V512, VEX_W;
3009 def PDZm : AVX5128I<opc, MRMSrcMem, (outs VR512:$dst), (ins f512mem:$src),
3010 !strconcat(OpcodeStr,
3011 "pd\t{$src, $dst|$dst, $src}"),
3012 [(set VR512:$dst, (OpNode (memopv16f32 addr:$src)))]>,
3013 EVEX, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
3016 /// avx512_fp_unop_p_int - AVX-512 intrinsics unops in packed forms.
3017 multiclass avx512_fp_unop_p_int<bits<8> opc, string OpcodeStr,
3018 Intrinsic V16F32Int, Intrinsic V8F64Int> {
3019 let isCodeGenOnly = 1 in {
3020 def PSZr_Int : AVX5128I<opc, MRMSrcReg, (outs VR512:$dst), (ins VR512:$src),
3021 !strconcat(OpcodeStr,
3022 "ps\t{$src, $dst|$dst, $src}"),
3023 [(set VR512:$dst, (V16F32Int VR512:$src))]>,
3025 def PSZm_Int : AVX5128I<opc, MRMSrcMem, (outs VR512:$dst), (ins f512mem:$src),
3026 !strconcat(OpcodeStr,
3027 "ps\t{$src, $dst|$dst, $src}"),
3029 (V16F32Int (memopv16f32 addr:$src)))]>, EVEX,
3030 EVEX_V512, EVEX_CD8<32, CD8VF>;
3031 def PDZr_Int : AVX5128I<opc, MRMSrcReg, (outs VR512:$dst), (ins VR512:$src),
3032 !strconcat(OpcodeStr,
3033 "pd\t{$src, $dst|$dst, $src}"),
3034 [(set VR512:$dst, (V8F64Int VR512:$src))]>,
3035 EVEX, EVEX_V512, VEX_W;
3036 def PDZm_Int : AVX5128I<opc, MRMSrcMem, (outs VR512:$dst), (ins f512mem:$src),
3037 !strconcat(OpcodeStr,
3038 "pd\t{$src, $dst|$dst, $src}"),
3040 (V8F64Int (memopv8f64 addr:$src)))]>,
3041 EVEX, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
3042 } // isCodeGenOnly = 1
3045 /// avx512_fp_unop_s - AVX-512 unops in scalar form.
3046 multiclass avx512_fp_unop_s<bits<8> opc, string OpcodeStr> {
3047 let hasSideEffects = 0 in {
3048 def SSZr : AVX5128I<opc, MRMSrcReg, (outs FR32X:$dst),
3049 (ins FR32X:$src1, FR32X:$src2),
3050 !strconcat(OpcodeStr,
3051 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3053 let mayLoad = 1 in {
3054 def SSZm : AVX5128I<opc, MRMSrcMem, (outs FR32X:$dst),
3055 (ins FR32X:$src1, f32mem:$src2),
3056 !strconcat(OpcodeStr,
3057 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3058 []>, EVEX_4V, EVEX_CD8<32, CD8VT1>;
3059 let isCodeGenOnly = 1 in
3060 def SSZm_Int : AVX5128I<opc, MRMSrcMem, (outs VR128X:$dst),
3061 (ins VR128X:$src1, ssmem:$src2),
3062 !strconcat(OpcodeStr,
3063 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3064 []>, EVEX_4V, EVEX_CD8<32, CD8VT1>;
3066 def SDZr : AVX5128I<opc, MRMSrcReg, (outs FR64X:$dst),
3067 (ins FR64X:$src1, FR64X:$src2),
3068 !strconcat(OpcodeStr,
3069 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), []>,
3071 let mayLoad = 1 in {
3072 def SDZm : AVX5128I<opc, MRMSrcMem, (outs FR64X:$dst),
3073 (ins FR64X:$src1, f64mem:$src2),
3074 !strconcat(OpcodeStr,
3075 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), []>,
3076 EVEX_4V, VEX_W, EVEX_CD8<64, CD8VT1>;
3077 let isCodeGenOnly = 1 in
3078 def SDZm_Int : AVX5128I<opc, MRMSrcMem, (outs VR128X:$dst),
3079 (ins VR128X:$src1, sdmem:$src2),
3080 !strconcat(OpcodeStr,
3081 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3082 []>, EVEX_4V, VEX_W, EVEX_CD8<64, CD8VT1>;
3087 defm VRCP14 : avx512_fp_unop_s<0x4D, "vrcp14">,
3088 avx512_fp_unop_p<0x4C, "vrcp14", X86frcp>,
3089 avx512_fp_unop_p_int<0x4C, "vrcp14",
3090 int_x86_avx512_rcp14_ps_512, int_x86_avx512_rcp14_pd_512>;
3092 defm VRSQRT14 : avx512_fp_unop_s<0x4F, "vrsqrt14">,
3093 avx512_fp_unop_p<0x4E, "vrsqrt14", X86frsqrt>,
3094 avx512_fp_unop_p_int<0x4E, "vrsqrt14",
3095 int_x86_avx512_rsqrt14_ps_512, int_x86_avx512_rsqrt14_pd_512>;
3097 def : Pat<(int_x86_avx512_rsqrt14_ss VR128X:$src),
3098 (COPY_TO_REGCLASS (VRSQRT14SSZr (f32 (IMPLICIT_DEF)),
3099 (COPY_TO_REGCLASS VR128X:$src, FR32)),
3101 def : Pat<(int_x86_avx512_rsqrt14_ss sse_load_f32:$src),
3102 (VRSQRT14SSZm_Int (v4f32 (IMPLICIT_DEF)), sse_load_f32:$src)>;
3104 def : Pat<(int_x86_avx512_rcp14_ss VR128X:$src),
3105 (COPY_TO_REGCLASS (VRCP14SSZr (f32 (IMPLICIT_DEF)),
3106 (COPY_TO_REGCLASS VR128X:$src, FR32)),
3108 def : Pat<(int_x86_avx512_rcp14_ss sse_load_f32:$src),
3109 (VRCP14SSZm_Int (v4f32 (IMPLICIT_DEF)), sse_load_f32:$src)>;
3111 let AddedComplexity = 20, Predicates = [HasERI] in {
3112 defm VRCP28 : avx512_fp_unop_s<0xCB, "vrcp28">,
3113 avx512_fp_unop_p<0xCA, "vrcp28", X86frcp>,
3114 avx512_fp_unop_p_int<0xCA, "vrcp28",
3115 int_x86_avx512_rcp28_ps_512, int_x86_avx512_rcp28_pd_512>;
3117 defm VRSQRT28 : avx512_fp_unop_s<0xCD, "vrsqrt28">,
3118 avx512_fp_unop_p<0xCC, "vrsqrt28", X86frsqrt>,
3119 avx512_fp_unop_p_int<0xCC, "vrsqrt28",
