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

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