1 ; RUN: llc < %s -march=x86 -mattr=+mmx,+ssse3,-avx | FileCheck %s --check-prefix=ALL --check-prefix=X86 2 ; RUN: llc < %s -march=x86 -mattr=+mmx,+avx | FileCheck %s --check-prefix=ALL --check-prefix=X86 3 ; RUN: llc < %s -march=x86-64 -mattr=+mmx,+ssse3,-avx | FileCheck %s --check-prefix=ALL --check-prefix=X64 4 ; RUN: llc < %s -march=x86-64 -mattr=+mmx,+avx | FileCheck %s --check-prefix=ALL --check-prefix=X64 5 6 declare x86_mmx @llvm.x86.ssse3.phadd.w(x86_mmx, x86_mmx) nounwind readnone 7 8 define i64 @test1(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 9 ; ALL-LABEL: @test1 10 ; ALL: phaddw 11 entry: 12 %0 = bitcast <1 x i64> %b to <4 x i16> 13 %1 = bitcast <1 x i64> %a to <4 x i16> 14 %2 = bitcast <4 x i16> %1 to x86_mmx 15 %3 = bitcast <4 x i16> %0 to x86_mmx 16 %4 = tail call x86_mmx @llvm.x86.ssse3.phadd.w(x86_mmx %2, x86_mmx %3) nounwind readnone 17 %5 = bitcast x86_mmx %4 to <4 x i16> 18 %6 = bitcast <4 x i16> %5 to <1 x i64> 19 %7 = extractelement <1 x i64> %6, i32 0 20 ret i64 %7 21 } 22 23 declare x86_mmx @llvm.x86.mmx.pcmpgt.d(x86_mmx, x86_mmx) nounwind readnone 24 25 define i64 @test88(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 26 ; ALL-LABEL: @test88 27 ; ALL: pcmpgtd 28 entry: 29 %0 = bitcast <1 x i64> %b to <2 x i32> 30 %1 = bitcast <1 x i64> %a to <2 x i32> 31 %mmx_var.i = bitcast <2 x i32> %1 to x86_mmx 32 %mmx_var1.i = bitcast <2 x i32> %0 to x86_mmx 33 %2 = tail call x86_mmx @llvm.x86.mmx.pcmpgt.d(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 34 %3 = bitcast x86_mmx %2 to <2 x i32> 35 %4 = bitcast <2 x i32> %3 to <1 x i64> 36 %5 = extractelement <1 x i64> %4, i32 0 37 ret i64 %5 38 } 39 40 declare x86_mmx @llvm.x86.mmx.pcmpgt.w(x86_mmx, x86_mmx) nounwind readnone 41 42 define i64 @test87(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 43 ; ALL-LABEL: @test87 44 ; ALL: pcmpgtw 45 entry: 46 %0 = bitcast <1 x i64> %b to <4 x i16> 47 %1 = bitcast <1 x i64> %a to <4 x i16> 48 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 49 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 50 %2 = tail call x86_mmx @llvm.x86.mmx.pcmpgt.w(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 51 %3 = bitcast x86_mmx %2 to <4 x i16> 52 %4 = bitcast <4 x i16> %3 to <1 x i64> 53 %5 = extractelement <1 x i64> %4, i32 0 54 ret i64 %5 55 } 56 57 declare x86_mmx @llvm.x86.mmx.pcmpgt.b(x86_mmx, x86_mmx) nounwind readnone 58 59 define i64 @test86(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 60 ; ALL-LABEL: @test86 61 ; ALL: pcmpgtb 62 entry: 63 %0 = bitcast <1 x i64> %b to <8 x i8> 64 %1 = bitcast <1 x i64> %a to <8 x i8> 65 %mmx_var.i = bitcast <8 x i8> %1 to x86_mmx 66 %mmx_var1.i = bitcast <8 x i8> %0 to x86_mmx 67 %2 = tail call x86_mmx @llvm.x86.mmx.pcmpgt.b(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 68 %3 = bitcast x86_mmx %2 to <8 x i8> 69 %4 = bitcast <8 x i8> %3 to <1 x i64> 70 %5 = extractelement <1 x i64> %4, i32 0 71 ret i64 %5 72 } 73 74 declare x86_mmx @llvm.x86.mmx.pcmpeq.d(x86_mmx, x86_mmx) nounwind readnone 75 76 define i64 @test85(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 77 ; ALL-LABEL: @test85 78 ; ALL: pcmpeqd 79 entry: 80 %0 = bitcast <1 x i64> %b to <2 x i32> 81 %1 = bitcast <1 x i64> %a to <2 x i32> 82 %mmx_var.i = bitcast <2 x i32> %1 to x86_mmx 83 %mmx_var1.i = bitcast <2 x i32> %0 to x86_mmx 84 %2 = tail call x86_mmx @llvm.x86.mmx.pcmpeq.d(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 85 %3 = bitcast x86_mmx %2 to <2 x i32> 86 %4 = bitcast <2 x i32> %3 to <1 x i64> 87 %5 = extractelement <1 x i64> %4, i32 0 88 ret i64 %5 89 } 90 91 declare x86_mmx @llvm.x86.mmx.pcmpeq.w(x86_mmx, x86_mmx) nounwind readnone 92 93 define i64 @test84(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 94 ; ALL-LABEL: @test84 95 ; ALL: pcmpeqw 96 entry: 97 %0 = bitcast <1 x i64> %b to <4 x i16> 98 %1 = bitcast <1 x i64> %a to <4 x i16> 99 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 100 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 101 %2 = tail call x86_mmx @llvm.x86.mmx.pcmpeq.w(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 102 %3 = bitcast x86_mmx %2 to <4 x i16> 103 %4 = bitcast <4 x i16> %3 to <1 x i64> 104 %5 = extractelement <1 x i64> %4, i32 0 105 ret i64 %5 106 } 107 108 declare x86_mmx @llvm.x86.mmx.pcmpeq.b(x86_mmx, x86_mmx) nounwind readnone 109 110 define i64 @test83(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 111 ; ALL-LABEL: @test83 112 ; ALL: pcmpeqb 113 entry: 114 %0 = bitcast <1 x i64> %b to <8 x i8> 115 %1 = bitcast <1 x i64> %a to <8 x i8> 116 %mmx_var.i = bitcast <8 x i8> %1 to x86_mmx 117 %mmx_var1.i = bitcast <8 x i8> %0 to x86_mmx 118 %2 = tail call x86_mmx @llvm.x86.mmx.pcmpeq.b(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 119 %3 = bitcast x86_mmx %2 to <8 x i8> 120 %4 = bitcast <8 x i8> %3 to <1 x i64> 121 %5 = extractelement <1 x i64> %4, i32 0 122 ret i64 %5 123 } 124 125 declare x86_mmx @llvm.x86.mmx.punpckldq(x86_mmx, x86_mmx) nounwind readnone 126 127 define i64 @test82(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 128 ; ALL-LABEL: @test82 129 ; X86: punpckldq {{.*#+}} mm0 = mm0[0],mem[0] 130 ; X64: punpckldq {{.*#+}} mm0 = mm0[0],mm1[0] 131 entry: 132 %0 = bitcast <1 x i64> %b to <2 x i32> 133 %1 = bitcast <1 x i64> %a to <2 x i32> 134 %mmx_var.i = bitcast <2 x i32> %1 to x86_mmx 135 %mmx_var1.i = bitcast <2 x i32> %0 to x86_mmx 136 %2 = tail call x86_mmx @llvm.x86.mmx.punpckldq(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 137 %3 = bitcast x86_mmx %2 to <2 x i32> 138 %4 = bitcast <2 x i32> %3 to <1 x i64> 139 %5 = extractelement <1 x i64> %4, i32 0 140 ret i64 %5 141 } 142 143 declare x86_mmx @llvm.x86.mmx.punpcklwd(x86_mmx, x86_mmx) nounwind readnone 144 145 define i64 @test81(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 146 ; ALL-LABEL: @test81 147 ; X86: punpcklwd {{.*#+}} mm0 = mm0[0],mem[0],mm0[1],mem[1] 148 ; X64: punpcklwd {{.*#+}} mm0 = mm0[0],mm1[0],mm0[1],mm1[1] 149 entry: 150 %0 = bitcast <1 x i64> %b to <4 x i16> 151 %1 = bitcast <1 x i64> %a to <4 x i16> 152 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 153 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 154 %2 = tail call x86_mmx @llvm.x86.mmx.punpcklwd(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 155 %3 = bitcast x86_mmx %2 to <4 x i16> 156 %4 = bitcast <4 x i16> %3 to <1 x i64> 157 %5 = extractelement <1 x i64> %4, i32 0 158 ret i64 %5 159 } 160 161 declare x86_mmx @llvm.x86.mmx.punpcklbw(x86_mmx, x86_mmx) nounwind readnone 162 163 define i64 @test80(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 164 ; ALL-LABEL: @test80 165 ; X86: punpcklbw {{.*#+}} mm0 = mm0[0],mem[0],mm0[1],mem[1],mm0[2],mem[2],mm0[3],mem[3] 166 ; X64: punpcklbw {{.*#+}} mm0 = mm0[0],mm1[0],mm0[1],mm1[1],mm0[2],mm1[2],mm0[3],mm1[3] 167 entry: 168 %0 = bitcast <1 x i64> %b to <8 x i8> 169 %1 = bitcast <1 x i64> %a to <8 x i8> 170 %mmx_var.i = bitcast <8 x i8> %1 to x86_mmx 171 %mmx_var1.i = bitcast <8 x i8> %0 to x86_mmx 172 %2 = tail call x86_mmx @llvm.x86.mmx.punpcklbw(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 173 %3 = bitcast x86_mmx %2 to <8 x i8> 174 %4 = bitcast <8 x i8> %3 to <1 x i64> 175 %5 = extractelement <1 x i64> %4, i32 0 176 ret i64 %5 177 } 178 179 declare x86_mmx @llvm.x86.mmx.punpckhdq(x86_mmx, x86_mmx) nounwind readnone 180 181 define i64 @test79(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 182 ; ALL-LABEL: @test79 183 ; X86: punpckhdq {{.*#+}} mm0 = mm0[1],mem[1] 184 ; X64: punpckhdq {{.*#+}} mm0 = mm0[1],mm1[1] 185 entry: 186 %0 = bitcast <1 x i64> %b to <2 x i32> 187 %1 = bitcast <1 x i64> %a to <2 x i32> 188 %mmx_var.i = bitcast <2 x i32> %1 to x86_mmx 189 %mmx_var1.i = bitcast <2 x i32> %0 to x86_mmx 190 %2 = tail call x86_mmx @llvm.x86.mmx.punpckhdq(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 191 %3 = bitcast x86_mmx %2 to <2 x i32> 192 %4 = bitcast <2 x i32> %3 to <1 x i64> 193 %5 = extractelement <1 x i64> %4, i32 0 194 ret i64 %5 195 } 196 197 declare x86_mmx @llvm.x86.mmx.punpckhwd(x86_mmx, x86_mmx) nounwind readnone 198 199 define i64 @test78(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 200 ; ALL-LABEL: @test78 201 ; X86: punpckhwd {{.