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 -mavx512ifma -mavx512vbmi -mclwb -mpcommit -mmwaitx" } */
3/* { dg-add-options bind_pic_locally } */
4
5#include <mm_malloc.h>
6
7/* Test that the intrinsics compile without optimization.  All of them are
8   defined as inline functions in {,x,e,p,t,s,w,a,b,i}mmintrin.h, mm3dnow.h,
9   fma4intrin.h, xopintrin.h, abmintrin.h, bmiintrin.h, tbmintrin.h,
10   lwpintrin.h, fmaintrin.h and mm_malloc.h that reference the proper
11   builtin functions.
12
13   Defining away "extern" and "__inline" results in all of them being compiled
14   as proper functions.  */
15
16#define extern
17#define __inline
18
19#include <x86intrin.h>
20
21#define _CONCAT(x,y) x ## y
22
23#define test_0(func, type, imm)						\
24  type _CONCAT(_,func) (int const I)					\
25  { return func (imm); }
26
27#define test_1(func, type, op1_type, imm)				\
28  type _CONCAT(_,func) (op1_type A, int const I)			\
29  { return func (A, imm); }
30
31#define test_1x(func, type, op1_type, imm1, imm2)			\
32  type _CONCAT(_,func) (op1_type A, int const I, int const L)		\
33  { return func (A, imm1, imm2); }
34
35#define test_1y(func, type, op1_type, imm1, imm2, imm3)			\
36  type _CONCAT(_,func) (op1_type A, int const I, int const L, int const R)\
37  { return func (A, imm1, imm2, imm3); }
38
39#define test_2(func, type, op1_type, op2_type, imm)			\
40  type _CONCAT(_,func) (op1_type A, op2_type B, int const I)		\
41  { return func (A, B, imm); }
42
43#define test_2x(func, type, op1_type, op2_type, imm1, imm2)		\
44  type _CONCAT(_,func) (op1_type A, op2_type B, int const I, int const L) \
45  { return func (A, B, imm1, imm2); }
46
47#define test_2y(func, type, op1_type, op2_type, imm1, imm2, imm3)	 \
48  type _CONCAT(_,func) (op1_type A, op2_type B, int const I, int const L,\
49			int const R)					 \
50  { return func (A, B, imm1, imm2, imm3); }
51
52#define test_2vx(func, op1_type, op2_type, imm1, imm2)     \
53  _CONCAT(_,func) (op1_type A, op2_type B, int const I, int const L) \
54  { func (A, B, imm1, imm2); }
55
56#define test_3(func, type, op1_type, op2_type, op3_type, imm)		\
57  type _CONCAT(_,func) (op1_type A, op2_type B,				\
58			op3_type C, int const I)			\
59  { return func (A, B, C, imm); }
60
61#define test_3x(func, type, op1_type, op2_type, op3_type, imm1, imm2)		\
62  type _CONCAT(_,func) (op1_type A, op2_type B,				\
63			op3_type C, int const I, int const L)			\
64  { return func (A, B, C, imm1, imm2); }
65
66#define test_3y(func, type, op1_type, op2_type, op3_type, imm1, imm2, imm3)	\
67  type _CONCAT(_,func) (op1_type A, op2_type B,				\
68			op3_type C, int const I, int const L, int const R)	\
69  { return func (A, B, C, imm1, imm2, imm3); }
70
71#define test_3v(func, op1_type, op2_type, op3_type, imm)		\
72  int _CONCAT(_,func) (op1_type A, op2_type B,				\
73		       op3_type C, int const I)				\
74  { func (A, B, C, imm); }
75
76#define test_3vx(func, op1_type, op2_type, op3_type, imm1, imm2)   \
77  int _CONCAT(_,func) (op1_type A, op2_type B,			   \
78		       op3_type C, int const I, int const L)       \
79  { func (A, B, C, imm1, imm2); }
80
81#define test_4(func, type, op1_type, op2_type, op3_type, op4_type, imm)	\
82  type _CONCAT(_,func) (op1_type A, op2_type B,				\
83			op3_type C, op4_type D, int const I)		\
84  { return func (A, B, C, D, imm); }
85
86#define test_4x(func, type, op1_type, op2_type, op3_type, op4_type, imm1, imm2)	\
87  type _CONCAT(_,func) (op1_type A, op2_type B,				\
88			op3_type C, op4_type D, int const I, int const L)		\
89  { return func (A, B, C, D, imm1, imm2); }
90
91#define test_4y(func, type, op1_type, op2_type, op3_type, op4_type, imm1, imm2, imm3)	\
92  type _CONCAT(_,func) (op1_type A, op2_type B,	op3_type C,		\
93			op4_type D, int const I, int const L, int const R)		\
94  { return func (A, B, C, D, imm1, imm2, imm3); }
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/* Following intrinsics require immediate arguments.  They
103   are defined as macros for non-optimized compilations. */
104
105/* ammintrin.h */
106test_1x (_mm_extracti_si64, __m128i, __m128i, 1, 1)
107test_2x (_mm_inserti_si64, __m128i, __m128i, __m128i, 1, 1)
108
109/* immintrin.h */
110test_2 (_mm256_blend_pd, __m256d, __m256d, __m256d, 1)
111test_2 (_mm256_blend_ps, __m256, __m256, __m256, 1)
112test_2 (_mm256_dp_ps, __m256, __m256, __m256, 1)
113test_2 (_mm256_shuffle_pd, __m256d, __m256d, __m256d, 1)
114test_2 (_mm256_shuffle_ps, __m256, __m256, __m256, 1)
115test_2 (_mm_cmp_sd, __m128d, __m128d, __m128d, 1)
116test_2 (_mm_cmp_ss, __m128, __m128, __m128, 1)
117test_2 (_mm_cmp_pd, __m128d, __m128d, __m128d, 1)
118test_2 (_mm_cmp_ps, __m128, __m128, __m128, 1)
119test_2 (_mm256_cmp_pd, __m256d, __m256d, __m256d, 1)
120test_2 (_mm256_cmp_ps, __m256, __m256, __m256, 1)
121test_1 (_mm256_extractf128_pd, __m128d, __m256d, 1)
