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

   1 // REQUIRES: powerpc-registered-target
   2
   3 // RUN: %clang -S -emit-llvm -target powerpc64-gnu-linux -mcpu=pwr8 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \
   4 // RUN:   -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-gnu-linux -mcpu=pwr8 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \
   6 // RUN:   -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 powerpc64-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 --check-prefixes=CHECK,CHECK-BE
  10 // RUN: %clang -S -emit-llvm -target powerpc64le-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 --check-prefixes=CHECK,CHECK-LE
  12
  13 // RUN: %clang -S -emit-llvm -target powerpc64-ibm-aix -mcpu=pwr8 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \
  14 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -o - | llvm-cxxfilt -n | FileCheck %s --check-prefixes=CHECK,CHECK-BE
  15
  16 // RUN: %clang -x c++ -S -emit-llvm -target powerpc64le-gnu-linux -mcpu=pwr8 -ffreestanding -DNO_WARN_X86_INTRINSICS %s \
  17 // RUN:   -fno-discard-value-names -mllvm -disable-llvm-optzns -fsyntax-only
  18
  19 #include <tmmintrin.h>
  20
  21 __m64 res, m1, m2;
  22 __m128i resi, mi1, mi2;
  23
  24 void __attribute__((noinline))
  25 test_abs() {
  26   resi = _mm_abs_epi16(mi1);
  27   resi = _mm_abs_epi32(mi1);
  28   resi = _mm_abs_epi8(mi1);
  29   res = _mm_abs_pi16(m1);
  30   res = _mm_abs_pi32(m1);
  31   res = _mm_abs_pi8(m1);
  32 }
  33
  34 // CHECK-LABEL: @test_abs
  35
  36 // CHECK-LABEL: define available_externally <2 x i64> @_mm_abs_epi16
  37 // CHECK: call <8 x i16> @vec_abs(short vector[8])
  38
  39 // CHECK-LABEL: define available_externally <2 x i64> @_mm_abs_epi32
  40 // CHECK: call <4 x i32> @vec_abs(int vector[4])
  41
  42 // CHECK-LABEL: define available_externally <2 x i64> @_mm_abs_epi8
  43 // CHECK: call <16 x i8> @vec_abs(signed char vector[16])
  44
  45 // CHECK-LABEL: define available_externally i64 @_mm_abs_pi16
  46 // CHECK: %[[ABS:[0-9a-zA-Z_.]+]] = call <8 x i16> @vec_abs(short vector[8])
  47 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <8 x i16> %[[ABS]] to <2 x i64>
  48 // CHECK: extractelement <2 x i64> %[[CAST]], i32 0
  49
  50 // CHECK-LABEL: define available_externally i64 @_mm_abs_pi32
  51 // CHECK: %[[ABS:[0-9a-zA-Z_.]+]] = call <4 x i32> @vec_abs(int vector[4])
  52 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <4 x i32> %[[ABS]] to <2 x i64>
  53 // CHECK: extractelement <2 x i64> %[[CAST]], i32 0
  54
  55 // CHECK-LABEL: define available_externally i64 @_mm_abs_pi8
  56 // CHECK: %[[ABS:[0-9a-zA-Z_.]+]] = call <16 x i8> @vec_abs(signed char vector[16])
  57 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <16 x i8> %[[ABS]] to <2 x i64>
  58 // CHECK: extractelement <2 x i64> %[[CAST]], i32 0
  59
  60 void __attribute__((noinline))
  61 test_alignr() {
  62   resi = _mm_alignr_epi8(mi1, mi2, 1U);
  63   res = _mm_alignr_pi8(m1, m2, 1U);
  64 }
  65
  66 // CHECK-LABEL: @test_alignr
  67
  68 // CHECK-LABEL: define available_externally <2 x i64> @_mm_alignr_epi8
  69 // CHECK: %[[CONST:[0-9a-zA-Z_.]+]] = call i1 @llvm.is.constant.i32(i32 %0)
  70 // CHECK: br i1 %[[CONST]]
  71 // CHECK-BE: call <16 x i8> @vec_sld(unsigned char vector[16], unsigned char vector[16], unsigned int)
  72 // CHECK-LE: call <16 x i8> @vec_reve(unsigned char vector[16])
  73 // CHECK-LE: call <16 x i8> @vec_reve(unsigned char vector[16])
  74 // CHECk-LE: call <16 x i8> @vec_sld(unsigned char vector[16], unsigned char vector[16], unsigned int)