*#+}} mm0 = mm0[2],mem[2],mm0[3],mem[3] 202 ; X64: punpckhwd {{.*#+}} mm0 = mm0[2],mm1[2],mm0[3],mm1[3] 203 entry: 204 %0 = bitcast <1 x i64> %b to <4 x i16> 205 %1 = bitcast <1 x i64> %a to <4 x i16> 206 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 207 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 208 %2 = tail call x86_mmx @llvm.x86.mmx.punpckhwd(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 209 %3 = bitcast x86_mmx %2 to <4 x i16> 210 %4 = bitcast <4 x i16> %3 to <1 x i64> 211 %5 = extractelement <1 x i64> %4, i32 0 212 ret i64 %5 213 } 214 215 declare x86_mmx @llvm.x86.mmx.punpckhbw(x86_mmx, x86_mmx) nounwind readnone 216 217 define i64 @test77(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 218 ; ALL-LABEL: @test77 219 ; X86: punpckhbw {{.*#+}} mm0 = mm0[4],mem[4],mm0[5],mem[5],mm0[6],mem[6],mm0[7],mem[7] 220 ; X64: punpckhbw {{.*#+}} mm0 = mm0[4],mm1[4],mm0[5],mm1[5],mm0[6],mm1[6],mm0[7],mm1[7] 221 entry: 222 %0 = bitcast <1 x i64> %b to <8 x i8> 223 %1 = bitcast <1 x i64> %a to <8 x i8> 224 %mmx_var.i = bitcast <8 x i8> %1 to x86_mmx 225 %mmx_var1.i = bitcast <8 x i8> %0 to x86_mmx 226 %2 = tail call x86_mmx @llvm.x86.mmx.punpckhbw(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 227 %3 = bitcast x86_mmx %2 to <8 x i8> 228 %4 = bitcast <8 x i8> %3 to <1 x i64> 229 %5 = extractelement <1 x i64> %4, i32 0 230 ret i64 %5 231 } 232 233 declare x86_mmx @llvm.x86.mmx.packuswb(x86_mmx, x86_mmx) nounwind readnone 234 235 define i64 @test76(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 236 ; ALL-LABEL: @test76 237 ; ALL: packuswb 238 entry: 239 %0 = bitcast <1 x i64> %b to <4 x i16> 240 %1 = bitcast <1 x i64> %a to <4 x i16> 241 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 242 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 243 %2 = tail call x86_mmx @llvm.x86.mmx.packuswb(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 244 %3 = bitcast x86_mmx %2 to <8 x i8> 245 %4 = bitcast <8 x i8> %3 to <1 x i64> 246 %5 = extractelement <1 x i64> %4, i32 0 247 ret i64 %5 248 } 249 250 declare x86_mmx @llvm.x86.mmx.packssdw(x86_mmx, x86_mmx) nounwind readnone 251 252 define i64 @test75(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 253 ; ALL-LABEL: @test75 254 ; ALL: packssdw 255 entry: 256 %0 = bitcast <1 x i64> %b to <2 x i32> 257 %1 = bitcast <1 x i64> %a to <2 x i32> 258 %mmx_var.i = bitcast <2 x i32> %1 to x86_mmx 259 %mmx_var1.i = bitcast <2 x i32> %0 to x86_mmx 260 %2 = tail call x86_mmx @llvm.x86.mmx.packssdw(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 261 %3 = bitcast x86_mmx %2 to <4 x i16> 262 %4 = bitcast <4 x i16> %3 to <1 x i64> 263 %5 = extractelement <1 x i64> %4, i32 0 264 ret i64 %5 265 } 266 267 declare x86_mmx @llvm.x86.mmx.packsswb(x86_mmx, x86_mmx) nounwind readnone 268 269 define i64 @test74(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 270 ; ALL-LABEL: @test74 271 ; ALL: packsswb 272 entry: 273 %0 = bitcast <1 x i64> %b to <4 x i16> 274 %1 = bitcast <1 x i64> %a to <4 x i16> 275 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 276 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 277 %2 = tail call x86_mmx @llvm.x86.mmx.packsswb(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 278 %3 = bitcast x86_mmx %2 to <8 x i8> 279 %4 = bitcast <8 x i8> %3 to <1 x i64> 280 %5 = extractelement <1 x i64> %4, i32 0 281 ret i64 %5 282 } 283 284 declare x86_mmx @llvm.x86.mmx.psrai.d(x86_mmx, i32) nounwind readnone 285 286 define i64 @test73(<1 x i64> %a) nounwind readnone optsize ssp { 287 ; ALL-LABEL: @test73 288 ; ALL: psrad 289 entry: 290 %0 = bitcast <1 x i64> %a to <2 x i32> 291 %mmx_var.i = bitcast <2 x i32> %0 to x86_mmx 292 %1 = tail call x86_mmx @llvm.x86.mmx.psrai.d(x86_mmx %mmx_var.i, i32 3) nounwind 293 %2 = bitcast x86_mmx %1 to <2 x i32> 294 %3 = bitcast <2 x i32> %2 to <1 x i64> 295 %4 = extractelement <1 x i64> %3, i32 0 