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

   1 // REQUIRES: powerpc-registered-target
   2
   3 // RUN: %clang -S -emit-llvm -target powerpc64le-unknown-linux-gnu -mcpu=pwr8 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \
   4 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s
   5 // RUN: %clang -S -emit-llvm -target powerpc64-unknown-linux-gnu -mcpu=pwr8 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \
   6 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s
   7
   8 // RUN: %clang -S -emit-llvm -target powerpc64le-unknown-freebsd13.0 -mcpu=pwr8 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \
   9 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s
  10 // RUN: %clang -S -emit-llvm -target powerpc64-unknown-freebsd13.0 -mcpu=pwr8 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \
  11 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s
  12
  13 // RUN: %clang -S -emit-llvm -target powerpc64le-unknown-linux-gnu -mcpu=pwr10 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \
  14 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefix=P10
  15 // RUN: %clang -S -emit-llvm -target powerpc64-unknown-linux-gnu -mcpu=pwr10 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \
  16 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefix=P10
  17
  18 // RUN: %clang -x c++ -S -emit-llvm -target powerpc64le-unknown-linux-gnu -mcpu=pwr8 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \
  19 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -fsyntax-only
  20 // RUN: %clang -x c++ -S -emit-llvm -target powerpc64le-unknown-linux-gnu -mcpu=pwr10 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \
  21 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -fsyntax-only
  22
  23 // RUN: %clang -S -emit-llvm -target powerpc64le-unknown-freebsd13.0 -mcpu=pwr8 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \
  24 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s
  25 // RUN: %clang -S -emit-llvm -target powerpc64-unknown-freebsd13.0 -mcpu=pwr8 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \
  26 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s
  27
  28 // RUN: %clang -S -emit-llvm -target powerpc64-ibm-aix -mcpu=pwr10 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \
  29 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefix=P10
  30 // RUN: %clang -S -emit-llvm -target powerpc64-ibm-aix -mcpu=pwr8 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \
  31 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s
  32
  33 #include <smmintrin.h>
  34
  35 __m128 mn1, mn2;
  36 __m128d md1, md2;
  37 __m128i mi, m1, m2;
  38
  39 void __attribute__((noinline))
  40 test_extract() {
  41   _mm_extract_epi8(mi, 0);
  42   _mm_extract_epi32(mi, 0);
  43   _mm_extract_epi64(mi, 0);
  44   _mm_extract_ps((__m128)mi, 0);
  45 }
  46
  47 // CHECK-LABEL: @test_extract
  48
  49 // CHECK-LABEL: define available_externally signext i32 @_mm_extract_epi8(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef signext %{{[0-9a-zA-Z_.]+}})
  50 // CHECK: %[[AND:[0-9a-zA-Z_.]+]] = and i32 %{{[0-9a-zA-Z_.]+}}, 15
  51 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = extractelement <16 x i8> %{{[0-9a-zA-Z_.]+}}, i32 %[[AND]]
  52 // CHECK: zext i8 %[[EXT]] to i32
  53
  54 // CHECK-LABEL: define available_externally signext i32 @_mm_extract_epi32(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef signext %{{[0-9a-zA-Z_.]+}})
  55 // CHECK: %[[AND:[0-9a-zA-Z_.]+]] = and i32 %{{[0-9a-zA-Z_.]+}}, 3
  56 // CHECK: extractelement <4 x i32> %{{[0-9a-zA-Z_.]+}}, i32 %[[AND]]
  57
  58 // CHECK-LABEL: define available_externally signext i32 @_mm_extract_epi64(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef signext %{{[0-9a-zA-Z_.]+}})
