clang/test/CodeGen/PowerPC/ppc-mmintrin.c

   1 // REQUIRES: powerpc-registered-target
   2
   3 // RUN: %clang -S -emit-llvm -target powerpc64-unknown-linux-gnu -mcpu=pwr8 -DNO_WARN_X86_INTRINSICS %s \
   4 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK-P8,CHECK,CHECK-BE
   5 // RUN: %clang -S -emit-llvm -target powerpc64le-unknown-linux-gnu -mcpu=pwr8 -DNO_WARN_X86_INTRINSICS %s \
   6 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK-P8,CHECK,CHECK-LE
   7 // RUN: %clang -S -emit-llvm -target powerpc64-unknown-linux-gnu -mcpu=pwr9 -DNO_WARN_X86_INTRINSICS %s \
   8 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK-P9,CHECK,CHECK-BE
   9 // RUN: %clang -S -emit-llvm -target powerpc64le-unknown-linux-gnu -mcpu=pwr9 -DNO_WARN_X86_INTRINSICS %s \
  10 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n| FileCheck %s --check-prefixes=CHECK-P9,CHECK,CHECK-LE
  11
  12 // RUN: %clang -S -emit-llvm -target powerpc64-unknown-freebsd13.0 -mcpu=pwr8 -DNO_WARN_X86_INTRINSICS %s \
  13 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK-P8,CHECK,CHECK-BE
  14 // RUN: %clang -S -emit-llvm -target powerpc64le-unknown-freebsd13.0 -mcpu=pwr8 -DNO_WARN_X86_INTRINSICS %s \
  15 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK-P8,CHECK,CHECK-LE
  16 // RUN: %clang -S -emit-llvm -target powerpc64-unknown-freebsd13.0 -mcpu=pwr9 -DNO_WARN_X86_INTRINSICS %s \
  17 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK-P9,CHECK,CHECK-BE
  18 // RUN: %clang -S -emit-llvm -target powerpc64le-unknown-freebsd13.0 -mcpu=pwr9 -DNO_WARN_X86_INTRINSICS %s \
  19 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n| FileCheck %s --check-prefixes=CHECK-P9,CHECK,CHECK-LE
  20
  21 // RUN: %clang -S -emit-llvm -target powerpc64-ibm-aix -mcpu=pwr8 -DNO_WARN_X86_INTRINSICS %s \
  22 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK-P8,CHECK,CHECK-BE
  23 // RUN: %clang -S -emit-llvm -target powerpc64-ibm-aix -mcpu=pwr9 -DNO_WARN_X86_INTRINSICS %s \
  24 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK-P9,CHECK,CHECK-BE
  25
  26 #include <mmintrin.h>
  27
  28 unsigned long long int ull1, ull2;
  29 int i1, i2;
  30 short s[4];
  31 signed char c[8];
  32 long long int ll1;
  33 __m64 m1, m2, res;
  34
  35 void __attribute__((noinline))
  36 test_add() {
  37   res = _mm_add_pi32(m1, m2);
  38   res = _mm_add_pi16(m1, m2);
  39   res = _mm_add_pi8(m1, m2);
  40   res = _mm_adds_pu16(m1, m2);
  41   res = _mm_adds_pu8(m1, m2);
  42   res = _mm_adds_pi16(m1, m2);
  43   res = _mm_adds_pi8(m1, m2);
  44 }
  45
  46 // CHECK-LABEL: @test_add
  47
  48 // CHECK-LABEL: define available_externally i64 @_mm_add_pi32
  49 // CHECK-P9: call <2 x i64> @vec_splats(unsigned long long)
  50 // CHECK-P9: call <2 x i64> @vec_splats(unsigned long long)
  51 // CHECK-P9: call <4 x i32> @vec_add(int vector[4], int vector[4])
  52 // CHECK-P8: add nsw i32 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
  53 // CHECK-P8: add nsw i32 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
  54
  55 // CHECK-LABEL: define available_externally i64 @_mm_add_pi16
  56 // CHECK: call <2 x i64> @vec_splats
  57 // CHECK: call <2 x i64> @vec_splats
  58 // CHECK: call <8 x i16> @vec_add(short vector[8], short vector[8])
  59
  60 // CHECK-LABEL: define available_externally i64 @_mm_add_pi8
  61 // CHECK: call <2 x i64> @vec_splats
  62 // CHECK: call <2 x i64> @vec_splats
  63 // CHECK: call <16 x i8> @vec_add(signed char vector[16], signed char vector[16])
  64
  65 // CHECK-LABEL: define available_externally i64 @_mm_adds_pu16
  66 // CHECK: call <2 x i64> @vec_splats
  67 // CHECK: call <2 x i64> @vec_splats
  68 // CHECK: call <8 x i16> @vec_adds(unsigned short vector[8], unsigned short vector[8])
  69
  70 // CHECK-LABEL: define available_externally i64 @_mm_adds_pu8
  71 // CHECK: call <2 x i64> @vec_splats
  72 // CHECK: call <2 x i64> @vec_splats
  73 // CHECK: call <16 x i8> @vec_adds(unsigned char vector[16], unsigned char vector[16])
  74
  75 // CHECK-LABEL: define available_externally i64 @_mm_adds_pi16
  76 // CHECK: call <2 x i64> @vec_splats
  77 // CHECK: call <2 x i64> @vec_splats
  78 // CHECK: call <8 x i16> @vec_adds(short vector[8], short vector[8])
  79
  80 // CHECK-LABEL: define available_externally i64 @_mm_adds_pi8
  81 // CHECK: call <2 x i64> @vec_splats
  82 // CHECK: call <2 x i64> @vec_splats
  83 // CHECK: call <16 x i8> @vec_adds(signed char vector[16], signed char vector[16])
  84
  85 void __attribute__((noinline))
  86 test_alt_name_add() {
  87   res = _m_paddb(m1, m2);
  88   res = _m_paddd(m1, m2);
  89   res = _m_paddsb(m1, m2);
  90   res = _m_paddsw(m1, m2);
