test/CodeGen/X86/bswap-vector.ll

   1 ; RUN: llc < %s -mcpu=x86-64 | FileCheck %s --check-prefix=CHECK-NOSSSE3
   2 ; RUN: llc < %s -mcpu=core2 | FileCheck %s --check-prefix=CHECK-SSSE3
   3 ; RUN: llc < %s -mcpu=core-avx2 | FileCheck %s --check-prefix=CHECK-AVX2
   4 ; RUN: llc < %s -mcpu=core-avx2 -x86-experimental-vector-widening-legalization | FileCheck %s --check-prefix=CHECK-WIDE-AVX2
   5
   6 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
   7 target triple = "x86_64-unknown-linux-gnu"
   8
   9 declare <8 x i16> @llvm.bswap.v8i16(<8 x i16>)
  10 declare <4 x i32> @llvm.bswap.v4i32(<4 x i32>)
  11 declare <2 x i64> @llvm.bswap.v2i64(<2 x i64>)
  12
  13 define <8 x i16> @test1(<8 x i16> %v) {
  14 ; CHECK-NOSSSE3-LABEL: test1:
  15 ; CHECK-NOSSSE3:       # BB#0: # %entry
  16 ; CHECK-NOSSSE3-NEXT:    pextrw $7, %xmm0, %eax
  17 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
  18 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm1
  19 ; CHECK-NOSSSE3-NEXT:    pextrw $3, %xmm0, %eax
  20 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
  21 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm2
  22 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1],xmm2[2],xmm1[2],xmm2[3],xmm1[3]
  23 ; CHECK-NOSSSE3-NEXT:    pextrw $5, %xmm0, %eax
  24 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
  25 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm3
  26 ; CHECK-NOSSSE3-NEXT:    pextrw $1, %xmm0, %eax
  27 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
  28 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm1
  29 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
  30 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
  31 ; CHECK-NOSSSE3-NEXT:    pextrw $6, %xmm0, %eax
  32 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
  33 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm2
  34 ; CHECK-NOSSSE3-NEXT:    pextrw $2, %xmm0, %eax
  35 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
  36 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm3
  37 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
  38 ; CHECK-NOSSSE3-NEXT:    pextrw $4, %xmm0, %eax
  39 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
  40 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm2
  41 ; CHECK-NOSSSE3-NEXT:    movd %xmm0, %eax
  42 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
  43 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm0
  44 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
  45 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
  46 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
  47 ; CHECK-NOSSSE3-NEXT:    retq
  48 ;
  49 ; CHECK-SSSE3-LABEL: test1:
  50 ; CHECK-SSSE3:       # BB#0: # %entry
  51 ; CHECK-SSSE3-NEXT:    pshufb {{.*#+}} xmm0 = xmm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
  52 ; CHECK-SSSE3-NEXT:    retq
  53 ;
  54 ; CHECK-AVX2-LABEL: test1:
  55 ; CHECK-AVX2:       # BB#0: # %entry
  56 ; CHECK-AVX2-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
  57 ; CHECK-AVX2-NEXT:    retq
  58 ;
  59 ; CHECK-WIDE-AVX2-LABEL: test1:
  60 ; CHECK-WIDE-AVX2:       # BB#0: # %entry
  61 ; CHECK-WIDE-AVX2-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