296 ret i64 %4 297 } 298 299 declare x86_mmx @llvm.x86.mmx.psrai.w(x86_mmx, i32) nounwind readnone 300 301 define i64 @test72(<1 x i64> %a) nounwind readnone optsize ssp { 302 ; ALL-LABEL: @test72 303 ; ALL: psraw 304 entry: 305 %0 = bitcast <1 x i64> %a to <4 x i16> 306 %mmx_var.i = bitcast <4 x i16> %0 to x86_mmx 307 %1 = tail call x86_mmx @llvm.x86.mmx.psrai.w(x86_mmx %mmx_var.i, i32 3) nounwind 308 %2 = bitcast x86_mmx %1 to <4 x i16> 309 %3 = bitcast <4 x i16> %2 to <1 x i64> 310 %4 = extractelement <1 x i64> %3, i32 0 311 ret i64 %4 312 } 313 314 declare x86_mmx @llvm.x86.mmx.psrli.q(x86_mmx, i32) nounwind readnone 315 316 define i64 @test71(<1 x i64> %a) nounwind readnone optsize ssp { 317 ; ALL-LABEL: @test71 318 ; ALL: psrlq 319 entry: 320 %0 = extractelement <1 x i64> %a, i32 0 321 %mmx_var.i = bitcast i64 %0 to x86_mmx 322 %1 = tail call x86_mmx @llvm.x86.mmx.psrli.q(x86_mmx %mmx_var.i, i32 3) nounwind 323 %2 = bitcast x86_mmx %1 to i64 324 ret i64 %2 325 } 326 327 declare x86_mmx @llvm.x86.mmx.psrli.d(x86_mmx, i32) nounwind readnone 328 329 define i64 @test70(<1 x i64> %a) nounwind readnone optsize ssp { 330 ; ALL-LABEL: @test70 331 ; ALL: psrld 332 entry: 333 %0 = bitcast <1 x i64> %a to <2 x i32> 334 %mmx_var.i = bitcast <2 x i32> %0 to x86_mmx 335 %1 = tail call x86_mmx @llvm.x86.mmx.psrli.d(x86_mmx %mmx_var.i, i32 3) nounwind 336 %2 = bitcast x86_mmx %1 to <2 x i32> 337 %3 = bitcast <2 x i32> %2 to <1 x i64> 338 %4 = extractelement <1 x i64> %3, i32 0 339 ret i64 %4 340 } 341 342 declare x86_mmx @llvm.x86.mmx.psrli.w(x86_mmx, i32) nounwind readnone 343 344 define i64 @test69(<1 x i64> %a) nounwind readnone optsize ssp { 345 ; ALL-LABEL: @test69 346 ; ALL: psrlw 347 entry: 348 %0 = bitcast <1 x i64> %a to <4 x i16> 349 %mmx_var.i = bitcast <4 x i16> %0 to x86_mmx 350 %1 = tail call x86_mmx @llvm.x86.mmx.psrli.w(x86_mmx %mmx_var.i, i32 3) nounwind 351 %2 = bitcast x86_mmx %1 to <4 x i16> 352 %3 = bitcast <4 x i16> %2 to <1 x i64> 353 %4 = extractelement <1 x i64> %3, i32 0 354 ret i64 %4 355 } 356 357 declare x86_mmx @llvm.x86.mmx.pslli.q(x86_mmx, i32) nounwind readnone 358 359 define i64 @test68(<1 x i64> %a) nounwind readnone optsize ssp { 360 ; ALL-LABEL: @test68 361 ; ALL: psllq 362 entry: 363 %0 = extractelement <1 x i64> %a, i32 0 364 %mmx_var.i = bitcast i64 %0 to x86_mmx 365 %1 = tail call x86_mmx @llvm.x86.mmx.pslli.q(x86_mmx %mmx_var.i, i32 3) nounwind 366 %2 = bitcast x86_mmx %1 to i64 367 ret i64 %2 368 } 369 370 declare x86_mmx @llvm.x86.mmx.pslli.d(x86_mmx, i32) nounwind readnone 371 372 define i64 @test67(<1 x i64> %a) nounwind readnone optsize ssp { 373 ; ALL-LABEL: @test67 374 ; ALL: pslld 375 entry: 376 %0 = bitcast <1 x i64> %a to <2 x i32> 377 %mmx_var.i = bitcast <2 x i32> %0 to x86_mmx 378 %1 = tail call x86_mmx @llvm.x86.mmx.pslli.d(x86_mmx %mmx_var.i, i32 3) nounwind 379 %2 = bitcast x86_mmx %1 to <2 x i32> 380 %3 = bitcast <2 x i32> %2 to <1 x i64> 381 %4 = extractelement <1 x i64> %3, i32 0 382 ret i64 %4 383 } 384 385 declare x86_mmx @llvm.x86.mmx.pslli.w(x86_mmx, i32) nounwind readnone 386 387 define i64 @test66(<1 x i64> %a) nounwind readnone optsize ssp { 388 ; ALL-LABEL: @test66 389 ; ALL: psllw 390 entry: 391 %0 = bitcast <1 x i64> %a to <4 x i16> 392 %mmx_var.i = bitcast <4 x i16> %0 to x86_mmx 393 %1 = tail call x86_mmx @llvm.x86.mmx.pslli.w(x86_mmx %mmx_var.i, i32 3) nounwind 394 %2 = bitcast x86_mmx %1 to <4 x i16> 395 %3 = bitcast <4 x i16> %2 to <1 x i64> 396 %4 = extractelement <1 x i64> %3, i32 0 397 ret i64 %4 398 } 399 400 declare x86_mmx @llvm.x86.mmx.psra.d(x86_mmx, x86_mmx) nounwind readnone 401 402 define i64 @test65(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 403 ; ALL-LABEL: @test65 404 ; ALL: psrad 405 entry: 406 %0 = bitcast <1 x i64> %a to <2 x i32> 407 %mmx_var.i = bitcast <2 x i32> %0 to x86_mmx 408 %1 = extractelement <1 x i64> %b, i32 0 409 %mmx_var1.i = bitcast i64 %1 to x86_mmx 410 %2 = tail call x86_mmx @llvm.x86.mmx.psra.d(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 411 %3 = bitcast x86_mmx %2 to <2 x i32> 412 %4 = bitcast <2 x i32> %3 to <1 x i64> 413 %5 = extractelement <1 x i64> %4, i32 0 414 ret i64 %5 415 } 416 417 declare x86_mmx @llvm.x86.mmx.psra.w(x86_mmx, x86_mmx) nounwind readnone 418 419 define i64 @test64(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 420 ; ALL-LABEL: @test64 421 ; ALL: psraw 422 entry: 423 %0 = bitcast <1 x i64> %a to <4 x i16> 424 %mmx_var.i = bitcast <4 x i16> %0 to x86_mmx 425 %1 = extractelement <1 x i64> %b, i32 0 426 %mmx_var1.i = bitcast i64 %1 to x86_mmx 427 %2 = tail call x86_mmx @llvm.x86.mmx.psra.w(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 428 %3 = bitcast x86_mmx %2 to <4 x i16> 429 %4 = bitcast <4 x i16> %3 to <1 x i64> 430 %5 = extractelement <1 x i64> %4, i32 0 431 ret i64 %5 432 } 433 434 declare x86_mmx @llvm.x86.mmx.psrl.q(x86_mmx, x86_mmx) nounwind readnone 435 436 define i64 @test63(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 437 ; ALL-LABEL: @test63 438 ; ALL: psrlq 439 entry: 440 %0 = extractelement <1 x i64> %a, i32 0 441 %mmx_var.i = bitcast i64 %0 to x86_mmx 442 %1 = extractelement <1 x i64> %b, i32 0 443 %mmx_var1.i = bitcast i64 %1 to x86_mmx 444 %2 = tail call x86_mmx @llvm.x86.mmx.psrl.q(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 445 %3 = bitcast x86_mmx %2 to i64 446 ret i64 %3 447 } 448 449 declare x86_mmx @llvm.x86.mmx.psrl.d(x86_mmx, x86_mmx) nounwind readnone 450 451 define i64 @test62(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 452 ; ALL-LABEL: @test62 453 ; ALL: psrld 454 entry: 455 %0 = bitcast <1 x i64> %a to <2 x i32> 456 %mmx_var.i = bitcast <2 x i32> %0 to x86_mmx 457 %1 = extractelement <1 x i64> %b, i32 0 458 %mmx_var1.i = bitcast i64 %1 to x86_mmx 459 %2 = tail call x86_mmx @llvm.x86.mmx.psrl.d(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 460 %3 = bitcast x86_mmx %2 to <2 x i32> 461 %4 = bitcast <2 x i32> %3 to <1 x i64> 462 %5 = extractelement <1 x i64> %4, i32 0 463 ret i64 %5 464 } 465 466 declare x86_mmx @llvm.x86.mmx.psrl.w(x86_mmx, x86_mmx) nounwind readnone 467 468 define i64 @test61(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 469 ; ALL-LABEL: @test61 470 ; ALL: psrlw 471 entry: 472 %0 = bitcast <1 x i64> %a to <4 x i16> 473 %mmx_var.i = bitcast <4 x i16> %0 to x86_mmx 474 %1 = extractelement <1 x i64> %b, i32 0 475 %mmx_var1.i = bitcast i64 %1 to x86_mmx 476 %2 = tail call x86_mmx @llvm.x86.mmx.psrl.w(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 477 %3 = bitcast x86_mmx %2 to <4 x i16> 478 %4 = bitcast <4 x i16> %3 to <1 x i64> 479 %5 = extractelement <1 x i64> %4, i32 0 480 ret i64 %5 481 } 482 483 declare x86_mmx @llvm.x86.mmx.psll.q(x86_mmx, x86_mmx) nounwind readnone 484 485 define i64 @test60(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 486 ; ALL-LABEL: @test60 487 ; ALL: psllq 488 entry: 489 %0 = extractelement <1 x i64> %a, i32 0 490 %mmx_var.i = bitcast i64 %0 to x86_mmx 491 %1 = extractelement <1 x i64> %b, i32 0 492 %mmx_var1.i = bitcast i64 %1 to x86_mmx 493 %2 = tail call x86_mmx @llvm.x86.mmx.psll.q(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 494 %3 = bitcast x86_mmx %2 to i64 495 ret i64 %3 496 } 497 498 declare x86_mmx @llvm.x86.mmx.psll.d(x86_mmx, x86_mmx) nounwind readnone 499 500 define i64 @test59(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 501 ; ALL-LABEL: @test59 502 ; ALL: pslld 503 entry: 504 %0 = bitcast <1 x i64> %a to <2 x i32> 505 %mmx_var.i = bitcast <2 x i32> %0 to x86_mmx 506 %1 = extractelement <1 x i64> %b, i32 0 507 %mmx_var1.i = bitcast i64 %1 to x86_mmx 508 %2 = tail call x86_mmx @llvm.x86.mmx.psll.d(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 509 %3 = bitcast x86_mmx %2 to <2 x i32> 510 %4 = bitcast <2 x i32> %3 to <1 x i64> 511 %5 = extractelement <1 x i64> %4, i32 0 512 ret i64 %5 513 } 514 515 declare x86_mmx @llvm.x86.mmx.psll.w(x86_mmx, x86_mmx) nounwind readnone 516 517 define i64 @test58(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 518 ; ALL-LABEL: @test58 519 ; ALL: psllw 520 