  91   res = _m_paddusb(m1, m2);
  92   res = _m_paddusw(m1, m2);
  93   res = _m_paddw(m1, m2);
  94 }
  95
  96 // CHECK-LABEL: @test_alt_name_add
  97
  98 // CHECK-LABEL: define available_externally i64 @_m_paddb
  99 // CHECK: call i64 @_mm_add_pi8
 100
 101 // CHECK-LABEL: define available_externally i64 @_m_paddd
 102 // CHECK: call i64 @_mm_add_pi32
 103
 104 // CHECK-LABEL: define available_externally i64 @_m_paddsb
 105 // CHECK: call i64 @_mm_adds_pi8
 106
 107 // CHECK-LABEL: define available_externally i64 @_m_paddsw
 108 // CHECK: call i64 @_mm_adds_pi16
 109
 110 // CHECK-LABEL: define available_externally i64 @_m_paddusb
 111 // CHECK: call i64 @_mm_adds_pu8
 112
 113 // CHECK-LABEL: define available_externally i64 @_m_paddusw
 114 // CHECK: call i64 @_mm_adds_pu16
 115
 116 // CHECK-LABEL: define available_externally i64 @_m_paddw
 117 // CHECK: call i64 @_mm_add_pi16
 118
 119 void __attribute__((noinline))
 120 test_cmp() {
 121   res = _mm_cmpeq_pi32(m1, m2);
 122   res = _mm_cmpeq_pi16(m1, m2);
 123   res = _mm_cmpeq_pi8(m1, m2);
 124   res = _mm_cmpgt_pi32(m1, m2);
 125   res = _mm_cmpgt_pi16(m1, m2);
 126   res = _mm_cmpgt_pi8(m1, m2);
 127 }
 128
 129 // CHECK-LABEL: @test_cmp
 130
 131 // CHECK-LABEL: define available_externally i64 @_mm_cmpeq_pi32
 132 // CHECK-P9: call <2 x i64> @vec_splats(unsigned long long)
 133 // CHECK-P9: call <2 x i64> @vec_splats(unsigned long long)
 134 // CHECK-P9: call <4 x i32> @vec_cmpeq(int vector[4], int vector[4])
 135 // CHECK-P8: %[[CMP1:[0-9a-zA-Z_.]+]] = icmp eq i32 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 136 // CHECK-P8: select i1 %[[CMP1]], i32 -1, i32 0
 137 // CHECK-P8: %[[CMP2:[0-9a-zA-Z_.]+]] = icmp eq i32 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 138 // CHECK-P8: select i1 %[[CMP2]], i32 -1, i32 0
 139
 140 // CHECK-LABEL: define available_externally i64 @_mm_cmpeq_pi16
 141 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 142 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 143 // CHECK: call <8 x i16> @vec_cmpeq(short vector[8], short vector[8])
 144
 145 // CHECK-LABEL: define available_externally i64 @_mm_cmpeq_pi8
 146 // CHECK: call i64 asm "cmpb $0,$1,$2;\0A", "=r,r,r"
 147
 148 // CHECK-LABEL: define available_externally i64 @_mm_cmpgt_pi32
 149 // CHECK-P9: call <2 x i64> @vec_splats(unsigned long long)
 150 // CHECK-P9: call <2 x i64> @vec_splats(unsigned long long)
 151 // CHECK-P9: call <4 x i32> @vec_cmpgt(int vector[4], int vector[4])
 152 // CHECK-P8: %[[CMP1:[0-9a-zA-Z_.]+]] = icmp sgt i32 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 153 // CHECK-P8: select i1 %[[CMP1]], i32 -1, i32 0
 154 // CHECK-P8: [[CMP2:[0-9a-zA-Z_.]+]] = icmp sgt i32 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 155 // CHECK-P8: select i1 %[[CMP2]], i32 -1, i32 0
 156
 157 // CHECK-LABEL: define available_externally i64 @_mm_cmpgt_pi16
 158 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 159 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 160 // CHECK: call <8 x i16> @vec_cmpgt(short vector[8], short vector[8])
 161
 162 // CHECK-LABEL: define available_externally i64 @_mm_cmpgt_pi8
 163 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 164 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 165 // CHECK: call <16 x i8> @vec_cmpgt(signed char vector[16], signed char vector[16])
 166
 167 void __attribute__((noinline))
 168 test_alt_name_cmp() {
 169   res = _m_pcmpeqb(m1, m2);
 170   res = _m_pcmpeqd(m1, m2);
 171   res = _m_pcmpeqw(m1, m2);
 172   res = _m_pcmpgtb(m1, m2);
 173   res = _m_pcmpgtd(m1, m2);
 174   res = _m_pcmpgtw(m1, m2);
 175 }
 176
 177 // CHECK-LABEL: @test_alt_name_cmp
 178
 179 // CHECK-LABEL: define available_externally i64 @_m_pcmpeqb
 180 // CHECK: call i64 @_mm_cmpeq_pi8
 181
 182 // CHECK-LABEL: define available_externally i64 @_m_pcmpeqd
 183 // CHECK: call i64 @_mm_cmpeq_pi32
 184
 185 // CHECK-LABEL: define available_externally i64 @_m_pcmpeqw
 186 // CHECK: call i64 @_mm_cmpeq_pi16
 187
 188 // CHECK-LABEL: define available_externally i64 @_m_pcmpgtb
 189 // CHECK: call i64 @_mm_cmpgt_pi8
 190
 191 // CHECK-LABEL: define available_externally i64 @_m_pcmpgtd
 192 // CHECK: call i64 @_mm_cmpgt_pi32
 193
 194 // CHECK-LABEL: define available_externally i64 @_m_pcmpgtw
 195 // CHECK: call i64 @_mm_cmpgt_pi16
 196
 197 void __attribute__((noinline))
 198 test_convert() {
 199   ll1 = _mm_cvtm64_si64(m1);
 200   m1 = _mm_cvtsi32_si64(i1);
 201   m1 = _mm_cvtsi64_m64(ll1);
 202   i1 = _mm_cvtsi64_si32(m1);
 203 }
 204
 205 // CHECK-LABEL: @test_convert
 206
 207 // CHECK-LABEL: define available_externally i64 @_mm_cvtm64_si64
