1 // REQUIRES: powerpc-registered-target 2 3 // RUN: %clang -Xclang -no-opaque-pointers -S -emit-llvm -target powerpc64-unknown-linux-gnu -mcpu=pwr8 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \ 4 // RUN: -ffp-contract=off -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK,CHECK-BE 5 // RUN: %clang -Xclang -no-opaque-pointers -S -emit-llvm -target powerpc64le-unknown-linux-gnu -mcpu=pwr8 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \ 6 // RUN: -ffp-contract=off -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK,CHECK-LE 7 8 // RUN: %clang -Xclang -no-opaque-pointers -S -emit-llvm -target powerpc64le-unknown-linux-gnu -mcpu=pwr10 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \ 9 // RUN: -ffp-contract=off -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK-P10-LE 10 11 // CHECK-BE-DAG: @_mm_movemask_pd.__perm_mask = internal constant <4 x i32> <i32 -2139062144, i32 -2139062144, i32 -2139062144, i32 -2139078656>, align 16 12 // CHECK-BE-DAG: @_mm_shuffle_epi32.__permute_selectors = internal constant [4 x i32] [i32 66051, i32 67438087, i32 134810123, i32 202182159], align 4 13 // CHECK-BE-DAG: @_mm_shufflehi_epi16.__permute_selectors = internal constant [4 x i16] [i16 2057, i16 2571, i16 3085, i16 3599], align 2 14 // CHECK-BE-DAG: @_mm_shufflelo_epi16.__permute_selectors = internal constant [4 x i16] [i16 1, i16 515, i16 1029, i16 1543], align 2 15 16 // CHECK-LE-DAG: @_mm_movemask_pd.__perm_mask = internal constant <4 x i32> <i32 -2139094976, i32 -2139062144, i32 -2139062144, i32 -2139062144>, align 16 17 // CHECK-LE-DAG: @_mm_shuffle_epi32.__permute_selectors = internal constant [4 x i32] [i32 50462976, i32 117835012, i32 185207048, i32 252579084], align 4 18 // CHECK-LE-DAG: @_mm_shufflehi_epi16.__permute_selectors = internal constant [4 x i16] [i16 2312, i16 2826, i16 3340, i16 3854], align 2 19 // CHECK-LE-DAG: @_mm_shufflelo_epi16.__permute_selectors = internal constant [4 x i16] [i16 256, i16 770, i16 1284, i16 1798], align 2 20 21 #include <emmintrin.h> 22 23 __m128i resi, mi1, mi2; 24 __m128i *mip; 25 double dp[2]; 26 __m128d resd, md1, md2; 27 __m64 res64, m641, m642; 28 __m128 res, m1; 29 int i; 30 char chs[16]; 31 int is[4]; 32 short ss[8]; 33 long long i64s[2]; 34 35 void __attribute__((noinline)) 36 test_add() { 37 resi = _mm_add_epi64(mi1, mi2); 38 resi = _mm_add_epi32(mi1, mi2); 39 resi = _mm_add_epi16(mi1, mi2); 40 resi = _mm_add_epi8(mi1, mi2); 41 resd = _mm_add_pd(md1, md2); 42 resd = _mm_add_sd(md1, md2); 43 res64 = _mm_add_si64(m641, m642); 44 resi = _mm_adds_epi16(mi1, mi2); 45 resi = _mm_adds_epi8(mi1, mi2); 46 resi = _mm_adds_epu16(mi1, mi2); 47 resi = _mm_adds_epu8(mi1, mi2); 48 } 49 50 // CHECK-LABEL: @test_add 51 52 // CHECK-LABEL: define available_externally <2 x i64> @_mm_add_epi64 53 // CHECK: add <2 x i64> 54 55 // CHECK-LABEL: define available_externally <2 x i64> @_mm_add_epi32 56 // CHECK: add <4 x i32> 57 58 // CHECK-LABEL: define available_externally <2 x i64> @_mm_add_epi16 59 // CHECK: add <8 x i16> 60 61 // CHECK-LABEL: define available_externally <2 x i64> @_mm_add_epi8 62 // CHECK: add <16 x i8> 63 64 // CHECK-LABEL: define available_externally <2 x double> @_mm_add_pd 65 // CHECK: fadd <2 x double> 66 67 // CHECK-LABEL: define available_externally <2 x double> @_mm_add_sd 68 // CHECK: fadd double 69 70 // CHECK-LABEL: define available_externally i64 @_mm_add_si64 71 // CHECK: add i64 72 73 // CHECK-LABEL: define available_externally <2 x i64> @_mm_adds_epi16 74 // CHECK: call <8 x i16> @vec_adds(short vector[8], short vector[8]) 75 76 // CHECK-LABEL: define available_externally <2 x i64> @_mm_adds_epi8 77 // CHECK: call <16 x i8> @vec_adds(signed char vector[16], signed char vector[16]) 78 79 // CHECK-LABEL: define available_externally <2 x i64> @_mm_adds_epu16 80 // CHECK: call <8 x i16> @vec_adds(unsigned short vector[8], unsigned short vector[8]) 81 82 // CHECK-LABEL: define available_externally <2 x i64> @_mm_adds_epu8 83 // CHECK: call <16 x i8> @vec_adds(unsigned char vector[16], unsigned char vector[16]) 84 85 void __attribute__((noinline)) 86 test_avg() { 87 resi = _mm_avg_epu16(mi1, mi2); 88 resi = _mm_avg_epu8(mi1, mi2); 89 } 90 91 // CHECK-LABEL: @test_avg 92 93 // CHECK-LABEL: define available_externally <2 x i64> @_mm_avg_epu16 94 // CHECK: call <8 x i16> @vec_avg(unsigned short vector[8], unsigned short vector[8]) 95 96 // CHECK-LABEL: define available_externally <2 x i64> @_mm_avg_epu8 97 // CHECK: call <16 x i8> @vec_avg(unsigned char vector[16], unsigned char vector[16]) 98 99 void __attribute__((noinline)) 100 test_bs() { 101 resi = _mm_bslli_si128(mi1, i); 102 resi = _mm_bsrli_si128(mi1, i); 103 } 104 105 // CHECK-LABEL: @test_bs 106 107 // CHECK-LABEL: define available_externally <2 x i64> @_mm_bslli_si128 108 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16 109 // CHECK: br i1 %[[CMP]] 110 // CHECK: call <16 x i8> @vec_sld(unsigned char vector[16], unsigned char vector[16], unsigned int)(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef zeroinitializer, i32 noundef zeroext %{{[0-9a-zA-Z_.]+}}) 111 // CHECK: store <16 x i8> zeroinitializer, <16 x i8>* %{{[0-9a-zA-Z_.]+}}, align 16 112 113 // CHECK-LABEL: define available_externally <2 x i64> @_mm_bsrli_si128 114 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16 115 // CHECK: br i1 %[[CMP]] 116 // CHECK-LE: call i1 @llvm.is.constant 117 // CHECK-LE: %[[SUB:[0-9a-zA-Z_.]+]] = sub nsw i32 16, %{{[0-9a-zA-Z_.]+}} 118 // CHECK-LE: call <16 x i8> @vec_sld(unsigned char vector[16], unsigned char vector[16], unsigned int)(<16 x i8> noundef zeroinitializer, <16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef zeroext %[[SUB]]) 119 // CHECK-LE: %[[MUL:[0-9a-zA-Z_.]+]] = mul nsw i32 %{{[0-9a-zA-Z_.]+}}, 8 120 // CHECK-LE: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %[[MUL]] to i8 121 // CHECK-LE: call <16 x i8> @vec_splats(unsigned char)(i8 noundef zeroext %[[TRUNC]]) 122 // CHECK-LE: call <16 x i8> @vec_sro(unsigned char vector[16], unsigned char vector[16]) 123 // CHECK-LE: store <16 x i8> zeroinitializer, <16 x i8>* %{{[0-9a-zA-Z_.]+}}, align 16 124 // CHECK-BE: %[[MUL:[0-9a-zA-Z_.]+]] = mul nsw i32 %{{[0-9a-zA-Z_.]+}}, 8 125 // CHECK-BE: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %[[MUL]] to i8 126 // CHECK-BE: call <16 x i8> @vec_splats(unsigned char)(i8 noundef zeroext %[[TRUNC]]) 127 // CHECK-BE: call <16 x i8> @vec_slo(unsigned char vector[16], unsigned char vector[16]) 128 // CHECK-BE: store <16 x i8> zeroinitializer, <16 x i8>* %{{[0-9a-zA-Z_.]+}}, align 16 129 130 void __attribute__((noinline)) 131 test_cast() { 132 res = _mm_castpd_ps(md1); 133 resi = _mm_castpd_si128(md1); 134 resd = _mm_castps_pd(m1); 135 resi = _mm_castps_si128(m1); 136 resd = _mm_castsi128_pd(mi1); 137 res = _mm_castsi128_ps(mi1); 138 } 139 140 // CHECK-LABEL: @test_cast 141 142 // CHECK-LABEL: define available_externally <4 x float> @_mm_castpd_ps 143 // CHECK: bitcast <2 x double> %{{[0-9a-zA-Z_.]+}} to <4 x float> 144 145 // CHECK-LABEL: define available_externally <2 x i64> @_mm_castpd_si128 146 // CHECK: bitcast <2 x double> %{{[0-9a-zA-Z_.]+}} to <2 x i64> 147 148 // CHECK-LABEL: define available_externally <2 x double> @_mm_castps_pd 149 // CHECK: bitcast <4 x float> %{{[0-9a-zA-Z_.]+}} to <2 x double> 150 151 // CHECK-LABEL: define available_externally <2 x i64> @_mm_castps_si128 152 // CHECK: bitcast <4 x float> %{{[0-9a-zA-Z_.]+}} to <2 x i64> 153 154 // CHECK-LABEL: define available_externally <2 x double> @_mm_castsi128_pd 155 // CHECK: bitcast <2 x i64> %{{[0-9a-zA-Z_.]+}} to <2 x double> 156 157 // CHECK-LABEL: define available_externally <4 x float> @_mm_castsi128_ps 158 // CHECK: bitcast <2 x i64> %{{[0-9a-zA-Z_.]+}} to <4 x float> 159 160 void __attribute__((noinline)) 161 test_cmp() { 162 resi = _mm_cmpeq_epi32(mi1, mi2); 163 resi = _mm_cmpeq_epi16(mi1, mi2); 164 resi = _mm_cmpeq_epi8(mi1, mi2); 165 resi = _mm_cmpgt_epi32(mi1, mi2); 166 resi = _mm_cmpgt_epi16(mi1, mi2); 167 resi = _mm_cmpgt_epi8(mi1, mi2); 168 resi = _mm_cmplt_epi32(mi1, mi2); 169 resi = _mm_cmplt_epi16(mi1, mi2); 170 resi = _mm_cmplt_epi8(mi1, mi2); 171 resd = _mm_cmpeq_pd(md1, md2); 172 resd = _mm_cmpeq_sd(md1, md2); 173 resd = _mm_cmpge_pd(md1, md2); 174 resd = _mm_cmpge_sd(md1, md2); 175 resd = _mm_cmpgt_pd(md1, md2); 176 resd = _mm_cmpgt_sd(md1, md2); 177 resd = _mm_cmple_pd(md1, md2); 178 resd = _mm_cmple_sd(md1, md2); 179 resd = _mm_cmplt_pd(md1, md2); 180 resd = _mm_cmplt_sd(md1, md2); 181 resd = _mm_cmpneq_pd(md1, md2); 182 resd = _mm_cmpneq_sd(md1, md2); 183 resd = _mm_cmpnge_pd(md1, md2); 184 resd = _mm_cmpnge_sd(md1, md2); 185 resd = _mm_cmpngt_pd(md1, md2); 186 resd = _mm_cmpngt_sd(md1, md2); 187 resd = _mm_cmpnle_pd(md1, md2); 188 resd = _mm_cmpnle_sd(md1, md2); 189 resd = _mm_cmpnlt_pd(md1, md2); 190 resd = _mm_cmpnlt_sd(md1, md2); 191 resd = _mm_cmpord_pd(md1, md2); 192 resd = _mm_cmpord_sd(md1, md2); 193 resd = _mm_cmpunord_pd(md1, md2); 194 resd = _mm_cmpunord_sd(md1, md2); 195 } 196 197 // CHECK-LABEL: @test_cmp 198 199 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmpeq_epi32 200 // CHECK: call <4 x i32> @vec_cmpeq(int vector[4], int vector[4]) 201 202 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmpeq_epi16 203 // CHECK: call <8 x i16> @vec_cmpeq(short vector[8], short vector[8]) 204 205 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmpeq_epi8 206 // CHECK: call <16 x i8> @vec_cmpeq(signed char vector[16], signed char vector[16]) 207 208 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmpgt_epi32 209 // CHECK: call <4 x i32> @vec_cmpgt(int vector[4], int vector[4]) 210 211 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmpgt_epi16 212 // CHECK: call <8 x i16> @vec_cmpgt(short vector[8], short vector[8]) 213 214 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmpgt_epi8 215 // CHECK: call <16 x i8> @vec_cmpgt(signed char vector[16], signed char vector[16]) 216 217 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmplt_epi32 218 // CHECK: call <4 x i32> @vec_cmplt(int vector[4], int vector[4]) 219 220 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmplt_epi16 221 // CHECK: call <8 x i16> @vec_cmplt(short vector[8], short vector[8]) 222 223 