1 |
691 |
jeremybenn |
/* 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 |
|
|
|
5 |
|
|
#include <mm_malloc.h>
|
6 |
|
|
|
7 |
|
|
/* Test that the intrinsics compile with optimization. All of them
|
8 |
|
|
are defined as inline functions in {,x,e,p,t,s,w,a,b,i}mmintrin.h,
|
9 |
|
|
mm3dnow.h, fma4intrin.h, xopintrin.h, abmintrin.h, bmiintrin.h,
|
10 |
|
|
tbmintrin.h, lwpintrin.h, popcntintrin.h, fmaintrin.h and mm_malloc.h
|
11 |
|
|
that reference the proper builtin functions.
|
12 |
|
|
|
13 |
|
|
Defining away "extern" and "__inline" results in all of them being
|
14 |
|
|
compiled as proper functions. */
|
15 |
|
|
|
16 |
|
|
#define extern
|
17 |
|
|
#define __inline
|
18 |
|
|
|
19 |
|
|
#define _CONCAT(x,y) x ## y
|
20 |
|
|
|
21 |
|
|
#define test_1(func, type, op1_type, imm) \
|
22 |
|
|
type _CONCAT(_,func) (op1_type A, int const I) \
|
23 |
|
|
{ return func (A, imm); }
|
24 |
|
|
|
25 |
|
|
#define test_1x(func, type, op1_type, imm1, imm2) \
|
26 |
|
|
type _CONCAT(_,func) (op1_type A, int const I, int const L) \
|
27 |
|
|
{ return func (A, imm1, imm2); }
|
28 |
|
|
|
29 |
|
|
#define test_2(func, type, op1_type, op2_type, imm) \
|
30 |
|
|
type _CONCAT(_,func) (op1_type A, op2_type B, int const I) \
|
31 |
|
|
{ return func (A, B, imm); }
|
32 |
|
|
|
33 |
|
|
#define test_2x(func, type, op1_type, op2_type, imm1, imm2) \
|
34 |
|
|
type _CONCAT(_,func) (op1_type A, op2_type B, int const I, int const L) \
|
35 |
|
|
{ return func (A, B, imm1, imm2); }
|
36 |
|
|
|
37 |
|
|
#define test_3(func, type, op1_type, op2_type, op3_type, imm) \
|
38 |
|
|
type _CONCAT(_,func) (op1_type A, op2_type B, \
|
39 |
|
|
op3_type C, int const I) \
|
40 |
|
|
{ return func (A, B, C, imm); }
|
41 |
|
|
|
42 |
|
|
#define test_4(func, type, op1_type, op2_type, op3_type, op4_type, imm) \
|
43 |
|
|
type _CONCAT(_,func) (op1_type A, op2_type B, \
|
44 |
|
|
op3_type C, op4_type D, int const I) \
|
45 |
|
|
{ return func (A, B, C, D, imm); }
|
46 |
|
|
|
47 |
|
|
|
48 |
|
|
#ifndef DIFFERENT_PRAGMAS
|
49 |
|
|
#pragma GCC target ("sse4a,3dnow,avx,avx2,fma4,xop,aes,pclmul,popcnt,abm,lzcnt,bmi,bmi2,tbm,lwp,fsgsbase,rdrnd,f16c")
|
50 |
|
|
#endif
|
51 |
|
|
|
52 |
|
|
/* Following intrinsics require immediate arguments. They
|
53 |
|
|
are defined as macros for non-optimized compilations. */
|
54 |
|
|
|
55 |
|
|
/* mmintrin.h (MMX). */