 208 // CHECK: %[[RESULT:[0-9a-zA-Z_.]+]] = load i64, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 209 // CHECK-NEXT: ret i64 %[[RESULT]]
 210
 211 // CHECK-LABEL: define available_externally i64 @_mm_cvtsi32_si64
 212 // CHECK: %[[LOAD:[0-9a-zA-Z_.]+]] = load i32, ptr %{{[0-9a-zA-Z_.]+}}
 213 // CHECK-NEXT: %[[RESULT:[0-9a-zA-Z_.]+]] = zext i32 %[[LOAD]] to i64
 214 // CHECK-NEXT: ret i64 %[[RESULT]]
 215
 216 // CHECK-LABEL: define available_externally i64 @_mm_cvtsi64_m64
 217 // CHECK: %[[RESULT:[0-9a-zA-Z_.]+]] = load i64, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 218 // CHECK-NEXT: ret i64 %[[RESULT]]
 219
 220 // CHECK-LABEL: define available_externally signext i32 @_mm_cvtsi64_si32
 221 // CHECK: %[[LOAD:[0-9a-zA-Z_.]+]] = load i64, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 222 // CHECK-NEXT: %[[RESULT:[0-9a-zA-Z_.]+]] = trunc i64 %[[LOAD]] to i32
 223 // CHECK-NEXT: ret i32 %[[RESULT]]
 224
 225 void __attribute__((noinline))
 226 test_alt_name_convert() {
 227   m1 = _m_from_int(i1);
 228   m1 = _m_from_int64(ll1);
 229   i1 = _m_to_int(m1);
 230   ll1 = _m_to_int64(m1);
 231 }
 232
 233 // CHECK-LABEL: @test_alt_name_convert
 234
 235 // CHECK-LABEL: define available_externally i64 @_m_from_int
 236 // CHECK: call i64 @_mm_cvtsi32_si64
 237
 238 // CHECK-LABEL: define available_externally i64 @_m_from_int64
 239 // CHECK: %[[RESULT:[0-9a-zA-Z_.]+]] = load i64, ptr %{{[0-9a-zA-Z_.]+}}
 240 // CHECK-NEXT: ret i64 %[[RESULT]]
 241
 242 // CHECK-LABEL: define available_externally signext i32 @_m_to_int
 243 // CHECK: call signext i32 @_mm_cvtsi64_si32
 244
 245 // CHECK-LABEL: define available_externally i64 @_m_to_int64
 246 // CHECK: %[[RESULT:[0-9a-zA-Z_.]+]] = load i64, ptr %{{[0-9a-zA-Z_.]+}}
 247 // CHECK-NEXT: ret i64 %[[RESULT]]
 248
 249 void __attribute__((noinline))
 250 test_empty() {
 251   _mm_empty();
 252   _m_empty();
 253 }
 254
 255 // CHECK-LABEL: @test_empty
 256
 257 // CHECK-LABEL: define available_externally void @_mm_empty
 258 // CHECK-NEXT: entry
 259 // CHECK-NEXT: ret void
 260
 261 // CHECK-LABEL: define available_externally void @_m_empty
 262 // CHECK-NEXT: entry
 263 // CHECK-NEXT: ret void
 264
 265 void __attribute__((noinline))
 266 test_logic() {
 267   res = _mm_and_si64(m1, m2);
 268   res = _mm_andnot_si64(m1, m2);
 269   res = _mm_or_si64(m1, m2);
 270   res = _mm_xor_si64(m1, m2);
 271 }
 272
 273 // CHECK-LABEL: @test_logic
 274
 275 // CHECK-LABEL: define available_externally i64 @_mm_and_si64
 276 // CHECK: and i64 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 277
 278 // CHECK-LABEL: define available_externally i64 @_mm_andnot_si64
 279 // CHECK: %[[XOR:[0-9a-zA-Z_.]+]] = xor i64 %{{[0-9a-zA-Z_.]+}}, -1
 280 // CHECK: and i64 %[[XOR]], %{{[0-9a-zA-Z_.]+}}
 281
 282 // CHECK-LABEL: define available_externally i64 @_mm_or_si64
 283 // CHECK: or i64 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 284
 285 // CHECK-LABEL: define available_externally i64 @_mm_xor_si64
 286 // CHECK: xor i64 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 287
 288 void __attribute__((noinline))
 289 test_alt_name_logic() {
 290   res = _m_pand(m1, m2);
 291   res = _m_pandn(m1, m2);
 292   res = _m_por(m1, m2);
 293   res = _m_pxor(m1, m2);
 294 }
 295
 296 // CHECK-LABEL: @test_alt_name_logic
 297
 298 // CHECK-LABEL: define available_externally i64 @_m_pand
 299 // CHECK: call i64 @_mm_and_si64
 300
 301 // CHECK-LABEL: define available_externally i64 @_m_pandn
 302 // CHECK: call i64 @_mm_andnot_si64
 303
 304 // CHECK-LABEL: define available_externally i64 @_m_por
 305 // CHECK: call i64 @_mm_or_si64
 306
 307 // CHECK-LABEL: define available_externally i64 @_m_pxor
 308 // CHECK: call i64 @_mm_xor_si64
 309
 310 void __attribute__((noinline))
 311 test_madd() {
 312   res = _mm_madd_pi16(m1, m2);
 313   res = _m_pmaddwd(m1, m2);
 314 }
 315
 316 // CHECK-LABEL: @test_madd
 317
 318 // CHECK-LABEL: define available_externally i64 @_mm_madd_pi16
 319 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 320 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 321 // CHECK: call <4 x i32> @vec_vmsumshm
 322
 323 // CHECK-LABEL: define available_externally i64 @_m_pmaddwd
 324 // CHECK: call i64 @_mm_madd_pi16
 325
 326 void __attribute__((noinline))
 327 test_mul() {
 328   res = _mm_mulhi_pi16(m1, m2);
 329   res = _mm_mullo_pi16(m1, m2);
 330 }
 331
 332 // CHECK-LABEL: @test_mul
 333
 334 // CHECK-LABEL: define available_externally i64 @_mm_mulhi_pi16
 335 // CHECK-BE: store <16 x i8> <i8 0, i8 1, i8 16, i8 17, i8 4, i8 5, i8 20, i8 21, i8 0, i8 1, i8 16, i8 17, i8 4, i8 5, i8 20, i8 21>
