1 /* { dg-do compile } */ 2 /* { dg-options "-O0 -Werror-implicit-function-declaration -march=k8 -m3dnow -mavx -mavx2 -msse4a -maes -mpclmul -mavx512bw" } */ 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,g,a,b}mmintrin.h and 9 mm3dnow.h that reference the proper builtin functions. Defining away 10 "extern" and "__inline" results in all of them being compiled as proper 11 functions. */ 12 13 #define extern 14 #define __inline 15 16 #include <wmmintrin.h> 17 #include <immintrin.h> 18 #include <ammintrin.h> 19 #include <mm3dnow.h> 20 21 #define _CONCAT(x,y) x ## y 22 23 #define test_1(func, type, op1_type, imm) \ 24 type _CONCAT(_,func) (op1_type A, int const I) \ 25 { return func (A, imm); } 26 27 #define test_1x(func, type, op1_type, imm1, imm2) \ 28 type _CONCAT(_,func) (op1_type A, int const I, int const L) \ 29 { return func (A, imm1, imm2); } 30 31 #define test_2(func, type, op1_type, op2_type, imm) \ 32 type _CONCAT(_,func) (op1_type A, op2_type B, int const I) \ 33 { return func (A, B, imm); } 34 35 #define test_2x(func, type, op1_type, op2_type, imm1, imm2) \ 36 type _CONCAT(_,func) (op1_type A, op2_type B, int const I, int const L) \ 37 { return func (A, B, imm1, imm2); } 38 39 #define test_3(func, type, op1_type, op2_type, op3_type, imm) \ 40 type _CONCAT(_,func) (op1_type A, op2_type B, \ 41 op3_type C, int const I) \ 42 { return func (A, B, C, imm); } 43 44 #define test_4(func, type, op1_type, op2_type, op3_type, op4_type, imm) \ 45 type _CONCAT(_,func) (op1_type A, op2_type B, \ 46 op3_type C, op4_type D, int const I) \ 47 { return func (A, B, C, D, imm); } 48 49 50 /* Following intrinsics require immediate arguments. They 51 are defined as macros for non-optimized compilations. */ 52 53 /* ammintrin.h */ 54 test_1x (_mm_extracti_si64, __m128i, __m128i, 1, 1) 55 test_2x (_mm_inserti_si64, __m128i, __m128i, __m128i, 1, 1) 56 57 /* immintrin.h */ 58 test_2 (_mm256_blend_pd, __m256d, __m256d, __m256d, 1) 59 test_2 (_mm256_blend_ps, __m256, __m256, __m256, 1) 60 test_2 (_mm256_dp_ps, __m256, __m256, __m256, 1) 61 test_2 (_mm256_shuffle_pd, __m256d, __m256d, __m256d, 1) 62 test_2 (_mm256_shuffle_ps, __m256, __m256, __m256, 1) 63 test_2 (_mm_cmp_sd, __m128d, __m128d, __m128d, 1) 64 test_2 (_mm_cmp_ss, __m128, __m128, __m128, 1) 65 test_2 (_mm_cmp_pd, __m128d, __m128d, __m128d, 1) 66 test_2 (_mm_cmp_ps, __m128, __m128, __m128, 1) 67 test_2 (_mm256_cmp_pd, __m256d, __m256d, __m256d, 1) 68 test_2 (_mm256_cmp_ps, __m256, __m256, __m256, 1) 69 test_1 (_mm256_extractf128_pd, __m128d, __m256d, 1) 70 test_1 (_mm256_extractf128_ps, __m128, __m256, 1) 71 test_1 (_mm256_extractf128_si256, __m128i, __m256i, 1) 72 test_1 (_mm256_extract_epi8, int, __m256i, 20) 73 test_1 (_mm256_extract_epi16, int, __m256i, 13) 74 test_1 (_mm256_extract_epi32, int, __m256i, 6) 75 #ifdef __x86_64__ 76 test_1 (_mm256_extract_epi64, long long, __m256i, 2) 77 #endif 78 test_1 (_mm_permute_pd, __m128d, __m128d, 1) 79 test_1 (_mm256_permute_pd, __m256d, __m256d, 1) 80 test_1 (_mm_permute_ps, __m128, __m128, 1) 81 test_1 (_mm256_permute_ps, __m256, __m256, 1) 82 test_2 (_mm256_permute2f128_pd, __m256d, __m256d, __m256d, 1) 83 test_2 (_mm256_permute2f128_ps, __m256, __m256, __m256, 1) 84 test_2 (_mm256_permute2f128_si256, __m256i, __m256i, __m256i, 1) 85 test_2 (_mm256_insertf128_pd, __m256d, __m256d, __m128d, 1) 86 test_2 (_mm256_insertf128_ps, __m256, __m256, __m128, 1) 87 