Initial support for AVX-512{VL,BW,DQ}
[official-gcc.git] / gcc / testsuite / gcc.target / i386 / sse-14.c
blob28475c443024f76036a3881833b99395550d0824
1 /* { dg-do compile } */
2 /* { dg-options "-O0 -Werror-implicit-function-declaration -march=k8 -msse4a -m3dnow -mavx -mavx2 -mfma4 -mxop -maes -mpclmul -mpopcnt -mabm -mlzcnt -mbmi -mbmi2 -mtbm -mlwp -mfsgsbase -mrdrnd -mf16c -mfma -mrtm -mrdseed -mprfchw -madx -mfxsr -mxsaveopt -mavx512f -mavx512er -mavx512cd -mavx512pf -msha -mprefetchwt1 -mxsavec -mxsaves -mclflushopt -mavx512dq -mavx512bw -mavx512vl" } */
4 #include <mm_malloc.h>
6 /* Test that the intrinsics compile without optimization. All of them are
7 defined as inline functions in {,x,e,p,t,s,w,a,b,i}mmintrin.h, mm3dnow.h,
8 fma4intrin.h, xopintrin.h, abmintrin.h, bmiintrin.h, tbmintrin.h,
9 lwpintrin.h, fmaintrin.h and mm_malloc.h that reference the proper
10 builtin functions.
12 Defining away "extern" and "__inline" results in all of them being compiled
13 as proper functions. */
15 #define extern
16 #define __inline
18 #include <x86intrin.h>
20 #define _CONCAT(x,y) x ## y
22 #define test_0(func, type, imm) \
23 type _CONCAT(_,func) (int const I) \
24 { return func (imm); }
26 #define test_1(func, type, op1_type, imm) \
27 type _CONCAT(_,func) (op1_type A, int const I) \
28 { return func (A, imm); }
30 #define test_1x(func, type, op1_type, imm1, imm2) \
31 type _CONCAT(_,func) (op1_type A, int const I, int const L) \
32 { return func (A, imm1, imm2); }
34 #define test_1y(func, type, op1_type, imm1, imm2, imm3) \
35 type _CONCAT(_,func) (op1_type A, int const I, int const L, int const R)\
36 { return func (A, imm1, imm2, imm3); }
38 #define test_2(func, type, op1_type, op2_type, imm) \
39 type _CONCAT(_,func) (op1_type A, op2_type B, int const I) \
40 { return func (A, B, imm); }
42 #define test_2x(func, type, op1_type, op2_type, imm1, imm2) \
43 type _CONCAT(_,func) (op1_type A, op2_type B, int const I, int const L) \
44 { return func (A, B, imm1, imm2); }
46 #define test_2y(func, type, op1_type, op2_type, imm1, imm2, imm3) \
47 type _CONCAT(_,func) (op1_type A, op2_type B, int const I, int const L,\
48 int const R) \
49 { return func (A, B, imm1, imm2, imm3); }
51 #define test_2vx(func, op1_type, op2_type, imm1, imm2) \
52 _CONCAT(_,func) (op1_type A, op2_type B, int const I, int const L) \
53 { func (A, B, imm1, imm2); }
55 #define test_3(func, type, op1_type, op2_type, op3_type, imm) \
56 type _CONCAT(_,func) (op1_type A, op2_type B, \
57 op3_type C, int const I) \
58 { return func (A, B, C, imm); }
60 #define test_3x(func, type, op1_type, op2_type, op3_type, imm1, imm2) \
61 type _CONCAT(_,func) (op1_type A, op2_type B, \
62 op3_type C, int const I, int const L) \
63 { return func (A, B, C, imm1, imm2); }
65 #define test_3y(func, type, op1_type, op2_type, op3_type, imm1, imm2, imm3) \
66 type _CONCAT(_,func) (op1_type A, op2_type B, \
67 op3_type C, int const I, int const L, int const R) \
68 { return func (A, B, C, imm1, imm2, imm3); }
70 #define test_3v(func, op1_type, op2_type, op3_type, imm) \
71 _CONCAT(_,func) (op1_type A, op2_type B, \
72 op3_type C, int const I) \
73 { func (A, B, C, imm); }
75 #define test_3vx(func, op1_type, op2_type, op3_type, imm1, imm2) \
76 _CONCAT(_,func) (op1_type A, op2_type B, \
77 op3_type C, int const I, int const L) \
78 { func (A, B, C, imm1, imm2); }
80 #define test_4(func, type, op1_type, op2_type, op3_type, op4_type, imm) \
81 type _CONCAT(_,func) (op1_type A, op2_type B, \
82 op3_type C, op4_type D, int const I) \
83 { return func (A, B, C, D, imm); }
85 #define test_4x(func, type, op1_type, op2_type, op3_type, op4_type, imm1, imm2) \
86 type _CONCAT(_,func) (op1_type A, op2_type B, \
87 op3_type C, op4_type D, int const I, int const L) \
88 { return func (A, B, C, D, imm1, imm2); }
90 #define test_4y(func, type, op1_type, op2_type, op3_type, op4_type, imm1, imm2, imm3) \
91 type _CONCAT(_,func) (op1_type A, op2_type B, op3_type C, \
92 op4_type D, int const I, int const L, int const R) \
93 { return func (A, B, C, D, imm1, imm2, imm3); }
95 #define test_4v(func, op1_type, op2_type, op3_type, op4_type, imm) \
96 _CONCAT(_,func) (op1_type A, op2_type B, \
97 op3_type C, op4_type D, int const I) \
98 { func (A, B, C, D, imm); }
101 /* Following intrinsics require immediate arguments. They
102 are defined as macros for non-optimized compilations. */
104 /* ammintrin.h */
105 test_1x (_mm_extracti_si64, __m128i, __m128i, 1, 1)
106 test_2x (_mm_inserti_si64, __m128i, __m128i, __m128i, 1, 1)
108 /* immintrin.h */
109 test_2 (_mm256_blend_pd, __m256d, __m256d, __m256d, 1)
110 test_2 (_mm256_blend_ps, __m256, __m256, __m256, 1)
111 test_2 (_mm256_dp_ps, __m256, __m256, __m256, 1)
112 test_2 (_mm256_shuffle_pd, __m256d, __m256d, __m256d, 1)
113 test_2 (_mm256_shuffle_ps, __m256, __m256, __m256, 1)
114 test_2 (_mm_cmp_sd, __m128d, __m128d, __m128d, 1)
115 test_2 (_mm_cmp_ss, __m128, __m128, __m128, 1)
