test/CodeGen/X86/bswap-vector.ll

   1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
   2 ; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu | FileCheck %s --check-prefix=CHECK-ALL --check-prefix=CHECK-SSE --check-prefix=CHECK-NOSSSE3
   3 ; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mattr=+ssse3 | FileCheck %s --check-prefix=CHECK-ALL --check-prefix=CHECK-SSE --check-prefix=CHECK-SSSE3
   4 ; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mattr=+avx2 | FileCheck %s --check-prefix=CHECK-ALL --check-prefix=CHECK-AVX --check-prefix=CHECK-AVX2
   5
   6 declare <8 x i16> @llvm.bswap.v8i16(<8 x i16>)
   7 declare <4 x i32> @llvm.bswap.v4i32(<4 x i32>)
   8 declare <2 x i64> @llvm.bswap.v2i64(<2 x i64>)
   9
  10 define <8 x i16> @test1(<8 x i16> %v) {
  11 ; CHECK-NOSSSE3-LABEL: test1:
  12 ; CHECK-NOSSSE3:       # %bb.0: # %entry
  13 ; CHECK-NOSSSE3-NEXT:    pxor %xmm1, %xmm1
  14 ; CHECK-NOSSSE3-NEXT:    movdqa %xmm0, %xmm2
  15 ; CHECK-NOSSSE3-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
  16 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[1,0,3,2,4,5,6,7]
  17 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,4,7,6]
  18 ; CHECK-NOSSSE3-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
  19 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[1,0,3,2,4,5,6,7]
  20 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,4,7,6]
  21 ; CHECK-NOSSSE3-NEXT:    packuswb %xmm2, %xmm0
  22 ; CHECK-NOSSSE3-NEXT:    retq
  23 ;
  24 ; CHECK-SSSE3-LABEL: test1:
  25 ; CHECK-SSSE3:       # %bb.0: # %entry
  26 ; CHECK-SSSE3-NEXT:    pshufb {{.*#+}} xmm0 = xmm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
  27 ; CHECK-SSSE3-NEXT:    retq
  28 ;
  29 ; CHECK-AVX-LABEL: test1:
  30 ; CHECK-AVX:       # %bb.0: # %entry
  31 ; CHECK-AVX-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
  32 ; CHECK-AVX-NEXT:    retq
  33 entry:
  34   %r = call <8 x i16> @llvm.bswap.v8i16(<8 x i16> %v)
  35   ret <8 x i16> %r
  36 }
  37
  38 define <4 x i32> @test2(<4 x i32> %v) {
  39 ; CHECK-NOSSSE3-LABEL: test2:
  40 ; CHECK-NOSSSE3:       # %bb.0:
  41 ; CHECK-NOSSSE3-NEXT:    pxor %xmm1, %xmm1
  42 ; CHECK-NOSSSE3-NEXT:    movdqa %xmm0, %xmm2
  43 ; CHECK-NOSSSE3-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
  44 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[3,2,1,0,4,5,6,7]
  45 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,7,6,5,4]
  46 ; CHECK-NOSSSE3-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
  47 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]
  48 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4]
  49 ; CHECK-NOSSSE3-NEXT:    packuswb %xmm2, %xmm0
  50 ; CHECK-NOSSSE3-NEXT:    retq
  51 ;
  52 ; CHECK-SSSE3-LABEL: test2:
  53 ; CHECK-SSSE3:       # %bb.0:
  54 ; CHECK-SSSE3-NEXT:    pshufb {{.*#+}} xmm0 = xmm0[3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12]
  55 ; CHECK-SSSE3-NEXT:    retq
  56 ;