  62 ; CHECK-WIDE-AVX2-NEXT:    retq
  63 entry:
  64   %r = call <8 x i16> @llvm.bswap.v8i16(<8 x i16> %v)
  65   ret <8 x i16> %r
  66 }
  67
  68 define <4 x i32> @test2(<4 x i32> %v) {
  69 ; CHECK-NOSSSE3-LABEL: test2:
  70 ; CHECK-NOSSSE3:       # BB#0: # %entry
  71 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[3,1,2,3]
  72 ; CHECK-NOSSSE3-NEXT:    movd %xmm1, %eax
  73 ; CHECK-NOSSSE3-NEXT:    bswapl %eax
  74 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm1
  75 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,1,2,3]
  76 ; CHECK-NOSSSE3-NEXT:    movd %xmm2, %eax
  77 ; CHECK-NOSSSE3-NEXT:    bswapl %eax
  78 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm2
  79 ; CHECK-NOSSSE3-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
  80 ; CHECK-NOSSSE3-NEXT:    movd %xmm0, %eax
  81 ; CHECK-NOSSSE3-NEXT:    bswapl %eax
  82 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm1
  83 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
  84 ; CHECK-NOSSSE3-NEXT:    movd %xmm0, %eax
  85 ; CHECK-NOSSSE3-NEXT:    bswapl %eax
  86 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm0
  87 ; CHECK-NOSSSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
  88 ; CHECK-NOSSSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
  89 ; CHECK-NOSSSE3-NEXT:    movdqa %xmm1, %xmm0
  90 ; CHECK-NOSSSE3-NEXT:    retq
  91 ;
  92 ; CHECK-SSSE3-LABEL: test2:
  93 ; CHECK-SSSE3:       # BB#0: # %entry
  94 ; CHECK-SSSE3-NEXT:    pshufb {{.*#+}} xmm0 = xmm0[3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12]
  95 ; CHECK-SSSE3-NEXT:    retq
  96 ;
  97 ; CHECK-AVX2-LABEL: test2:
  98 ; CHECK-AVX2:       # BB#0: # %entry
  99 ; CHECK-AVX2-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12]
 100 ; CHECK-AVX2-NEXT:    retq
 101 ;
 102 ; CHECK-WIDE-AVX2-LABEL: test2:
 103 ; CHECK-WIDE-AVX2:       # BB#0: # %entry
 104 ; CHECK-WIDE-AVX2-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12]
 105 ; CHECK-WIDE-AVX2-NEXT:    retq
 106 entry:
 107   %r = call <4 x i32> @llvm.bswap.v4i32(<4 x i32> %v)
 108   ret <4 x i32> %r
 109 }
 110
 111 define <2 x i64> @test3(<2 x i64> %v) {
 112 ; CHECK-NOSSSE3-LABEL: test3:
 113 ; CHECK-NOSSSE3:       # BB#0: # %entry
 114 ; CHECK-NOSSSE3-NEXT:    movd %xmm0, %rax
 115 ; CHECK-NOSSSE3-NEXT:    bswapq %rax
 116 ; CHECK-NOSSSE3-NEXT:    movd %rax, %xmm1
 117 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
 118 ; CHECK-NOSSSE3-NEXT:    movd %xmm0, %rax
 119 ; CHECK-NOSSSE3-NEXT:    bswapq %rax
 120 ; CHECK-NOSSSE3-NEXT:    movd %rax, %xmm0
 121 ; CHECK-NOSSSE3-NEXT:    punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]
 122 ; CHECK-NOSSSE3-NEXT:    movdqa %xmm1, %xmm0
 123 ; CHECK-NOSSSE3-NEXT:    retq
 124 ;
 125 ; CHECK-SSSE3-LABEL: test3:
 126 ; CHECK-SSSE3:       # BB#0: # %entry
 127 ; CHECK-SSSE3-NEXT:    pshufb {{.*#+}} xmm0 = xmm0[7,6,5,4,3,2,1,0,15,14,13,12,11,10,9,8]
 128 ; CHECK-SSSE3-NEXT:    retq
 129 ;
 130 ; CHECK-AVX2-LABEL: test3:
 131 ; CHECK-AVX2:       # BB#0: # %entry
 132 ; CHECK-AVX2-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[7,6,5,4,3,2,1,0,15,14,13,12,11,10,9,8]
 133 ; CHECK-AVX2-NEXT:    retq
 134 ;
 135 ; CHECK-WIDE-AVX2-LABEL: test3:
 136 ; CHECK-WIDE-AVX2:       # BB#0: # %entry
 137 ; CHECK-WIDE-AVX2-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[7,6,5,4,3,2,1,0,15,14,13,12,11,10,9,8]
 138 ; CHECK-WIDE-AVX2-NEXT:    retq
 139 entry:
 140   %r = call <2 x i64> @llvm.bswap.v2i64(<2 x i64> %v)
 141   ret <2 x i64> %r
 142 }
 143
 144 declare <16 x i16> @llvm.bswap.v16i16(<16 x i16>)
 145 declare <8 x i32> @llvm.bswap.v8i32(<8 x i32>)
 146 declare <4 x i64> @llvm.bswap.v4i64(<4 x i64>)
 147
 148 define <16 x i16> @test4(<16 x i16> %v) {
 149 ; CHECK-NOSSSE3-LABEL: test4:
 150 ; CHECK-NOSSSE3:       # BB#0: # %entry
 151 ; CHECK-NOSSSE3-NEXT:    pextrw $7, %xmm0, %eax
 152 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
 153 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm2
 154 ; CHECK-NOSSSE3-NEXT:    pextrw $3, %xmm0, %eax
 155 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
 156 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm3
 157 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
 158 ; CHECK-NOSSSE3-NEXT:    pextrw $5, %xmm0, %eax
 159 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
 160 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm4
 161 ; CHECK-NOSSSE3-NEXT:    pextrw $1, %xmm0, %eax
 162 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
 163 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm2
 164 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
 165 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
 166 ; CHECK-NOSSSE3-NEXT:    pextrw $6, %xmm0, %eax
 167 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
 168 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm3
 169 ; CHECK-NOSSSE3-NEXT:    pextrw $2, %xmm0, %eax
 170 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
 171 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm4
 172 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
 173 ; CHECK-NOSSSE3-NEXT:    pextrw $4, %xmm0, %eax
 174 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
 175 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm3
 176 ; CHECK-NOSSSE3-NEXT:    movd %xmm0, %eax
 177 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
 178 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm0
 179 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1],xmm0[2],xmm3[2],xmm0[3],xmm3[3]
 180 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm4[0],xmm0[1],xmm4[1],xmm0[2],xmm4[2],xmm0[3],xmm4[3]
 181 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
 182 ; CHECK-NOSSSE3-NEXT:    pextrw $7, %xmm1, %eax
 183 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
 184 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm2
 185 ; CHECK-NOSSSE3-NEXT:    pextrw $3, %xmm1, %eax
 186 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
 187 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm3
 188 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1],xmm3[2],xmm2[2],xmm3[3],xmm2[3]
 189 ; CHECK-NOSSSE3-NEXT:    pextrw $5, %xmm1, %eax
 190 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
 191 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm4
 192 ; CHECK-NOSSSE3-NEXT:    pextrw $1, %xmm1, %eax
 193 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
 194 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm2
