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 -x c++ -S -emit-llvm -target powerpc64le-unknown-linux-gnu -mcpu=pwr8 -DNO_WARN_X86_INTRINSICS %s \
  13 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -fsyntax-only
  14 // RUN: %clang -x c++ -S -emit-llvm -target powerpc64le-unknown-linux-gnu -mcpu=pwr9 -DNO_WARN_X86_INTRINSICS %s \
  15 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -fsyntax-only
  16
  17 // RUN: %clang -S -emit-llvm -target powerpc64-unknown-freebsd13.0 -mcpu=pwr8 -DNO_WARN_X86_INTRINSICS %s \
  18 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK-P8,CHECK,CHECK-BE
  19 // RUN: %clang -S -emit-llvm -target powerpc64le-unknown-freebsd13.0 -mcpu=pwr8 -DNO_WARN_X86_INTRINSICS %s \
  20 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK-P8,CHECK,CHECK-LE
  21 // RUN: %clang -S -emit-llvm -target powerpc64-unknown-freebsd13.0 -mcpu=pwr9 -DNO_WARN_X86_INTRINSICS %s \
  22 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK-P9,CHECK,CHECK-BE
  23 // RUN: %clang -S -emit-llvm -target powerpc64le-unknown-freebsd13.0 -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-LE
  25
  26 // RUN: %clang -S -emit-llvm -target powerpc64-ibm-aix -mcpu=pwr8 -DNO_WARN_X86_INTRINSICS %s \
  27 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK-P8,CHECK,CHECK-BE
  28 // RUN: %clang -S -emit-llvm -target powerpc64-ibm-aix -mcpu=pwr9 -DNO_WARN_X86_INTRINSICS %s \
  29 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK-P9,CHECK,CHECK-BE
  30
  31 #include <mmintrin.h>
  32
  33 unsigned long long int ull1, ull2;
  34 int i1, i2;
  35 short s[4];
  36 signed char c[8];
  37 long long int ll1;
  38 __m64 m1, m2, res;
  39
  40 void __attribute__((noinline))
  41 test_add() {
  42   res = _mm_add_pi32(m1, m2);
  43   res = _mm_add_pi16(m1, m2);
  44   res = _mm_add_pi8(m1, m2);
  45   res = _mm_adds_pu16(m1, m2);
  46   res = _mm_adds_pu8(m1, m2);
  47   res = _mm_adds_pi16(m1, m2);
  48   res = _mm_adds_pi8(m1, m2);
  49 }
  50
  51 // CHECK-LABEL: @test_add
  52
  53 // CHECK-LABEL: define available_externally i64 @_mm_add_pi32
  54 // CHECK-P9: call <2 x i64> @vec_splats(unsigned long long)
  55 // CHECK-P9: call <2 x i64> @vec_splats(unsigned long long)
  56 // CHECK-P9: call <4 x i32> @vec_add(int vector[4], int vector[4])
  57 // CHECK-P8: add nsw i32 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
  58 // CHECK-P8: add nsw i32 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
  59
  60 // CHECK-LABEL: define available_externally i64 @_mm_add_pi16
  61 // CHECK: call <2 x i64> @vec_splats
  62 // CHECK: call <2 x i64> @vec_splats
  63 // CHECK: call <8 x i16> @vec_add(short vector[8], short vector[8])
  64
  65 // CHECK-LABEL: define available_externally i64 @_mm_add_pi8
  66 // CHECK: call <2 x i64> @vec_splats
  67 // CHECK: call <2 x i64> @vec_splats
  68 // CHECK: call <16 x i8> @vec_add(signed char vector[16], signed char vector[16])
  69
  70 // CHECK-LABEL: define available_externally i64 @_mm_adds_pu16
  71 // CHECK: call <2 x i64> @vec_splats
  72 // CHECK: call <2 x i64> @vec_splats
  73 // CHECK: call <8 x i16> @vec_adds(unsigned short vector[8], unsigned short vector[8])
  74
  75 // CHECK-LABEL: define available_externally i64 @_mm_adds_pu8
  76 // CHECK: call <2 x i64> @vec_splats
  77 // CHECK: call <2 x i64> @vec_splats
  78 // CHECK: call <16 x i8> @vec_adds(unsigned char vector[16], unsigned char vector[16])
  79
  80 // CHECK-LABEL: define available_externally i64 @_mm_adds_pi16
  81 // CHECK: call <2 x i64> @vec_splats
  82 // CHECK: call <2 x i64> @vec_splats
  83 // CHECK: call <8 x i16> @vec_adds(short vector[8], short vector[8])
  84
  85 // CHECK-LABEL: define available_externally i64 @_mm_adds_pi8
  86 // CHECK: call <2 x i64> @vec_splats
  87 // CHECK: call <2 x i64> @vec_splats
  88 // CHECK: call <16 x i8> @vec_adds(signed char vector[16], signed char vector[16])
  89
  90 void __attribute__((noinline))
  91 test_alt_name_add() {
  92   res = _m_paddb(m1, m2);
  93   res = _m_paddd(m1, m2);
  94   res = _m_paddsb(m1, m2);
  95   res = _m_paddsw(m1, m2);
  96   res = _m_paddusb(m1, m2);
  97   res = _m_paddusw(m1, m2);
  98   res = _m_paddw(m1, m2);
  99 }
 100
 101 // CHECK-LABEL: @test_alt_name_add
 102
 103 // CHECK-LABEL: define available_externally i64 @_m_paddb
 104 // CHECK: call i64 @_mm_add_pi8
 105
 106 // CHECK-LABEL: define available_externally i64 @_m_paddd
 107 // CHECK: call i64 @_mm_add_pi32
 108
 109 // CHECK-LABEL: define available_externally i64 @_m_paddsb
