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 -S -emit-llvm -target powerpc64-ibm-aix -mcpu=pwr8 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \
  12 // RUN:   -ffp-contract=off -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK,CHECK-BE
  13 // RUN: %clang -S -emit-llvm -target powerpc64-ibm-aix -mcpu=pwr10 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \
  14 // RUN:   -ffp-contract=off -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK-P10
  15
  16 // CHECK-BE-DAG: @_mm_movemask_pd.__perm_mask = internal constant <4 x i32> <i32 -2139062144, i32 -2139062144, i32 -2139062144, i32 -2139078656>, align 16
  17 // CHECK-BE-DAG: @_mm_shuffle_epi32.__permute_selectors = internal constant [4 x i32] [i32 66051, i32 67438087, i32 134810123, i32 202182159], align 4
  18 // CHECK-BE-DAG: @_mm_shufflehi_epi16.__permute_selectors = internal constant [4 x i16] [i16 2057, i16 2571, i16 3085, i16 3599], align 2
  19 // CHECK-BE-DAG: @_mm_shufflelo_epi16.__permute_selectors = internal constant [4 x i16] [i16 1, i16 515, i16 1029, i16 1543], align 2
  20
  21 // CHECK-LE-DAG: @_mm_movemask_pd.__perm_mask = internal constant <4 x i32> <i32 -2139094976, i32 -2139062144, i32 -2139062144, i32 -2139062144>, align 16
  22 // CHECK-LE-DAG: @_mm_shuffle_epi32.__permute_selectors = internal constant [4 x i32] [i32 50462976, i32 117835012, i32 185207048, i32 252579084], align 4
  23 // CHECK-LE-DAG: @_mm_shufflehi_epi16.__permute_selectors = internal constant [4 x i16] [i16 2312, i16 2826, i16 3340, i16 3854], align 2
  24 // CHECK-LE-DAG: @_mm_shufflelo_epi16.__permute_selectors = internal constant [4 x i16] [i16 256, i16 770, i16 1284, i16 1798], align 2
  25
  26 #include <emmintrin.h>
  27
  28 __m128i resi, mi1, mi2;
  29 __m128i *mip;
  30 double dp[2];
  31 __m128d resd, md1, md2;
  32 __m64 res64, m641, m642;
  33 __m128 res, m1;
  34 int i;
  35 char chs[16];
  36 int is[4];
  37 short ss[8];
  38 long long i64s[2];
  39
  40 void __attribute__((noinline))
  41 test_add() {
  42   resi = _mm_add_epi64(mi1, mi2);
  43   resi = _mm_add_epi32(mi1, mi2);
  44   resi = _mm_add_epi16(mi1, mi2);
  45   resi = _mm_add_epi8(mi1, mi2);
  46   resd = _mm_add_pd(md1, md2);
  47   resd = _mm_add_sd(md1, md2);
  48   res64 = _mm_add_si64(m641, m642);
  49   resi = _mm_adds_epi16(mi1, mi2);
  50   resi = _mm_adds_epi8(mi1, mi2);
  51   resi = _mm_adds_epu16(mi1, mi2);
  52   resi = _mm_adds_epu8(mi1, mi2);
  53 }
  54
  55 // CHECK-LABEL: @test_add
  56
  57 // CHECK-LABEL: define available_externally <2 x i64> @_mm_add_epi64
  58 // CHECK: add <2 x i64>
  59
  60 // CHECK-LABEL: define available_externally <2 x i64> @_mm_add_epi32
  61 // CHECK: add <4 x i32>
  62
  63 // CHECK-LABEL: define available_externally <2 x i64> @_mm_add_epi16
  64 // CHECK: add <8 x i16>
  65
  66 // CHECK-LABEL: define available_externally <2 x i64> @_mm_add_epi8
  67 // CHECK: add <16 x i8>
  68
  69 // CHECK-LABEL: define available_externally <2 x double> @_mm_add_pd
  70 // CHECK: fadd <2 x double>
  71
  72 // CHECK-LABEL: define available_externally <2 x double> @_mm_add_sd
  73 // CHECK: fadd double
  74
  75 // CHECK-LABEL: define available_externally i64 @_mm_add_si64
  76 // CHECK: add i64
  77
  78 // CHECK-LABEL: define available_externally <2 x i64> @_mm_adds_epi16
  79 // CHECK: call <8 x i16> @vec_adds(short vector[8], short vector[8])
  80
  81 // CHECK-LABEL: define available_externally <2 x i64> @_mm_adds_epi8
  82 // CHECK: call <16 x i8> @vec_adds(signed char vector[16], signed char vector[16])
  83
  84 // CHECK-LABEL: define available_externally <2 x i64> @_mm_adds_epu16
  85 // CHECK: call <8 x i16> @vec_adds(unsigned short vector[8], unsigned short vector[8])
  86
  87 // CHECK-LABEL: define available_externally <2 x i64> @_mm_adds_epu8
  88 // CHECK: call <16 x i8> @vec_adds(unsigned char vector[16], unsigned char vector[16])
  89
  90 void __attribute__((noinline))
  91 test_avg() {
  92   resi = _mm_avg_epu16(mi1, mi2);
  93   resi = _mm_avg_epu8(mi1, mi2);
  94 }
  95
  96 // CHECK-LABEL: @test_avg
  97
  98 // CHECK-LABEL: define available_externally <2 x i64> @_mm_avg_epu16
  99 // CHECK: call <8 x i16> @vec_avg(unsigned short vector[8], unsigned short vector[8])
 100
 101 // CHECK-LABEL: define available_externally <2 x i64> @_mm_avg_epu8
 102 // CHECK: call <16 x i8> @vec_avg(unsigned char vector[16], unsigned char vector[16])
 103
 104 void __attribute__((noinline))
 105 test_bs() {
 106   resi = _mm_bslli_si128(mi1, i);
 107   resi = _mm_bsrli_si128(mi1, i);
 108 }
 109
 110 // CHECK-LABEL: @test_bs
 111
 112 // CHECK-LABEL: define available_externally <2 x i64> @_mm_bslli_si128
 113 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16
 114 // CHECK: br i1 %[[CMP]]
 115 // 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_.]+}})
 116 // CHECK: store <16 x i8> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 117
 118 // CHECK-LABEL: define available_externally <2 x i64> @_mm_bsrli_si128
 119 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16
 120 // CHECK: br i1 %[[CMP]]
 121 // CHECK-LE: call i1 @llvm.is.constant
 122 // CHECK-LE: %[[SUB:[0-9a-zA-Z_.]+]] = sub nsw i32 16, %{{[0-9a-zA-Z_.]+}}
 123 // 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]])
 124 // CHECK-LE: %[[MUL:[0-9a-zA-Z_.]+]] = mul nsw i32 %{{[0-9a-zA-Z_.]+}}, 8
 125 // CHECK-LE: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %[[MUL]] to i8
 126 // CHECK-LE: call <16 x i8> @vec_splats(unsigned char)(i8 noundef zeroext %[[TRUNC]])
 127 // CHECK-LE: call <16 x i8> @vec_sro(unsigned char vector[16], unsigned char vector[16])
 128 // CHECK-LE: store <16 x i8> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 129 // CHECK-BE: %[[MUL:[0-9a-zA-Z_.]+]] = mul nsw i32 %{{[0-9a-zA-Z_.]+}}, 8
 130 // CHECK-BE: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %[[MUL]] to i8
 131 // CHECK-BE: call <16 x i8> @vec_splats(unsigned char)(i8 noundef zeroext %[[TRUNC]])
 132 // CHECK-BE: call <16 x i8> @vec_slo(unsigned char vector[16], unsigned char vector[16])
 133 // CHECK-BE: store <16 x i8> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 134
 135 void __attribute__((noinline))
 136 test_cast() {
 137   res = _mm_castpd_ps(md1);
 138   resi = _mm_castpd_si128(md1);
 139   resd = _mm_castps_pd(m1);
 140   resi = _mm_castps_si128(m1);
 141   resd = _mm_castsi128_pd(mi1);
 142   res = _mm_castsi128_ps(mi1);
 143 }
 144
 145 // CHECK-LABEL: @test_cast
 146
 147 // CHECK-LABEL: define available_externally <4 x float> @_mm_castpd_ps
 148
 149 // CHECK-LABEL: define available_externally <2 x i64> @_mm_castpd_si128
 150
 151 // CHECK-LABEL: define available_externally <2 x double> @_mm_castps_pd
 152
 153 // CHECK-LABEL: define available_externally <2 x i64> @_mm_castps_si128
 154
 155 // CHECK-LABEL: define available_externally <2 x double> @_mm_castsi128_pd
 156
 157 // CHECK-LABEL: define available_externally <4 x float> @_mm_castsi128_ps
 158
 159 void __attribute__((noinline))
 160 test_cmp() {
 161   resi = _mm_cmpeq_epi32(mi1, mi2);
 162   resi = _mm_cmpeq_epi16(mi1, mi2);
 163   resi = _mm_cmpeq_epi8(mi1, mi2);
 164   resi = _mm_cmpgt_epi32(mi1, mi2);
 165   resi = _mm_cmpgt_epi16(mi1, mi2);
 166   resi = _mm_cmpgt_epi8(mi1, mi2);
 167   resi = _mm_cmplt_epi32(mi1, mi2);
 168   resi = _mm_cmplt_epi16(mi1, mi2);
 169   resi = _mm_cmplt_epi8(mi1, mi2);
 170   resd = _mm_cmpeq_pd(md1, md2);
 171   resd = _mm_cmpeq_sd(md1, md2);
 172   resd = _mm_cmpge_pd(md1, md2);
 173   resd = _mm_cmpge_sd(md1, md2);
 174   resd = _mm_cmpgt_pd(md1, md2);
 175   resd = _mm_cmpgt_sd(md1, md2);
 176   resd = _mm_cmple_pd(md1, md2);
 177   resd = _mm_cmple_sd(md1, md2);
 178   resd = _mm_cmplt_pd(md1, md2);
 179   resd = _mm_cmplt_sd(md1, md2);
 180   resd = _mm_cmpneq_pd(md1, md2);
 181   resd = _mm_cmpneq_sd(md1, md2);
 182   resd = _mm_cmpnge_pd(md1, md2);
 183   resd = _mm_cmpnge_sd(md1, md2);
 184   resd = _mm_cmpngt_pd(md1, md2);
 185   resd = _mm_cmpngt_sd(md1, md2);
 186   resd = _mm_cmpnle_pd(md1, md2);
 187   resd = _mm_cmpnle_sd(md1, md2);
 188   resd = _mm_cmpnlt_pd(md1, md2);
 189   resd = _mm_cmpnlt_sd(md1, md2);
 190   resd = _mm_cmpord_pd(md1, md2);
 191   resd = _mm_cmpord_sd(md1, md2);
 192   resd = _mm_cmpunord_pd(md1, md2);
 193   resd = _mm_cmpunord_sd(md1, md2);
 194 }
 195
 196 // CHECK-LABEL: @test_cmp
 197
 198 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmpeq_epi32
 199 // CHECK: call <4 x i32> @vec_cmpeq(int vector[4], int vector[4])
 200
 201 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmpeq_epi16
 202 // CHECK: call <8 x i16> @vec_cmpeq(short vector[8], short vector[8])
 203
 204 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmpeq_epi8
 205 // CHECK: call <16 x i8> @vec_cmpeq(signed char vector[16], signed char vector[16])
 206
 207 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmpgt_epi32
 208 // CHECK: call <4 x i32> @vec_cmpgt(int vector[4], int vector[4])
 209
 210 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmpgt_epi16
 211 // CHECK: call <8 x i16> @vec_cmpgt(short vector[8], short vector[8])
 212
 213 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmpgt_epi8
 214 // CHECK: call <16 x i8> @vec_cmpgt(signed char vector[16], signed char vector[16])
 215
 216 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmplt_epi32
 217 // CHECK: call <4 x i32> @vec_cmplt(int vector[4], int vector[4])
 218
 219 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmplt_epi16
 220 // CHECK: call <8 x i16> @vec_cmplt(short vector[8], short vector[8])