  57 ; CHECK-AVX-LABEL: test2:
  58 ; CHECK-AVX:       # %bb.0:
  59 ; CHECK-AVX-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12]
  60 ; CHECK-AVX-NEXT:    retq
  61   %r = call <4 x i32> @llvm.bswap.v4i32(<4 x i32> %v)
  62   ret <4 x i32> %r
  63 }
  64
  65 define <4 x i32> @or_bswap(<4 x i32> %x, <4 x i32> %y, <4 x i32>* %p1, <4 x i32>* %p2) {
  66 ; CHECK-NOSSSE3-LABEL: or_bswap:
  67 ; CHECK-NOSSSE3:       # %bb.0:
  68 ; CHECK-NOSSSE3-NEXT:    por %xmm1, %xmm0
  69 ; CHECK-NOSSSE3-NEXT:    pxor %xmm1, %xmm1
  70 ; CHECK-NOSSSE3-NEXT:    movdqa %xmm0, %xmm2
  71 ; CHECK-NOSSSE3-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
  72 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[3,2,1,0,4,5,6,7]
  73 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,7,6,5,4]
  74 ; CHECK-NOSSSE3-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
  75 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]
  76 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4]
  77 ; CHECK-NOSSSE3-NEXT:    packuswb %xmm2, %xmm0
  78 ; CHECK-NOSSSE3-NEXT:    retq
  79 ;
  80 ; CHECK-SSSE3-LABEL: or_bswap:
  81 ; CHECK-SSSE3:       # %bb.0:
  82 ; CHECK-SSSE3-NEXT:    por %xmm1, %xmm0
  83 ; CHECK-SSSE3-NEXT:    pshufb {{.*#+}} xmm0 = xmm0[3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12]
  84 ; CHECK-SSSE3-NEXT:    retq
  85 ;
  86 ; CHECK-AVX-LABEL: or_bswap:
  87 ; CHECK-AVX:       # %bb.0:
  88 ; CHECK-AVX-NEXT:    vpor %xmm1, %xmm0, %xmm0
  89 ; CHECK-AVX-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12]
  90 ; CHECK-AVX-NEXT:    retq
  91   %xt = call <4 x i32> @llvm.bswap.v4i32(<4 x i32> %x)
  92   %yt = call <4 x i32> @llvm.bswap.v4i32(<4 x i32> %y)
  93   %r = or <4 x i32> %xt, %yt
  94   ret <4 x i32> %r
  95 }
  96
  97 define <2 x i64> @test3(<2 x i64> %v) {
  98 ; CHECK-NOSSSE3-LABEL: test3:
  99 ; CHECK-NOSSSE3:       # %bb.0: # %entry
 100 ; CHECK-NOSSSE3-NEXT:    pxor %xmm1, %xmm1
 101 ; CHECK-NOSSSE3-NEXT:    movdqa %xmm0, %xmm2
 102 ; CHECK-NOSSSE3-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
 103 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm2 = xmm2[2,3,0,1]
 104 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[3,2,1,0,4,5,6,7]
 105 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,7,6,5,4]
 106 ; CHECK-NOSSSE3-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
 107 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
 108 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]
 109 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4]
 110 ; CHECK-NOSSSE3-NEXT:    packuswb %xmm2, %xmm0
 111 ; CHECK-NOSSSE3-NEXT:    retq
 112 ;
 113 ; CHECK-SSSE3-LABEL: test3:
 114 ; CHECK-SSSE3:       # %bb.0: # %entry
 115 ; CHECK-SSSE3-NEXT:    pshufb {{.*#+}} xmm0 = xmm0[7,6,5,4,3,2,1,0,15,14,13,12,11,10,9,8]
 116 ; CHECK-SSSE3-NEXT:    retq
 117 ;
 118 ; CHECK-AVX-LABEL: test3:
 119 ; CHECK-AVX:       # %bb.0: # %entry