116 test_2 (_mm_cmp_pd, __m128d, __m128d, __m128d, 1)
117 test_2 (_mm_cmp_ps, __m128, __m128, __m128, 1)
118 test_2 (_mm256_cmp_pd, __m256d, __m256d, __m256d, 1)
119 test_2 (_mm256_cmp_ps, __m256, __m256, __m256, 1)
120 test_1 (_mm256_extractf128_pd, __m128d, __m256d, 1)
121 test_1 (_mm256_extractf128_ps, __m128, __m256, 1)
122 test_1 (_mm256_extractf128_si256, __m128i, __m256i, 1)
123 test_1 (_mm256_extract_epi8, int, __m256i, 20)
124 test_1 (_mm256_extract_epi16, int, __m256i, 13)
125 test_1 (_mm256_extract_epi32, int, __m256i, 6)
126 #ifdef __x86_64__
127 test_1 (_mm256_extract_epi64, long long, __m256i, 2)
128 #endif
129 test_1 (_mm_permute_pd, __m128d, __m128d, 1)
130 test_1 (_mm256_permute_pd, __m256d, __m256d, 1)
131 test_1 (_mm_permute_ps, __m128, __m128, 1)
132 test_1 (_mm256_permute_ps, __m256, __m256, 1)
133 test_2 (_mm256_permute2f128_pd, __m256d, __m256d, __m256d, 1)
134 test_2 (_mm256_permute2f128_ps, __m256, __m256, __m256, 1)
135 test_2 (_mm256_permute2f128_si256, __m256i, __m256i, __m256i, 1)
136 test_2 (_mm256_insertf128_pd, __m256d, __m256d, __m128d, 1)
137 test_2 (_mm256_insertf128_ps, __m256, __m256, __m128, 1)
138 test_2 (_mm256_insertf128_si256, __m256i, __m256i, __m128i, 1)
139 test_2 (_mm256_insert_epi8, __m256i, __m256i, int, 30)
140 test_2 (_mm256_insert_epi16, __m256i, __m256i, int, 7)
141 test_2 (_mm256_insert_epi32, __m256i, __m256i, int, 3)
142 #ifdef __x86_64__
143 test_2 (_mm256_insert_epi64, __m256i, __m256i, long long, 1)
144 #endif
145 test_1 (_mm256_round_pd, __m256d, __m256d, 9)
146 test_1 (_mm256_round_ps, __m256, __m256, 9)
147 test_1 (_cvtss_sh, unsigned short, float, 1)
148 test_1 (_mm_cvtps_ph, __m128i, __m128, 1)
149 test_1 (_mm256_cvtps_ph, __m128i, __m256, 1)
150 test_0 (_xabort, void, 1)
151 test_1 (_mm512_cvt_roundepi32_ps, __m512, __m512i, 9)
152 test_1 (_mm512_cvt_roundepu32_ps, __m512, __m512i, 9)
153 test_1 (_mm512_cvt_roundpd_epi32, __m256i, __m512d, 9)
154 test_1 (_mm512_cvt_roundpd_epu32, __m256i, __m512d, 9)
155 test_1 (_mm512_cvt_roundpd_ps, __m256, __m512d, 9)
156 test_1 (_mm512_cvt_roundph_ps, __m512, __m256i, 8)
157 test_1 (_mm512_cvt_roundps_epi32, __m512i, __m512, 9)
158 test_1 (_mm512_cvt_roundps_epu32, __m512i, __m512, 9)
159 test_1 (_mm512_cvt_roundps_pd, __m512d, __m256, 8)
160 test_1 (_mm512_cvtps_ph, __m256i, __m512, 1)
161 test_1 (_mm512_cvtt_roundpd_epi32, __m256i, __m512d, 8)
162 test_1 (_mm512_cvtt_roundpd_epu32, __m256i, __m512d, 8)
163 test_1 (_mm512_cvtt_roundps_epi32, __m512i, __m512, 8)
164 test_1 (_mm512_cvtt_roundps_epu32, __m512i, __m512, 8)
165 test_1 (_mm512_extractf32x4_ps, __m128, __m512, 1)
166 test_1 (_mm512_extractf64x4_pd, __m256d, __m512d, 1)
167 test_1 (_mm512_extracti32x4_epi32, __m128i, __m512i, 1)
168 test_1 (_mm512_extracti64x4_epi64, __m256i, __m512i, 1)
169 test_1 (_mm512_getexp_round_pd, __m512d, __m512d, 8)
170 test_1 (_mm512_getexp_round_ps, __m512, __m512, 8)
171 test_1y (_mm512_getmant_round_pd, __m512d, __m512d, 1, 1, 8)
172 test_1y (_mm512_getmant_round_ps, __m512, __m512, 1, 1, 8)
173 test_1 (_mm512_permute_pd, __m512d, __m512d, 1)
174 test_1 (_mm512_permute_ps, __m512, __m512, 1)
175 test_1 (_mm512_permutex_epi64, __m512i, __m512i, 1)
176 test_1 (_mm512_permutex_pd, __m512d, __m512d, 1)
177 test_1 (_mm512_rol_epi32, __m512i, __m512i, 1)
178 test_1 (_mm512_rol_epi64, __m512i, __m512i, 1)
179 test_1 (_mm512_ror_epi32, __m512i, __m512i, 1)
180 test_1 (_mm512_ror_epi64, __m512i, __m512i, 1)
181 test_1 (_mm512_shuffle_epi32, __m512i, __m512i, 1)
182 test_1 (_mm512_slli_epi32, __m512i, __m512i, 1)
183 test_1 (_mm512_slli_epi64, __m512i, __m512i, 1)
184 test_1 (_mm512_sqrt_round_pd, __m512d, __m512d, 9)
185 test_1 (_mm512_sqrt_round_ps, __m512, __m512, 9)
186 test_1 (_mm512_srai_epi32, __m512i, __m512i, 1)
187 test_1 (_mm512_srai_epi64, __m512i, __m512i, 1)
188 test_1 (_mm512_srli_epi32, __m512i, __m512i, 1)
189 test_1 (_mm512_srli_epi64, __m512i, __m512i, 1)
190 test_1 (_mm_cvt_roundsd_i32, int, __m128d, 9)
191 test_1 (_mm_cvt_roundsd_u32, unsigned, __m128d, 9)
192 test_1 (_mm_cvt_roundss_i32, int, __m128, 9)
193 test_1 (_mm_cvt_roundss_u32, unsigned, __m128, 9)
194 test_1 (_mm_cvtt_roundsd_i32, int, __m128d, 8)
195 test_1 (_mm_cvtt_roundsd_u32, unsigned, __m128d, 8)
196 test_1 (_mm_cvtt_roundss_i32, int, __m128, 8)
197 test_1 (_mm_cvtt_roundss_u32, unsigned, __m128, 8)
198 test_1x (_mm512_getmant_pd, __m512d, __m512d, 1, 1)
199 test_1x (_mm512_getmant_ps, __m512, __m512, 1, 1)
200 test_1x (_mm512_roundscale_round_pd, __m512d, __m512d, 1, 8)
201 test_1x (_mm512_roundscale_round_ps, __m512, __m512, 1, 8)
202 test_1x (_mm_cvt_roundi32_ss, __m128, __m128, 1, 9)
203 test_2 (_mm512_add_round_pd, __m512d, __m512d, __m512d, 9)
204 test_2 (_mm512_add_round_ps, __m512, __m512, __m512, 9)
205 test_2 (_mm512_alignr_epi32, __m512i, __m512i, __m512i, 1)
206 test_2 (_mm512_alignr_epi64, __m512i, __m512i, __m512i, 1)
207 test_2 (_mm512_cmp_epi32_mask, __mmask16, __m512i, __m512i, 1)
