1 /* Same as sse-14, except converted to use #pragma GCC option.  */
2 /* { dg-do compile } */
3 /* { dg-options "-O0 -Werror-implicit-function-declaration -march=k8" } */
4 /* { dg-add-options bind_pic_locally } */
5 
6 #include <mm_malloc.h>
7 
8 /* Test that the intrinsics compile with optimization.  All of them
9    are defined as inline functions in {,x,e,p,t,s,w,a,b,i}mmintrin.h,
10    mm3dnow.h, fma4intrin.h, xopintrin.h, abmintrin.h, bmiintrin.h,
11    tbmintrin.h, lwpintrin.h, popcntintrin.h, fmaintrin.h and mm_malloc.h
12    that reference the proper builtin functions.
13 
14    Defining away "extern" and "__inline" results in all of them being
15    compiled as proper functions.  */
16 
17 #define extern
18 #define __inline
19 
20 #define _CONCAT(x,y) x ## y
21 
22 #define test_0(func, type, imm)						\
23   type _CONCAT(_,func) (int const I)					\
24   { return func (imm); }
25 
26 #define test_1(func, type, op1_type, imm)				\
27   type _CONCAT(_,func) (op1_type A, int const I)			\
28   { return func (A, imm); }
29 
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); }
33 
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); }
37 
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); }
41 
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); }
45 
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); }
50 
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); }
54 
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); }
59 
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); }
64 
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); }
69 
70 #define test_3v(func, op1_type, op2_type, op3_type, imm)		\
71   int _CONCAT(_,func) (op1_type A, op2_type B,				\
72 		       op3_type C, int const I)				\
73   { func (A, B, C, imm); }
74 
75 #define test_3vx(func, op1_type, op2_type, op3_type, imm1, imm2)   \
76   int _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); }
79 
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); }
84 
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); }
89 
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); }
94 
95 
96 #define test_4v(func, op1_type, op2_type, op3_type, op4_type, imm)	\
97   int _CONCAT(_,func) (op1_type A, op2_type B,				\
98 		       op3_type C, op4_type D, int const I)		\
99   { func (A, B, C, D, imm); }
100 
101 
102 #ifndef DIFFERENT_PRAGMAS
103 #pragma GCC target ("sse4a,3dnow,avx,avx2,fma4,xop,aes,pclmul,popcnt,abm,lzcnt,bmi,bmi2,tbm,lwp,fsgsbase,rdrnd,f16c,rtm,rdseed,prfchw,adx,fxsr,xsaveopt,avx512f,avx512er,avx512cd,avx512pf,sha,prefetchwt1,avx512vl,avx512bw,avx512dq,avx512vbmi,avx512ifma")
104 #endif
105 
106 /* Following intrinsics require immediate arguments.  They
107    are defined as macros for non-optimized compilations. */
108 
109 /* mmintrin.h (MMX).  */
110 #ifdef DIFFERENT_PRAGMAS
111 #pragma GCC target ("mmx")
112 #endif
113 #include <mmintrin.h>
114 
115 /* mm3dnow.h (3DNOW).  */
116 #ifdef DIFFERENT_PRAGMAS
117 #pragma GCC target ("3dnow")
118 #endif
119 #include <mm3dnow.h>
120 
121 /* xmmintrin.h (SSE).  */
122 #ifdef DIFFERENT_PRAGMAS
123 #pragma GCC target ("sse")
124 #endif
125 #include <xmmintrin.h>
126 test_2 (_mm_shuffle_ps, __m128, __m128, __m128, 1)
127 test_1 (_mm_extract_pi16, int, __m64, 1)
128 test_1 (_m_pextrw, int, __m64, 1)
129 test_2 (_mm_insert_pi16, __m64, __m64, int, 1)
130 test_2 (_m_pinsrw, __m64, __m64, int, 1)
131 test_1 (_mm_shuffle_pi16, __m64, __m64, 1)
132 test_1 (_m_pshufw, __m64, __m64, 1)
133 test_1 (_mm_prefetch, void, void *, _MM_HINT_NTA)
134 
135 /* emmintrin.h (SSE2).  */
136 #ifdef DIFFERENT_PRAGMAS
137 #pragma GCC target ("sse2")
138 #endif
139 #include <emmintrin.h>
140 test_2 (_mm_shuffle_pd, __m128d, __m128d, __m128d, 1)
141 test_1 (_mm_bsrli_si128, __m128i, __m128i, 1)
142 test_1 (_mm_bslli_si128, __m128i, __m128i, 1)
143 test_1 (_mm_srli_si128, __m128i, __m128i, 1)
144 test_1 (_mm_slli_si128, __m128i, __m128i, 1)
145 test_1 (_mm_extract_epi16, int, __m128i, 1)
146 test_2 (_mm_insert_epi16, __m128i, __m128i, int, 1)
147 test_1 (_mm_shufflehi_epi16, __m128i, __m128i, 1)
148 test_1 (_mm_shufflelo_epi16, __m128i, __m128i, 1)
149 test_1 (_mm_shuffle_epi32, __m128i, __m128i, 1)
150 
151 /* pmmintrin.h (SSE3).  */
152 #ifdef DIFFERENT_PRAGMAS
153 #pragma GCC target ("sse3")
154 #endif
155 #include <pmmintrin.h>
156 
157 /* tmmintrin.h (SSSE3).  */
158 #ifdef DIFFERENT_PRAGMAS
159 #pragma GCC target ("ssse3")
160 #endif
161 #include <tmmintrin.h>
162 test_2 (_mm_alignr_epi8, __m128i, __m128i, __m128i, 1)
163 test_2 (_mm_alignr_pi8, __m64, __m64, __m64, 1)
164 
165 /* ammintrin.h (SSE4A).  */
166 #ifdef DIFFERENT_PRAGMAS
167 #pragma GCC target ("sse4a")
168 #endif
169 #include <ammintrin.h>
170 test_1x (_mm_extracti_si64, __m128i, __m128i, 1, 1)
171 test_2x (_mm_inserti_si64, __m128i, __m128i, __m128i, 1, 1)
172 
173 /* Note, nmmintrin.h includes smmintrin.h, and smmintrin.h
174    checks for the #ifdef.  So just set the option to SSE4.2.  */
175 #ifdef DIFFERENT_PRAGMAS
176 #pragma GCC target ("sse4.2")
177 #endif
178 #include <nmmintrin.h>
179 /* smmintrin.h (SSE4.2).  */