122test_1 (_mm256_extractf128_ps, __m128, __m256, 1)
123test_1 (_mm256_extractf128_si256, __m128i, __m256i, 1)
124test_1 (_mm256_extract_epi8, int, __m256i, 20)
125test_1 (_mm256_extract_epi16, int, __m256i, 13)
126test_1 (_mm256_extract_epi32, int, __m256i, 6)
127#ifdef __x86_64__
128test_1 (_mm256_extract_epi64, long long, __m256i, 2)
129#endif
130test_1 (_mm_permute_pd, __m128d, __m128d, 1)
131test_1 (_mm256_permute_pd, __m256d, __m256d, 1)
132test_1 (_mm_permute_ps, __m128, __m128, 1)
133test_1 (_mm256_permute_ps, __m256, __m256, 1)
134test_2 (_mm256_permute2f128_pd, __m256d, __m256d, __m256d, 1)
135test_2 (_mm256_permute2f128_ps, __m256, __m256, __m256, 1)
136test_2 (_mm256_permute2f128_si256, __m256i, __m256i, __m256i, 1)
137test_2 (_mm256_insertf128_pd, __m256d, __m256d, __m128d, 1)
138test_2 (_mm256_insertf128_ps, __m256, __m256, __m128, 1)
139test_2 (_mm256_insertf128_si256, __m256i, __m256i, __m128i, 1)
140test_2 (_mm256_insert_epi8, __m256i, __m256i, int, 30)
141test_2 (_mm256_insert_epi16, __m256i, __m256i, int, 7)
142test_2 (_mm256_insert_epi32, __m256i, __m256i, int, 3)
143#ifdef __x86_64__
144test_2 (_mm256_insert_epi64, __m256i, __m256i, long long, 1)
145#endif
146test_1 (_mm256_round_pd, __m256d, __m256d, 9)
147test_1 (_mm256_round_ps, __m256, __m256, 9)
148test_1 (_cvtss_sh, unsigned short, float, 1)
149test_1 (_mm_cvtps_ph, __m128i, __m128, 1)
150test_1 (_mm256_cvtps_ph, __m128i, __m256, 1)
151test_0 (_xabort, void, 1)
152test_1 (_mm512_cvt_roundepi32_ps, __m512, __m512i, 9)
153test_1 (_mm512_cvt_roundepu32_ps, __m512, __m512i, 9)
154test_1 (_mm512_cvt_roundpd_epi32, __m256i, __m512d, 9)
155test_1 (_mm512_cvt_roundpd_epu32, __m256i, __m512d, 9)
156test_1 (_mm512_cvt_roundpd_ps, __m256, __m512d, 9)
157test_1 (_mm512_cvt_roundph_ps, __m512, __m256i, 8)
158test_1 (_mm512_cvt_roundps_epi32, __m512i, __m512, 9)
159test_1 (_mm512_cvt_roundps_epu32, __m512i, __m512, 9)
160test_1 (_mm512_cvt_roundps_pd, __m512d, __m256, 8)
161test_1 (_mm512_cvtps_ph, __m256i, __m512, 1)
162test_1 (_mm512_cvtt_roundpd_epi32, __m256i, __m512d, 8)
163test_1 (_mm512_cvtt_roundpd_epu32, __m256i, __m512d, 8)
164test_1 (_mm512_cvtt_roundps_epi32, __m512i, __m512, 8)
165test_1 (_mm512_cvtt_roundps_epu32, __m512i, __m512, 8)
166test_1 (_mm512_extractf32x4_ps, __m128, __m512, 1)
167test_1 (_mm512_extractf64x4_pd, __m256d, __m512d, 1)
168test_1 (_mm512_extracti32x4_epi32, __m128i, __m512i, 1)
169test_1 (_mm512_extracti64x4_epi64, __m256i, __m512i, 1)
170test_1 (_mm512_getexp_round_pd, __m512d, __m512d, 8)
171test_1 (_mm512_getexp_round_ps, __m512, __m512, 8)
172test_1y (_mm512_getmant_round_pd, __m512d, __m512d, 1, 1, 8)
173test_1y (_mm512_getmant_round_ps, __m512, __m512, 1, 1, 8)
174test_1 (_mm512_permute_pd, __m512d, __m512d, 1)
175test_1 (_mm512_permute_ps, __m512, __m512, 1)
176test_1 (_mm512_permutex_epi64, __m512i, __m512i, 1)
177test_1 (_mm512_permutex_pd, __m512d, __m512d, 1)
178test_1 (_mm512_rol_epi32, __m512i, __m512i, 1)
179test_1 (_mm512_rol_epi64, __m512i, __m512i, 1)
180test_1 (_mm512_ror_epi32, __m512i, __m512i, 1)
181test_1 (_mm512_ror_epi64, __m512i, __m512i, 1)
182test_1 (_mm512_shuffle_epi32, __m512i, __m512i, 1)
183test_1 (_mm512_slli_epi32, __m512i, __m512i, 1)
184test_1 (_mm512_slli_epi64, __m512i, __m512i, 1)
185test_1 (_mm512_sqrt_round_pd, __m512d, __m512d, 9)
186test_1 (_mm512_sqrt_round_ps, __m512, __m512, 9)
187test_1 (_mm512_srai_epi32, __m512i, __m512i, 1)
188test_1 (_mm512_srai_epi64, __m512i, __m512i, 1)
189test_1 (_mm512_srli_epi32, __m512i, __m512i, 1)
190test_1 (_mm512_srli_epi64, __m512i, __m512i, 1)
191test_1 (_mm_cvt_roundsd_i32, int, __m128d, 9)
192test_1 (_mm_cvt_roundsd_u32, unsigned, __m128d, 9)
193test_1 (_mm_cvt_roundss_i32, int, __m128, 9)
194test_1 (_mm_cvt_roundss_u32, unsigned, __m128, 9)
195test_1 (_mm_cvtt_roundsd_i32, int, __m128d, 8)
196test_1 (_mm_cvtt_roundsd_u32, unsigned, __m128d, 8)
197test_1 (_mm_cvtt_roundss_i32, int, __m128, 8)
198test_1 (_mm_cvtt_roundss_u32, unsigned, __m128, 8)
199test_1x (_mm512_getmant_pd, __m512d, __m512d, 1, 1)
200test_1x (_mm512_getmant_ps, __m512, __m512, 1, 1)
201test_1x (_mm512_roundscale_round_pd, __m512d, __m512d, 1, 8)
202test_1x (_mm512_roundscale_round_ps, __m512, __m512, 1, 8)
203test_1x (_mm_cvt_roundi32_ss, __m128, __m128, 1, 9)
204test_2 (_mm512_add_round_pd, __m512d, __m512d, __m512d, 9)
205test_2 (_mm512_add_round_ps, __m512, __m512, __m512, 9)
206test_2 (_mm512_alignr_epi32, __m512i, __m512i, __m512i, 1)
207test_2 (_mm512_alignr_epi64, __m512i, __m512i, __m512i, 1)
208test_2 (_mm512_cmp_epi32_mask, __mmask16, __m512i, __m512i, 1)
209test_2 (_mm512_cmp_epi64_mask, __mmask8, __m512i, __m512i, 1)
210test_2 (_mm512_cmp_epu32_mask, __mmask16, __m512i, __m512i, 1)
211test_2 (_mm512_cmp_epu64_mask, __mmask8, __m512i, __m512i, 1)
212test_2 (_mm512_cmp_pd_mask, __mmask8, __m512d, __m512d, 1)