 221
 222 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmplt_epi8
 223 // CHECK: call <16 x i8> @vec_cmplt(signed char vector[16], signed char vector[16])
 224
 225 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpeq_pd
 226 // CHECK: call <2 x i64> @vec_cmpeq(double vector[2], double vector[2])
 227
 228 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpeq_sd
 229 // CHECK: call <2 x double> @vec_splats(double)
 230 // CHECK: call <2 x double> @vec_splats(double)
 231 // CHECK: call <2 x i64> @vec_cmpeq(double vector[2], double vector[2])
 232 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}})
 233
 234 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpge_pd
 235 // CHECK: call <2 x i64> @vec_cmpge(double vector[2], double vector[2])
 236
 237 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpge_sd
 238 // CHECK: call <2 x double> @vec_splats(double)
 239 // CHECK: call <2 x double> @vec_splats(double)
 240 // CHECK: call <2 x i64> @vec_cmpge(double vector[2], double vector[2])
 241 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}})
 242
 243 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpgt_pd
 244 // CHECK: call <2 x i64> @vec_cmpgt(double vector[2], double vector[2])
 245
 246 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpgt_sd
 247 // CHECK: call <2 x double> @vec_splats(double)
 248 // CHECK: call <2 x double> @vec_splats(double)
 249 // CHECK: call <2 x i64> @vec_cmpgt(double vector[2], double vector[2])
 250 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}})
 251
 252 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmple_pd
 253 // CHECK: call <2 x i64> @vec_cmple(double vector[2], double vector[2])
 254
 255 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmple_sd
 256 // CHECK: call <2 x double> @vec_splats(double)
 257 // CHECK: call <2 x double> @vec_splats(double)
 258 // CHECK: call <2 x i64> @vec_cmple(double vector[2], double vector[2])
 259 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}})
 260
 261 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmplt_pd
 262 // CHECK: call <2 x i64> @vec_cmplt(double vector[2], double vector[2])
 263
 264 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmplt_sd
 265 // CHECK: call <2 x double> @vec_splats(double)
 266 // CHECK: call <2 x double> @vec_splats(double)
 267 // CHECK: call <2 x i64> @vec_cmplt(double vector[2], double vector[2])
 268 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}})
 269
 270 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpneq_pd
 271 // CHECK: call <2 x i64> @vec_cmpeq(double vector[2], double vector[2])
 272 // CHECK: call <2 x double> @vec_nor(double vector[2], double vector[2])
 273
 274 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpneq_sd
 275 // CHECK: call <2 x double> @vec_splats(double)
 276 // CHECK: call <2 x double> @vec_splats(double)
 277 // CHECK: call <2 x i64> @vec_cmpeq(double vector[2], double vector[2])
 278 // CHECK: call <2 x double> @vec_nor(double vector[2], double vector[2])
 279 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}})
 280
 281 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpnge_pd
 282 // CHECK: call <2 x i64> @vec_cmplt(double vector[2], double vector[2])
 283
 284 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpnge_sd
 285 // CHECK: call <2 x double> @vec_splats(double)
 286 // CHECK: call <2 x double> @vec_splats(double)
 287 // CHECK: call <2 x i64> @vec_cmplt(double vector[2], double vector[2])
 288 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}})
 289
 290 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpngt_pd
 291 // CHECK: call <2 x i64> @vec_cmple(double vector[2], double vector[2])
 292
 293 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpngt_sd
 294 // CHECK: call <2 x double> @vec_splats(double)
 295 // CHECK: call <2 x double> @vec_splats(double)
 296 // CHECK: call <2 x i64> @vec_cmple(double vector[2], double vector[2])
 297 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}})
 298
 299 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpnle_pd
 300 // CHECK: call <2 x i64> @vec_cmpgt(double vector[2], double vector[2])
 301
 302 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpnle_sd
 303 // CHECK: call <2 x double> @vec_splats(double)
 304 // CHECK: call <2 x double> @vec_splats(double)
 305 // CHECK: call <2 x i64> @vec_cmpge(double vector[2], double vector[2])
 306 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}})
 307
 308 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpnlt_pd
 309 // CHECK: call <2 x i64> @vec_cmpge(double vector[2], double vector[2])
 310
 311 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpnlt_sd
 312 // CHECK: call <2 x double> @vec_splats(double)
 313 // CHECK: call <2 x double> @vec_splats(double)
 314 // CHECK: call <2 x i64> @vec_cmpge(double vector[2], double vector[2])
 315 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}})
 316
 317 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpord_pd
 318 // CHECK: call <2 x i64> @vec_cmpeq(double vector[2], double vector[2])
 319 // CHECK: call <2 x i64> @vec_cmpeq(double vector[2], double vector[2])
 320 // CHECK: call <2 x i64> @vec_and(unsigned long long vector[2], unsigned long long vector[2])
 321
 322 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpord_sd
 323 // CHECK: call <2 x double> @vec_splats(double)
 324 // CHECK: call <2 x double> @vec_splats(double)
 325 // CHECK: call <2 x double> @_mm_cmpord_pd(<2 x double> noundef %{{[0-9a-zA-Z_.]+}}, <2 x double> noundef %{{[0-9a-zA-Z_.]+}})
 326 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}})
 327
 328 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpunord_pd
 329 // CHECK: call <2 x i64> @vec_cmpeq(double vector[2], double vector[2])
 330 // CHECK: call <2 x i64> @vec_cmpeq(double vector[2], double vector[2])
 331 // CHECK: call <2 x i64> @vec_nor(unsigned long long vector[2], unsigned long long vector[2])
 332 // CHECK: call <2 x i64> @vec_orc(unsigned long long vector[2], unsigned long long vector[2])
 333
 334 // CHECK-LABEL: define available_externally <2 x double> @_mm_cmpunord_sd
 335 // CHECK: call <2 x double> @vec_splats(double)
 336 // CHECK: call <2 x double> @vec_splats(double)
 337 // CHECK: call <2 x double> @_mm_cmpunord_pd(<2 x double> noundef %{{[0-9a-zA-Z_.]+}}, <2 x double> noundef %{{[0-9a-zA-Z_.]+}})
 338 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}})
 339
 340 void __attribute__((noinline))
 341 test_comi() {
 342   i = _mm_comieq_sd(md1, md2);
 343   i = _mm_comige_sd(md1, md2);
 344   i = _mm_comigt_sd(md1, md2);
 345   i = _mm_comile_sd(md1, md2);
 346   i = _mm_comilt_sd(md1, md2);
 347   i = _mm_comineq_sd(md1, md2);
 348 }
 349
 350 // CHECK-LABEL: @test_comi
 351
 352 // CHECK-LABEL: define available_externally signext i32 @_mm_comieq_sd
 353 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = fcmp oeq double
 354 // CHECK: zext i1 %[[CMP]] to i32
 355
 356 // CHECK-LABEL: define available_externally signext i32 @_mm_comige_sd
 357 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = fcmp oge double
 358 // CHECK: zext i1 %[[CMP]] to i32
 359
 360 // CHECK-LABEL: define available_externally signext i32 @_mm_comigt_sd
 361 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = fcmp ogt double
 362 // CHECK: zext i1 %[[CMP]] to i32
 363
 364 // CHECK-LABEL: define available_externally signext i32 @_mm_comile_sd
 365 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = fcmp ole double
 366 // CHECK: zext i1 %[[CMP]] to i32
 367
 368 // CHECK-LABEL: define available_externally signext i32 @_mm_comilt_sd
 369 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = fcmp olt double
 370 // CHECK: zext i1 %[[CMP]] to i32
 371
 372 // CHECK-LABEL: define available_externally signext i32 @_mm_comineq_sd
 373 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = fcmp une double
 374 // CHECK: zext i1 %[[CMP]] to i32
 375
 376 void __attribute__((noinline))
 377 test_control() {
 378   _mm_clflush(dp);
 379   _mm_lfence();
 380   _mm_mfence();
 381   _mm_pause();
 382 }
 383
 384 // CHECK-LABEL: @test_control
 385
 386 // CHECK-LABEL: define available_externally void @_mm_clflush
 387 // CHECK: call void asm sideeffect "dcbf 0,$0", "b,~{memory}"(ptr %{{[0-9a-zA-Z_.]+}})
 388
 389 // CHECK-LABEL: define available_externally void @_mm_lfence()
 390 // CHECK: fence release
 391
 392 // CHECK-LABEL: define available_externally void @_mm_mfence()
 393 // CHECK: fence seq_cst
 394
 395 // CHECK-LABEL: define available_externally void @_mm_pause()
 396 // CHECK: call i64 asm sideeffect "\09mfppr\09$0;   or 31,31,31;   isync;   lwsync;   isync;   mtppr\09$0;", "=r,~{memory}"()
 397
 398 void __attribute__((noinline))
 399 test_converts() {
 400   resd = _mm_cvtepi32_pd(mi1);
 401   res = _mm_cvtepi32_ps(mi1);
 402   resi = _mm_cvtpd_epi32(md1);
 403   res64 = _mm_cvtpd_pi32(md1);
 404   res = _mm_cvtpd_ps(md1);
 405   resd = _mm_cvtpi32_pd(res64);
 406   resi = _mm_cvtps_epi32(m1);
 407   resd = _mm_cvtps_pd(m1);
 408   *dp = _mm_cvtsd_f64(md1);
 409   i = _mm_cvtsd_si32(md1);
 410   i64s[0] = _mm_cvtsd_si64(md1);
 411   i64s[0] = _mm_cvtsd_si64x(md1);
 412   res = _mm_cvtsd_ss(m1, md2);
 413   i = _mm_cvtsi128_si32(mi1);
 414   i64s[0] = _mm_cvtsi128_si64(mi1);
 415   i64s[0] = _mm_cvtsi128_si64x(mi1);
 416   resd = _mm_cvtsi32_sd(md1, i);
 417   resi = _mm_cvtsi32_si128(i);
 418   resd = _mm_cvtsi64_sd(md1, i64s[1]);
 419   resi = _mm_cvtsi64_si128(i64s[1]);
 420   resd = _mm_cvtsi64x_sd(md1, i64s[1]);