180 test_1 (_mm_round_pd, __m128d, __m128d, 9)
181 test_1 (_mm_round_ps, __m128, __m128, 9)
182 test_2 (_mm_round_sd, __m128d, __m128d, __m128d, 9)
183 test_2 (_mm_round_ss, __m128, __m128, __m128, 9)
184 
185 test_2 (_mm_blend_epi16, __m128i, __m128i, __m128i, 1)
186 test_2 (_mm_blend_ps, __m128, __m128, __m128, 1)
187 test_2 (_mm_blend_pd, __m128d, __m128d, __m128d, 1)
188 test_2 (_mm_dp_ps, __m128, __m128, __m128, 1)
189 test_2 (_mm_dp_pd, __m128d, __m128d, __m128d, 1)
190 test_2 (_mm_insert_ps, __m128, __m128, __m128, 1)
191 test_1 (_mm_extract_ps, int, __m128, 1)
192 test_2 (_mm_insert_epi8, __m128i, __m128i, int, 1)
193 test_2 (_mm_insert_epi32, __m128i, __m128i, int, 1)
194 #ifdef __x86_64__
195 test_2 (_mm_insert_epi64, __m128i, __m128i, long long, 1)
196 #endif
197 test_1 (_mm_extract_epi8, int, __m128i, 1)
198 test_1 (_mm_extract_epi32, int, __m128i, 1)
199 #ifdef __x86_64__
200 test_1 (_mm_extract_epi64, long long, __m128i, 1)
201 #endif
202 test_2 (_mm_mpsadbw_epu8, __m128i, __m128i, __m128i, 1)
203 test_2 (_mm_cmpistrm, __m128i, __m128i, __m128i, 1)
204 test_2 (_mm_cmpistri, int, __m128i, __m128i, 1)
205 test_4 (_mm_cmpestrm, __m128i, __m128i, int, __m128i, int, 1)
206 test_4 (_mm_cmpestri, int, __m128i, int, __m128i, int, 1)
207 test_2 (_mm_cmpistra, int, __m128i, __m128i, 1)
208 test_2 (_mm_cmpistrc, int, __m128i, __m128i, 1)
209 test_2 (_mm_cmpistro, int, __m128i, __m128i, 1)
210 test_2 (_mm_cmpistrs, int, __m128i, __m128i, 1)
211 test_2 (_mm_cmpistrz, int, __m128i, __m128i, 1)
212 test_4 (_mm_cmpestra, int, __m128i, int, __m128i, int, 1)
213 test_4 (_mm_cmpestrc, int, __m128i, int, __m128i, int, 1)
214 test_4 (_mm_cmpestro, int, __m128i, int, __m128i, int, 1)
215 test_4 (_mm_cmpestrs, int, __m128i, int, __m128i, int, 1)
216 test_4 (_mm_cmpestrz, int, __m128i, int, __m128i, int, 1)
217 
218 /* immintrin.h (AVX/AVX2/RDRND/FSGSBASE/F16C/RTM/AVX512F/SHA) */
219 #ifdef DIFFERENT_PRAGMAS
220 #pragma GCC target ("avx,avx2,rdrnd,fsgsbase,f16c,rtm,avx512f,avx512er,avx512cd,avx512pf,sha,avx512vl,avx512bw,avx512dq,avx512ifma,avx512vbmi")
221 #endif
222 #include <immintrin.h>
223 test_1 (_cvtss_sh, unsigned short, float, 1)
224 test_1 (_mm_cvtps_ph, __m128i, __m128, 1)
225 test_1 (_mm256_cvtps_ph, __m128i, __m256, 1)
226 
227 /* avxintrin.h */
228 test_2 (_mm256_blend_pd, __m256d, __m256d, __m256d, 1)
229 test_2 (_mm256_blend_ps, __m256, __m256, __m256, 1)
230 test_2 (_mm256_dp_ps, __m256, __m256, __m256, 1)
231 test_2 (_mm256_shuffle_pd, __m256d, __m256d, __m256d, 1)
232 test_2 (_mm256_shuffle_ps, __m256, __m256, __m256, 1)
233 test_2 (_mm_cmp_sd, __m128d, __m128d, __m128d, 1)
234 test_2 (_mm_cmp_ss, __m128, __m128, __m128, 1)
235 test_2 (_mm_cmp_pd, __m128d, __m128d, __m128d, 1)
236 test_2 (_mm_cmp_ps, __m128, __m128, __m128, 1)
237 test_2 (_mm256_cmp_pd, __m256d, __m256d, __m256d, 1)
238 test_2 (_mm256_cmp_ps, __m256, __m256, __m256, 1)
239 test_1 (_mm256_extractf128_pd, __m128d, __m256d, 1)
240 test_1 (_mm256_extractf128_ps, __m128, __m256, 1)
241 test_1 (_mm256_extractf128_si256, __m128i, __m256i, 1)
242 test_1 (_mm256_extract_epi8, int, __m256i, 20)
243 test_1 (_mm256_extract_epi16, int, __m256i, 13)
244 test_1 (_mm256_extract_epi32, int, __m256i, 6)
245 #ifdef __x86_64__
246 test_1 (_mm256_extract_epi64, long long, __m256i, 2)
247 #endif
248 test_1 (_mm_permute_pd, __m128d, __m128d, 1)
249 test_1 (_mm256_permute_pd, __m256d, __m256d, 1)
250 test_1 (_mm_permute_ps, __m128, __m128, 1)
251 test_1 (_mm256_permute_ps, __m256, __m256, 1)
252 test_2 (_mm256_permute2f128_pd, __m256d, __m256d, __m256d, 1)
253 test_2 (_mm256_permute2f128_ps, __m256, __m256, __m256, 1)
254 test_2 (_mm256_permute2f128_si256, __m256i, __m256i, __m256i, 1)
255 test_2 (_mm256_insertf128_pd, __m256d, __m256d, __m128d, 1)
256 test_2 (_mm256_insertf128_ps, __m256, __m256, __m128, 1)
257 test_2 (_mm256_insertf128_si256, __m256i, __m256i, __m128i, 1)
258 test_2 (_mm256_insert_epi8, __m256i, __m256i, int, 30)
259 test_2 (_mm256_insert_epi16, __m256i, __m256i, int, 7)
260 test_2 (_mm256_insert_epi32, __m256i, __m256i, int, 3)
261 #ifdef __x86_64__
262 test_2 (_mm256_insert_epi64, __m256i, __m256i, long long, 1)
263 #endif
264 test_1 (_mm256_round_pd, __m256d, __m256d, 9)
265 test_1 (_mm256_round_ps, __m256, __m256, 9)
266 
267 /* avx2intrin.h */
268 test_2 ( _mm256_mpsadbw_epu8, __m256i, __m256i, __m256i, 1)
269 test_2 ( _mm256_alignr_epi8, __m256i, __m256i, __m256i, 1)
270 test_2 ( _mm256_blend_epi16, __m256i, __m256i, __m256i, 1)
271 test_1 ( _mm256_shuffle_epi32, __m256i, __m256i, 1)
272 test_1 ( _mm256_shufflehi_epi16, __m256i, __m256i, 1)
273 test_1 ( _mm256_shufflelo_epi16, __m256i, __m256i, 1)
274 test_1 ( _mm256_bslli_epi128, __m256i, __m256i, 8)
275 test_1 ( _mm256_bsrli_epi128, __m256i, __m256i, 8)
276 test_1 ( _mm256_slli_si256, __m256i, __m256i, 8)
277 test_1 ( _mm256_srli_si256, __m256i, __m256i, 8)
278 test_2 ( _mm_blend_epi32, __m128i, __m128i, __m128i, 1)
279 test_2 ( _mm256_blend_epi32, __m256i, __m256i, __m256, 1)
280 test_1 ( _mm256_permute4x64_pd, __m256d, __m256d, 1)
281 test_1 ( _mm256_permute4x64_epi64, __m256i, __m256i, 1)