208 test_2 (_mm512_cmp_epi64_mask, __mmask8, __m512i, __m512i, 1)
209 test_2 (_mm512_cmp_epu32_mask, __mmask16, __m512i, __m512i, 1)
210 test_2 (_mm512_cmp_epu64_mask, __mmask8, __m512i, __m512i, 1)
211 test_2 (_mm512_cmp_pd_mask, __mmask8, __m512d, __m512d, 1)
212 test_2 (_mm512_cmp_ps_mask, __mmask16, __m512, __m512, 1)
213 test_2 (_mm512_div_round_pd, __m512d, __m512d, __m512d, 9)
214 test_2 (_mm512_div_round_ps, __m512, __m512, __m512, 9)
215 test_2 (_mm512_i32gather_epi32, __m512i, __m512i, void const *, 1)
216 test_2 (_mm512_i32gather_epi64, __m512i, __m256i, void const *, 1)
217 test_2 (_mm512_i32gather_pd, __m512d, __m256i, void const *, 1)
218 test_2 (_mm512_i32gather_ps, __m512, __m512i, void const *, 1)
219 test_2 (_mm512_i64gather_epi32, __m256i, __m512i, void const *, 1)
220 test_2 (_mm512_i64gather_epi64, __m512i, __m512i, void const *, 1)
221 test_2 (_mm512_i64gather_pd, __m512d, __m512i, void const *, 1)
222 test_2 (_mm512_i64gather_ps, __m256, __m512i, void const *, 1)
223 test_2 (_mm512_insertf32x4, __m512, __m512, __m128, 1)
224 test_2 (_mm512_insertf64x4, __m512d, __m512d, __m256d, 1)
225 test_2 (_mm512_inserti32x4, __m512i, __m512i, __m128i, 1)
226 test_2 (_mm512_inserti64x4, __m512i, __m512i, __m256i, 1)
227 test_2 (_mm512_maskz_cvt_roundepi32_ps, __m512, __mmask16, __m512i, 9)
228 test_2 (_mm512_maskz_cvt_roundepu32_ps, __m512, __mmask16, __m512i, 9)
229 test_2 (_mm512_maskz_cvt_roundpd_epi32, __m256i, __mmask8, __m512d, 9)
230 test_2 (_mm512_maskz_cvt_roundpd_epu32, __m256i, __mmask8, __m512d, 9)
231 test_2 (_mm512_maskz_cvt_roundpd_ps, __m256, __mmask8, __m512d, 9)
232 test_2 (_mm512_maskz_cvt_roundph_ps, __m512, __mmask16, __m256i, 8)
233 test_2 (_mm512_maskz_cvt_roundps_epi32, __m512i, __mmask16, __m512, 9)
234 test_2 (_mm512_maskz_cvt_roundps_epu32, __m512i, __mmask16, __m512, 9)
235 test_2 (_mm512_maskz_cvt_roundps_pd, __m512d, __mmask8, __m256, 8)
236 test_2 (_mm512_maskz_cvtps_ph, __m256i, __mmask16, __m512, 1)
237 test_2 (_mm512_maskz_cvtt_roundpd_epi32, __m256i, __mmask8, __m512d, 8)
238 test_2 (_mm512_maskz_cvtt_roundpd_epu32, __m256i, __mmask8, __m512d, 8)
239 test_2 (_mm512_maskz_cvtt_roundps_epi32, __m512i, __mmask16, __m512, 8)
240 test_2 (_mm512_maskz_cvtt_roundps_epu32, __m512i, __mmask16, __m512, 8)
241 test_2 (_mm512_maskz_extractf32x4_ps, __m128, __mmask8, __m512, 1)
242 test_2 (_mm512_maskz_extractf64x4_pd, __m256d, __mmask8, __m512d, 1)
243 test_2 (_mm512_maskz_extracti32x4_epi32, __m128i, __mmask8, __m512i, 1)
244 test_2 (_mm512_maskz_extracti64x4_epi64, __m256i, __mmask8, __m512i, 1)
245 test_2 (_mm512_maskz_getexp_round_pd, __m512d, __mmask8, __m512d, 8)
246 test_2 (_mm512_maskz_getexp_round_ps, __m512, __mmask16, __m512, 8)
247 test_2y (_mm512_maskz_getmant_round_pd, __m512d, __mmask8, __m512d, 1, 1, 8)
248 test_2y (_mm512_maskz_getmant_round_ps, __m512, __mmask16, __m512, 1, 1, 8)
249 test_2 (_mm512_maskz_permute_pd, __m512d, __mmask8, __m512d, 1)
250 test_2 (_mm512_maskz_permute_ps, __m512, __mmask16, __m512, 1)
251 test_2 (_mm512_maskz_permutex_epi64, __m512i, __mmask8, __m512i, 1)
252 test_2 (_mm512_maskz_permutex_pd, __m512d, __mmask8, __m512d, 1)
253 test_2 (_mm512_maskz_rol_epi32, __m512i, __mmask16, __m512i, 1)
254 test_2 (_mm512_maskz_rol_epi64, __m512i, __mmask8, __m512i, 1)
255 test_2 (_mm512_maskz_ror_epi32, __m512i, __mmask16, __m512i, 1)
256 test_2 (_mm512_maskz_ror_epi64, __m512i, __mmask8, __m512i, 1)
257 test_2 (_mm512_maskz_shuffle_epi32, __m512i, __mmask16, __m512i, 1)
258 test_2 (_mm512_maskz_slli_epi32, __m512i, __mmask16, __m512i, 1)
259 test_2 (_mm512_maskz_slli_epi64, __m512i, __mmask8, __m512i, 1)
260 test_2 (_mm512_maskz_sqrt_round_pd, __m512d, __mmask8, __m512d, 9)
261 test_2 (_mm512_maskz_sqrt_round_ps, __m512, __mmask16, __m512, 9)
262 test_2 (_mm512_maskz_srai_epi32, __m512i, __mmask16, __m512i, 1)
263 test_2 (_mm512_maskz_srai_epi64, __m512i, __mmask8, __m512i, 1)
264 test_2 (_mm512_maskz_srli_epi32, __m512i, __mmask16, __m512i, 1)
265 test_2 (_mm512_maskz_srli_epi64, __m512i, __mmask8, __m512i, 1)
266 test_2 (_mm512_max_round_pd, __m512d, __m512d, __m512d, 8)
267 test_2 (_mm512_max_round_ps, __m512, __m512, __m512, 8)
268 test_2 (_mm512_min_round_pd, __m512d, __m512d, __m512d, 8)
269 test_2 (_mm512_min_round_ps, __m512, __m512, __m512, 8)
270 test_2 (_mm512_mul_round_pd, __m512d, __m512d, __m512d, 9)
271 test_2 (_mm512_mul_round_ps, __m512, __m512, __m512, 9)
272 test_2 (_mm512_scalef_round_pd, __m512d, __m512d, __m512d, 9)
273 test_2 (_mm512_scalef_round_ps, __m512, __m512, __m512, 9)
274 test_2 (_mm512_shuffle_f32x4, __m512, __m512, __m512, 1)
275 test_2 (_mm512_shuffle_f64x2, __m512d, __m512d, __m512d, 1)
276 test_2 (_mm512_shuffle_i32x4, __m512i, __m512i, __m512i, 1)
277 test_2 (_mm512_shuffle_i64x2, __m512i, __m512i, __m512i, 1)
278 test_2 (_mm512_shuffle_pd, __m512d, __m512d, __m512d, 1)
279 test_2 (_mm512_shuffle_ps, __m512, __m512, __m512, 1)
280 test_2 (_mm512_sub_round_pd, __m512d, __m512d, __m512d, 9)