213test_2 (_mm512_cmp_ps_mask, __mmask16, __m512, __m512, 1)
214test_2 (_mm512_div_round_pd, __m512d, __m512d, __m512d, 9)
215test_2 (_mm512_div_round_ps, __m512, __m512, __m512, 9)
216test_2 (_mm512_i32gather_epi32, __m512i, __m512i, void const *, 1)
217test_2 (_mm512_i32gather_epi64, __m512i, __m256i, void const *, 1)
218test_2 (_mm512_i32gather_pd, __m512d, __m256i, void const *, 1)
219test_2 (_mm512_i32gather_ps, __m512, __m512i, void const *, 1)
220test_2 (_mm512_i64gather_epi32, __m256i, __m512i, void const *, 1)
221test_2 (_mm512_i64gather_epi64, __m512i, __m512i, void const *, 1)
222test_2 (_mm512_i64gather_pd, __m512d, __m512i, void const *, 1)
223test_2 (_mm512_i64gather_ps, __m256, __m512i, void const *, 1)
224test_2 (_mm512_insertf32x4, __m512, __m512, __m128, 1)
225test_2 (_mm512_insertf64x4, __m512d, __m512d, __m256d, 1)
226test_2 (_mm512_inserti32x4, __m512i, __m512i, __m128i, 1)
227test_2 (_mm512_inserti64x4, __m512i, __m512i, __m256i, 1)
228test_2 (_mm512_maskz_cvt_roundepi32_ps, __m512, __mmask16, __m512i, 9)
229test_2 (_mm512_maskz_cvt_roundepu32_ps, __m512, __mmask16, __m512i, 9)
230test_2 (_mm512_maskz_cvt_roundpd_epi32, __m256i, __mmask8, __m512d, 9)
231test_2 (_mm512_maskz_cvt_roundpd_epu32, __m256i, __mmask8, __m512d, 9)
232test_2 (_mm512_maskz_cvt_roundpd_ps, __m256, __mmask8, __m512d, 9)
233test_2 (_mm512_maskz_cvt_roundph_ps, __m512, __mmask16, __m256i, 8)
234test_2 (_mm512_maskz_cvt_roundps_epi32, __m512i, __mmask16, __m512, 9)
235test_2 (_mm512_maskz_cvt_roundps_epu32, __m512i, __mmask16, __m512, 9)
236test_2 (_mm512_maskz_cvt_roundps_pd, __m512d, __mmask8, __m256, 8)
237test_2 (_mm512_maskz_cvtps_ph, __m256i, __mmask16, __m512, 1)
238test_2 (_mm512_maskz_cvtt_roundpd_epi32, __m256i, __mmask8, __m512d, 8)
239test_2 (_mm512_maskz_cvtt_roundpd_epu32, __m256i, __mmask8, __m512d, 8)
240test_2 (_mm512_maskz_cvtt_roundps_epi32, __m512i, __mmask16, __m512, 8)
241test_2 (_mm512_maskz_cvtt_roundps_epu32, __m512i, __mmask16, __m512, 8)
242test_2 (_mm512_maskz_extractf32x4_ps, __m128, __mmask8, __m512, 1)
243test_2 (_mm512_maskz_extractf64x4_pd, __m256d, __mmask8, __m512d, 1)
244test_2 (_mm512_maskz_extracti32x4_epi32, __m128i, __mmask8, __m512i, 1)
245test_2 (_mm512_maskz_extracti64x4_epi64, __m256i, __mmask8, __m512i, 1)
246test_2 (_mm512_maskz_getexp_round_pd, __m512d, __mmask8, __m512d, 8)
247test_2 (_mm512_maskz_getexp_round_ps, __m512, __mmask16, __m512, 8)
248test_2y (_mm512_maskz_getmant_round_pd, __m512d, __mmask8, __m512d, 1, 1, 8)
249test_2y (_mm512_maskz_getmant_round_ps, __m512, __mmask16, __m512, 1, 1, 8)
250test_2 (_mm512_maskz_permute_pd, __m512d, __mmask8, __m512d, 1)
251test_2 (_mm512_maskz_permute_ps, __m512, __mmask16, __m512, 1)
252test_2 (_mm512_maskz_permutex_epi64, __m512i, __mmask8, __m512i, 1)
253test_2 (_mm512_maskz_permutex_pd, __m512d, __mmask8, __m512d, 1)
254test_2 (_mm512_maskz_rol_epi32, __m512i, __mmask16, __m512i, 1)
255test_2 (_mm512_maskz_rol_epi64, __m512i, __mmask8, __m512i, 1)
256test_2 (_mm512_maskz_ror_epi32, __m512i, __mmask16, __m512i, 1)
257test_2 (_mm512_maskz_ror_epi64, __m512i, __mmask8, __m512i, 1)
258test_2 (_mm512_maskz_shuffle_epi32, __m512i, __mmask16, __m512i, 1)
259test_2 (_mm512_maskz_slli_epi32, __m512i, __mmask16, __m512i, 1)
260test_2 (_mm512_maskz_slli_epi64, __m512i, __mmask8, __m512i, 1)
261test_2 (_mm512_maskz_sqrt_round_pd, __m512d, __mmask8, __m512d, 9)
262test_2 (_mm512_maskz_sqrt_round_ps, __m512, __mmask16, __m512, 9)
263test_2 (_mm512_maskz_srai_epi32, __m512i, __mmask16, __m512i, 1)
264test_2 (_mm512_maskz_srai_epi64, __m512i, __mmask8, __m512i, 1)
265test_2 (_mm512_maskz_srli_epi32, __m512i, __mmask16, __m512i, 1)
266test_2 (_mm512_maskz_srli_epi64, __m512i, __mmask8, __m512i, 1)
267test_2 (_mm512_max_round_pd, __m512d, __m512d, __m512d, 8)
268test_2 (_mm512_max_round_ps, __m512, __m512, __m512, 8)
269test_2 (_mm512_min_round_pd, __m512d, __m512d, __m512d, 8)
270test_2 (_mm512_min_round_ps, __m512, __m512, __m512, 8)
271test_2 (_mm512_mul_round_pd, __m512d, __m512d, __m512d, 9)
272test_2 (_mm512_mul_round_ps, __m512, __m512, __m512, 9)
273test_2 (_mm512_scalef_round_pd, __m512d, __m512d, __m512d, 9)
274test_2 (_mm512_scalef_round_ps, __m512, __m512, __m512, 9)
275test_2 (_mm512_shuffle_f32x4, __m512, __m512, __m512, 1)
276test_2 (_mm512_shuffle_f64x2, __m512d, __m512d, __m512d, 1)
277test_2 (_mm512_shuffle_i32x4, __m512i, __m512i, __m512i, 1)
278test_2 (_mm512_shuffle_i64x2, __m512i, __m512i, __m512i, 1)
279test_2 (_mm512_shuffle_pd, __m512d, __m512d, __m512d, 1)
280test_2 (_mm512_shuffle_ps, __m512, __m512, __m512, 1)
281test_2 (_mm512_sub_round_pd, __m512d, __m512d, __m512d, 9)
282test_2 (_mm512_sub_round_ps, __m512, __m512, __m512, 9)
283test_2 (_mm_add_round_sd, __m128d, __m128d, __m128d, 9)
284test_2 (_mm_add_round_ss, __m128, __m128, __m128, 9)
285test_2 (_mm_cmp_sd_mask, __mmask8, __m128d, __m128d, 1)