282 test_2 ( _mm256_permute2x128_si256, __m256i, __m256i, __m256i, 1)
283 test_1 ( _mm256_extracti128_si256, __m128i, __m256i, 1)
284 test_2 ( _mm256_inserti128_si256, __m256i, __m256i, __m128i, 1)
285 test_2 ( _mm_i32gather_pd, __m128d, double const *, __m128i, 1)
286 test_2 ( _mm256_i32gather_pd, __m256d, double const *, __m128i, 1)
287 test_2 ( _mm_i64gather_pd, __m128d, double const *, __m128i, 1)
288 test_2 ( _mm256_i64gather_pd, __m256d, double const *, __m256i, 1)
289 test_2 ( _mm_i32gather_ps, __m128, float const *, __m128i, 1)
290 test_2 ( _mm256_i32gather_ps, __m256, float const *, __m256i, 1)
291 test_2 ( _mm_i64gather_ps, __m128, float const *, __m128i, 1)
292 test_2 ( _mm256_i64gather_ps, __m128, float const *, __m256i, 1)
293 test_2 ( _mm_i32gather_epi64, __m128i, long long int const *, __m128i, 1)
294 test_2 ( _mm256_i32gather_epi64, __m256i, long long int const *, __m128i, 1)
295 test_2 ( _mm_i64gather_epi64, __m128i, long long int const *, __m128i, 1)
296 test_2 ( _mm256_i64gather_epi64,  __m256i, long long int const *, __m256i, 1)
297 test_2 ( _mm_i32gather_epi32, __m128i, int const *, __m128i, 1)
298 test_2 ( _mm256_i32gather_epi32, __m256i, int const *, __m256i, 1)
299 test_2 ( _mm_i64gather_epi32, __m128i, int const *, __m128i, 1)
300 test_2 ( _mm256_i64gather_epi32, __m128i, int const *, __m256i, 1)
301 
302 /* rtmintrin.h */
303 test_0 ( _xabort, void, 1)
304 
305 /* avx512fintrin.h */
306 test_1 (_mm512_cvt_roundepi32_ps, __m512, __m512i, 9)
307 test_1 (_mm512_cvt_roundepu32_ps, __m512, __m512i, 9)
308 test_1 (_mm512_cvt_roundpd_epi32, __m256i, __m512d, 9)
309 test_1 (_mm512_cvt_roundpd_epu32, __m256i, __m512d, 9)
310 test_1 (_mm512_cvt_roundpd_ps, __m256, __m512d, 9)
311 test_1 (_mm512_cvt_roundph_ps, __m512, __m256i, 8)
312 test_1 (_mm512_cvt_roundps_epi32, __m512i, __m512, 9)
313 test_1 (_mm512_cvt_roundps_epu32, __m512i, __m512, 9)
314 test_1 (_mm512_cvt_roundps_pd, __m512d, __m256, 8)
315 test_1 (_mm512_cvtps_ph, __m256i, __m512, 1)
316 test_1 (_mm512_cvtt_roundpd_epi32, __m256i, __m512d, 8)
317 test_1 (_mm512_cvtt_roundpd_epu32, __m256i, __m512d, 8)
318 test_1 (_mm512_cvtt_roundps_epi32, __m512i, __m512, 8)
319 test_1 (_mm512_cvtt_roundps_epu32, __m512i, __m512, 8)
320 test_1 (_mm512_extractf32x4_ps, __m128, __m512, 1)
321 test_1 (_mm512_extractf64x4_pd, __m256d, __m512d, 1)
322 test_1 (_mm512_extracti32x4_epi32, __m128i, __m512i, 1)
323 test_1 (_mm512_extracti64x4_epi64, __m256i, __m512i, 1)
324 test_1 (_mm512_getexp_round_pd, __m512d, __m512d, 8)
325 test_1 (_mm512_getexp_round_ps, __m512, __m512, 8)
326 test_1y (_mm512_getmant_round_pd, __m512d, __m512d, 1, 1, 8)
327 test_1y (_mm512_getmant_round_ps, __m512, __m512, 1, 1, 8)
328 test_1 (_mm512_permute_pd, __m512d, __m512d, 1)
329 test_1 (_mm512_permute_ps, __m512, __m512, 1)
330 test_1 (_mm512_permutex_epi64, __m512i, __m512i, 1)
331 test_1 (_mm512_permutex_pd, __m512d, __m512d, 1)
332 test_1 (_mm512_rol_epi32, __m512i, __m512i, 1)
333 test_1 (_mm512_rol_epi64, __m512i, __m512i, 1)
334 test_1 (_mm512_ror_epi32, __m512i, __m512i, 1)
335 test_1 (_mm512_ror_epi64, __m512i, __m512i, 1)
336 test_1 (_mm512_shuffle_epi32, __m512i, __m512i, 1)
337 test_1 (_mm512_slli_epi32, __m512i, __m512i, 1)
338 test_1 (_mm512_slli_epi64, __m512i, __m512i, 1)
339 test_1 (_mm512_sqrt_round_pd, __m512d, __m512d, 9)
340 test_1 (_mm512_sqrt_round_ps, __m512, __m512, 9)
341 test_1 (_mm512_srai_epi32, __m512i, __m512i, 1)
342 test_1 (_mm512_srai_epi64, __m512i, __m512i, 1)
343 test_1 (_mm512_srli_epi32, __m512i, __m512i, 1)
344 test_1 (_mm512_srli_epi64, __m512i, __m512i, 1)
345 test_1 (_mm_cvt_roundsd_i32, int, __m128d, 9)
346 test_1 (_mm_cvt_roundsd_u32, unsigned, __m128d, 9)
347 test_1 (_mm_cvt_roundss_i32, int, __m128, 9)
348 test_1 (_mm_cvt_roundss_u32, unsigned, __m128, 9)
349 test_1 (_mm_cvtt_roundsd_i32, int, __m128d, 8)
350 test_1 (_mm_cvtt_roundsd_u32, unsigned, __m128d, 8)
351 test_1 (_mm_cvtt_roundss_i32, int, __m128, 8)
352 test_1 (_mm_cvtt_roundss_u32, unsigned, __m128, 8)
353 test_1x (_mm512_getmant_pd, __m512d, __m512d, 1, 1)
354 test_1x (_mm512_getmant_ps, __m512, __m512, 1, 1)
355 test_1x (_mm_cvt_roundi32_ss, __m128, __m128, 1, 9)
356 test_2 (_mm512_add_round_pd, __m512d, __m512d, __m512d, 9)
357 test_2 (_mm512_add_round_ps, __m512, __m512, __m512, 9)
358 test_2 (_mm512_alignr_epi32, __m512i, __m512i, __m512i, 1)
359 test_2 (_mm512_alignr_epi64, __m512i, __m512i, __m512i, 1)
360 test_2 (_mm512_cmp_epi32_mask, __mmask16, __m512i, __m512i, 1)
361 test_2 (_mm512_cmp_epi64_mask, __mmask8, __m512i, __m512i, 1)
362 test_2 (_mm512_cmp_epu32_mask, __mmask16, __m512i, __m512i, 1)
363 test_2 (_mm512_cmp_epu64_mask, __mmask8, __m512i, __m512i, 1)
364 test_2 (_mm512_cmp_pd_mask, __mmask8, __m512d, __m512d, 1)
365 test_2 (_mm512_cmp_ps_mask, __mmask16, __m512, __m512, 1)
366 test_2 (_mm512_div_round_pd, __m512d, __m512d, __m512d, 9)
367 test_2 (_mm512_div_round_ps, __m512, __m512, __m512, 9)
368 test_2 (_mm512_i32gather_epi32, __m512i, __m512i, void const *, 1)
369 test_2 (_mm512_i32gather_epi64, __m512i, __m256i, void const *, 1)