 421   resi = _mm_cvtsi64x_si128(i64s[1]);
 422   resd = _mm_cvtss_sd(md1, m1);
 423   resi = _mm_cvttpd_epi32(md1);
 424   res64 = _mm_cvttpd_pi32(md1);
 425   resi = _mm_cvttps_epi32(m1);
 426   i = _mm_cvttsd_si32(md1);
 427   i64s[0] = _mm_cvttsd_si64(md1);
 428   i64s[0] = _mm_cvttsd_si64x(md1);
 429 }
 430
 431 // CHECK-LABEL: @test_converts
 432
 433 // CHECK-LABEL: define available_externally <2 x double> @_mm_cvtepi32_pd
 434 // CHECK: call <2 x i64> @vec_unpackh(int vector[4])
 435 // CHECK: %[[CONV:[0-9a-zA-Z_.]+]] = sitofp <2 x i64> %{{[0-9a-zA-Z_.]+}} to <2 x double>
 436 // CHECK: fmul <2 x double> %[[CONV]], <double 1.000000e+00, double 1.000000e+00>
 437
 438 // CHECK-LABEL: define available_externally <4 x float> @_mm_cvtepi32_ps
 439 // CHECK: call <4 x float> @llvm.ppc.altivec.vcfsx(<4 x i32> %{{[0-9a-zA-Z_.]+}}, i32 0)
 440
 441 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtpd_epi32
 442 // CHECK: call <2 x double> @vec_rint(double vector[2])
 443 // CHECK: store <4 x i32> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 444 // CHECK: call <4 x i32> asm "xvcvdpsxws ${0:x},${1:x}", "=^wa,^wa"(<2 x double> %{{[0-9a-zA-Z_.]+}})
 445 // CHECK-LE: call <4 x i32> @vec_mergeo(int vector[4], int vector[4])
 446 // CHECK-BE: call <4 x i32> @vec_mergee(int vector[4], int vector[4])
 447 // 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)
 448
 449 // CHECK-LABEL: define available_externally i64 @_mm_cvtpd_pi32
 450 // CHECK: call <2 x i64> @_mm_cvtpd_epi32(<2 x double> noundef %{{[0-9a-zA-Z_.]+}})
 451 // CHECK: extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 0
 452
 453 // CHECK-LABEL: define available_externally <4 x float> @_mm_cvtpd_ps
 454 // CHECK: store <4 x i32> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 455 // CHECK: call <4 x i32> asm "xvcvdpsp ${0:x},${1:x}", "=^wa,^wa"(<2 x double> %{{[0-9a-zA-Z_.]+}})
 456 // CHECK-LE: call <4 x i32> @vec_mergeo(int vector[4], int vector[4])
 457 // CHECK-BE: call <4 x i32> @vec_mergee(int vector[4], int vector[4])
 458 // 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)
 459
 460 // CHECK-LABEL: define available_externally <2 x double> @_mm_cvtpi32_pd
 461 // CHECK: call <2 x i64> @vec_splats(unsigned long long)
 462 // CHECK: call <2 x i64> @vec_unpackl(int vector[4])
 463 // CHECK: %[[CONV:[0-9a-zA-Z_.]+]] = sitofp <2 x i64> %{{[0-9a-zA-Z._]+}} to <2 x double>
 464 // CHECK: fmul <2 x double> %[[CONV]], <double 1.000000e+00, double 1.000000e+00>
 465
 466 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtps_epi32
 467 // CHECK: call <4 x float> @vec_rint(float vector[4])
 468 // CHECK: call <4 x i32> @llvm.ppc.altivec.vctsxs(<4 x float> %{{[0-9a-zA-Z_.]+}}, i32 0)
 469
 470 // CHECK-LABEL: define available_externally <2 x double> @_mm_cvtps_pd
 471 // CHECK-BE: call <4 x float> @vec_vmrghw(float vector[4], float vector[4])
 472 // CHECK-BE: call <2 x double> asm " xvcvspdp ${0:x},${1:x}", "=^wa,^wa"(<4 x float> %{{[0-9a-zA-Z_.]+}})
 473 // 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>
 474 // 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>
 475 // CHECK-LE: call <2 x double> asm " xvcvspdp ${0:x},${1:x}", "=^wa,^wa"(<4 x float> %{{[0-9a-zA-Z_.]+}})
 476
 477 // CHECK-LABEL: define available_externally double @_mm_cvtsd_f64
 478 // CHECK: extractelement <2 x double> %{{[0-9a-zA-Z_.]+}}, i32 0
 479
 480 // CHECK-LABEL: define available_externally signext i32 @_mm_cvtsd_si32
 481 // CHECK: call <2 x double> @vec_rint(double vector[2])
 482 // CHECK: fptosi double %{{[0-9a-zA-Z_.]+}} to i32
 483
 484 // CHECK-LABEL: define available_externally i64 @_mm_cvtsd_si64
 485 // CHECK: call <2 x double> @vec_rint(double vector[2])
 486 // CHECK: fptosi double %{{[0-9a-zA-Z_.]+}} to i64
 487
 488 // CHECK-LABEL: define available_externally i64 @_mm_cvtsd_si64x
 489 // CHECK: call i64 @_mm_cvtsd_si64(<2 x double> noundef %{{[0-9a-zA-Z_.]+}})
 490
 491 // CHECK-LABEL: define available_externally <4 x float> @_mm_cvtsd_ss
 492 // CHECK-BE: %[[EXT:[0-9a-zA-Z_.]+]] = extractelement <2 x double> %{{[0-9a-zA-Z_.]+}}, i32 0
 493 // CHECK-BE: %[[TRUNC:[0-9a-zA-Z_.]+]] = fptrunc double %[[EXT]] to float
 494 // CHECK-BE: insertelement <4 x float> %{{[0-9a-zA-Z_.]+}}, float %[[TRUNC]], i32 0
 495 // 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)
 496 // 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>
 497 // CHECK-LE: call <4 x float> asm "xscvdpsp ${0:x},${1:x}", "=^wa,^wa"(<2 x double> %{{[0-9a-zA-Z_.]+}})
 498 // 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>
 499
 500 // CHECK-LABEL: define available_externally signext i32 @_mm_cvtsi128_si32
 501 // CHECK: extractelement <4 x i32> %{{[0-9a-zA-Z_.]+}}, i32 0
 502
 503 // CHECK-LABEL: define available_externally i64 @_mm_cvtsi128_si64
 504 // CHECK: extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 0
 505
 506 // CHECK-LABEL: define available_externally i64 @_mm_cvtsi128_si64x
 507 // CHECK: extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 0
 508
 509 // CHECK-LABEL: define available_externally <2 x double> @_mm_cvtsi32_sd
 510 // CHECK: sitofp i32 %{{[0-9a-zA-Z_.]+}} to double
 511
 512 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtsi32_si128
 513 // 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_.]+}})
 514
 515 // CHECK-LABEL: define available_externally <2 x double> @_mm_cvtsi64_sd
 516 // CHECK: sitofp i64 %{{[0-9a-zA-Z_.]+}} to double
 517
 518 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtsi64_si128
 519 // CHECK: %[[INS:[0-9a-zA-Z_.]+]] = insertelement <2 x i64> undef, i64 %{{[0-9a-zA-Z_.]+}}, i32 0
 520 // CHECK: insertelement <2 x i64> %[[INS]], i64 0, i32 1
 521
 522 // CHECK-LABEL: define available_externally <2 x double> @_mm_cvtsi64x_sd
 523 // CHECK: call <2 x double> @_mm_cvtsi64_sd(<2 x double> noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef %{{[0-9a-zA-Z_.]+}})
 524
 525 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtsi64x_si128
 526 // CHECK: %[[INS:[0-9a-zA-Z_.]+]] = insertelement <2 x i64> undef, i64 %{{[0-9a-zA-Z_.]+}}, i32 0
 527 // CHECK: insertelement <2 x i64> %[[INS]], i64 0, i32 1
 528
 529 // CHECK-LABEL: define available_externally <2 x double> @_mm_cvtss_sd
 530 // CHECK-BE: fpext float %{{[0-9a-zA-Z_.]+}} to double
 531 // 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)
 532 // CHECK-LE: call <2 x double> asm "xscvspdp ${0:x},${1:x}", "=^wa,^wa"(<4 x float> %{{[0-9a-zA-Z_.]+}})
 533 // CHECK-LE: call <2 x double> @vec_mergel(double vector[2], double vector[2])
 534
 535 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvttpd_epi32
 536 // CHECK: call <4 x i32> asm "xvcvdpsxws ${0:x},${1:x}", "=^wa,^wa"
 537 // CHECK-LE: call <4 x i32> @vec_mergeo(int vector[4], int vector[4])
 538 // CHECK-BE: call <4 x i32> @vec_mergee(int vector[4], int vector[4])
 539 // 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)
 540
 541 // CHECK-LABEL: define available_externally i64 @_mm_cvttpd_pi32
 542 // CHECK: call <2 x i64> @_mm_cvttpd_epi32(<2 x double> noundef %{{[0-9a-zA-Z_.]+}})
 543
 544 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvttps_epi32
 545 // CHECK: call <4 x i32> @llvm.ppc.altivec.vctsxs(<4 x float> %{{[0-9a-zA-Z_.]+}}, i32 0)
 546
 547 // CHECK-LABEL: define available_externally signext i32 @_mm_cvttsd_si32
 548 // CHECK: fptosi double %{{[0-9a-zA-Z_.]+}} to i32
 549
 550 // CHECK-LABEL: define available_externally i64 @_mm_cvttsd_si64
 551 // CHECK: fptosi double %{{[0-9a-zA-Z_.]+}} to i64
 552
 553 // CHECK-LABEL: define available_externally i64 @_mm_cvttsd_si64x
 554 // CHECK: call i64 @_mm_cvttsd_si64(<2 x double> noundef %{{[0-9a-zA-Z_.]+}})
 555
 556 void __attribute__((noinline))
 557 test_div() {
 558   resd = _mm_div_pd(md1, md2);
 559   resd = _mm_div_sd(md1, md2);
 560 }
 561
 562 // CHECK-LABEL: @test_div
 563
 564 // CHECK-LABEL: define available_externally <2 x double> @_mm_div_pd
 565 // CHECK: fdiv <2 x double>
 566
 567 // CHECK-LABEL: define available_externally <2 x double> @_mm_div_sd
 568 // CHECK: fdiv double
 569
 570 void __attribute__((noinline))
 571 test_extract() {
 572   i = _mm_extract_epi16(mi1, i);
 573 }
 574
 575 // CHECK-LABEL: @test_extract
 576
 577 // CHECK-LABEL: define available_externally signext i32 @_mm_extract_epi16
 578 // CHECK: %[[AND:[0-9a-zA-Z_.]+]] = and i32 %{{[0-9a-zA-Z_.]+}}, 7
 579 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = extractelement <8 x i16> %{{[0-9a-zA-Z_.]+}}, i32 %[[AND]]
 580 // CHECK: zext i16 %[[EXT]] to i32
 581
 582 void __attribute__((noinline))
 583 test_insert() {
 584   resi = _mm_insert_epi16 (mi1, i, is[0]);
 585 }
 586
 587 // CHECK-LABEL: @test_insert
 588
 589 // CHECK-LABEL: define available_externally <2 x i64> @_mm_insert_epi16
 590 // CHECK: trunc i32 %{{[0-9a-zA-Z_.]+}} to i16
 591 // CHECK: and i32 %{{[0-9a-zA-Z_.]+}}, 7
 592
 593 void __attribute__((noinline))
 594 test_load() {
 595   resd = _mm_load_pd(dp);
 596   resd = _mm_load_pd1(dp);
 597   resd = _mm_load_sd(dp);
 598   resi = _mm_load_si128(mip);
 599   resd = _mm_load1_pd(dp);
 600   resd = _mm_loadh_pd(md1, dp);
 601   resi = _mm_loadl_epi64(mip);
 602   resd = _mm_loadl_pd(md1, dp);
 603   resd = _mm_loadr_pd(dp);
 604   resd = _mm_loadu_pd(dp);
 605   resi = _mm_loadu_si128(mip);
 606 }
 607
 608 // CHECK-LABEL: @test_load
 609
 610 // CHECK-LABEL: define available_externally <2 x double> @_mm_load_pd
 611 // CHECK: call <16 x i8> @vec_ld(long, unsigned char vector[16] const*)(i64 noundef 0, ptr noundef %{{[0-9a-zA-Z_.]+}})
 612