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmplt_epi8 224 // CHECK: call <16 x i8> @vec_cmplt(signed char vector[16], signed char vector[16]) 225 226 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpeq_pd 227 // CHECK: call <2 x i64> @vec_cmpeq(double vector[2], double vector[2]) 228 229 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpeq_sd 230 // CHECK: call <2 x double> @vec_splats(double) 231 // CHECK: call <2 x double> @vec_splats(double) 232 // CHECK: call <2 x i64> @vec_cmpeq(double vector[2], double vector[2]) 233 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}}) 234 235 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpge_pd 236 // CHECK: call <2 x i64> @vec_cmpge(double vector[2], double vector[2]) 237 238 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpge_sd 239 // CHECK: call <2 x double> @vec_splats(double) 240 // CHECK: call <2 x double> @vec_splats(double) 241 // CHECK: call <2 x i64> @vec_cmpge(double vector[2], double vector[2]) 242 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}}) 243 244 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpgt_pd 245 // CHECK: call <2 x i64> @vec_cmpgt(double vector[2], double vector[2]) 246 247 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpgt_sd 248 // CHECK: call <2 x double> @vec_splats(double) 249 // CHECK: call <2 x double> @vec_splats(double) 250 // CHECK: call <2 x i64> @vec_cmpgt(double vector[2], double vector[2]) 251 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}}) 252 253 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmple_pd 254 // CHECK: call <2 x i64> @vec_cmple(double vector[2], double vector[2]) 255 256 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmple_sd 257 // CHECK: call <2 x double> @vec_splats(double) 258 // CHECK: call <2 x double> @vec_splats(double) 259 // CHECK: call <2 x i64> @vec_cmple(double vector[2], double vector[2]) 260 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}}) 261 262 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmplt_pd 263 // CHECK: call <2 x i64> @vec_cmplt(double vector[2], double vector[2]) 264 265 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmplt_sd 266 // CHECK: call <2 x double> @vec_splats(double) 267 // CHECK: call <2 x double> @vec_splats(double) 268 // CHECK: call <2 x i64> @vec_cmplt(double vector[2], double vector[2]) 269 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}}) 270 271 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpneq_pd 272 // CHECK: call <2 x i64> @vec_cmpeq(double vector[2], double vector[2]) 273 // CHECK: call <2 x double> @vec_nor(double vector[2], double vector[2]) 274 275 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpneq_sd 276 // CHECK: call <2 x double> @vec_splats(double) 277 // CHECK: call <2 x double> @vec_splats(double) 278 // CHECK: call <2 x i64> @vec_cmpeq(double vector[2], double vector[2]) 279 // CHECK: call <2 x double> @vec_nor(double vector[2], double vector[2]) 280 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}}) 281 282 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpnge_pd 283 // CHECK: call <2 x i64> @vec_cmplt(double vector[2], double vector[2]) 284 285 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpnge_sd 286 // CHECK: call <2 x double> @vec_splats(double) 287 // CHECK: call <2 x double> @vec_splats(double) 288 // CHECK: call <2 x i64> @vec_cmplt(double vector[2], double vector[2]) 289 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}}) 290 291 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpngt_pd 292 // CHECK: call <2 x i64> @vec_cmple(double vector[2], double vector[2]) 293 294 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpngt_sd 295 // CHECK: call <2 x double> @vec_splats(double) 296 // CHECK: call <2 x double> @vec_splats(double) 297 // CHECK: call <2 x i64> @vec_cmple(double vector[2], double vector[2]) 298 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}}) 299 300 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpnle_pd 301 // CHECK: call <2 x i64> @vec_cmpgt(double vector[2], double vector[2]) 302 303 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpnle_sd 304 // CHECK: call <2 x double> @vec_splats(double) 305 // CHECK: call <2 x double> @vec_splats(double) 306 // CHECK: call <2 x i64> @vec_cmpge(double vector[2], double vector[2]) 307 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}}) 308 309 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpnlt_pd 310 // CHECK: call <2 x i64> @vec_cmpge(double vector[2], double vector[2]) 311 312 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpnlt_sd 313 // CHECK: call <2 x double> @vec_splats(double) 314 // CHECK: call <2 x double> @vec_splats(double) 315 // CHECK: call <2 x i64> @vec_cmpge(double vector[2], double vector[2]) 316 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}}) 317 318 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpord_pd 319 // CHECK: call <2 x i64> @vec_cmpeq(double vector[2], double vector[2]) 320 // CHECK: call <2 x i64> @vec_cmpeq(double vector[2], double vector[2]) 321 // CHECK: call <2 x i64> @vec_and(unsigned long long vector[2], unsigned long long vector[2]) 322 323 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpord_sd 324 // CHECK: call <2 x double> @vec_splats(double) 325 // CHECK: call <2 x double> @vec_splats(double) 326 // CHECK: call <2 x double> @_mm_cmpord_pd(<2 x double> noundef %{{[0-9a-zA-Z_.]+}}, <2 x double> noundef %{{[0-9a-zA-Z_.]+}}) 327 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}}) 328 329 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpunord_pd 330 // CHECK: call <2 x i64> @vec_cmpeq(double vector[2], double vector[2]) 331 // CHECK: call <2 x i64> @vec_cmpeq(double vector[2], double vector[2]) 332 // CHECK: call <2 x i64> @vec_nor(unsigned long long vector[2], unsigned long long vector[2]) 333 // CHECK: call <2 x i64> @vec_orc(unsigned long long vector[2], unsigned long long vector[2]) 334 335 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpunord_sd 336 // CHECK: call <2 x double> @vec_splats(double) 337 // CHECK: call <2 x double> @vec_splats(double) 338 // CHECK: call <2 x double> @_mm_cmpunord_pd(<2 x double> noundef %{{[0-9a-zA-Z_.]+}}, <2 x double> noundef %{{[0-9a-zA-Z_.]+}}) 339 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}}) 340 341 void __attribute__((noinline)) 342 test_comi() { 343 i = _mm_comieq_sd(md1, md2); 344 i = _mm_comige_sd(md1, md2); 345 i = _mm_comigt_sd(md1, md2); 346 i = _mm_comile_sd(md1, md2); 347 i = _mm_comilt_sd(md1, md2); 348 i = _mm_comineq_sd(md1, md2); 349 } 350 351 // CHECK-LABEL: @test_comi 352 353 // CHECK-LABEL: define available_externally signext i32 @_mm_comieq_sd 354 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = fcmp oeq double 355 // CHECK: zext i1 %[[CMP]] to i32 356 357 // CHECK-LABEL: define available_externally signext i32 @_mm_comige_sd 358 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = fcmp oge double 359 // CHECK: zext i1 %[[CMP]] to i32 360 361 // CHECK-LABEL: define available_externally signext i32 @_mm_comigt_sd 362 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = fcmp ogt double 363 // CHECK: zext i1 %[[CMP]] to i32 364 365 // CHECK-LABEL: define available_externally signext i32 @_mm_comile_sd 366 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = fcmp ole double 367 // CHECK: zext i1 %[[CMP]] to i32 368 369 // CHECK-LABEL: define available_externally signext i32 @_mm_comilt_sd 370 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = fcmp olt double 371 // CHECK: zext i1 %[[CMP]] to i32 372 373 // CHECK-LABEL: define available_externally signext i32 @_mm_comineq_sd 374 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = fcmp une double 375 // CHECK: zext i1 %[[CMP]] to i32 376 377 void __attribute__((noinline)) 378 test_control() { 379 _mm_clflush(dp); 380 _mm_lfence(); 381 _mm_mfence(); 382 _mm_pause(); 383 } 384 385 // CHECK-LABEL: @test_control 386 387 // CHECK-LABEL: define available_externally void @_mm_clflush 388 // CHECK: call void asm sideeffect "dcbf 0,$0", "b,~{memory}"(i8* %{{[0-9a-zA-Z_.]+}}) 389 390 // CHECK-LABEL: define available_externally void @_mm_lfence() 391 // CHECK: fence release 392 393 // CHECK-LABEL: define available_externally void @_mm_mfence() 394 // CHECK: fence seq_cst 395 396 // CHECK-LABEL: define available_externally void @_mm_pause() 397 // CHECK: call i64 asm sideeffect "\09mfppr\09$0; or 31,31,31; isync; lwsync; isync; mtppr\09$0;", "=r,~{memory}"() 398 399 void __attribute__((noinline)) 400 test_converts() { 401 resd = _mm_cvtepi32_pd(mi1); 402 res = _mm_cvtepi32_ps(mi1); 403 resi = _mm_cvtpd_epi32(md1); 404 res64 = _mm_cvtpd_pi32(md1); 405 res = _mm_cvtpd_ps(md1); 406 resd = _mm_cvtpi32_pd(res64); 407 resi = _mm_cvtps_epi32(m1); 408 resd = _mm_cvtps_pd(m1); 409 *dp = _mm_cvtsd_f64(md1); 410 i = _mm_cvtsd_si32(md1); 411 i64s[0] = _mm_cvtsd_si64(md1); 412 i64s[0] = _mm_cvtsd_si64x(md1); 413 res = _mm_cvtsd_ss(m1, md2); 414 i = _mm_cvtsi128_si32(mi1); 415 i64s[0] = _mm_cvtsi128_si64(mi1); 416 i64s[0] = _mm_cvtsi128_si64x(mi1); 417 resd = _mm_cvtsi32_sd(md1, i); 418 resi = _mm_cvtsi32_si128(i); 419 resd = _mm_cvtsi64_sd(md1, i64s[1]); 420 resi = _mm_cvtsi64_si128(i64s[1]); 421 resd = _mm_cvtsi64x_sd(md1, i64s[1]); 422 resi = _mm_cvtsi64x_si128(i64s[1]); 423 resd = _mm_cvtss_sd(md1, m1); 424 resi = _mm_cvttpd_epi32(md1); 425 res64 = _mm_cvttpd_pi32(md1); 426 resi = _mm_cvttps_epi32(m1); 427 i = _mm_cvttsd_si32(md1); 428 i64s[0] = _mm_cvttsd_si64(md1); 429 i64s[0] = _mm_cvttsd_si64x(md1); 430 } 431 432 // CHECK-LABEL: @test_converts 433 434 // CHECK-LABEL: define available_externally <2 x double> @_mm_cvtepi32_pd 435 // CHECK: call <2 x i64> @vec_unpackh(int vector[4]) 436 // CHECK: %[[CONV:[0-9a-zA-Z_.]+]] = sitofp <2 x i64> %{{[0-9a-zA-Z_.]+}} to <2 x double> 437 // CHECK: fmul <2 x double> %[[CONV]], <double 1.000000e+00, double 1.000000e+00> 438 439 // CHECK-LABEL: define available_externally <4 x float> @_mm_cvtepi32_ps 440 // CHECK: call <4 x float> @llvm.ppc.altivec.vcfsx(<4 x i32> %{{[0-9a-zA-Z_.]+}}, i32 0) 441 442 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtpd_epi32 443 // CHECK: call <2 x double> @vec_rint(double vector[2]) 444 // CHECK: store <4 x i32> zeroinitializer, <4 x i32>* %{{[0-9a-zA-Z_.]