|
56 |
|
|
#ifdef DIFFERENT_PRAGMAS
|
57 |
|
|
#pragma GCC target ("mmx")
|
58 |
|
|
#endif
|
59 |
|
|
#include <mmintrin.h>
|
60 |
|
|
|
61 |
|
|
/* mm3dnow.h (3DNOW). */
|
62 |
|
|
#ifdef DIFFERENT_PRAGMAS
|
63 |
|
|
#pragma GCC target ("3dnow")
|
64 |
|
|
#endif
|
65 |
|
|
#include <mm3dnow.h>
|
66 |
|
|
|
67 |
|
|
/* xmmintrin.h (SSE). */
|
68 |
|
|
#ifdef DIFFERENT_PRAGMAS
|
69 |
|
|
#pragma GCC target ("sse")
|
70 |
|
|
#endif
|
71 |
|
|
#include <xmmintrin.h>
|
72 |
|
|
test_2 (_mm_shuffle_ps, __m128, __m128, __m128, 1)
|
73 |
|
|
test_1 (_mm_extract_pi16, int, __m64, 1)
|
74 |
|
|
test_1 (_m_pextrw, int, __m64, 1)
|
75 |
|
|
test_2 (_mm_insert_pi16, __m64, __m64, int, 1)
|
76 |
|
|
test_2 (_m_pinsrw, __m64, __m64, int, 1)
|
77 |
|
|
test_1 (_mm_shuffle_pi16, __m64, __m64, 1)
|
78 |
|
|
test_1 (_m_pshufw, __m64, __m64, 1)
|
79 |
|
|
test_1 (_mm_prefetch, void, void *, _MM_HINT_NTA)
|
80 |
|
|
|
81 |
|
|
/* emmintrin.h (SSE2). */
|
82 |
|
|
#ifdef DIFFERENT_PRAGMAS
|
83 |
|
|
#pragma GCC target ("sse2")
|
84 |
|
|
#endif
|
85 |
|
|
#include <emmintrin.h>
|
86 |
|
|
test_2 (_mm_shuffle_pd, __m128d, __m128d, __m128d, 1)
|
87 |
|
|
test_1 (_mm_srli_si128, __m128i, __m128i, 1)
|
88 |
|
|
test_1 (_mm_slli_si128, __m128i, __m128i, 1)
|
89 |
|
|
test_1 (_mm_extract_epi16, int, __m128i, 1)
|
90 |
|
|
test_2 (_mm_insert_epi16, __m128i, __m128i, int, 1)
|
91 |
|
|
test_1 (_mm_shufflehi_epi16, __m128i, __m128i, 1)
|
92 |
|
|
test_1 (_mm_shufflelo_epi16, __m128i, __m128i, 1)
|
93 |
|
|
test_1 (_mm_shuffle_epi32, __m128i, __m128i, 1)
|
94 |
|
|
|
95 |
|
|
/* pmmintrin.h (SSE3). */
|
96 |
|
|
#ifdef DIFFERENT_PRAGMAS
|
97 |
|
|
#pragma GCC target ("sse3")
|
98 |
|
|
#endif
|
99 |
|
|
#include <pmmintrin.h>
|
100 |
|
|
|
101 |
|
|
/* tmmintrin.h (SSSE3). */
|
102 |
|
|
#ifdef DIFFERENT_PRAGMAS
|
103 |
|
|
#pragma GCC target ("ssse3")
|
104 |
|
|
#endif
|
105 |
|
|
#include <tmmintrin.h>
|
106 |
|
|
test_2 (_mm_alignr_epi8, __m128i, __m128i, __m128i, 1)
|
107 |
|
|
test_2 (_mm_alignr_pi8, __m64, __m64, __m64, 1)
|
108 |
|
|
|
109 |
|
|
/* ammintrin.h (SSE4A). */
|
110 |
|
|
#ifdef DIFFERENT_PRAGMAS
|
111 |
|
|
#pragma GCC target ("sse4a")
|
112 |
|
|
#endif
|
113 |
|
|
#include <ammintrin.h>
|
114 |
|
|
test_1x (_mm_extracti_si64, __m128i, __m128i, 1, 1)
|
115 |
|
|
test_2x (_mm_inserti_si64, __m128i, __m128i, __m128i, 1, 1)
|