 613 // CHECK-LABEL: define available_externally <2 x double> @_mm_load_pd1
 614 // CHECK: call <2 x double> @_mm_load1_pd(ptr noundef %{{[0-9a-zA-Z_.]+}})
 615
 616 // CHECK-LABEL: define available_externally <2 x double> @_mm_load_sd
 617 // CHECK: call <2 x double> @_mm_set_sd(double noundef %{{[0-9a-zA-Z_.]+}})
 618
 619 // CHECK-LABEL: define available_externally <2 x i64> @_mm_load_si128
 620 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load ptr, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 621 // CHECK: load <2 x i64>, ptr %[[ADDR]], align 16
 622
 623 // CHECK-LABEL: define available_externally <2 x double> @_mm_load1_pd
 624 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load ptr, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 625 // CHECK: %[[VAL:[0-9a-zA-Z_.]+]] = load double, ptr %[[ADDR]]
 626 // CHECK: call <2 x double> @vec_splats(double)(double noundef %[[VAL]])
 627
 628 // CHECK-LABEL: define available_externally <2 x double> @_mm_loadh_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 %{{[0-9a-zA-Z_.]+}}
 631 // CHECK: %[[VEC:[0-9a-zA-Z_.]+]] = load <2 x double>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 632 // CHECK: insertelement <2 x double> %[[VEC]], double %[[VAL]], i32 1
 633
 634 // CHECK-LABEL: define available_externally <2 x i64> @_mm_loadl_epi64
 635 // CHECK: call <2 x i64> @_mm_set_epi64(i64 noundef 0, i64 noundef %{{[0-9a-zA-Z_.]+}})
 636
 637 // CHECK-LABEL: define available_externally <2 x double> @_mm_loadl_pd
 638 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load ptr, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 639 // CHECK: %[[ADDR2:[0-9a-zA-Z_.]+]] = load double, ptr %[[ADDR]]
 640 // CHECK: %[[VEC:[0-9a-zA-Z_.]+]] = load <2 x double>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 641 // CHECK: insertelement <2 x double> %[[VEC]], double %[[ADDR2]], i32 0
 642
 643 // CHECK-LABEL: define available_externally <2 x double> @_mm_loadr_pd
 644 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load ptr, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 645 // CHECK: call <2 x double> @_mm_load_pd(ptr noundef %[[ADDR]])
 646 // CHECK: shufflevector <2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i32> <i32 1, i32 2>
 647
 648 // CHECK-LABEL: define available_externally <2 x double> @_mm_loadu_pd
 649 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load ptr, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 650 // CHECK: call <2 x double> @vec_vsx_ld(int, double const*)(i32 noundef signext 0, ptr noundef %[[ADDR]])
 651
 652 // CHECK-LABEL: define available_externally <2 x i64> @_mm_loadu_si128
 653 // CHECK: load ptr, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 654 // CHECK: call <4 x i32> @vec_vsx_ld(int, int const*)(i32 noundef signext 0, ptr noundef %{{[0-9a-zA-Z_.]+}})
 655
 656 void __attribute__((noinline))
 657 test_logical() {
 658   resd = _mm_and_pd(md1, md2);
 659   resi = _mm_and_si128(mi1, mi2);
 660   resd = _mm_andnot_pd(md1, md2);
 661   resi = _mm_andnot_si128(mi1, mi2);
 662   resd = _mm_xor_pd(md1, md2);
 663   resi = _mm_xor_si128(mi1, mi2);
 664   resd = _mm_or_pd(md1, md2);
 665   resi = _mm_or_si128(mi1, mi2);
 666 }
 667
 668 // CHECK-LABEL: @test_logical
 669
 670 // CHECK-LABEL: define available_externally <2 x double> @_mm_and_pd
 671 // CHECK: call <2 x double> @vec_and(double vector[2], double vector[2])
 672
 673 // CHECK-LABEL: define available_externally <2 x i64> @_mm_and_si128
 674 // CHECK: call <2 x i64> @vec_and(long long vector[2], long long vector[2])
 675
 676 // CHECK-LABEL: define available_externally <2 x double> @_mm_andnot_pd
 677 // CHECK: call <2 x double> @vec_andc(double vector[2], double vector[2])
 678
 679 // CHECK-LABEL: define available_externally <2 x i64> @_mm_andnot_si128
 680 // CHECK: call <2 x i64> @vec_andc(long long vector[2], long long vector[2])
 681
 682 // CHECK-LABEL: define available_externally <2 x double> @_mm_xor_pd
 683 // CHECK: call <2 x double> @vec_xor(double vector[2], double vector[2])
 684
 685 // CHECK-LABEL: define available_externally <2 x i64> @_mm_xor_si128
 686 // CHECK: call <2 x i64> @vec_xor(long long vector[2], long long vector[2])
 687
 688 // CHECK-LABEL: define available_externally <2 x double> @_mm_or_pd
 689 // CHECK: call <2 x double> @vec_or(double vector[2], double vector[2])
 690
 691 // CHECK-LABEL: define available_externally <2 x i64> @_mm_or_si128
 692 // CHECK: call <2 x i64> @vec_or(long long vector[2], long long vector[2])
 693
 694 void __attribute__((noinline))
 695 test_max() {
 696   resi = _mm_max_epi16(mi1, mi2);
 697   resi = _mm_max_epu8(mi1, mi2);
 698   resd = _mm_max_pd(md1, md2);
 699   resd = _mm_max_sd(md1, md2);
 700 }
 701
 702 // CHECK-LABEL: @test_max
 703
 704 // CHECK-LABEL: define available_externally <2 x i64> @_mm_max_epi16
 705 // CHECK: call <8 x i16> @vec_max(short vector[8], short vector[8])
 706
 707 // CHECK-LABEL: define available_externally <2 x i64> @_mm_max_epu8
 708 // CHECK: call <16 x i8> @vec_max(unsigned char vector[16], unsigned char vector[16])
 709
 710 // CHECK-LABEL: define available_externally <2 x double> @_mm_max_pd
 711 // CHECK: call <2 x double> @vec_max(double vector[2], double vector[2])
 712
 713 // CHECK-LABEL: define available_externally <2 x double> @_mm_max_sd
 714 // CHECK: call <2 x double> @vec_splats(double)
 715 // CHECK: call <2 x double> @vec_splats(double)
 716 // CHECK: call <2 x double> @vec_max(double vector[2], double vector[2])
 717 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}})
 718
 719 void __attribute__((noinline))
 720 test_min() {
 721   resi = _mm_min_epi16(mi1, mi2);
 722   resi = _mm_min_epu8(mi1, mi2);
 723   resd = _mm_min_pd(md1, md2);
 724   resd = _mm_min_sd(md1, md2);
 725 }
 726
 727 // CHECK-LABEL: @test_min
 728
 729 // CHECK-LABEL: define available_externally <2 x i64> @_mm_min_epi16
 730 // CHECK: call <8 x i16> @vec_min(short vector[8], short vector[8])
 731
 732 // CHECK-LABEL: define available_externally <2 x i64> @_mm_min_epu8
 733 // CHECK: call <16 x i8> @vec_min(unsigned char vector[16], unsigned char vector[16])
 734
 735 // CHECK-LABEL: define available_externally <2 x double> @_mm_min_pd
 736 // CHECK: call <2 x double> @vec_min(double vector[2], double vector[2])
 737
 738 // CHECK-LABEL: define available_externally <2 x double> @_mm_min_sd
 739 // CHECK: call <2 x double> @vec_splats(double)
 740 // CHECK: call <2 x double> @vec_splats(double)
 741 // CHECK: call <2 x double> @vec_min(double vector[2], double vector[2])
 742 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}})
 743
 744 void __attribute__((noinline))
 745 test_move() {
 746   resi = _mm_move_epi64(mi1);
 747   resd = _mm_move_sd(md1, md2);
 748   i = _mm_movemask_epi8(mi1);
 749   i = _mm_movemask_pd(md1);
 750   res64 = _mm_movepi64_pi64(mi1);
 751   resi = _mm_movpi64_epi64(m641);
 752   _mm_maskmoveu_si128(mi1, mi2, chs);
 753 }
 754
 755 // CHECK-LABEL: @test_move
 756
 757 // CHECK-LABEL: define available_externally <2 x i64> @_mm_move_epi64
 758 // CHECK: call <2 x i64> @_mm_set_epi64(i64 noundef 0, i64 noundef %{{[0-9a-zA-Z_.]+}})
 759
 760 // CHECK-LABEL: define available_externally <2 x double> @_mm_move_sd
 761 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = extractelement <2 x double> %{{[0-9a-zA-Z_.]+}}, i32 0
 762 // CHECK: insertelement <2 x double> %{{[0-9a-zA-Z_.]+}}, double %[[EXT]], i32 0
 763
 764 // CHECK-P10-LABEL: define available_externally signext i32 @_mm_movemask_epi8
 765 // CHECK-P10: call zeroext i32 @vec_extractm(unsigned char vector[16])(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}})
 766
 767 // CHECK-LABEL: define available_externally signext i32 @_mm_movemask_epi8
 768 // 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>)
 769 // CHECK-LE: %[[VAL:[0-9a-zA-Z_.]+]] = extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 1
 770 // CHECK-BE: %[[VAL:[0-9a-zA-Z_.]+]] = extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 0
 771 // CHECK: trunc i64 %[[VAL]] to i32
 772
 773 // CHECK-P10-LABEL: define available_externally signext i32 @_mm_movemask_pd
 774 // CHECK-P10: call zeroext i32 @vec_extractm(unsigned long long vector[2])(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 775
 776 // CHECK-LABEL: define available_externally signext i32 @_mm_movemask_pd
 777 // 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>))
 778 // CHECK-LE: extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 1
 779 // 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>))
 780 // CHECK-BE: extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 0
 781
 782 // CHECK-LABEL: define available_externally i64 @_mm_movepi64_pi64
 783 // CHECK: extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 0
 784
 785 // CHECK-LABEL: define available_externally <2 x i64> @_mm_movpi64_epi64
 786 // CHECK: call <2 x i64> @_mm_set_epi64(i64 noundef 0, i64 noundef %{{[0-9a-zA-Z_.]+}})
 787
 788 // CHECK-LABEL: define available_externally void @_mm_maskmoveu_si128
 789 // CHECK: call <2 x i64> @_mm_loadu_si128(ptr noundef %{{[0-9a-zA-Z_.]+}})
 790 // CHECK: call <16 x i8> @vec_cmpgt(unsigned char vector[16], unsigned char vector[16])
 791 // CHECK: call <16 x i8> @vec_sel(unsigned char vector[16], unsigned char vector[16], unsigned char vector[16])
 792 // CHECK: call void @_mm_storeu_si128(ptr noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 793
 794 void __attribute__((noinline))
 795 test_mul() {
 796   resi = _mm_mul_epu32(mi1, mi2);