 110 // CHECK: call i64 @_mm_adds_pi8
 111
 112 // CHECK-LABEL: define available_externally i64 @_m_paddsw
 113 // CHECK: call i64 @_mm_adds_pi16
 114
 115 // CHECK-LABEL: define available_externally i64 @_m_paddusb
 116 // CHECK: call i64 @_mm_adds_pu8
 117
 118 // CHECK-LABEL: define available_externally i64 @_m_paddusw
 119 // CHECK: call i64 @_mm_adds_pu16
 120
 121 // CHECK-LABEL: define available_externally i64 @_m_paddw
 122 // CHECK: call i64 @_mm_add_pi16
 123
 124 void __attribute__((noinline))
 125 test_cmp() {
 126   res = _mm_cmpeq_pi32(m1, m2);
 127   res = _mm_cmpeq_pi16(m1, m2);
 128   res = _mm_cmpeq_pi8(m1, m2);
 129   res = _mm_cmpgt_pi32(m1, m2);
 130   res = _mm_cmpgt_pi16(m1, m2);
 131   res = _mm_cmpgt_pi8(m1, m2);
 132 }
 133
 134 // CHECK-LABEL: @test_cmp
 135
 136 // CHECK-LABEL: define available_externally i64 @_mm_cmpeq_pi32
 137 // CHECK-P9: call <2 x i64> @vec_splats(unsigned long long)
 138 // CHECK-P9: call <2 x i64> @vec_splats(unsigned long long)
 139 // CHECK-P9: call <4 x i32> @vec_cmpeq(int vector[4], int vector[4])
 140 // CHECK-P8: %[[CMP1:[0-9a-zA-Z_.]+]] = icmp eq i32 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 141 // CHECK-P8: select i1 %[[CMP1]], i32 -1, i32 0
 142 // CHECK-P8: %[[CMP2:[0-9a-zA-Z_.]+]] = icmp eq i32 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 143 // CHECK-P8: select i1 %[[CMP2]], i32 -1, i32 0
 144
 145 // CHECK-LABEL: define available_externally i64 @_mm_cmpeq_pi16
 146 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 147 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 148 // CHECK: call <8 x i16> @vec_cmpeq(short vector[8], short vector[8])
 149
 150 // CHECK-LABEL: define available_externally i64 @_mm_cmpeq_pi8
 151 // CHECK: call i64 asm "cmpb $0,$1,$2;\0A", "=r,r,r"
 152
 153 // CHECK-LABEL: define available_externally i64 @_mm_cmpgt_pi32
 154 // CHECK-P9: call <2 x i64> @vec_splats(unsigned long long)
 155 // CHECK-P9: call <2 x i64> @vec_splats(unsigned long long)
 156 // CHECK-P9: call <4 x i32> @vec_cmpgt(int vector[4], int vector[4])
 157 // CHECK-P8: %[[CMP1:[0-9a-zA-Z_.]+]] = icmp sgt i32 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 158 // CHECK-P8: select i1 %[[CMP1]], i32 -1, i32 0
 159 // CHECK-P8: [[CMP2:[0-9a-zA-Z_.]+]] = icmp sgt i32 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 160 // CHECK-P8: select i1 %[[CMP2]], i32 -1, i32 0
 161
 162 // CHECK-LABEL: define available_externally i64 @_mm_cmpgt_pi16
 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 <8 x i16> @vec_cmpgt(short vector[8], short vector[8])
 166
 167 // CHECK-LABEL: define available_externally i64 @_mm_cmpgt_pi8
 168 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 169 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 170 // CHECK: call <16 x i8> @vec_cmpgt(signed char vector[16], signed char vector[16])
 171
 172 void __attribute__((noinline))
 173 test_alt_name_cmp() {
 174   res = _m_pcmpeqb(m1, m2);
 175   res = _m_pcmpeqd(m1, m2);
 176   res = _m_pcmpeqw(m1, m2);
 177   res = _m_pcmpgtb(m1, m2);
 178   res = _m_pcmpgtd(m1, m2);
 179   res = _m_pcmpgtw(m1, m2);
 180 }
 181
 182 // CHECK-LABEL: @test_alt_name_cmp
 183
 184 // CHECK-LABEL: define available_externally i64 @_m_pcmpeqb
 185 // CHECK: call i64 @_mm_cmpeq_pi8
 186
 187 // CHECK-LABEL: define available_externally i64 @_m_pcmpeqd
 188 // CHECK: call i64 @_mm_cmpeq_pi32
 189
 190 // CHECK-LABEL: define available_externally i64 @_m_pcmpeqw
 191 // CHECK: call i64 @_mm_cmpeq_pi16
 192
 193 // CHECK-LABEL: define available_externally i64 @_m_pcmpgtb
 194 // CHECK: call i64 @_mm_cmpgt_pi8
 195
 196 // CHECK-LABEL: define available_externally i64 @_m_pcmpgtd
 197 // CHECK: call i64 @_mm_cmpgt_pi32
 198
 199 // CHECK-LABEL: define available_externally i64 @_m_pcmpgtw
 200 // CHECK: call i64 @_mm_cmpgt_pi16
 201
 202 void __attribute__((noinline))
 203 test_convert() {
 204   ll1 = _mm_cvtm64_si64(m1);
 205   m1 = _mm_cvtsi32_si64(i1);
 206   m1 = _mm_cvtsi64_m64(ll1);
 207   i1 = _mm_cvtsi64_si32(m1);
 208 }
 209
 210 // CHECK-LABEL: @test_convert
 211
 212 // CHECK-LABEL: define available_externally i64 @_mm_cvtm64_si64
 213 // CHECK: %[[RESULT:[0-9a-zA-Z_.]+]] = load i64, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 214 // CHECK-NEXT: ret i64 %[[RESULT]]
 215
 216 // CHECK-LABEL: define available_externally i64 @_mm_cvtsi32_si64
 217 // CHECK: %[[LOAD:[0-9a-zA-Z_.]+]] = load i32, ptr %{{[0-9a-zA-Z_.]+}}