116 |
|
|
|
117 |
|
|
/* Note, nmmintrin.h includes smmintrin.h, and smmintrin.h
|
118 |
|
|
checks for the #ifdef. So just set the option to SSE4.2. */
|
119 |
|
|
#ifdef DIFFERENT_PRAGMAS
|
120 |
|
|
#pragma GCC target ("sse4.2")
|
121 |
|
|
#endif
|
122 |
|
|
#include <nmmintrin.h>
|
123 |
|
|
/* smmintrin.h (SSE4.2). */
|
124 |
|
|
test_1 (_mm_round_pd, __m128d, __m128d, 1)
|
125 |
|
|
test_1 (_mm_round_ps, __m128, __m128, 1)
|
126 |
|
|
test_2 (_mm_round_sd, __m128d, __m128d, __m128d, 1)
|
127 |
|
|
test_2 (_mm_round_ss, __m128, __m128, __m128, 1)
|
128 |
|
|
|
129 |
|
|
test_2 (_mm_blend_epi16, __m128i, __m128i, __m128i, 1)
|
130 |
|
|
test_2 (_mm_blend_ps, __m128, __m128, __m128, 1)
|
131 |
|
|
test_2 (_mm_blend_pd, __m128d, __m128d, __m128d, 1)
|
132 |
|
|
test_2 (_mm_dp_ps, __m128, __m128, __m128, 1)
|
133 |
|
|
test_2 (_mm_dp_pd, __m128d, __m128d, __m128d, 1)
|
134 |
|
|
test_2 (_mm_insert_ps, __m128, __m128, __m128, 1)
|
135 |
|
|
test_1 (_mm_extract_ps, int, __m128, 1)
|
136 |
|
|
test_2 (_mm_insert_epi8, __m128i, __m128i, int, 1)
|
137 |
|
|
test_2 (_mm_insert_epi32, __m128i, __m128i, int, 1)
|
138 |
|
|
#ifdef __x86_64__
|
139 |
|
|
test_2 (_mm_insert_epi64, __m128i, __m128i, long long, 1)
|
140 |
|
|
#endif
|
141 |
|
|
test_1 (_mm_extract_epi8, int, __m128i, 1)
|
142 |
|
|
test_1 (_mm_extract_epi32, int, __m128i, 1)
|
143 |
|
|
#ifdef __x86_64__
|
144 |
|
|
test_1 (_mm_extract_epi64, long long, __m128i, 1)
|
145 |
|
|
#endif
|
146 |
|
|
test_2 (_mm_mpsadbw_epu8, __m128i, __m128i, __m128i, 1)
|
147 |
|
|
test_2 (_mm_cmpistrm, __m128i, __m128i, __m128i, 1)
|
148 |
|
|
test_2 (_mm_cmpistri, int, __m128i, __m128i, 1)
|
149 |
|
|
test_4 (_mm_cmpestrm, __m128i, __m128i, int, __m128i, int, 1)
|
150 |
|
|
test_4 (_mm_cmpestri, int, __m128i, int, __m128i, int, 1)
|
151 |
|
|
test_2 (_mm_cmpistra, int, __m128i, __m128i, 1)
|
152 |
|
|
test_2 (_mm_cmpistrc, int, __m128i, __m128i, 1)
|
153 |
|
|
test_2 (_mm_cmpistro, int, __m128i, __m128i, 1)
|
154 |
|
|
test_2 (_mm_cmpistrs, int, __m128i, __m128i, 1)
|
155 |
|
|
test_2 (_mm_cmpistrz, int, __m128i, __m128i, 1)
|
156 |
|
|
test_4 (_mm_cmpestra, int, __m128i, int, __m128i, int, 1)
|
157 |
|
|
test_4 (_mm_cmpestrc, int, __m128i, int, __m128i, int, 1)
|
158 |
|
|
test_4 (_mm_cmpestro, int, __m128i, int, __m128i, int, 1)
|
159 |
|
|
test_4 (_mm_cmpestrs, int, __m128i, int, __m128i, int, 1)