+}}, align 16 445 // CHECK: call <4 x i32> asm "xvcvdpsxws ${0:x},${1:x}", "=^wa,^wa"(<2 x double> %{{[0-9a-zA-Z_.]+}}) 446 // CHECK-LE: call <4 x i32> @vec_mergeo(int vector[4], int vector[4]) 447 // CHECK-BE: call <4 x i32> @vec_mergee(int vector[4], int vector[4]) 448 // CHECK: call <4 x i32> @vec_vpkudum(long long vector[2], long long vector[2])(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef zeroinitializer) 449 450 // CHECK-LABEL: define available_externally i64 @_mm_cvtpd_pi32 451 // CHECK: call <2 x i64> @_mm_cvtpd_epi32(<2 x double> noundef %{{[0-9a-zA-Z_.]+}}) 452 // CHECK: extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 0 453 454 // CHECK-LABEL: define available_externally <4 x float> @_mm_cvtpd_ps 455 // CHECK: store <4 x i32> zeroinitializer, <4 x i32>* %{{[0-9a-zA-Z_.]+}}, align 16 456 // CHECK: call <4 x i32> asm "xvcvdpsp ${0:x},${1:x}", "=^wa,^wa"(<2 x double> %{{[0-9a-zA-Z_.]+}}) 457 // CHECK-LE: call <4 x i32> @vec_mergeo(int vector[4], int vector[4]) 458 // CHECK-BE: call <4 x i32> @vec_mergee(int vector[4], int vector[4]) 459 // CHECK: call <4 x i32> @vec_vpkudum(long long vector[2], long long vector[2])(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef zeroinitializer) 460 461 // CHECK-LABEL: define available_externally <2 x double> @_mm_cvtpi32_pd 462 // CHECK: call <2 x i64> @vec_splats(unsigned long long) 463 // CHECK: call <2 x i64> @vec_unpackl(int vector[4]) 464 // CHECK: %[[CONV:[0-9a-zA-Z_.]+]] = sitofp <2 x i64> %{{[0-9a-zA-Z._]+}} to <2 x double> 465 // CHECK: fmul <2 x double> %[[CONV]], <double 1.000000e+00, double 1.000000e+00> 466 467 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtps_epi32 468 // CHECK: call <4 x float> @vec_rint(float vector[4]) 469 // CHECK: call <4 x i32> @llvm.ppc.altivec.vctsxs(<4 x float> %{{[0-9a-zA-Z_.]+}}, i32 0) 470 471 // CHECK-LABEL: define available_externally <2 x double> @_mm_cvtps_pd 472 // CHECK-BE: call <4 x float> @vec_vmrghw(float vector[4], float vector[4]) 473 // CHECK-BE: call <2 x double> asm " xvcvspdp ${0:x},${1:x}", "=^wa,^wa"(<4 x float> %{{[0-9a-zA-Z_.]+}}) 474 // CHECK-LE: shufflevector <4 x i32> %{{[0-9a-zA-Z_.]+}}, <4 x i32> %{{[0-9a-zA-Z_.]+}}, <4 x i32> <i32 5, i32 6, i32 7, i32 0> 475 // CHECK-LE: shufflevector <4 x i32> %{{[0-9a-zA-Z_.]+}}, <4 x i32> %{{[0-9a-zA-Z_.]+}}, <4 x i32> <i32 6, i32 7, i32 0, i32 1> 476 // CHECK-LE: call <2 x double> asm " xvcvspdp ${0:x},${1:x}", "=^wa,^wa"(<4 x float> %{{[0-9a-zA-Z_.]+}}) 477 478 // CHECK-LABEL: define available_externally double @_mm_cvtsd_f64 479 // CHECK: extractelement <2 x double> %{{[0-9a-zA-Z_.]+}}, i32 0 480 481 // CHECK-LABEL: define available_externally signext i32 @_mm_cvtsd_si32 482 // CHECK: call <2 x double> @vec_rint(double vector[2]) 483 // CHECK: fptosi double %{{[0-9a-zA-Z_.]+}} to i32 484 485 // CHECK-LABEL: define available_externally i64 @_mm_cvtsd_si64 486 // CHECK: call <2 x double> @vec_rint(double vector[2]) 487 // CHECK: fptosi double %{{[0-9a-zA-Z_.]+}} to i64 488 489 // CHECK-LABEL: define available_externally i64 @_mm_cvtsd_si64x 490 // CHECK: call i64 @_mm_cvtsd_si64(<2 x double> noundef %{{[0-9a-zA-Z_.]+}}) 491 492 // CHECK-LABEL: define available_externally <4 x float> @_mm_cvtsd_ss 493 // CHECK-BE: %[[EXT:[0-9a-zA-Z_.]+]] = extractelement <2 x double> %{{[0-9a-zA-Z_.]+}}, i32 0 494 // CHECK-BE: %[[TRUNC:[0-9a-zA-Z_.]+]] = fptrunc double %[[EXT]] to float 495 // CHECK-BE: insertelement <4 x float> %{{[0-9a-zA-Z_.]+}}, float %[[TRUNC]], i32 0 496 // CHECK-LE: call <2 x double> @vec_splat(double vector[2], unsigned int)(<2 x double> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef zeroext 0) 497 // CHECK-LE: shufflevector <4 x i32> %{{[0-9a-zA-Z_.]+}}, <4 x i32> %{{[0-9a-zA-Z_.]+}}, <4 x i32> <i32 5, i32 6, i32 7, i32 0> 498 // CHECK-LE: call <4 x float> asm "xscvdpsp ${0:x},${1:x}", "=^wa,^wa"(<2 x double> %{{[0-9a-zA-Z_.]+}}) 499 // CHECK-LE: shufflevector <4 x i32> %{{[0-9a-zA-Z_.]+}}, <4 x i32> %{{[0-9a-zA-Z_.]+}}, <4 x i32> <i32 7, i32 0, i32 1, i32 2> 500 501 // CHECK-LABEL: define available_externally signext i32 @_mm_cvtsi128_si32 502 // CHECK: extractelement <4 x i32> %{{[0-9a-zA-Z_.]+}}, i32 0 503 504 // CHECK-LABEL: define available_externally i64 @_mm_cvtsi128_si64 505 // CHECK: extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 0 506 507 // CHECK-LABEL: define available_externally i64 @_mm_cvtsi128_si64x 508 // CHECK: extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 0 509 510 // CHECK-LABEL: define available_externally <2 x double> @_mm_cvtsi32_sd 511 // CHECK: sitofp i32 %{{[0-9a-zA-Z_.]+}} to double 512 513 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtsi32_si128 514 // CHECK: call <2 x i64> @_mm_set_epi32(i32 noundef signext 0, i32 noundef signext 0, i32 noundef signext 0, i32 noundef signext %{{[0-9a-zA-Z_.]+}}) 515 516 // CHECK-LABEL: define available_externally <2 x double> @_mm_cvtsi64_sd 517 // CHECK: sitofp i64 %{{[0-9a-zA-Z_.]+}} to double 518 519 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtsi64_si128 520 // CHECK: %[[INS:[0-9a-zA-Z_.]+]] = insertelement <2 x i64> undef, i64 %{{[0-9a-zA-Z_.]+}}, i32 0 521 // CHECK: insertelement <2 x i64> %[[INS]], i64 0, i32 1 522 523 // CHECK-LABEL: define available_externally <2 x double> @_mm_cvtsi64x_sd 524 // CHECK: call <2 x double> @_mm_cvtsi64_sd(<2 x double> noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef %{{[0-9a-zA-Z_.]+}}) 525 526 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtsi64x_si128 527 // CHECK: %[[INS:[0-9a-zA-Z_.]+]] = insertelement <2 x i64> undef, i64 %{{[0-9a-zA-Z_.]+}}, i32 0 528 // CHECK: insertelement <2 x i64> %[[INS]], i64 0, i32 1 529 530 // CHECK-LABEL: define available_externally <2 x double> @_mm_cvtss_sd 531 // CHECK-BE: fpext float %{{[0-9a-zA-Z_.]+}} to double 532 // CHECK-LE: call <4 x float> @vec_splat(float vector[4], unsigned int)(<4 x float> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef zeroext 0) 533 // CHECK-LE: call <2 x double> asm "xscvspdp ${0:x},${1:x}", "=^wa,^wa"(<4 x float> %{{[0-9a-zA-Z_.]+}}) 534 // CHECK-LE: call <2 x double> @vec_mergel(double vector[2], double vector[2]) 535 536 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvttpd_epi32 537 // CHECK: call <4 x i32> asm "xvcvdpsxws ${0:x},${1:x}", "=^wa,^wa" 538 // CHECK-LE: call <4 x i32> @vec_mergeo(int vector[4], int vector[4]) 539 // CHECK-BE: call <4 x i32> @vec_mergee(int vector[4], int vector[4]) 540 // CHECK: call <4 x i32> @vec_vpkudum(long long vector[2], long long vector[2])(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef zeroinitializer) 541 542 // CHECK-LABEL: define available_externally i64 @_mm_cvttpd_pi32 543 // CHECK: call <2 x i64> @_mm_cvttpd_epi32(<2 x double> noundef %{{[0-9a-zA-Z_.]+}}) 544 545 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvttps_epi32 546 // CHECK: call <4 x i32> @llvm.ppc.altivec.vctsxs(<4 x float> %{{[0-9a-zA-Z_.]+}}, i32 0) 547 548 // CHECK-LABEL: define available_externally signext i32 @_mm_cvttsd_si32 549 // CHECK: fptosi double %{{[0-9a-zA-Z_.]+}} to i32 550 551 // CHECK-LABEL: define available_externally i64 @_mm_cvttsd_si64 552 // CHECK: fptosi double %{{[0-9a-zA-Z_.]+}} to i64 553 554 // CHECK-LABEL: define available_externally i64 @_mm_cvttsd_si64x 555 // CHECK: call i64 @_mm_cvttsd_si64(<2 x double> noundef %{{[0-9a-zA-Z_.]+}}) 556 557 void __attribute__((noinline)) 558 test_div() { 559 resd = _mm_div_pd(md1, md2); 560 resd = _mm_div_sd(md1, md2); 561 } 562 563 // CHECK-LABEL: @test_div 564 565 // CHECK-LABEL: define available_externally <2 x double> @_mm_div_pd 566 // CHECK: fdiv <2 x double> 567 568 // CHECK-LABEL: define available_externally <2 x double> @_mm_div_sd 569 // CHECK: fdiv double 570 571 void __attribute__((noinline)) 572 test_extract() { 573 i = _mm_extract_epi16(mi1, i); 574 } 575 576 // CHECK-LABEL: @test_extract 577 578 // CHECK-LABEL: define available_externally signext i32 @_mm_extract_epi16 579 // CHECK: %[[AND:[0-9a-zA-Z_.]+]] = and i32 %{{[0-9a-zA-Z_.]+}}, 7 580 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = extractelement <8 x i16> %{{[0-9a-zA-Z_.]+}}, i32 %[[AND]] 581 // CHECK: zext i16 %[[EXT]] to i32 582 583 void __attribute__((noinline)) 584 test_insert() { 585 resi = _mm_insert_epi16 (mi1, i, is[0]); 586 } 587 588 // CHECK-LABEL: @test_insert 589 590 // CHECK-LABEL: define available_externally <2 x i64> @_mm_insert_epi16 591 // CHECK: trunc i32 %{{[0-9a-zA-Z_.]+}} to i16 592 // CHECK: and i32 %{{[0-9a-zA-Z_.]+}}, 7 593 594 void __attribute__((noinline)) 595 test_load() { 596 resd = _mm_load_pd(dp); 597 resd = _mm_load_pd1(dp); 598 resd = _mm_load_sd(dp); 599 resi = _mm_load_si128(mip); 600 resd = _mm_load1_pd(dp); 601 resd = _mm_loadh_pd(md1, dp); 602 resi = _mm_loadl_epi64(mip); 603 resd = _mm_loadl_pd(md1, dp); 604 resd = _mm_loadr_pd(dp); 605 resd = _mm_loadu_pd(dp); 606 resi = _mm_loadu_si128(mip); 607 } 608 609 // CHECK-LABEL: @test_load 610 611 // CHECK-LABEL: define available_externally <2 x double> @_mm_load_pd 612 // CHECK: call <16 x i8> @vec_ld(long, unsigned char vector[16] const*)(i64 noundef 0, <16 x i8>* noundef %{{[0-9a-zA-Z_.]+}}) 613 614 // CHECK-LABEL: define available_externally <2 x double> @_mm_load_pd1 615 // CHECK: call <2 x double> @_mm_load1_pd(double* noundef %{{[0-9a-zA-Z_.]+}}) 616 617 // CHECK-LABEL: define available_externally <2 x double> @_mm_load_sd 618 // CHECK: call <2 x double> @_mm_set_sd(double noundef %{{[0-9a-zA-Z_.]+}}) 619 620 // CHECK-LABEL: define available_externally <2 x i64> @_mm_load_si128 621 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load <2 x i64>*, <2 x i64>** %{{[0-9a-zA-Z_.]+}}, align 8 622 // CHECK: load <2 x i64>, <2 x i64>* %[[ADDR]], align 16 623 624 // CHECK-LABEL: define available_externally <2 x double> @_mm_load1_pd 625 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load double*, double** %{{[0-9a-zA-Z_.]+}}, align 8 626 // CHECK: %[[VAL:[0-9a-zA-Z_.]+]] = load double, double* %[[ADDR]], align 8 627 // CHECK: call <2 x double> @vec_splats(double)(double noundef %[[VAL]]) 628 629 // CHECK-LABEL: define available_externally <2 x double> @_mm_loadh_pd 630 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load double*, double** %{{[0-9a-zA-Z_.]