 336 // 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>
 337 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 338 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 339 // CHECK: call <4 x i32> @vec_vmulesh
 340 // CHECK: call <4 x i32> @vec_vmulosh
 341 // CHECK: call <4 x i32> @vec_perm(int vector[4], int vector[4], unsigned char vector[16])
 342
 343 // CHECK-LABEL: define available_externally i64 @_mm_mullo_pi16
 344 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 345 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 346 // CHECK: mul <8 x i16> %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 347
 348 void __attribute__((noinline))
 349 test_alt_name_mul() {
 350   res = _m_pmulhw(m1, m2);
 351   res = _m_pmullw(m1, m2);
 352 }
 353
 354 // CHECK-LABEL: @test_alt_name_mul
 355
 356 // CHECK-LABEL: define available_externally i64 @_m_pmulhw
 357 // CHECK: call i64 @_mm_mulhi_pi16
 358
 359 // CHECK-LABEL: define available_externally i64 @_m_pmullw
 360 // CHECK: call i64 @_mm_mullo_pi16
 361
 362 void __attribute__((noinline))
 363 test_packs() {
 364   res = _mm_packs_pu16((__m64)ull1, (__m64)ull2);
 365   res = _mm_packs_pi16((__m64)ull1, (__m64)ull2);
 366   res = _mm_packs_pi32((__m64)ull1, (__m64)ull2);
 367 }
 368
 369 // CHECK-LABEL: @test_packs
 370
 371 // CHECK-LABEL: define available_externally i64 @_mm_packs_pu16
 372 // CHECK: call <8 x i16> @vec_cmplt
 373 // CHECK: call <16 x i8> @vec_packs(unsigned short vector[8], unsigned short vector[8])
 374 // CHECK: call <16 x i8> @vec_pack(bool vector[8], bool vector[8])
 375 // CHECK: call <16 x i8> @vec_sel(unsigned char vector[16], unsigned char vector[16], bool vector[16])(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef zeroinitializer, <16 x i8> noundef %{{[0-9a-zA-Z_.]+}})
 376
 377 // CHECK-LABEL: define available_externally i64 @_mm_packs_pi16
 378 // CHECK: call <16 x i8> @vec_packs(short vector[8], short vector[8])
 379
 380 // CHECK-LABEL: define available_externally i64 @_mm_packs_pi32
 381 // CHECK: call <8 x i16> @vec_packs(int vector[4], int vector[4])
 382
 383 void __attribute__((noinline))
 384 test_alt_name_packs() {
 385   res = _m_packssdw(m1, m2);
 386   res = _m_packsswb(m1, m2);
 387   res = _m_packuswb(m1, m2);
 388 }
 389
 390 // CHECK-LABEL: @test_alt_name_packs
 391
 392 // CHECK-LABEL: define available_externally i64 @_m_packssdw
 393 // CHECK: call i64 @_mm_packs_pi32
 394
 395 // CHECK-LABEL: define available_externally i64 @_m_packsswb
 396 // CHECK: call i64 @_mm_packs_pi16
 397
 398 // CHECK-LABEL: define available_externally i64 @_m_packuswb
 399 // CHECK: call i64 @_mm_packs_pu16
 400
 401 void __attribute__((noinline))
 402 test_set() {
 403   m1 = _mm_set_pi32(2134, -128);
 404   m1 = _mm_set_pi16(2134, -128, 1234, 6354);
 405   m1 = _mm_set_pi8(-128, 10, 0, 123, -1, -5, 127, 5);
 406 }
 407
 408 // CHECK-LABEL: @test_set
 409
 410 // CHECK-LABEL: define available_externally i64 @_mm_set_pi32
 411 // CHECK-COUNT-2: store i32 %{{[0-9a-zA-Z_.]+}}, ptr
 412
 413 // CHECK-LABEL: define available_externally i64 @_mm_set_pi16
 414 // CHECK-COUNT-4: store i16 %{{[0-9a-zA-Z_.]+}}, ptr
 415
 416 // CHECK-LABEL: define available_externally i64 @_mm_set_pi8
 417 // CHECK-COUNT-8: store i8 %{{[0-9a-zA-Z_.]+}}, ptr
 418
 419 void __attribute__((noinline))
 420 test_set1() {
 421   res = _mm_set1_pi32(i1);
 422   res = _mm_set1_pi16(s[0]);
 423   res = _mm_set1_pi8(c[0]);
 424 }
 425
 426 // CHECK-LABEL: @test_set1
 427
 428 // CHECK-LABEL: define available_externally i64 @_mm_set1_pi32
 429 // CHECK-COUNT-2: store i32 %{{[0-9a-zA-Z_.]+}}, ptr
 430
 431 // CHECK-LABEL: define available_externally i64 @_mm_set1_pi16
 432 // CHECK-P9: call <8 x i16> @vec_splats(short)
 433 // CHECK-P9: extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 0
 434 // CHECK-P8: %[[ADDR1:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 435 // CHECK-P8: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR1]], align 8
 436 // CHECK-P8: %[[ADDR2:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 437 // CHECK-P8: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR2]], align 2
 438 // CHECK-P8: %[[ADDR3:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 2
 439 // CHECK-P8: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR3]], align 4
 440 // CHECK-P8: %[[ADDR4:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 3
 441 // CHECK-P8: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR4]], align 2
 442
 443
 444 // CHECK-LABEL: define available_externally i64 @_mm_set1_pi8
 445 // CHECK: call <16 x i8> @vec_splats(signed char)