3120 int_x86_avx512_rsqrt28_ps_512, int_x86_avx512_rsqrt28_pd_512>;
3123 let Predicates = [HasERI] in {
3124 def : Pat<(int_x86_avx512_rsqrt28_ss VR128X:$src),
3125 (COPY_TO_REGCLASS (VRSQRT28SSZr (f32 (IMPLICIT_DEF)),
3126 (COPY_TO_REGCLASS VR128X:$src, FR32)),
3128 def : Pat<(int_x86_avx512_rsqrt28_ss sse_load_f32:$src),
3129 (VRSQRT28SSZm_Int (v4f32 (IMPLICIT_DEF)), sse_load_f32:$src)>;
3131 def : Pat<(int_x86_avx512_rcp28_ss VR128X:$src),
3132 (COPY_TO_REGCLASS (VRCP28SSZr (f32 (IMPLICIT_DEF)),
3133 (COPY_TO_REGCLASS VR128X:$src, FR32)),
3135 def : Pat<(int_x86_avx512_rcp28_ss sse_load_f32:$src),
3136 (VRCP28SSZm_Int (v4f32 (IMPLICIT_DEF)), sse_load_f32:$src)>;
3138 multiclass avx512_sqrt_packed<bits<8> opc, string OpcodeStr, SDNode OpNode,
3139 Intrinsic V16F32Int, Intrinsic V8F64Int,
3140 OpndItins itins_s, OpndItins itins_d> {
3141 def PSZrr :AVX512PSI<opc, MRMSrcReg, (outs VR512:$dst), (ins VR512:$src),
3142 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3143 [(set VR512:$dst, (v16f32 (OpNode VR512:$src)))], itins_s.rr>,
3147 def PSZrm : AVX512PSI<opc, MRMSrcMem, (outs VR512:$dst), (ins f512mem:$src),
3148 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3150 (OpNode (v16f32 (bitconvert (memopv16f32 addr:$src)))))],
3151 itins_s.rm>, EVEX, EVEX_V512, EVEX_CD8<32, CD8VF>;
3153 def PDZrr : AVX512PDI<opc, MRMSrcReg, (outs VR512:$dst), (ins VR512:$src),
3154 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3155 [(set VR512:$dst, (v8f64 (OpNode VR512:$src)))], itins_d.rr>,
3159 def PDZrm : AVX512PDI<opc, MRMSrcMem, (outs VR512:$dst), (ins f512mem:$src),
3160 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3161 [(set VR512:$dst, (OpNode
3162 (v8f64 (bitconvert (memopv16f32 addr:$src)))))],
3163 itins_d.rm>, EVEX, EVEX_V512, EVEX_CD8<64, CD8VF>;
3165 let isCodeGenOnly = 1 in {
3166 def PSZr_Int : AVX512PSI<opc, MRMSrcReg, (outs VR512:$dst), (ins VR512:$src),
3167 !strconcat(OpcodeStr,
3168 "ps\t{$src, $dst|$dst, $src}"),
3169 [(set VR512:$dst, (V16F32Int VR512:$src))]>,
3171 def PSZm_Int : AVX512PSI<opc, MRMSrcMem, (outs VR512:$dst), (ins f512mem:$src),
3172 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
3174 (V16F32Int (memopv16f32 addr:$src)))]>, EVEX,
3175 EVEX_V512, EVEX_CD8<32, CD8VF>;
3176 def PDZr_Int : AVX512PDI<opc, MRMSrcReg, (outs VR512:$dst), (ins VR512:$src),
3177 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
3178 [(set VR512:$dst, (V8F64Int VR512:$src))]>,
3179 EVEX, EVEX_V512, VEX_W;
3180 def PDZm_Int : AVX512PDI<opc, MRMSrcMem, (outs VR512:$dst), (ins f512mem:$src),
3181 !strconcat(OpcodeStr,
3182 "pd\t{$src, $dst|$dst, $src}"),
3183 [(set VR512:$dst, (V8F64Int (memopv8f64 addr:$src)))]>,
3184 EVEX, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
3185 } // isCodeGenOnly = 1
3188 multiclass avx512_sqrt_scalar<bits<8> opc, string OpcodeStr,
3189 Intrinsic F32Int, Intrinsic F64Int,
3190 OpndItins itins_s, OpndItins itins_d> {
3191 def SSZr : SI<opc, MRMSrcReg, (outs FR32X:$dst),
3192 (ins FR32X:$src1, FR32X:$src2),
3193 !strconcat(OpcodeStr,
3194 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3195 [], itins_s.rr>, XS, EVEX_4V;
3196 let isCodeGenOnly = 1 in
3197 def SSZr_Int : SIi8<opc, MRMSrcReg, (outs VR128X:$dst),
3198 (ins VR128X:$src1, VR128X:$src2),
3199 !strconcat(OpcodeStr,
3200 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3202 (F32Int VR128X:$src1, VR128X:$src2))],
3203 itins_s.rr>, XS, EVEX_4V;
3204 let mayLoad = 1 in {
3205 def SSZm : SI<opc, MRMSrcMem, (outs FR32X:$dst),
3206 (ins FR32X:$src1, f32mem:$src2),
3207 !strconcat(OpcodeStr,
3208 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3209 [], itins_s.rm>, XS, EVEX_4V, EVEX_CD8<32, CD8VT1>;
3210 let isCodeGenOnly = 1 in
3211 def SSZm_Int : SIi8<opc, MRMSrcMem, (outs VR128X:$dst),
3212 (ins VR128X:$src1, ssmem:$src2),
3213 !strconcat(OpcodeStr,
3214 "ss\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3216 (F32Int VR128X:$src1, sse_load_f32:$src2))],
3217 itins_s.rm>, XS, EVEX_4V, EVEX_CD8<32, CD8VT1>;
3219 def SDZr : SI<opc, MRMSrcReg, (outs FR64X:$dst),
3220 (ins FR64X:$src1, FR64X:$src2),
3221 !strconcat(OpcodeStr,
3222 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), []>,
3224 let isCodeGenOnly = 1 in
3225 def SDZr_Int : SIi8<opc, MRMSrcReg, (outs VR128X:$dst),
3226 (ins VR128X:$src1, VR128X:$src2),
3227 !strconcat(OpcodeStr,