286test_2 (_mm_cmp_ss_mask, __mmask8, __m128, __m128, 1)
287#ifdef __x86_64__
288test_2 (_mm_cvt_roundi64_sd, __m128d, __m128d, long long, 9)
289test_2 (_mm_cvt_roundi64_ss, __m128, __m128, long long, 9)
290#endif
291test_2 (_mm_cvt_roundsd_ss, __m128, __m128, __m128d, 9)
292test_2 (_mm_cvt_roundss_sd, __m128d, __m128d, __m128, 8)
293test_2 (_mm_cvt_roundu32_ss, __m128, __m128, unsigned, 9)
294#ifdef __x86_64__
295test_2 (_mm_cvt_roundu64_sd, __m128d, __m128d, unsigned long long, 9)
296test_2 (_mm_cvt_roundu64_ss, __m128, __m128, unsigned long long, 9)
297#endif
298test_2 (_mm_div_round_sd, __m128d, __m128d, __m128d, 9)
299test_2 (_mm_div_round_ss, __m128, __m128, __m128, 9)
300test_2 (_mm_getexp_round_sd, __m128d, __m128d, __m128d, 8)
301test_2 (_mm_getexp_round_ss, __m128, __m128, __m128, 8)
302test_2y (_mm_getmant_round_sd, __m128d, __m128d, __m128d, 1, 1, 8)
303test_2y (_mm_getmant_round_ss, __m128, __m128, __m128, 1, 1, 8)
304test_2 (_mm_mul_round_sd, __m128d, __m128d, __m128d, 9)
305test_2 (_mm_mul_round_ss, __m128, __m128, __m128, 9)
306test_2 (_mm_scalef_round_sd, __m128d, __m128d, __m128d, 9)
307test_2 (_mm_scalef_round_ss, __m128, __m128, __m128, 9)
308test_2 (_mm_sqrt_round_sd, __m128d, __m128d, __m128d, 9)
309test_2 (_mm_sqrt_round_ss, __m128, __m128, __m128, 9)
310test_2 (_mm_sub_round_sd, __m128d, __m128d, __m128d, 9)
311test_2 (_mm_sub_round_ss, __m128, __m128, __m128, 9)
312test_2x (_mm512_cmp_round_pd_mask, __mmask8, __m512d, __m512d, 1, 8)
313test_2x (_mm512_cmp_round_ps_mask, __mmask16, __m512, __m512, 1, 8)
314test_2x (_mm512_maskz_roundscale_round_pd, __m512d, __mmask8, __m512d, 1, 8)
315test_2x (_mm512_maskz_roundscale_round_ps, __m512, __mmask16, __m512, 1, 8)
316test_2x (_mm_cmp_round_sd_mask, __mmask8, __m128d, __m128d, 1, 8)
317test_2x (_mm_cmp_round_ss_mask, __mmask8, __m128, __m128, 1, 8)
318test_2x (_mm_comi_round_sd, int, __m128d, __m128d, 1, 8)
319test_2x (_mm_comi_round_ss, int, __m128, __m128, 1, 8)
320test_2x (_mm_roundscale_round_sd, __m128d, __m128d, __m128d, 1, 8)
321test_2x (_mm_roundscale_round_ss, __m128, __m128, __m128, 1, 8)
322test_3 (_mm512_fmadd_round_pd, __m512d, __m512d, __m512d, __m512d, 9)
323test_3 (_mm512_fmadd_round_ps, __m512, __m512, __m512, __m512, 9)
324test_3 (_mm512_fmaddsub_round_pd, __m512d, __m512d, __m512d, __m512d, 9)
325test_3 (_mm512_fmaddsub_round_ps, __m512, __m512, __m512, __m512, 9)
326test_3 (_mm512_fmsub_round_pd, __m512d, __m512d, __m512d, __m512d, 9)
327test_3 (_mm512_fmsub_round_ps, __m512, __m512, __m512, __m512, 9)
328test_3 (_mm512_fmsubadd_round_pd, __m512d, __m512d, __m512d, __m512d, 9)
329test_3 (_mm512_fmsubadd_round_ps, __m512, __m512, __m512, __m512, 9)
330test_3 (_mm512_fnmadd_round_pd, __m512d, __m512d, __m512d, __m512d, 9)
331test_3 (_mm512_fnmadd_round_ps, __m512, __m512, __m512, __m512, 9)
332test_3 (_mm512_fnmsub_round_pd, __m512d, __m512d, __m512d, __m512d, 9)
333test_3 (_mm512_fnmsub_round_ps, __m512, __m512, __m512, __m512, 9)
334test_3 (_mm512_mask_cmp_epi32_mask, __mmask16, __mmask16, __m512i, __m512i, 1)
335test_3 (_mm512_mask_cmp_epi64_mask, __mmask8, __mmask8, __m512i, __m512i, 1)
336test_3 (_mm512_mask_cmp_epu32_mask, __mmask16, __mmask16, __m512i, __m512i, 1)
337test_3 (_mm512_mask_cmp_epu64_mask, __mmask8, __mmask8, __m512i, __m512i, 1)
338test_3 (_mm512_mask_cmp_pd_mask, __mmask8, __mmask8, __m512d, __m512d, 1)
339test_3 (_mm512_mask_cmp_ps_mask, __mmask16, __mmask16, __m512, __m512, 1)
340test_3 (_mm512_mask_cvt_roundepi32_ps, __m512, __m512, __mmask16, __m512i, 9)
341test_3 (_mm512_mask_cvt_roundepu32_ps, __m512, __m512, __mmask16, __m512i, 9)
342test_3 (_mm512_mask_cvt_roundpd_epi32, __m256i, __m256i, __mmask8, __m512d, 9)
343test_3 (_mm512_mask_cvt_roundpd_epu32, __m256i, __m256i, __mmask8, __m512d, 9)
344test_3 (_mm512_mask_cvt_roundpd_ps, __m256, __m256, __mmask8, __m512d, 9)
345test_3 (_mm512_mask_cvt_roundph_ps, __m512, __m512, __mmask16, __m256i, 8)
346test_3 (_mm512_mask_cvt_roundps_epi32, __m512i, __m512i, __mmask16, __m512, 9)
347test_3 (_mm512_mask_cvt_roundps_epu32, __m512i, __m512i, __mmask16, __m512, 9)
348test_3 (_mm512_mask_cvt_roundps_pd, __m512d, __m512d, __mmask8, __m256, 8)
349test_3 (_mm512_mask_cvtps_ph, __m256i, __m256i, __mmask16, __m512, 1)
350test_3 (_mm512_mask_cvtt_roundpd_epi32, __m256i, __m256i, __mmask8, __m512d, 8)
351test_3 (_mm512_mask_cvtt_roundpd_epu32, __m256i, __m256i, __mmask8, __m512d, 8)
352test_3 (_mm512_mask_cvtt_roundps_epi32, __m512i, __m512i, __mmask16, __m512, 8)
353test_3 (_mm512_mask_cvtt_roundps_epu32, __m512i, __m512i, __mmask16, __m512, 8)
354test_3 (_mm512_mask_extractf32x4_ps, __m128, __m128, __mmask8, __m512, 1)
355test_3 (_mm512_mask_extractf64x4_pd, __m256d, __m256d, __mmask8, __m512d, 1)
356test_3 (_mm512_mask_extracti32x4_epi32, __m128i, __m128i, __mmask8, __m512i, 1)
357test_3 (_mm512_mask_extracti64x4_epi64, __m256i, __m256i, __mmask8, __m512i, 1)