  75 // CHECK-LE: call <16 x i8> @vec_reve(unsigned char vector[16])
  76 // CHECK: store <16 x i8> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
  77 // CHECK: store <2 x i64> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
  78 // CHECK: %[[SUB:[0-9a-zA-Z_.]+]] = sub i32 %{{[0-9a-zA-Z_.]+}}, 16
  79 // CHECK: %[[MUL:[0-9a-zA-Z_.]+]] = mul i32 %[[SUB]], 8
  80 // CHECK: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %[[MUL]] to i8
  81 // CHECK: call <16 x i8> @vec_splats(unsigned char)(i8 noundef zeroext %[[TRUNC]])
  82 // CHECK-BE: call <16 x i8> @vec_slo(unsigned char vector[16], unsigned char vector[16])
  83 // CHECK-LE: call <16 x i8> @vec_sro(unsigned char vector[16], unsigned char vector[16])
  84 // CHECK: %[[SUB2:[0-9a-zA-Z_.]+]] = sub i32 16, %{{[0-9a-zA-Z_.]+}}
  85 // CHECK: %[[MUL2:[0-9a-zA-Z_.]+]] = mul i32 %[[SUB2]], 8
  86 // CHECK-BE: %[[TRUNC:[0-9a-zA-Z_.]+]] = trunc i32 %[[MUL2]] to i8
  87 // CHECK-BE: call <16 x i8> @vec_splats(unsigned char)(i8 noundef zeroext %[[TRUNC]])
  88 // CHECK-BE: mul i32 %{{[0-9a-zA-Z_.]+}}, 8
  89 // CHECK-BE: call <16 x i8> @vec_sro(unsigned char vector[16], unsigned char vector[16])
  90 // CHECK-BE: call <16 x i8> @vec_slo(unsigned char vector[16], unsigned char vector[16])
  91 // CHECK-BE: call <16 x i8> @vec_or(unsigned char vector[16], unsigned char vector[16])
  92 // CHECK-LE: %[[MUL3:[0-9a-zA-Z_.]+]] = mul i32 %{{[0-9a-zA-Z_.]+}}, 8
  93 // CHECK-LE: trunc i32 %[[MUL3]] to i8
  94 // CHECK-LE: call <16 x i8> @vec_splats(unsigned char)
  95 // CHECK-LE: call <16 x i8> @vec_slo(unsigned char vector[16], unsigned char vector[16])
  96 // CHECK-LE: call <16 x i8> @vec_sro(unsigned char vector[16], unsigned char vector[16])
  97 // CHECK-LE: call <16 x i8> @vec_or(unsigned char vector[16], unsigned char vector[16])
  98
  99 // CHECK-LABEL: define available_externally i64 @_mm_alignr_pi8
 100 // CHECK: %[[CMP:[0-9a-zA-Z_.]+]] = icmp ult i32 %{{[0-9a-zA-Z_.]+}}, 16
 101 // CHECK: br i1 %[[CMP]]
 102 // CHECK-BE: call <16 x i8> @vec_slo(unsigned char vector[16], unsigned char vector[16])
 103 // CHECK-LE: call <16 x i8> @vec_sro(unsigned char vector[16], unsigned char vector[16])
 104 // CHECK: extractelement <2 x i64> %{{[0-9a-zA-Z_.]+}}, i32 0
 105 // CHECK: store i64 0, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 106 // CHECK: store i64 0, ptr %{{[0-9a-zA-Z_.]+}}, align 8
 107
 108 void __attribute__((noinline))
 109 test_hadd() {
 110   resi = _mm_hadd_epi16(mi1, mi2);
 111   resi = _mm_hadd_epi32(mi1, mi2);
 112   res = _mm_hadd_pi16(m1, m2);
 113   res = _mm_hadd_pi32(m1, m2);
 114   resi = _mm_hadds_epi16(mi1, mi2);
 115   res = _mm_hadds_pi16(m1, m2);
 116 }
 117
 118 // CHECK-LABEL: @test_hadd
 119
 120 // CHECK-LABEL: define available_externally <2 x i64> @_mm_hadd_epi16
 121 // CHECK: store <16 x i8> <i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13, i8 16, i8 17, i8 20, i8 21, i8 24, i8 25, i8 28, i8 29>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 122 // CHECK: store <16 x i8> <i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15, i8 18, i8 19, i8 22, i8 23, i8 26, i8 27, i8 30, i8 31>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 123 // CHECK: call <8 x i16> @vec_perm(short vector[8], short vector[8], unsigned char vector[16])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13, i8 16, i8 17, i8 20, i8 21, i8 24, i8 25, i8 28, i8 29>)