370 test_2 (_mm512_i32gather_pd, __m512d, __m256i, void const *, 1)
371 test_2 (_mm512_i32gather_ps, __m512, __m512i, void const *, 1)
372 test_2 (_mm512_i64gather_epi32, __m256i, __m512i, void const *, 1)
373 test_2 (_mm512_i64gather_epi64, __m512i, __m512i, void const *, 1)
374 test_2 (_mm512_i64gather_pd, __m512d, __m512i, void const *, 1)
375 test_2 (_mm512_i64gather_ps, __m256, __m512i, void const *, 1)
376 test_2 (_mm512_insertf32x4, __m512, __m512, __m128, 1)
377 test_2 (_mm512_insertf64x4, __m512d, __m512d, __m256d, 1)
378 test_2 (_mm512_inserti32x4, __m512i, __m512i, __m128i, 1)
379 test_2 (_mm512_inserti64x4, __m512i, __m512i, __m256i, 1)
380 test_2 (_mm512_maskz_cvt_roundepi32_ps, __m512, __mmask16, __m512i, 9)
381 test_2 (_mm512_maskz_cvt_roundepu32_ps, __m512, __mmask16, __m512i, 9)
382 test_2 (_mm512_maskz_cvt_roundpd_epi32, __m256i, __mmask8, __m512d, 9)
383 test_2 (_mm512_maskz_cvt_roundpd_epu32, __m256i, __mmask8, __m512d, 9)
384 test_2 (_mm512_maskz_cvt_roundpd_ps, __m256, __mmask8, __m512d, 9)
385 test_2 (_mm512_maskz_cvt_roundph_ps, __m512, __mmask16, __m256i, 8)
386 test_2 (_mm512_maskz_cvt_roundps_epi32, __m512i, __mmask16, __m512, 9)
387 test_2 (_mm512_maskz_cvt_roundps_epu32, __m512i, __mmask16, __m512, 9)
388 test_2 (_mm512_maskz_cvt_roundps_pd, __m512d, __mmask8, __m256, 8)
389 test_2 (_mm512_maskz_cvtps_ph, __m256i, __mmask16, __m512, 1)
390 test_2 (_mm512_maskz_cvtt_roundpd_epi32, __m256i, __mmask8, __m512d, 8)
391 test_2 (_mm512_maskz_cvtt_roundpd_epu32, __m256i, __mmask8, __m512d, 8)
392 test_2 (_mm512_maskz_cvtt_roundps_epi32, __m512i, __mmask16, __m512, 8)
393 test_2 (_mm512_maskz_cvtt_roundps_epu32, __m512i, __mmask16, __m512, 8)
394 test_2 (_mm512_maskz_extractf32x4_ps, __m128, __mmask8, __m512, 1)
395 test_2 (_mm512_maskz_extractf64x4_pd, __m256d, __mmask8, __m512d, 1)
396 test_2 (_mm512_maskz_extracti32x4_epi32, __m128i, __mmask8, __m512i, 1)
397 test_2 (_mm512_maskz_extracti64x4_epi64, __m256i, __mmask8, __m512i, 1)
398 test_2 (_mm512_maskz_getexp_round_pd, __m512d, __mmask8, __m512d, 8)
399 test_2 (_mm512_maskz_getexp_round_ps, __m512, __mmask16, __m512, 8)
400 test_2y (_mm512_maskz_getmant_round_pd, __m512d, __mmask8, __m512d, 1, 1, 8)
401 test_2y (_mm512_maskz_getmant_round_ps, __m512, __mmask16, __m512, 1, 1, 8)
402 test_2 (_mm512_maskz_permute_pd, __m512d, __mmask8, __m512d, 1)
403 test_2 (_mm512_maskz_permute_ps, __m512, __mmask16, __m512, 1)
404 test_2 (_mm512_maskz_permutex_epi64, __m512i, __mmask8, __m512i, 1)
405 test_2 (_mm512_maskz_permutex_pd, __m512d, __mmask8, __m512d, 1)
406 test_2 (_mm512_maskz_rol_epi32, __m512i, __mmask16, __m512i, 1)
407 test_2 (_mm512_maskz_rol_epi64, __m512i, __mmask8, __m512i, 1)
408 test_2 (_mm512_maskz_ror_epi32, __m512i, __mmask16, __m512i, 1)
409 test_2 (_mm512_maskz_ror_epi64, __m512i, __mmask8, __m512i, 1)
410 test_2 (_mm512_maskz_shuffle_epi32, __m512i, __mmask16, __m512i, 1)
411 test_2 (_mm512_maskz_slli_epi32, __m512i, __mmask16, __m512i, 1)
412 test_2 (_mm512_maskz_slli_epi64, __m512i, __mmask8, __m512i, 1)
413 test_2 (_mm512_maskz_sqrt_round_pd, __m512d, __mmask8, __m512d, 9)
414 test_2 (_mm512_maskz_sqrt_round_ps, __m512, __mmask16, __m512, 9)
415 test_2 (_mm512_maskz_srai_epi32, __m512i, __mmask16, __m512i, 1)
416 test_2 (_mm512_maskz_srai_epi64, __m512i, __mmask8, __m512i, 1)
417 test_2 (_mm512_maskz_srli_epi32, __m512i, __mmask16, __m512i, 1)
418 test_2 (_mm512_maskz_srli_epi64, __m512i, __mmask8, __m512i, 1)
419 test_2 (_mm512_max_round_pd, __m512d, __m512d, __m512d, 8)
420 test_2 (_mm512_max_round_ps, __m512, __m512, __m512, 8)
421 test_2 (_mm512_min_round_pd, __m512d, __m512d, __m512d, 8)
422 test_2 (_mm512_min_round_ps, __m512, __m512, __m512, 8)
423 test_2 (_mm512_mul_round_pd, __m512d, __m512d, __m512d, 9)
424 test_2 (_mm512_mul_round_ps, __m512, __m512, __m512, 9)
425 test_2 (_mm512_scalef_round_pd, __m512d, __m512d, __m512d, 9)
426 test_2 (_mm512_scalef_round_ps, __m512, __m512, __m512, 9)
427 test_2 (_mm512_shuffle_f32x4, __m512, __m512, __m512, 1)
428 test_2 (_mm512_shuffle_f64x2, __m512d, __m512d, __m512d, 1)
429 test_2 (_mm512_shuffle_i32x4, __m512i, __m512i, __m512i, 1)
430 test_2 (_mm512_shuffle_i64x2, __m512i, __m512i, __m512i, 1)
431 test_2 (_mm512_shuffle_pd, __m512d, __m512d, __m512d, 1)
432 test_2 (_mm512_shuffle_ps, __m512, __m512, __m512, 1)
433 test_2 (_mm512_sub_round_pd, __m512d, __m512d, __m512d, 9)
434 test_2 (_mm512_sub_round_ps, __m512, __m512, __m512, 9)
435 test_2 (_mm_cmp_sd_mask, __mmask8, __m128d, __m128d, 1)
436 test_2 (_mm_cmp_ss_mask, __mmask8, __m128, __m128, 1)
437 #ifdef __x86_64__
438 test_2 (_mm_cvt_roundi64_sd, __m128d, __m128d, long long, 9)
439 test_2 (_mm_cvt_roundi64_ss, __m128, __m128, long long, 9)
440 #endif
441 test_2 (_mm_cvt_roundu32_ss, __m128, __m128, unsigned, 9)
442 #ifdef __x86_64__
443 test_2 (_mm_cvt_roundu64_sd, __m128d, __m128d, unsigned long long, 9)
444 test_2 (_mm_cvt_roundu64_ss, __m128, __m128, unsigned long long, 9)
445 #endif
446 test_2x (_mm512_cmp_round_pd_mask, __mmask8, __m512d, __m512d, 1, 8)
447 test_2x (_mm512_cmp_round_ps_mask, __mmask16, __m512, __m512, 1, 8)