358test_3 (_mm512_mask_getexp_round_pd, __m512d, __m512d, __mmask8, __m512d, 8)
359test_3 (_mm512_mask_getexp_round_ps, __m512, __m512, __mmask16, __m512, 8)
360test_3y (_mm512_mask_getmant_round_pd, __m512d, __m512d, __mmask8, __m512d, 1, 1, 8)
361test_3y (_mm512_mask_getmant_round_ps, __m512, __m512, __mmask16, __m512, 1, 1, 8)
362test_3 (_mm512_mask_permute_pd, __m512d, __m512d, __mmask8, __m512d, 1)
363test_3 (_mm512_mask_permute_ps, __m512, __m512, __mmask16, __m512, 1)
364test_3 (_mm512_mask_permutex_epi64, __m512i, __m512i, __mmask8, __m512i, 1)
365test_3 (_mm512_mask_permutex_pd, __m512d, __m512d, __mmask8, __m512d, 1)
366test_3 (_mm512_mask_rol_epi32, __m512i, __m512i, __mmask16, __m512i, 1)
367test_3 (_mm512_mask_rol_epi64, __m512i, __m512i, __mmask8, __m512i, 1)
368test_3 (_mm512_mask_ror_epi32, __m512i, __m512i, __mmask16, __m512i, 1)
369test_3 (_mm512_mask_ror_epi64, __m512i, __m512i, __mmask8, __m512i, 1)
370test_3 (_mm512_mask_shuffle_epi32, __m512i, __m512i, __mmask16, __m512i, 1)
371test_3 (_mm512_mask_slli_epi32, __m512i, __m512i, __mmask16, __m512i, 1)
372test_3 (_mm512_mask_slli_epi64, __m512i, __m512i, __mmask8, __m512i, 1)
373test_3 (_mm512_mask_sqrt_round_pd, __m512d, __m512d, __mmask8, __m512d, 9)
374test_3 (_mm512_mask_sqrt_round_ps, __m512, __m512, __mmask16, __m512, 9)
375test_3 (_mm512_mask_srai_epi32, __m512i, __m512i, __mmask16, __m512i, 1)
376test_3 (_mm512_mask_srai_epi64, __m512i, __m512i, __mmask8, __m512i, 1)
377test_3 (_mm512_mask_srli_epi32, __m512i, __m512i, __mmask16, __m512i, 1)
378test_3 (_mm512_mask_srli_epi64, __m512i, __m512i, __mmask8, __m512i, 1)
379test_3 (_mm512_maskz_add_round_pd, __m512d, __mmask8, __m512d, __m512d, 9)
380test_3 (_mm512_maskz_add_round_ps, __m512, __mmask16, __m512, __m512, 9)
381test_3 (_mm512_maskz_alignr_epi32, __m512i, __mmask16, __m512i, __m512i, 1)
382test_3 (_mm512_maskz_alignr_epi64, __m512i, __mmask8, __m512i, __m512i, 1)
383test_3 (_mm512_maskz_div_round_pd, __m512d, __mmask8, __m512d, __m512d, 9)
384test_3 (_mm512_maskz_div_round_ps, __m512, __mmask16, __m512, __m512, 9)
385test_3 (_mm512_maskz_insertf32x4, __m512, __mmask16, __m512, __m128, 1)
386test_3 (_mm512_maskz_insertf64x4, __m512d, __mmask8, __m512d, __m256d, 1)
387test_3 (_mm512_maskz_inserti32x4, __m512i, __mmask16, __m512i, __m128i, 1)
388test_3 (_mm512_maskz_inserti64x4, __m512i, __mmask8, __m512i, __m256i, 1)
389test_3 (_mm512_maskz_max_round_pd, __m512d, __mmask8, __m512d, __m512d, 8)
390test_3 (_mm512_maskz_max_round_ps, __m512, __mmask16, __m512, __m512, 8)
391test_3 (_mm512_maskz_min_round_pd, __m512d, __mmask8, __m512d, __m512d, 8)
392test_3 (_mm512_maskz_min_round_ps, __m512, __mmask16, __m512, __m512, 8)
393test_3 (_mm512_maskz_mul_round_pd, __m512d, __mmask8, __m512d, __m512d, 9)
394test_3 (_mm512_maskz_mul_round_ps, __m512, __mmask16, __m512, __m512, 9)
395test_3 (_mm512_maskz_scalef_round_pd, __m512d, __mmask8, __m512d, __m512d, 9)
396test_3 (_mm512_maskz_scalef_round_ps, __m512, __mmask16, __m512, __m512, 9)
397test_3 (_mm512_maskz_shuffle_f32x4, __m512, __mmask16, __m512, __m512, 1)
398test_3 (_mm512_maskz_shuffle_f64x2, __m512d, __mmask8, __m512d, __m512d, 1)
399test_3 (_mm512_maskz_shuffle_i32x4, __m512i, __mmask16, __m512i, __m512i, 1)
400test_3 (_mm512_maskz_shuffle_i64x2, __m512i, __mmask8, __m512i, __m512i, 1)
401test_3 (_mm512_maskz_shuffle_pd, __m512d, __mmask8, __m512d, __m512d, 1)
402test_3 (_mm512_maskz_shuffle_ps, __m512, __mmask16, __m512, __m512, 1)
403test_3 (_mm512_maskz_sub_round_pd, __m512d, __mmask8, __m512d, __m512d, 9)
404test_3 (_mm512_maskz_sub_round_ps, __m512, __mmask16, __m512, __m512, 9)
405test_3 (_mm512_ternarylogic_epi32, __m512i, __m512i, __m512i, __m512i, 1)
406test_3 (_mm512_ternarylogic_epi64, __m512i, __m512i, __m512i, __m512i, 1)
407test_3 (_mm_fmadd_round_sd, __m128d, __m128d, __m128d, __m128d, 9)
408test_3 (_mm_fmadd_round_ss, __m128, __m128, __m128, __m128, 9)
409test_3 (_mm_fmsub_round_sd, __m128d, __m128d, __m128d, __m128d, 9)
410test_3 (_mm_fmsub_round_ss, __m128, __m128, __m128, __m128, 9)
411test_3 (_mm_fnmadd_round_sd, __m128d, __m128d, __m128d, __m128d, 9)
412test_3 (_mm_fnmadd_round_ss, __m128, __m128, __m128, __m128, 9)
413test_3 (_mm_fnmsub_round_sd, __m128d, __m128d, __m128d, __m128d, 9)
414test_3 (_mm_fnmsub_round_ss, __m128, __m128, __m128, __m128, 9)
415test_3 (_mm_mask_cmp_sd_mask, __mmask8, __mmask8, __m128d, __m128d, 1)
416test_3 (_mm_mask_cmp_ss_mask, __mmask8, __mmask8, __m128, __m128, 1)
417test_3v (_mm512_i32scatter_epi32, void *, __m512i, __m512i, 1)
418test_3v (_mm512_i32scatter_epi64, void *, __m256i, __m512i, 1)
419test_3v (_mm512_i32scatter_pd, void *, __m256i, __m512d, 1)
420test_3v (_mm512_i32scatter_ps, void *, __m512i, __m512, 1)
421test_3v (_mm512_i64scatter_epi32, void *, __m512i, __m256i, 1)
422test_3v (_mm512_i64scatter_epi64, void *, __m512i, __m512i, 1)