281 test_2 (_mm512_sub_round_ps, __m512, __m512, __m512, 9)
282 test_2 (_mm_add_round_sd, __m128d, __m128d, __m128d, 9)
283 test_2 (_mm_add_round_ss, __m128, __m128, __m128, 9)
284 test_2 (_mm_cmp_sd_mask, __mmask8, __m128d, __m128d, 1)
285 test_2 (_mm_cmp_ss_mask, __mmask8, __m128, __m128, 1)
286 #ifdef __x86_64__
287 test_2 (_mm_cvt_roundi64_sd, __m128d, __m128d, long long, 9)
288 test_2 (_mm_cvt_roundi64_ss, __m128, __m128, long long, 9)
289 #endif
290 test_2 (_mm_cvt_roundsd_ss, __m128, __m128, __m128d, 9)
291 test_2 (_mm_cvt_roundss_sd, __m128d, __m128d, __m128, 8)
292 test_2 (_mm_cvt_roundu32_ss, __m128, __m128, unsigned, 9)
293 #ifdef __x86_64__
294 test_2 (_mm_cvt_roundu64_sd, __m128d, __m128d, unsigned long long, 9)
295 test_2 (_mm_cvt_roundu64_ss, __m128, __m128, unsigned long long, 9)
296 #endif
297 test_2 (_mm_div_round_sd, __m128d, __m128d, __m128d, 9)
298 test_2 (_mm_div_round_ss, __m128, __m128, __m128, 9)
299 test_2 (_mm_getexp_round_sd, __m128d, __m128d, __m128d, 8)
300 test_2 (_mm_getexp_round_ss, __m128, __m128, __m128, 8)
301 test_2y (_mm_getmant_round_sd, __m128d, __m128d, __m128d, 1, 1, 8)
302 test_2y (_mm_getmant_round_ss, __m128, __m128, __m128, 1, 1, 8)
303 test_2 (_mm_mul_round_sd, __m128d, __m128d, __m128d, 9)
304 test_2 (_mm_mul_round_ss, __m128, __m128, __m128, 9)
305 test_2 (_mm_scalef_round_sd, __m128d, __m128d, __m128d, 9)
306 test_2 (_mm_scalef_round_ss, __m128, __m128, __m128, 9)
307 test_2 (_mm_sqrt_round_sd, __m128d, __m128d, __m128d, 9)
308 test_2 (_mm_sqrt_round_ss, __m128, __m128, __m128, 9)
309 test_2 (_mm_sub_round_sd, __m128d, __m128d, __m128d, 9)
310 test_2 (_mm_sub_round_ss, __m128, __m128, __m128, 9)
311 test_2x (_mm512_cmp_round_pd_mask, __mmask8, __m512d, __m512d, 1, 8)
312 test_2x (_mm512_cmp_round_ps_mask, __mmask16, __m512, __m512, 1, 8)
313 test_2x (_mm512_maskz_roundscale_round_pd, __m512d, __mmask8, __m512d, 1, 8)
314 test_2x (_mm512_maskz_roundscale_round_ps, __m512, __mmask16, __m512, 1, 8)
315 test_2x (_mm_cmp_round_sd_mask, __mmask8, __m128d, __m128d, 1, 8)
316 test_2x (_mm_cmp_round_ss_mask, __mmask8, __m128, __m128, 1, 8)
317 test_2x (_mm_comi_round_sd, int, __m128d, __m128d, 1, 8)
318 test_2x (_mm_comi_round_ss, int, __m128, __m128, 1, 8)
319 test_2x (_mm_roundscale_round_sd, __m128d, __m128d, __m128d, 1, 8)
320 test_2x (_mm_roundscale_round_ss, __m128, __m128, __m128, 1, 8)
321 test_3 (_mm512_fmadd_round_pd, __m512d, __m512d, __m512d, __m512d, 9)
322 test_3 (_mm512_fmadd_round_ps, __m512, __m512, __m512, __m512, 9)
323 test_3 (_mm512_fmaddsub_round_pd, __m512d, __m512d, __m512d, __m512d, 9)
324 test_3 (_mm512_fmaddsub_round_ps, __m512, __m512, __m512, __m512, 9)
325 test_3 (_mm512_fmsub_round_pd, __m512d, __m512d, __m512d, __m512d, 9)
326 test_3 (_mm512_fmsub_round_ps, __m512, __m512, __m512, __m512, 9)
327 test_3 (_mm512_fmsubadd_round_pd, __m512d, __m512d, __m512d, __m512d, 9)
328 test_3 (_mm512_fmsubadd_round_ps, __m512, __m512, __m512, __m512, 9)
329 test_3 (_mm512_fnmadd_round_pd, __m512d, __m512d, __m512d, __m512d, 9)
330 test_3 (_mm512_fnmadd_round_ps, __m512, __m512, __m512, __m512, 9)
331 test_3 (_mm512_fnmsub_round_pd, __m512d, __m512d, __m512d, __m512d, 9)
332 test_3 (_mm512_fnmsub_round_ps, __m512, __m512, __m512, __m512, 9)
333 test_3 (_mm512_mask_cmp_epi32_mask, __mmask16, __mmask16, __m512i, __m512i, 1)
334 test_3 (_mm512_mask_cmp_epi64_mask, __mmask8, __mmask8, __m512i, __m512i, 1)
335 test_3 (_mm512_mask_cmp_epu32_mask, __mmask16, __mmask16, __m512i, __m512i, 1)
336 test_3 (_mm512_mask_cmp_epu64_mask, __mmask8, __mmask8, __m512i, __m512i, 1)
337 test_3 (_mm512_mask_cmp_pd_mask, __mmask8, __mmask8, __m512d, __m512d, 1)
338 test_3 (_mm512_mask_cmp_ps_mask, __mmask16, __mmask16, __m512, __m512, 1)
339 test_3 (_mm512_mask_cvt_roundepi32_ps, __m512, __m512, __mmask16, __m512i, 9)
340 test_3 (_mm512_mask_cvt_roundepu32_ps, __m512, __m512, __mmask16, __m512i, 9)
341 test_3 (_mm512_mask_cvt_roundpd_epi32, __m256i, __m256i, __mmask8, __m512d, 9)
342 test_3 (_mm512_mask_cvt_roundpd_epu32, __m256i, __m256i, __mmask8, __m512d, 9)
343 test_3 (_mm512_mask_cvt_roundpd_ps, __m256, __m256, __mmask8, __m512d, 9)
344 test_3 (_mm512_mask_cvt_roundph_ps, __m512, __m512, __mmask16, __m256i, 8)
345 test_3 (_mm512_mask_cvt_roundps_epi32, __m512i, __m512i, __mmask16, __m512, 9)
346 test_3 (_mm512_mask_cvt_roundps_epu32, __m512i, __m512i, __mmask16, __m512, 9)
347 test_3 (_mm512_mask_cvt_roundps_pd, __m512d, __m512d, __mmask8, __m256, 8)
348 test_3 (_mm512_mask_cvtps_ph, __m256i, __m256i, __mmask16, __m512, 1)
349 test_3 (_mm512_mask_cvtt_roundpd_epi32, __m256i, __m256i, __mmask8, __m512d, 8)
350 test_3 (_mm512_mask_cvtt_roundpd_epu32, __m256i, __m256i, __mmask8, __m512d, 8)
351 test_3 (_mm512_mask_cvtt_roundps_epi32, __m512i, __m512i, __mmask16, __m512, 8)
352 test_3 (_mm512_mask_cvtt_roundps_epu32, __m512i, __m512i, __mmask16, __m512, 8)
353 test_3 (_mm512_mask_extractf32x4_ps, __m128, __m128, __mmask8, __m512, 1)
