1 /* PR target/87767 */
2 /* { dg-do compile } */
3 /* { dg-options "-O2 -mavx512f -mavx512vl" } */
4 /* { dg-additional-options "-fno-PIE" { target ia32 } } */
5 /* { dg-additional-options "-mdynamic-no-pic" { target { *-*-darwin* && ia32 } } }
6 /* { dg-final { scan-assembler-times "vadd\[^\n\]*\\\{1to2\\\}" 1 } } */
7 /* { dg-final { scan-assembler-times "vadd\[^\n\]*\\\{1to4\\\}" 2 } } */
8 /* { dg-final { scan-assembler-times "vadd\[^\n\]*\\\{1to8\\\}" 2 } } */
9 /* { dg-final { scan-assembler-times "vadd\[^\n\]*\\\{1to16\\\}" 1 } } */
10 /* { dg-final { scan-assembler-times "vsub\[^\n\]*\\\{1to2\\\}" 1 } } */
11 /* { dg-final { scan-assembler-times "vsub\[^\n\]*\\\{1to4\\\}" 2 } } */
12 /* { dg-final { scan-assembler-times "vsub\[^\n\]*\\\{1to8\\\}" 2 } } */
13 /* { dg-final { scan-assembler-times "vsub\[^\n\]*\\\{1to16\\\}" 1 } } */
14 /* { dg-final { scan-assembler-times "vmul\[^\n\]*\\\{1to2\\\}" 1 } } */
15 /* { dg-final { scan-assembler-times "vmul\[^\n\]*\\\{1to4\\\}" 2 } } */
16 /* { dg-final { scan-assembler-times "vmul\[^\n\]*\\\{1to8\\\}" 2 } } */
17 /* { dg-final { scan-assembler-times "vmul\[^\n\]*\\\{1to16\\\}" 1 } } */
18 /* { dg-final { scan-assembler-times "vdiv\[^\n\]*\\\{1to2\\\}" 1 } } */
19 /* { dg-final { scan-assembler-times "vdiv\[^\n\]*\\\{1to4\\\}" 2 } } */
20 /* { dg-final { scan-assembler-times "vdiv\[^\n\]*\\\{1to8\\\}" 2 } } */
21 /* { dg-final { scan-assembler-times "vdiv\[^\n\]*\\\{1to16\\\}" 1 } } */
22 /* { dg-final { scan-assembler-times "vfmadd\[^\n\]*\\\{1to2\\\}" 1 } } */
23 /* { dg-final { scan-assembler-times "vfmadd\[^\n\]*\\\{1to4\\\}" 2 } } */
24 /* { dg-final { scan-assembler-times "vfmadd\[^\n\]*\\\{1to8\\\}" 2 } } */
25 /* { dg-final { scan-assembler-times "vfmadd\[^\n\]*\\\{1to16\\\}" 1 } } */
26 /* { dg-final { scan-assembler-times "vfmsub\[^\n\]*\\\{1to2\\\}" 1 } } */
27 /* { dg-final { scan-assembler-times "vfmsub\[^\n\]*\\\{1to4\\\}" 2 } } */
28 /* { dg-final { scan-assembler-times "vfmsub\[^\n\]*\\\{1to8\\\}" 2 } } */
29 /* { dg-final { scan-assembler-times "vfmsub\[^\n\]*\\\{1to16\\\}" 1 } } */
30 /* { dg-final { scan-assembler-times "vfnmadd\[^\n\]*\\\{1to2\\\}" 1 } } */
31 /* { dg-final { scan-assembler-times "vfnmadd\[^\n\]*\\\{1to4\\\}" 2 } } */
32 /* { dg-final { scan-assembler-times "vfnmadd\[^\n\]*\\\{1to8\\\}" 2 } } */
33 /* { dg-final { scan-assembler-times "vfnmadd\[^\n\]*\\\{1to16\\\}" 1 } } */
34 /* { dg-final { scan-assembler-times "vfnmsub\[^\n\]*\\\{1to2\\\}" 1 } } */
35 /* { dg-final { scan-assembler-times "vfnmsub\[^\n\]*\\\{1to4\\\}" 2 } } */
36 /* { dg-final { scan-assembler-times "vfnmsub\[^\n\]*\\\{1to8\\\}" 2 } } */
37 /* { dg-final { scan-assembler-times "vfnmsub\[^\n\]*\\\{1to16\\\}" 1 } } */
38