+}}, align 8 631 // CHECK: %[[VAL:[0-9a-zA-Z_.]+]] = load double, double* %{{[0-9a-zA-Z_.]+}}, align 8 632 // CHECK: %[[VEC:[0-9a-zA-Z_.]+]] = load <2 x double>, <2 x double>* %{{[0-9a-zA-Z_.]+}}, align 16 633 // CHECK: insertelement <2 x double> %[[VEC]], double %[[VAL]], i32 1 634 635 // CHECK-LABEL: define available_externally <2 x i64> @_mm_loadl_epi64 636 // CHECK: call <2 x i64> @_mm_set_epi64(i64 noundef 0, i64 noundef %{{[0-9a-zA-Z_.]+}}) 637 638 // CHECK-LABEL: define available_externally <2 x double> @_mm_loadl_pd 639 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load double*, double** %{{[0-9a-zA-Z_.]+}}, align 8 640 // CHECK: %[[ADDR2:[0-9a-zA-Z_.]+]] = load double, double* %[[ADDR]], align 8 641 // CHECK: %[[VEC:[0-9a-zA-Z_.]+]] = load <2 x double>, <2 x double>* %{{[0-9a-zA-Z_.]+}}, align 16 642 // CHECK: insertelement <2 x double> %[[VEC]], double %[[ADDR2]], i32 0 643 644 // CHECK-LABEL: define available_externally <2 x double> @_mm_loadr_pd 645 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load double*, double** %{{[0-9a-zA-Z_.]+}}, align 8 646 // CHECK: call <2 x double> @_mm_load_pd(double* noundef %[[ADDR]]) 647 // CHECK: shufflevector <2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i32> <i32 1, i32 2> 648 649 // CHECK-LABEL: define available_externally <2 x double> @_mm_loadu_pd 650 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load double*, double** %{{[0-9a-zA-Z_.]+}}, align 8 651 // CHECK: call <2 x double> @vec_vsx_ld(int, double const*)(i32 noundef signext 0, double* noundef %[[ADDR]]) 652 653 // CHECK-LABEL: define available_externally <2 x i64> @_mm_loadu_si128 654 // CHECK: load <2 x i64>*, <2 x i64>** %{{[0-9a-zA-Z_.]+}}, align 8 655 // CHECK: call <4 x i32> @vec_vsx_ld(int, int const*)(i32 noundef signext 0, i32* noundef %{{[0-9a-zA-Z_.]+}}) 656 657 void __attribute__((noinline)) 658 test_logical() { 659 resd = _mm_and_pd(md1, md2); 660 resi = _mm_and_si128(mi1, mi2); 661 resd = _mm_andnot_pd(md1, md2); 662 resi = _mm_andnot_si128(mi1, mi2); 663 resd = _mm_xor_pd(md1, md2); 664 resi = _mm_xor_si128(mi1, mi2); 665 resd = _mm_or_pd(md1, md2); 666 resi = _mm_or_si128(mi1, mi2); 667 } 668 669 // CHECK-LABEL: @test_logical 670 671 // CHECK-LABEL: define available_externally <2 x double> @_mm_and_pd 672 // CHECK: call <2 x double> @vec_and(double vector[2], double vector[2]) 673 674 // CHECK-LABEL: define available_externally <2 x i64> @_mm_and_si128 675 // CHECK: call <2 x i64> @vec_and(long long vector[2], long long vector[2]) 676 677 // CHECK-LABEL: define available_externally <2 x double> @_mm_andnot_pd 678 // CHECK: call <2 x double> @vec_andc(double vector[2], double vector[2]) 679 680 // CHECK-LABEL: define available_externally <2 x i64> @_mm_andnot_si128 681 // CHECK: call <2 x i64> @vec_andc(long long vector[2], long long vector[2]) 682 683 // CHECK-LABEL: define available_externally <2 x double> @_mm_xor_pd 684 // CHECK: call <2 x double> @vec_xor(double vector[2], double vector[2]) 685 686 // CHECK-LABEL: define available_externally <2 x i64> @_mm_xor_si128 687 // CHECK: call <2 x i64> @vec_xor(long long vector[2], long long vector[2]) 688 689 // CHECK-LABEL: define available_externally <2 x double> @_mm_or_pd 690 // CHECK: call <2 x double> @vec_or(double vector[2], double vector[2]) 691 692 // CHECK-LABEL: define available_externally <2 x i64> @_mm_or_si128 693 // CHECK: call <2 x i64> @vec_or(long long vector[2], long long vector[2]) 694 695 void __attribute__((noinline)) 696 test_max() { 697 resi = _mm_max_epi16(mi1, mi2); 698 resi = _mm_max_epu8(mi1, mi2); 699 resd = _mm_max_pd(md1, md2); 700 resd = _mm_max_sd(md1, md2); 701 } 702 703 // CHECK-LABEL: @test_max 704 705 // CHECK-LABEL: define available_externally <2 x i64> @_mm_max_epi16 706 // CHECK: call <8 x i16> @vec_max(short vector[8], short vector[8]) 707 708 // CHECK-LABEL: define available_externally <2 x i64> @_mm_max_epu8 709 // CHECK: call <16 x i8> @vec_max(unsigned char vector[16], unsigned char vector[16]) 710 711 // CHECK-LABEL: define available_externally <2 x double> @_mm_max_pd 712 // CHECK: call <2 x double> @vec_max(double vector[2], double vector[2]) 713 714 // CHECK-LABEL: define available_externally <2 x double> @_mm_max_sd 715 // CHECK: call <2 x double> @vec_splats(double) 716 // CHECK: call <2 x double> @vec_splats(double) 717 // CHECK: call <2 x double> @vec_max(double vector[2], double vector[2]) 718 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}}) 719 720 void __attribute__((noinline)) 721 test_min() { 722 resi = _mm_min_epi16(mi1, mi2); 723 resi = _mm_min_epu8(mi1, mi2); 724 resd = _mm_min_pd(md1, md2); 725 resd = _mm_min_sd(md1, md2); 726 } 727 728 // CHECK-LABEL: @test_min 729 730 // CHECK-LABEL: define available_externally <2 x i64> @_mm_min_epi16 731 // CHECK: call <8 x i16> @vec_min(short vector[8], short vector[8]) 732 733 // CHECK-LABEL: define available_externally <2 x i64> @_mm_min_epu8 734 // CHECK: call <16 x i8> @vec_min(unsigned char vector[16], unsigned char vector[16]) 735 736 // CHECK-LABEL: define available_externally <2 x double> @_mm_min_pd 737 // CHECK: call <2 x double> @vec_min(double vector[2], double vector[2]) 738 739 // CHECK-LABEL: define available_externally <2 x double> @_mm_min_sd 740 // CHECK: call <2 x double> @vec_splats(double) 741 // CHECK: call <2 x double> @vec_splats(double) 742 // CHECK: call <2 x double> @vec_min(double vector[2], double vector[2]) 743 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}}) 744 745 void __attribute__((noinline)) 746 test_move() { 747 resi = _mm_move_epi64(mi1); 748 resd = _mm_move_sd(md1, md2); 749 i = _mm_movemask_epi8(mi1); 750 i = _mm_movemask_pd(md1); 751 res64 = _mm_movepi64_pi64(mi1); 752 resi = _mm_movpi64_epi64(m641); 753 _mm_maskmoveu_si128(mi1, mi2, chs); 754 } 755 756 // CHECK-LABEL: @test_move 757 758 // CHECK-LABEL: define available_externally <2 x i64> @_mm_move_epi64 759 // CHECK: call <2 x i64> @_mm_set_epi64(i64 noundef 0, i64 noundef %{{[0-9a-zA-Z_.]+}}) 760 761 // CHECK-LABEL: define available_externally <2 x double> @_mm_move_sd 762 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = extractelement <2 x double> %{{[0-9a-zA-Z_.]+}}, i32 0 763 // CHECK: insertelement <2 x double> %{{[0-9a-zA-Z_.]+}}, double %[[EXT]], i32 0 764 765 // CHECK-P10-LE-LABEL: define available_externally signext i32 @_mm_movemask_epi8 766 // CHECK-P10-LE: call zeroext i32 @vec_extractm(unsigned char vector[16])(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}) 767 768 // CHECK-LABEL: define available_externally signext i32 @_mm_movemask_epi8 769 // CHECK: call <2 x i64> @vec_vbpermq(unsigned char vector[16], unsigned char vector[16])(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 120, i8 112, i8 104, i8 96, i8 88, i8 80, i8 72, i8 64, i8 56, i8 48, i8 40, i8 32, i8 24, i8 16, i8 8, i8 0>) 770 // CHECK-LE: %[[VAL:[0-9a-zA-Z_.]+]] = extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 1 771 // CHECK-BE: %[[VAL:[0-9a-zA-Z_.]+]] = extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 0 772 // CHECK: trunc i64 %[[VAL]] to i32 773 774 // CHECK-P10-LE-LABEL: define available_externally signext i32 @_mm_movemask_pd 775 // CHECK-P10-LE: call zeroext i32 @vec_extractm(unsigned long long vector[2])(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}) 776 777 // CHECK-LABEL: define available_externally signext i32 @_mm_movemask_pd 778 // CHECK-LE: call <2 x i64> @vec_vbpermq(unsigned char vector[16], unsigned char vector[16])(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef bitcast (<4 x i32> <i32 -2139094976, i32 -2139062144, i32 -2139062144, i32 -2139062144> to <16 x i8>)) 779 // CHECK-LE: extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 1 780 // CHECK-BE: call <2 x i64> @vec_vbpermq(unsigned char vector[16], unsigned char vector[16])(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef bitcast (<4 x i32> <i32 -2139062144, i32 -2139062144, i32 -2139062144, i32 -2139078656> to <16 x i8>)) 781 // CHECK-BE: extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 0 782 783 // CHECK-LABEL: define available_externally i64 @_mm_movepi64_pi64 784 // CHECK: extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 0 785 786 // CHECK-LABEL: define available_externally <2 x i64> @_mm_movpi64_epi64 787 // CHECK: call <2 x i64> @_mm_set_epi64(i64 noundef 0, i64 noundef %{{[0-9a-zA-Z_.]+}}) 788 789 // CHECK-LABEL: define available_externally void @_mm_maskmoveu_si128 790 // CHECK: call <2 x i64> @_mm_loadu_si128(<2 x i64>* noundef %{{[0-9a-zA-Z_.]+}}) 791 // CHECK: call <16 x i8> @vec_cmpgt(unsigned char vector[16], unsigned char vector[16]) 792 // CHECK: call <16 x i8> @vec_sel(unsigned char vector[16], unsigned char vector[16], unsigned char vector[16]) 793 // CHECK: call void @_mm_storeu_si128(<2 x i64>* noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}}) 794 795 void __attribute__((noinline)) 796 test_mul() { 797 resi = _mm_mul_epu32(mi1, mi2); 798 resd = _mm_mul_pd(md1, md2); 799 resd = _mm_mul_sd(md1, md2); 800 res64 = _mm_mul_su32(m641, m642); 801 resi = _mm_mulhi_epi16(mi1, mi2); 802 resi = _mm_mulhi_epu16(mi1, mi2); 803 resi = _mm_mullo_epi16(mi1, mi2); 804 } 805 806 // CHECK-LABEL: @test_mul 807 808 // CHECK-LABEL: define available_externally <2 x i64> @_mm_mul_epu32 809 // CHECK-LE: call <2 x i64> asm "vmulouw $0,$1,$2", "=v,v,v"(<2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i64> %{{[0-9a-zA-Z_.]+}}) 810 // CHECK-BE: call <2 x i64> asm "vmuleuw $0,$1,$2", "=v,v,v"(<2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i64> %{{[0-9a-zA-Z_.]+}}) 811 812 // CHECK-LABEL: define available_externally <2 x double> @_mm_mul_pd 813 // CHECK: fmul <2 x double> 814 815 // CHECK-LABEL: define available_externally <2 x double> @_mm_mul_sd 816 // CHECK: fmul double 817 818 // CHECK-LABEL: define available_externally i64 @_mm_mul_su32 819 // CHECK: trunc i64 %{{[0-9a-zA-Z_.]+}} to i32 820 // CHECK: trunc i64 %{{[0-9a-zA-Z_.]+}} to i32 821 // CHECK: %[[EXT1:[0-9a-zA-Z_.]+]] = zext i32 %{{[0-9a-zA-Z_.]+}} to i64 822 // CHECK: %[[EXT2:[0-9a-zA-Z_.]+]] = zext i32 %{{[0-9a-zA-Z_.]