 797   resd = _mm_mul_pd(md1, md2);
 798   resd = _mm_mul_sd(md1, md2);
 799   res64 = _mm_mul_su32(m641, m642);
 800   resi = _mm_mulhi_epi16(mi1, mi2);
 801   resi = _mm_mulhi_epu16(mi1, mi2);
 802   resi = _mm_mullo_epi16(mi1, mi2);
 803 }
 804
 805 // CHECK-LABEL: @test_mul
 806
 807 // CHECK-LABEL: define available_externally <2 x i64> @_mm_mul_epu32
 808 // 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_.]+}})
 809 // 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_.]+}})
 810
 811 // CHECK-LABEL: define available_externally <2 x double> @_mm_mul_pd
 812 // CHECK: fmul <2 x double>
 813
 814 // CHECK-LABEL: define available_externally <2 x double> @_mm_mul_sd
 815 // CHECK: fmul double
 816
 817 // CHECK-LABEL: define available_externally i64 @_mm_mul_su32
 818 // CHECK: trunc i64 %{{[0-9a-zA-Z_.]+}} to i32
 819 // CHECK: trunc i64 %{{[0-9a-zA-Z_.]+}} to i32
 820 // CHECK: %[[EXT1:[0-9a-zA-Z_.]+]] = zext i32 %{{[0-9a-zA-Z_.]+}} to i64
 821 // CHECK: %[[EXT2:[0-9a-zA-Z_.]+]] = zext i32 %{{[0-9a-zA-Z_.]+}} to i64
 822 // CHECK: mul i64 %[[EXT1]], %[[EXT2]]
 823
 824 // CHECK-LABEL: define available_externally <2 x i64> @_mm_mulhi_epi16
 825 // 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
 826 // 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
 827 // CHECK: call <4 x i32> @vec_vmulesh(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}})
 828 // CHECK: call <4 x i32> @vec_vmulosh(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}})
 829 // CHECK: call <4 x i32> @vec_perm(int vector[4], int vector[4], unsigned char vector[16])
 830
 831 // CHECK-LABEL: define available_externally <2 x i64> @_mm_mulhi_epu16
 832 // 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
 833 // 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
 834 // CHECK: call <4 x i32> @vec_vmuleuh(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}})
 835 // CHECK: call <4 x i32> @vec_vmulouh(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}})
 836 // CHECK: call <4 x i32> @vec_perm(unsigned int vector[4], unsigned int vector[4], unsigned char vector[16])
 837
 838 // CHECK-LABEL: define available_externally <2 x i64> @_mm_mullo_epi16
 839 // CHECK: mul <8 x i16>
 840
 841 void __attribute__((noinline))
 842 test_pack() {
 843   resi = _mm_packs_epi16(mi1, mi2);
 844   resi = _mm_packs_epi32(mi1, mi2);
 845   resi = _mm_packus_epi16(mi1, mi2);
 846 }
 847
 848 // CHECK-LABEL: @test_pack
 849
 850 // CHECK-LABEL: define available_externally <2 x i64> @_mm_packs_epi16
 851 // CHECK: call <16 x i8> @vec_packs(short vector[8], short vector[8])
 852
 853 // CHECK-LABEL: define available_externally <2 x i64> @_mm_packs_epi32
 854 // CHECK: call <8 x i16> @vec_packs(int vector[4], int vector[4])
 855
 856 // CHECK-LABEL: define available_externally <2 x i64> @_mm_packus_epi16
 857 // CHECK: call <16 x i8> @vec_packsu(short vector[8], short vector[8])
 858
 859 void __attribute__((noinline))
 860 test_sad() {
 861   resi = _mm_sad_epu8(mi1, mi2);
 862 }
 863
 864 // CHECK-LABEL: @test_sad
 865
 866 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sad_epu8
 867 // CHECK: call <16 x i8> @vec_min(unsigned char vector[16], unsigned char vector[16])
 868 // CHECK: call <16 x i8> @vec_max(unsigned char vector[16], unsigned char vector[16])
 869 // CHECK: call <16 x i8> @vec_sub(unsigned char vector[16], unsigned char vector[16])
 870 // 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)
 871 // CHECK-LE: call <4 x i32> asm "vsum2sws $0,$1,$2", "=v,v,v"(<4 x i32> %11, <4 x i32> zeroinitializer)
 872 // CHECK-BE: call <4 x i32> @vec_sum2s(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <4 x i32> noundef zeroinitializer)
 873 // CHECK-BE: call <4 x i32> @vec_sld(int vector[4], int vector[4], unsigned int)
 874
 875 void __attribute__((noinline))
 876 test_set() {
 877   resi = _mm_set_epi16(ss[7], ss[6], ss[5], ss[4], ss[3], ss[2], ss[1], ss[0]);
 878   resi = _mm_set_epi32(is[3], is[2], is[1], is[0]);
 879   resi = _mm_set_epi64(m641, m642);
 880   resi = _mm_set_epi64x(i64s[0], i64s[1]);
 881   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]);
 882   resd = _mm_set_pd(dp[0], dp[1]);
 883   resd = _mm_set_pd1(dp[0]);
 884   resd = _mm_set_sd(dp[0]);
 885   resi = _mm_set1_epi16(ss[0]);
 886   resi = _mm_set1_epi32(i);
 887   resi = _mm_set1_epi64(m641);
 888   resi = _mm_set1_epi64x(i64s[0]);
 889   resi = _mm_set1_epi8(chs[0]);
 890   resd = _mm_set1_pd(dp[0]);
 891   resi = _mm_setr_epi16(ss[7], ss[6], ss[5], ss[4], ss[3], ss[2], ss[1], ss[0]);
 892   resi = _mm_setr_epi32(is[3], is[2], is[1], is[0]);
 893   resi = _mm_setr_epi64(m641, m642);
 894   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]);
 895   resd = _mm_setr_pd(dp[0], dp[1]);
 896   resd = _mm_setzero_pd();
 897   resi = _mm_setzero_si128();
 898 }
 899
 900 // CHECK-LABEL: @test_set
 901
 902 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set_epi16
 903 // CHECK-COUNT-8: store i16 {{[0-9a-zA-Z_%.]+}}, ptr {{[0-9a-zA-Z_%.]+}}, align 2
 904 // CHECK: insertelement <8 x i16> undef, i16 {{[0-9a-zA-Z_%.]+}}, i32 0
 905 // CHECK-COUNT-7: insertelement <8 x i16> {{[0-9a-zA-Z_%.]+}}, i16 {{[0-9a-zA-Z_%.]+}}, i32 {{[1-7]}}
 906
 907 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set_epi32
 908 // CHECK-COUNT-4: store i32 {{[0-9a-zA-Z_%.]+}}, ptr {{[0-9a-zA-Z_%.]+}}, align 4
 909 // CHECK: insertelement <4 x i32> undef, i32 {{[0-9a-zA-Z_%.]+}}, i32 0
 910 // CHECK-COUNT-3: insertelement <4 x i32> {{[0-9a-zA-Z_%.]+}}, i32 {{[0-9a-zA-Z_%.]+}}, i32 {{[1-3]}}
 911
 912 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set_epi64
 913 // CHECK: call <2 x i64> @_mm_set_epi64x(i64 noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef %{{[0-9a-zA-Z_.]+}})
 914
 915 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set_epi64x
 916 // CHECK: %[[VEC:[0-9a-zA-Z_.]+]] = insertelement <2 x i64> undef, i64 %{{[0-9a-zA-Z_.]+}}, i32 0
 917 // CHECK: insertelement <2 x i64> %[[VEC]], i64 %{{[0-9a-zA-Z_.]+}}, i32 1
 918
 919 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set_epi8
 920 // CHECK-COUNT-16: store i8 {{[0-9a-zA-Z_%.]+}}, ptr {{[0-9a-zA-Z_%.]+}}, align 1
 921 // CHECK: insertelement <16 x i8> undef, i8 {{[0-9a-zA-Z_%.]+}}, i32 {{[0-9]+}}
 922 // CHECK-COUNT-15: {{[0-9a-zA-Z_%.]+}} = insertelement <16 x i8> {{[0-9a-zA-Z_%.]+}}, i8 {{[0-9a-zA-Z_%.]+}}, i32 {{[0-9]+}}
 923
 924 // CHECK-LABEL: define available_externally <2 x double> @_mm_set_pd
 925 // CHECK: %[[VEC:[0-9a-zA-Z_.]+]] = insertelement <2 x double> undef, double %{{[0-9a-zA-Z_.]+}}, i32 0
 926 // CHECK: insertelement <2 x double> %[[VEC]], double %{{[0-9a-zA-Z_.]+}}, i32 1
 927
 928 // CHECK-LABEL: define available_externally <2 x double> @_mm_set_pd1
 929 // CHECK: call <2 x double> @_mm_set1_pd(double noundef %{{[0-9a-zA-Z_.]+}})
 930
 931 // CHECK-LABEL: define available_externally <2 x double> @_mm_set_sd
 932 // CHECK: %[[VEC:[0-9a-zA-Z_.]+]] = insertelement <2 x double> undef, double %{{[0-9a-zA-Z_.]+}}, i32 0
 933 // CHECK: insertelement <2 x double> %[[VEC]], double 0.000000e+00, i32 1
 934
 935 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set1_epi16
 936 // CHECK-COUNT-8: load i16, ptr %{{[0-9a-zA-Z_.]+}}, align 2
 937 // CHECK: call <2 x i64> @_mm_set_epi16
 938
 939 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set1_epi32
 940 // CHECK-COUNT-4: load i32, ptr %{{[0-9a-zA-Z_.]+}}, align 4
 941 // CHECK: call <2 x i64> @_mm_set_epi32
 942
 943 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set1_epi64
 944 // CHECK: %[[VAL1:[0-9a-zA-Z_.]+]] = load i64, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 945 // CHECK: %[[VAL2:[0-9a-zA-Z_.]+]] = load i64, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 946 // CHECK: call <2 x i64> @_mm_set_epi64(i64 noundef %[[VAL1]], i64 noundef %[[VAL2]])
 947
 948 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set1_epi64x
 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_epi64x(i64 noundef %[[VAL1]], i64 noundef %[[VAL2]])
 952
 953 // CHECK-LABEL: define available_externally <2 x i64> @_mm_set1_epi8
 954 // CHECK-COUNT-16: load i8, ptr %{{[0-9a-zA-Z_.]+}}, align 1
 955 // CHECK: call <2 x i64> @_mm_set_epi8
 956
 957 // CHECK-LABEL: define available_externally <2 x double> @_mm_set1_pd
 958 // CHECK: %[[VEC:[0-9a-zA-Z_.]+]] = insertelement <2 x double> undef, double %{{[0-9a-zA-Z_.]+}}, i32 0
 959 // CHECK: insertelement <2 x double> %[[VEC]], double %{{[0-9a-zA-Z_.]+}}, i32 1
 960
 961 // CHECK-LABEL: define available_externally <2 x i64> @_mm_setr_epi16
 962 // CHECK-COUNT-8: load i16, ptr {{[0-9a-zA-Z_%.]+}}, align 2
 963 // CHECK: call <2 x i64> @_mm_set_epi16
 964
 965 // CHECK-LABEL: define available_externally <2 x i64> @_mm_setr_epi32
 966 // CHECK-COUNT-4: load i32, ptr {{[0-9a-zA-Z_%.]+}}, align 4
 967 // CHECK: call <2 x i64> @_mm_set_epi32
 968
 969 // CHECK-LABEL: define available_externally <2 x i64> @_mm_setr_epi64
 970 // CHECK: %[[VAL1:[0-9a-zA-Z_.]+]] = load i64, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 971 // CHECK: %[[VAL2:[0-9a-zA-Z_.]+]] = load i64, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 972 // CHECK: call <2 x i64> @_mm_set_epi64(i64 noundef %[[VAL1]], i64 noundef %[[VAL2]])
 973
 974 // CHECK-LABEL: define available_externally <2 x i64> @_mm_setr_epi8
 975 // CHECK-COUNT-16: load i8, ptr {{[0-9a-zA-Z_%.]+}}, align 1
 976 // CHECK: call <2 x i64> @_mm_set_epi8
 977
 978 // CHECK-LABEL: define available_externally <2 x double> @_mm_setr_pd
 979 // CHECK: %[[VEC:[0-9a-zA-Z_.]+]] = insertelement <2 x double> undef, double %{{[0-9a-zA-Z_.]+}}, i32 0