 124 // CHECK: call <8 x i16> @vec_perm(short vector[8], short vector[8], unsigned char vector[16])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15, i8 18, i8 19, i8 22, i8 23, i8 26, i8 27, i8 30, i8 31>)
 125 // CHECK: call <8 x i16> @vec_add(short vector[8], short vector[8])
 126
 127 // CHECK-LABEL: define available_externally <2 x i64> @_mm_hadd_epi32
 128 // CHECK: store <16 x i8> <i8 0, i8 1, i8 2, i8 3, i8 8, i8 9, i8 10, i8 11, i8 16, i8 17, i8 18, i8 19, i8 24, i8 25, i8 26, i8 27>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 129 // CHECK: store <16 x i8> <i8 4, i8 5, i8 6, i8 7, i8 12, i8 13, i8 14, i8 15, i8 20, i8 21, i8 22, i8 23, i8 28, i8 29, i8 30, i8 31>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 130 // CHECK: call <4 x i32> @vec_perm(int vector[4], int vector[4], unsigned char vector[16])(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 0, i8 1, i8 2, i8 3, i8 8, i8 9, i8 10, i8 11, i8 16, i8 17, i8 18, i8 19, i8 24, i8 25, i8 26, i8 27>)
 131 // CHECK: call <4 x i32> @vec_perm(int vector[4], int vector[4], unsigned char vector[16])(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 4, i8 5, i8 6, i8 7, i8 12, i8 13, i8 14, i8 15, i8 20, i8 21, i8 22, i8 23, i8 28, i8 29, i8 30, i8 31>)
 132 // CHECK: call <4 x i32> @vec_add(int vector[4], int vector[4])
 133
 134 // CHECK-LABEL: define available_externally i64 @_mm_hadd_pi16
 135 // CHECK: store <16 x i8> <i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13, i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 136 // CHECK: store <16 x i8> <i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15, i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 137 // CHECK: call <8 x i16> @vec_perm(short vector[8], short vector[8], unsigned char vector[16])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15, i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15>)
 138 // CHECK: call <8 x i16> @vec_perm(short vector[8], short vector[8], unsigned char vector[16])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13, i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13>)
 139 // CHECK: call <8 x i16> @vec_add(short vector[8], short vector[8])
 140 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <8 x i16> %{{[0-9a-zA-Z_.]+}} to <2 x i64>
 141 // CHECK: extractelement <2 x i64> %[[CAST]], i32 1
 142
 143 // CHECK-LABEL: define available_externally i64 @_mm_hadd_pi32
 144 // CHECK: store <16 x i8> <i8 0, i8 1, i8 2, i8 3, i8 8, i8 9, i8 10, i8 11, i8 0, i8 1, i8 2, i8 3, i8 8, i8 9, i8 10, i8 11>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 145 // CHECK: store <16 x i8> <i8 4, i8 5, i8 6, i8 7, i8 12, i8 13, i8 14, i8 15, i8 4, i8 5, i8 6, i8 7, i8 12, i8 13, i8 14, i8 15>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 146 // CHECK: call <4 x i32> @vec_perm(int vector[4], int vector[4], unsigned char vector[16])(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 4, i8 5, i8 6, i8 7, i8 12, i8 13, i8 14, i8 15, i8 4, i8 5, i8 6, i8 7, i8 12, i8 13, i8 14, i8 15>)
 147 // CHECK: call <4 x i32> @vec_perm(int vector[4], int vector[4], unsigned char vector[16])(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 0, i8 1, i8 2, i8 3, i8 8, i8 9, i8 10, i8 11, i8 0, i8 1, i8 2, i8 3, i8 8, i8 9, i8 10, i8 11>)
 148 // CHECK: call <4 x i32> @vec_add(int vector[4], int vector[4])
 149 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <4 x i32> %{{[0-9a-zA-Z_.]+}} to <2 x i64>
 150 // CHECK: extractelement <2 x i64> %[[CAST]], i32 1
 151
 152 // CHECK-LABEL: define available_externally <2 x i64> @_mm_hadds_epi16