448 test_2x (_mm512_maskz_roundscale_round_pd, __m512d, __mmask8, __m512d, 1, 8)
449 test_2x (_mm512_maskz_roundscale_round_ps, __m512, __mmask16, __m512, 1, 8)
450 test_2x (_mm_cmp_round_sd_mask, __mmask8, __m128d, __m128d, 1, 8)
451 test_2x (_mm_cmp_round_ss_mask, __mmask8, __m128, __m128, 1, 8)
452 test_2x (_mm_comi_round_sd, int, __m128d, __m128d, 1, 8)
453 test_2x (_mm_comi_round_ss, int, __m128, __m128, 1, 8)
454 test_3 (_mm512_fmadd_round_pd, __m512d, __m512d, __m512d, __m512d, 9)
455 test_3 (_mm512_fmadd_round_ps, __m512, __m512, __m512, __m512, 9)
456 test_3 (_mm512_fmaddsub_round_pd, __m512d, __m512d, __m512d, __m512d, 9)
457 test_3 (_mm512_fmaddsub_round_ps, __m512, __m512, __m512, __m512, 9)
458 test_3 (_mm512_fmsub_round_pd, __m512d, __m512d, __m512d, __m512d, 9)
459 test_3 (_mm512_fmsub_round_ps, __m512, __m512, __m512, __m512, 9)
460 test_3 (_mm512_fmsubadd_round_pd, __m512d, __m512d, __m512d, __m512d, 9)
461 test_3 (_mm512_fmsubadd_round_ps, __m512, __m512, __m512, __m512, 9)
462 test_3 (_mm512_fnmadd_round_pd, __m512d, __m512d, __m512d, __m512d, 9)
463 test_3 (_mm512_fnmadd_round_ps, __m512, __m512, __m512, __m512, 9)
464 test_3 (_mm512_fnmsub_round_pd, __m512d, __m512d, __m512d, __m512d, 9)
465 test_3 (_mm512_fnmsub_round_ps, __m512, __m512, __m512, __m512, 9)
466 test_3 (_mm512_mask_cmp_epi32_mask, __mmask16, __mmask16, __m512i, __m512i, 1)
467 test_3 (_mm512_mask_cmp_epi64_mask, __mmask8, __mmask8, __m512i, __m512i, 1)
468 test_3 (_mm512_mask_cmp_epu32_mask, __mmask16, __mmask16, __m512i, __m512i, 1)
469 test_3 (_mm512_mask_cmp_epu64_mask, __mmask8, __mmask8, __m512i, __m512i, 1)
470 test_3 (_mm512_mask_cmp_pd_mask, __mmask8, __mmask8, __m512d, __m512d, 1)
471 test_3 (_mm512_mask_cmp_ps_mask, __mmask16, __mmask16, __m512, __m512, 1)
472 test_3 (_mm512_mask_cvt_roundepi32_ps, __m512, __m512, __mmask16, __m512i, 9)
473 test_3 (_mm512_mask_cvt_roundepu32_ps, __m512, __m512, __mmask16, __m512i, 9)
474 test_3 (_mm512_mask_cvt_roundpd_epi32, __m256i, __m256i, __mmask8, __m512d, 9)
475 test_3 (_mm512_mask_cvt_roundpd_epu32, __m256i, __m256i, __mmask8, __m512d, 9)
476 test_3 (_mm512_mask_cvt_roundpd_ps, __m256, __m256, __mmask8, __m512d, 9)
477 test_3 (_mm512_mask_cvt_roundph_ps, __m512, __m512, __mmask16, __m256i, 8)
478 test_3 (_mm512_mask_cvt_roundps_epi32, __m512i, __m512i, __mmask16, __m512, 9)
479 test_3 (_mm512_mask_cvt_roundps_epu32, __m512i, __m512i, __mmask16, __m512, 9)
480 test_3 (_mm512_mask_cvt_roundps_pd, __m512d, __m512d, __mmask8, __m256, 8)
481 test_3 (_mm512_mask_cvtps_ph, __m256i, __m256i, __mmask16, __m512, 1)
482 test_3 (_mm512_mask_cvtt_roundpd_epi32, __m256i, __m256i, __mmask8, __m512d, 8)
483 test_3 (_mm512_mask_cvtt_roundpd_epu32, __m256i, __m256i, __mmask8, __m512d, 8)
484 test_3 (_mm512_mask_cvtt_roundps_epi32, __m512i, __m512i, __mmask16, __m512, 8)
485 test_3 (_mm512_mask_cvtt_roundps_epu32, __m512i, __m512i, __mmask16, __m512, 8)
486 test_3 (_mm512_mask_extractf32x4_ps, __m128, __m128, __mmask8, __m512, 1)
487 test_3 (_mm512_mask_extractf64x4_pd, __m256d, __m256d, __mmask8, __m512d, 1)
488 test_3 (_mm512_mask_extracti32x4_epi32, __m128i, __m128i, __mmask8, __m512i, 1)
489 test_3 (_mm512_mask_extracti64x4_epi64, __m256i, __m256i, __mmask8, __m512i, 1)
490 test_3 (_mm512_mask_getexp_round_pd, __m512d, __m512d, __mmask8, __m512d, 8)
491 test_3 (_mm512_mask_getexp_round_ps, __m512, __m512, __mmask16, __m512, 8)
492 test_3y (_mm512_mask_getmant_round_pd, __m512d, __m512d, __mmask8, __m512d, 1, 1, 8)
493 test_3y (_mm512_mask_getmant_round_ps, __m512, __m512, __mmask16, __m512, 1, 1, 8)
494 test_3 (_mm512_mask_permute_pd, __m512d, __m512d, __mmask8, __m512d, 1)
495 test_3 (_mm512_mask_permute_ps, __m512, __m512, __mmask16, __m512, 1)
496 test_3 (_mm512_mask_permutex_epi64, __m512i, __m512i, __mmask8, __m512i, 1)
497 test_3 (_mm512_mask_permutex_pd, __m512d, __m512d, __mmask8, __m512d, 1)
498 test_3 (_mm512_mask_rol_epi32, __m512i, __m512i, __mmask16, __m512i, 1)
499 test_3 (_mm512_mask_rol_epi64, __m512i, __m512i, __mmask8, __m512i, 1)
500 test_3 (_mm512_mask_ror_epi32, __m512i, __m512i, __mmask16, __m512i, 1)
501 test_3 (_mm512_mask_ror_epi64, __m512i, __m512i, __mmask8, __m512i, 1)
502 test_3 (_mm512_mask_shuffle_epi32, __m512i, __m512i, __mmask16, __m512i, 1)
503 test_3 (_mm512_mask_slli_epi32, __m512i, __m512i, __mmask16, __m512i, 1)
504 test_3 (_mm512_mask_slli_epi64, __m512i, __m512i, __mmask8, __m512i, 1)
505 test_3 (_mm512_mask_sqrt_round_pd, __m512d, __m512d, __mmask8, __m512d, 9)
506 test_3 (_mm512_mask_sqrt_round_ps, __m512, __m512, __mmask16, __m512, 9)
507 test_3 (_mm512_mask_srai_epi32, __m512i, __m512i, __mmask16, __m512i, 1)
508 test_3 (_mm512_mask_srai_epi64, __m512i, __m512i, __mmask8, __m512i, 1)
509 test_3 (_mm512_mask_srli_epi32, __m512i, __m512i, __mmask16, __m512i, 1)
510 test_3 (_mm512_mask_srli_epi64, __m512i, __m512i, __mmask8, __m512i, 1)
511 test_3 (_mm512_maskz_add_round_pd, __m512d, __mmask8, __m512d, __m512d, 9)
512 test_3 (_mm512_maskz_add_round_ps, __m512, __mmask16, __m512, __m512, 9)
513 test_3 (_mm512_maskz_alignr_epi32, __m512i, __mmask16, __m512i, __m512i, 1)