  59 // CHECK: %[[AND:[0-9a-zA-Z_.]+]] = and i32 %{{[0-9a-zA-Z_.]+}}, 1
  60 // CHECK: %[[EXT:[0-9a-zA-Z_.]+]] = extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 %[[AND]]
  61 // CHECK: trunc i64 %[[EXT]] to i32
  62
  63 // CHECK-LABEL: define available_externally signext i32 @_mm_extract_ps(<4 x float> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef signext %{{[0-9a-zA-Z_.]+}})
  64 // CHECK: %[[AND:[0-9a-zA-Z_.]+]] = and i32 %{{[0-9a-zA-Z_.]+}}, 3
  65 // CHECK: extractelement <4 x i32> %{{[0-9a-zA-Z_.]+}}, i32 %[[AND]]
  66
  67 void __attribute__((noinline))
  68 test_blend() {
  69   _mm_blend_epi16(m1, m2, 0);
  70   _mm_blendv_epi8(m1, m2, mi);
  71   _mm_blend_ps(mn1, mn2, 0);
  72   _mm_blendv_ps(mn1, mn2, mn1);
  73   _mm_blend_pd(md1, md2, 0);
  74   _mm_blendv_pd(md1, md2, md1);
  75 }
  76
  77 // CHECK-LABEL: @test_blend
  78
  79 // CHECK-LABEL: define available_externally <2 x i64> @_mm_blend_epi16(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef signext %{{[0-9a-zA-Z_.]+}})
  80 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i8
  81 // CHECK: call <16 x i8> @vec_splats(unsigned char)(i8 noundef zeroext %[[TRUNC]])
  82 // CHECK: call <16 x i8> @llvm.ppc.altivec.vgbbd(<16 x i8> %{{[0-9a-zA-Z_.]+}})
  83 // CHECK: %[[PACK:[0-9a-zA-Z_.]+]] = call <8 x i16> @vec_unpackh(signed char vector[16])
  84 // CHECK: store <8 x i16> %[[PACK]], ptr %{{[0-9a-zA-Z_.]+}}, align 16
  85 // BE: %[[REVE:[0-9a-zA-Z_.]+]] = call <8 x i16> @vec_reve(unsigned short vector[8])
  86 // BE: store <8 x i16> %[[REVE]], ptr %{{[0-9a-zA-Z_.]+}}, align 16
  87 // CHECK: call <8 x i16> @vec_sel(unsigned short vector[8], unsigned short vector[8], unsigned short vector[8])
  88
  89 // P10-LABEL: define available_externally <2 x i64> @_mm_blendv_epi8(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
  90 // P10: call <16 x i8> @vec_blendv(signed char vector[16], signed char vector[16], unsigned char vector[16])(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef %{{[0-9a-zA-Z_.]+}})
  91
  92 // CHECK-LABEL: define available_externally <2 x i64> @_mm_blendv_epi8(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
  93 // CHECK: call <16 x i8> @vec_splats(unsigned char)(i8 noundef zeroext 7)
  94 // CHECK: call <16 x i8> @vec_sra(unsigned char vector[16], unsigned char vector[16])
  95 // CHECK: call <16 x i8> @vec_sel(signed char vector[16], signed char vector[16], unsigned char vector[16])
  96
  97 void __attribute__((noinline))
  98 test_insert() {
  99   _mm_insert_epi8(m1, 1, 0);
 100   _mm_insert_epi32(m1, 1, 0);
 101   _mm_insert_epi64(m1, 0xFFFFFFFF1L, 0);
 102 }
 103
 104 // CHECK-LABEL: @test_insert
 105
 106 // CHECK-LABEL: define available_externally <2 x i64> @_mm_insert_epi8(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef signext %{{[0-9a-zA-Z_.]+}}, i32 noundef signext %{{[0-9a-zA-Z_.]+}})
 107 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %{{[0-9a-zA-Z_.]+}} to i8
 108 // CHECK: %[[AND:[0-9a-zA-Z_.]+]] = and i32 %{{[0-9a-zA-Z_.]+}}, 15
 109 // CHECK: insertelement <16 x i8> %{{[0-9a-zA-Z_.]+}}, i8 %[[TRUNC]], i32 %[[AND]]
 110
 111 // CHECK-LABEL: define available_externally <2 x i64> @_mm_insert_epi32(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef signext %{{[0-9a-zA-Z_.]+}}, i32 noundef signext %{{[0-9a-zA-Z_.]+}})
 112 // CHECK: %[[AND:[0-9a-zA-Z_.]+]] = and i32 %{{[0-9a-zA-Z_.]+}}, 3
 113 // CHECK: insertelement <4 x i32> %{{[0-9a-zA-Z_.]+}}, i32 %{{[0-9a-zA-Z_.]+}}, i32 %[[AND]]
 114
 115 // CHECK-LABEL: define available_externally <2 x i64> @_mm_insert_epi64(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, i64 noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef signext %{{[0-9a-zA-Z_.]+}})
 116 // CHECK: %[[AND:[0-9a-zA-Z_.]+]] = and i32 %{{[0-9a-zA-Z_.]+}}, 1