3228 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3230 (F64Int VR128X:$src1, VR128X:$src2))],
3231 itins_s.rr>, XD, EVEX_4V, VEX_W;
3232 let mayLoad = 1 in {
3233 def SDZm : SI<opc, MRMSrcMem, (outs FR64X:$dst),
3234 (ins FR64X:$src1, f64mem:$src2),
3235 !strconcat(OpcodeStr,
3236 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"), []>,
3237 XD, EVEX_4V, VEX_W, EVEX_CD8<64, CD8VT1>;
3238 let isCodeGenOnly = 1 in
3239 def SDZm_Int : SIi8<opc, MRMSrcMem, (outs VR128X:$dst),
3240 (ins VR128X:$src1, sdmem:$src2),
3241 !strconcat(OpcodeStr,
3242 "sd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3244 (F64Int VR128X:$src1, sse_load_f64:$src2))]>,
3245 XD, EVEX_4V, VEX_W, EVEX_CD8<64, CD8VT1>;
3250 defm VSQRT : avx512_sqrt_scalar<0x51, "sqrt",
3251 int_x86_avx512_sqrt_ss, int_x86_avx512_sqrt_sd,
3252 SSE_SQRTSS, SSE_SQRTSD>,
3253 avx512_sqrt_packed<0x51, "vsqrt", fsqrt,
3254 int_x86_avx512_sqrt_ps_512, int_x86_avx512_sqrt_pd_512,
3255 SSE_SQRTPS, SSE_SQRTPD>;
3257 let Predicates = [HasAVX512] in {
3258 def : Pat<(f32 (fsqrt FR32X:$src)),
3259 (VSQRTSSZr (f32 (IMPLICIT_DEF)), FR32X:$src)>;
3260 def : Pat<(f32 (fsqrt (load addr:$src))),
3261 (VSQRTSSZm (f32 (IMPLICIT_DEF)), addr:$src)>,
3262 Requires<[OptForSize]>;
3263 def : Pat<(f64 (fsqrt FR64X:$src)),
3264 (VSQRTSDZr (f64 (IMPLICIT_DEF)), FR64X:$src)>;
3265 def : Pat<(f64 (fsqrt (load addr:$src))),
3266 (VSQRTSDZm (f64 (IMPLICIT_DEF)), addr:$src)>,
3267 Requires<[OptForSize]>;
3269 def : Pat<(f32 (X86frsqrt FR32X:$src)),
3270 (VRSQRT14SSZr (f32 (IMPLICIT_DEF)), FR32X:$src)>;
3271 def : Pat<(f32 (X86frsqrt (load addr:$src))),
3272 (VRSQRT14SSZm (f32 (IMPLICIT_DEF)), addr:$src)>,
3273 Requires<[OptForSize]>;
3275 def : Pat<(f32 (X86frcp FR32X:$src)),
3276 (VRCP14SSZr (f32 (IMPLICIT_DEF)), FR32X:$src)>;
3277 def : Pat<(f32 (X86frcp (load addr:$src))),
3278 (VRCP14SSZm (f32 (IMPLICIT_DEF)), addr:$src)>,
3279 Requires<[OptForSize]>;
3281 def : Pat<(int_x86_sse_sqrt_ss VR128X:$src),
3282 (COPY_TO_REGCLASS (VSQRTSSZr (f32 (IMPLICIT_DEF)),
3283 (COPY_TO_REGCLASS VR128X:$src, FR32)),
3285 def : Pat<(int_x86_sse_sqrt_ss sse_load_f32:$src),
3286 (VSQRTSSZm_Int (v4f32 (IMPLICIT_DEF)), sse_load_f32:$src)>;
3288 def : Pat<(int_x86_sse2_sqrt_sd VR128X:$src),
3289 (COPY_TO_REGCLASS (VSQRTSDZr (f64 (IMPLICIT_DEF)),
3290 (COPY_TO_REGCLASS VR128X:$src, FR64)),
3292 def : Pat<(int_x86_sse2_sqrt_sd sse_load_f64:$src),
3293 (VSQRTSDZm_Int (v2f64 (IMPLICIT_DEF)), sse_load_f64:$src)>;
3297 multiclass avx512_fp_unop_rm<bits<8> opcps, bits<8> opcpd, string OpcodeStr,
3298 X86MemOperand x86memop, RegisterClass RC,
3299 PatFrag mem_frag32, PatFrag mem_frag64,
3300 Intrinsic V4F32Int, Intrinsic V2F64Int,
3302 let ExeDomain = SSEPackedSingle in {
3303 // Intrinsic operation, reg.
3304 // Vector intrinsic operation, reg
3305 def PSr : AVX512AIi8<opcps, MRMSrcReg,
3306 (outs RC:$dst), (ins RC:$src1, i32i8imm:$src2),
3307 !strconcat(OpcodeStr,
3308 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3309 [(set RC:$dst, (V4F32Int RC:$src1, imm:$src2))]>;
3311 // Vector intrinsic operation, mem
3312 def PSm : AVX512AIi8<opcps, MRMSrcMem,
3313 (outs RC:$dst), (ins x86memop:$src1, i32i8imm:$src2),
3314 !strconcat(OpcodeStr,
3315 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3317 (V4F32Int (mem_frag32 addr:$src1),imm:$src2))]>,
3318 EVEX_CD8<32, VForm>;
3319 } // ExeDomain = SSEPackedSingle
3321 let ExeDomain = SSEPackedDouble in {
3322 // Vector intrinsic operation, reg
3323 def PDr : AVX512AIi8<opcpd, MRMSrcReg,
3324 (outs RC:$dst), (ins RC:$src1, i32i8imm:$src2),
3325 !strconcat(OpcodeStr,
3326 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3327 [(set RC:$dst, (V2F64Int RC:$src1, imm:$src2))]>;
3329 // Vector intrinsic operation, mem
3330 def PDm : AVX512AIi8<opcpd, MRMSrcMem,
3331 (outs RC:$dst), (ins x86memop:$src1, i32i8imm:$src2),
3332 !strconcat(OpcodeStr,
3333 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3335 (V2F64Int (mem_frag64 addr:$src1),imm:$src2))]>,
3336 EVEX_CD8<64, VForm>;
3337 } // ExeDomain = SSEPackedDouble
3340 multiclass avx512_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3344 let ExeDomain = GenericDomain in {
3346 let hasSideEffects = 0 in
3347 def SSr : AVX512AIi8<opcss, MRMSrcReg,
3348 (outs FR32X:$dst), (ins FR32X:$src1, FR32X:$src2, i32i8imm:$src3),
3349 !strconcat(OpcodeStr,
3350 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
3353 // Intrinsic operation, reg.