 120 ; CHECK-AVX-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[7,6,5,4,3,2,1,0,15,14,13,12,11,10,9,8]
 121 ; CHECK-AVX-NEXT:    retq
 122 entry:
 123   %r = call <2 x i64> @llvm.bswap.v2i64(<2 x i64> %v)
 124   ret <2 x i64> %r
 125 }
 126
 127 declare <16 x i16> @llvm.bswap.v16i16(<16 x i16>)
 128 declare <8 x i32> @llvm.bswap.v8i32(<8 x i32>)
 129 declare <4 x i64> @llvm.bswap.v4i64(<4 x i64>)
 130
 131 define <16 x i16> @test4(<16 x i16> %v) {
 132 ; CHECK-NOSSSE3-LABEL: test4:
 133 ; CHECK-NOSSSE3:       # %bb.0: # %entry
 134 ; CHECK-NOSSSE3-NEXT:    pxor %xmm2, %xmm2
 135 ; CHECK-NOSSSE3-NEXT:    movdqa %xmm0, %xmm3
 136 ; CHECK-NOSSSE3-NEXT:    punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
 137 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm3 = xmm3[1,0,3,2,4,5,6,7]
 138 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,5,4,7,6]
 139 ; CHECK-NOSSSE3-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
 140 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[1,0,3,2,4,5,6,7]
 141 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,4,7,6]
 142 ; CHECK-NOSSSE3-NEXT:    packuswb %xmm3, %xmm0
 143 ; CHECK-NOSSSE3-NEXT:    movdqa %xmm1, %xmm3
 144 ; CHECK-NOSSSE3-NEXT:    punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
 145 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm3 = xmm3[1,0,3,2,4,5,6,7]
 146 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,5,4,7,6]
 147 ; CHECK-NOSSSE3-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
 148 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[1,0,3,2,4,5,6,7]
 149 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,5,4,7,6]
 150 ; CHECK-NOSSSE3-NEXT:    packuswb %xmm3, %xmm1
 151 ; CHECK-NOSSSE3-NEXT:    retq
 152 ;
 153 ; CHECK-SSSE3-LABEL: test4:
 154 ; CHECK-SSSE3:       # %bb.0: # %entry
 155 ; CHECK-SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
 156 ; CHECK-SSSE3-NEXT:    pshufb %xmm2, %xmm0
 157 ; CHECK-SSSE3-NEXT:    pshufb %xmm2, %xmm1
 158 ; CHECK-SSSE3-NEXT:    retq
 159 ;
 160 ; CHECK-AVX-LABEL: test4:
 161 ; CHECK-AVX:       # %bb.0: # %entry
 162 ; CHECK-AVX-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]
 163 ; CHECK-AVX-NEXT:    retq
 164 entry:
 165   %r = call <16 x i16> @llvm.bswap.v16i16(<16 x i16> %v)
 166   ret <16 x i16> %r
 167 }
 168
 169 define <8 x i32> @test5(<8 x i32> %v) {
 170 ; CHECK-NOSSSE3-LABEL: test5:
 171 ; CHECK-NOSSSE3:       # %bb.0: # %entry
 172 ; CHECK-NOSSSE3-NEXT:    pxor %xmm2, %xmm2
 173 ; CHECK-NOSSSE3-NEXT:    movdqa %xmm0, %xmm3
 174 ; CHECK-NOSSSE3-NEXT:    punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
 175 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm3 = xmm3[3,2,1,0,4,5,6,7]
 176 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,7,6,5,4]
 177 ; CHECK-NOSSSE3-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
 178 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]
 179 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4]
 180 ; CHECK-NOSSSE3-NEXT:    packuswb %xmm3, %xmm0