 117 // CHECK: insertelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i64 %{{[0-9a-zA-Z_.]+}}, i32 %[[AND:[0-9a-zA-Z_.]+]]
 118
 119 void __attribute__((noinline))
 120 test_convert() {
 121   _mm_cvtepi16_epi32(m1);
 122   _mm_cvtepi16_epi64(m1);
 123   _mm_cvtepi32_epi64(m1);
 124   _mm_cvtepi8_epi16(m1);
 125   _mm_cvtepi8_epi32(m1);
 126   _mm_cvtepi8_epi64(m1);
 127   _mm_cvtepu16_epi32(m1);
 128   _mm_cvtepu16_epi64(m1);
 129   _mm_cvtepu32_epi64(m1);
 130   _mm_cvtepu8_epi16(m1);
 131   _mm_cvtepu8_epi32(m1);
 132   _mm_cvtepu8_epi64(m1);
 133 }
 134
 135 // CHECK-LABEL: @test_convert
 136
 137 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtepi16_epi32(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 138 // CHECK: call <4 x i32> @vec_unpackh(short vector[8])
 139
 140 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtepi16_epi64(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 141 // CHECK: call <4 x i32> @vec_unpackh(short vector[8])
 142 // CHECK: call <2 x i64> @vec_unpackh(int vector[4])
 143
 144 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtepi32_epi64(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 145 // CHECK: call <2 x i64> @vec_unpackh(int vector[4])
 146
 147 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtepi8_epi16(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 148 // CHECK: call <8 x i16> @vec_unpackh(signed char vector[16])
 149
 150 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtepi8_epi32(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 151 // CHECK: call <8 x i16> @vec_unpackh(signed char vector[16])
 152 // CHECK: call <4 x i32> @vec_unpackh(short vector[8])
 153
 154 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtepi8_epi64(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 155 // CHECK: call <8 x i16> @vec_unpackh(signed char vector[16])
 156 // CHECK: call <4 x i32> @vec_unpackh(short vector[8])
 157 // CHECK: call <2 x i64> @vec_unpackh(int vector[4])
 158
 159 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtepu16_epi32(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 160 // LE: call <8 x i16> @vec_mergeh(unsigned short vector[8], unsigned short vector[8])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef zeroinitializer)
 161 // BE: call <8 x i16> @vec_mergeh(unsigned short vector[8], unsigned short vector[8])(<8 x i16> noundef zeroinitializer, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}})
 162
 163 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtepu16_epi64(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 164 // LE: call <8 x i16> @vec_mergeh(unsigned short vector[8], unsigned short vector[8])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef zeroinitializer)
 165 // LE: call <4 x i32> @vec_mergeh(unsigned int vector[4], unsigned int vector[4])(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <4 x i32> noundef zeroinitializer)
 166 // BE: call <8 x i16> @vec_mergeh(unsigned short vector[8], unsigned short vector[8])(<8 x i16> noundef zeroinitializer, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}})
 167 // BE: call <4 x i32> @vec_mergeh(unsigned int vector[4], unsigned int vector[4])(<4 x i32> noundef zeroinitializer, <4 x i32> noundef %{{[0-9a-zA-Z_.]+}})
 168
 169 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtepu32_epi64(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 170 // LE: call <4 x i32> @vec_mergeh(unsigned int vector[4], unsigned int vector[4])(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <4 x i32> noundef zeroinitializer)
 171 // BE: call <4 x i32> @vec_mergeh(unsigned int vector[4], unsigned int vector[4])(<4 x i32> noundef zeroinitializer, <4 x i32> noundef %{{[0-9a-zA-Z_.]+}})
 172