423test_3v (_mm512_i64scatter_pd, void *, __m512i, __m512d, 1)
424test_3v (_mm512_i64scatter_ps, void *, __m512i, __m256, 1)
425test_3x (_mm512_mask_roundscale_round_pd, __m512d, __m512d, __mmask8, __m512d, 1, 8)
426test_3x (_mm512_mask_roundscale_round_ps, __m512, __m512, __mmask16, __m512, 1, 8)
427test_3x (_mm_fixupimm_round_sd, __m128d, __m128d, __m128d, __m128i, 1, 8)
428test_3x (_mm_fixupimm_round_ss, __m128, __m128, __m128, __m128i, 1, 8)
429test_3x (_mm_mask_cmp_round_sd_mask, __mmask8, __mmask8, __m128d, __m128d, 1, 8)
430test_3x (_mm_mask_cmp_round_ss_mask, __mmask8, __mmask8, __m128, __m128, 1, 8)
431test_4 (_mm512_mask3_fmadd_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9)
432test_4 (_mm512_mask3_fmadd_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9)
433test_4 (_mm512_mask3_fmaddsub_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9)
434test_4 (_mm512_mask3_fmaddsub_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9)
435test_4 (_mm512_mask3_fmsub_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9)
436test_4 (_mm512_mask3_fmsub_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9)
437test_4 (_mm512_mask3_fmsubadd_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9)
438test_4 (_mm512_mask3_fmsubadd_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9)
439test_4 (_mm512_mask3_fnmadd_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9)
440test_4 (_mm512_mask3_fnmadd_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9)
441test_4 (_mm512_mask3_fnmsub_round_pd, __m512d, __m512d, __m512d, __m512d, __mmask8, 9)
442test_4 (_mm512_mask3_fnmsub_round_ps, __m512, __m512, __m512, __m512, __mmask16, 9)
443test_4 (_mm512_mask_add_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
444test_4 (_mm512_mask_add_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
445test_4 (_mm512_mask_alignr_epi32, __m512i, __m512i, __mmask16, __m512i, __m512i, 1)
446test_4 (_mm512_mask_alignr_epi64, __m512i, __m512i, __mmask8, __m512i, __m512i, 1)
447test_4 (_mm512_mask_div_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
448test_4 (_mm512_mask_div_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
449test_4 (_mm512_mask_fmadd_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
450test_4 (_mm512_mask_fmadd_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
451test_4 (_mm512_mask_fmaddsub_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
452test_4 (_mm512_mask_fmaddsub_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
453test_4 (_mm512_mask_fmsub_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
454test_4 (_mm512_mask_fmsub_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
455test_4 (_mm512_mask_fmsubadd_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
456test_4 (_mm512_mask_fmsubadd_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
457test_4 (_mm512_mask_fnmadd_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
458test_4 (_mm512_mask_fnmadd_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
459test_4 (_mm512_mask_fnmsub_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
460test_4 (_mm512_mask_fnmsub_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
461test_4 (_mm512_mask_i32gather_epi32, __m512i, __m512i, __mmask16, __m512i, void const *, 1)
462test_4 (_mm512_mask_i32gather_epi64, __m512i, __m512i, __mmask8, __m256i, void const *, 1)
463test_4 (_mm512_mask_i32gather_pd, __m512d, __m512d, __mmask8, __m256i, void const *, 1)
464test_4 (_mm512_mask_i32gather_ps, __m512, __m512, __mmask16, __m512i, void const *, 1)
465test_4 (_mm512_mask_i64gather_epi32, __m256i, __m256i, __mmask8, __m512i, void const *, 1)
466test_4 (_mm512_mask_i64gather_epi64, __m512i, __m512i, __mmask8, __m512i, void const *, 1)
467test_4 (_mm512_mask_i64gather_pd, __m512d, __m512d, __mmask8, __m512i, void const *, 1)
468test_4 (_mm512_mask_i64gather_ps, __m256, __m256, __mmask8, __m512i, void const *, 1)
469test_4 (_mm512_mask_insertf32x4, __m512, __m512, __mmask16, __m512, __m128, 1)
470test_4 (_mm512_mask_insertf64x4, __m512d, __m512d, __mmask8, __m512d, __m256d, 1)
471test_4 (_mm512_mask_inserti32x4, __m512i, __m512i, __mmask16, __m512i, __m128i, 1)
472test_4 (_mm512_mask_inserti64x4, __m512i, __m512i, __mmask8, __m512i, __m256i, 1)
473test_4 (_mm512_mask_max_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 8)
474test_4 (_mm512_mask_max_round_ps, __m512, __m512, __mmask16, __m512, __m512, 8)
475test_4 (_mm512_mask_min_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 8)
476test_4 (_mm512_mask_min_round_ps, __m512, __m512, __mmask16, __m512, __m512, 8)
477test_4 (_mm512_mask_mul_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
478test_4 (_mm512_mask_mul_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
479test_4 (_mm512_mask_scalef_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