 181 ; CHECK-NOSSSE3-NEXT:    movdqa %xmm1, %xmm3
 182 ; CHECK-NOSSSE3-NEXT:    punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
 183 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm3 = xmm3[3,2,1,0,4,5,6,7]
 184 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,7,6,5,4]
 185 ; CHECK-NOSSSE3-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
 186 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[3,2,1,0,4,5,6,7]
 187 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,7,6,5,4]
 188 ; CHECK-NOSSSE3-NEXT:    packuswb %xmm3, %xmm1
 189 ; CHECK-NOSSSE3-NEXT:    retq
 190 ;
 191 ; CHECK-SSSE3-LABEL: test5:
 192 ; CHECK-SSSE3:       # %bb.0: # %entry
 193 ; CHECK-SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [3,2,1,0,7,6,5,4,11,10,9,8,15,14,13,12]
 194 ; CHECK-SSSE3-NEXT:    pshufb %xmm2, %xmm0
 195 ; CHECK-SSSE3-NEXT:    pshufb %xmm2, %xmm1
 196 ; CHECK-SSSE3-NEXT:    retq
 197 ;
 198 ; CHECK-AVX-LABEL: test5:
 199 ; CHECK-AVX:       # %bb.0: # %entry
 200 ; CHECK-AVX-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]
 201 ; CHECK-AVX-NEXT:    retq
 202 entry:
 203   %r = call <8 x i32> @llvm.bswap.v8i32(<8 x i32> %v)
 204   ret <8 x i32> %r
 205 }
 206
 207 define <4 x i64> @test6(<4 x i64> %v) {
 208 ; CHECK-NOSSSE3-LABEL: test6:
 209 ; CHECK-NOSSSE3:       # %bb.0: # %entry
 210 ; CHECK-NOSSSE3-NEXT:    pxor %xmm2, %xmm2
 211 ; CHECK-NOSSSE3-NEXT:    movdqa %xmm0, %xmm3
 212 ; CHECK-NOSSSE3-NEXT:    punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
 213 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[2,3,0,1]
 214 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm3 = xmm3[3,2,1,0,4,5,6,7]
 215 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,7,6,5,4]
 216 ; CHECK-NOSSSE3-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3],xmm0[4],xmm2[4],xmm0[5],xmm2[5],xmm0[6],xmm2[6],xmm0[7],xmm2[7]
 217 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
 218 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[3,2,1,0,4,5,6,7]
 219 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,7,6,5,4]
 220 ; CHECK-NOSSSE3-NEXT:    packuswb %xmm3, %xmm0
 221 ; CHECK-NOSSSE3-NEXT:    movdqa %xmm1, %xmm3
 222 ; CHECK-NOSSSE3-NEXT:    punpckhbw {{.*#+}} xmm3 = xmm3[8],xmm2[8],xmm3[9],xmm2[9],xmm3[10],xmm2[10],xmm3[11],xmm2[11],xmm3[12],xmm2[12],xmm3[13],xmm2[13],xmm3[14],xmm2[14],xmm3[15],xmm2[15]
 223 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm3 = xmm3[2,3,0,1]
 224 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm3 = xmm3[3,2,1,0,4,5,6,7]
 225 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm3 = xmm3[0,1,2,3,7,6,5,4]
 226 ; CHECK-NOSSSE3-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[1],xmm2[1],xmm1[2],xmm2[2],xmm1[3],xmm2[3],xmm1[4],xmm2[4],xmm1[5],xmm2[5],xmm1[6],xmm2[6],xmm1[7],xmm2[7]
 227 ; CHECK-NOSSSE3-NEXT:    pshufd {{.*#+}} xmm1 = xmm1[2,3,0,1]
 228 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm1 = xmm1[3,2,1,0,4,5,6,7]