 173 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtepu8_epi16(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 174 // LE: call <16 x i8> @vec_mergeh(unsigned char vector[16], unsigned char vector[16])(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef zeroinitializer)
 175 // BE: call <16 x i8> @vec_mergeh(unsigned char vector[16], unsigned char vector[16])(<16 x i8> noundef zeroinitializer, <16 x i8> noundef %{{[0-9a-zA-Z_.]+}})
 176
 177 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtepu8_epi32(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 178 // LE: call <16 x i8> @vec_mergeh(unsigned char vector[16], unsigned char vector[16])(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef zeroinitializer)
 179 // LE: call <8 x i16> @vec_mergeh(unsigned short vector[8], unsigned short vector[8])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef zeroinitializer)
 180 // BE: call <16 x i8> @vec_mergeh(unsigned char vector[16], unsigned char vector[16])(<16 x i8> noundef zeroinitializer, <16 x i8> noundef %{{[0-9a-zA-Z_.]+}})
 181 // BE: call <8 x i16> @vec_mergeh(unsigned short vector[8], unsigned short vector[8])(<8 x i16> noundef zeroinitializer, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}})
 182
 183 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cvtepu8_epi64(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 184 // CHECK: call <16 x i8> @vec_mergeh(unsigned char vector[16], unsigned char vector[16])
 185 // CHECK: call <8 x i16> @vec_mergeh(unsigned short vector[8], unsigned short vector[8])
 186 // CHECK: call <4 x i32> @vec_mergeh(unsigned int vector[4], unsigned int vector[4])
 187
 188 void __attribute__((noinline))
 189 test_minmax() {
 190   _mm_max_epi32(m1, m2);
 191   _mm_max_epi8(m1, m2);
 192   _mm_max_epu16(m1, m2);
 193   _mm_max_epu32(m1, m2);
 194   _mm_min_epi32(m1, m2);
 195   _mm_min_epi8(m1, m2);
 196   _mm_min_epu16(m1, m2);
 197   _mm_min_epu32(m1, m2);
 198   _mm_minpos_epu16(m1);
 199 }
 200
 201 // CHECK-LABEL: @test_minmax
 202
 203 // CHECK-LABEL: define available_externally <2 x i64> @_mm_max_epi32(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 204 // CHECK: call <4 x i32> @vec_max(int vector[4], int vector[4])
 205
 206 // CHECK-LABEL: define available_externally <2 x i64> @_mm_max_epi8(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 207 // CHECK: call <16 x i8> @vec_max(signed char vector[16], signed char vector[16])
 208
 209 // CHECK-LABEL: define available_externally <2 x i64> @_mm_max_epu16(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 210 // CHECK: call <8 x i16> @vec_max(unsigned short vector[8], unsigned short vector[8])
 211
 212 // CHECK-LABEL: define available_externally <2 x i64> @_mm_max_epu32(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 213 // CHECK: call <4 x i32> @vec_max(unsigned int vector[4], unsigned int vector[4])
 214
 215 // CHECK-LABEL: define available_externally <2 x i64> @_mm_min_epi32(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 216 // CHECK: call <4 x i32> @vec_min(int vector[4], int vector[4])
 217
 218 // CHECK-LABEL: define available_externally <2 x i64> @_mm_min_epi8(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 219 // CHECK: call <16 x i8> @vec_min(signed char vector[16], signed char vector[16])
 220
 221 // CHECK-LABEL: define available_externally <2 x i64> @_mm_min_epu16(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 222 // CHECK: call <8 x i16> @vec_min(unsigned short vector[8], unsigned short vector[8])
 223
 224 // CHECK-LABEL: define available_externally <2 x i64> @_mm_min_epu32(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 225 // CHECK: call <4 x i32> @vec_min(unsigned int vector[4], unsigned int vector[4])
 226