 980 // CHECK: insertelement <2 x double> %[[VEC]], double %{{[0-9a-zA-Z_.]+}}, i32 1
 981
 982 // CHECK-LABEL: define available_externally <2 x double> @_mm_setzero_pd()
 983 // CHECK: call <4 x i32> @vec_splats(int)(i32 noundef signext 0)
 984
 985 // CHECK-LABEL: define available_externally <2 x i64> @_mm_setzero_si128()
 986 // CHECK: store <4 x i32> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 987
 988 void __attribute__((noinline))
 989 test_shuffle() {
 990   resi = _mm_shuffle_epi32(mi1, i);
 991   resd = _mm_shuffle_pd(md1, md2, i);
 992   resi = _mm_shufflehi_epi16(mi1, i);
 993   resi = _mm_shufflelo_epi16(mi1, i);
 994 }
 995
 996 // CHECK-LABEL: @test_shuffle
 997
 998 // CHECK-LABEL: define available_externally <2 x i64> @_mm_shuffle_epi32
 999 // CHECK: %[[AND:[0-9a-zA-Z_.]+]] = and i32 %{{[0-9a-zA-Z_.]+}}, 3
1000 // CHECK: sext i32 %[[AND]] to i64
1001 // CHECK: %[[SHR:[0-9a-zA-Z_.]+]] = ashr i32 %{{[0-9a-zA-Z_.]+}}, 2
1002 // CHECK: %[[AND2:[0-9a-zA-Z_.]+]] = and i32 %[[SHR]], 3
1003 // CHECK: sext i32 %[[AND2]] to i64
1004 // CHECK: %[[SHR2:[0-9a-zA-Z_.]+]] = ashr i32 %{{[0-9a-zA-Z_.]+}}, 4
1005 // CHECK: %[[AND3:[0-9a-zA-Z_.]+]] = and i32 %[[SHR2]], 3
1006 // CHECK: sext i32 %[[AND3]] to i64
1007 // CHECK: %[[SHR:[0-9a-zA-Z_.]+]] = ashr i32 %{{[0-9a-zA-Z_.]+}}, 6
1008 // CHECK: %[[AND4:[0-9a-zA-Z_.]+]] = and i32 %[[SHR]], 3
1009 // CHECK: sext i32 %[[AND4]] to i64
1010 // CHECK: getelementptr inbounds [4 x i32], ptr @_mm_shuffle_epi32.__permute_selectors, i64 0, i64 %{{[0-9a-zA-Z_.]+}}
1011 // CHECK: insertelement <4 x i32> %{{[0-9a-zA-Z_.]+}}, i32 %{{[0-9a-zA-Z_.]+}}, i32 0
1012 // CHECK: getelementptr inbounds [4 x i32], ptr @_mm_shuffle_epi32.__permute_selectors, i64 0, i64 %{{[0-9a-zA-Z_.]+}}
1013 // CHECK: insertelement <4 x i32> %{{[0-9a-zA-Z_.]+}}, i32 %{{[0-9a-zA-Z_.]+}}, i32 1
1014 // CHECK: getelementptr inbounds [4 x i32], ptr @_mm_shuffle_epi32.__permute_selectors, i64 0, i64 %{{[0-9a-zA-Z_.]+}}
1015 // CHECK: %[[ADD:[0-9a-zA-Z_.]+]] = add i32 %{{[0-9a-zA-Z_.]+}}, 269488144
1016 // CHECK: insertelement <4 x i32> %{{[0-9a-zA-Z_.]+}}, i32 %[[ADD]], i32 2
1017 // CHECK: getelementptr inbounds [4 x i32], ptr @_mm_shuffle_epi32.__permute_selectors, i64 0, i64 %{{[0-9a-zA-Z_.]+}}
1018 // CHECK: add i32 %{{[0-9a-zA-Z_.]+}}, 269488144
1019 // CHECK: call <4 x i32> @vec_perm(int vector[4], int vector[4], unsigned char vector[16])
1020
1021 // CHECK-LABEL: define available_externally <2 x double> @_mm_shuffle_pd
1022 // CHECK: and i32 %{{[0-9a-zA-Z_.]+}}, 3
1023 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp eq i32 %{{[0-9a-zA-Z_.]+}}, 0
1024 // CHECK: br i1 %[[CMP]]
1025 // CHECK: call <2 x double> @vec_mergeh(double vector[2], double vector[2])
1026 // CHECK: %[[CMP2:[0-9a-zA-Z_.]+]] = icmp eq i32 %{{[0-9a-zA-Z_.]+}}, 1
1027 // CHECK: br i1 %[[CMP2]]
1028 // CHECK: shufflevector <2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i32> <i32 1, i32 2>
1029 // CHECK: %[[CMP3:[0-9a-zA-Z_.]+]] = icmp eq i32 %{{[0-9a-zA-Z_.]+}}, 2
1030 // CHECK: br i1 %[[CMP3]]
1031 // CHECK: shufflevector <2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i32> <i32 0, i32 3>
1032 // CHECK: call <2 x double> @vec_mergel(double vector[2], double vector[2])
1033
1034 // CHECK-LABEL: define available_externally <2 x i64> @_mm_shufflehi_epi16
1035 // CHECK: %[[AND:[0-9a-zA-Z_.]+]] = and i32 %{{[0-9a-zA-Z_.]+}}, 3
1036 // CHECK: sext i32 %[[AND]] to i64
1037 // CHECK: %[[SHR:[0-9a-zA-Z_.]+]] = ashr i32 %{{[0-9a-zA-Z_.]+}}, 2
1038 // CHECK: %[[AND2:[0-9a-zA-Z_.]+]] = and i32 %[[SHR]], 3
1039 // CHECK: sext i32 %[[AND2]] to i64
1040 // CHECK: %[[SHR2:[0-9a-zA-Z_.]+]] = ashr i32 %{{[0-9a-zA-Z_.]+}}, 4
1041 // CHECK: %[[AND3:[0-9a-zA-Z_.]+]] = and i32 %[[SHR2]], 3
1042 // CHECK: sext i32 %[[AND3]] to i64
1043 // CHECK: %[[SHR3:[0-9a-zA-Z_.]+]] = ashr i32 %{{[0-9a-zA-Z_.]+}}, 6
1044 // CHECK: %[[AND4:[0-9a-zA-Z_.]+]] = and i32 %[[SHR3]], 3
1045 // CHECK: sext i32 %[[AND4]] to i64
1046 // CHECK-LE: store <2 x i64> <i64 1663540288323457296, i64 0>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
1047 // CHECK-BE: store <2 x i64> <i64 1157726452361532951, i64 0>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
1048 // CHECK-COUNT-4: getelementptr inbounds [4 x i16], ptr @_mm_shufflehi_epi16.__permute_selectors, i64 0, i64 {{[0-9a-zA-Z_%.]+}}
1049 // CHECK: call <2 x i64> @vec_perm(unsigned long long vector[2], unsigned long long vector[2], unsigned char vector[16])
1050
1051 // CHECK-LABEL: define available_externally <2 x i64> @_mm_shufflelo_epi16
1052 // CHECK: %[[AND:[0-9a-zA-Z_.]+]] = and i32 {{[0-9a-zA-Z_%.]+}}, 3
1053 // CHECK: sext i32 %[[AND]] to i64
1054 // CHECK: %[[SHR:[0-9a-zA-Z_.]+]] = ashr i32 {{[0-9a-zA-Z_%.]+}}, 2
1055 // CHECK: %[[AND2:[0-9a-zA-Z_.]+]] = and i32 %[[SHR]], 3
1056 // CHECK: sext i32 %[[AND2]] to i64
1057 // CHECK: %[[SHR2:[0-9a-zA-Z_.]+]] = ashr i32 {{[0-9a-zA-Z_%.]+}}, 4
1058 // CHECK: %[[AND3:[0-9a-zA-Z_.]+]] = and i32 %[[SHR2]], 3
1059 // CHECK: sext i32 %[[AND3]] to i64
1060 // CHECK: %[[SHR3:[0-9a-zA-Z_.]+]] = ashr i32 {{[0-9a-zA-Z_%.]+}}, 6
1061 // CHECK: %[[AND4:[0-9a-zA-Z_.]+]] = and i32 %[[SHR3]], 3
1062 // CHECK: sext i32 %[[AND4]] to i64
1063 // CHECK-LE: store <2 x i64> <i64 0, i64 2242261671028070680>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
1064 // CHECK-BE: store <2 x i64> <i64 0, i64 1736447835066146335>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
1065 // CHECK-COUNT-4: getelementptr inbounds [4 x i16], ptr @_mm_shufflelo_epi16.__permute_selectors, i64 0, i64 {{[0-9a-zA-Z_%.]+}}
1066 // CHECK: call <2 x i64> @vec_perm(unsigned long long vector[2], unsigned long long vector[2], unsigned char vector[16])
1067
1068 void __attribute__((noinline))
1069 test_sll() {
1070   resi = _mm_sll_epi16(mi1, mi2);
1071   resi = _mm_sll_epi32(mi1, mi2);
1072   resi = _mm_sll_epi64(mi1, mi2);
1073   resi = _mm_slli_epi16(mi1, i);
1074   resi = _mm_slli_epi32(mi1, i);
1075   resi = _mm_slli_epi64(mi1, i);
1076   resi = _mm_slli_si128(mi1, i);
1077 }
1078
1079 // CHECK-LABEL: @test_sll
1080
1081 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sll_epi16
1082 // 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
1083 // CHECK-LE: call <8 x i16> @vec_splat(unsigned short vector[8], unsigned int)
1084 // CHECK-BE: call <8 x i16> @vec_splat(unsigned short vector[8], unsigned int)
1085 // 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>)
1086 // CHECK: call <8 x i16> @vec_sl(unsigned short vector[8], unsigned short vector[8])
1087 // CHECK: call <8 x i16> @vec_sel(unsigned short vector[8], unsigned short vector[8], bool vector[8])
1088
1089 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sll_epi32
1090 // 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)
1091 // 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)
1092 // 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>)
1093 // CHECK: call <4 x i32> @vec_sl(unsigned int vector[4], unsigned int vector[4])
1094 // CHECK: call <4 x i32> @vec_sel(unsigned int vector[4], unsigned int vector[4], bool vector[4])
1095
1096 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sll_epi64
1097 // 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)
1098 // 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>)
1099 // CHECK: call <2 x i64> @vec_sl(unsigned long long vector[2], unsigned long long vector[2])
1100 // CHECK: call <2 x i64> @vec_sel(unsigned long long vector[2], unsigned long long vector[2], bool vector[2])
1101
1102 // CHECK-LABEL: define available_externally <2 x i64> @_mm_slli_epi16
1103 // CHECK: store <8 x i16> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
1104 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp sge i32 %{{[0-9a-zA-Z_.]+}}, 0
1105 // CHECK: br i1 %[[CMP]]
1106 // CHECK: %[[CMP2:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16
1107 // CHECK: br i1 %[[CMP2]]
1108 // CHECK: call i1 @llvm.is.constant
1109 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i8
1110 // CHECK: call <8 x i16> @vec_splat_s16(signed char)(i8 noundef signext %[[TRUNC]])
1111 // CHECK: %[[TRUNC2:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i16
1112 // CHECK: call <8 x i16> @vec_splats(unsigned short)(i16 noundef zeroext %[[TRUNC2]])
1113 // CHECK: call <8 x i16> @vec_sl(short vector[8], unsigned short vector[8])
1114
1115 // CHECK-LABEL: define available_externally <2 x i64> @_mm_slli_epi32
1116 // CHECK: store <4 x i32> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
1117 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp sge i32 %{{[0-9a-zA-Z_.]+}}, 0
1118 // CHECK: br i1 %[[CMP]]
1119 // CHECK: %[[CMP2:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 32
1120 // CHECK: br i1 %[[CMP2]]
1121 // CHECK: call i1 @llvm.is.constant
1122 // CHECK: %[[CMP3:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16
1123 // CHECK: br i1 %[[CMP3]]
1124 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i8
1125 // CHECK: call <4 x i32> @vec_splat_s32(signed char)(i8 noundef signext %[[TRUNC]])
1126 // CHECK: call <4 x i32> @vec_splats(unsigned int)
1127 // CHECK: call <4 x i32> @vec_sl(int vector[4], unsigned int vector[4])
1128
1129 // CHECK-LABEL: define available_externally <2 x i64> @_mm_slli_epi64
1130 // CHECK: store <2 x i64> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
1131 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp sge i32 %{{[0-9a-zA-Z_.]+}}, 0
1132 // CHECK: br i1 %[[CMP]]
1133 // CHECK: %[[CMP2:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 64