 195 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm4[0],xmm2[1],xmm4[1],xmm2[2],xmm4[2],xmm2[3],xmm4[3]
 196 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1],xmm2[2],xmm3[2],xmm2[3],xmm3[3]
 197 ; CHECK-NOSSSE3-NEXT:    pextrw $6, %xmm1, %eax
 198 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
 199 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm3
 200 ; CHECK-NOSSSE3-NEXT:    pextrw $2, %xmm1, %eax
 201 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
 202 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm4
 203 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm3[0],xmm4[1],xmm3[1],xmm4[2],xmm3[2],xmm4[3],xmm3[3]
 204 ; CHECK-NOSSSE3-NEXT:    pextrw $4, %xmm1, %eax
 205 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
 206 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm3
 207 ; CHECK-NOSSSE3-NEXT:    movd %xmm1, %eax
 208 ; CHECK-NOSSSE3-NEXT:    rolw $8, %ax
 209 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm1
 210 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1],xmm1[2],xmm3[2],xmm1[3],xmm3[3]
 211 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm4[0],xmm1[1],xmm4[1],xmm1[2],xmm4[2],xmm1[3],xmm4[3]
 212 ; CHECK-NOSSSE3-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3]
 213 ; CHECK-NOSSSE3-NEXT:    retq
 214 ;
 215 ; CHECK-SSSE3-LABEL: test4:
 216 ; CHECK-SSSE3:       # BB#0: # %entry
 217 ; CHECK-SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
 218 ; CHECK-SSSE3-NEXT:    pshufb %xmm2, %xmm0
 219 ; CHECK-SSSE3-NEXT:    pshufb %xmm2, %xmm1
 220 ; CHECK-SSSE3-NEXT:    retq
 221 ;
 222 ; CHECK-AVX2-LABEL: test4:
 223 ; CHECK-AVX2:       # BB#0: # %entry
 224 ; CHECK-AVX2-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14,17,16,19,18,21,20,23,22,25,24,27,26,29,28,31,30]
 225 ; CHECK-AVX2-NEXT:    retq
 226 ;
 227 ; CHECK-WIDE-AVX2-LABEL: test4:
 228 ; CHECK-WIDE-AVX2:       # BB#0: # %entry
 229 ; CHECK-WIDE-AVX2-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14,17,16,19,18,21,20,23,22,25,24,27,26,29,28,31,30]
 230 ; CHECK-WIDE-AVX2-NEXT:    retq
 231 entry:
 232   %r = call <16 x i16> @llvm.bswap.v16i16(<16 x i16> %v)
 233   ret <16 x i16> %r
 234 }
 235
 236 define <8 x i32> @test5(<8 x i32> %v) {
 237 ; CHECK-NOSSSE3-LABEL: test5:
 238 ; CHECK-NOSSSE3:       # BB#0: # %entry
 239 ; CHECK-NOSSSE3-NEXT:    movdqa %xmm0, %xmm2
 240 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[3,1,2,3]
 241 ; CHECK-NOSSSE3-NEXT:    movd %xmm0, %eax
 242 ; CHECK-NOSSSE3-NEXT:    bswapl %eax
 243 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm0
 244 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[1,1,2,3]
 245 ; CHECK-NOSSSE3-NEXT:    movd %xmm3, %eax
 246 ; CHECK-NOSSSE3-NEXT:    bswapl %eax
 247 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm3
 248 ; CHECK-NOSSSE3-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
 249 ; CHECK-NOSSSE3-NEXT:    movd %xmm2, %eax
 250 ; CHECK-NOSSSE3-NEXT:    bswapl %eax
 251 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm0
 252 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[2,3,0,1]
 253 ; CHECK-NOSSSE3-NEXT:    movd %xmm2, %eax
 254 ; CHECK-NOSSSE3-NEXT:    bswapl %eax
 255 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm2