 218 // CHECK-NEXT: %[[RESULT:[0-9a-zA-Z_.]+]] = zext i32 %[[LOAD]] to i64
 219 // CHECK-NEXT: ret i64 %[[RESULT]]
 220
 221 // CHECK-LABEL: define available_externally i64 @_mm_cvtsi64_m64
 222 // CHECK: %[[RESULT:[0-9a-zA-Z_.]+]] = load i64, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 223 // CHECK-NEXT: ret i64 %[[RESULT]]
 224
 225 // CHECK-LABEL: define available_externally signext i32 @_mm_cvtsi64_si32
 226 // CHECK: %[[LOAD:[0-9a-zA-Z_.]+]] = load i64, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 227 // CHECK-NEXT: %[[RESULT:[0-9a-zA-Z_.]+]] = trunc i64 %[[LOAD]] to i32
 228 // CHECK-NEXT: ret i32 %[[RESULT]]
 229
 230 void __attribute__((noinline))
 231 test_alt_name_convert() {
 232   m1 = _m_from_int(i1);
 233   m1 = _m_from_int64(ll1);
 234   i1 = _m_to_int(m1);
 235   ll1 = _m_to_int64(m1);
 236 }
 237
 238 // CHECK-LABEL: @test_alt_name_convert
 239
 240 // CHECK-LABEL: define available_externally i64 @_m_from_int
 241 // CHECK: call i64 @_mm_cvtsi32_si64
 242
 243 // CHECK-LABEL: define available_externally i64 @_m_from_int64
 244 // CHECK: %[[RESULT:[0-9a-zA-Z_.]+]] = load i64, ptr %{{[0-9a-zA-Z_.]+}}
 245 // CHECK-NEXT: ret i64 %[[RESULT]]
 246
 247 // CHECK-LABEL: define available_externally signext i32 @_m_to_int
 248 // CHECK: call signext i32 @_mm_cvtsi64_si32
 249
 250 // CHECK-LABEL: define available_externally i64 @_m_to_int64
 251 // CHECK: %[[RESULT:[0-9a-zA-Z_.]+]] = load i64, ptr %{{[0-9a-zA-Z_.]+}}
 252 // CHECK-NEXT: ret i64 %[[RESULT]]
 253
 254 void __attribute__((noinline))
 255 test_empty() {
 256   _mm_empty();
 257   _m_empty();
 258 }
 259
 260 // CHECK-LABEL: @test_empty
 261
 262 // CHECK-LABEL: define available_externally void @_mm_empty
 263 // CHECK-NEXT: entry
 264 // CHECK-NEXT: ret void
 265
 266 // CHECK-LABEL: define available_externally void @_m_empty
 267 // CHECK-NEXT: entry
 268 // CHECK-NEXT: ret void
 269
 270 void __attribute__((noinline))
 271 test_logic() {
 272   res = _mm_and_si64(m1, m2);
 273   res = _mm_andnot_si64(m1, m2);
 274   res = _mm_or_si64(m1, m2);
 275   res = _mm_xor_si64(m1, m2);
 276 }
 277
 278 // CHECK-LABEL: @test_logic
 279
 280 // CHECK-LABEL: define available_externally i64 @_mm_and_si64
 281 // CHECK: and i64 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 282
 283 // CHECK-LABEL: define available_externally i64 @_mm_andnot_si64
 284 // CHECK: %[[XOR:[0-9a-zA-Z_.]+]] = xor i64 %{{[0-9a-zA-Z_.]+}}, -1
 285 // CHECK: and i64 %[[XOR]], %{{[0-9a-zA-Z_.]+}}
 286
 287 // CHECK-LABEL: define available_externally i64 @_mm_or_si64
 288 // CHECK: or i64 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 289
 290 // CHECK-LABEL: define available_externally i64 @_mm_xor_si64
 291 // CHECK: xor i64 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 292
 293 void __attribute__((noinline))
 294 test_alt_name_logic() {
 295   res = _m_pand(m1, m2);
 296   res = _m_pandn(m1, m2);
 297   res = _m_por(m1, m2);
 298   res = _m_pxor(m1, m2);
 299 }
 300
 301 // CHECK-LABEL: @test_alt_name_logic
 302
 303 // CHECK-LABEL: define available_externally i64 @_m_pand
 304 // CHECK: call i64 @_mm_and_si64
 305
 306 // CHECK-LABEL: define available_externally i64 @_m_pandn
 307 // CHECK: call i64 @_mm_andnot_si64
 308
 309 // CHECK-LABEL: define available_externally i64 @_m_por
 310 // CHECK: call i64 @_mm_or_si64
 311
 312 // CHECK-LABEL: define available_externally i64 @_m_pxor
 313 // CHECK: call i64 @_mm_xor_si64
 314
 315 void __attribute__((noinline))
 316 test_madd() {
 317   res = _mm_madd_pi16(m1, m2);
 318   res = _m_pmaddwd(m1, m2);
 319 }
 320
 321 // CHECK-LABEL: @test_madd
 322
 323 // CHECK-LABEL: define available_externally i64 @_mm_madd_pi16
 324 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 325 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 326 // CHECK: call <4 x i32> @vec_vmsumshm
 327
 328 // CHECK-LABEL: define available_externally i64 @_m_pmaddwd
 329 // CHECK: call i64 @_mm_madd_pi16
 330
 331 void __attribute__((noinline))
 332 test_mul() {
 333   res = _mm_mulhi_pi16(m1, m2);
 334   res = _mm_mullo_pi16(m1, m2);
 335 }
 336
 337 // CHECK-LABEL: @test_mul
 338
 339 // CHECK-LABEL: define available_externally i64 @_mm_mulhi_pi16
 340 // 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>
 341 // 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>
 342 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 343 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 344 // CHECK: call <4 x i32> @vec_vmulesh
 345 // CHECK: call <4 x i32> @vec_vmulosh