 446 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <16 x i8> %{{[0-9a-zA-Z_.]+}} to <2 x i64>
 447 // CHECK: extractelement <2 x i64> %[[CAST]], i32 0
 448
 449 void __attribute__((noinline))
 450 test_setr() {
 451   res = _mm_setr_pi32(i1, i2);
 452   res = _mm_setr_pi16(s[0], s[1], s[2], s[3]);
 453   res = _mm_setr_pi8(c[0], c[1], c[2], c[3], c[4], c[5], c[6], c[7]);
 454 }
 455
 456 // CHECK-LABEL: @test_setr
 457
 458 // CHECK-LABEL: define available_externally i64 @_mm_setr_pi32
 459 // CHECK: %[[ADDR1:[0-9a-zA-Z_.]+]] = getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 460 // CHECK: store i32 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR1]], align 8
 461 // CHECK: %[[ADDR2:[0-9a-zA-Z_.]+]] = getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 462 // CHECK: store i32 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR2]], align 4
 463
 464 // CHECK-LABEL: define available_externally i64 @_mm_setr_pi16
 465 // CHECK: call i64 @_mm_set_pi16
 466
 467 // CHECK-LABEL: define available_externally i64 @_mm_setr_pi8
 468 // CHECK: call i64 @_mm_set_pi8
 469
 470 void __attribute__((noinline))
 471 test_setzero() {
 472   res = _mm_setzero_si64();
 473 }
 474
 475 // CHECK-LABEL: @test_setzero
 476
 477 // CHECK-LABEL: define available_externally i64 @_mm_setzero_si64
 478 // CHECK: entry
 479 // CHECK-NEXT: ret i64 0
 480
 481 void __attribute__((noinline))
 482 test_sll() {
 483   res = _mm_sll_pi16(m1, m2);
 484   res = _mm_sll_pi32(m1, m2);
 485   res = _mm_sll_si64(m1, m2);
 486   res = _mm_slli_pi16(m1, i1);
 487   res = _mm_slli_pi32(m1, i1);
 488   res = _mm_slli_si64(m1, i1);
 489 }
 490
 491 // CHECK-LABEL: @test_sll
 492
 493 // CHECK-LABEL: define available_externally i64 @_mm_sll_pi16
 494 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp ule i64 %{{[0-9a-zA-Z_.]+}}, 15
 495 // CHECK-NEXT: br i1 %[[CMP]]
 496 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 497 // CHECK: trunc i64 %{{[0-9a-zA-Z_.]+}} to i16
 498 // CHECK: call <8 x i16> @vec_splats(unsigned short)
 499 // CHECK: call <8 x i16> @vec_sl(short vector[8], unsigned short vector[8])
 500 // CHECK: store i64 0, ptr
 501
 502 // CHECK-LABEL: define available_externally i64 @_mm_sll_pi32
 503 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i64 %{{[0-9a-zA-Z_.]+}} to i32
 504 // CHECK: shl i32 %{{[0-9a-zA-Z_.]+}}, %[[TRUNC]]
 505 // CHECK: trunc i64 %{{[0-9a-zA-Z_.]+}} to i32
 506 // CHECK: shl i32 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 507
 508 // CHECK-LABEL: define available_externally i64 @_mm_sll_si64
 509 // CHECK: shl i64 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 510
 511 // CHECK-LABEL: define available_externally i64 @_mm_slli_pi16
 512 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = sext i32 %{{[0-9a-zA-Z_.]+}} to i64
 513 // CHECK: call i64 @_mm_sll_pi16(i64 noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef %[[EXT]])
 514
 515 // CHECK-LABEL: define available_externally i64 @_mm_slli_pi32
 516 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = sext i32 %{{[0-9a-zA-Z_.]+}} to i64
 517 // CHECK: call i64 @_mm_sll_pi32(i64 noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef %[[EXT]])
 518
 519 // CHECK-LABEL: define available_externally i64 @_mm_slli_si64
 520 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = zext i32 %{{[0-9a-zA-Z_.]+}} to i64
 521 // CHECK: shl i64 %{{[0-9a-zA-Z_.]+}}, %[[EXT]]
 522
 523 void __attribute__((noinline))
 524 test_alt_name_sll() {
 525   res = _m_pslld(m1, m2);
 526   res = _m_pslldi(m1, i1);
 527   res = _m_psllq(m1, m2);
 528   res = _m_psllqi(m1, i1);
 529   res = _m_psllw(m1, m2);
 530   res = _m_psllwi(m1, i1);
 531 }
 532
 533 // CHECK-LABEL: @test_alt_name_sll
 534
 535 // CHECK-LABEL: define available_externally i64 @_m_pslld
 536 // CHECK: call i64 @_mm_sll_pi32
 537
 538 // CHECK-LABEL: define available_externally i64 @_m_pslldi
 539 // CHECK: call i64 @_mm_slli_pi32
 540
 541 // CHECK-LABEL: define available_externally i64 @_m_psllq
 542 // CHECK: call i64 @_mm_sll_si64
 543
 544 // CHECK-LABEL: define available_externally i64 @_m_psllqi
 545 // CHECK: call i64 @_mm_slli_si64
 546
 547 // CHECK-LABEL: define available_externally i64 @_m_psllw
 548 // CHECK: call i64 @_mm_sll_pi16
 549
 550 // CHECK-LABEL: define available_externally i64 @_m_psllwi
 551 // CHECK: call i64 @_mm_slli_pi16
 552
 553 void __attribute__((noinline))
 554 test_sra() {
 555   res = _mm_sra_pi32(m1, m2);
 556   res = _mm_sra_pi16(m1, m2);
 557   res = _mm_srai_pi32(m1, i1);
 558   res = _mm_srai_pi16(m1, i1);
 559 }
 560
 561 // CHECK-LABEL: @test_sra
 562
 563 // CHECK-LABEL: define available_externally i64 @_mm_sra_pi32