514 test_3 (_mm512_maskz_alignr_epi64, __m512i, __mmask8, __m512i, __m512i, 1)
515 test_3 (_mm512_maskz_div_round_pd, __m512d, __mmask8, __m512d, __m512d, 9)
516 test_3 (_mm512_maskz_div_round_ps, __m512, __mmask16, __m512, __m512, 9)
517 test_3 (_mm512_maskz_insertf32x4, __m512, __mmask16, __m512, __m128, 1)
518 test_3 (_mm512_maskz_insertf64x4, __m512d, __mmask8, __m512d, __m256d, 1)
519 test_3 (_mm512_maskz_inserti32x4, __m512i, __mmask16, __m512i, __m128i, 1)
520 test_3 (_mm512_maskz_inserti64x4, __m512i, __mmask8, __m512i, __m256i, 1)
521 test_3 (_mm512_maskz_max_round_pd, __m512d, __mmask8, __m512d, __m512d, 8)
522 test_3 (_mm512_maskz_max_round_ps, __m512, __mmask16, __m512, __m512, 8)
523 test_3 (_mm512_maskz_min_round_pd, __m512d, __mmask8, __m512d, __m512d, 8)
524 test_3 (_mm512_maskz_min_round_ps, __m512, __mmask16, __m512, __m512, 8)
525 test_3 (_mm512_maskz_mul_round_pd, __m512d, __mmask8, __m512d, __m512d, 9)
526 test_3 (_mm512_maskz_mul_round_ps, __m512, __mmask16, __m512, __m512, 9)
527 test_3 (_mm512_maskz_scalef_round_pd, __m512d, __mmask8, __m512d, __m512d, 9)
528 test_3 (_mm512_maskz_scalef_round_ps, __m512, __mmask16, __m512, __m512, 9)
529 test_3 (_mm512_maskz_shuffle_f32x4, __m512, __mmask16, __m512, __m512, 1)
530 test_3 (_mm512_maskz_shuffle_f64x2, __m512d, __mmask8, __m512d, __m512d, 1)
531 test_3 (_mm512_maskz_shuffle_i32x4, __m512i, __mmask16, __m512i, __m512i, 1)
532 test_3 (_mm512_maskz_shuffle_i64x2, __m512i, __mmask8, __m512i, __m512i, 1)
533 test_3 (_mm512_maskz_shuffle_pd, __m512d, __mmask8, __m512d, __m512d, 1)
534 test_3 (_mm512_maskz_shuffle_ps, __m512, __mmask16, __m512, __m512, 1)
535 test_3 (_mm512_maskz_sub_round_pd, __m512d, __mmask8, __m512d, __m512d, 9)
536 test_3 (_mm512_maskz_sub_round_ps, __m512, __mmask16, __m512, __m512, 9)
537 test_3 (_mm512_ternarylogic_epi32, __m512i, __m512i, __m512i, __m512i, 1)
538 test_3 (_mm512_ternarylogic_epi64, __m512i, __m512i, __m512i, __m512i, 1)
539 test_3 (_mm_mask_cmp_sd_mask, __mmask8, __mmask8, __m128d, __m128d, 1)
540 test_3 (_mm_mask_cmp_ss_mask, __mmask8, __mmask8, __m128, __m128, 1)
541 test_3v (_mm512_i32scatter_epi32, void *, __m512i, __m512i, 1)
542 test_3v (_mm512_i32scatter_epi64, void *, __m256i, __m512i, 1)
543 test_3v (_mm512_i32scatter_pd, void *, __m256i, __m512d, 1)
544 test_3v (_mm512_i32scatter_ps, void *, __m512i, __m512, 1)
545 test_3v (_mm512_i64scatter_epi32, void *, __m512i, __m256i, 1)
546 test_3v (_mm512_i64scatter_epi64, void *, __m512i, __m512i, 1)
547 test_3v (_mm512_i64scatter_pd, void *, __m512i, __m512d, 1)
548 test_3v (_mm512_i64scatter_ps, void *, __m512i, __m256, 1)
549 test_3x (_mm512_mask_roundscale_round_pd, __m512d, __m512d, __mmask8, __m512d, 1, 8)
550 test_3x (_mm512_mask_roundscale_round_ps, __m512, __m512, __mmask16, __m512, 1, 8)
551 test_3x (_mm512_mask_cmp_round_pd_mask, __mmask8, __mmask8, __m512d, __m512d, 1, 8)
552 test_3x (_mm512_mask_cmp_round_ps_mask, __mmask16, __mmask16, __m512, __m512, 1, 8)
553 test_3x (_mm_fixupimm_round_sd, __m128d, __m128d, __m128d, __m128i, 1, 8)
554 test_3x (_mm_fixupimm_round_ss, __m128, __m128, __m128, __m128i, 1, 8)
555 test_3x (_mm_mask_cmp_round_sd_mask, __mmask8, __mmask8, __m128d, __m128d, 1, 8)
556 test_3x (_mm_mask_cmp_round_ss_mask, __mmask8, __mmask8, __m128, __m128, 1, 8)
557 test_4 (_mm512_mask3_fmadd_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9)
558 test_4 (_mm512_mask3_fmadd_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9)
559 test_4 (_mm512_mask3_fmaddsub_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9)
560 test_4 (_mm512_mask3_fmaddsub_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9)
561 test_4 (_mm512_mask3_fmsub_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9)
562 test_4 (_mm512_mask3_fmsub_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9)
563 test_4 (_mm512_mask3_fmsubadd_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9)
564 test_4 (_mm512_mask3_fmsubadd_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9)
565 test_4 (_mm512_mask3_fnmadd_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9)
566 test_4 (_mm512_mask3_fnmadd_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9)
567 test_4 (_mm512_mask3_fnmsub_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9)
568 test_4 (_mm512_mask3_fnmsub_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9)
569 test_4 (_mm512_mask_add_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
570 test_4 (_mm512_mask_add_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
571 test_4 (_mm512_mask_alignr_epi32, __m512i, __m512i, __mmask16, __m512i, __m512i, 1)
572 test_4 (_mm512_mask_alignr_epi64, __m512i, __m512i, __mmask8, __m512i, __m512i, 1)
573 test_4 (_mm512_mask_div_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
574 test_4 (_mm512_mask_div_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
575 test_4 (_mm512_mask_fmadd_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
576 test_4 (_mm512_mask_fmadd_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
577 test_4 (_mm512_mask_fmaddsub_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