|
160 |
|
|
test_4 (_mm_cmpestrz, int, __m128i, int, __m128i, int, 1)
|
161 |
|
|
|
162 |
|
|
/* immintrin.h (AVX/AVX2/RDRND/FSGSBASE/F16C) */
|
163 |
|
|
#ifdef DIFFERENT_PRAGMAS
|
164 |
|
|
#pragma GCC target ("avx,avx2,rdrnd,fsgsbase,f16c")
|
165 |
|
|
#endif
|
166 |
|
|
#include <immintrin.h>
|
167 |
|
|
test_1 (_cvtss_sh, unsigned short, float, 1)
|
168 |
|
|
test_1 (_mm_cvtps_ph, __m128i, __m128, 1)
|
169 |
|
|
test_1 (_mm256_cvtps_ph, __m128i, __m256, 1)
|
170 |
|
|
|
171 |
|
|
/* avxintrin.h */
|
172 |
|
|
test_2 (_mm256_blend_pd, __m256d, __m256d, __m256d, 1)
|
173 |
|
|
test_2 (_mm256_blend_ps, __m256, __m256, __m256, 1)
|
174 |
|
|
test_2 (_mm256_dp_ps, __m256, __m256, __m256, 1)
|
175 |
|
|
test_2 (_mm256_shuffle_pd, __m256d, __m256d, __m256d, 1)
|
176 |
|
|
test_2 (_mm256_shuffle_ps, __m256, __m256, __m256, 1)
|
177 |
|
|
test_2 (_mm_cmp_sd, __m128d, __m128d, __m128d, 1)
|
178 |
|
|
test_2 (_mm_cmp_ss, __m128, __m128, __m128, 1)
|
179 |
|
|
test_2 (_mm_cmp_pd, __m128d, __m128d, __m128d, 1)
|
180 |
|
|
test_2 (_mm_cmp_ps, __m128, __m128, __m128, 1)
|
181 |
|
|
test_2 (_mm256_cmp_pd, __m256d, __m256d, __m256d, 1)
|
182 |
|
|
test_2 (_mm256_cmp_ps, __m256, __m256, __m256, 1)
|
183 |
|
|
test_1 (_mm256_extractf128_pd, __m128d, __m256d, 1)
|
184 |
|
|
test_1 (_mm256_extractf128_ps, __m128, __m256, 1)
|
185 |
|
|
test_1 (_mm256_extractf128_si256, __m128i, __m256i, 1)
|
186 |
|
|
test_1 (_mm256_extract_epi8, int, __m256i, 20)
|
187 |
|
|
test_1 (_mm256_extract_epi16, int, __m256i, 13)
|
188 |
|
|
test_1 (_mm256_extract_epi32, int, __m256i, 6)
|
189 |
|
|
#ifdef __x86_64__
|
190 |
|
|
test_1 (_mm256_extract_epi64, long long, __m256i, 2)
|
191 |
|
|
#endif
|
192 |
|
|
test_1 (_mm_permute_pd, __m128d, __m128d, 1)
|
193 |
|
|
test_1 (_mm256_permute_pd, __m256d, __m256d, 1)
|
194 |
|
|
test_1 (_mm_permute_ps, __m128, __m128, 1)
|
195 |
|
|
test_1 (_mm256_permute_ps, __m256, __m256, 1)
|
196 |
|
|
test_2 (_mm256_permute2f128_pd, __m256d, __m256d, __m256d, 1)
|
197 |
|
|
test_2 (_mm256_permute2f128_ps, __m256, __m256, __m256, 1)
|
198 |
|
|
test_2 (_mm256_permute2f128_si256, __m256i, __m256i, __m256i, 1)
|
199 |
|
|
test_2 (_mm256_insertf128_pd, __m256d, __m256d, __m128d, 1)
|
200 |
|
|
test_2 (_mm256_insertf128_ps, __m256, __m256, __m128, 1)
|
201 |
|
|
test_2 (_mm256_insertf128_si256, __m256i, __m256i, __m128i, 1)
|