480test_4 (_mm512_mask_scalef_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
481test_4 (_mm512_mask_shuffle_f32x4, __m512, __m512, __mmask16, __m512, __m512, 1)
482test_4 (_mm512_mask_shuffle_f64x2, __m512d, __m512d, __mmask8, __m512d, __m512d, 1)
483test_4 (_mm512_mask_shuffle_i32x4, __m512i, __m512i, __mmask16, __m512i, __m512i, 1)
484test_4 (_mm512_mask_shuffle_i64x2, __m512i, __m512i, __mmask8, __m512i, __m512i, 1)
485test_4 (_mm512_mask_shuffle_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 1)
486test_4 (_mm512_mask_shuffle_ps, __m512, __m512, __mmask16, __m512, __m512, 1)
487test_4 (_mm512_mask_sub_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512d, 9)
488test_4 (_mm512_mask_sub_round_ps, __m512, __m512, __mmask16, __m512, __m512, 9)
489test_4 (_mm512_mask_ternarylogic_epi32, __m512i, __m512i, __mmask16, __m512i, __m512i, 1)
490test_4 (_mm512_mask_ternarylogic_epi64, __m512i, __m512i, __mmask8, __m512i, __m512i, 1)
491test_4 (_mm512_maskz_fmadd_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9)
492test_4 (_mm512_maskz_fmadd_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9)
493test_4 (_mm512_maskz_fmaddsub_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9)
494test_4 (_mm512_maskz_fmaddsub_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9)
495test_4 (_mm512_maskz_fmsub_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9)
496test_4 (_mm512_maskz_fmsub_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9)
497test_4 (_mm512_maskz_fmsubadd_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9)
498test_4 (_mm512_maskz_fmsubadd_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9)
499test_4 (_mm512_maskz_fnmadd_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9)
500test_4 (_mm512_maskz_fnmadd_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9)
501test_4 (_mm512_maskz_fnmsub_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512d, 9)
502test_4 (_mm512_maskz_fnmsub_round_ps, __m512, __mmask16, __m512, __m512, __m512, 9)
503test_4 (_mm512_maskz_ternarylogic_epi32, __m512i, __mmask16, __m512i, __m512i, __m512i, 1)
504test_4 (_mm512_maskz_ternarylogic_epi64, __m512i, __mmask8, __m512i, __m512i, __m512i, 1)
505test_4v (_mm512_mask_i32scatter_epi32, void *, __mmask16, __m512i, __m512i, 1)
506test_4v (_mm512_mask_i32scatter_epi64, void *, __mmask8, __m256i, __m512i, 1)
507test_4v (_mm512_mask_i32scatter_pd, void *, __mmask8, __m256i, __m512d, 1)
508test_4v (_mm512_mask_i32scatter_ps, void *, __mmask16, __m512i, __m512, 1)
509test_4v (_mm512_mask_i64scatter_epi32, void *, __mmask8, __m512i, __m256i, 1)
510test_4v (_mm512_mask_i64scatter_epi64, void *, __mmask8, __m512i, __m512i, 1)
511test_4v (_mm512_mask_i64scatter_pd, void *, __mmask8, __m512i, __m512d, 1)
512test_4v (_mm512_mask_i64scatter_ps, void *, __mmask8, __m512i, __m256, 1)
513test_4x (_mm512_mask_fixupimm_round_pd, __m512d, __m512d, __mmask8, __m512d, __m512i, 1, 8)
514test_4x (_mm512_mask_fixupimm_round_ps, __m512, __m512, __mmask16, __m512, __m512i, 1, 8)
515test_4x (_mm512_maskz_fixupimm_round_pd, __m512d, __mmask8, __m512d, __m512d, __m512i, 1, 8)
516test_4x (_mm512_maskz_fixupimm_round_ps, __m512, __mmask16, __m512, __m512, __m512i, 1, 8)
517test_4x (_mm_mask_fixupimm_round_sd, __m128d, __m128d, __mmask8, __m128d, __m128i, 1, 8)
518test_4x (_mm_mask_fixupimm_round_ss, __m128, __m128, __mmask8, __m128, __m128i, 1, 8)
519test_4x (_mm_maskz_fixupimm_round_sd, __m128d, __mmask8, __m128d, __m128d, __m128i, 1, 8)
520test_4x (_mm_maskz_fixupimm_round_ss, __m128, __mmask8, __m128, __m128, __m128i, 1, 8)
521
522/* avx512pfintrin.h */
523test_3vx (_mm512_mask_prefetch_i32gather_ps, __m512i, __mmask16, void const *, 1, _MM_HINT_T0)
524test_3vx (_mm512_mask_prefetch_i32scatter_ps, void const *, __mmask16, __m512i, 1, _MM_HINT_T0)
525test_3vx (_mm512_mask_prefetch_i64gather_ps, __m512i, __mmask8, void const *, 1, _MM_HINT_T0)
526test_3vx (_mm512_mask_prefetch_i64scatter_ps, void const *, __mmask8, __m512i, 1, _MM_HINT_T0)
527test_3vx (_mm512_mask_prefetch_i32gather_pd, __m256i, __mmask8, void const *, 1, _MM_HINT_T0)
528test_3vx (_mm512_mask_prefetch_i32scatter_pd, void const *, __mmask8, __m256i, 1, _MM_HINT_T0)
529test_3vx (_mm512_mask_prefetch_i64gather_pd, __m512i, __mmask8, void const *, 1, _MM_HINT_T0)
530test_3vx (_mm512_mask_prefetch_i64scatter_pd, void const *, __mmask8, __m512i, 1, _MM_HINT_T0)
531
532/* avx512erintrin.h */
533test_1 (_mm512_exp2a23_round_pd, __m512d, __m512d, 8)
534test_1 (_mm512_exp2a23_round_ps, __m512, __m512, 8)
535test_1 (_mm512_rcp28_round_pd, __m512d, __m512d, 8)
536test_1 (_mm512_rcp28_round_ps, __m512, __m512, 8)
537test_1 (_mm512_rsqrt28_round_pd, __m512d, __m512d, 8)
538test_1 (_mm512_rsqrt28_round_ps, __m512, __m512, 8)
539test_2 (_mm512_maskz_exp2a23_round_pd, __m512d, __mmask8, __m512d, 8)
540test_2 (_mm512_maskz_exp2a23_round_ps, __m512, __mmask16, __m512, 8)