 256 ; CHECK-NOSSSE3-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
 257 ; CHECK-NOSSSE3-NEXT:    punpckldq {{.*#+}} xmm0 = xmm0[0],xmm3[0],xmm0[1],xmm3[1]
 258 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[3,1,2,3]
 259 ; CHECK-NOSSSE3-NEXT:    movd %xmm2, %eax
 260 ; CHECK-NOSSSE3-NEXT:    bswapl %eax
 261 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm2
 262 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[1,1,2,3]
 263 ; CHECK-NOSSSE3-NEXT:    movd %xmm3, %eax
 264 ; CHECK-NOSSSE3-NEXT:    bswapl %eax
 265 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm3
 266 ; CHECK-NOSSSE3-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
 267 ; CHECK-NOSSSE3-NEXT:    movd %xmm1, %eax
 268 ; CHECK-NOSSSE3-NEXT:    bswapl %eax
 269 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm2
 270 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,3,0,1]
 271 ; CHECK-NOSSSE3-NEXT:    movd %xmm1, %eax
 272 ; CHECK-NOSSSE3-NEXT:    bswapl %eax
 273 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm1
 274 ; CHECK-NOSSSE3-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
 275 ; CHECK-NOSSSE3-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm3[0],xmm2[1],xmm3[1]
 276 ; CHECK-NOSSSE3-NEXT:    movdqa %xmm2, %xmm1
 277 ; CHECK-NOSSSE3-NEXT:    retq
 278 ;
 279 ; CHECK-SSSE3-LABEL: test5:
 280 ; CHECK-SSSE3:       # BB#0: # %entry
 281 ; CHECK-SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12]
 282 ; CHECK-SSSE3-NEXT:    pshufb %xmm2, %xmm0
 283 ; CHECK-SSSE3-NEXT:    pshufb %xmm2, %xmm1
 284 ; CHECK-SSSE3-NEXT:    retq
 285 ;
 286 ; CHECK-AVX2-LABEL: test5:
 287 ; CHECK-AVX2:       # BB#0: # %entry
 288 ; CHECK-AVX2-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12,19,18,17,16,23,22,21,20,27,26,25,24,31,30,29,28]
 289 ; CHECK-AVX2-NEXT:    retq
 290 ;
 291 ; CHECK-WIDE-AVX2-LABEL: test5:
 292 ; CHECK-WIDE-AVX2:       # BB#0: # %entry
 293 ; CHECK-WIDE-AVX2-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12,19,18,17,16,23,22,21,20,27,26,25,24,31,30,29,28]
 294 ; CHECK-WIDE-AVX2-NEXT:    retq
 295 entry:
 296   %r = call <8 x i32> @llvm.bswap.v8i32(<8 x i32> %v)
 297   ret <8 x i32> %r
 298 }
 299
 300 define <4 x i64> @test6(<4 x i64> %v) {
 301 ; CHECK-NOSSSE3-LABEL: test6:
 302 ; CHECK-NOSSSE3:       # BB#0: # %entry
 303 ; CHECK-NOSSSE3-NEXT:    movd %xmm0, %rax
 304 ; CHECK-NOSSSE3-NEXT:    bswapq %rax
 305 ; CHECK-NOSSSE3-NEXT:    movd %rax, %xmm2
 306 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
 307 ; CHECK-NOSSSE3-NEXT:    movd %xmm0, %rax
 308 ; CHECK-NOSSSE3-NEXT:    bswapq %rax
 309 ; CHECK-NOSSSE3-NEXT:    movd %rax, %xmm0
 310 ; CHECK-NOSSSE3-NEXT:    punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm0[0]
 311 ; CHECK-NOSSSE3-NEXT:    movd %xmm1, %rax
 312 ; CHECK-NOSSSE3-NEXT:    bswapq %rax
 313 ; CHECK-NOSSSE3-NEXT:    movd %rax, %xmm3
 314 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[2,3,0,1]
 315 ; CHECK-NOSSSE3-NEXT:    movd %xmm0, %rax
 316 ; CHECK-NOSSSE3-NEXT:    bswapq %rax
 317 ; CHECK-NOSSSE3-NEXT:    movd %rax, %xmm0
 318 ; CHECK-NOSSSE3-NEXT:    punpcklqdq {{.*#+}} xmm3 = xmm3[0],xmm0[0]
 319 ; CHECK-NOSSSE3-NEXT:    movdqa %xmm2, %xmm0