202 |
|
|
test_2 (_mm256_insert_epi8, __m256i, __m256i, int, 30)
|
203 |
|
|
test_2 (_mm256_insert_epi16, __m256i, __m256i, int, 7)
|
204 |
|
|
test_2 (_mm256_insert_epi32, __m256i, __m256i, int, 3)
|
205 |
|
|
#ifdef __x86_64__
|
206 |
|
|
test_2 (_mm256_insert_epi64, __m256i, __m256i, long long, 1)
|
207 |
|
|
#endif
|
208 |
|
|
test_1 (_mm256_round_pd, __m256d, __m256d, 1)
|
209 |
|
|
test_1 (_mm256_round_ps, __m256, __m256, 1)
|
210 |
|
|
|
211 |
|
|
/* avx2intrin.h */
|
212 |
|
|
test_2 ( _mm256_mpsadbw_epu8, __m256i, __m256i, __m256i, 1)
|
213 |
|
|
test_2 ( _mm256_alignr_epi8, __m256i, __m256i, __m256i, 1)
|
214 |
|
|
test_2 ( _mm256_blend_epi16, __m256i, __m256i, __m256i, 1)
|
215 |
|
|
test_1 ( _mm256_shuffle_epi32, __m256i, __m256i, 1)
|
216 |
|
|
test_1 ( _mm256_shufflehi_epi16, __m256i, __m256i, 1)
|
217 |
|
|
test_1 ( _mm256_shufflelo_epi16, __m256i, __m256i, 1)
|
218 |
|
|
test_1 ( _mm256_slli_si256, __m256i, __m256i, 8)
|
219 |
|
|
test_1 ( _mm256_srli_si256, __m256i, __m256i, 8)
|
220 |
|
|
test_2 ( _mm_blend_epi32, __m128i, __m128i, __m128i, 1)
|
221 |
|
|
test_2 ( _mm256_blend_epi32, __m256i, __m256i, __m256, 1)
|
222 |
|
|
test_1 ( _mm256_permute4x64_pd, __m256d, __m256d, 1)
|
223 |
|
|
test_1 ( _mm256_permute4x64_epi64, __m256i, __m256i, 1)
|
224 |
|
|
test_2 ( _mm256_permute2x128_si256, __m256i, __m256i, __m256i, 1)
|
225 |
|
|
test_1 ( _mm256_extracti128_si256, __m128i, __m256i, 1)
|
226 |
|
|
test_2 ( _mm256_inserti128_si256, __m256i, __m256i, __m128i, 1)
|
227 |
|
|
test_2 ( _mm_i32gather_pd, __m128d, double const *, __m128i, 1)
|
228 |
|
|
test_2 ( _mm256_i32gather_pd, __m256d, double const *, __m128i, 1)
|
229 |
|
|
test_2 ( _mm_i64gather_pd, __m128d, double const *, __m128i, 1)
|
230 |
|
|
test_2 ( _mm256_i64gather_pd, __m256d, double const *, __m256i, 1)
|
231 |
|
|
test_2 ( _mm_i32gather_ps, __m128, float const *, __m128i, 1)
|
232 |
|
|
test_2 ( _mm256_i32gather_ps, __m256, float const *, __m256i, 1)
|
233 |
|
|
test_2 ( _mm_i64gather_ps, __m128, float const *, __m128i, 1)
|
234 |
|
|
test_2 ( _mm256_i64gather_ps, __m128, float const *, __m256i, 1)
|
235 |
|
|
test_2 ( _mm_i32gather_epi64, __m128i, long long int const *, __m128i, 1)
|
236 |
|
|
test_2 ( _mm256_i32gather_epi64, __m256i, long long int const *, __m128i, 1)
|
237 |
|
|
test_2 ( _mm_i64gather_epi64, __m128i, long long int const *, __m128i, 1)