 153 // CHECK: store <4 x i32> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 154 // CHECK: store <4 x i32> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 155 // CHECK: call <4 x i32> @vec_sum4s(short vector[8], int vector[4])
 156 // CHECK: call <4 x i32> @vec_sum4s(short vector[8], int vector[4])
 157 // CHECK: call <8 x i16> @vec_packs(int vector[4], int vector[4])
 158
 159 // CHECK-LABEL: define available_externally i64 @_mm_hadds_pi16
 160 // CHECK: call <4 x i32> @vec_sum4s(short vector[8], int vector[4])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <4 x i32> noundef zeroinitializer)
 161 // CHECK: call <8 x i16> @vec_packs(int vector[4], int vector[4])
 162 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <8 x i16> %{{[0-9a-zA-Z_.]+}} to <2 x i64>
 163 // CHECK: extractelement <2 x i64> %[[CAST]], i32 1
 164
 165 void __attribute__((noinline))
 166 test_hsub() {
 167   resi = _mm_hsub_epi16(mi1, mi2);
 168   resi = _mm_hsub_epi32(mi1, mi2);
 169   res = _mm_hsub_pi16(m1, m2);
 170   res = _mm_hsub_pi32(m1, m2);
 171   resi = _mm_hsubs_epi16(mi1, mi2);
 172   res = _mm_hsubs_pi16(m1, m2);
 173 }
 174
 175 // CHECK-LABEL: @test_hsub
 176
 177 // CHECK-LABEL: define available_externally <2 x i64> @_mm_hsub_epi16
 178 // CHECK: store <16 x i8> <i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13, i8 16, i8 17, i8 20, i8 21, i8 24, i8 25, i8 28, i8 29>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 179 // CHECK: store <16 x i8> <i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15, i8 18, i8 19, i8 22, i8 23, i8 26, i8 27, i8 30, i8 31>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 180 // CHECK: call <8 x i16> @vec_perm(short vector[8], short vector[8], unsigned char vector[16])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13, i8 16, i8 17, i8 20, i8 21, i8 24, i8 25, i8 28, i8 29>)
 181 // CHECK: call <8 x i16> @vec_perm(short vector[8], short vector[8], unsigned char vector[16])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15, i8 18, i8 19, i8 22, i8 23, i8 26, i8 27, i8 30, i8 31>)
 182 // CHECK: call <8 x i16> @vec_sub(short vector[8], short vector[8])
 183
 184 // CHECK-LABEL: define available_externally <2 x i64> @_mm_hsub_epi32
 185 // CHECK: store <16 x i8> <i8 0, i8 1, i8 2, i8 3, i8 8, i8 9, i8 10, i8 11, i8 16, i8 17, i8 18, i8 19, i8 24, i8 25, i8 26, i8 27>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 186 // CHECK: store <16 x i8> <i8 4, i8 5, i8 6, i8 7, i8 12, i8 13, i8 14, i8 15, i8 20, i8 21, i8 22, i8 23, i8 28, i8 29, i8 30, i8 31>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 187 // CHECK: call <4 x i32> @vec_perm(int vector[4], int vector[4], unsigned char vector[16])(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 0, i8 1, i8 2, i8 3, i8 8, i8 9, i8 10, i8 11, i8 16, i8 17, i8 18, i8 19, i8 24, i8 25, i8 26, i8 27>)
 188 // CHECK: call <4 x i32> @vec_perm(int vector[4], int vector[4], unsigned char vector[16])(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 4, i8 5, i8 6, i8 7, i8 12, i8 13, i8 14, i8 15, i8 20, i8 21, i8 22, i8 23, i8 28, i8 29, i8 30, i8 31>)
 189 // CHECK: call <4 x i32> @vec_sub(int vector[4], int vector[4])
 190
 191 // CHECK-LABEL: define available_externally i64 @_mm_hsub_pi16
 192 // CHECK: store <16 x i8> <i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13, i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 193 // CHECK: store <16 x i8> <i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15, i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 194 // CHECK: call <8 x i16> @vec_perm(short vector[8], short vector[8], unsigned char vector[16])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15, i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15>)