 346 // CHECK: call <4 x i32> @vec_perm(int vector[4], int vector[4], unsigned char vector[16])
 347
 348 // CHECK-LABEL: define available_externally i64 @_mm_mullo_pi16
 349 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 350 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 351 // CHECK: mul <8 x i16> %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 352
 353 void __attribute__((noinline))
 354 test_alt_name_mul() {
 355   res = _m_pmulhw(m1, m2);
 356   res = _m_pmullw(m1, m2);
 357 }
 358
 359 // CHECK-LABEL: @test_alt_name_mul
 360
 361 // CHECK-LABEL: define available_externally i64 @_m_pmulhw
 362 // CHECK: call i64 @_mm_mulhi_pi16
 363
 364 // CHECK-LABEL: define available_externally i64 @_m_pmullw
 365 // CHECK: call i64 @_mm_mullo_pi16
 366
 367 void __attribute__((noinline))
 368 test_packs() {
 369   res = _mm_packs_pu16((__m64)ull1, (__m64)ull2);
 370   res = _mm_packs_pi16((__m64)ull1, (__m64)ull2);
 371   res = _mm_packs_pi32((__m64)ull1, (__m64)ull2);
 372 }
 373
 374 // CHECK-LABEL: @test_packs
 375
 376 // CHECK-LABEL: define available_externally i64 @_mm_packs_pu16
 377 // CHECK: call <8 x i16> @vec_cmplt
 378 // CHECK: call <16 x i8> @vec_packs(unsigned short vector[8], unsigned short vector[8])
 379 // CHECK: call <16 x i8> @vec_pack(bool vector[8], bool vector[8])
 380 // 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_.]+}})
 381
 382 // CHECK-LABEL: define available_externally i64 @_mm_packs_pi16
 383 // CHECK: call <16 x i8> @vec_packs(short vector[8], short vector[8])
 384
 385 // CHECK-LABEL: define available_externally i64 @_mm_packs_pi32
 386 // CHECK: call <8 x i16> @vec_packs(int vector[4], int vector[4])
 387
 388 void __attribute__((noinline))
 389 test_alt_name_packs() {
 390   res = _m_packssdw(m1, m2);
 391   res = _m_packsswb(m1, m2);
 392   res = _m_packuswb(m1, m2);
 393 }
 394
 395 // CHECK-LABEL: @test_alt_name_packs
 396
 397 // CHECK-LABEL: define available_externally i64 @_m_packssdw
 398 // CHECK: call i64 @_mm_packs_pi32
 399
 400 // CHECK-LABEL: define available_externally i64 @_m_packsswb
 401 // CHECK: call i64 @_mm_packs_pi16
 402
 403 // CHECK-LABEL: define available_externally i64 @_m_packuswb
 404 // CHECK: call i64 @_mm_packs_pu16
 405
 406 void __attribute__((noinline))
 407 test_set() {
 408   m1 = _mm_set_pi32(2134, -128);
 409   m1 = _mm_set_pi16(2134, -128, 1234, 6354);
 410   m1 = _mm_set_pi8(-128, 10, 0, 123, -1, -5, 127, 5);
 411 }
 412
 413 // CHECK-LABEL: @test_set
 414
 415 // CHECK-LABEL: define available_externally i64 @_mm_set_pi32
 416 // CHECK-COUNT-2: store i32 %{{[0-9a-zA-Z_.]+}}, ptr
 417
 418 // CHECK-LABEL: define available_externally i64 @_mm_set_pi16
 419 // CHECK-COUNT-4: store i16 %{{[0-9a-zA-Z_.]+}}, ptr
 420
 421 // CHECK-LABEL: define available_externally i64 @_mm_set_pi8
 422 // CHECK-COUNT-8: store i8 %{{[0-9a-zA-Z_.]+}}, ptr
 423
 424 void __attribute__((noinline))
 425 test_set1() {
 426   res = _mm_set1_pi32(i1);
 427   res = _mm_set1_pi16(s[0]);
 428   res = _mm_set1_pi8(c[0]);
 429 }
 430
 431 // CHECK-LABEL: @test_set1
 432
 433 // CHECK-LABEL: define available_externally i64 @_mm_set1_pi32
 434 // CHECK-COUNT-2: store i32 %{{[0-9a-zA-Z_.]+}}, ptr
 435
 436 // CHECK-LABEL: define available_externally i64 @_mm_set1_pi16
 437 // CHECK-P9: call <8 x i16> @vec_splats(short)
 438 // CHECK-P9: extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 0
 439 // CHECK-P8: %[[ADDR1:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 440 // CHECK-P8: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR1]], align 8
 441 // CHECK-P8: %[[ADDR2:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 442 // CHECK-P8: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR2]], align 2
 443 // CHECK-P8: %[[ADDR3:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 2
 444 // CHECK-P8: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR3]], align 4
 445 // CHECK-P8: %[[ADDR4:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 3
 446 // CHECK-P8: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR4]], align 2
 447
 448
 449 // CHECK-LABEL: define available_externally i64 @_mm_set1_pi8
 450 // CHECK: call <16 x i8> @vec_splats(signed char)
 451 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <16 x i8> %{{[0-9a-zA-Z_.]+}} to <2 x i64>
 452 // CHECK: extractelement <2 x i64> %[[CAST]], i32 0
 453
 454 void __attribute__((noinline))
 455 test_setr() {
 456   res = _mm_setr_pi32(i1, i2);
 457   res = _mm_setr_pi16(s[0], s[1], s[2], s[3]);