39 #include<immintrin.h>
40
41 #define CONSTANT 101
42
43 #define FOO(VTYPE, OP_NAME, LEN, SUFFIX, MTYPE) \
44 VTYPE \
45 __attribute__ ((noipa)) \
46 _mm##LEN##_foo_##OP_NAME##_##SUFFIX (VTYPE dst, VTYPE src, MTYPE m) \
47 { \
48 return _mm##LEN##_mask_##OP_NAME##_##SUFFIX (dst, m, src, \
49 _mm##LEN##_set1_##SUFFIX (CONSTANT)); \
50 } \
51
52 #define FOOZ(VTYPE, OP_NAME, LEN, SUFFIX, MTYPE) \
53 VTYPE \
54 __attribute__ ((noipa)) \
55 _mm##LEN##_fooz_##OP_NAME##_##SUFFIX (VTYPE dst, VTYPE src, MTYPE m) \
56 { \
57 return _mm##LEN##_maskz_##OP_NAME##_##SUFFIX (m, dst, src, \
58 _mm##LEN##_set1_##SUFFIX (CONSTANT)); \
59 } \
60
61 FOO (__m512, add, 512, ps, __mmask16);
62 FOO (__m256, add, 256, ps, __mmask8);
63 FOO (__m128, add,, ps, __mmask8);
64
65 FOO (__m512, sub, 512, ps, __mmask16);
66 FOO (__m256, sub, 256, ps, __mmask8);
67 FOO (__m128, sub,, ps, __mmask8);
68
69 FOO (__m512, mul, 512, ps, __mmask16);
70 FOO (__m256, mul, 256, ps, __mmask8);
71 FOO (__m128, mul,, ps, __mmask8);
72
73 FOO (__m512, div, 512, ps, __mmask16);
74 FOO (__m256, div, 256, ps, __mmask8);
75 FOO (__m128, div,, ps, __mmask8);
76
77 FOOZ (__m512, fmadd, 512, ps, __mmask16);
78 FOOZ (__m256, fmadd, 256, ps, __mmask8);
79 FOOZ (__m128, fmadd,, ps, __mmask8);
80
81 FOOZ (__m512, fmsub, 512, ps, __mmask16);
82 FOOZ (__m256, fmsub, 256, ps, __mmask8);
83 FOOZ (__m128, fmsub,, ps, __mmask8);
84
85 FOOZ (__m512, fnmadd, 512, ps, __mmask16);
86 FOOZ (__m256, fnmadd, 256, ps, __mmask8);
87 FOOZ (__m128, fnmadd,, ps, __mmask8);
88
89 FOOZ (__m512, fnmsub, 512, ps, __mmask16);
90 FOOZ (__m256, fnmsub, 256, ps, __mmask8);
91 FOOZ (__m128, fnmsub,, ps, __mmask8);
92
93 FOO (__m512d, add, 512, pd, __mmask8);
94 FOO (__m256d, add, 256, pd, __mmask8);
95 FOO (__m128d, add,, pd, __mmask8);
96
97 FOO (__m512d, sub, 512, pd, __mmask8);
98 FOO (__m256d, sub, 256, pd, __mmask8);
99 FOO (__m128d, sub,, pd, __mmask8);
100
101 FOO (__m512d, mul, 512, pd, __mmask8);
102 FOO (__m256d, mul, 256, pd, __mmask8);
103 FOO (__m128d, mul,, pd, __mmask8);
104
105 FOO (__m512d, div, 512, pd, __mmask8);
106 FOO (__m256d, div, 256, pd, __mmask8);
107 FOO (__m128d, div,, pd, __mmask8);
108
109 FOOZ (__m512d, fmadd, 512, pd, __mmask8);
110 FOOZ (__m256d, fmadd, 256, pd, __mmask8);
111 FOOZ (__m128d, fmadd,, pd, __mmask8);
112
113 FOOZ (__m512d, fmsub, 512, pd, __mmask8);
114 FOOZ (__m256d, fmsub, 256, pd, __mmask8);
115 FOOZ (__m128d, fmsub,, pd, __mmask8);
116
117 FOOZ (__m512d, fnmadd, 512, pd, __mmask8);
118 FOOZ (__m256d, fnmadd, 256, pd, __mmask8);
119 FOOZ (__m128d, fnmadd,, pd, __mmask8);
120
121 FOOZ (__m512d, fnmsub, 512, pd, __mmask8);
122 FOOZ (__m256d, fnmsub, 256, pd, __mmask8);
123 FOOZ (__m128d, fnmsub,, pd, __mmask8);