354 test_3 (_mm512_mask_extractf64x4_pd, __m256d, __m256d, __mmask8, __m512d, 1)
355 test_3 (_mm512_mask_extracti32x4_epi32, __m128i, __m128i, __mmask8, __m512i, 1)
356 test_3 (_mm512_mask_extracti64x4_epi64, __m256i, __m256i, __mmask8, __m512i, 1)
357 test_3 (_mm512_mask_getexp_round_pd, __m512d, __m512d, __mmask8, __m512d, 8)
358 test_3 (_mm512_mask_getexp_round_ps, __m512, __m512, __mmask16, __m512, 8)
359 test_3y (_mm512_mask_getmant_round_pd, __m512d, __m512d, __mmask8, __m512d, 1, 1, 8)
360 test_3y (_mm512_mask_getmant_round_ps, __m512, __m512, __mmask16, __m512, 1, 1, 8)
361 test_3 (_mm512_mask_permute_pd, __m512d, __m512d, __mmask8, __m512d, 1)
362 test_3 (_mm512_mask_permute_ps, __m512, __m512, __mmask16, __m512, 1)
363 test_3 (_mm512_mask_permutex_epi64, __m512i, __m512i, __mmask8, __m512i, 1)
364 test_3 (_mm512_mask_permutex_pd, __m512d, __m512d, __mmask8, __m512d, 1)
365 test_3 (_mm512_mask_rol_epi32, __m512i, __m512i, __mmask16, __m512i, 1)
366 test_3 (_mm512_mask_rol_epi64, __m512i, __m512i, __mmask8, __m512i, 1)
367 test_3 (_mm512_mask_ror_epi32, __m512i, __m512i, __mmask16, __m512i, 1)
368 test_3 (_mm512_mask_ror_epi64, __m512i, __m512i, __mmask8, __m512i, 1)
369 test_3 (_mm512_mask_shuffle_epi32, __m512i, __m512i, __mmask16, __m512i, 1)
370 test_3 (_mm512_mask_slli_epi32, __m512i, __m512i, __mmask16, __m512i, 1)
371 test_3 (_mm512_mask_slli_epi64, __m512i, __m512i, __mmask8, __m512i, 1)
372 test_3 (_mm512_mask_sqrt_round_pd, __m512d, __m512d, __mmask8, __m512d, 9)
373 test_3 (_mm512_mask_sqrt_round_ps, __m512, __m512, __mmask16, __m512, 9)
374 test_3 (_mm512_mask_srai_epi32, __m512i, __m512i, __mmask16, __m512i, 1)
375 test_3 (_mm512_mask_srai_epi64, __m512i, __m512i, __mmask8, __m512i, 1)
376 test_3 (_mm512_mask_srli_epi32, __m512i, __m512i, __mmask16, __m512i, 1)
377 test_3 (_mm512_mask_srli_epi64, __m512i, __m512i, __mmask8, __m512i, 1)
378 test_3 (_mm512_maskz_add_round_pd, __m512d, __mmask8, __m512d, __m512d, 9)
379 test_3 (_mm512_maskz_add_round_ps, __m512, __mmask16, __m512, __m512, 9)
380 test_3 (_mm512_maskz_alignr_epi32, __m512i, __mmask16, __m512i, __m512i, 1)
381 test_3 (_mm512_maskz_alignr_epi64, __m512i, __mmask8, __m512i, __m512i, 1)
382 test_3 (_mm512_maskz_div_round_pd, __m512d, __mmask8, __m512d, __m512d, 9)
383 test_3 (_mm512_maskz_div_round_ps, __m512, __mmask16, __m512, __m512, 9)
384 test_3 (_mm512_maskz_insertf32x4, __m512, __mmask16, __m512, __m128, 1)
385 test_3 (_mm512_maskz_insertf64x4, __m512d, __mmask8, __m512d, __m256d, 1)
386 test_3 (_mm512_maskz_inserti32x4, __m512i, __mmask16, __m512i, __m128i, 1)
387 test_3 (_mm512_maskz_inserti64x4, __m512i, __mmask8, __m512i, __m256i, 1)
388 test_3 (_mm512_maskz_max_round_pd, __m512d, __mmask8, __m512d, __m512d, 8)
389 test_3 (_mm512_maskz_max_round_ps, __m512, __mmask16, __m512, __m512, 8)
390 test_3 (_mm512_maskz_min_round_pd, __m512d, __mmask8, __m512d, __m512d, 8)
391 test_3 (_mm512_maskz_min_round_ps, __m512, __mmask16, __m512, __m512, 8)
392 test_3 (_mm512_maskz_mul_round_pd, __m512d, __mmask8, __m512d, __m512d, 9)
393 test_3 (_mm512_maskz_mul_round_ps, __m512, __mmask16, __m512, __m512, 9)
394 test_3 (_mm512_maskz_scalef_round_pd, __m512d, __mmask8, __m512d, __m512d, 9)
395 test_3 (_mm512_maskz_scalef_round_ps, __m512, __mmask16, __m512, __m512, 9)
396 test_3 (_mm512_maskz_shuffle_f32x4, __m512, __mmask16, __m512, __m512, 1)
397 test_3 (_mm512_maskz_shuffle_f64x2, __m512d, __mmask8, __m512d, __m512d, 1)
398 test_3 (_mm512_maskz_shuffle_i32x4, __m512i, __mmask16, __m512i, __m512i, 1)
399 test_3 (_mm512_maskz_shuffle_i64x2, __m512i, __mmask8, __m512i, __m512i, 1)
400 test_3 (_mm512_maskz_shuffle_pd, __m512d, __mmask8, __m512d, __m512d, 1)
401 test_3 (_mm512_maskz_shuffle_ps, __m512, __mmask16, __m512, __m512, 1)
402 test_3 (_mm512_maskz_sub_round_pd, __m512d, __mmask8, __m512d, __m512d, 9)
403 test_3 (_mm512_maskz_sub_round_ps, __m512, __mmask16, __m512, __m512, 9)
404 test_3 (_mm512_ternarylogic_epi32, __m512i, __m512i, __m512i, __m512i, 1)
405 test_3 (_mm512_ternarylogic_epi64, __m512i, __m512i, __m512i, __m512i, 1)
406 test_3 (_mm_fmadd_round_sd, __m128d, __m128d, __m128d, __m128d, 9)
407 test_3 (_mm_fmadd_round_ss, __m128, __m128, __m128, __m128, 9)
408 test_3 (_mm_fmsub_round_sd, __m128d, __m128d, __m128d, __m128d, 9)
409 test_3 (_mm_fmsub_round_ss, __m128, __m128, __m128, __m128, 9)
410 test_3 (_mm_fnmadd_round_sd, __m128d, __m128d, __m128d, __m128d, 9)
411 test_3 (_mm_fnmadd_round_ss, __m128, __m128, __m128, __m128, 9)
412 test_3 (_mm_fnmsub_round_sd, __m128d, __m128d, __m128d, __m128d, 9)
413 test_3 (_mm_fnmsub_round_ss, __m128, __m128, __m128, __m128, 9)
414 test_3 (_mm_mask_cmp_sd_mask, __mmask8, __mmask8, __m128d, __m128d, 1)
415 test_3 (_mm_mask_cmp_ss_mask, __mmask8, __mmask8, __m128, __m128, 1)
416 test_3v (_mm512_i32scatter_epi32, void *, __m512i, __m512i, 1)
417 test_3v (_mm512_i32scatter_epi64, void *, __m256i, __m512i, 1)