 458   res = _mm_setr_pi8(c[0], c[1], c[2], c[3], c[4], c[5], c[6], c[7]);
 459 }
 460
 461 // CHECK-LABEL: @test_setr
 462
 463 // CHECK-LABEL: define available_externally i64 @_mm_setr_pi32
 464 // CHECK: %[[ADDR1:[0-9a-zA-Z_.]+]] = getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 465 // CHECK: store i32 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR1]], align 8
 466 // CHECK: %[[ADDR2:[0-9a-zA-Z_.]+]] = getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 467 // CHECK: store i32 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR2]], align 4
 468
 469 // CHECK-LABEL: define available_externally i64 @_mm_setr_pi16
 470 // CHECK: call i64 @_mm_set_pi16
 471
 472 // CHECK-LABEL: define available_externally i64 @_mm_setr_pi8
 473 // CHECK: call i64 @_mm_set_pi8
 474
 475 void __attribute__((noinline))
 476 test_setzero() {
 477   res = _mm_setzero_si64();
 478 }
 479
 480 // CHECK-LABEL: @test_setzero
 481
 482 // CHECK-LABEL: define available_externally i64 @_mm_setzero_si64
 483 // CHECK: entry
 484 // CHECK-NEXT: ret i64 0
 485
 486 void __attribute__((noinline))
 487 test_sll() {
 488   res = _mm_sll_pi16(m1, m2);
 489   res = _mm_sll_pi32(m1, m2);
 490   res = _mm_sll_si64(m1, m2);
 491   res = _mm_slli_pi16(m1, i1);
 492   res = _mm_slli_pi32(m1, i1);
 493   res = _mm_slli_si64(m1, i1);
 494 }
 495
 496 // CHECK-LABEL: @test_sll
 497
 498 // CHECK-LABEL: define available_externally i64 @_mm_sll_pi16
 499 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp ule i64 %{{[0-9a-zA-Z_.]+}}, 15
 500 // CHECK-NEXT: br i1 %[[CMP]]
 501 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 502 // CHECK: trunc i64 %{{[0-9a-zA-Z_.]+}} to i16
 503 // CHECK: call <8 x i16> @vec_splats(unsigned short)
 504 // CHECK: call <8 x i16> @vec_sl(short vector[8], unsigned short vector[8])
 505 // CHECK: store i64 0, ptr
 506
 507 // CHECK-LABEL: define available_externally i64 @_mm_sll_pi32
 508 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i64 %{{[0-9a-zA-Z_.]+}} to i32
 509 // CHECK: shl i32 %{{[0-9a-zA-Z_.]+}}, %[[TRUNC]]
 510 // CHECK: trunc i64 %{{[0-9a-zA-Z_.]+}} to i32
 511 // CHECK: shl i32 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 512
 513 // CHECK-LABEL: define available_externally i64 @_mm_sll_si64
 514 // CHECK: shl i64 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 515
 516 // CHECK-LABEL: define available_externally i64 @_mm_slli_pi16
 517 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = sext i32 %{{[0-9a-zA-Z_.]+}} to i64
 518 // CHECK: call i64 @_mm_sll_pi16(i64 noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef %[[EXT]])
 519
 520 // CHECK-LABEL: define available_externally i64 @_mm_slli_pi32
 521 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = sext i32 %{{[0-9a-zA-Z_.]+}} to i64
 522 // CHECK: call i64 @_mm_sll_pi32(i64 noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef %[[EXT]])
 523
 524 // CHECK-LABEL: define available_externally i64 @_mm_slli_si64
 525 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = zext i32 %{{[0-9a-zA-Z_.]+}} to i64
 526 // CHECK: shl i64 %{{[0-9a-zA-Z_.]+}}, %[[EXT]]
 527
 528 void __attribute__((noinline))
 529 test_alt_name_sll() {
 530   res = _m_pslld(m1, m2);
 531   res = _m_pslldi(m1, i1);
 532   res = _m_psllq(m1, m2);
 533   res = _m_psllqi(m1, i1);
 534   res = _m_psllw(m1, m2);
 535   res = _m_psllwi(m1, i1);
 536 }
 537
 538 // CHECK-LABEL: @test_alt_name_sll
 539
 540 // CHECK-LABEL: define available_externally i64 @_m_pslld
 541 // CHECK: call i64 @_mm_sll_pi32
 542
 543 // CHECK-LABEL: define available_externally i64 @_m_pslldi
 544 // CHECK: call i64 @_mm_slli_pi32
 545
 546 // CHECK-LABEL: define available_externally i64 @_m_psllq
 547 // CHECK: call i64 @_mm_sll_si64
 548
 549 // CHECK-LABEL: define available_externally i64 @_m_psllqi
 550 // CHECK: call i64 @_mm_slli_si64
 551
 552 // CHECK-LABEL: define available_externally i64 @_m_psllw
 553 // CHECK: call i64 @_mm_sll_pi16
 554
 555 // CHECK-LABEL: define available_externally i64 @_m_psllwi
 556 // CHECK: call i64 @_mm_slli_pi16
 557
 558 void __attribute__((noinline))
 559 test_sra() {
 560   res = _mm_sra_pi32(m1, m2);
 561   res = _mm_sra_pi16(m1, m2);
 562   res = _mm_srai_pi32(m1, i1);
 563   res = _mm_srai_pi16(m1, i1);
 564 }
 565
 566 // CHECK-LABEL: @test_sra
 567
 568 // CHECK-LABEL: define available_externally i64 @_mm_sra_pi32
 569 // CHECK: %[[TRUNC1:[0-9a-zA-Z_.]+]] = trunc i64 %{{[0-9a-zA-Z_.]+}} to i32