 195 // CHECK: call <8 x i16> @vec_perm(short vector[8], short vector[8], unsigned char vector[16])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13, i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13>)
 196 // CHECK: call <8 x i16> @vec_sub(short vector[8], short vector[8])
 197 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <8 x i16> %{{[0-9a-zA-Z_.]+}} to <2 x i64>
 198 // CHECK: extractelement <2 x i64> %[[CAST]], i32 1
 199
 200 // CHECK-LABEL: define available_externally i64 @_mm_hsub_pi32
 201 // CHECK: store <16 x i8> <i8 0, i8 1, i8 2, i8 3, i8 8, i8 9, i8 10, i8 11, i8 0, i8 1, i8 2, i8 3, i8 8, i8 9, i8 10, i8 11>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 202 // CHECK: store <16 x i8> <i8 4, i8 5, i8 6, i8 7, i8 12, i8 13, i8 14, i8 15, i8 4, i8 5, i8 6, i8 7, i8 12, i8 13, i8 14, i8 15>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 203 // CHECK: call <4 x i32> @vec_perm(int vector[4], int vector[4], unsigned char vector[16])(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 4, i8 5, i8 6, i8 7, i8 12, i8 13, i8 14, i8 15, i8 4, i8 5, i8 6, i8 7, i8 12, i8 13, i8 14, i8 15>)
 204 // CHECK: call <4 x i32> @vec_perm(int vector[4], int vector[4], unsigned char vector[16])(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 0, i8 1, i8 2, i8 3, i8 8, i8 9, i8 10, i8 11, i8 0, i8 1, i8 2, i8 3, i8 8, i8 9, i8 10, i8 11>)
 205 // CHECK: call <4 x i32> @vec_sub(int vector[4], int vector[4])
 206 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <4 x i32> %{{[0-9a-zA-Z_.]+}} to <2 x i64>
 207 // CHECK: extractelement <2 x i64> %[[CAST]], i32 1
 208
 209 // CHECK-LABEL: define available_externally <2 x i64> @_mm_hsubs_epi16
 210 // CHECK: store <16 x i8> <i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13, i8 16, i8 17, i8 20, i8 21, i8 24, i8 25, i8 28, i8 29>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 211 // CHECK: store <16 x i8> <i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15, i8 18, i8 19, i8 22, i8 23, i8 26, i8 27, i8 30, i8 31>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 212 // CHECK: call <8 x i16> @vec_perm(short vector[8], short vector[8], unsigned char vector[16])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13, i8 16, i8 17, i8 20, i8 21, i8 24, i8 25, i8 28, i8 29>)
 213 // CHECK: call <8 x i16> @vec_perm(short vector[8], short vector[8], unsigned char vector[16])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15, i8 18, i8 19, i8 22, i8 23, i8 26, i8 27, i8 30, i8 31>)
 214 // CHECK: call <8 x i16> @vec_subs(short vector[8], short vector[8])
 215
 216 // CHECK-LABEL: define available_externally i64 @_mm_hsubs_pi16
 217 // CHECK: store <16 x i8> <i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13, i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 218 // CHECK: store <16 x i8> <i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15, i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 219 // CHECK: call <8 x i16> @vec_perm(short vector[8], short vector[8], unsigned char vector[16])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13, i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13>)
 220 // CHECK: call <8 x i16> @vec_perm(short vector[8], short vector[8], unsigned char vector[16])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15, i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15>)