418 test_3v (_mm512_i32scatter_pd, void *, __m256i, __m512d, 1)
419 test_3v (_mm512_i32scatter_ps, void *, __m512i, __m512, 1)
420 test_3v (_mm512_i64scatter_epi32, void *, __m512i, __m256i, 1)
421 test_3v (_mm512_i64scatter_epi64, void *, __m512i, __m512i, 1)
422 test_3v (_mm512_i64scatter_pd, void *, __m512i, __m512d, 1)
423 test_3v (_mm512_i64scatter_ps, void *, __m512i, __m256, 1)
424 test_3x (_mm512_mask_roundscale_round_pd, __m512d, __m512d, __mmask8, __m512d, 1, 8)
425 test_3x (_mm512_mask_roundscale_round_ps, __m512, __m512, __mmask16, __m512, 1, 8)
426 test_3x (_mm_fixupimm_round_sd, __m128d, __m128d, __m128d, __m128i, 1, 8)
427 test_3x (_mm_fixupimm_round_ss, __m128, __m128, __m128, __m128i, 1, 8)
428 test_3x (_mm_mask_cmp_round_sd_mask, __mmask8, __mmask8, __m128d, __m128d, 1, 8)
429 test_3x (_mm_mask_cmp_round_ss_mask, __mmask8, __mmask8, __m128, __m128, 1, 8)
430 test_4 (_mm512_mask3_fmadd_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9)
431 test_4 (_mm512_mask3_fmadd_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9)
432 test_4 (_mm512_mask3_fmaddsub_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9)
433 test_4 (_mm512_mask3_fmaddsub_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9)
434 test_4 (_mm512_mask3_fmsub_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9)
435 test_4 (_mm512_mask3_fmsub_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9)
436 test_4 (_mm512_mask3_fmsubadd_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9)
437 test_4 (_mm512_mask3_fmsubadd_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9)
438 test_4 (_mm512_mask3_fnmadd_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9)
439 test_4 (_mm512_mask3_fnmadd_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9)
440 test_4 (_mm512_mask3_fnmsub_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9)
441 test_4 (_mm512_mask3_fnmsub_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9)
442 test_4 (_mm512_mask_add_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
443 test_4 (_mm512_mask_add_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
444 test_4 (_mm512_mask_alignr_epi32, __m512i, __m512i, __mmask16, __m512i, __m512i, 1)
445 test_4 (_mm512_mask_alignr_epi64, __m512i, __m512i, __mmask8, __m512i, __m512i, 1)
446 test_4 (_mm512_mask_div_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
447 test_4 (_mm512_mask_div_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
448 test_4 (_mm512_mask_fmadd_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
449 test_4 (_mm512_mask_fmadd_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
450 test_4 (_mm512_mask_fmaddsub_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
451 test_4 (_mm512_mask_fmaddsub_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
452 test_4 (_mm512_mask_fmsub_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
453 test_4 (_mm512_mask_fmsub_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
454 test_4 (_mm512_mask_fmsubadd_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
455 test_4 (_mm512_mask_fmsubadd_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
456 test_4 (_mm512_mask_fnmadd_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
457 test_4 (_mm512_mask_fnmadd_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
458 test_4 (_mm512_mask_fnmsub_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
459 test_4 (_mm512_mask_fnmsub_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
460 test_4 (_mm512_mask_i32gather_epi32, __m512i, __m512i, __mmask16, __m512i, void const *, 1)
461 test_4 (_mm512_mask_i32gather_epi64, __m512i, __m512i, __mmask8, __m256i, void const *, 1)
462 test_4 (_mm512_mask_i32gather_pd, __m512d, __m512d, __mmask8, __m256i, void const *, 1)
463 test_4 (_mm512_mask_i32gather_ps, __m512, __m512, __mmask16, __m512i, void const *, 1)
464 test_4 (_mm512_mask_i64gather_epi32, __m256i, __m256i, __mmask8, __m512i, void const *, 1)
465 test_4 (_mm512_mask_i64gather_epi64, __m512i, __m512i, __mmask8, __m512i, void const *, 1)
466 test_4 (_mm512_mask_i64gather_pd, __m512d, __m512d, __mmask8, __m512i, void const *, 1)
467 test_4 (_mm512_mask_i64gather_ps, __m256, __m256, __mmask8, __m512i, void const *, 1)
468 test_4 (_mm512_mask_insertf32x4, __m512, __m512, __mmask16, __m512, __m128, 1)
469 test_4 (_mm512_mask_insertf64x4, __m512d, __m512d, __mmask8, __m512d, __m256d, 1)
470 test_4 (_mm512_mask_inserti32x4, __m512i, __m512i, __mmask16, __m512i, __m128i, 1)
471 test_4 (_mm512_mask_inserti64x4, __m512i, __m512i, __mmask8, __m512i, __m256i, 1)
472 test_4 (_mm512_mask_max_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 8)
473 test_4 (_mm512_mask_max_round_ps, __m512, __m512, __mmask16, __m512, __m512, 8)
474 test_4 (_mm512_mask_min_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 8)
475 test_4 (_mm512_mask_min_round_ps, __m512, __m512, __mmask16, __m512, __m512, 8)
476 test_4 (_mm512_mask_mul_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