+}} to i64 823 // CHECK: mul i64 %[[EXT1]], %[[EXT2]] 824 825 // CHECK-LABEL: define available_externally <2 x i64> @_mm_mulhi_epi16 826 // CHECK-LE: store <16 x i8> <i8 2, i8 3, i8 18, i8 19, i8 6, i8 7, i8 22, i8 23, i8 10, i8 11, i8 26, i8 27, i8 14, i8 15, i8 30, i8 31>, <16 x i8>* %{{[0-9a-zA-Z_.]+}}, align 16 827 // CHECK-BE: store <16 x i8> <i8 0, i8 1, i8 16, i8 17, i8 4, i8 5, i8 20, i8 21, i8 8, i8 9, i8 24, i8 25, i8 12, i8 13, i8 28, i8 29>, <16 x i8>* %{{[0-9a-zA-Z_.]+}}, align 16 828 // CHECK: call <4 x i32> @vec_vmulesh(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}) 829 // CHECK: call <4 x i32> @vec_vmulosh(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}) 830 // CHECK: call <4 x i32> @vec_perm(int vector[4], int vector[4], unsigned char vector[16]) 831 832 // CHECK-LABEL: define available_externally <2 x i64> @_mm_mulhi_epu16 833 // CHECK-LE: store <16 x i8> <i8 2, i8 3, i8 18, i8 19, i8 6, i8 7, i8 22, i8 23, i8 10, i8 11, i8 26, i8 27, i8 14, i8 15, i8 30, i8 31>, <16 x i8>* %{{[0-9a-zA-Z_.]+}}, align 16 834 // CHECK-BE: store <16 x i8> <i8 0, i8 1, i8 16, i8 17, i8 4, i8 5, i8 20, i8 21, i8 8, i8 9, i8 24, i8 25, i8 12, i8 13, i8 28, i8 29>, <16 x i8>* %{{[0-9a-zA-Z_.]+}}, align 16 835 // CHECK: call <4 x i32> @vec_vmuleuh(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}) 836 // CHECK: call <4 x i32> @vec_vmulouh(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}) 837 // CHECK: call <4 x i32> @vec_perm(unsigned int vector[4], unsigned int vector[4], unsigned char vector[16]) 838 839 // CHECK-LABEL: define available_externally <2 x i64> @_mm_mullo_epi16 840 // CHECK: mul <8 x i16> 841 842 void __attribute__((noinline)) 843 test_pack() { 844 resi = _mm_packs_epi16(mi1, mi2); 845 resi = _mm_packs_epi32(mi1, mi2); 846 resi = _mm_packus_epi16(mi1, mi2); 847 } 848 849 // CHECK-LABEL: @test_pack 850 851 // CHECK-LABEL: define available_externally <2 x i64> @_mm_packs_epi16 852 // CHECK: call <16 x i8> @vec_packs(short vector[8], short vector[8]) 853 854 // CHECK-LABEL: define available_externally <2 x i64> @_mm_packs_epi32 855 // CHECK: call <8 x i16> @vec_packs(int vector[4], int vector[4]) 856 857 // CHECK-LABEL: define available_externally <2 x i64> @_mm_packus_epi16 858 // CHECK: call <16 x i8> @vec_packsu(short vector[8], short vector[8]) 859 860 void __attribute__((noinline)) 861 test_sad() { 862 resi = _mm_sad_epu8(mi1, mi2); 863 } 864 865 // CHECK-LABEL: @test_sad 866 867 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sad_epu8 868 // CHECK: call <16 x i8> @vec_min(unsigned char vector[16], unsigned char vector[16]) 869 // CHECK: call <16 x i8> @vec_max(unsigned char vector[16], unsigned char vector[16]) 870 // CHECK: call <16 x i8> @vec_sub(unsigned char vector[16], unsigned char vector[16]) 871 // CHECK: call <4 x i32> @vec_sum4s(unsigned char vector[16], unsigned int vector[4])(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, <4 x i32> noundef zeroinitializer) 872 // CHECK-LE: call <4 x i32> asm "vsum2sws $0,$1,$2", "=v,v,v"(<4 x i32> %11, <4 x i32> zeroinitializer) 873 // CHECK-BE: call <4 x i32> @vec_sum2s(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <4 x i32> noundef zeroinitializer) 874 // CHECK-BE: call <4 x i32> @vec_sld(int vector[4], int vector[4], unsigned int) 875 876 void __attribute__((noinline)) 877 test_set() { 878 resi = _mm_set_epi16(ss[7], ss[6], ss[5], ss[4], ss[3], ss[2], ss[1], ss[0]); 879 resi = _mm_set_epi32(is[3], is[2], is[1], is[0]); 880 resi = _mm_set_epi64(m641, m642); 881 resi = _mm_set_epi64x(i64s[0], i64s[1]); 882 resi = _mm_set_epi8(chs[15], chs[14], chs[13], chs[12], chs[11], chs[10], chs[9], chs[8], chs[7], chs[6], chs[5], chs[4], chs[3], chs[2], chs[1], chs[0]); 883 resd = _mm_set_pd(dp[0], dp[1]); 884 resd = _mm_set_pd1(dp[0]); 885 resd = _mm_set_sd(dp[0]); 886 resi = _mm_set1_epi16(ss[0]); 887 resi = _mm_set1_epi32(i); 888 resi = _mm_set1_epi64(m641); 889 resi = _mm_set1_epi64x(i64s[0]); 890 resi = _mm_set1_epi8(chs[0]); 891 resd = _mm_set1_pd(dp[0]); 892 resi = _mm_setr_epi16(ss[7], ss[6], ss[5], ss[4], ss[3], ss[2], ss[1], ss[0]); 893 resi = _mm_setr_epi32(is[3], is[2], is[1], is[0]); 894 resi = _mm_setr_epi64(m641, m642); 895 resi = _mm_setr_epi8(chs[15], chs[14], chs[13], chs[12], chs[11], chs[10], chs[9], chs[8], chs[7], chs[6], chs[5], chs[4], chs[3], chs[2], chs[1], chs[0]); 896 resd = _mm_setr_pd(dp[0], dp[1]); 897 resd = _mm_setzero_pd(); 898 resi = _mm_setzero_si128(); 899 } 900 901 // CHECK-LABEL: @test_set 902 903 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set_epi16 904 // CHECK-COUNT-8: store i16 {{[0-9a-zA-Z_%.]+}}, i16* {{[0-9a-zA-Z_%.]+}}, align 2 905 // CHECK: insertelement <8 x i16> undef, i16 {{[0-9a-zA-Z_%.]+}}, i32 0 906 // CHECK-COUNT-7: insertelement <8 x i16> {{[0-9a-zA-Z_%.]+}}, i16 {{[0-9a-zA-Z_%.]+}}, i32 {{[1-7]}} 907 908 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set_epi32 909 // CHECK-COUNT-4: store i32 {{[0-9a-zA-Z_%.]+}}, i32* {{[0-9a-zA-Z_%.]+}}, align 4 910 // CHECK: insertelement <4 x i32> undef, i32 {{[0-9a-zA-Z_%.]+}}, i32 0 911 // CHECK-COUNT-3: insertelement <4 x i32> {{[0-9a-zA-Z_%.]+}}, i32 {{[0-9a-zA-Z_%.]+}}, i32 {{[1-3]}} 912 913 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set_epi64 914 // CHECK: call <2 x i64> @_mm_set_epi64x(i64 noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef %{{[0-9a-zA-Z_.]+}}) 915 916 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set_epi64x 917 // CHECK: %[[VEC:[0-9a-zA-Z_.]+]] = insertelement <2 x i64> undef, i64 %{{[0-9a-zA-Z_.]+}}, i32 0 918 // CHECK: insertelement <2 x i64> %[[VEC]], i64 %{{[0-9a-zA-Z_.]+}}, i32 1 919 920 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set_epi8 921 // CHECK-COUNT-16: store i8 {{[0-9a-zA-Z_%.]+}}, i8* {{[0-9a-zA-Z_%.]+}}, align 1 922 // CHECK: insertelement <16 x i8> undef, i8 {{[0-9a-zA-Z_%.]+}}, i32 {{[0-9]+}} 923 // CHECK-COUNT-15: {{[0-9a-zA-Z_%.]+}} = insertelement <16 x i8> {{[0-9a-zA-Z_%.]+}}, i8 {{[0-9a-zA-Z_%.]+}}, i32 {{[0-9]+}} 924 925 // CHECK-LABEL: define available_externally <2 x double> @_mm_set_pd 926 // CHECK: %[[VEC:[0-9a-zA-Z_.]+]] = insertelement <2 x double> undef, double %{{[0-9a-zA-Z_.]+}}, i32 0 927 // CHECK: insertelement <2 x double> %[[VEC]], double %{{[0-9a-zA-Z_.]+}}, i32 1 928 929 // CHECK-LABEL: define available_externally <2 x double> @_mm_set_pd1 930 // CHECK: call <2 x double> @_mm_set1_pd(double noundef %{{[0-9a-zA-Z_.]+}}) 931 932 // CHECK-LABEL: define available_externally <2 x double> @_mm_set_sd 933 // CHECK: %[[VEC:[0-9a-zA-Z_.]+]] = insertelement <2 x double> undef, double %{{[0-9a-zA-Z_.]+}}, i32 0 934 // CHECK: insertelement <2 x double> %[[VEC]], double 0.000000e+00, i32 1 935 936 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set1_epi16 937 // CHECK-COUNT-8: load i16, i16* %{{[0-9a-zA-Z_.]+}}, align 2 938 // CHECK: call <2 x i64> @_mm_set_epi16 939 940 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set1_epi32 941 // CHECK-COUNT-4: load i32, i32* %{{[0-9a-zA-Z_.]+}}, align 4 942 // CHECK: call <2 x i64> @_mm_set_epi32 943 944 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set1_epi64 945 // CHECK: %[[VAL1:[0-9a-zA-Z_.]+]] = load i64, i64* %{{[0-9a-zA-Z_.]+}}, align 8 946 // CHECK: %[[VAL2:[0-9a-zA-Z_.]+]] = load i64, i64* %{{[0-9a-zA-Z_.]+}}, align 8 947 // CHECK: call <2 x i64> @_mm_set_epi64(i64 noundef %[[VAL1]], i64 noundef %[[VAL2]]) 948 949 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set1_epi64x 950 // CHECK: %[[VAL1:[0-9a-zA-Z_.]+]] = load i64, i64* %{{[0-9a-zA-Z_.]+}}, align 8 951 // CHECK: %[[VAL2:[0-9a-zA-Z_.]+]] = load i64, i64* %{{[0-9a-zA-Z_.]+}}, align 8 952 // CHECK: call <2 x i64> @_mm_set_epi64x(i64 noundef %[[VAL1]], i64 noundef %[[VAL2]]) 953 954 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set1_epi8 955 // CHECK-COUNT-16: load i8, i8* %{{[0-9a-zA-Z_.]+}}, align 1 956 // CHECK: call <2 x i64> @_mm_set_epi8 957 958 // CHECK-LABEL: define available_externally <2 x double> @_mm_set1_pd 959 // CHECK: %[[VEC:[0-9a-zA-Z_.]+]] = insertelement <2 x double> undef, double %{{[0-9a-zA-Z_.]+}}, i32 0 960 // CHECK: insertelement <2 x double> %[[VEC]], double %{{[0-9a-zA-Z_.]+}}, i32 1 961 962 // CHECK-LABEL: define available_externally <2 x i64> @_mm_setr_epi16 963 // CHECK-COUNT-8: load i16, i16* {{[0-9a-zA-Z_%.]+}}, align 2 964 // CHECK: call <2 x i64> @_mm_set_epi16 965 966 // CHECK-LABEL: define available_externally <2 x i64> @_mm_setr_epi32 967 // CHECK-COUNT-4: load i32, i32* {{[0-9a-zA-Z_%.]+}}, align 4 968 // CHECK: call <2 x i64> @_mm_set_epi32 969 970 // CHECK-LABEL: define available_externally <2 x i64> @_mm_setr_epi64 971 // CHECK: %[[VAL1:[0-9a-zA-Z_.]+]] = load i64, i64* %{{[0-9a-zA-Z_.]+}}, align 8 972 // CHECK: %[[VAL2:[0-9a-zA-Z_.]+]] = load i64, i64* %{{[0-9a-zA-Z_.]+}}, align 8 973 // CHECK: call <2 x i64> @_mm_set_epi64(i64 noundef %[[VAL1]], i64 noundef %[[VAL2]]) 974 975 // CHECK-LABEL: define available_externally <2 x i64> @_mm_setr_epi8 976 // CHECK-COUNT-16: load i8, i8* {{[0-9a-zA-Z_%.]+}}, align 1 977 // CHECK: call <2 x i64> @_mm_set_epi8 978 979 // CHECK-LABEL: define available_externally <2 x double> @_mm_setr_pd 980 // CHECK: %[[VEC:[0-9a-zA-Z_.]+]] = insertelement <2 x double> undef, double %{{[0-9a-zA-Z_.]+}}, i32 0 981 // CHECK: insertelement <2 x double> %[[VEC]], double %{{[0-9a-zA-Z_.]+}}, i32 1 982 983 // CHECK-LABEL: define available_externally <2 x double> @_mm_setzero_pd() 984 // CHECK: call <4 x i32> @vec_splats(int)(i32 noundef signext 0) 985 986 // CHECK-LABEL: define available_externally <2 x i64> @_mm_setzero_si128() 987 // CHECK: store <4 x i32> zeroinitializer, <4 x i32>* %{{[0-9a-zA-Z_.]+}}, align 16 988 989 void __attribute__((noinline)) 990 test_shuffle() { 991 resi = _mm_shuffle_epi32(mi1, i); 992 resd = _mm_shuffle_pd(md1, md2, i); 993 resi = _mm_shufflehi_epi16(mi1, i); 994 resi = _mm_shufflelo_epi16(mi1, i); 995 } 996 997 // CHECK-LABEL: @test_shuffle 998 999 // CHECK-LABEL: define available_externally <2 x i64> @_mm_shuffle_epi32 1000 // CHECK: %[[AND:[0-9a-zA-Z_.]+]] = and i32 %{{[0-9a-zA-Z_.]+}}, 3 1001 // CHECK: sext i32 %[[AND]] to i64 1002 // CHECK: %[[SHR:[0-9a-zA-Z_.]+]] = ashr i32 %{{[0-9a-zA-Z_.]+}}, 2 1003 // CHECK: %[[AND2:[0-9a-zA-Z_.]+]] = and i32 %[[SHR]], 3 1004 // CHECK: sext i32 %[[AND2]] to i64 1005 // CHECK: %[[SHR2:[0-9a-zA-Z_.]