 320 ; CHECK-NOSSSE3-NEXT:    movdqa %xmm3, %xmm1
 321 ; CHECK-NOSSSE3-NEXT:    retq
 322 ;
 323 ; CHECK-SSSE3-LABEL: test6:
 324 ; CHECK-SSSE3:       # BB#0: # %entry
 325 ; CHECK-SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [7,6,5,4,3,2,1,0,15,14,13,12,11,10,9,8]
 326 ; CHECK-SSSE3-NEXT:    pshufb %xmm2, %xmm0
 327 ; CHECK-SSSE3-NEXT:    pshufb %xmm2, %xmm1
 328 ; CHECK-SSSE3-NEXT:    retq
 329 ;
 330 ; CHECK-AVX2-LABEL: test6:
 331 ; CHECK-AVX2:       # BB#0: # %entry
 332 ; CHECK-AVX2-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[7,6,5,4,3,2,1,0,15,14,13,12,11,10,9,8,23,22,21,20,19,18,17,16,31,30,29,28,27,26,25,24]
 333 ; CHECK-AVX2-NEXT:    retq
 334 ;
 335 ; CHECK-WIDE-AVX2-LABEL: test6:
 336 ; CHECK-WIDE-AVX2:       # BB#0: # %entry
 337 ; CHECK-WIDE-AVX2-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[7,6,5,4,3,2,1,0,15,14,13,12,11,10,9,8,23,22,21,20,19,18,17,16,31,30,29,28,27,26,25,24]
 338 ; CHECK-WIDE-AVX2-NEXT:    retq
 339 entry:
 340   %r = call <4 x i64> @llvm.bswap.v4i64(<4 x i64> %v)
 341   ret <4 x i64> %r
 342 }
 343
 344 declare <4 x i16> @llvm.bswap.v4i16(<4 x i16>)
 345
 346 define <4 x i16> @test7(<4 x i16> %v) {
 347 ; CHECK-NOSSSE3-LABEL: test7:
 348 ; CHECK-NOSSSE3:       # BB#0: # %entry
 349 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[3,1,2,3]
 350 ; CHECK-NOSSSE3-NEXT:    movd %xmm1, %eax
 351 ; CHECK-NOSSSE3-NEXT:    bswapl %eax
 352 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm1
 353 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,1,2,3]
 354 ; CHECK-NOSSSE3-NEXT:    movd %xmm2, %eax
 355 ; CHECK-NOSSSE3-NEXT:    bswapl %eax
 356 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm2
 357 ; CHECK-NOSSSE3-NEXT:    punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
 358 ; CHECK-NOSSSE3-NEXT:    movd %xmm0, %eax
 359 ; CHECK-NOSSSE3-NEXT:    bswapl %eax
 360 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm1
 361 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
 362 ; CHECK-NOSSSE3-NEXT:    movd %xmm0, %eax
 363 ; CHECK-NOSSSE3-NEXT:    bswapl %eax
 364 ; CHECK-NOSSSE3-NEXT:    movd %eax, %xmm0
 365 ; CHECK-NOSSSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
 366 ; CHECK-NOSSSE3-NEXT:    punpckldq {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1]
 367 ; CHECK-NOSSSE3-NEXT:    psrld $16, %xmm1
 368 ; CHECK-NOSSSE3-NEXT:    movdqa %xmm1, %xmm0
 369 ; CHECK-NOSSSE3-NEXT:    retq
 370 ;
 371 ; CHECK-SSSE3-LABEL: test7:
 372 ; CHECK-SSSE3:       # BB#0: # %entry
 373 ; CHECK-SSSE3-NEXT:    pshufb {{.*#+}} xmm0 = xmm0[3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12]
 374 ; CHECK-SSSE3-NEXT:    psrld $16, %xmm0
 375 ; CHECK-SSSE3-NEXT:    retq
 376 ;
 377 ; CHECK-AVX2-LABEL: test7:
 378 ; CHECK-AVX2:       # BB#0: # %entry
 379 ; CHECK-AVX2-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12]
 380 ; CHECK-AVX2-NEXT:    vpsrld $16, %xmm0, %xmm0
 381 ; CHECK-AVX2-NEXT:    retq
 382 ;
 383 ; CHECK-WIDE-AVX2-LABEL: test7:
 384 ; CHECK-WIDE-AVX2:       # BB#0: # %entry
 385 ; CHECK-WIDE-AVX2-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
 386 ; CHECK-WIDE-AVX2-NEXT:    retq
 387 entry:
 388   %r = call <4 x i16> @llvm.bswap.v4i16(<4 x i16> %v)
 389   ret <4 x i16> %r
 390 }