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