3354 let isCodeGenOnly = 1 in
3355 def SSr_Int : AVX512AIi8<opcss, MRMSrcReg,
3356 (outs VR128X:$dst), (ins VR128X:$src1, VR128X:$src2, i32i8imm:$src3),
3357 !strconcat(OpcodeStr,
3358 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
3359 [(set VR128X:$dst, (F32Int VR128X:$src1, VR128X:$src2, imm:$src3))]>;
3361 // Intrinsic operation, mem.
3362 def SSm : AVX512AIi8<opcss, MRMSrcMem, (outs VR128X:$dst),
3363 (ins VR128X:$src1, ssmem:$src2, i32i8imm:$src3),
3364 !strconcat(OpcodeStr,
3365 "ss\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
3366 [(set VR128X:$dst, (F32Int VR128X:$src1,
3367 sse_load_f32:$src2, imm:$src3))]>,
3368 EVEX_CD8<32, CD8VT1>;
3371 let hasSideEffects = 0 in
3372 def SDr : AVX512AIi8<opcsd, MRMSrcReg,
3373 (outs FR64X:$dst), (ins FR64X:$src1, FR64X:$src2, i32i8imm:$src3),
3374 !strconcat(OpcodeStr,
3375 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
3378 // Intrinsic operation, reg.
3379 let isCodeGenOnly = 1 in
3380 def SDr_Int : AVX512AIi8<opcsd, MRMSrcReg,
3381 (outs VR128X:$dst), (ins VR128X:$src1, VR128X:$src2, i32i8imm:$src3),
3382 !strconcat(OpcodeStr,
3383 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
3384 [(set VR128X:$dst, (F64Int VR128X:$src1, VR128X:$src2, imm:$src3))]>,
3387 // Intrinsic operation, mem.
3388 def SDm : AVX512AIi8<opcsd, MRMSrcMem,
3389 (outs VR128X:$dst), (ins VR128X:$src1, sdmem:$src2, i32i8imm:$src3),
3390 !strconcat(OpcodeStr,
3391 "sd\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
3393 (F64Int VR128X:$src1, sse_load_f64:$src2, imm:$src3))]>,
3394 VEX_W, EVEX_CD8<64, CD8VT1>;
3395 } // ExeDomain = GenericDomain
3398 multiclass avx512_rndscale<bits<8> opc, string OpcodeStr,
3399 X86MemOperand x86memop, RegisterClass RC,
3400 PatFrag mem_frag, Domain d> {
3401 let ExeDomain = d in {
3402 // Intrinsic operation, reg.
3403 // Vector intrinsic operation, reg
3404 def r : AVX512AIi8<opc, MRMSrcReg,
3405 (outs RC:$dst), (ins RC:$src1, i32i8imm:$src2),
3406 !strconcat(OpcodeStr,
3407 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3410 // Vector intrinsic operation, mem
3411 def m : AVX512AIi8<opc, MRMSrcMem,
3412 (outs RC:$dst), (ins x86memop:$src1, i32i8imm:$src2),
3413 !strconcat(OpcodeStr,
3414 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3420 defm VRNDSCALEPSZ : avx512_rndscale<0x08, "vrndscaleps", f512mem, VR512,
3421 memopv16f32, SSEPackedSingle>, EVEX_V512,
3422 EVEX_CD8<32, CD8VF>;
3424 def : Pat<(v16f32 (int_x86_avx512_mask_rndscale_ps_512 (v16f32 VR512:$src1),
3425 imm:$src2, (bc_v16f32 (v16i32 immAllZerosV)), (i16 -1),
3427 (VRNDSCALEPSZr VR512:$src1, imm:$src2)>;
3430 defm VRNDSCALEPDZ : avx512_rndscale<0x09, "vrndscalepd", f512mem, VR512,
3431 memopv8f64, SSEPackedDouble>, EVEX_V512,
3432 VEX_W, EVEX_CD8<64, CD8VF>;
3434 def : Pat<(v8f64 (int_x86_avx512_mask_rndscale_pd_512 (v8f64 VR512:$src1),
3435 imm:$src2, (bc_v8f64 (v16i32 immAllZerosV)), (i8 -1),
3437 (VRNDSCALEPDZr VR512:$src1, imm:$src2)>;
3439 multiclass avx512_rndscale_scalar<bits<8> opc, string OpcodeStr,
3440 Operand x86memop, RegisterClass RC, Domain d> {
3441 let ExeDomain = d in {
3442 def r : AVX512AIi8<opc, MRMSrcReg,
3443 (outs RC:$dst), (ins RC:$src1, RC:$src2, i32i8imm:$src3),
3444 !strconcat(OpcodeStr,
3445 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3448 def m : AVX512AIi8<opc, MRMSrcMem,
3449 (outs RC:$dst), (ins RC:$src1, x86memop:$src2, i32i8imm:$src3),
3450 !strconcat(OpcodeStr,
3451 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3456 defm VRNDSCALESS : avx512_rndscale_scalar<0x0A, "vrndscaless", ssmem, FR32X,
3457 SSEPackedSingle>, EVEX_CD8<32, CD8VT1>;
3459 defm VRNDSCALESD : avx512_rndscale_scalar<0x0B, "vrndscalesd", sdmem, FR64X,
3460 SSEPackedDouble>, EVEX_CD8<64, CD8VT1>;
3462 def : Pat<(ffloor FR32X:$src),
3463 (VRNDSCALESSr (f32 (IMPLICIT_DEF)), FR32X:$src, (i32 0x1))>;
3464 def : Pat<(f64 (ffloor FR64X:$src)),
3465 (VRNDSCALESDr (f64 (IMPLICIT_DEF)), FR64X:$src, (i32 0x1))>;
3466 def : Pat<(f32 (fnearbyint FR32X:$src)),
3467 (VRNDSCALESSr (f32 (IMPLICIT_DEF)), FR32X:$src, (i32 0xC))>;
3468 def : Pat<(f64 (fnearbyint FR64X:$src)),