578 test_4 (_mm512_mask_fmaddsub_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
579 test_4 (_mm512_mask_fmsub_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
580 test_4 (_mm512_mask_fmsub_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
581 test_4 (_mm512_mask_fmsubadd_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
582 test_4 (_mm512_mask_fmsubadd_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
583 test_4 (_mm512_mask_fnmadd_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
584 test_4 (_mm512_mask_fnmadd_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
585 test_4 (_mm512_mask_fnmsub_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
586 test_4 (_mm512_mask_fnmsub_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
587 test_4 (_mm512_mask_i32gather_epi32, __m512i, __m512i, __mmask16, __m512i, void const *, 1)
588 test_4 (_mm512_mask_i32gather_epi64, __m512i, __m512i, __mmask8, __m256i, void const *, 1)
589 test_4 (_mm512_mask_i32gather_pd, __m512d, __m512d, __mmask8, __m256i, void const *, 1)
590 test_4 (_mm512_mask_i32gather_ps, __m512, __m512, __mmask16, __m512i, void const *, 1)
591 test_4 (_mm512_mask_i64gather_epi32, __m256i, __m256i, __mmask8, __m512i, void const *, 1)
592 test_4 (_mm512_mask_i64gather_epi64, __m512i, __m512i, __mmask8, __m512i, void const *, 1)
593 test_4 (_mm512_mask_i64gather_pd, __m512d, __m512d, __mmask8, __m512i, void const *, 1)
594 test_4 (_mm512_mask_i64gather_ps, __m256, __m256, __mmask8, __m512i, void const *, 1)
595 test_4 (_mm512_mask_insertf32x4, __m512, __m512, __mmask16, __m512, __m128, 1)
596 test_4 (_mm512_mask_insertf64x4, __m512d, __m512d, __mmask8, __m512d, __m256d, 1)
597 test_4 (_mm512_mask_inserti32x4, __m512i, __m512i, __mmask16, __m512i, __m128i, 1)
598 test_4 (_mm512_mask_inserti64x4, __m512i, __m512i, __mmask8, __m512i, __m256i, 1)
599 test_4 (_mm512_mask_max_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 8)
600 test_4 (_mm512_mask_max_round_ps, __m512, __m512, __mmask16, __m512, __m512, 8)
601 test_4 (_mm512_mask_min_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 8)
602 test_4 (_mm512_mask_min_round_ps, __m512, __m512, __mmask16, __m512, __m512, 8)
603 test_4 (_mm512_mask_mul_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
604 test_4 (_mm512_mask_mul_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
605 test_4 (_mm512_mask_scalef_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
606 test_4 (_mm512_mask_scalef_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
607 test_4 (_mm512_mask_shuffle_f32x4, __m512, __m512, __mmask16, __m512, __m512, 1)
608 test_4 (_mm512_mask_shuffle_f64x2, __m512d, __m512d, __mmask8, __m512d, __m512d, 1)
609 test_4 (_mm512_mask_shuffle_i32x4, __m512i, __m512i, __mmask16, __m512i, __m512i, 1)
610 test_4 (_mm512_mask_shuffle_i64x2, __m512i, __m512i, __mmask8, __m512i, __m512i, 1)
611 test_4 (_mm512_mask_shuffle_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 1)
612 test_4 (_mm512_mask_shuffle_ps, __m512, __m512, __mmask16, __m512, __m512, 1)
613 test_4 (_mm512_mask_sub_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
614 test_4 (_mm512_mask_sub_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
615 test_4 (_mm512_mask_ternarylogic_epi32, __m512i, __m512i, __mmask16, __m512i, __m512i, 1)
616 test_4 (_mm512_mask_ternarylogic_epi64, __m512i, __m512i, __mmask8, __m512i, __m512i, 1)
617 test_4 (_mm512_maskz_fmadd_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9)
618 test_4 (_mm512_maskz_fmadd_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9)
619 test_4 (_mm512_maskz_fmaddsub_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9)
620 test_4 (_mm512_maskz_fmaddsub_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9)
621 test_4 (_mm512_maskz_fmsub_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9)
622 test_4 (_mm512_maskz_fmsub_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9)
623 test_4 (_mm512_maskz_fmsubadd_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9)
624 test_4 (_mm512_maskz_fmsubadd_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9)
625 test_4 (_mm512_maskz_fnmadd_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9)
626 test_4 (_mm512_maskz_fnmadd_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9)
627 test_4 (_mm512_maskz_fnmsub_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9)
628 test_4 (_mm512_maskz_fnmsub_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9)
629 test_4 (_mm512_maskz_ternarylogic_epi32, __m512i, __mmask16, __m512i, __m512i, __m512i, 1)
630 test_4 (_mm512_maskz_ternarylogic_epi64, __m512i, __mmask8, __m512i, __m512i, __m512i, 1)
631 test_4v (_mm512_mask_i32scatter_epi32, void *, __mmask16, __m512i, __m512i, 1)
632 test_4v (_mm512_mask_i32scatter_epi64, void *, __mmask8, __m256i, __m512i, 1)
633 test_4v (_mm512_mask_i32scatter_pd, void *, __mmask8, __m256i, __m512d, 1)
634 test_4v (_mm512_mask_i32scatter_ps, void *, __mmask16, __m512i, __m512, 1)
635 test_4v (_mm512_mask_i64scatter_epi32, void *, __mmask8, __m512i, __m256i, 1)