 570 // CHECK: ashr i32 %{{[0-9a-zA-Z_.]+}}, %[[TRUNC1]]
 571 // CHECK: %[[TRUNC2:[0-9a-zA-Z_.]+]] = trunc i64 %{{[0-9a-zA-Z_.]+}} to i32
 572 // CHECK: ashr i32 %{{[0-9a-zA-Z_.]+}}, %[[TRUNC2]]
 573
 574 // CHECK-LABEL: define available_externally i64 @_mm_sra_pi16
 575 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp ule i64 %{{[0-9a-zA-Z_.]+}}, 15
 576 // CHECK: br i1 %[[CMP]]
 577 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 578 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i64 %{{[0-9a-zA-Z_.]+}} to i16
 579 // CHECK: call <8 x i16> @vec_splats(unsigned short)(i16 noundef zeroext %[[TRUNC]])
 580 // CHECK: call <8 x i16> @vec_sra(short vector[8], unsigned short vector[8])
 581 // CHECK: store i64 0, ptr
 582
 583 // CHECK-LABEL: define available_externally i64 @_mm_srai_pi32
 584 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = sext i32 %{{[0-9a-zA-Z_.]+}} to i64
 585 // CHECK: call i64 @_mm_sra_pi32(i64 noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef %[[EXT]])
 586
 587 // CHECK-LABEL: define available_externally i64 @_mm_srai_pi16
 588 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = sext i32 %{{[0-9a-zA-Z_.]+}} to i64
 589 // CHECK: call i64 @_mm_sra_pi16(i64 noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef %[[EXT]])
 590
 591 void __attribute__((noinline))
 592 test_alt_name_sra() {
 593   res = _m_psrad(m1, m2);
 594   res = _m_psraw(m1, m2);
 595   res = _m_psradi(m1, i1);
 596   res = _m_psrawi(m1, i1);
 597 }
 598
 599 // CHECK-LABEL: @test_alt_name_sra
 600
 601 // CHECK-LABEL: define available_externally i64 @_m_psrad
 602 // CHECK: call i64 @_mm_sra_pi32
 603
 604 // CHECK-LABEL: define available_externally i64 @_m_psraw
 605 // CHECK: call i64 @_mm_sra_pi16
 606
 607 // CHECK-LABEL: define available_externally i64 @_m_psradi
 608 // CHECK: call i64 @_mm_srai_pi32
 609
 610 // CHECK-LABEL: define available_externally i64 @_m_psrawi
 611 // CHECK: call i64 @_mm_srai_pi16
 612
 613 void __attribute__((noinline))
 614 test_srl() {
 615   res = _mm_srl_si64(m1, m2);
 616   res = _mm_srl_pi32(m1, m2);
 617   res = _mm_srl_pi16(m1, m2);
 618   res = _mm_srli_si64(m1, i1);
 619   res = _mm_srli_pi32(m1, i1);
 620   res = _mm_srli_pi16(m1, i1);
 621 }
 622
 623 // CHECK-LABEL: @test_srl
 624
 625 // CHECK-LABEL: define available_externally i64 @_mm_srl_si64
 626 // CHECK: lshr i64 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 627
 628 // CHECK-LABEL: define available_externally i64 @_mm_srl_pi32
 629 // CHECK: %[[TRUNC1:[0-9a-zA-Z_.]+]] = trunc i64 %{{[0-9a-zA-Z_.]+}} to i32
 630 // CHECK: lshr i32 %{{[0-9a-zA-Z_.]+}}, %[[TRUNC1]]
 631 // CHECK: %[[TRUNC2:[0-9a-zA-Z_.]+]] = trunc i64 %{{[0-9a-zA-Z_.]+}} to i32
 632 // CHECK: lshr i32 %{{[0-9a-zA-Z_.]+}}, %[[TRUNC2]]
 633
 634 // CHECK-LABEL: define available_externally i64 @_mm_srl_pi16
 635 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp ule i64 %{{[0-9a-zA-Z_.]+}}, 15
 636 // CHECK: br i1 %[[CMP]]
 637 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 638 // CHECK: trunc i64 %{{[0-9a-zA-Z_.]+}} to i16
 639 // CHECK: call <8 x i16> @vec_splats(unsigned short)
 640 // CHECK: call <8 x i16> @vec_sr(unsigned short vector[8], unsigned short vector[8])
 641 // CHECK: store i64 0, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 642
 643 // CHECK-LABEL: define available_externally i64 @_mm_srli_si64
 644 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = zext i32 %{{[0-9a-zA-Z_.]+}} to i64
 645 // CHECK: lshr i64 %{{[0-9a-zA-Z_.]+}}, %[[EXT]]
 646
 647 // CHECK-LABEL: define available_externally i64 @_mm_srli_pi32
 648 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = sext i32 %{{[0-9a-zA-Z_.]+}} to i64
 649 // CHECK: call i64 @_mm_srl_pi32(i64 noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef %[[EXT]])
 650
 651 // CHECK-LABEL: define available_externally i64 @_mm_srli_pi16
 652 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = sext i32 %{{[0-9a-zA-Z_.]+}} to i64
 653 // CHECK: call i64 @_mm_srl_pi16(i64 noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef %[[EXT]])
 654
 655 void __attribute__((noinline))
 656 test_alt_name_srl() {
 657   res = _m_psrlq(m1, m2);
 658   res = _m_psrld(m1, m2);
 659   res = _m_psrlw(m1, m2);
 660   res = _m_psrlqi(m1, i1);
 661   res = _m_psrldi(m1, i1);
 662   res = _m_psrlwi(m1, i1);
 663 }
 664
 665 // CHECK-LABEL: @test_alt_name_srl
 666
 667 // CHECK-LABEL: define available_externally i64 @_m_psrlq
 668 // CHECK: call i64 @_mm_srl_si64
 669