3469 (VRNDSCALESDr (f64 (IMPLICIT_DEF)), FR64X:$src, (i32 0xC))>;
3470 def : Pat<(f32 (fceil FR32X:$src)),
3471 (VRNDSCALESSr (f32 (IMPLICIT_DEF)), FR32X:$src, (i32 0x2))>;
3472 def : Pat<(f64 (fceil FR64X:$src)),
3473 (VRNDSCALESDr (f64 (IMPLICIT_DEF)), FR64X:$src, (i32 0x2))>;
3474 def : Pat<(f32 (frint FR32X:$src)),
3475 (VRNDSCALESSr (f32 (IMPLICIT_DEF)), FR32X:$src, (i32 0x4))>;
3476 def : Pat<(f64 (frint FR64X:$src)),
3477 (VRNDSCALESDr (f64 (IMPLICIT_DEF)), FR64X:$src, (i32 0x4))>;
3478 def : Pat<(f32 (ftrunc FR32X:$src)),
3479 (VRNDSCALESSr (f32 (IMPLICIT_DEF)), FR32X:$src, (i32 0x3))>;
3480 def : Pat<(f64 (ftrunc FR64X:$src)),
3481 (VRNDSCALESDr (f64 (IMPLICIT_DEF)), FR64X:$src, (i32 0x3))>;
3483 def : Pat<(v16f32 (ffloor VR512:$src)),
3484 (VRNDSCALEPSZr VR512:$src, (i32 0x1))>;
3485 def : Pat<(v16f32 (fnearbyint VR512:$src)),
3486 (VRNDSCALEPSZr VR512:$src, (i32 0xC))>;
3487 def : Pat<(v16f32 (fceil VR512:$src)),
3488 (VRNDSCALEPSZr VR512:$src, (i32 0x2))>;
3489 def : Pat<(v16f32 (frint VR512:$src)),
3490 (VRNDSCALEPSZr VR512:$src, (i32 0x4))>;
3491 def : Pat<(v16f32 (ftrunc VR512:$src)),
3492 (VRNDSCALEPSZr VR512:$src, (i32 0x3))>;
3494 def : Pat<(v8f64 (ffloor VR512:$src)),
3495 (VRNDSCALEPDZr VR512:$src, (i32 0x1))>;
3496 def : Pat<(v8f64 (fnearbyint VR512:$src)),
3497 (VRNDSCALEPDZr VR512:$src, (i32 0xC))>;
3498 def : Pat<(v8f64 (fceil VR512:$src)),
3499 (VRNDSCALEPDZr VR512:$src, (i32 0x2))>;
3500 def : Pat<(v8f64 (frint VR512:$src)),
3501 (VRNDSCALEPDZr VR512:$src, (i32 0x4))>;
3502 def : Pat<(v8f64 (ftrunc VR512:$src)),
3503 (VRNDSCALEPDZr VR512:$src, (i32 0x3))>;
3505 //-------------------------------------------------
3506 // Integer truncate and extend operations
3507 //-------------------------------------------------
3509 multiclass avx512_trunc_sat<bits<8> opc, string OpcodeStr,
3510 RegisterClass dstRC, RegisterClass srcRC,
3511 RegisterClass KRC, X86MemOperand x86memop> {
3512 def rr : AVX512XS8I<opc, MRMDestReg, (outs dstRC:$dst),
3514 !strconcat(OpcodeStr,"\t{$src, $dst|$dst, $src}"),
3517 def krr : AVX512XS8I<opc, MRMDestReg, (outs dstRC:$dst),
3518 (ins KRC:$mask, srcRC:$src),
3519 !strconcat(OpcodeStr,
3520 "\t{$src, ${dst} {${mask}} {z}|${dst} {${mask}} {z}, $src}"),
3523 def mr : AVX512XS8I<opc, MRMDestMem, (outs), (ins x86memop:$dst, srcRC:$src),
3524 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3527 defm VPMOVQB : avx512_trunc_sat<0x32, "vpmovqb", VR128X, VR512, VK8WM,
3528 i128mem>, EVEX_V512, EVEX_CD8<8, CD8VO>;
3529 defm VPMOVSQB : avx512_trunc_sat<0x22, "vpmovsqb", VR128X, VR512, VK8WM,
3530 i128mem>, EVEX_V512, EVEX_CD8<8, CD8VO>;
3531 defm VPMOVUSQB : avx512_trunc_sat<0x12, "vpmovusqb", VR128X, VR512, VK8WM,
3532 i128mem>, EVEX_V512, EVEX_CD8<8, CD8VO>;
3533 defm VPMOVQW : avx512_trunc_sat<0x34, "vpmovqw", VR128X, VR512, VK8WM,
3534 i128mem>, EVEX_V512, EVEX_CD8<16, CD8VQ>;
3535 defm VPMOVSQW : avx512_trunc_sat<0x24, "vpmovsqw", VR128X, VR512, VK8WM,
3536 i128mem>, EVEX_V512, EVEX_CD8<16, CD8VQ>;
3537 defm VPMOVUSQW : avx512_trunc_sat<0x14, "vpmovusqw", VR128X, VR512, VK8WM,
3538 i128mem>, EVEX_V512, EVEX_CD8<16, CD8VQ>;
3539 defm VPMOVQD : avx512_trunc_sat<0x35, "vpmovqd", VR256X, VR512, VK8WM,
3540 i256mem>, EVEX_V512, EVEX_CD8<32, CD8VH>;
3541 defm VPMOVSQD : avx512_trunc_sat<0x25, "vpmovsqd", VR256X, VR512, VK8WM,
3542 i256mem>, EVEX_V512, EVEX_CD8<32, CD8VH>;
3543 defm VPMOVUSQD : avx512_trunc_sat<0x15, "vpmovusqd", VR256X, VR512, VK8WM,
3544 i256mem>, EVEX_V512, EVEX_CD8<32, CD8VH>;
3545 defm VPMOVDW : avx512_trunc_sat<0x33, "vpmovdw", VR256X, VR512, VK16WM,
3546 i256mem>, EVEX_V512, EVEX_CD8<16, CD8VH>;
3547 defm VPMOVSDW : avx512_trunc_sat<0x23, "vpmovsdw", VR256X, VR512, VK16WM,
3548 i256mem>, EVEX_V512, EVEX_CD8<16, CD8VH>;
3549 defm VPMOVUSDW : avx512_trunc_sat<0x13, "vpmovusdw", VR256X, VR512, VK16WM,
3550 i256mem>, EVEX_V512, EVEX_CD8<16, CD8VH>;
3551 defm VPMOVDB : avx512_trunc_sat<0x31, "vpmovdb", VR128X, VR512, VK16WM,
3552 i128mem>, EVEX_V512, EVEX_CD8<8, CD8VQ>;
3553 defm VPMOVSDB : avx512_trunc_sat<0x21, "vpmovsdb", VR128X, VR512, VK16WM,
3554 i128mem>, EVEX_V512, EVEX_CD8<8, CD8VQ>;