 227 // CHECK-LABEL: define available_externally <2 x i64> @_mm_minpos_epu16(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 228 // CHECK: call void @llvm.memset.p0.i64(ptr align 16 %{{[0-9a-zA-Z_.]+}}, i8 0, i64 16, i1 false)
 229 // CHECK: extractelement <8 x i16> %{{[0-9a-zA-Z_.]+}}, i16 %{{[0-9a-zA-Z_.]+}}
 230 // CHECK: %[[VEXT:[0-9a-zA-Z_.]+]] = extractelement <8 x i16> %{{[0-9a-zA-Z_.]+}}, i64 %{{[0-9a-zA-Z_.]+}}
 231 // CHECK: zext i16 %[[VEXT]] to i32
 232 // CHECK: zext i16 %{{[0-9a-zA-Z_.]+}} to i32
 233 // CHECK: extractelement <8 x i16> %{{[0-9a-zA-Z_.]+}}, i64 %{{[0-9a-zA-Z_.]+}}
 234 // CHECK: add i64 %{{[0-9a-zA-Z_.]+}}, 1
 235
 236 void __attribute__((noinline))
 237 test_round() {
 238   _mm_round_ps(mn1, 0);
 239   _mm_round_ss(mn1, mn2, 0);
 240   _mm_round_pd(md1, 0);
 241   _mm_round_sd(md1, md2, 0);
 242 }
 243
 244 // CHECK-LABEL: @test_round
 245
 246 // CHECK-LABEL: define available_externally <4 x float> @_mm_round_ps(<4 x float> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef signext %{{[0-9a-zA-Z_.]+}})
 247 // CHECK: call double @llvm.ppc.readflm()
 248 // CHECK: call void @llvm.ppc.mtfsf(i32 3, double %{{[0-9a-zA-Z_.]+}})
 249 // CHECK: %{{[0-9a-zA-Z_.]+}} = call <4 x float> asm "", "=^wa,0"
 250 // CHECK: call double @llvm.ppc.readflm()
 251 // P10: call double @llvm.ppc.mffsl()
 252 // CHECK: call double @llvm.ppc.setrnd(i32 0)
 253 // CHECK: %{{[0-9a-zA-Z_.]+}} = call <4 x float> asm "", "=^wa,0"
 254 // CHECK: call <4 x float> @vec_rint(float vector[4])
 255 // CHECK: call void asm sideeffect "", "^wa"
 256 // CHECK: call double @llvm.ppc.setrnd(i32 %{{[0-9a-zA-Z_.]+}})
 257 // CHECK: call <4 x float> @vec_floor(float vector[4])
 258 // CHECK: call <4 x float> @vec_ceil(float vector[4])
 259 // CHECK: call <4 x float> @vec_trunc(float vector[4])
 260 // CHECK: call <4 x float> @vec_rint(float vector[4])
 261 // CHECK: call void asm sideeffect "", "^wa"
 262 // CHECK: call double @llvm.ppc.readflm()
 263 // P10: call double @llvm.ppc.mffsl()
 264 // CHECK: call void @llvm.ppc.mtfsf(i32 3, double %{{[0-9a-zA-Z_.]+}})
 265
 266 // CHECK-LABEL: define available_externally <4 x float> @_mm_round_ss(<4 x float> noundef %{{[0-9a-zA-Z_.]+}}, <4 x float> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef signext %{{[0-9a-zA-Z_.]+}})
 267 // CHECK: call <4 x float> @_mm_round_ps(<4 x float> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef signext %{{[0-9a-zA-Z_.]+}})
 268 // CHECK: extractelement <4 x float> %{{[0-9a-zA-Z_.]+}}, i32 0
 269
 270 // CHECK-LABEL: define available_externally <2 x double> @_mm_round_pd(<2 x double> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef signext %{{[0-9a-zA-Z_.]+}})
 271 // CHECK: call double @llvm.ppc.readflm()
 272 // CHECK: call void @llvm.ppc.mtfsf(i32 3, double %{{[0-9a-zA-Z_.]+}})
 273 // CHECK: %{{[0-9a-zA-Z_.]+}} = call <2 x double> asm "", "=^wa,0"
 274 // CHECK: call double @llvm.ppc.readflm()
 275 // P10: call double @llvm.ppc.mffsl()
 276 // CHECK: call double @llvm.ppc.setrnd(i32 0)
 277 // CHECK: %{{[0-9a-zA-Z_.]+}} = call <2 x double> asm "", "=^wa,0"
 278 // CHECK: call <2 x double> @vec_rint(double vector[2])
 279 // CHECK: call void asm sideeffect "", "^wa"
 280 // CHECK: call double @llvm.ppc.setrnd(i32 %{{[0-9a-zA-Z_.]+}})
 281 // CHECK: call <2 x double> @vec_floor(double vector[2])
 282 // CHECK: call <2 x double> @vec_ceil(double vector[2])
 283 // CHECK: call <2 x double> @vec_trunc(double vector[2])
 284 // CHECK: call <2 x double> @vec_rint(double vector[2])
 285 // CHECK: call void asm sideeffect "", "^wa"
 286 // CHECK: call double @llvm.ppc.readflm()
 287 // P10: call double @llvm.ppc.mffsl()
 288 // CHECK: call void @llvm.ppc.mtfsf(i32 3, double %{{[0-9a-zA-Z_.]+}})
 289