 221 // CHECK: call <8 x i16> @vec_subs(short vector[8], short vector[8])
 222 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <8 x i16> %{{[0-9a-zA-Z_.]+}} to <2 x i64>
 223 // CHECK: extractelement <2 x i64> %[[CAST]], i32 1
 224
 225 void __attribute__((noinline))
 226 test_shuffle() {
 227   resi = _mm_shuffle_epi8(mi1, mi2);
 228   res = _mm_shuffle_pi8(m1, m2);
 229 }
 230
 231 // CHECK-LABEL: @test_shuffle
 232
 233 // CHECK-LABEL: define available_externally <2 x i64> @_mm_shuffle_epi8
 234 // CHECK: store <16 x i8> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 235 // CHECK: call <16 x i8> @vec_cmplt(signed char vector[16], signed char vector[16])(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef zeroinitializer)
 236 // CHECK: call <16 x i8> @vec_perm(signed char vector[16], signed char vector[16], unsigned char vector[16])
 237 // CHECK: call <16 x i8> @vec_sel(signed char vector[16], signed char vector[16], bool vector[16])(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef zeroinitializer, <16 x i8> noundef %{{[0-9a-zA-Z_.]+}})
 238
 239 // CHECK-LABEL: define available_externally i64 @_mm_shuffle_pi8
 240 // CHECK: call <16 x i8> @vec_cmplt(signed char vector[16], signed char vector[16])(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef zeroinitializer)
 241 // CHECK: call <16 x i8> @vec_perm(signed char vector[16], signed char vector[16], unsigned char vector[16])
 242 // CHECK: call <16 x i8> @vec_sel(signed char vector[16], signed char vector[16], bool vector[16])(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef zeroinitializer, <16 x i8> noundef %{{[0-9a-zA-Z_.]+}})
 243 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <16 x i8> %{{[0-9a-zA-Z_.]+}} to <2 x i64>
 244 // CHECK: extractelement <2 x i64> %[[CAST]], i32 0
 245
 246 void __attribute__((noinline))
 247 test_sign() {
 248   resi = _mm_sign_epi8(mi1, mi2);
 249   resi = _mm_sign_epi16(mi1, mi2);
 250   resi = _mm_sign_epi32(mi1, mi2);
 251   res = _mm_sign_pi8(m1, m2);
 252   res = _mm_sign_pi16(m1, m2);
 253   res = _mm_sign_pi32(m1, m2);
 254 }
 255
 256 // CHECK-LABEL: @test_sign
 257
 258 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sign_epi8
 259 // CHECK: call <16 x i8> @vec_cmplt(signed char vector[16], signed char vector[16])(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef zeroinitializer)
 260 // CHECK: call <16 x i8> @vec_cmpgt(signed char vector[16], signed char vector[16])(<16 x i8> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef zeroinitializer)
 261 // CHECK: call <16 x i8> @vec_neg(signed char vector[16])
 262 // CHECK: call <16 x i8> @vec_add(signed char vector[16], signed char vector[16])
 263 // CHECK: call <16 x i8> @vec_mul(signed char vector[16], signed char vector[16])
 264
 265 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sign_epi16
 266 // CHECK: call <8 x i16> @vec_cmplt(short vector[8], short vector[8])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef zeroinitializer)
 267 // CHECK: call <8 x i16> @vec_cmpgt(short vector[8], short vector[8])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef zeroinitializer)
 268 // CHECK: call <8 x i16> @vec_neg(short vector[8])
 269 // CHECK: call <8 x i16> @vec_add(short vector[8], short vector[8])
 270 // CHECK: call <8 x i16> @vec_mul(short vector[8], short vector[8])
 271
 272 // CHECK-LABEL: define available_externally <2 x i64> @_mm_sign_epi32
 273 // CHECK: store <4 x i32> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 274 // CHECK: call <4 x i32> @vec_cmplt(int vector[4], int vector[4])(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <4 x i32> noundef zeroinitializer)