 670 // CHECK-LABEL: define available_externally i64 @_m_psrld
 671 // CHECK: call i64 @_mm_srl_pi32
 672
 673 // CHECK-LABEL: define available_externally i64 @_m_psrlw
 674 // CHECK: call i64 @_mm_srl_pi16
 675
 676 // CHECK-LABEL: define available_externally i64 @_m_psrlqi
 677 // CHECK: call i64 @_mm_srli_si64
 678
 679 // CHECK-LABEL: define available_externally i64 @_m_psrldi
 680 // CHECK: call i64 @_mm_srli_pi32
 681
 682 // CHECK-LABEL: define available_externally i64 @_m_psrlwi
 683 // CHECK: call i64 @_mm_srli_pi16
 684
 685 void __attribute__((noinline))
 686 test_sub() {
 687   res = _mm_sub_pi32(m1, m2);
 688   res = _mm_sub_pi16(m1, m2);
 689   res = _mm_sub_pi8(m1, m2);
 690   res = _mm_subs_pi16(m1, m2);
 691   res = _mm_subs_pi8(m1, m2);
 692   res = _mm_subs_pu16(m1, m2);
 693   res = _mm_subs_pu8(m1, m2);
 694 }
 695
 696 // CHECK-LABEL: @test_sub
 697
 698 // CHECK-LABEL: define available_externally i64 @_mm_sub_pi32
 699 // CHECK-P8-COUNT-2: getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 700 // CHECK-P8: sub nsw i32 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 701 // CHECK-P8-COUNT-2: getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 702 // CHECK-P8: sub nsw i32 %{{[0-9a-zA-Z_.]+}}, %{{[0-9a-zA-Z_.]+}}
 703 // CHECK-P9: call <2 x i64> @vec_splats(unsigned long long)
 704 // CHECK-P9: call <2 x i64> @vec_splats(unsigned long long)
 705 // CHECK-P9: call <4 x i32> @vec_sub(int vector[4], int vector[4])
 706
 707 // CHECK-LABEL: define available_externally i64 @_mm_sub_pi16
 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 <8 x i16> @vec_sub(short vector[8], short vector[8])
 711
 712 // CHECK-LABEL: define available_externally i64 @_mm_sub_pi8
 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 <16 x i8> @vec_sub(signed char vector[16], signed char vector[16])
 716
 717 // CHECK-LABEL: define available_externally i64 @_mm_subs_pi16
 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 <8 x i16> @vec_subs(short vector[8], short vector[8])
 721
 722 // CHECK-LABEL: define available_externally i64 @_mm_subs_pi8
 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 <16 x i8> @vec_subs(signed char vector[16], signed char vector[16])
 726
 727 // CHECK-LABEL: define available_externally i64 @_mm_subs_pu16
 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 <8 x i16> @vec_subs(unsigned short vector[8], unsigned short vector[8])
 731
 732 // CHECK-LABEL: define available_externally i64 @_mm_subs_pu8
 733 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 734 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 735 // CHECK: call <16 x i8> @vec_subs(unsigned char vector[16], unsigned char vector[16])
 736
 737 void __attribute__((noinline))
 738 test_alt_name_sub() {
 739   res = _m_psubd(m1, m2);
 740   res = _m_psubw(m1, m2);
 741   res = _m_psubb(m1, m2);
 742   res = _m_psubsw(m1, m2);
 743   res = _m_psubsb(m1, m2);
 744   res = _m_psubusw(m1, m2);
 745   res = _m_psubusb(m1, m2);
 746 }
 747
 748 // CHECK-LABEL: @test_alt_name_sub
 749
 750 // CHECK-LABEL: define available_externally i64 @_m_psubd
 751 // CHECK: call i64 @_mm_sub_pi32
 752
 753 // CHECK-LABEL: define available_externally i64 @_m_psubw
 754 // CHECK: call i64 @_mm_sub_pi16
 755
 756 // CHECK-LABEL: define available_externally i64 @_m_psubb
 757 // CHECK: call i64 @_mm_sub_pi8
 758
 759 // CHECK-LABEL: define available_externally i64 @_m_psubsw
 760 // CHECK: call i64 @_mm_subs_pi16
 761
 762 // CHECK-LABEL: define available_externally i64 @_m_psubsb
 763 // CHECK: call i64 @_mm_subs_pi8
 764
 765 // CHECK-LABEL: define available_externally i64 @_m_psubusw
 766 // CHECK: call i64 @_mm_subs_pu16
 767
 768 // CHECK-LABEL: define available_externally i64 @_m_psubusb
 769 // CHECK: call i64 @_mm_subs_pu8
 770
 771 void __attribute__((noinline))
 772 test_unpack() {
 773   res = _mm_unpackhi_pi32(m1, m2);
 774   res = _mm_unpackhi_pi16(m1, m2);
 775   res = _mm_unpackhi_pi8(m1, m2);
 776   res = _mm_unpacklo_pi32(m1, m2);
 777   res = _mm_unpacklo_pi16(m1, m2);
 778   res = _mm_unpacklo_pi8(m1, m2);
 779 }
 780
 781 // CHECK-LABEL: @test_unpack
 782
 783 // CHECK-LABEL: define available_externally i64 @_mm_unpackhi_pi32
 784 // CHECK: getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 785 // CHECK: %[[ADDR1:[0-9a-zA-Z_.]+]] = getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 786 // CHECK: store i32 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR1]], align 8