3555 defm VPMOVUSDB : avx512_trunc_sat<0x11, "vpmovusdb", VR128X, VR512, VK16WM,
3556 i128mem>, EVEX_V512, EVEX_CD8<8, CD8VQ>;
3558 def : Pat<(v16i8 (X86vtrunc (v8i64 VR512:$src))), (VPMOVQBrr VR512:$src)>;
3559 def : Pat<(v8i16 (X86vtrunc (v8i64 VR512:$src))), (VPMOVQWrr VR512:$src)>;
3560 def : Pat<(v16i16 (X86vtrunc (v16i32 VR512:$src))), (VPMOVDWrr VR512:$src)>;
3561 def : Pat<(v16i8 (X86vtrunc (v16i32 VR512:$src))), (VPMOVDBrr VR512:$src)>;
3562 def : Pat<(v8i32 (X86vtrunc (v8i64 VR512:$src))), (VPMOVQDrr VR512:$src)>;
3564 def : Pat<(v16i8 (X86vtruncm VK16WM:$mask, (v16i32 VR512:$src))),
3565 (VPMOVDBkrr VK16WM:$mask, VR512:$src)>;
3566 def : Pat<(v16i16 (X86vtruncm VK16WM:$mask, (v16i32 VR512:$src))),
3567 (VPMOVDWkrr VK16WM:$mask, VR512:$src)>;
3568 def : Pat<(v8i16 (X86vtruncm VK8WM:$mask, (v8i64 VR512:$src))),
3569 (VPMOVQWkrr VK8WM:$mask, VR512:$src)>;
3570 def : Pat<(v8i32 (X86vtruncm VK8WM:$mask, (v8i64 VR512:$src))),
3571 (VPMOVQDkrr VK8WM:$mask, VR512:$src)>;
3574 multiclass avx512_extend<bits<8> opc, string OpcodeStr, RegisterClass DstRC,
3575 RegisterClass SrcRC, SDNode OpNode, PatFrag mem_frag,
3576 X86MemOperand x86memop, ValueType OpVT, ValueType InVT> {
3578 def rr : AVX5128I<opc, MRMSrcReg, (outs DstRC:$dst),
3580 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3581 [(set DstRC:$dst, (OpVT (OpNode (InVT SrcRC:$src))))]>, EVEX;
3582 def rm : AVX5128I<opc, MRMSrcMem, (outs DstRC:$dst),
3583 (ins x86memop:$src),
3584 !strconcat(OpcodeStr,"\t{$src, $dst|$dst, $src}"),
3586 (OpVT (OpNode (InVT (bitconvert (mem_frag addr:$src))))))]>,
3590 defm VPMOVZXBDZ: avx512_extend<0x31, "vpmovzxbd", VR512, VR128X, X86vzext,
3591 memopv2i64, i128mem, v16i32, v16i8>, EVEX_V512,
3593 defm VPMOVZXBQZ: avx512_extend<0x32, "vpmovzxbq", VR512, VR128X, X86vzext,
3594 memopv2i64, i128mem, v8i64, v16i8>, EVEX_V512,
3596 defm VPMOVZXWDZ: avx512_extend<0x33, "vpmovzxwd", VR512, VR256X, X86vzext,
3597 memopv4i64, i256mem, v16i32, v16i16>, EVEX_V512,
3598 EVEX_CD8<16, CD8VH>;
3599 defm VPMOVZXWQZ: avx512_extend<0x34, "vpmovzxwq", VR512, VR128X, X86vzext,
3600 memopv2i64, i128mem, v8i64, v8i16>, EVEX_V512,
3601 EVEX_CD8<16, CD8VQ>;
3602 defm VPMOVZXDQZ: avx512_extend<0x35, "vpmovzxdq", VR512, VR256X, X86vzext,
3603 memopv4i64, i256mem, v8i64, v8i32>, EVEX_V512,
3604 EVEX_CD8<32, CD8VH>;
3606 defm VPMOVSXBDZ: avx512_extend<0x21, "vpmovsxbd", VR512, VR128X, X86vsext,
3607 memopv2i64, i128mem, v16i32, v16i8>, EVEX_V512,
3609 defm VPMOVSXBQZ: avx512_extend<0x22, "vpmovsxbq", VR512, VR128X, X86vsext,
3610 memopv2i64, i128mem, v8i64, v16i8>, EVEX_V512,
3612 defm VPMOVSXWDZ: avx512_extend<0x23, "vpmovsxwd", VR512, VR256X, X86vsext,
3613 memopv4i64, i256mem, v16i32, v16i16>, EVEX_V512,
3614 EVEX_CD8<16, CD8VH>;
3615 defm VPMOVSXWQZ: avx512_extend<0x24, "vpmovsxwq", VR512, VR128X, X86vsext,
3616 memopv2i64, i128mem, v8i64, v8i16>, EVEX_V512,
3617 EVEX_CD8<16, CD8VQ>;
3618 defm VPMOVSXDQZ: avx512_extend<0x25, "vpmovsxdq", VR512, VR256X, X86vsext,
3619 memopv4i64, i256mem, v8i64, v8i32>, EVEX_V512,
3620 EVEX_CD8<32, CD8VH>;
3622 //===----------------------------------------------------------------------===//
3623 // GATHER - SCATTER Operations
3625 multiclass avx512_gather<bits<8> opc, string OpcodeStr, RegisterClass KRC,
3626 RegisterClass RC, X86MemOperand memop> {
3628 Constraints = "@earlyclobber $dst, $src1 = $dst, $mask = $mask_wb" in
3629 def rm : AVX5128I<opc, MRMSrcMem, (outs RC:$dst, KRC:$mask_wb),
3630 (ins RC:$src1, KRC:$mask, memop:$src2),
3631 !strconcat(OpcodeStr,
3632 "\t{$src2, ${dst} {${mask}}|${dst} {${mask}}, $src2}"),
3635 defm VGATHERDPDZ : avx512_gather<0x92, "vgatherdpd", VK8WM, VR512, vy64xmem>,
3636 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT1>;
3637 defm VGATHERDPSZ : avx512_gather<0x92, "vgatherdps", VK16WM, VR512, vz32mem>,
3638 EVEX_V512, EVEX_CD8<32, CD8VT1>;
3640 defm VGATHERQPDZ : avx512_gather<0x93, "vgatherqpd", VK8WM, VR512, vz64mem>,
3641 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT1>;
3642 defm VGATHERQPSZ : avx512_gather<0x93, "vgatherqps", VK8WM, VR256X, vz64mem>,
3643 EVEX_V512, EVEX_CD8<32, CD8VT1>;