+]] = ashr i32 %{{[0-9a-zA-Z_.]+}}, 4 1006 // CHECK: %[[AND3:[0-9a-zA-Z_.]+]] = and i32 %[[SHR2]], 3 1007 // CHECK: sext i32 %[[AND3]] to i64 1008 // CHECK: %[[SHR:[0-9a-zA-Z_.]+]] = ashr i32 %{{[0-9a-zA-Z_.]+}}, 6 1009 // CHECK: %[[AND4:[0-9a-zA-Z_.]+]] = and i32 %[[SHR]], 3 1010 // CHECK: sext i32 %[[AND4]] to i64 1011 // CHECK: getelementptr inbounds [4 x i32], [4 x i32]* @_mm_shuffle_epi32.__permute_selectors, i64 0, i64 %{{[0-9a-zA-Z_.]+}} 1012 // CHECK: insertelement <4 x i32> %{{[0-9a-zA-Z_.]+}}, i32 %{{[0-9a-zA-Z_.]+}}, i32 0 1013 // CHECK: getelementptr inbounds [4 x i32], [4 x i32]* @_mm_shuffle_epi32.__permute_selectors, i64 0, i64 %{{[0-9a-zA-Z_.]+}} 1014 // CHECK: insertelement <4 x i32> %{{[0-9a-zA-Z_.]+}}, i32 %{{[0-9a-zA-Z_.]+}}, i32 1 1015 // CHECK: getelementptr inbounds [4 x i32], [4 x i32]* @_mm_shuffle_epi32.__permute_selectors, i64 0, i64 %{{[0-9a-zA-Z_.]+}} 1016 // CHECK: %[[ADD:[0-9a-zA-Z_.]+]] = add i32 %{{[0-9a-zA-Z_.]+}}, 269488144 1017 // CHECK: insertelement <4 x i32> %{{[0-9a-zA-Z_.]+}}, i32 %[[ADD]], i32 2 1018 // CHECK: getelementptr inbounds [4 x i32], [4 x i32]* @_mm_shuffle_epi32.__permute_selectors, i64 0, i64 %{{[0-9a-zA-Z_.]+}} 1019 // CHECK: add i32 %{{[0-9a-zA-Z_.]+}}, 269488144 1020 // CHECK: call <4 x i32> @vec_perm(int vector[4], int vector[4], unsigned char vector[16]) 1021 1022 // CHECK-LABEL: define available_externally <2 x double> @_mm_shuffle_pd 1023 // CHECK: and i32 %{{[0-9a-zA-Z_.]+}}, 3 1024 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp eq i32 %{{[0-9a-zA-Z_.]+}}, 0 1025 // CHECK: br i1 %[[CMP]] 1026 // CHECK: call <2 x double> @vec_mergeh(double vector[2], double vector[2]) 1027 // CHECK: %[[CMP2:[0-9a-zA-Z_.]+]] = icmp eq i32 %{{[0-9a-zA-Z_.]+}}, 1 1028 // CHECK: br i1 %[[CMP2]] 1029 // CHECK: shufflevector <2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i32> <i32 1, i32 2> 1030 // CHECK: %[[CMP3:[0-9a-zA-Z_.]+]] = icmp eq i32 %{{[0-9a-zA-Z_.]+}}, 2 1031 // CHECK: br i1 %[[CMP3]] 1032 // CHECK: shufflevector <2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i32> <i32 0, i32 3> 1033 // CHECK: call <2 x double> @vec_mergel(double vector[2], double vector[2]) 1034 1035 // CHECK-LABEL: define available_externally <2 x i64> @_mm_shufflehi_epi16 1036 // CHECK: %[[AND:[0-9a-zA-Z_.]+]] = and i32 %{{[0-9a-zA-Z_.]+}}, 3 1037 // CHECK: sext i32 %[[AND]] to i64 1038 // CHECK: %[[SHR:[0-9a-zA-Z_.]+]] = ashr i32 %{{[0-9a-zA-Z_.]+}}, 2 1039 // CHECK: %[[AND2:[0-9a-zA-Z_.]+]] = and i32 %[[SHR]], 3 1040 // CHECK: sext i32 %[[AND2]] to i64 1041 // CHECK: %[[SHR2:[0-9a-zA-Z_.]+]] = ashr i32 %{{[0-9a-zA-Z_.]+}}, 4 1042 // CHECK: %[[AND3:[0-9a-zA-Z_.]+]] = and i32 %[[SHR2]], 3 1043 // CHECK: sext i32 %[[AND3]] to i64 1044 // CHECK: %[[SHR3:[0-9a-zA-Z_.]+]] = ashr i32 %{{[0-9a-zA-Z_.]+}}, 6 1045 // CHECK: %[[AND4:[0-9a-zA-Z_.]+]] = and i32 %[[SHR3]], 3 1046 // CHECK: sext i32 %[[AND4]] to i64 1047 // CHECK-LE: store <2 x i64> <i64 1663540288323457296, i64 0>, <2 x i64>* %{{[0-9a-zA-Z_.]+}}, align 16 1048 // CHECK-BE: store <2 x i64> <i64 1157726452361532951, i64 0>, <2 x i64>* %{{[0-9a-zA-Z_.]+}}, align 16 1049 // CHECK-COUNT-4: getelementptr inbounds [4 x i16], [4 x i16]* @_mm_shufflehi_epi16.__permute_selectors, i64 0, i64 {{[0-9a-zA-Z_%.]+}} 1050 // CHECK: call <2 x i64> @vec_perm(unsigned long long vector[2], unsigned long long vector[2], unsigned char vector[16]) 1051 1052 // CHECK-LABEL: define available_externally <2 x i64> @_mm_shufflelo_epi16 1053 // CHECK: %[[AND:[0-9a-zA-Z_.]+]] = and i32 {{[0-9a-zA-Z_%.]+}}, 3 1054 // CHECK: sext i32 %[[AND]] to i64 1055 // CHECK: %[[SHR:[0-9a-zA-Z_.]+]] = ashr i32 {{[0-9a-zA-Z_%.]+}}, 2 1056 // CHECK: %[[AND2:[0-9a-zA-Z_.]+]] = and i32 %[[SHR]], 3 1057 // CHECK: sext i32 %[[AND2]] to i64 1058 // CHECK: %[[SHR2:[0-9a-zA-Z_.]+]] = ashr i32 {{[0-9a-zA-Z_%.]+}}, 4 1059 // CHECK: %[[AND3:[0-9a-zA-Z_.]+]] = and i32 %[[SHR2]], 3 1060 // CHECK: sext i32 %[[AND3]] to i64 1061 // CHECK: %[[SHR3:[0-9a-zA-Z_.]+]] = ashr i32 {{[0-9a-zA-Z_%.]+}}, 6 1062 // CHECK: %[[AND4:[0-9a-zA-Z_.]+]] = and i32 %[[SHR3]], 3 1063 // CHECK: sext i32 %[[AND4]] to i64 1064 // CHECK-LE: store <2 x i64> <i64 0, i64 2242261671028070680>, <2 x i64>* %{{[0-9a-zA-Z_.]+}}, align 16 1065 // CHECK-BE: store <2 x i64> <i64 0, i64 1736447835066146335>, <2 x i64>* %{{[0-9a-zA-Z_.]+}}, align 16 1066 // CHECK-COUNT-4: getelementptr inbounds [4 x i16], [4 x i16]* @_mm_shufflelo_epi16.__permute_selectors, i64 0, i64 {{[0-9a-zA-Z_%.]+}} 1067 // CHECK: call <2 x i64> @vec_perm(unsigned long long vector[2], unsigned long long vector[2], unsigned char vector[16]) 1068 1069 void __attribute__((noinline)) 1070 test_sll() { 1071 resi = _mm_sll_epi16(mi1, mi2); 1072 resi = _mm_sll_epi32(mi1, mi2); 1073 resi = _mm_sll_epi64(mi1, mi2); 1074 resi = _mm_slli_epi16(mi1, i); 1075 resi = _mm_slli_epi32(mi1, i); 1076 resi = _mm_slli_epi64(mi1, i); 1077 resi = _mm_slli_si128(mi1, i); 1078 } 1079 1080 // CHECK-LABEL: @test_sll 1081 1082 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sll_epi16 1083 // CHECK: store <8 x i16> <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>, <8 x i16>* %{{[0-9a-zA-Z_.]+}}, align 16 1084 // CHECK-LE: call <8 x i16> @vec_splat(unsigned short vector[8], unsigned int) 1085 // CHECK-BE: call <8 x i16> @vec_splat(unsigned short vector[8], unsigned int) 1086 // CHECK: call <8 x i16> @vec_cmple(unsigned short vector[8], unsigned short vector[8])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>) 1087 // CHECK: call <8 x i16> @vec_sl(unsigned short vector[8], unsigned short vector[8]) 1088 // CHECK: call <8 x i16> @vec_sel(unsigned short vector[8], unsigned short vector[8], bool vector[8]) 1089 1090 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sll_epi32 1091 // CHECK-LE: call <4 x i32> @vec_splat(unsigned int vector[4], unsigned int)(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef zeroext 0) 1092 // CHECK-BE: call <4 x i32> @vec_splat(unsigned int vector[4], unsigned int)(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef zeroext 1) 1093 // CHECK: call <4 x i32> @vec_cmplt(unsigned int vector[4], unsigned int vector[4])(<4 x i32> noundef {{[0-9a-zA-Z_%.]+}}, <4 x i32> noundef <i32 32, i32 32, i32 32, i32 32>) 1094 // CHECK: call <4 x i32> @vec_sl(unsigned int vector[4], unsigned int vector[4]) 1095 // CHECK: call <4 x i32> @vec_sel(unsigned int vector[4], unsigned int vector[4], bool vector[4]) 1096 1097 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sll_epi64 1098 // CHECK: call <2 x i64> @vec_splat(unsigned long long vector[2], unsigned int)(<2 x i64> noundef {{[0-9a-zA-Z_%.]+}}, i32 noundef zeroext 0) 1099 // CHECK: call <2 x i64> @vec_cmplt(unsigned long long vector[2], unsigned long long vector[2])(<2 x i64> noundef {{[0-9a-zA-Z_%.]+}}, <2 x i64> noundef <i64 64, i64 64>) 1100 // CHECK: call <2 x i64> @vec_sl(unsigned long long vector[2], unsigned long long vector[2]) 1101 // CHECK: call <2 x i64> @vec_sel(unsigned long long vector[2], unsigned long long vector[2], bool vector[2]) 1102 1103 // CHECK-LABEL: define available_externally <2 x i64> @_mm_slli_epi16 1104 // CHECK: store <8 x i16> zeroinitializer, <8 x i16>* %{{[0-9a-zA-Z_.]+}}, align 16 1105 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp sge i32 %{{[0-9a-zA-Z_.]+}}, 0 1106 // CHECK: br i1 %[[CMP]] 1107 // CHECK: %[[CMP2:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16 1108 // CHECK: br i1 %[[CMP2]] 1109 // CHECK: call i1 @llvm.is.constant 1110 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i8 1111 // CHECK: call <8 x i16> @vec_splat_s16(signed char)(i8 noundef signext %[[TRUNC]]) 1112 // CHECK: %[[TRUNC2:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i16 1113 // CHECK: call <8 x i16> @vec_splats(unsigned short)(i16 noundef zeroext %[[TRUNC2]]) 1114 // CHECK: call <8 x i16> @vec_sl(short vector[8], unsigned short vector[8]) 1115 1116 // CHECK-LABEL: define available_externally <2 x i64> @_mm_slli_epi32 1117 // CHECK: store <4 x i32> zeroinitializer, <4 x i32>* %{{[0-9a-zA-Z_.]+}}, align 16 1118 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp sge i32 %{{[0-9a-zA-Z_.]+}}, 0 1119 // CHECK: br i1 %[[CMP]] 1120 // CHECK: %[[CMP2:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 32 1121 // CHECK: br i1 %[[CMP2]] 1122 // CHECK: call i1 @llvm.is.constant 1123 // CHECK: %[[CMP3:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16 1124 // CHECK: br i1 %[[CMP3]] 1125 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i8 1126 // CHECK: call <4 x i32> @vec_splat_s32(signed char)(i8 noundef signext %[[TRUNC]]) 1127 // CHECK: call <4 x i32> @vec_splats(unsigned int) 1128 // CHECK: call <4 x i32> @vec_sl(int vector[4], unsigned int vector[4]) 1129 1130 // CHECK-LABEL: define available_externally <2 x i64> @_mm_slli_epi64 1131 // CHECK: store <2 x i64> zeroinitializer, <2 x i64>* %{{[0-9a-zA-Z_.]+}}, align 16 1132 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp sge i32 %{{[0-9a-zA-Z_.]+}}, 0 1133 // CHECK: br i1 %[[CMP]] 1134 // CHECK: %[[CMP2:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 64 1135 // CHECK: br i1 %[[CMP2]] 1136 // CHECK: call i1 @llvm.is.constant 1137 // CHECK: %[[CMP3:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16 1138 // CHECK: br i1 %[[CMP3]] 1139 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i8 1140 // CHECK: call <4 x i32> @vec_splat_s32(signed char)(i8 noundef signext %[[TRUNC]]) 1141 // CHECK: call <4 x i32> @vec_splats(unsigned int) 1142 // CHECK: call <2 x i64> @vec_sl(long long vector[2], unsigned long long vector[2]) 1143 1144 // CHECK-LABEL: define available_externally <2 x i64> @_mm_slli_si128 1145 // CHECK: store <16 x i8> zeroinitializer, <16 x i8>* %{{[0-9a-zA-Z_.]+}}, align 16 1146 // CHECK-BE: %[[SUB:[0-9a-zA-Z_.]+]] = sub nsw i32 16, %{{[0-9a-zA-Z_.]+}} 1147 // CHECK-BE: call <16 x i8> @vec_sld(unsigned char vector[16], unsigned char vector[16], unsigned int)(<16 x i8> noundef zeroinitializer, <16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef zeroext %[[SUB]]) 1148 // CHECK-LE: call <16 x i8> @vec_sld(unsigned char vector[16], unsigned char vector[16], unsigned int)(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef zeroinitializer, i32 noundef zeroext %{{[0-9a-zA-Z_.]