 787 // CHECK: getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 788 // CHECK: %[[ADDR2:[0-9a-zA-Z_.]+]] = getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 789 // CHECK: store i32 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR2]], align 4
 790
 791 // CHECK-LABEL: define available_externally i64 @_mm_unpackhi_pi16
 792 // CHECK: getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 2
 793 // CHECK: %[[ADDR1:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 794 // CHECK: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR1]], align 8
 795 // CHECK: getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 2
 796 // CHECK: %[[ADDR2:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 797 // CHECK: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR2]], align 2
 798 // CHECK: getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 3
 799 // CHECK: %[[ADDR3:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 2
 800 // CHECK: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR3]], align 4
 801 // CHECK: getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 3
 802 // CHECK: %[[ADDR4:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 3
 803 // CHECK: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR4]], align 2
 804
 805 // CHECK-LABEL: define available_externally i64 @_mm_unpackhi_pi8
 806 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 807 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 808 // CHECK: call <16 x i8> @vec_mergel(unsigned char vector[16], unsigned char vector[16])
 809
 810 // CHECK-LABEL: define available_externally i64 @_mm_unpacklo_pi32
 811 // CHECK: getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 812 // CHECK: %[[ADDR1:[0-9a-zA-Z_.]+]] = getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 813 // CHECK: store i32 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR1]], align 8
 814 // CHECK: getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 815 // CHECK: %[[ADDR2:[0-9a-zA-Z_.]+]] = getelementptr inbounds [2 x i32], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 816 // CHECK: store i32 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR2]], align 4
 817
 818 // CHECK-LABEL: define available_externally i64 @_mm_unpacklo_pi16
 819 // CHECK: getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 820 // CHECK: %[[ADDR1:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 821 // CHECK: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR1]], align 8
 822 // CHECK: getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 0
 823 // CHECK: %[[ADDR2:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 824 // CHECK: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR2]], align 2
 825 // CHECK: getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 826 // CHECK: %[[ADDR3:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 2
 827 // CHECK: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR3]], align 4
 828 // CHECK: getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 1
 829 // CHECK: %[[ADDR4:[0-9a-zA-Z_.]+]] = getelementptr inbounds [4 x i16], ptr %{{[0-9a-zA-Z_.]+}}, i64 0, i64 3
 830 // CHECK: store i16 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR4]], align 2
 831
 832 // CHECK-LABEL: define available_externally i64 @_mm_unpacklo_pi8
 833 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 834 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 835 // CHECK: call <16 x i8> @vec_mergel(unsigned char vector[16], unsigned char vector[16])
 836
 837 void __attribute__((noinline))
 838 test_alt_name_unpack() {
 839   res = _m_punpckhdq(m1, m2);
 840   res = _m_punpckhwd(m1, m2);
 841   res = _m_punpckhbw(m1, m2);
 842   res = _m_punpckldq(m1, m2);
 843   res = _m_punpcklwd(m1, m2);
 844   res = _m_punpcklbw(m1, m2);
 845 }
 846
 847 // CHECK-LABEL: @test_alt_name_unpack
 848
 849 // CHECK-LABEL: define available_externally i64 @_m_punpckhdq
 850 // CHECK: call i64 @_mm_unpackhi_pi32
 851
 852 // CHECK-LABEL: define available_externally i64 @_m_punpckhwd
 853 // CHECK: call i64 @_mm_unpackhi_pi16
 854
 855 // CHECK-LABEL: define available_externally i64 @_m_punpckhbw
 856 // CHECK: call i64 @_mm_unpackhi_pi8
 857
 858 // CHECK-LABEL: define available_externally i64 @_m_punpckldq
 859 // CHECK: call i64 @_mm_unpacklo_pi32
 860
 861 // CHECK-LABEL: define available_externally i64 @_m_punpcklwd
 862 // CHECK: call i64 @_mm_unpacklo_pi16
 863
 864 // CHECK-LABEL: define available_externally i64 @_m_punpcklbw
 865 // CHECK: call i64 @_mm_unpacklo_pi8