 275 // CHECK: call <4 x i32> @vec_cmpgt(int vector[4], int vector[4])(<4 x i32> noundef %{{[0-9a-zA-Z_.]+}}, <4 x i32> noundef zeroinitializer)
 276 // CHECK: call <4 x i32> @vec_neg(int vector[4])
 277 // CHECK: call <4 x i32> @vec_add(int vector[4], int vector[4])
 278 // CHECK: call <4 x i32> @vec_mul(int vector[4], int vector[4])
 279
 280 // CHECK-LABEL: define available_externally i64 @_mm_sign_pi8
 281 // CHECK: store <16 x i8> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 282 // CHECK: call <2 x i64> @_mm_sign_epi8
 283 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <16 x i8> %{{[0-9a-zA-Z_.]+}} to <2 x i64>
 284 // CHECK: extractelement <2 x i64> %[[CAST]], i32 0
 285
 286 // CHECK-LABEL: define available_externally i64 @_mm_sign_pi16
 287 // CHECK: store <8 x i16> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 288 // CHECK: call <2 x i64> @_mm_sign_epi16
 289 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <8 x i16> %{{[0-9a-zA-Z_.]+}} to <2 x i64>
 290 // CHECK: extractelement <2 x i64> %[[CAST]], i32 0
 291
 292 // CHECK-LABEL: define available_externally i64 @_mm_sign_pi32
 293 // CHECK: store <4 x i32> zeroinitializer, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 294 // CHECK: call <2 x i64> @_mm_sign_epi32
 295 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <4 x i32> %{{[0-9a-zA-Z_.]+}} to <2 x i64>
 296 // CHECK: extractelement <2 x i64> %[[CAST]], i32 0
 297
 298 void __attribute__((noinline))
 299 test_maddubs() {
 300   resi = _mm_maddubs_epi16(mi1, mi2);
 301   res = _mm_maddubs_pi16(m1, m2);
 302 }
 303
 304 // CHECK-LABEL: @test_maddubs
 305
 306 // CHECK-LABEL: define available_externally <2 x i64> @_mm_maddubs_epi16
 307 // CHECK: call <8 x i16> @vec_splats(short)(i16 noundef signext 255)
 308 // CHECK: call <8 x i16> @vec_unpackh(signed char vector[16])
 309 // CHECK: call <8 x i16> @vec_and(short vector[8], short vector[8])
 310 // CHECK: call <8 x i16> @vec_unpackl(signed char vector[16])
 311 // CHECK: call <8 x i16> @vec_and(short vector[8], short vector[8])
 312 // CHECK: call <8 x i16> @vec_unpackh(signed char vector[16])
 313 // CHECK: call <8 x i16> @vec_unpackl(signed char vector[16])
 314 // CHECK: call <8 x i16> @vec_mul(short vector[8], short vector[8])
 315 // CHECK: call <8 x i16> @vec_mul(short vector[8], short vector[8])
 316 // CHECK: store <16 x i8> <i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13, i8 16, i8 17, i8 20, i8 21, i8 24, i8 25, i8 28, i8 29>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 317 // CHECK: store <16 x i8> <i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15, i8 18, i8 19, i8 22, i8 23, i8 26, i8 27, i8 30, i8 31>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 318 // CHECK: call <8 x i16> @vec_perm(short vector[8], short vector[8], unsigned char vector[16])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13, i8 16, i8 17, i8 20, i8 21, i8 24, i8 25, i8 28, i8 29>)
 319 // CHECK: call <8 x i16> @vec_perm(short vector[8], short vector[8], unsigned char vector[16])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15, i8 18, i8 19, i8 22, i8 23, i8 26, i8 27, i8 30, i8 31>)
 320 // CHECK: call <8 x i16> @vec_adds(short vector[8], short vector[8])
 321
 322 // CHECK-LABEL: define available_externally i64 @_mm_maddubs_pi16
 323 // CHECK: call <8 x i16> @vec_unpackl(signed char vector[16])
 324 // CHECK: call <8 x i16> @vec_splats(short)(i16 noundef signext 255)
 325 // CHECK: call <8 x i16> @vec_and(short vector[8], short vector[8])