 564 // CHECK: %[[TRUNC1:[0-9a-zA-Z_.]+]] = trunc i64 %{{[0-9a-zA-Z_.]+}} to i32
 565 // CHECK: ashr i32 %{{[0-9a-zA-Z_.]+}}, %[[TRUNC1]]
 566 // CHECK: %[[TRUNC2:[0-9a-zA-Z_.]+]] = trunc i64 %{{[0-9a-zA-Z_.]+}} to i32
 567 // CHECK: ashr i32 %{{[0-9a-zA-Z_.]+}}, %[[TRUNC2]]
 568
 569 // CHECK-LABEL: define available_externally i64 @_mm_sra_pi16
 570 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp ule i64 %{{[0-9a-zA-Z_.]+}}, 15
 571 // CHECK: br i1 %[[CMP]]
 572 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 573 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i64 %{{[0-9a-zA-Z_.]+}} to i16
 574 // CHECK: call <8 x i16> @vec_splats(unsigned short)(i16 noundef zeroext %[[TRUNC]])
 575 // CHECK: call <8 x i16> @vec_sra(short vector[8], unsigned short vector[8])
 576 // CHECK: store i64 0, ptr
 577
 578 // CHECK-LABEL: define available_externally i64 @_mm_srai_pi32
 579 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = sext i32 %{{[0-9a-zA-Z_.]+}} to i64
 580 // CHECK: call i64 @_mm_sra_pi32(i64 noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef %[[EXT]])
 581
 582 // CHECK-LABEL: define available_externally i64 @_mm_srai_pi16
 583 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = sext i32 %{{[0-9a-zA-Z_.]+}} to i64
 584 // CHECK: call i64 @_mm_sra_pi16(i64 noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef %[[EXT]])
 585
 586 void __attribute__((noinline))
 587 test_alt_name_sra() {
 588   res = _m_psrad(m1, m2);
 589   res = _m_psraw(m1, m2);
 590   res = _m_psradi(m1, i1);
 591   res = _m_psrawi(m1, i1);
 592 }
 593
 594 // CHECK-LABEL: @test_alt_name_sra
 595
 596 // CHECK-LABEL: define available_externally i64 @_m_psrad
 597 // CHECK: call i64 @_mm_sra_pi32
 598
 599 // CHECK-LABEL: define available_externally i64 @_m_psraw
 600 // CHECK: call i64 @_mm_sra_pi16
 601
 602 // CHECK-LABEL: define available_externally i64 @_m_psradi
 603 // CHECK: call i64 @_mm_srai_pi32
 604
 605 // CHECK-LABEL: define available_externally i64 @_m_psrawi
 606 // CHECK: call i64 @_mm_srai_pi16
 607
 608 void __attribute__((noinline))
 609 test_srl() {
 610   res = _mm_srl_si64(m1, m2);
 611   res = _mm_srl_pi32(m1, m2);
 612   res = _mm_srl_pi16(m1, m2);
 613   res = _mm_srli_si64(m1, i1);
 614   res = _mm_srli_pi32(m1, i1);
 615   res = _mm_srli_pi16(m1, i1);
 616 }
 617
 618 // CHECK-LABEL: @test_srl
 619
 620 // CHECK-LABEL: define available_externally i64 @_mm_srl_si64
 621 // CHECK: lshr i64 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 622
 623 // CHECK-LABEL: define available_externally i64 @_mm_srl_pi32
 624 // CHECK: %[[TRUNC1:[0-9a-zA-Z_.]+]] = trunc i64 %{{[0-9a-zA-Z_.]+}} to i32
 625 // CHECK: lshr i32 %{{[0-9a-zA-Z_.]+}}, %[[TRUNC1]]
 626 // CHECK: %[[TRUNC2:[0-9a-zA-Z_.]+]] = trunc i64 %{{[0-9a-zA-Z_.]+}} to i32
 627 // CHECK: lshr i32 %{{[0-9a-zA-Z_.]+}}, %[[TRUNC2]]
 628
 629 // CHECK-LABEL: define available_externally i64 @_mm_srl_pi16
 630 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp ule i64 %{{[0-9a-zA-Z_.]+}}, 15
 631 // CHECK: br i1 %[[CMP]]
 632 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 633 // CHECK: trunc i64 %{{[0-9a-zA-Z_.]+}} to i16
 634 // CHECK: call <8 x i16> @vec_splats(unsigned short)
 635 // CHECK: call <8 x i16> @vec_sr(unsigned short vector[8], unsigned short vector[8])
 636 // CHECK: store i64 0, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 637
 638 // CHECK-LABEL: define available_externally i64 @_mm_srli_si64
 639 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = zext i32 %{{[0-9a-zA-Z_.]+}} to i64
 640 // CHECK: lshr i64 %{{[0-9a-zA-Z_.]+}}, %[[EXT]]
 641
 642 // CHECK-LABEL: define available_externally i64 @_mm_srli_pi32
 643 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = sext i32 %{{[0-9a-zA-Z_.]+}} to i64
 644 // CHECK: call i64 @_mm_srl_pi32(i64 noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef %[[EXT]])
 645
 646 // CHECK-LABEL: define available_externally i64 @_mm_srli_pi16
 647 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = sext i32 %{{[0-9a-zA-Z_.]+}} to i64
 648 // CHECK: call i64 @_mm_srl_pi16(i64 noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef %[[EXT]])
 649
 650 void __attribute__((noinline))
 651 test_alt_name_srl() {
 652   res = _m_psrlq(m1, m2);
 653   res = _m_psrld(m1, m2);
 654   res = _m_psrlw(m1, m2);
 655   res = _m_psrlqi(m1, i1);
 656   res = _m_psrldi(m1, i1);
 657   res = _m_psrlwi(m1, i1);
 658 }
 659
 660 // CHECK-LABEL: @test_alt_name_srl
 661
 662 // CHECK-LABEL: define available_externally i64 @_m_psrlq
 663 // CHECK: call i64 @_mm_srl_si64
 664
 665 // CHECK-LABEL: define available_externally i64 @_m_psrld