1134 // CHECK: br i1 %[[CMP2]]
1135 // CHECK: call i1 @llvm.is.constant
1136 // CHECK: %[[CMP3:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16
1137 // CHECK: br i1 %[[CMP3]]
1138 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i8
1139 // CHECK: call <4 x i32> @vec_splat_s32(signed char)(i8 noundef signext %[[TRUNC]])
1140 // CHECK: call <4 x i32> @vec_splats(unsigned int)
1141 // CHECK: call <2 x i64> @vec_sl(long long vector[2], unsigned long long vector[2])
1142
1143 // CHECK-LABEL: define available_externally <2 x i64> @_mm_slli_si128
1144 // CHECK: store <16 x i8> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
1145 // CHECK-BE: %[[SUB:[0-9a-zA-Z_.]+]] = sub nsw i32 16, %{{[0-9a-zA-Z_.]+}}
1146 // 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]])
1147 // 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_.]+}})
1148 // CHECK: store <16 x i8> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
1149
1150 void __attribute__((noinline))
1151 test_sqrt() {
1152   resd = _mm_sqrt_pd(md1);
1153   resd = _mm_sqrt_sd(md1, md2);
1154 }
1155
1156 // CHECK-LABEL: @test_sqrt
1157
1158 // CHECK-LABEL: define available_externally <2 x double> @_mm_sqrt_pd
1159 // CHECK: call <2 x double> @vec_sqrt(double vector[2])(<2 x double> noundef {{[0-9a-zA-Z_%.]+}})
1160
1161 // CHECK-LABEL: define available_externally <2 x double> @_mm_sqrt_sd
1162 // CHECK: %[[CALL:[0-9a-zA-Z_.]+]] = call <2 x double> @_mm_set1_pd(double noundef %{{[0-9a-zA-Z_.]+}})
1163 // CHECK: call <2 x double> @vec_sqrt(double vector[2])(<2 x double> noundef %{{[0-9a-zA-Z_.]+}})
1164 // CHECK: call <2 x double> @_mm_setr_pd(double noundef %{{[0-9a-zA-Z_.]+}}, double noundef %{{[0-9a-zA-Z_.]+}})
1165
1166 void __attribute__((noinline))
1167 test_sra() {
1168   resi = _mm_sra_epi16(mi1, mi2);
1169   resi = _mm_sra_epi32(mi1, mi2);
1170   resi = _mm_srai_epi16(mi1, i);
1171   resi = _mm_srai_epi32(mi1, i);
1172 }
1173
1174 // CHECK-LABEL: @test_sra
1175
1176 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sra_epi16
1177 // 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
1178 // 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)
1179 // 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)
1180 // 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>)
1181 // CHECK: call <8 x i16> @vec_sra(short vector[8], unsigned short vector[8])
1182
1183 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sra_epi32
1184 // CHECK: store <4 x i32> <i32 31, i32 31, i32 31, i32 31>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
1185 // 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)
1186 // 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)
1187 // 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>)
1188 // CHECK: call <4 x i32> @vec_sra(int vector[4], unsigned int vector[4])
1189
1190 // CHECK-LABEL: define available_externally <2 x i64> @_mm_srai_epi16
1191 // 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
1192 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16
1193 // CHECK: br i1 %[[CMP]]
1194 // CHECK: call i1 @llvm.is.constant
1195 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i8
1196 // CHECK: call <8 x i16> @vec_splat_s16(signed char)(i8 noundef signext %[[TRUNC]])
1197 // CHECK: %[[TRUNC2:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i16
1198 // CHECK: call <8 x i16> @vec_splats(unsigned short)(i16 noundef zeroext %{{[0-9a-zA-Z_.]+}})
1199 // CHECK: call <8 x i16> @vec_sra(short vector[8], unsigned short vector[8])
1200
1201 // CHECK-LABEL: define available_externally <2 x i64> @_mm_srai_epi32
1202 // CHECK: store <4 x i32> <i32 31, i32 31, i32 31, i32 31>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
1203 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 32
1204 // CHECK: br i1 %[[CMP]]
1205 // CHECK: call i1 @llvm.is.constant
1206 // CHECK: %[[CMP2:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16
1207 // CHECK: br i1 %[[CMP2]]
1208 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i8
1209 // CHECK: call <4 x i32> @vec_splat_s32(signed char)(i8 noundef signext %[[TRUNC]])
1210 // CHECK: call <4 x i32> @vec_splats(unsigned int)
1211 // CHECK: call <4 x i32> @vec_splats(unsigned int)
1212 // CHECK: call <4 x i32> @vec_sra(int vector[4], unsigned int vector[4])
1213
1214 void __attribute__((noinline))
1215 test_srl() {
1216   resi = _mm_srl_epi16(mi1, mi2);
1217   resi = _mm_srl_epi32(mi1, mi2);
1218   resi = _mm_srl_epi64(mi1, mi2);
1219   resi = _mm_srli_epi16(mi1, i);
1220   resi = _mm_srli_epi32(mi1, i);
1221   resi = _mm_srli_epi64(mi1, i);
1222   resi = _mm_srli_si128(mi1, i);
1223 }
1224
1225 // CHECK-LABEL: @test_srl
1226
1227 // CHECK-LABEL: define available_externally <2 x i64> @_mm_srl_epi16
1228 // 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
1229 // 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)
1230 // 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)
1231 // 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>)
1232 // CHECK: call <8 x i16> @vec_sr(unsigned short vector[8], unsigned short vector[8])
1233 // CHECK: call <8 x i16> @vec_sel(unsigned short vector[8], unsigned short vector[8], bool vector[8])
1234
1235 // CHECK-LABEL: define available_externally <2 x i64> @_mm_srl_epi32
1236 // 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)
1237 // 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)
1238 // 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>)
1239 // CHECK: call <4 x i32> @vec_sr(unsigned int vector[4], unsigned int vector[4])
1240 // CHECK: call <4 x i32> @vec_sel(unsigned int vector[4], unsigned int vector[4], bool vector[4])
1241
1242 // CHECK-LABEL: define available_externally <2 x i64> @_mm_srl_epi64
1243 // 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)
1244 // 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>)
1245 // CHECK: call <2 x i64> @vec_sr(unsigned long long vector[2], unsigned long long vector[2])
1246 // CHECK: call <2 x i64> @vec_sel(unsigned long long vector[2], unsigned long long vector[2], bool vector[2])
1247
1248 // CHECK-LABEL: define available_externally <2 x i64> @_mm_srli_epi16
1249 // CHECK: store <8 x i16> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
1250 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16
1251 // CHECK: br i1 %[[CMP]]
1252 // CHECK: call i1 @llvm.is.constant
1253 // CHECK: trunc i32 %{{[0-9a-zA-Z_.]+}} to i8
1254 // CHECK: call <8 x i16> @vec_splat_s16(signed char)
1255 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i16
1256 // CHECK: call <8 x i16> @vec_splats(unsigned short)(i16 noundef zeroext %[[TRUNC]])
1257 // CHECK: call <8 x i16> @vec_sr(short vector[8], unsigned short vector[8])
1258
1259 // CHECK-LABEL: define available_externally <2 x i64> @_mm_srli_epi32
1260 // CHECK: store <4 x i32> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
1261 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 32
1262 // CHECK: br i1 %[[CMP]]
1263 // CHECK: call i1 @llvm.is.constant
1264 // CHECK: %[[CMP2:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16
1265 // CHECK: br i1 %[[CMP2]]
1266 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i8
1267 // CHECK: call <4 x i32> @vec_splat_s32(signed char)
1268 // CHECK: call <4 x i32> @vec_splats(unsigned int)
1269 // CHECK: call <4 x i32> @vec_splats(unsigned int)
1270 // CHECK: call <4 x i32> @vec_sr(int vector[4], unsigned int vector[4])
1271
1272 // CHECK-LABEL: define available_externally <2 x i64> @_mm_srli_epi64
1273 // CHECK: store <2 x i64> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
1274 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 64
1275 // CHECK: br i1 %[[CMP]]
1276 // CHECK: call i1 @llvm.is.constant
1277 // CHECK: %[[CMP2:[0-9a-zA-Z_.]+]] = icmp slt i32 %{{[0-9a-zA-Z_.]+}}, 16
1278 // CHECK: br i1 %[[CMP2]]
1279 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i8
1280 // CHECK: call <4 x i32> @vec_splat_s32(signed char)(i8 noundef signext %[[TRUNC]])
1281 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = sext i32 %{{[0-9a-zA-Z_.]+}} to i64
1282 // CHECK: call <2 x i64> @vec_splats(unsigned long long)(i64 noundef %[[EXT]])
1283 // CHECK: call <4 x i32> @vec_splats(unsigned int)(i32 noundef zeroext %{{[0-9a-zA-Z_.]+}})
1284 // CHECK: call <2 x i64> @vec_sr(long long vector[2], unsigned long long vector[2])
1285
1286 // CHECK-LABEL: define available_externally <2 x i64> @_mm_srli_si128
1287 // CHECK: call <2 x i64> @_mm_bsrli_si128
1288
1289 void __attribute__((noinline))
1290 test_store() {
1291   _mm_store_pd(dp, md1);
1292   _mm_store_pd1(dp, md1);
1293   _mm_store_sd(dp, md1);
1294   _mm_store_si128(mip, mi1);
1295   _mm_store1_pd(dp, md1);
1296   _mm_storeh_pd(dp, md1);
1297   _mm_storel_epi64(mip, mi1);
1298   _mm_storel_pd(dp, md1);
1299   _mm_storer_pd(dp, md1);
1300   _mm_storeu_pd(dp, md1);
1301   _mm_storeu_si128(mip, mi1);
1302 }
1303
1304 // CHECK-LABEL: @test_store
1305
1306 // CHECK-LABEL: define available_externally void @_mm_store_pd
1307 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load ptr, ptr %{{[0-9a-zA-Z_.]+}}, align 8
1308 // 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]])
1309
1310 // CHECK-LABEL: define available_externally void @_mm_store_pd1
1311 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load ptr, ptr %{{[0-9a-zA-Z_.]+}}, align 8
1312 // CHECK: %[[ADDR2:[0-9a-zA-Z_.]+]] = load <2 x double>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
1313 // CHECK: call void @_mm_store1_pd(ptr noundef %[[ADDR]], <2 x double> noundef %[[ADDR2]])