477 test_4 (_mm512_mask_mul_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
478 test_4 (_mm512_mask_scalef_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
479 test_4 (_mm512_mask_scalef_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
480 test_4 (_mm512_mask_shuffle_f32x4, __m512, __m512, __mmask16, __m512, __m512, 1)
481 test_4 (_mm512_mask_shuffle_f64x2, __m512d, __m512d, __mmask8, __m512d, __m512d, 1)
482 test_4 (_mm512_mask_shuffle_i32x4, __m512i, __m512i, __mmask16, __m512i, __m512i, 1)
483 test_4 (_mm512_mask_shuffle_i64x2, __m512i, __m512i, __mmask8, __m512i, __m512i, 1)
484 test_4 (_mm512_mask_shuffle_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 1)
485 test_4 (_mm512_mask_shuffle_ps, __m512, __m512, __mmask16, __m512, __m512, 1)
486 test_4 (_mm512_mask_sub_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
487 test_4 (_mm512_mask_sub_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
488 test_4 (_mm512_mask_ternarylogic_epi32, __m512i, __m512i, __mmask16, __m512i, __m512i, 1)
489 test_4 (_mm512_mask_ternarylogic_epi64, __m512i, __m512i, __mmask8, __m512i, __m512i, 1)
490 test_4 (_mm512_maskz_fmadd_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9)
491 test_4 (_mm512_maskz_fmadd_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9)
492 test_4 (_mm512_maskz_fmaddsub_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9)
493 test_4 (_mm512_maskz_fmaddsub_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9)
494 test_4 (_mm512_maskz_fmsub_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9)
495 test_4 (_mm512_maskz_fmsub_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9)
496 test_4 (_mm512_maskz_fmsubadd_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9)
497 test_4 (_mm512_maskz_fmsubadd_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9)
498 test_4 (_mm512_maskz_fnmadd_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9)
499 test_4 (_mm512_maskz_fnmadd_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9)
500 test_4 (_mm512_maskz_fnmsub_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9)
501 test_4 (_mm512_maskz_fnmsub_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9)
502 test_4 (_mm512_maskz_ternarylogic_epi32, __m512i, __mmask16, __m512i, __m512i, __m512i, 1)
503 test_4 (_mm512_maskz_ternarylogic_epi64, __m512i, __mmask8, __m512i, __m512i, __m512i, 1)
504 test_4v (_mm512_mask_i32scatter_epi32, void *, __mmask16, __m512i, __m512i, 1)
505 test_4v (_mm512_mask_i32scatter_epi64, void *, __mmask8, __m256i, __m512i, 1)
506 test_4v (_mm512_mask_i32scatter_pd, void *, __mmask8, __m256i, __m512d, 1)
507 test_4v (_mm512_mask_i32scatter_ps, void *, __mmask16, __m512i, __m512, 1)
508 test_4v (_mm512_mask_i64scatter_epi32, void *, __mmask8, __m512i, __m256i, 1)
509 test_4v (_mm512_mask_i64scatter_epi64, void *, __mmask8, __m512i, __m512i, 1)
510 test_4v (_mm512_mask_i64scatter_pd, void *, __mmask8, __m512i, __m512d, 1)
511 test_4v (_mm512_mask_i64scatter_ps, void *, __mmask8, __m512i, __m256, 1)
512 test_4x (_mm512_mask_fixupimm_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512i, 1, 8)
513 test_4x (_mm512_mask_fixupimm_round_ps, __m512, __m512, __mmask16, __m512, __m512i, 1, 8)
514 test_4x (_mm512_maskz_fixupimm_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512i, 1, 8)
515 test_4x (_mm512_maskz_fixupimm_round_ps, __m512, __mmask16, __m512, __m512, __m512i, 1, 8)
516 test_4x (_mm_mask_fixupimm_round_sd, __m128d, __m128d, __mmask8, __m128d, __m128i, 1, 8)
517 test_4x (_mm_mask_fixupimm_round_ss, __m128, __m128, __mmask8, __m128, __m128i, 1, 8)
518 test_4x (_mm_maskz_fixupimm_round_sd, __m128d, __mmask8, __m128d, __m128d, __m128i, 1, 8)
519 test_4x (_mm_maskz_fixupimm_round_ss, __m128, __mmask8, __m128, __m128, __m128i, 1, 8)
521 /* avx512pfintrin.h */
522 test_3vx (_mm512_mask_prefetch_i32gather_ps, __m512i, __mmask16, void const *, 1, _MM_HINT_T0)
523 test_3vx (_mm512_mask_prefetch_i32scatter_ps, void const *, __mmask16, __m512i, 1, _MM_HINT_T0)
524 test_3vx (_mm512_mask_prefetch_i64gather_ps, __m512i, __mmask8, void const *, 1, _MM_HINT_T0)
525 test_3vx (_mm512_mask_prefetch_i64scatter_ps, void const *, __mmask8, __m512i, 1, _MM_HINT_T0)
526 test_3vx (_mm512_mask_prefetch_i32gather_pd, __m256i, __mmask8, void const *, 1, _MM_HINT_T0)
527 test_3vx (_mm512_mask_prefetch_i32scatter_pd, void const *, __mmask8, __m256i, 1, _MM_HINT_T0)
528 test_3vx (_mm512_mask_prefetch_i64gather_pd, __m512i, __mmask8, void const *, 1, _MM_HINT_T0)
529 test_3vx (_mm512_mask_prefetch_i64scatter_pd, void const *, __mmask8, __m512i, 1, _MM_HINT_T0)
531 /* avx512erintrin.h */
532 test_1 (_mm512_exp2a23_round_pd, __m512d, __m512d, 8)
533 test_1 (_mm512_exp2a23_round_ps, __m512, __m512, 8)
534 test_1 (_mm512_rcp28_round_pd, __m512d, __m512d, 8)
535 test_1 (_mm512_rcp28_round_ps, __m512, __m512, 8)
536 test_1 (_mm512_rsqrt28_round_pd, __m512d, __m512d, 8)
537 test_1 (_mm512_rsqrt28_round_ps, __m512, __m512, 8)