 666 // CHECK: call i64 @_mm_srl_pi32
 667
 668 // CHECK-LABEL: define available_externally i64 @_m_psrlw
 669 // CHECK: call i64 @_mm_srl_pi16
 670
 671 // CHECK-LABEL: define available_externally i64 @_m_psrlqi
 672 // CHECK: call i64 @_mm_srli_si64
 673
 674 // CHECK-LABEL: define available_externally i64 @_m_psrldi
 675 // CHECK: call i64 @_mm_srli_pi32
 676
 677 // CHECK-LABEL: define available_externally i64 @_m_psrlwi
 678 // CHECK: call i64 @_mm_srli_pi16
 679
 680 void __attribute__((noinline))
 681 test_sub() {
 682   res = _mm_sub_pi32(m1, m2);
 683   res = _mm_sub_pi16(m1, m2);
 684   res = _mm_sub_pi8(m1, m2);
 685   res = _mm_subs_pi16(m1, m2);
 686   res = _mm_subs_pi8(m1, m2);
 687   res = _mm_subs_pu16(m1, m2);
 688   res = _mm_subs_pu8(m1, m2);
 689 }
 690
 691 // CHECK-LABEL: @test_sub
 692
 693 // CHECK-LABEL: define available_externally i64 @_mm_sub_pi32
 694 // CHECK-P8-COUNT-2: getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 695 // CHECK-P8: sub nsw i32 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 696 // CHECK-P8-COUNT-2: getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 697 // CHECK-P8: sub nsw i32 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 698 // CHECK-P9: call <2 x i64> @vec_splats(unsigned long long)
 699 // CHECK-P9: call <2 x i64> @vec_splats(unsigned long long)
 700 // CHECK-P9: call <4 x i32> @vec_sub(int vector[4], int vector[4])
 701
 702 // CHECK-LABEL: define available_externally i64 @_mm_sub_pi16
 703 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 704 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 705 // CHECK: call <8 x i16> @vec_sub(short vector[8], short vector[8])
 706
 707 // CHECK-LABEL: define available_externally i64 @_mm_sub_pi8
 708 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 709 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 710 // CHECK: call <16 x i8> @vec_sub(signed char vector[16], signed char vector[16])
 711
 712 // CHECK-LABEL: define available_externally i64 @_mm_subs_pi16
 713 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 714 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 715 // CHECK: call <8 x i16> @vec_subs(short vector[8], short vector[8])
 716
 717 // CHECK-LABEL: define available_externally i64 @_mm_subs_pi8
 718 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 719 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 720 // CHECK: call <16 x i8> @vec_subs(signed char vector[16], signed char vector[16])
 721
 722 // CHECK-LABEL: define available_externally i64 @_mm_subs_pu16
 723 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 724 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 725 // CHECK: call <8 x i16> @vec_subs(unsigned short vector[8], unsigned short vector[8])
 726
 727 // CHECK-LABEL: define available_externally i64 @_mm_subs_pu8
 728 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 729 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 730 // CHECK: call <16 x i8> @vec_subs(unsigned char vector[16], unsigned char vector[16])
 731
 732 void __attribute__((noinline))
 733 test_alt_name_sub() {
 734   res = _m_psubd(m1, m2);
 735   res = _m_psubw(m1, m2);
 736   res = _m_psubb(m1, m2);
 737   res = _m_psubsw(m1, m2);
 738   res = _m_psubsb(m1, m2);
 739   res = _m_psubusw(m1, m2);
 740   res = _m_psubusb(m1, m2);
 741 }
 742
 743 // CHECK-LABEL: @test_alt_name_sub
 744
 745 // CHECK-LABEL: define available_externally i64 @_m_psubd
 746 // CHECK: call i64 @_mm_sub_pi32
 747
 748 // CHECK-LABEL: define available_externally i64 @_m_psubw
 749 // CHECK: call i64 @_mm_sub_pi16
 750
 751 // CHECK-LABEL: define available_externally i64 @_m_psubb
 752 // CHECK: call i64 @_mm_sub_pi8
 753
 754 // CHECK-LABEL: define available_externally i64 @_m_psubsw
 755 // CHECK: call i64 @_mm_subs_pi16
 756
 757 // CHECK-LABEL: define available_externally i64 @_m_psubsb
 758 // CHECK: call i64 @_mm_subs_pi8
 759
 760 // CHECK-LABEL: define available_externally i64 @_m_psubusw
 761 // CHECK: call i64 @_mm_subs_pu16
 762
 763 // CHECK-LABEL: define available_externally i64 @_m_psubusb
 764 // CHECK: call i64 @_mm_subs_pu8
 765
 766 void __attribute__((noinline))
 767 test_unpack() {
 768   res = _mm_unpackhi_pi32(m1, m2);
 769   res = _mm_unpackhi_pi16(m1, m2);
 770   res = _mm_unpackhi_pi8(m1, m2);
 771   res = _mm_unpacklo_pi32(m1, m2);
 772   res = _mm_unpacklo_pi16(m1, m2);
 773   res = _mm_unpacklo_pi8(m1, m2);
 774 }
 775
 776 // CHECK-LABEL: @test_unpack
 777
 778 // CHECK-LABEL: define available_externally i64 @_mm_unpackhi_pi32