1314
1315 // CHECK-LABEL: define available_externally void @_mm_store_sd
1316 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load ptr, ptr %{{[0-9a-zA-Z_.]+}}, align 8
1317 // CHECK: store double %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR]]
1318
1319 // CHECK-LABEL: define available_externally void @_mm_store_si128
1320 // CHECK: %[[LOAD:[0-9a-zA-Z_.]+]] = load ptr, ptr %{{[0-9a-zA-Z_.]+}}, align 8
1321 // 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]])
1322
1323 // CHECK-LABEL: define available_externally void @_mm_store1_pd
1324 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load ptr, ptr %{{[0-9a-zA-Z_.]+}}, align 8
1325 // CHECK: %[[VAL:[0-9a-zA-Z_.]+]] = load <2 x double>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
1326 // 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)
1327 // CHECK: call void @_mm_store_pd(ptr noundef %[[ADDR]], <2 x double> noundef %[[CALL]])
1328
1329 // CHECK-LABEL: define available_externally void @_mm_storeh_pd
1330 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load ptr, ptr %{{[0-9a-zA-Z_.]+}}, align 8
1331 // CHECK: store double %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR]]
1332
1333 // CHECK-LABEL: define available_externally void @_mm_storel_epi64
1334 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load ptr, ptr %{{[0-9a-zA-Z_.]+}}, align 8
1335 // CHECK: store i64 %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR]], align 8
1336
1337 // CHECK-LABEL: define available_externally void @_mm_storel_pd
1338 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load ptr, ptr %{{[0-9a-zA-Z_.]+}}, align 8
1339 // CHECK: %[[VAL:[0-9a-zA-Z_.]+]] = load <2 x double>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
1340 // CHECK: call void @_mm_store_sd(ptr noundef %[[ADDR]], <2 x double> noundef %[[VAL]])
1341
1342 // CHECK-LABEL: define available_externally void @_mm_storer_pd
1343 // CHECK: shufflevector <2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i64> %{{[0-9a-zA-Z_.]+}}, <2 x i32> <i32 1, i32 2>
1344 // CHECK: call void @_mm_store_pd(ptr noundef %{{[0-9a-zA-Z_.]+}}, <2 x double> noundef %{{[0-9a-zA-Z_.]+}})
1345
1346 // CHECK-LABEL: define available_externally void @_mm_storeu_pd
1347 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load ptr, ptr %{{[0-9a-zA-Z_.]+}}, align 8
1348 // CHECK: store <2 x double> %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR]], align 1
1349
1350 // CHECK-LABEL: define available_externally void @_mm_storeu_si128
1351 // CHECK: %[[ADDR:[0-9a-zA-Z_.]+]] = load ptr, ptr %{{[0-9a-zA-Z_.]+}}, align 8
1352 // CHECK: store <2 x i64> %{{[0-9a-zA-Z_.]+}}, ptr %[[ADDR]], align 1
1353
1354 void __attribute__((noinline))
1355 test_stream() {
1356   _mm_stream_pd(dp, md1);
1357   _mm_stream_si128(mip, mi1);
1358   _mm_stream_si32(is, i);
1359   _mm_stream_si64(i64s, i64s[1]);
1360 }
1361
1362 // CHECK-LABEL: @test_stream
1363
1364 // CHECK-LABEL: define available_externally void @_mm_stream_pd
1365 // CHECK: call void asm sideeffect "dcbtstt 0,$0", "b,~{memory}"(ptr %{{[0-9a-zA-Z_.]+}})
1366
1367 // CHECK-LABEL: define available_externally void @_mm_stream_si128
1368 // CHECK: call void asm sideeffect "dcbtstt 0,$0", "b,~{memory}"(ptr %{{[0-9a-zA-Z_.]+}})
1369
1370 // CHECK-LABEL: define available_externally void @_mm_stream_si32
1371 // CHECK: call void asm sideeffect "dcbtstt 0,$0", "b,~{memory}"(ptr %{{[0-9a-zA-Z_.]+}})
1372
1373 // CHECK-LABEL: define available_externally void @_mm_stream_si64
1374 // CHECK: call void asm sideeffect "\09dcbtstt\090,$0", "b,~{memory}"(ptr %{{[0-9a-zA-Z_.]+}})
1375
1376 void __attribute__((noinline))
1377 test_sub() {
1378   resi = _mm_sub_epi64(mi1, mi2);
1379   resi = _mm_sub_epi32(mi1, mi2);
1380   resi = _mm_sub_epi16(mi1, mi2);
1381   resi = _mm_sub_epi8(mi1, mi2);
1382   resd = _mm_sub_pd(md1, md2);
1383   resd = _mm_sub_sd(md1, md2);
1384   res64 = _mm_sub_si64(m641, m642);
1385   resi = _mm_subs_epi16(mi1, mi2);
1386   resi = _mm_subs_epi8(mi1, mi2);
1387   resi = _mm_subs_epu16(mi1, mi2);
1388   resi = _mm_subs_epu8(mi1, mi2);
1389 }
1390
1391 // CHECK-LABEL: @test_sub
1392
1393 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sub_epi64
1394 // CHECK: sub <2 x i64>
1395
1396 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sub_epi32
1397 // CHECK: sub <4 x i32>
1398
1399 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sub_epi16
1400 // CHECK: sub <8 x i16>
1401
1402 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sub_epi8
1403 // CHECK: sub <16 x i8>
1404
1405 // CHECK-LABEL: define available_externally <2 x double> @_mm_sub_pd
1406 // CHECK: fsub <2 x double>
1407
1408 // CHECK-LABEL: define available_externally <2 x double> @_mm_sub_sd
1409 // CHECK: fsub double
1410
1411 // CHECK-LABEL: define available_externally i64 @_mm_sub_si64
1412 // CHECK: sub i64
1413
1414 // CHECK-LABEL: define available_externally <2 x i64> @_mm_subs_epi16
1415 // CHECK: call <8 x i16> @vec_subs(short vector[8], short vector[8])
1416
1417 // CHECK-LABEL: define available_externally <2 x i64> @_mm_subs_epi8
1418 // CHECK: call <16 x i8> @vec_subs(signed char vector[16], signed char vector[16])
1419
1420 // CHECK-LABEL: define available_externally <2 x i64> @_mm_subs_epu16
1421 // CHECK: call <8 x i16> @vec_subs(unsigned short vector[8], unsigned short vector[8])
1422
1423 // CHECK-LABEL: define available_externally <2 x i64> @_mm_subs_epu8
1424 // CHECK: call <16 x i8> @vec_subs(unsigned char vector[16], unsigned char vector[16])
1425
1426 void __attribute__((noinline))
1427 test_ucomi() {
1428   i = _mm_ucomieq_sd(md1, md2);
1429   i = _mm_ucomige_sd(md1, md2);
1430   i = _mm_ucomigt_sd(md1, md2);
1431   i = _mm_ucomile_sd(md1, md2);
1432   i = _mm_ucomilt_sd(md1, md2);
1433   i = _mm_ucomineq_sd(md1, md2);
1434 }
1435
1436 // CHECK-LABEL: @test_ucomi
1437
1438 // CHECK-LABEL: define available_externally signext i32 @_mm_ucomieq_sd
1439 // CHECK: fcmp oeq double
1440
1441 // CHECK-LABEL: define available_externally signext i32 @_mm_ucomige_sd
1442 // CHECK: fcmp oge double
1443
1444 // CHECK-LABEL: define available_externally signext i32 @_mm_ucomigt_sd
1445 // CHECK: fcmp ogt double
1446
1447 // CHECK-LABEL: define available_externally signext i32 @_mm_ucomile_sd
1448 // CHECK: fcmp ole double
1449
1450 // CHECK-LABEL: define available_externally signext i32 @_mm_ucomilt_sd
1451 // CHECK: fcmp olt double
1452
1453 // CHECK-LABEL: define available_externally signext i32 @_mm_ucomineq_sd
1454 // CHECK: fcmp une double
1455
1456 void __attribute__((noinline))
1457 test_undefined() {
1458   resd = _mm_undefined_pd();
1459   resi = _mm_undefined_si128();
1460 }
1461
1462 // CHECK-LABEL: @test_undefined
1463
1464 // CHECK-LABEL: define available_externally <2 x double> @_mm_undefined_pd()
1465 // CHECK: %[[VAL:[0-9a-zA-Z_.]+]] = load <2 x double>, ptr %[[ADDR:[0-9a-zA-Z_.]+]], align 16
1466 // CHECK: store <2 x double> %[[VAL]], ptr %[[ADDR]], align 16
1467 // CHECK: load <2 x double>, ptr %[[ADDR]], align 16
1468
1469 // CHECK-LABEL: define available_externally <2 x i64> @_mm_undefined_si128()
1470 // CHECK: %[[VAL:[0-9a-zA-Z_.]+]] = load <2 x i64>, ptr %[[ADDR:[0-9a-zA-Z_.]+]], align 16
1471 // CHECK: store <2 x i64> %[[VAL]], ptr %[[ADDR]], align 16
1472 // CHECK: load <2 x i64>, ptr %[[ADDR]], align 16
1473
1474 void __attribute__((noinline))
1475 test_unpack() {
1476   resi = _mm_unpackhi_epi16(mi1, mi2);
1477   resi = _mm_unpackhi_epi32(mi1, mi2);
1478   resi = _mm_unpackhi_epi64(mi1, mi2);
1479   resi = _mm_unpackhi_epi8(mi1, mi2);
1480   resd = _mm_unpackhi_pd(md1, md2);
1481   resi = _mm_unpacklo_epi16(mi1, mi2);
1482   resi = _mm_unpacklo_epi32(mi1, mi2);
1483   resi = _mm_unpacklo_epi64(mi1, mi2);
1484   resi = _mm_unpacklo_epi8(mi1, mi2);
1485   resd = _mm_unpacklo_pd(md1, md2);
1486 }
1487
1488 // CHECK-LABEL: @test_unpack
1489
1490 // CHECK-LABEL: define available_externally <2 x i64> @_mm_unpackhi_epi16
1491 // CHECK: call <8 x i16> @vec_mergel(unsigned short vector[8], unsigned short vector[8])
1492
1493 // CHECK-LABEL: define available_externally <2 x i64> @_mm_unpackhi_epi32
1494 // CHECK: call <4 x i32> @vec_mergel(unsigned int vector[4], unsigned int vector[4])
1495
1496 // CHECK-LABEL: define available_externally <2 x i64> @_mm_unpackhi_epi64
1497 // CHECK: call <2 x i64> @vec_mergel(long long vector[2], long long vector[2])
1498
1499 // CHECK-LABEL: define available_externally <2 x i64> @_mm_unpackhi_epi8
1500 // CHECK: call <16 x i8> @vec_mergel(unsigned char vector[16], unsigned char vector[16])
1501
1502 // CHECK-LABEL: define available_externally <2 x double> @_mm_unpackhi_pd
1503 // CHECK: call <2 x double> @vec_mergel(double vector[2], double vector[2])
1504
1505 // CHECK-LABEL: define available_externally <2 x i64> @_mm_unpacklo_epi16
1506 // CHECK: call <8 x i16> @vec_mergeh(short vector[8], short vector[8])
1507
1508 // CHECK-LABEL: define available_externally <2 x i64> @_mm_unpacklo_epi32
1509 // CHECK: call <4 x i32> @vec_mergeh(int vector[4], int vector[4])
1510
1511 // CHECK-LABEL: define available_externally <2 x i64> @_mm_unpacklo_epi64
1512 // CHECK: call <2 x i64> @vec_mergeh(long long vector[2], long long vector[2])
1513
1514 // CHECK-LABEL: define available_externally <2 x i64> @_mm_unpacklo_epi8
1515 // CHECK: call <16 x i8> @vec_mergeh(unsigned char vector[16], unsigned char vector[16])
1516
1517 // CHECK-LABEL: define available_externally <2 x double> @_mm_unpacklo_pd
1518 // CHECK: call <2 x double> @vec_mergeh(double vector[2], double vector[2])