+}}) 1149 // CHECK: store <16 x i8> zeroinitializer, <16 x i8>* %{{[0-9a-zA-Z_.]+}}, align 16 1150 1151 void __attribute__((noinline)) 1152 test_sqrt() { 1153 resd = _mm_sqrt_pd(md1); 1154 resd = _mm_sqrt_sd(md1, md2); 1155 } 1156 1157 // CHECK-LABEL: @test_sqrt 1158 1159 // CHECK-LABEL: define available_externally <2 x double> @_mm_sqrt_pd 1160 // CHECK: call <2 x double> @vec_sqrt(double vector[2])(<2 x double> noundef {{[0-9a-zA-Z_%.]+}}) 1161 1162 // CHECK-LABEL: define available_externally <2 x double> @_mm_sqrt_sd 1163 // CHECK: %[[CALL:[0-9a-zA-Z_.]+]] = call <2 x double> @_mm_set1_pd(double noundef %{{[0-9a-zA-Z_.]+}}) 1164 // CHECK: call <2 x double> @vec_sqrt(double vector[2])(<2 x double> noundef %{{[0-9a-zA-Z_.]+}}) 1165 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}}) 1166 1167 void __attribute__((noinline)) 1168 test_sra() { 1169 resi = _mm_sra_epi16(mi1, mi2); 1170 resi = _mm_sra_epi32(mi1, mi2); 1171 resi = _mm_srai_epi16(mi1, i); 1172 resi = _mm_srai_epi32(mi1, i); 1173 } 1174 1175 // CHECK-LABEL: @test_sra 1176 1177 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sra_epi16 1178 // CHECK: store <8 x i16> <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>, <8 x i16>* %{{[0-9a-zA-Z_.]+}}, align 16 1179 // CHECK-LE: call <8 x i16> @vec_splat(unsigned short vector[8], unsigned int)(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef zeroext 0) 1180 // CHECK-BE: call <8 x i16> @vec_splat(unsigned short vector[8], unsigned int)(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef zeroext 3) 1181 // CHECK: call <8 x i16> @vec_min(unsigned short vector[8], unsigned short vector[8])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>) 1182 // CHECK: call <8 x i16> @vec_sra(short vector[8], unsigned short vector[8]) 1183 1184 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sra_epi32 1185 // CHECK: store <4 x i32> <i32 31, i32 31, i32 31, i32 31>, <4 x i32>* %{{[0-9a-zA-Z_.]+}}, align 16 1186 // CHECK-LE: call <4 x i32> @vec_splat(unsigned int vector[4], unsigned int)(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef zeroext 0) 1187 // CHECK-BE: call <4 x i32> @vec_splat(unsigned int vector[4], unsigned int)(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef zeroext 1) 1188 // CHECK: call <4 x i32> @vec_min(unsigned int vector[4], unsigned int vector[4])(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <4 x i32> noundef <i32 31, i32 31, i32 31, i32 31>) 1189 // CHECK: call <4 x i32> @vec_sra(int vector[4], unsigned int vector[4]) 1190 1191 // CHECK-LABEL: define available_externally <2 x i64> @_mm_srai_epi16 1192 // CHECK: store <8 x i16> <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>, <8 x i16>* %{{[0-9a-zA-Z_.]+}}, align 16 1193 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16 1194 // CHECK: br i1 %[[CMP]] 1195 // CHECK: call i1 @llvm.is.constant 1196 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i8 1197 // CHECK: call <8 x i16> @vec_splat_s16(signed char)(i8 noundef signext %[[TRUNC]]) 1198 // CHECK: %[[TRUNC2:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i16 1199 // CHECK: call <8 x i16> @vec_splats(unsigned short)(i16 noundef zeroext %{{[0-9a-zA-Z_.]+}}) 1200 // CHECK: call <8 x i16> @vec_sra(short vector[8], unsigned short vector[8]) 1201 1202 // CHECK-LABEL: define available_externally <2 x i64> @_mm_srai_epi32 1203 // CHECK: store <4 x i32> <i32 31, i32 31, i32 31, i32 31>, <4 x i32>* %{{[0-9a-zA-Z_.]+}}, align 16 1204 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 32 1205 // CHECK: br i1 %[[CMP]] 1206 // CHECK: call i1 @llvm.is.constant 1207 // CHECK: %[[CMP2:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16 1208 // CHECK: br i1 %[[CMP2]] 1209 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i8 1210 // CHECK: call <4 x i32> @vec_splat_s32(signed char)(i8 noundef signext %[[TRUNC]]) 1211 // CHECK: call <4 x i32> @vec_splats(unsigned int) 1212 // CHECK: call <4 x i32> @vec_splats(unsigned int) 1213 // CHECK: call <4 x i32> @vec_sra(int vector[4], unsigned int vector[4]) 1214 1215 void __attribute__((noinline)) 1216 test_srl() { 1217 resi = _mm_srl_epi16(mi1, mi2); 1218 resi = _mm_srl_epi32(mi1, mi2); 1219 resi = _mm_srl_epi64(mi1, mi2); 1220 resi = _mm_srli_epi16(mi1, i); 1221 resi = _mm_srli_epi32(mi1, i); 1222 resi = _mm_srli_epi64(mi1, i); 1223 resi = _mm_srli_si128(mi1, i); 1224 } 1225 1226 // CHECK-LABEL: @test_srl 1227 1228 // CHECK-LABEL: define available_externally <2 x i64> @_mm_srl_epi16 1229 // CHECK: store <8 x i16> <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>, <8 x i16>* %{{[0-9a-zA-Z_.]+}}, align 16 1230 // CHECK-LE: call <8 x i16> @vec_splat(unsigned short vector[8], unsigned int)(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef zeroext 0) 1231 // CHECK-BE: call <8 x i16> @vec_splat(unsigned short vector[8], unsigned int)(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef zeroext 3) 1232 // CHECK: call <8 x i16> @vec_cmple(unsigned short vector[8], unsigned short vector[8])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>) 1233 // CHECK: call <8 x i16> @vec_sr(unsigned short vector[8], unsigned short vector[8]) 1234 // CHECK: call <8 x i16> @vec_sel(unsigned short vector[8], unsigned short vector[8], bool vector[8]) 1235 1236 // CHECK-LABEL: define available_externally <2 x i64> @_mm_srl_epi32 1237 // CHECK-LE: call <4 x i32> @vec_splat(unsigned int vector[4], unsigned int)(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef zeroext 0) 1238 // CHECK-BE: call <4 x i32> @vec_splat(unsigned int vector[4], unsigned int)(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef zeroext 1) 1239 // CHECK: call <4 x i32> @vec_cmplt(unsigned int vector[4], unsigned int vector[4])(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <4 x i32> noundef <i32 32, i32 32, i32 32, i32 32>) 1240 // CHECK: call <4 x i32> @vec_sr(unsigned int vector[4], unsigned int vector[4]) 1241 // CHECK: call <4 x i32> @vec_sel(unsigned int vector[4], unsigned int vector[4], bool vector[4]) 1242 1243 // CHECK-LABEL: define available_externally <2 x i64> @_mm_srl_epi64 1244 // CHECK: call <2 x i64> @vec_splat(unsigned long long vector[2], unsigned int)(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef zeroext 0) 1245 // CHECK: call <2 x i64> @vec_cmplt(unsigned long long vector[2], unsigned long long vector[2])(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef <i64 64, i64 64>) 1246 // CHECK: call <2 x i64> @vec_sr(unsigned long long vector[2], unsigned long long vector[2]) 1247 // CHECK: call <2 x i64> @vec_sel(unsigned long long vector[2], unsigned long long vector[2], bool vector[2]) 1248 1249 // CHECK-LABEL: define available_externally <2 x i64> @_mm_srli_epi16 1250 // CHECK: store <8 x i16> zeroinitializer, <8 x i16>* %{{[0-9a-zA-Z_.]+}}, align 16 1251 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16 1252 // CHECK: br i1 %[[CMP]] 1253 // CHECK: call i1 @llvm.is.constant 1254 // CHECK: trunc i32 %{{[0-9a-zA-Z_.]+}} to i8 1255 // CHECK: call <8 x i16> @vec_splat_s16(signed char) 1256 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i16 1257 // CHECK: call <8 x i16> @vec_splats(unsigned short)(i16 noundef zeroext %[[TRUNC]]) 1258 // CHECK: call <8 x i16> @vec_sr(short vector[8], unsigned short vector[8]) 1259 1260 // CHECK-LABEL: define available_externally <2 x i64> @_mm_srli_epi32 1261 // CHECK: store <4 x i32> zeroinitializer, <4 x i32>* %{{[0-9a-zA-Z_.]+}}, align 16 1262 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 32 1263 // CHECK: br i1 %[[CMP]] 1264 // CHECK: call i1 @llvm.is.constant 1265 // CHECK: %[[CMP2:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16 1266 // CHECK: br i1 %[[CMP2]] 1267 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i8 1268 // CHECK: call <4 x i32> @vec_splat_s32(signed char) 1269 // CHECK: call <4 x i32> @vec_splats(unsigned int) 1270 // CHECK: call <4 x i32> @vec_splats(unsigned int) 1271 // CHECK: call <4 x i32> @vec_sr(int vector[4], unsigned int vector[4]) 1272 1273 // CHECK-LABEL: define available_externally <2 x i64> @_mm_srli_epi64 1274 // CHECK: store <2 x i64> zeroinitializer, <2 x i64>* %{{[0-9a-zA-Z_.]+}}, align 16 1275 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 64 1276 // CHECK: br i1 %[[CMP]] 1277 // CHECK: call i1 @llvm.is.constant 1278 // CHECK: %[[CMP2:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16 1279 // CHECK: br i1 %[[CMP2]] 1280 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i8 1281 // CHECK: call <4 x i32> @vec_splat_s32(signed char)(i8 noundef signext %[[TRUNC]]) 1282 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = sext i32 %{{[0-9a-zA-Z_.]+}} to i64 1283 // CHECK: call <2 x i64> @vec_splats(unsigned long long)(i64 noundef %[[EXT]]) 1284 // CHECK: call <4 x i32> @vec_splats(unsigned int)(i32 noundef zeroext %{{[0-9a-zA-Z_.]+}}) 1285 // CHECK: call <2 x i64> @vec_sr(long long vector[2], unsigned long long vector[2]) 1286 1287 // CHECK-LABEL: define available_externally <2 x i64> @_mm_srli_si128 1288 // CHECK: call <2 x i64> @_mm_bsrli_si128 1289 1290 void __attribute__((noinline)) 1291 test_store() { 1292 _mm_store_pd(dp, md1); 1293 _mm_store_pd1(dp, md1); 1294 _mm_store_sd(dp, md1); 1295 _mm_store_si128(mip, mi1); 1296 _mm_store1_pd(dp, md1); 1297 _mm_storeh_pd(dp, md1); 1298 _mm_storel_epi64(mip, mi1); 1299 _mm_storel_pd(dp, md1); 1300 _mm_storer_pd(dp, md1); 1301 _mm_storeu_pd(dp, md1); 1302 _mm_storeu_si128(mip, mi1); 1303 } 1304 1305 // CHECK-LABEL: @test_store 1306 1307 // CHECK-LABEL: define available_externally void @_mm_store_pd 1308 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load double*, double** %{{[0-9a-zA-Z_.]+}}, align 8 1309 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast double* %[[ADDR]] to <16 x i8>* 1310 // CHECK: call void @vec_st(unsigned char vector[16], long, unsigned char vector[16]*)(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef 0, <16 x i8>* noundef %[[CAST]]) 1311 1312 // CHECK-LABEL: define available_externally void @_mm_store_pd1 1313 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load double*, double** %{{[0-9a-zA-Z_.]+}}, align 8 1314 // CHECK: %[[ADDR2:[0-9a-zA-Z_.]