3645 defm VPGATHERDQZ : avx512_gather<0x90, "vpgatherdq", VK8WM, VR512, vy64xmem>,
3646 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT1>;
3647 defm VPGATHERDDZ : avx512_gather<0x90, "vpgatherdd", VK16WM, VR512, vz32mem>,
3648 EVEX_V512, EVEX_CD8<32, CD8VT1>;
3650 defm VPGATHERQQZ : avx512_gather<0x91, "vpgatherqq", VK8WM, VR512, vz64mem>,
3651 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT1>;
3652 defm VPGATHERQDZ : avx512_gather<0x91, "vpgatherqd", VK8WM, VR256X, vz64mem>,
3653 EVEX_V512, EVEX_CD8<32, CD8VT1>;
3655 multiclass avx512_scatter<bits<8> opc, string OpcodeStr, RegisterClass KRC,
3656 RegisterClass RC, X86MemOperand memop> {
3657 let mayStore = 1, Constraints = "$mask = $mask_wb" in
3658 def mr : AVX5128I<opc, MRMDestMem, (outs KRC:$mask_wb),
3659 (ins memop:$dst, KRC:$mask, RC:$src2),
3660 !strconcat(OpcodeStr,
3661 "\t{$src2, ${dst} {${mask}}|${dst} {${mask}}, $src2}"),
3665 defm VSCATTERDPDZ : avx512_scatter<0xA2, "vscatterdpd", VK8WM, VR512, vy64xmem>,
3666 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT1>;
3667 defm VSCATTERDPSZ : avx512_scatter<0xA2, "vscatterdps", VK16WM, VR512, vz32mem>,
3668 EVEX_V512, EVEX_CD8<32, CD8VT1>;
3670 defm VSCATTERQPDZ : avx512_scatter<0xA3, "vscatterqpd", VK8WM, VR512, vz64mem>,
3671 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT1>;
3672 defm VSCATTERQPSZ : avx512_scatter<0xA3, "vscatterqps", VK8WM, VR256X, vz64mem>,
3673 EVEX_V512, EVEX_CD8<32, CD8VT1>;
3675 defm VPSCATTERDQZ : avx512_scatter<0xA0, "vpscatterdq", VK8WM, VR512, vy64xmem>,
3676 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT1>;
3677 defm VPSCATTERDDZ : avx512_scatter<0xA0, "vpscatterdd", VK16WM, VR512, vz32mem>,
3678 EVEX_V512, EVEX_CD8<32, CD8VT1>;
3680 defm VPSCATTERQQZ : avx512_scatter<0xA1, "vpscatterqq", VK8WM, VR512, vz64mem>,
3681 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VT1>;
3682 defm VPSCATTERQDZ : avx512_scatter<0xA1, "vpscatterqd", VK8WM, VR256X, vz64mem>,
3683 EVEX_V512, EVEX_CD8<32, CD8VT1>;
3685 //===----------------------------------------------------------------------===//
3686 // VSHUFPS - VSHUFPD Operations
3688 multiclass avx512_shufp<RegisterClass RC, X86MemOperand x86memop,
3689 ValueType vt, string OpcodeStr, PatFrag mem_frag,
3691 def rmi : AVX512PIi8<0xC6, MRMSrcMem, (outs RC:$dst),
3692 (ins RC:$src1, x86memop:$src2, i8imm:$src3),
3693 !strconcat(OpcodeStr,
3694 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
3695 [(set RC:$dst, (vt (X86Shufp RC:$src1, (mem_frag addr:$src2),
3696 (i8 imm:$src3))))], d, IIC_SSE_SHUFP>,
3697 EVEX_4V, Sched<[WriteShuffleLd, ReadAfterLd]>;
3698 def rri : AVX512PIi8<0xC6, MRMSrcReg, (outs RC:$dst),
3699 (ins RC:$src1, RC:$src2, i8imm:$src3),
3700 !strconcat(OpcodeStr,
3701 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
3702 [(set RC:$dst, (vt (X86Shufp RC:$src1, RC:$src2,
3703 (i8 imm:$src3))))], d, IIC_SSE_SHUFP>,
3704 EVEX_4V, Sched<[WriteShuffle]>;
3707 defm VSHUFPSZ : avx512_shufp<VR512, f512mem, v16f32, "vshufps", memopv16f32,
3708 SSEPackedSingle>, EVEX_V512, EVEX_CD8<32, CD8VF>;
3709 defm VSHUFPDZ : avx512_shufp<VR512, f512mem, v8f64, "vshufpd", memopv8f64,
3710 SSEPackedDouble>, OpSize, VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>;
3712 def : Pat<(v16i32 (X86Shufp VR512:$src1, VR512:$src2, (i8 imm:$imm))),
3713 (VSHUFPSZrri VR512:$src1, VR512:$src2, imm:$imm)>;
3714 def : Pat<(v16i32 (X86Shufp VR512:$src1,
3715 (memopv16i32 addr:$src2), (i8 imm:$imm))),
3716 (VSHUFPSZrmi VR512:$src1, addr:$src2, imm:$imm)>;
3718 def : Pat<(v8i64 (X86Shufp VR512:$src1, VR512:$src2, (i8 imm:$imm))),
3719 (VSHUFPDZrri VR512:$src1, VR512:$src2, imm:$imm)>;
3720 def : Pat<(v8i64 (X86Shufp VR512:$src1,
3721 (memopv8i64 addr:$src2), (i8 imm:$imm))),
3722 (VSHUFPDZrmi VR512:$src1, addr:$src2, imm:$imm)>;
3724 multiclass avx512_alignr<string OpcodeStr, RegisterClass RC,
3725 X86MemOperand x86memop> {
3726 def rri : AVX512AIi8<0x03, MRMSrcReg, (outs RC:$dst),
3727 (ins RC:$src1, RC:$src2, i8imm:$src3),
3728 !strconcat(OpcodeStr,
3729 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
3732 def rmi : AVX512AIi8<0x03, MRMSrcMem, (outs RC:$dst),