636 test_4v (_mm512_mask_i64scatter_epi64, void *, __mmask8, __m512i, __m512i, 1)
637 test_4v (_mm512_mask_i64scatter_pd, void *, __mmask8, __m512i, __m512d, 1)
638 test_4v (_mm512_mask_i64scatter_ps, void *, __mmask8, __m512i, __m256, 1)
639 test_4x (_mm512_mask_fixupimm_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512i, 1, 8)
640 test_4x (_mm512_mask_fixupimm_round_ps, __m512, __m512, __mmask16, __m512, __m512i, 1, 8)
641 test_4x (_mm512_maskz_fixupimm_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512i, 1, 8)
642 test_4x (_mm512_maskz_fixupimm_round_ps, __m512, __mmask16, __m512, __m512, __m512i, 1, 8)
643 test_4x (_mm_mask_fixupimm_round_sd, __m128d, __m128d, __mmask8, __m128d, __m128i, 1, 8)
644 test_4x (_mm_mask_fixupimm_round_ss, __m128, __m128, __mmask8, __m128, __m128i, 1, 8)
645 test_4x (_mm_maskz_fixupimm_round_sd, __m128d, __mmask8, __m128d, __m128d, __m128i, 1, 8)
646 test_4x (_mm_maskz_fixupimm_round_ss, __m128, __mmask8, __m128, __m128, __m128i, 1, 8)
647 
648 /* avx512pfintrin.h */
649 test_3vx (_mm512_mask_prefetch_i32gather_ps, __m512i, __mmask16, void const *, 1, _MM_HINT_T0)
650 test_3vx (_mm512_mask_prefetch_i32scatter_ps, void const *, __mmask16, __m512i, 1, _MM_HINT_T0)
651 test_3vx (_mm512_mask_prefetch_i64gather_ps, __m512i, __mmask8, void const *, 1, _MM_HINT_T0)
652 test_3vx (_mm512_mask_prefetch_i64scatter_ps, void const *, __mmask8, __m512i, 1, _MM_HINT_T0)
653 
654 test_3vx (_mm512_mask_prefetch_i32gather_pd, __m256i, __mmask8, void const *, 1, _MM_HINT_T0)
655 test_3vx (_mm512_mask_prefetch_i32scatter_pd, void const *, __mmask8, __m256i, 1, _MM_HINT_T0)
656 test_3vx (_mm512_mask_prefetch_i64gather_pd, __m512i, __mmask8, long long *, 1, _MM_HINT_T0)
657 test_3vx (_mm512_mask_prefetch_i64scatter_pd, void const *, __mmask8, __m512i, 1, _MM_HINT_T0)
658 
659 /* avx512erintrin.h */
660 test_1 (_mm512_exp2a23_round_pd, __m512d, __m512d, 8)
661 test_1 (_mm512_exp2a23_round_ps, __m512, __m512, 8)
662 test_1 (_mm512_rcp28_round_pd, __m512d, __m512d, 8)
663 test_1 (_mm512_rcp28_round_ps, __m512, __m512, 8)
664 test_1 (_mm512_rsqrt28_round_pd, __m512d, __m512d, 8)
665 test_1 (_mm512_rsqrt28_round_ps, __m512, __m512, 8)
666 test_2 (_mm512_maskz_exp2a23_round_pd, __m512d, __mmask8, __m512d, 8)
667 test_2 (_mm512_maskz_exp2a23_round_ps, __m512, __mmask16, __m512, 8)
668 test_2 (_mm512_maskz_rcp28_round_pd, __m512d, __mmask8, __m512d, 8)
669 test_2 (_mm512_maskz_rcp28_round_ps, __m512, __mmask16, __m512, 8)
670 test_2 (_mm512_maskz_rsqrt28_round_pd, __m512d, __mmask8, __m512d, 8)
671 test_2 (_mm512_maskz_rsqrt28_round_ps, __m512, __mmask16, __m512, 8)
672 test_3 (_mm512_mask_exp2a23_round_pd, __m512d, __m512d, __mmask8, __m512d, 8)
673 test_3 (_mm512_mask_exp2a23_round_ps, __m512, __m512, __mmask16, __m512, 8)
674 test_3 (_mm512_mask_rcp28_round_pd, __m512d, __m512d, __mmask8, __m512d, 8)
675 test_3 (_mm512_mask_rcp28_round_ps, __m512, __m512, __mmask16, __m512, 8)
676 test_3 (_mm512_mask_rsqrt28_round_pd, __m512d, __m512d, __mmask8, __m512d, 8)
677 test_3 (_mm512_mask_rsqrt28_round_ps, __m512, __m512, __mmask16, __m512, 8)
678 test_2 (_mm_rcp28_round_sd, __m128d, __m128d, __m128d, 8)
679 test_2 (_mm_rcp28_round_ss, __m128, __m128, __m128, 8)
680 test_2 (_mm_rsqrt28_round_sd, __m128d, __m128d, __m128d, 8)
681 test_2 (_mm_rsqrt28_round_ss, __m128, __m128, __m128, 8)
682 
683 /* shaintrin.h */
684 test_2 (_mm_sha1rnds4_epu32, __m128i, __m128i, __m128i, 1)
685 
686 /* wmmintrin.h (AES/PCLMUL).  */
687 #ifdef DIFFERENT_PRAGMAS
688 #pragma GCC target ("aes,pclmul")
689 #endif
690 #include <wmmintrin.h>
691 test_1 (_mm_aeskeygenassist_si128, __m128i, __m128i, 1)
692 test_2 (_mm_clmulepi64_si128, __m128i, __m128i, __m128i, 1)
693 
694 /* popcnintrin.h (POPCNT).  */
695 #ifdef DIFFERENT_PRAGMAS
696 #pragma GCC target ("popcnt")
697 #endif
698 #include <popcntintrin.h>
699 
700 /* x86intrin.h (FMA4/XOP/LWP/BMI/BMI2/TBM/LZCNT/FMA). */
701 #ifdef DIFFERENT_PRAGMAS
702 #pragma GCC target ("fma4,xop,lwp,bmi,bmi2,tbm,lzcnt,fma,rdseed,prfchw,adx,fxsr,xsaveopt,xsavec,xsaves,clflushopt,clwb,pku")
703 #endif
704 #include <x86intrin.h>
705 /* xopintrin.h */
706 test_1 ( _mm_roti_epi8, __m128i, __m128i, 1)
707 test_1 ( _mm_roti_epi16, __m128i, __m128i, 1)
708 test_1 ( _mm_roti_epi32, __m128i, __m128i, 1)
709 test_1 ( _mm_roti_epi64, __m128i, __m128i, 1)
710 test_3 (_mm_permute2_pd, __m128d, __m128d, __m128d, __m128d, 1)
711 test_3 (_mm256_permute2_pd, __m256d, __m256d, __m256d, __m256d, 1)
712 test_3 (_mm_permute2_ps, __m128, __m128, __m128, __m128, 1)
713 test_3 (_mm256_permute2_ps, __m256, __m256, __m256, __m256, 1)
714 
715 /* lwpintrin.h */
716 test_2 ( __lwpval32, void, unsigned int, unsigned int, 1)
717 test_2 ( __lwpins32, unsigned char, unsigned int, unsigned int, 1)
718 #ifdef __x86_64__
719 test_2 ( __lwpval64, void, unsigned long long, unsigned int, 1)
720 test_2 ( __lwpins64, unsigned char, unsigned long long, unsigned int, 1)
721 #endif
722 
723 /* tbmintrin.h */
724 test_1 ( __bextri_u32, unsigned int, unsigned int, 1)
725 #ifdef __x86_64__
726 test_1 ( __bextri_u64, unsigned long long, unsigned long long, 1)
727 #endif
728