+]] = load <2 x double>, <2 x double>* %{{[0-9a-zA-Z_.]+}}, align 16 1315 // CHECK: call void @_mm_store1_pd(double* noundef %[[ADDR]], <2 x double> noundef %[[ADDR2]]) 1316 1317 // CHECK-LABEL: define available_externally void @_mm_store_sd 1318 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load double*, double** %{{[0-9a-zA-Z_.]+}}, align 8 1319 // CHECK: store double %{{[0-9a-zA-Z_.]+}}, double* %[[ADDR]], align 8 1320 1321 // CHECK-LABEL: define available_externally void @_mm_store_si128 1322 // CHECK: %[[LOAD:[0-9a-zA-Z_.]+]] = load <2 x i64>*, <2 x i64>** %{{[0-9a-zA-Z_.]+}}, align 8 1323 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <2 x i64>* %[[LOAD]] to <16 x i8>* 1324 // CHECK: call void @vec_st(unsigned char vector[16], long, unsigned char vector[16]*)(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef 0, <16 x i8>* noundef %[[CAST]]) 1325 1326 // CHECK-LABEL: define available_externally void @_mm_store1_pd 1327 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load double*, double** %{{[0-9a-zA-Z_.]+}}, align 8 1328 // CHECK: %[[VAL:[0-9a-zA-Z_.]+]] = load <2 x double>, <2 x double>* %{{[0-9a-zA-Z_.]+}}, align 16 1329 // CHECK: %[[CALL:[0-9a-zA-Z_.]+]] = call <2 x double> @vec_splat(double vector[2], unsigned int)(<2 x double> noundef %[[VAL]], i32 noundef zeroext 0) 1330 // CHECK: call void @_mm_store_pd(double* noundef %[[ADDR]], <2 x double> noundef %[[CALL]]) 1331 1332 // CHECK-LABEL: define available_externally void @_mm_storeh_pd 1333 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load double*, double** %{{[0-9a-zA-Z_.]+}}, align 8 1334 // CHECK: store double %{{[0-9a-zA-Z_.]+}}, double* %[[ADDR]], align 8 1335 1336 // CHECK-LABEL: define available_externally void @_mm_storel_epi64 1337 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load <2 x i64>*, <2 x i64>** %{{[0-9a-zA-Z_.]+}}, align 8 1338 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <2 x i64>* %[[ADDR]] to i64* 1339 // CHECK: store i64 %{{[0-9a-zA-Z_.]+}}, i64* %[[CAST]], align 8 1340 1341 // CHECK-LABEL: define available_externally void @_mm_storel_pd 1342 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load double*, double** %{{[0-9a-zA-Z_.]+}}, align 8 1343 // CHECK: %[[VAL:[0-9a-zA-Z_.]+]] = load <2 x double>, <2 x double>* %{{[0-9a-zA-Z_.]+}}, align 16 1344 // CHECK: call void @_mm_store_sd(double* noundef %[[ADDR]], <2 x double> noundef %[[VAL]]) 1345 1346 // CHECK-LABEL: define available_externally void @_mm_storer_pd 1347 // CHECK: shufflevector <2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i32> <i32 1, i32 2> 1348 // CHECK: call void @_mm_store_pd(double* noundef %{{[0-9a-zA-Z_.]+}}, <2 x double> noundef %{{[0-9a-zA-Z_.]+}}) 1349 1350 // CHECK-LABEL: define available_externally void @_mm_storeu_pd 1351 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load double*, double** %{{[0-9a-zA-Z_.]+}}, align 8 1352 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast double* %[[ADDR]] to <2 x double>* 1353 // CHECK: store <2 x double> %{{[0-9a-zA-Z_.]+}}, <2 x double>* %[[CAST]], align 1 1354 1355 // CHECK-LABEL: define available_externally void @_mm_storeu_si128 1356 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load <2 x i64>*, <2 x i64>** %{{[0-9a-zA-Z_.]+}}, align 8 1357 // CHECK: store <2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i64>* %[[ADDR]], align 1 1358 1359 void __attribute__((noinline)) 1360 test_stream() { 1361 _mm_stream_pd(dp, md1); 1362 _mm_stream_si128(mip, mi1); 1363 _mm_stream_si32(is, i); 1364 _mm_stream_si64(i64s, i64s[1]); 1365 } 1366 1367 // CHECK-LABEL: @test_stream 1368 1369 // CHECK-LABEL: define available_externally void @_mm_stream_pd 1370 // CHECK: call void asm sideeffect "dcbtstt 0,$0", "b,~{memory}"(double* %{{[0-9a-zA-Z_.]+}}) 1371 1372 // CHECK-LABEL: define available_externally void @_mm_stream_si128 1373 // CHECK: call void asm sideeffect "dcbtstt 0,$0", "b,~{memory}"(<2 x i64>* %{{[0-9a-zA-Z_.]+}}) 1374 1375 // CHECK-LABEL: define available_externally void @_mm_stream_si32 1376 // CHECK: call void asm sideeffect "dcbtstt 0,$0", "b,~{memory}"(i32* %{{[0-9a-zA-Z_.]+}}) 1377 1378 // CHECK-LABEL: define available_externally void @_mm_stream_si64 1379 // CHECK: call void asm sideeffect "\09dcbtstt\090,$0", "b,~{memory}"(i64* %{{[0-9a-zA-Z_.]+}}) 1380 1381 void __attribute__((noinline)) 1382 test_sub() { 1383 resi = _mm_sub_epi64(mi1, mi2); 1384 resi = _mm_sub_epi32(mi1, mi2); 1385 resi = _mm_sub_epi16(mi1, mi2); 1386 resi = _mm_sub_epi8(mi1, mi2); 1387 resd = _mm_sub_pd(md1, md2); 1388 resd = _mm_sub_sd(md1, md2); 1389 res64 = _mm_sub_si64(m641, m642); 1390 resi = _mm_subs_epi16(mi1, mi2); 1391 resi = _mm_subs_epi8(mi1, mi2); 1392 resi = _mm_subs_epu16(mi1, mi2); 1393 resi = _mm_subs_epu8(mi1, mi2); 1394 } 1395 1396 // CHECK-LABEL: @test_sub 1397 1398 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sub_epi64 1399 // CHECK: sub <2 x i64> 1400 1401 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sub_epi32 1402 // CHECK: sub <4 x i32> 1403 1404 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sub_epi16 1405 // CHECK: sub <8 x i16> 1406 1407 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sub_epi8 1408 // CHECK: sub <16 x i8> 1409 1410 // CHECK-LABEL: define available_externally <2 x double> @_mm_sub_pd 1411 // CHECK: fsub <2 x double> 1412 1413 // CHECK-LABEL: define available_externally <2 x double> @_mm_sub_sd 1414 // CHECK: fsub double 1415 1416 // CHECK-LABEL: define available_externally i64 @_mm_sub_si64 1417 // CHECK: sub i64 1418 1419 // CHECK-LABEL: define available_externally <2 x i64> @_mm_subs_epi16 1420 // CHECK: call <8 x i16> @vec_subs(short vector[8], short vector[8]) 1421 1422 // CHECK-LABEL: define available_externally <2 x i64> @_mm_subs_epi8 1423 // CHECK: call <16 x i8> @vec_subs(signed char vector[16], signed char vector[16]) 1424 1425 // CHECK-LABEL: define available_externally <2 x i64> @_mm_subs_epu16 1426 // CHECK: call <8 x i16> @vec_subs(unsigned short vector[8], unsigned short vector[8]) 1427 1428 // CHECK-LABEL: define available_externally <2 x i64> @_mm_subs_epu8 1429 // CHECK: call <16 x i8> @vec_subs(unsigned char vector[16], unsigned char vector[16]) 1430 1431 void __attribute__((noinline)) 1432 test_ucomi() { 1433 i = _mm_ucomieq_sd(md1, md2); 1434 i = _mm_ucomige_sd(md1, md2); 1435 i = _mm_ucomigt_sd(md1, md2); 1436 i = _mm_ucomile_sd(md1, md2); 1437 i = _mm_ucomilt_sd(md1, md2); 1438 i = _mm_ucomineq_sd(md1, md2); 1439 } 1440 1441 // CHECK-LABEL: @test_ucomi 1442 1443 // CHECK-LABEL: define available_externally signext i32 @_mm_ucomieq_sd 1444 // CHECK: fcmp oeq double 1445 1446 // CHECK-LABEL: define available_externally signext i32 @_mm_ucomige_sd 1447 // CHECK: fcmp oge double 1448 1449 // CHECK-LABEL: define available_externally signext i32 @_mm_ucomigt_sd 1450 // CHECK: fcmp ogt double 1451 1452 // CHECK-LABEL: define available_externally signext i32 @_mm_ucomile_sd 1453 // CHECK: fcmp ole double 1454 1455 // CHECK-LABEL: define available_externally signext i32 @_mm_ucomilt_sd 1456 // CHECK: fcmp olt double 1457 1458 // CHECK-LABEL: define available_externally signext i32 @_mm_ucomineq_sd 1459 // CHECK: fcmp une double 1460 1461 void __attribute__((noinline)) 1462 test_undefined() { 1463 resd = _mm_undefined_pd(); 1464 resi = _mm_undefined_si128(); 1465 } 1466 1467 // CHECK-LABEL: @test_undefined 1468 1469 // CHECK-LABEL: define available_externally <2 x double> @_mm_undefined_pd() 1470 // CHECK: %[[VAL:[0-9a-zA-Z_.]+]] = load <2 x double>, <2 x double>* %[[ADDR:[0-9a-zA-Z_.]+]], align 16 1471 // CHECK: store <2 x double> %[[VAL]], <2 x double>* %[[ADDR]], align 16 1472 // CHECK: load <2 x double>, <2 x double>* %[[ADDR]], align 16 1473 1474 // CHECK-LABEL: define available_externally <2 x i64> @_mm_undefined_si128() 1475 // CHECK: %[[VAL:[0-9a-zA-Z_.]+]] = load <2 x i64>, <2 x i64>* %[[ADDR:[0-9a-zA-Z_.]+]], align 16 1476 // CHECK: store <2 x i64> %[[VAL]], <2 x i64>* %[[ADDR]], align 16 1477 // CHECK: load <2 x i64>, <2 x i64>* %[[ADDR]], align 16 1478 1479 void __attribute__((noinline)) 1480 test_unpack() { 1481 resi = _mm_unpackhi_epi16(mi1, mi2); 1482 resi = _mm_unpackhi_epi32(mi1, mi2); 1483 resi = _mm_unpackhi_epi64(mi1, mi2); 1484 resi = _mm_unpackhi_epi8(mi1, mi2); 1485 resd = _mm_unpackhi_pd(md1, md2); 1486 resi = _mm_unpacklo_epi16(mi1, mi2); 1487 resi = _mm_unpacklo_epi32(mi1, mi2); 1488 resi = _mm_unpacklo_epi64(mi1, mi2); 1489 resi = _mm_unpacklo_epi8(mi1, mi2); 1490 resd = _mm_unpacklo_pd(md1, md2); 1491 } 1492 1493 // CHECK-LABEL: @test_unpack 1494 1495 // CHECK-LABEL: define available_externally <2 x i64> @_mm_unpackhi_epi16 1496 // CHECK: call <8 x i16> @vec_mergel(unsigned short vector[8], unsigned short vector[8]) 1497 1498 // CHECK-LABEL: define available_externally <2 x i64> @_mm_unpackhi_epi32 1499 // CHECK: call <4 x i32> @vec_mergel(unsigned int vector[4], unsigned int vector[4]) 1500 1501 // CHECK-LABEL: define available_externally <2 x i64> @_mm_unpackhi_epi64 1502 // CHECK: call <2 x i64> @vec_mergel(long long vector[2], long long vector[2]) 1503 1504 // CHECK-LABEL: define available_externally <2 x i64> @_mm_unpackhi_epi8 1505 // CHECK: call <16 x i8> @vec_mergel(unsigned char vector[16], unsigned char vector[16]) 1506 1507 // CHECK-LABEL: define available_externally <2 x double> @_mm_unpackhi_pd 1508 // CHECK: call <2 x double> @vec_mergel(double vector[2], double vector[2]) 1509 1510 // CHECK-LABEL: define available_externally <2 x i64> @_mm_unpacklo_epi16 1511 // CHECK: call <8 x i16> @vec_mergeh(short vector[8], short vector[8]) 1512 1513 // CHECK-LABEL: define available_externally <2 x i64> @_mm_unpacklo_epi32 1514 // CHECK: call <4 x i32> @vec_mergeh(int vector[4], int vector[4]) 1515 1516 // CHECK-LABEL: define available_externally <2 x i64> @_mm_unpacklo_epi64 1517 // CHECK: call <2 x i64> @vec_mergeh(long long vector[2], long long vector[2]) 1518 1519 // CHECK-LABEL: define available_externally <2 x i64> @_mm_unpacklo_epi8 1520 // CHECK: call <16 x i8> @vec_mergeh(unsigned char vector[16], unsigned char vector[16]) 1521 1522 // CHECK-LABEL: define available_externally <2 x double> @_mm_unpacklo_pd 1523 // CHECK: call <2 x double> @vec_mergeh(double vector[2], double vector[2]) 1524