541test_2 (_mm512_maskz_rcp28_round_pd, __m512d, __mmask8, __m512d, 8)
542test_2 (_mm512_maskz_rcp28_round_ps, __m512, __mmask16, __m512, 8)
543test_2 (_mm512_maskz_rsqrt28_round_pd, __m512d, __mmask8, __m512d, 8)
544test_2 (_mm512_maskz_rsqrt28_round_ps, __m512, __mmask16, __m512, 8)
545test_3 (_mm512_mask_exp2a23_round_pd, __m512d, __m512d, __mmask8, __m512d, 8)
546test_3 (_mm512_mask_exp2a23_round_ps, __m512, __m512, __mmask16, __m512, 8)
547test_3 (_mm512_mask_rcp28_round_pd, __m512d, __m512d, __mmask8, __m512d, 8)
548test_3 (_mm512_mask_rcp28_round_ps, __m512, __m512, __mmask16, __m512, 8)
549test_3 (_mm512_mask_rsqrt28_round_pd, __m512d, __m512d, __mmask8, __m512d, 8)
550test_3 (_mm512_mask_rsqrt28_round_ps, __m512, __m512, __mmask16, __m512, 8)
551
552/* shaintrin.h */
553test_2 (_mm_sha1rnds4_epu32, __m128i, __m128i, __m128i, 1)
554
555/* wmmintrin.h */
556test_1 (_mm_aeskeygenassist_si128, __m128i, __m128i, 1)
557test_2 (_mm_clmulepi64_si128, __m128i, __m128i, __m128i, 1)
558
559/* smmintrin.h */
560test_1 (_mm_round_pd, __m128d, __m128d, 9)
561test_1 (_mm_round_ps, __m128, __m128, 9)
562test_2 (_mm_round_sd, __m128d, __m128d, __m128d, 9)
563test_2 (_mm_round_ss, __m128, __m128, __m128, 9)
564
565test_2 (_mm_blend_epi16, __m128i, __m128i, __m128i, 1)
566test_2 (_mm_blend_ps, __m128, __m128, __m128, 1)
567test_2 (_mm_blend_pd, __m128d, __m128d, __m128d, 1)
568test_2 (_mm_dp_ps, __m128, __m128, __m128, 1)
569test_2 (_mm_dp_pd, __m128d, __m128d, __m128d, 1)
570test_2 (_mm_insert_ps, __m128, __m128, __m128, 1)
571test_1 (_mm_extract_ps, int, __m128, 1)
572test_2 (_mm_insert_epi8, __m128i, __m128i, int, 1)
573test_2 (_mm_insert_epi32, __m128i, __m128i, int, 1)
574#ifdef __x86_64__
575test_2 (_mm_insert_epi64, __m128i, __m128i, long long, 1)
576#endif
577test_1 (_mm_extract_epi8, int, __m128i, 1)
578test_1 (_mm_extract_epi32, int, __m128i, 1)
579#ifdef __x86_64__
580test_1 (_mm_extract_epi64, long long, __m128i, 1)
581#endif
582test_2 (_mm_mpsadbw_epu8, __m128i, __m128i, __m128i, 1)
583test_2 (_mm_cmpistrm, __m128i, __m128i, __m128i, 1)
584test_2 (_mm_cmpistri, int, __m128i, __m128i, 1)
585test_4 (_mm_cmpestrm, __m128i, __m128i, int, __m128i, int, 1)
586test_4 (_mm_cmpestri, int, __m128i, int, __m128i, int, 1)
587test_2 (_mm_cmpistra, int, __m128i, __m128i, 1)
588test_2 (_mm_cmpistrc, int, __m128i, __m128i, 1)
589test_2 (_mm_cmpistro, int, __m128i, __m128i, 1)
590test_2 (_mm_cmpistrs, int, __m128i, __m128i, 1)
591test_2 (_mm_cmpistrz, int, __m128i, __m128i, 1)
592test_4 (_mm_cmpestra, int, __m128i, int, __m128i, int, 1)
593test_4 (_mm_cmpestrc, int, __m128i, int, __m128i, int, 1)
594test_4 (_mm_cmpestro, int, __m128i, int, __m128i, int, 1)
595test_4 (_mm_cmpestrs, int, __m128i, int, __m128i, int, 1)
596test_4 (_mm_cmpestrz, int, __m128i, int, __m128i, int, 1)
597
598/* tmmintrin.h */
599test_2 (_mm_alignr_epi8, __m128i, __m128i, __m128i, 1)
600test_2 (_mm_alignr_pi8, __m64, __m64, __m64, 1)
601
602/* emmintrin.h */
603test_2 (_mm_shuffle_pd, __m128d, __m128d, __m128d, 1)
604test_1 (_mm_bsrli_si128, __m128i, __m128i, 1)
605test_1 (_mm_bslli_si128, __m128i, __m128i, 1)
606test_1 (_mm_srli_si128, __m128i, __m128i, 1)
607test_1 (_mm_slli_si128, __m128i, __m128i, 1)
608test_1 (_mm_extract_epi16, int, __m128i, 1)
609test_2 (_mm_insert_epi16, __m128i, __m128i, int, 1)
610test_1 (_mm_shufflehi_epi16, __m128i, __m128i, 1)
611test_1 (_mm_shufflelo_epi16, __m128i, __m128i, 1)
612test_1 (_mm_shuffle_epi32, __m128i, __m128i, 1)
613
614/* xmmintrin.h */
615test_2 (_mm_shuffle_ps, __m128, __m128, __m128, 1)
616test_1 (_mm_extract_pi16, int, __m64, 1)
617test_1 (_m_pextrw, int, __m64, 1)
618test_2 (_mm_insert_pi16, __m64, __m64, int, 1)
619test_2 (_m_pinsrw, __m64, __m64, int, 1)
620test_1 (_mm_shuffle_pi16, __m64, __m64, 1)
621test_1 (_m_pshufw, __m64, __m64, 1)
622test_1 (_mm_prefetch, void, void *, _MM_HINT_NTA)
623
624/* xopintrin.h */
625test_1 ( _mm_roti_epi8, __m128i, __m128i, 1)
626test_1 ( _mm_roti_epi16, __m128i, __m128i, 1)
627test_1 ( _mm_roti_epi32, __m128i, __m128i, 1)
628test_1 ( _mm_roti_epi64, __m128i, __m128i, 1)
629test_3 (_mm_permute2_pd, __m128d, __m128d, __m128d, __m128d, 1)
630test_3 (_mm256_permute2_pd, __m256d, __m256d, __m256d, __m256d, 1)
631test_3 (_mm_permute2_ps, __m128, __m128, __m128, __m128, 1)
632test_3 (_mm256_permute2_ps, __m256, __m256, __m256, __m256, 1)
633
634/* lwpintrin.h */
635test_2 ( __lwpval32, void, unsigned int, unsigned int, 1)
636test_2 ( __lwpins32, unsigned char, unsigned int, unsigned int, 1)
637#ifdef __x86_64__
638test_2 ( __lwpval64, void, unsigned long long, unsigned int, 1)
639test_2 ( __lwpins64, unsigned char, unsigned long long, unsigned int, 1)
640#endif
641
642/* tbmintrin.h */
643test_1 ( __bextri_u32, unsigned int, unsigned int, 1)
644#ifdef __x86_64__
645test_1 ( __bextri_u64, unsigned long long, unsigned long long, 1)
646#endif
647