538 test_2 (_mm512_maskz_exp2a23_round_pd, __m512d, __mmask8, __m512d, 8)
539 test_2 (_mm512_maskz_exp2a23_round_ps, __m512, __mmask16, __m512, 8)
540 test_2 (_mm512_maskz_rcp28_round_pd, __m512d, __mmask8, __m512d, 8)
541 test_2 (_mm512_maskz_rcp28_round_ps, __m512, __mmask16, __m512, 8)
542 test_2 (_mm512_maskz_rsqrt28_round_pd, __m512d, __mmask8, __m512d, 8)
543 test_2 (_mm512_maskz_rsqrt28_round_ps, __m512, __mmask16, __m512, 8)
544 test_3 (_mm512_mask_exp2a23_round_pd, __m512d, __m512d, __mmask8, __m512d, 8)
545 test_3 (_mm512_mask_exp2a23_round_ps, __m512, __m512, __mmask16, __m512, 8)
546 test_3 (_mm512_mask_rcp28_round_pd, __m512d, __m512d, __mmask8, __m512d, 8)
547 test_3 (_mm512_mask_rcp28_round_ps, __m512, __m512, __mmask16, __m512, 8)
548 test_3 (_mm512_mask_rsqrt28_round_pd, __m512d, __m512d, __mmask8, __m512d, 8)
549 test_3 (_mm512_mask_rsqrt28_round_ps, __m512, __m512, __mmask16, __m512, 8)
551 /* shaintrin.h */
552 test_2 (_mm_sha1rnds4_epu32, __m128i, __m128i, __m128i, 1)
554 /* wmmintrin.h */
555 test_1 (_mm_aeskeygenassist_si128, __m128i, __m128i, 1)
556 test_2 (_mm_clmulepi64_si128, __m128i, __m128i, __m128i, 1)
558 /* smmintrin.h */
559 test_1 (_mm_round_pd, __m128d, __m128d, 9)
560 test_1 (_mm_round_ps, __m128, __m128, 9)
561 test_2 (_mm_round_sd, __m128d, __m128d, __m128d, 9)
562 test_2 (_mm_round_ss, __m128, __m128, __m128, 9)
564 test_2 (_mm_blend_epi16, __m128i, __m128i, __m128i, 1)
565 test_2 (_mm_blend_ps, __m128, __m128, __m128, 1)
566 test_2 (_mm_blend_pd, __m128d, __m128d, __m128d, 1)
567 test_2 (_mm_dp_ps, __m128, __m128, __m128, 1)
568 test_2 (_mm_dp_pd, __m128d, __m128d, __m128d, 1)
569 test_2 (_mm_insert_ps, __m128, __m128, __m128, 1)
570 test_1 (_mm_extract_ps, int, __m128, 1)
571 test_2 (_mm_insert_epi8, __m128i, __m128i, int, 1)
572 test_2 (_mm_insert_epi32, __m128i, __m128i, int, 1)
573 #ifdef __x86_64__
574 test_2 (_mm_insert_epi64, __m128i, __m128i, long long, 1)
575 #endif
576 test_1 (_mm_extract_epi8, int, __m128i, 1)
577 test_1 (_mm_extract_epi32, int, __m128i, 1)
578 #ifdef __x86_64__
579 test_1 (_mm_extract_epi64, long long, __m128i, 1)
580 #endif
581 test_2 (_mm_mpsadbw_epu8, __m128i, __m128i, __m128i, 1)
582 test_2 (_mm_cmpistrm, __m128i, __m128i, __m128i, 1)
583 test_2 (_mm_cmpistri, int, __m128i, __m128i, 1)
584 test_4 (_mm_cmpestrm, __m128i, __m128i, int, __m128i, int, 1)
585 test_4 (_mm_cmpestri, int, __m128i, int, __m128i, int, 1)
586 test_2 (_mm_cmpistra, int, __m128i, __m128i, 1)
587 test_2 (_mm_cmpistrc, int, __m128i, __m128i, 1)
588 test_2 (_mm_cmpistro, int, __m128i, __m128i, 1)
589 test_2 (_mm_cmpistrs, int, __m128i, __m128i, 1)
590 test_2 (_mm_cmpistrz, int, __m128i, __m128i, 1)
591 test_4 (_mm_cmpestra, int, __m128i, int, __m128i, int, 1)
592 test_4 (_mm_cmpestrc, int, __m128i, int, __m128i, int, 1)
593 test_4 (_mm_cmpestro, int, __m128i, int, __m128i, int, 1)
594 test_4 (_mm_cmpestrs, int, __m128i, int, __m128i, int, 1)
595 test_4 (_mm_cmpestrz, int, __m128i, int, __m128i, int, 1)
597 /* tmmintrin.h */
598 test_2 (_mm_alignr_epi8, __m128i, __m128i, __m128i, 1)
599 test_2 (_mm_alignr_pi8, __m64, __m64, __m64, 1)
601 /* emmintrin.h */
602 test_2 (_mm_shuffle_pd, __m128d, __m128d, __m128d, 1)
603 test_1 (_mm_srli_si128, __m128i, __m128i, 1)
604 test_1 (_mm_slli_si128, __m128i, __m128i, 1)
605 test_1 (_mm_extract_epi16, int, __m128i, 1)
606 test_2 (_mm_insert_epi16, __m128i, __m128i, int, 1)
607 test_1 (_mm_shufflehi_epi16, __m128i, __m128i, 1)
608 test_1 (_mm_shufflelo_epi16, __m128i, __m128i, 1)
609 test_1 (_mm_shuffle_epi32, __m128i, __m128i, 1)
611 /* xmmintrin.h */
612 test_2 (_mm_shuffle_ps, __m128, __m128, __m128, 1)
613 test_1 (_mm_extract_pi16, int, __m64, 1)
614 test_1 (_m_pextrw, int, __m64, 1)
615 test_2 (_mm_insert_pi16, __m64, __m64, int, 1)
616 test_2 (_m_pinsrw, __m64, __m64, int, 1)
617 test_1 (_mm_shuffle_pi16, __m64, __m64, 1)
618 test_1 (_m_pshufw, __m64, __m64, 1)
619 test_1 (_mm_prefetch, void, void *, _MM_HINT_NTA)
621 /* xopintrin.h */
622 test_1 ( _mm_roti_epi8, __m128i, __m128i, 1)
623 test_1 ( _mm_roti_epi16, __m128i, __m128i, 1)
624 test_1 ( _mm_roti_epi32, __m128i, __m128i, 1)
625 test_1 ( _mm_roti_epi64, __m128i, __m128i, 1)
626 test_3 (_mm_permute2_pd, __m128d, __m128d, __m128d, __m128d, 1)
627 test_3 (_mm256_permute2_pd, __m256d, __m256d, __m256d, __m256d, 1)
628 test_3 (_mm_permute2_ps, __m128, __m128, __m128, __m128, 1)
629 test_3 (_mm256_permute2_ps, __m256, __m256, __m256, __m256, 1)
631 /* lwpintrin.h */
632 test_2 ( __lwpval32, void, unsigned int, unsigned int, 1)
633 test_2 ( __lwpins32, unsigned char, unsigned int, unsigned int, 1)
634 #ifdef __x86_64__
635 test_2 ( __lwpval64, void, unsigned long long, unsigned int, 1)
636 test_2 ( __lwpins64, unsigned char, unsigned long long, unsigned int, 1)
637 #endif
639 /* tbmintrin.h */
640 test_1 ( __bextri_u32, unsigned int, unsigned int, 1)
641 #ifdef __x86_64__
642 test_1 ( __bextri_u64, unsigned long long, unsigned long long, 1)
643 #endif