3733 (ins RC:$src1, x86memop:$src2, i8imm:$src3),
3734 !strconcat(OpcodeStr,
3735 "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
3738 defm VALIGND : avx512_alignr<"valignd", VR512, i512mem>,
3739 EVEX_V512, EVEX_CD8<32, CD8VF>;
3740 defm VALIGNQ : avx512_alignr<"valignq", VR512, i512mem>,
3741 VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>;
3743 def : Pat<(v16f32 (X86PAlignr VR512:$src1, VR512:$src2, (i8 imm:$imm))),
3744 (VALIGNDrri VR512:$src2, VR512:$src1, imm:$imm)>;
3745 def : Pat<(v8f64 (X86PAlignr VR512:$src1, VR512:$src2, (i8 imm:$imm))),
3746 (VALIGNQrri VR512:$src2, VR512:$src1, imm:$imm)>;
3747 def : Pat<(v16i32 (X86PAlignr VR512:$src1, VR512:$src2, (i8 imm:$imm))),
3748 (VALIGNDrri VR512:$src2, VR512:$src1, imm:$imm)>;
3749 def : Pat<(v8i64 (X86PAlignr VR512:$src1, VR512:$src2, (i8 imm:$imm))),
3750 (VALIGNQrri VR512:$src2, VR512:$src1, imm:$imm)>;
3752 multiclass avx512_vpabs<bits<8> opc, string OpcodeStr, RegisterClass RC,
3753 X86MemOperand x86memop> {
3754 def rr : AVX5128I<opc, MRMSrcReg, (outs RC:$dst), (ins RC:$src),
3755 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"), []>,
3757 def rm : AVX5128I<opc, MRMSrcMem, (outs VR512:$dst),
3758 (ins x86memop:$src),
3759 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"), []>,
3763 defm VPABSD : avx512_vpabs<0x1E, "vpabsd", VR512, i512mem>, EVEX_V512,
3764 EVEX_CD8<32, CD8VF>;
3765 defm VPABSQ : avx512_vpabs<0x1F, "vpabsq", VR512, i512mem>, EVEX_V512, VEX_W,
3766 EVEX_CD8<64, CD8VF>;
3768 multiclass avx512_conflict<bits<8> opc, string OpcodeStr,
3769 RegisterClass RC, RegisterClass KRC,
3770 X86MemOperand x86memop,
3771 X86MemOperand x86scalar_mop, string BrdcstStr> {
3772 def rr : AVX5128I<opc, MRMSrcReg, (outs RC:$dst),
3774 !strconcat(OpcodeStr, "\t{$src, ${dst} |${dst}, $src}"),
3776 def rm : AVX5128I<opc, MRMSrcMem, (outs RC:$dst),
3777 (ins x86memop:$src),
3778 !strconcat(OpcodeStr, "\t{$src, ${dst}|${dst}, $src}"),
3780 def rmb : AVX5128I<opc, MRMSrcMem, (outs RC:$dst),
3781 (ins x86scalar_mop:$src),
3782 !strconcat(OpcodeStr, "\t{${src}", BrdcstStr,
3783 ", ${dst}|${dst}, ${src}", BrdcstStr, "}"),
3785 def rrkz : AVX5128I<opc, MRMSrcReg, (outs RC:$dst),
3786 (ins KRC:$mask, RC:$src),
3787 !strconcat(OpcodeStr,
3788 "\t{$src, ${dst} {${mask}} {z}|${dst} {${mask}} {z}, $src}"),
3790 def rmkz : AVX5128I<opc, MRMSrcMem, (outs RC:$dst),
3791 (ins KRC:$mask, x86memop:$src),
3792 !strconcat(OpcodeStr,
3793 "\t{$src, ${dst} {${mask}} {z}|${dst} {${mask}} {z}, $src}"),
3795 def rmbkz : AVX5128I<opc, MRMSrcMem, (outs RC:$dst),
3796 (ins KRC:$mask, x86scalar_mop:$src),
3797 !strconcat(OpcodeStr, "\t{${src}", BrdcstStr,
3798 ", ${dst} {${mask}} {z}|${dst} {${mask}} {z}, ${src}",
3800 []>, EVEX, EVEX_KZ, EVEX_B;
3802 let Constraints = "$src1 = $dst" in {
3803 def rrk : AVX5128I<opc, MRMSrcReg, (outs RC:$dst),
3804 (ins RC:$src1, KRC:$mask, RC:$src2),
3805 !strconcat(OpcodeStr,
3806 "\t{$src2, ${dst} {${mask}}|${dst} {${mask}}, $src2}"),
3808 def rmk : AVX5128I<opc, MRMSrcMem, (outs RC:$dst),
3809 (ins RC:$src1, KRC:$mask, x86memop:$src2),
3810 !strconcat(OpcodeStr,
3811 "\t{$src2, ${dst} {${mask}}|${dst} {${mask}}, $src2}"),
3813 def rmbk : AVX5128I<opc, MRMSrcMem, (outs RC:$dst),
3814 (ins RC:$src1, KRC:$mask, x86scalar_mop:$src2),
3815 !strconcat(OpcodeStr, "\t{${src2}", BrdcstStr,
3816 ", ${dst} {${mask}}|${dst} {${mask}}, ${src2}", BrdcstStr, "}"),
3817 []>, EVEX, EVEX_K, EVEX_B;
3821 let Predicates = [HasCDI] in {
3822 defm VPCONFLICTD : avx512_conflict<0xC4, "vpconflictd", VR512, VK16WM,
3823 i512mem, i32mem, "{1to16}">,
3824 EVEX_V512, EVEX_CD8<32, CD8VF>;
3827 defm VPCONFLICTQ : avx512_conflict<0xC4, "vpconflictq", VR512, VK8WM,
3828 i512mem, i64mem, "{1to8}">,
3829 EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
3833 def : Pat<(int_x86_avx512_mask_conflict_d_512 VR512:$src2, VR512:$src1,
3835 (VPCONFLICTDrrk VR512:$src1,
3836 (v16i1 (COPY_TO_REGCLASS GR16:$mask, VK16WM)), VR512:$src2)>;
3838 def : Pat<(int_x86_avx512_mask_conflict_q_512 VR512:$src2, VR512:$src1,
3840 (VPCONFLICTQrrk VR512:$src1,
3841 (v8i1 (COPY_TO_REGCLASS GR8:$mask, VK8WM)), VR512:$src2)>;