test_2 (_mm256_insertf128_si256, __m256i, __m256i, __m128i, 1) 88 test_2 (_mm256_insert_epi8, __m256i, __m256i, int, 30) 89 test_2 (_mm256_insert_epi16, __m256i, __m256i, int, 7) 90 test_2 (_mm256_insert_epi32, __m256i, __m256i, int, 3) 91 #ifdef __x86_64__ 92 test_2 (_mm256_insert_epi64, __m256i, __m256i, long long, 1) 93 #endif 94 test_1 (_mm256_round_pd, __m256d, __m256d, 1) 95 test_1 (_mm256_round_ps, __m256, __m256, 1) 96 97 /* wmmintrin.h */ 98 test_1 (_mm_aeskeygenassist_si128, __m128i, __m128i, 1) 99 test_2 (_mm_clmulepi64_si128, __m128i, __m128i, __m128i, 1) 100 101 /* smmintrin.h */ 102 test_1 (_mm_round_pd, __m128d, __m128d, 1) 103 test_1 (_mm_round_ps, __m128, __m128, 1) 104 test_2 (_mm_round_sd, __m128d, __m128d, __m128d, 1) 105 test_2 (_mm_round_ss, __m128, __m128, __m128, 1) 106 107 test_2 (_mm_blend_epi16, __m128i, __m128i, __m128i, 1) 108 test_2 (_mm_blend_ps, __m128, __m128, __m128, 1) 109 test_2 (_mm_blend_pd, __m128d, __m128d, __m128d, 1) 110 test_2 (_mm_dp_ps, __m128, __m128, __m128, 1) 111 test_2 (_mm_dp_pd, __m128d, __m128d, __m128d, 1) 112 test_2 (_mm_insert_ps, __m128, __m128, __m128, 1) 113 test_1 (_mm_extract_ps, int, __m128, 1) 114 test_2 (_mm_insert_epi8, __m128i, __m128i, int, 1) 115 test_2 (_mm_insert_epi32, __m128i, __m128i, int, 1) 116 #ifdef __x86_64__ 117 test_2 (_mm_insert_epi64, __m128i, __m128i, long long, 1) 118 #endif 119 test_1 (_mm_extract_epi8, int, __m128i, 1) 120 test_1 (_mm_extract_epi32, int, __m128i, 1) 121 #ifdef __x86_64__ 122 test_1 (_mm_extract_epi64, long long, __m128i, 1) 123 #endif 124 test_2 (_mm_mpsadbw_epu8, __m128i, __m128i, __m128i, 1) 125 test_2 (_mm_cmpistrm, __m128i, __m128i, __m128i, 1) 126 test_2 (_mm_cmpistri, int, __m128i, __m128i, 1) 127 test_4 (_mm_cmpestrm, __m128i, __m128i, int, __m128i, int, 1) 128 test_4 (_mm_cmpestri, int, __m128i, int, __m128i, int, 1) 129 test_2 (_mm_cmpistra, int, __m128i, __m128i, 1) 130 test_2 (_mm_cmpistrc, int, __m128i, __m128i, 1) 131 test_2 (_mm_cmpistro, int, __m128i, __m128i, 1) 132 test_2 (_mm_cmpistrs, int, __m128i, __m128i, 1) 133 test_2 (_mm_cmpistrz, int, __m128i, __m128i, 1) 134 test_4 (_mm_cmpestra, int, __m128i, int, __m128i, int, 1) 135 test_4 (_mm_cmpestrc, int, __m128i, int, __m128i, int, 1) 136 test_4 (_mm_cmpestro, int, __m128i, int, __m128i, int, 1) 137 test_4 (_mm_cmpestrs, int, __m128i, int, __m128i, int, 1) 138 test_4 (_mm_cmpestrz, int, __m128i, int, __m128i, int, 1) 139 140 /* tmmintrin.h */ 141 test_2 (_mm_alignr_epi8, __m128i, __m128i, __m128i, 1) 142 test_2 (_mm_alignr_pi8, __m64, __m64, __m64, 1) 143 144 /* emmintrin.h */ 145 test_2 (_mm_shuffle_pd, __m128d, __m128d, __m128d, 1) 146 test_1 (_mm_srli_si128, __m128i, __m128i, 1) 147 test_1 (_mm_slli_si128, __m128i, __m128i, 1) 148 test_1 (_mm_extract_epi16, int, __m128i, 1) 149 test_2 (_mm_insert_epi16, __m128i, __m128i, int, 1) 150 test_1 (_mm_shufflehi_epi16, __m128i, __m128i, 1) 151 test_1 (_mm_shufflelo_epi16, __m128i, __m128i, 1) 152 test_1 (_mm_shuffle_epi32, __m128i, __m128i, 1) 153 154 /* xmmintrin.h */ 155 test_2 (_mm_shuffle_ps, __m128, __m128, __m128, 1) 156 test_1 (_mm_extract_pi16, int, __m64, 1) 157 test_1 (_m_pextrw, int, __m64, 1) 158 test_2 (_mm_insert_pi16, __m64, __m64, int, 1) 159 test_2 (_m_pinsrw, __m64, __m64, int, 1) 160 test_1 (_mm_shuffle_pi16, __m64, __m64, 1) 161 test_1 (_m_pshufw, __m64, __m64, 1) 162 test_1 (_mm_prefetch, void, void *, _MM_HINT_NTA) 163 164