|
238 |
|
|
test_2 ( _mm256_i64gather_epi64, __m256i, long long int const *, __m256i, 1)
|
239 |
|
|
test_2 ( _mm_i32gather_epi32, __m128i, int const *, __m128i, 1)
|
240 |
|
|
test_2 ( _mm256_i32gather_epi32, __m256i, int const *, __m256i, 1)
|
241 |
|
|
test_2 ( _mm_i64gather_epi32, __m128i, int const *, __m128i, 1)
|
242 |
|
|
test_2 ( _mm256_i64gather_epi32, __m128i, int const *, __m256i, 1)
|
243 |
|
|
|
244 |
|
|
/* wmmintrin.h (AES/PCLMUL). */
|
245 |
|
|
#ifdef DIFFERENT_PRAGMAS
|
246 |
|
|
#pragma GCC target ("aes,pclmul")
|
247 |
|
|
#endif
|
248 |
|
|
#include <wmmintrin.h>
|
249 |
|
|
test_1 (_mm_aeskeygenassist_si128, __m128i, __m128i, 1)
|
250 |
|
|
test_2 (_mm_clmulepi64_si128, __m128i, __m128i, __m128i, 1)
|
251 |
|
|
|
252 |
|
|
/* popcnintrin.h (POPCNT). */
|
253 |
|
|
#ifdef DIFFERENT_PRAGMAS
|
254 |
|
|
#pragma GCC target ("popcnt")
|
255 |
|
|
#endif
|
256 |
|
|
#include <popcntintrin.h>
|
257 |
|
|
|
258 |
|
|
/* x86intrin.h (FMA4/XOP/LWP/BMI/BMI2/TBM/LZCNT/FMA). */
|
259 |
|
|
#ifdef DIFFERENT_PRAGMAS
|
260 |
|
|
#pragma GCC target ("fma4,xop,lwp,bmi,bmi2,tbm,lzcnt,fma")
|
261 |
|
|
#endif
|
262 |
|
|
#include <x86intrin.h>
|
263 |
|
|
/* xopintrin.h */
|
264 |
|
|
test_1 ( _mm_roti_epi8, __m128i, __m128i, 1)
|
265 |
|
|
test_1 ( _mm_roti_epi16, __m128i, __m128i, 1)
|
266 |
|
|
test_1 ( _mm_roti_epi32, __m128i, __m128i, 1)
|
267 |
|
|
test_1 ( _mm_roti_epi64, __m128i, __m128i, 1)
|
268 |
|
|
test_3 (_mm_permute2_pd, __m128d, __m128d, __m128d, __m128d, 1)
|
269 |
|
|
test_3 (_mm256_permute2_pd, __m256d, __m256d, __m256d, __m256d, 1)
|
270 |
|
|
test_3 (_mm_permute2_ps, __m128, __m128, __m128, __m128, 1)
|
271 |
|
|
test_3 (_mm256_permute2_ps, __m256, __m256, __m256, __m256, 1)
|
272 |
|
|
|
273 |
|
|
/* lwpintrin.h */
|
274 |
|
|
test_2 ( __lwpval32, void, unsigned int, unsigned int, 1)
|
275 |
|
|
test_2 ( __lwpins32, unsigned char, unsigned int, unsigned int, 1)
|
276 |
|
|
#ifdef __x86_64__
|
277 |
|
|
test_2 ( __lwpval64, void, unsigned long long, unsigned int, 1)
|
278 |
|
|
test_2 ( __lwpins64, unsigned char, unsigned long long, unsigned int, 1)
|
279 |
|
|
#endif
|
280 |
|
|
|
281 |
|
|
/* tbmintrin.h */
|
282 |
|
|
test_1 ( __bextri_u32, unsigned int, unsigned int, 1)
|
283 |
|
|
#ifdef __x86_64__
|
284 |
|
|
test_1 ( __bextri_u64, unsigned long long, unsigned long long, 1)
|
285 |
|
|
#endif
|