 326 // CHECK: call <8 x i16> @vec_unpackl(signed char vector[16])
 327 // CHECK: call <8 x i16> @vec_mul(short vector[8], short vector[8])
 328 // CHECK: store <16 x i8> <i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13, i8 16, i8 17, i8 20, i8 21, i8 24, i8 25, i8 28, i8 29>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 329 // CHECK: store <16 x i8> <i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15, i8 18, i8 19, i8 22, i8 23, i8 26, i8 27, i8 30, i8 31>, ptr %{{[0-9a-zA-Z_.]+}}, align 16
 330 // CHECK: call <8 x i16> @vec_perm(short vector[8], short vector[8], unsigned char vector[16])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 0, i8 1, i8 4, i8 5, i8 8, i8 9, i8 12, i8 13, i8 16, i8 17, i8 20, i8 21, i8 24, i8 25, i8 28, i8 29>)
 331 // CHECK: call <8 x i16> @vec_perm(short vector[8], short vector[8], unsigned char vector[16])(<8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <8 x i16> noundef %{{[0-9a-zA-Z_.]+}}, <16 x i8> noundef <i8 2, i8 3, i8 6, i8 7, i8 10, i8 11, i8 14, i8 15, i8 18, i8 19, i8 22, i8 23, i8 26, i8 27, i8 30, i8 31>)
 332 // CHECK: call <8 x i16> @vec_adds(short vector[8], short vector[8])
 333 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <8 x i16> %{{[0-9a-zA-Z_.]+}} to <2 x i64>
 334 // CHECK: extractelement <2 x i64> %[[CAST]], i32 0
 335
 336 void __attribute__((noinline))
 337 test_mulhrs() {
 338   resi = _mm_mulhrs_epi16(mi1, mi2);
 339   res = _mm_mulhrs_pi16(m1, m2);
 340 }
 341
 342 // CHECK-LABEL: @test_mulhrs
 343
 344 // CHECK-LABEL: define available_externally <2 x i64> @_mm_mulhrs_epi16
 345 // CHECK: call <4 x i32> @vec_unpackh(short vector[8])
 346 // CHECK: call <4 x i32> @vec_unpackh(short vector[8])
 347 // CHECK: call <4 x i32> @vec_mul(int vector[4], int vector[4])
 348 // CHECK: call <4 x i32> @vec_unpackl(short vector[8])
 349 // CHECK: call <4 x i32> @vec_unpackl(short vector[8])
 350 // CHECK: call <4 x i32> @vec_mul(int vector[4], int vector[4])
 351 // CHECK: call <4 x i32> @vec_splats(unsigned int)(i32 noundef zeroext 14)
 352 // CHECK: call <4 x i32> @vec_sr(int vector[4], unsigned int vector[4])
 353 // CHECK: call <4 x i32> @vec_sr(int vector[4], unsigned int vector[4])
 354 // CHECK: call <4 x i32> @vec_splats(int)(i32 noundef signext 1)
 355 // CHECK: call <4 x i32> @vec_add(int vector[4], int vector[4])
 356 // CHECK: call <4 x i32> @vec_sr(int vector[4], unsigned int vector[4])
 357 // CHECK: call <4 x i32> @vec_add(int vector[4], int vector[4])
 358 // CHECK: call <4 x i32> @vec_sr(int vector[4], unsigned int vector[4])
 359 // CHECK: %[[PACK:[0-9a-zA-Z_.]+]] = call <8 x i16> @vec_pack(int vector[4], int vector[4])
 360
 361 // CHECK-LABEL: define available_externally i64 @_mm_mulhrs_pi16
 362 // CHECK: call <4 x i32> @vec_unpackh(short vector[8])
 363 // CHECK: call <4 x i32> @vec_unpackh(short vector[8])
 364 // CHECK: call <4 x i32> @vec_mul(int vector[4], int vector[4])
 365 // CHECK: call <4 x i32> @vec_splats(unsigned int)(i32 noundef zeroext 14)
 366 // CHECK: call <4 x i32> @vec_sr(int vector[4], unsigned int vector[4])
 367 // CHECK: call <4 x i32> @vec_splats(int)(i32 noundef signext 1)
 368 // CHECK: call <4 x i32> @vec_add(int vector[4], int vector[4])
 369 // CHECK: call <4 x i32> @vec_sr(int vector[4], unsigned int vector[4])
 370 // CHECK: call <8 x i16> @vec_pack(int vector[4], int vector[4])
 371 // CHECK: %[[CAST:[0-9a-zA-Z_.]+]] = bitcast <8 x i16> %{{[0-9a-zA-Z_.]+}} to <2 x i64>
 372 // CHECK: extractelement <2 x i64> %[[CAST]], i32 0