 229 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm1 = xmm1[0,1,2,3,7,6,5,4]
 230 ; CHECK-NOSSSE3-NEXT:    packuswb %xmm3, %xmm1
 231 ; CHECK-NOSSSE3-NEXT:    retq
 232 ;
 233 ; CHECK-SSSE3-LABEL: test6:
 234 ; CHECK-SSSE3:       # %bb.0: # %entry
 235 ; CHECK-SSSE3-NEXT:    movdqa {{.*#+}} xmm2 = [7,6,5,4,3,2,1,0,15,14,13,12,11,10,9,8]
 236 ; CHECK-SSSE3-NEXT:    pshufb %xmm2, %xmm0
 237 ; CHECK-SSSE3-NEXT:    pshufb %xmm2, %xmm1
 238 ; CHECK-SSSE3-NEXT:    retq
 239 ;
 240 ; CHECK-AVX-LABEL: test6:
 241 ; CHECK-AVX:       # %bb.0: # %entry
 242 ; CHECK-AVX-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]
 243 ; CHECK-AVX-NEXT:    retq
 244 entry:
 245   %r = call <4 x i64> @llvm.bswap.v4i64(<4 x i64> %v)
 246   ret <4 x i64> %r
 247 }
 248
 249 declare <4 x i16> @llvm.bswap.v4i16(<4 x i16>)
 250
 251 define <4 x i16> @test7(<4 x i16> %v) {
 252 ; CHECK-NOSSSE3-LABEL: test7:
 253 ; CHECK-NOSSSE3:       # %bb.0: # %entry
 254 ; CHECK-NOSSSE3-NEXT:    pxor %xmm1, %xmm1
 255 ; CHECK-NOSSSE3-NEXT:    movdqa %xmm0, %xmm2
 256 ; CHECK-NOSSSE3-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm1[8],xmm2[9],xmm1[9],xmm2[10],xmm1[10],xmm2[11],xmm1[11],xmm2[12],xmm1[12],xmm2[13],xmm1[13],xmm2[14],xmm1[14],xmm2[15],xmm1[15]
 257 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm2 = xmm2[1,0,3,2,4,5,6,7]
 258 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm2 = xmm2[0,1,2,3,5,4,7,6]
 259 ; CHECK-NOSSSE3-NEXT:    punpcklbw {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3],xmm0[4],xmm1[4],xmm0[5],xmm1[5],xmm0[6],xmm1[6],xmm0[7],xmm1[7]
 260 ; CHECK-NOSSSE3-NEXT:    pshuflw {{.*#+}} xmm0 = xmm0[1,0,3,2,4,5,6,7]
 261 ; CHECK-NOSSSE3-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,4,7,6]
 262 ; CHECK-NOSSSE3-NEXT:    packuswb %xmm2, %xmm0
 263 ; CHECK-NOSSSE3-NEXT:    retq
 264 ;
 265 ; CHECK-SSSE3-LABEL: test7:
 266 ; CHECK-SSSE3:       # %bb.0: # %entry
 267 ; CHECK-SSSE3-NEXT:    pshufb {{.*#+}} xmm0 = xmm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
 268 ; CHECK-SSSE3-NEXT:    retq
 269 ;
 270 ; CHECK-AVX-LABEL: test7:
 271 ; CHECK-AVX:       # %bb.0: # %entry
 272 ; CHECK-AVX-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14]
 273 ; CHECK-AVX-NEXT:    retq
 274 entry:
 275   %r = call <4 x i16> @llvm.bswap.v4i16(<4 x i16> %v)
 276   ret <4 x i16> %r
 277 }
 278
 279 ;
 280 ; Double BSWAP -> Identity
 281 ;
 282
 283 define <8 x i16> @identity_v8i16(<8 x i16> %v) {
 284 ; CHECK-ALL-LABEL: identity_v8i16:
 285 ; CHECK-ALL:       # %bb.0: # %entry
 286 ; CHECK-ALL-NEXT:    retq
 287 entry:
 288   %bs1 = call <8 x i16> @llvm.bswap.v8i16(<8 x i16> %v)
 289   %bs2 = call <8 x i16> @llvm.bswap.v8i16(<8 x i16> %bs1)
 290   ret <8 x i16> %bs2
 291 }
 292
 293 define <4 x i32> @identity_v4i32(<4 x i32> %v) {
 294 ; CHECK-ALL-LABEL: identity_v4i32:
 295 ; CHECK-ALL:       # %bb.0: # %entry
 296 ; CHECK-ALL-NEXT:    retq
 297 entry:
 298   %bs1 = call <4 x i32> @llvm.bswap.v4i32(<4 x i32> %v)
 299   %bs2 = call <4 x i32> @llvm.bswap.v4i32(<4 x i32> %bs1)
 300   ret <4 x i32> %bs2
 301 }
 302
 303 define <2 x i64> @identity_v2i64(<2 x i64> %v) {
 304 ; CHECK-ALL-LABEL: identity_v2i64:
 305 ; CHECK-ALL:       # %bb.0: # %entry
 306 ; CHECK-ALL-NEXT:    retq
 307 entry:
 308   %bs1 = call <2 x i64> @llvm.bswap.v2i64(<2 x i64> %v)
 309   %bs2 = call <2 x i64> @llvm.bswap.v2i64(<2 x i64> %bs1)
 310   ret <2 x i64> %bs2
 311 }
 312
 313 define <16 x i16> @identity_v16i16(<16 x i16> %v) {
 314 ; CHECK-ALL-LABEL: identity_v16i16:
 315 ; CHECK-ALL:       # %bb.0: # %entry
 316 ; CHECK-ALL-NEXT:    retq
 317 entry:
 318   %bs1 = call <16 x i16> @llvm.bswap.v16i16(<16 x i16> %v)
 319   %bs2 = call <16 x i16> @llvm.bswap.v16i16(<16 x i16> %bs1)
 320   ret <16 x i16> %bs2
 321 }
 322
 323 define <8 x i32> @identity_v8i32(<8 x i32> %v) {
 324 ; CHECK-ALL-LABEL: identity_v8i32:
 325 ; CHECK-ALL:       # %bb.0: # %entry
 326 ; CHECK-ALL-NEXT:    retq
 327 entry:
 328   %bs1 = call <8 x i32> @llvm.bswap.v8i32(<8 x i32> %v)
 329   %bs2 = call <8 x i32> @llvm.bswap.v8i32(<8 x i32> %bs1)
 330   ret <8 x i32> %bs2
 331 }
 332
 333 define <4 x i64> @identity_v4i64(<4 x i64> %v) {
 334 ; CHECK-ALL-LABEL: identity_v4i64:
 335 ; CHECK-ALL:       # %bb.0: # %entry
 336 ; CHECK-ALL-NEXT:    retq
 337 entry:
 338   %bs1 = call <4 x i64> @llvm.bswap.v4i64(<4 x i64> %v)
 339   %bs2 = call <4 x i64> @llvm.bswap.v4i64(<4 x i64> %bs1)
 340   ret <4 x i64> %bs2
 341 }
 342
 343 define <4 x i16> @identity_v4i16(<4 x i16> %v) {
 344 ; CHECK-ALL-LABEL: identity_v4i16:
 345 ; CHECK-ALL:       # %bb.0: # %entry
 346 ; CHECK-ALL-NEXT:    retq
 347 entry:
 348   %bs1 = call <4 x i16> @llvm.bswap.v4i16(<4 x i16> %v)
 349   %bs2 = call <4 x i16> @llvm.bswap.v4i16(<4 x i16> %bs1)
 350   ret <4 x i16> %bs2
 351 }
 352
 353 ;
 354 ; Constant Folding
 355 ;
 356
 357 define <8 x i16> @fold_v8i16() {
 358 ; CHECK-SSE-LABEL: fold_v8i16:
 359 ; CHECK-SSE:       # %bb.0: # %entry
 360 ; CHECK-SSE-NEXT:    movaps {{.*#+}} xmm0 = [0,256,65535,512,65023,1024,64511,1536]
 361 ; CHECK-SSE-NEXT:    retq
 362 ;
 363 ; CHECK-AVX-LABEL: fold_v8i16:
 364 ; CHECK-AVX:       # %bb.0: # %entry
 365 ; CHECK-AVX-NEXT:    vmovaps {{.*#+}} xmm0 = [0,256,65535,512,65023,1024,64511,1536]
 366 ; CHECK-AVX-NEXT:    retq
 367 entry:
 368   %r = call <8 x i16> @llvm.bswap.v8i16(<8 x i16> <i16 0, i16 1, i16 -1, i16 2, i16 -3, i16 4, i16 -5, i16 6>)
 369   ret <8 x i16> %r
 370 }
 371
 372 define <4 x i32> @fold_v4i32() {
 373 ; CHECK-SSE-LABEL: fold_v4i32:
 374 ; CHECK-SSE:       # %bb.0: # %entry
 375 ; CHECK-SSE-NEXT:    movaps {{.*#+}} xmm0 = [0,4294967295,33554432,4261412863]
 376 ; CHECK-SSE-NEXT:    retq
 377 ;
 378 ; CHECK-AVX-LABEL: fold_v4i32:
 379 ; CHECK-AVX:       # %bb.0: # %entry
 380 ; CHECK-AVX-NEXT:    vmovaps {{.*#+}} xmm0 = [0,4294967295,33554432,4261412863]
 381 ; CHECK-AVX-NEXT:    retq
 382 entry:
 383   %r = call <4 x i32> @llvm.bswap.v4i32(<4 x i32> <i32 0, i32 -1, i32 2, i32 -3>)
 384   ret <4 x i32> %r
 385 }
 386
 387 define <2 x i64> @fold_v2i64() {
 388 ; CHECK-SSE-LABEL: fold_v2i64:
 389 ; CHECK-SSE:       # %bb.0: # %entry
 390 ; CHECK-SSE-NEXT:    movaps {{.*#+}} xmm0 = [18374686479671623680,18446744073709551615]
 391 ; CHECK-SSE-NEXT:    retq
 392 ;
 393 ; CHECK-AVX-LABEL: fold_v2i64:
 394 ; CHECK-AVX:       # %bb.0: # %entry
 395 ; CHECK-AVX-NEXT:    vmovaps {{.*#+}} xmm0 = [18374686479671623680,18446744073709551615]
 396 ; CHECK-AVX-NEXT:    retq
 397 entry:
 398   %r = call <2 x i64> @llvm.bswap.v2i64(<2 x i64> <i64 255, i64 -1>)
 399   ret <2 x i64> %r
 400 }
 401
 402 define <16 x i16> @fold_v16i16() {
 403 ; CHECK-SSE-LABEL: fold_v16i16:
 404 ; CHECK-SSE:       # %bb.0: # %entry
 405 ; CHECK-SSE-NEXT:    movaps {{.*#+}} xmm0 = [0,256,65535,512,65023,1024,64511,1536]
 406 ; CHECK-SSE-NEXT:    movaps {{.*#+}} xmm1 = [63999,2048,63487,2560,62975,3072,62463,3584]
 407 ; CHECK-SSE-NEXT:    retq
 408 ;
 409 ; CHECK-AVX-LABEL: fold_v16i16:
 410 ; CHECK-AVX:       # %bb.0: # %entry
 411 ; CHECK-AVX-NEXT:    vmovaps {{.*#+}} ymm0 = [0,256,65535,512,65023,1024,64511,1536,63999,2048,63487,2560,62975,3072,62463,3584]
 412 ; CHECK-AVX-NEXT:    retq
 413 entry:
 414   %r = call <16 x i16> @llvm.bswap.v16i16(<16 x i16> <i16 0, i16 1, i16 -1, i16 2, i16 -3, i16 4, i16 -5, i16 6, i16 -7, i16 8, i16 -9, i16 10, i16 -11, i16 12, i16 -13, i16 14>)
 415   ret <16 x i16> %r
 416 }
 417
 418 define <8 x i32> @fold_v8i32() {
 419 ; CHECK-SSE-LABEL: fold_v8i32:
 420 ; CHECK-SSE:       # %bb.0: # %entry
 421 ; CHECK-SSE-NEXT:    movaps {{.*#+}} xmm0 = [0,16777216,4294967295,33554432]
 422 ; CHECK-SSE-NEXT:    movaps {{.*#+}} xmm1 = [4261412863,67108864,4227858431,100663296]
 423 ; CHECK-SSE-NEXT:    retq
 424 ;
 425 ; CHECK-AVX-LABEL: fold_v8i32:
 426 ; CHECK-AVX:       # %bb.0: # %entry
 427 ; CHECK-AVX-NEXT:    vmovaps {{.*#+}} ymm0 = [0,16777216,4294967295,33554432,4261412863,67108864,4227858431,100663296]
 428 ; CHECK-AVX-NEXT:    retq
 429 entry:
 430   %r = call <8 x i32> @llvm.bswap.v8i32(<8 x i32> <i32 0, i32 1, i32 -1, i32 2, i32 -3, i32 4, i32 -5, i32 6>)
 431   ret <8 x i32> %r
 432 }
 433
 434 define <4 x i64> @fold_v4i64() {
 435 ; CHECK-SSE-LABEL: fold_v4i64:
 436 ; CHECK-SSE:       # %bb.0: # %entry
 437 ; CHECK-SSE-NEXT:    movaps {{.*#+}} xmm0 = [18374686479671623680,18446744073709551615]
 438 ; CHECK-SSE-NEXT:    movaps {{.*#+}} xmm1 = [18446462598732840960,72056494526300160]
 439 ; CHECK-SSE-NEXT:    retq
 440 ;
 441 ; CHECK-AVX-LABEL: fold_v4i64:
 442 ; CHECK-AVX:       # %bb.0: # %entry
 443 ; CHECK-AVX-NEXT:    vmovaps {{.*#+}} ymm0 = [18374686479671623680,18446744073709551615,18446462598732840960,72056494526300160]
 444 ; CHECK-AVX-NEXT:    retq
 445 entry:
 446   %r = call <4 x i64> @llvm.bswap.v4i64(<4 x i64> <i64 255, i64 -1, i64 65535, i64 16776960>)
 447   ret <4 x i64> %r
 448 }