 779 // CHECK: getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 780 // CHECK: %[[ADDR1:[0-9a-zA-Z_.]+]] = getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 781 // CHECK: store i32 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR1]], align 8
 782 // CHECK: getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 783 // CHECK: %[[ADDR2:[0-9a-zA-Z_.]+]] = getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 784 // CHECK: store i32 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR2]], align 4
 785
 786 // CHECK-LABEL: define available_externally i64 @_mm_unpackhi_pi16
 787 // CHECK: getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 2
 788 // CHECK: %[[ADDR1:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 789 // CHECK: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR1]], align 8
 790 // CHECK: getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 2
 791 // CHECK: %[[ADDR2:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 792 // CHECK: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR2]], align 2
 793 // CHECK: getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 3
 794 // CHECK: %[[ADDR3:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 2
 795 // CHECK: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR3]], align 4
 796 // CHECK: getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 3
 797 // CHECK: %[[ADDR4:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 3
 798 // CHECK: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR4]], align 2
 799
 800 // CHECK-LABEL: define available_externally i64 @_mm_unpackhi_pi8
 801 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 802 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 803 // CHECK: call <16 x i8> @vec_mergel(unsigned char vector[16], unsigned char vector[16])
 804
 805 // CHECK-LABEL: define available_externally i64 @_mm_unpacklo_pi32
 806 // CHECK: getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 807 // CHECK: %[[ADDR1:[0-9a-zA-Z_.]+]] = getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 808 // CHECK: store i32 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR1]], align 8
 809 // CHECK: getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 810 // CHECK: %[[ADDR2:[0-9a-zA-Z_.]+]] = getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 811 // CHECK: store i32 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR2]], align 4
 812
 813 // CHECK-LABEL: define available_externally i64 @_mm_unpacklo_pi16
 814 // CHECK: getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 815 // CHECK: %[[ADDR1:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 816 // CHECK: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR1]], align 8
 817 // CHECK: getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 818 // CHECK: %[[ADDR2:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 819 // CHECK: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR2]], align 2
 820 // CHECK: getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 821 // CHECK: %[[ADDR3:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 2
 822 // CHECK: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR3]], align 4
 823 // CHECK: getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 824 // CHECK: %[[ADDR4:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 3
 825 // CHECK: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR4]], align 2
 826
 827 // CHECK-LABEL: define available_externally i64 @_mm_unpacklo_pi8
 828 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 829 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 830 // CHECK: call <16 x i8> @vec_mergel(unsigned char vector[16], unsigned char vector[16])
 831
 832 void __attribute__((noinline))
 833 test_alt_name_unpack() {
 834   res = _m_punpckhdq(m1, m2);
 835   res = _m_punpckhwd(m1, m2);
 836   res = _m_punpckhbw(m1, m2);
 837   res = _m_punpckldq(m1, m2);
 838   res = _m_punpcklwd(m1, m2);
 839   res = _m_punpcklbw(m1, m2);
 840 }
 841
 842 // CHECK-LABEL: @test_alt_name_unpack
 843
 844 // CHECK-LABEL: define available_externally i64 @_m_punpckhdq
 845 // CHECK: call i64 @_mm_unpackhi_pi32
 846
 847 // CHECK-LABEL: define available_externally i64 @_m_punpckhwd
 848 // CHECK: call i64 @_mm_unpackhi_pi16
 849
 850 // CHECK-LABEL: define available_externally i64 @_m_punpckhbw
 851 // CHECK: call i64 @_mm_unpackhi_pi8
 852
 853 // CHECK-LABEL: define available_externally i64 @_m_punpckldq
 854 // CHECK: call i64 @_mm_unpacklo_pi32
 855
 856 // CHECK-LABEL: define available_externally i64 @_m_punpcklwd
 857 // CHECK: call i64 @_mm_unpacklo_pi16
 858
 859 // CHECK-LABEL: define available_externally i64 @_m_punpcklbw
 860 // CHECK: call i64 @_mm_unpacklo_pi8