 290 // CHECK-LABEL: define available_externally <2 x double> @_mm_round_sd(<2 x double> noundef %{{[0-9a-zA-Z_.]+}}, <2 x double> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef signext %{{[0-9a-zA-Z_.]+}})
 291 // CHECK: call <2 x double> @_mm_round_pd(<2 x double> noundef %{{[0-9a-zA-Z_.]+}}, i32 noundef signext %{{[0-9a-zA-Z_.]+}})
 292 // CHECK: extractelement <2 x double> %{{[0-9a-zA-Z_.]+}}, i32 0
 293 // CHECK: extractelement <2 x double> %{{[0-9a-zA-Z_.]+}}, i32 1
 294
 295 void __attribute__((noinline))
 296 test_testing() {
 297   _mm_testc_si128(m1, m2);
 298   _mm_testnzc_si128(m1, m2);
 299   _mm_testz_si128(m1, m2);
 300 }
 301
 302 // CHECK-LABEL: @test_testing
 303
 304 // CHECK-LABEL: define available_externally signext i32 @_mm_testc_si128(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 305 // CHECK: call <16 x i8> @vec_nor(unsigned char vector[16], unsigned char vector[16])
 306 // CHECK: call <16 x i8> @vec_and(unsigned char vector[16], unsigned char vector[16])
 307 // CHECK: call signext i32 @vec_all_eq(unsigned char vector[16], unsigned char vector[16])(<16 x i8> noundef %call1, <16 x i8> noundef zeroinitializer)
 308
 309 // CHECK-LABEL: define available_externally signext i32 @_mm_testnzc_si128(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 310 // CHECK: call signext i32 @_mm_testz_si128(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 311 // CHECK: call signext i32 @_mm_testc_si128(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 312 // CHECK: zext i1 %{{[0-9a-zA-Z_.]+}} to i32
 313
 314 // CHECK-LABEL: define available_externally signext i32 @_mm_testz_si128(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 315 // CHECK: call <16 x i8> @vec_and(unsigned char vector[16], unsigned char vector[16])
 316 // CHECK: call signext i32 @vec_all_eq(unsigned char vector[16], unsigned char vector[16])(<16 x i8> noundef %call, <16 x i8> noundef zeroinitializer)
 317
 318 void __attribute__((noinline))
 319 test_compare() {
 320   _mm_cmpeq_epi64(m1, m2);
 321   _mm_cmpgt_epi64(m1, m2);
 322 }
 323
 324 // CHECK-LABEL: @test_compare
 325
 326 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmpeq_epi64(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 327 // CHECK: call <2 x i64> @vec_cmpeq(long long vector[2], long long vector[2])
 328
 329 // CHECK-LABEL: define available_externally <2 x i64> @_mm_cmpgt_epi64(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 330 // CHECK: call <2 x i64> @vec_cmpgt(long long vector[2], long long vector[2])
 331
 332 void __attribute__((noinline))
 333 test_mul() {
 334   _mm_mul_epi32(m1, m2);
 335   _mm_mullo_epi32(m1, m2);
 336 }
 337
 338 // CHECK-LABEL: @test_mul
 339
 340 // CHECK-LABEL: define available_externally <2 x i64> @_mm_mul_epi32(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 341 // CHECK: %call = call <2 x i64> @vec_mule(int vector[4], int vector[4])
 342
 343 // CHECK-LABEL: define available_externally <2 x i64> @_mm_mullo_epi32(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 344 // CHECK: %call = call <4 x i32> @vec_mul(unsigned int vector[4], unsigned int vector[4])
 345
 346 void __attribute__((noinline))
 347 test_packus() {
 348   _mm_packus_epi32(m1, m2);
 349 }
 350
 351 // CHECK-LABEL: @test_packus
 352
 353 // CHECK-LABEL: define available_externally <2 x i64> @_mm_packus_epi32(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}}, <2 x i64> noundef %{{[0-9a-zA-Z_.]+}})
 354 // CHECK: call <8 x i16> @vec_packsu(int vector[4], int vector[4])(<4 x i32> noundef %1, <4 x i32> noundef %3)
 355
 356 // To test smmintrin.h includes tmmintrin.h
 357
 358 void __attribute__((noinline))
 359 test_abs_ssse3() {
 360   _mm_abs_epi16(m1);
 361 }
 362
 363 // CHECK-LABEL: @test_abs_ssse3
 364
 365 // CHECK-LABEL: define available_externally <2 x i64> @_mm_abs_epi16(<2 x i64> noundef %{{[0-9a-zA-Z_.]+}})