entry: 521 %0 = bitcast <1 x i64> %a to <4 x i16> 522 %mmx_var.i = bitcast <4 x i16> %0 to x86_mmx 523 %1 = extractelement <1 x i64> %b, i32 0 524 %mmx_var1.i = bitcast i64 %1 to x86_mmx 525 %2 = tail call x86_mmx @llvm.x86.mmx.psll.w(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 526 %3 = bitcast x86_mmx %2 to <4 x i16> 527 %4 = bitcast <4 x i16> %3 to <1 x i64> 528 %5 = extractelement <1 x i64> %4, i32 0 529 ret i64 %5 530 } 531 532 declare x86_mmx @llvm.x86.mmx.pxor(x86_mmx, x86_mmx) nounwind readnone 533 534 define i64 @test56(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 535 ; ALL-LABEL: @test56 536 ; ALL: pxor 537 entry: 538 %0 = bitcast <1 x i64> %b to <2 x i32> 539 %1 = bitcast <1 x i64> %a to <2 x i32> 540 %mmx_var.i = bitcast <2 x i32> %1 to x86_mmx 541 %mmx_var1.i = bitcast <2 x i32> %0 to x86_mmx 542 %2 = tail call x86_mmx @llvm.x86.mmx.pxor(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 543 %3 = bitcast x86_mmx %2 to <2 x i32> 544 %4 = bitcast <2 x i32> %3 to <1 x i64> 545 %5 = extractelement <1 x i64> %4, i32 0 546 ret i64 %5 547 } 548 549 declare x86_mmx @llvm.x86.mmx.por(x86_mmx, x86_mmx) nounwind readnone 550 551 define i64 @test55(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 552 ; ALL-LABEL: @test55 553 ; ALL: por 554 entry: 555 %0 = bitcast <1 x i64> %b to <2 x i32> 556 %1 = bitcast <1 x i64> %a to <2 x i32> 557 %mmx_var.i = bitcast <2 x i32> %1 to x86_mmx 558 %mmx_var1.i = bitcast <2 x i32> %0 to x86_mmx 559 %2 = tail call x86_mmx @llvm.x86.mmx.por(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 560 %3 = bitcast x86_mmx %2 to <2 x i32> 561 %4 = bitcast <2 x i32> %3 to <1 x i64> 562 %5 = extractelement <1 x i64> %4, i32 0 563 ret i64 %5 564 } 565 566 declare x86_mmx @llvm.x86.mmx.pandn(x86_mmx, x86_mmx) nounwind readnone 567 568 define i64 @test54(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 569 ; ALL-LABEL: @test54 570 ; ALL: pandn 571 entry: 572 %0 = bitcast <1 x i64> %b to <2 x i32> 573 %1 = bitcast <1 x i64> %a to <2 x i32> 574 %mmx_var.i = bitcast <2 x i32> %1 to x86_mmx 575 %mmx_var1.i = bitcast <2 x i32> %0 to x86_mmx 576 %2 = tail call x86_mmx @llvm.x86.mmx.pandn(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 577 %3 = bitcast x86_mmx %2 to <2 x i32> 578 %4 = bitcast <2 x i32> %3 to <1 x i64> 579 %5 = extractelement <1 x i64> %4, i32 0 580 ret i64 %5 581 } 582 583 declare x86_mmx @llvm.x86.mmx.pand(x86_mmx, x86_mmx) nounwind readnone 584 585 define i64 @test53(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 586 ; ALL-LABEL: @test53 587 ; ALL: pand 588 entry: 589 %0 = bitcast <1 x i64> %b to <2 x i32> 590 %1 = bitcast <1 x i64> %a to <2 x i32> 591 %mmx_var.i = bitcast <2 x i32> %1 to x86_mmx 592 %mmx_var1.i = bitcast <2 x i32> %0 to x86_mmx 593 %2 = tail call x86_mmx @llvm.x86.mmx.pand(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 594 %3 = bitcast x86_mmx %2 to <2 x i32> 595 %4 = bitcast <2 x i32> %3 to <1 x i64> 596 %5 = extractelement <1 x i64> %4, i32 0 597 ret i64 %5 598 } 599 600 declare x86_mmx @llvm.x86.mmx.pmull.w(x86_mmx, x86_mmx) nounwind readnone 601 602 define i64 @test52(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 603 ; ALL-LABEL: @test52 604 ; ALL: pmullw 605 entry: 606 %0 = bitcast <1 x i64> %b to <4 x i16> 607 %1 = bitcast <1 x i64> %a to <4 x i16> 608 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 609 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 610 %2 = tail call x86_mmx @llvm.x86.mmx.pmull.w(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 611 %3 = bitcast x86_mmx %2 to <4 x i16> 612 %4 = bitcast <4 x i16> %3 to <1 x i64> 613 %5 = extractelement <1 x i64> %4, i32 0 614 ret i64 %5 615 } 616 617 define i64 @test51(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 618 ; ALL-LABEL: @test51 619 ; ALL: pmullw 620 entry: 621 %0 = bitcast <1 x i64> %b to <4 x i16> 622 %1 = bitcast <1 x i64> %a to <4 x i16> 623 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 624 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 625 %2 = tail call x86_mmx @llvm.x86.mmx.pmull.w(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 626 %3 = bitcast x86_mmx %2 to <4 x i16> 627 %4 = bitcast <4 x i16> %3 to <1 x i64> 628 %5 = extractelement <1 x i64> %4, i32 0 629 ret i64 %5 630 } 631 632 declare x86_mmx @llvm.x86.mmx.pmulh.w(x86_mmx, x86_mmx) nounwind readnone 633 634 define i64 @test50(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 635 ; ALL-LABEL: @test50 636 ; ALL: pmulhw 637 entry: 638 %0 = bitcast <1 x i64> %b to <4 x i16> 639 %1 = bitcast <1 x i64> %a to <4 x i16> 640 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 641 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 642 %2 = tail call x86_mmx @llvm.x86.mmx.pmulh.w(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 643 %3 = bitcast x86_mmx %2 to <4 x i16> 644 %4 = bitcast <4 x i16> %3 to <1 x i64> 645 %5 = extractelement <1 x i64> %4, i32 0 646 ret i64 %5 647 } 648 649 declare x86_mmx @llvm.x86.mmx.pmadd.wd(x86_mmx, x86_mmx) nounwind readnone 650 651 define i64 @test49(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 652 ; ALL-LABEL: @test49 653 ; ALL: pmaddwd 654 entry: 655 %0 = bitcast <1 x i64> %b to <4 x i16> 656 %1 = bitcast <1 x i64> %a to <4 x i16> 657 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 658 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 659 %2 = tail call x86_mmx @llvm.x86.mmx.pmadd.wd(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 660 %3 = bitcast x86_mmx %2 to <2 x i32> 661 %4 = bitcast <2 x i32> %3 to <1 x i64> 662 %5 = extractelement <1 x i64> %4, i32 0 663 ret i64 %5 664 } 665 666 declare x86_mmx @llvm.x86.mmx.psubus.w(x86_mmx, x86_mmx) nounwind readnone 667 668 define i64 @test48(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 669 ; ALL-LABEL: @test48 670 ; ALL: psubusw 671 entry: 672 %0 = bitcast <1 x i64> %b to <4 x i16> 673 %1 = bitcast <1 x i64> %a to <4 x i16> 674 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 675 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 676 %2 = tail call x86_mmx @llvm.x86.mmx.psubus.w(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 677 %3 = bitcast x86_mmx %2 to <4 x i16> 678 %4 = bitcast <4 x i16> %3 to <1 x i64> 679 %5 = extractelement <1 x i64> %4, i32 0 680 ret i64 %5 681 } 682 683 declare x86_mmx @llvm.x86.mmx.psubus.b(x86_mmx, x86_mmx) nounwind readnone 684 685 define i64 @test47(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 686 ; ALL-LABEL: @test47 687 ; ALL: psubusb 688 entry: 689 %0 = bitcast <1 x i64> %b to <8 x i8> 690 %1 = bitcast <1 x i64> %a to <8 x i8> 691 %mmx_var.i = bitcast <8 x i8> %1 to x86_mmx 692 %mmx_var1.i = bitcast <8 x i8> %0 to x86_mmx 693 %2 = tail call x86_mmx @llvm.x86.mmx.psubus.b(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 694 %3 = bitcast x86_mmx %2 to <8 x i8> 695 %4 = bitcast <8 x i8> %3 to <1 x i64> 696 %5 = extractelement <1 x i64> %4, i32 0 697 ret i64 %5 698 } 699 700 declare x86_mmx @llvm.x86.mmx.psubs.w(x86_mmx, x86_mmx) nounwind readnone 701 702 define i64 @test46(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 703 ; ALL-LABEL: @test46 704 ; ALL: psubsw 705 entry: 706 %0 = bitcast <1 x i64> %b to <4 x i16> 707 %1 = bitcast <1 x i64> %a to <4 x i16> 708 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 709 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 710 %2 = tail call x86_mmx @llvm.x86.mmx.psubs.w(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 711 %3 = bitcast x86_mmx %2 to <4 x i16> 712 %4 = bitcast <4 x i16> %3 to <1 x i64> 713 %5 = extractelement <1 x i64> %4, i32 0 714 ret i64 %5 715 } 716 717 declare x86_mmx @llvm.x86.mmx.psubs.b(x86_mmx, x86_mmx) nounwind readnone 718 719 define i64 @test45(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 720 ; ALL-LABEL: @test45 721 ; ALL: psubsb 722 entry: 723 %0 = bitcast <1 x i64> %b to <8 x i8> 724 %1 = bitcast <1 x i64> %a to <8 x i8> 725 %mmx_var.i = bitcast <8 x i8> %1 to x86_mmx 726 %mmx_var1.i = bitcast <8 x i8> %0 to x86_mmx 727 %2 = tail call x86_mmx @llvm.x86.mmx.psubs.b(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 728 %3 = bitcast x86_mmx %2 to <8 x i8> 729 %4 = bitcast <8 x i8> %3 to <1 x i64> 730 %5 = extractelement <1 x i64> %4, i32 0 731 ret i64 %5 732 } 733 734 define i64 @test44(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 735 ; ALL-LABEL: @test44 736 ; ALL: psubq 737 entry: 738 %0 = extractelement <1 x i64> %a, i32 0 739 %mmx_var = bitcast i64 %0 to x86_mmx 740 %1 = extractelement <1 x i64> %b, i32 0 741 %mmx_var1 = bitcast i64 %1 to x86_mmx 742 %2 = tail call x86_mmx @llvm.x86.mmx.psub.q(x86_mmx %mmx_var, x86_mmx %mmx_var1) 743 %3 = bitcast x86_mmx %2 to i64 744 ret i64 %3 745 } 746 747 declare x86_mmx @llvm.x86.mmx.psub.q(x86_mmx, x86_mmx) nounwind readnone 748 749 declare x86_mmx @llvm.x86.mmx.psub.d(x86_mmx, x86_mmx) nounwind readnone 750 751 define i64 @test43(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 752 ; ALL-LABEL: @test43 753 ; ALL: psubd 754 entry: 755 %0 = bitcast <1 x i64> %b to <2 x i32> 756 %1 = bitcast <1 x i64> %a to <2 x i32> 757 %mmx_var.i = bitcast <2 x i32> %1 to x86_mmx 758 %mmx_var1.i = bitcast <2 x i32> %0 to x86_mmx 759 %2 = tail call x86_mmx @llvm.x86.mmx.psub.d(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 760 %3 = bitcast x86_mmx %2 to <2 x i32> 761 %4 = bitcast <2 x i32> %3 to <1 x i64> 762 %5 = extractelement <1 x i64> %4, i32 0 763 ret i64 %5 764 } 765 766 declare x86_mmx @llvm.x86.mmx.psub.w(x86_mmx, x86_mmx) nounwind readnone 767 768 define i64 @test42(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 769 ; ALL-LABEL: @test42 770 ; ALL: psubw 771 entry: 772 %0 = bitcast <1 x i64> %b to <4 x i16> 773 %1 = bitcast <1 x i64> %a to <4 x i16> 774 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 775 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 776 %2 = tail call x86_mmx @llvm.x86.mmx.psub.w(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 777 %3 = bitcast x86_mmx %2 to <4 x i16> 778 %4 = bitcast <4 x i16> %3 to <1 x i64> 779 %5 = extractelement <1 x i64> %4, i32 0 780 ret i64 %5 781 } 782 783 declare x86_mmx @llvm.x86.mmx.psub.b(x86_mmx, x86_mmx) nounwind readnone 784 785 define i64 @test41(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 786 ; ALL-LABEL: @test41 787 ; ALL: psubb 788 entry: 789 %0 = bitcast <1 x i64> %b to <8 x i8> 790 %1 = bitcast <1 x i64> %a to <8 x i8> 791 %mmx_var.i = bitcast <8 x i8> %1 to x86_mmx 792 %mmx_var1.i = bitcast <8 x i8> %0 to x86_mmx 793 %2 = tail call x86_mmx @llvm.x86.mmx.psub.b(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 794 %3 = bitcast x86_mmx %2 to <8 x i8> 795 %4 = bitcast <8 x i8> %3 to <1 x i64> 796 %5 = extractelement <1 x i64> %4, i32 0 797 ret i64 %5 798 } 799 800 declare x86_mmx @llvm.x86.mmx.paddus.w(x86_mmx, x86_mmx) nounwind readnone 801 802 define i64 @test40(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 803 ; ALL-LABEL: @test40 804 ; ALL: paddusw 805 entry: 806 %0 = bitcast <1 x i64> %b to <4 x i16> 807 %1 = bitcast <1 x i64> %a to <4 x i16> 808 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 809 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 810 %2 = tail call x86_mmx @llvm.x86.mmx.paddus.w(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 811 %3 = bitcast x86_mmx %2 to <4 x i16> 812 %4 = bitcast <4 x i16> %3 to <1 x i64> 813 %5 = extractelement <1 x i64> %4, i32 0 814 ret i64 %5 815 } 816 817 declare x86_mmx @llvm.x86.mmx.paddus.b(x86_mmx, x86_mmx) nounwind readnone 818 819 define i64 @test39(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 820 ; ALL-LABEL: @test39 821 ; ALL: paddusb 822 entry: 823 %0 = bitcast <1 x i64> %b to <8 x i8> 824 %1 = bitcast <1 x i64> %a to <8 x i8> 825 %mmx_var.i = bitcast <8 x i8> %1 to x86_mmx 826 %mmx_var1.i = bitcast <8 x i8> %0 to x86_mmx 827 %2 = tail call x86_mmx @llvm.x86.mmx.paddus.b(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 828 %3 = bitcast x86_mmx %2 to <8 x i8> 829 %4 = bitcast <8 x i8> %3 to <1 x i64> 830 %5 = extractelement <1 x i64> %4, i32 0 831 ret i64 %5 832 } 833 834 declare x86_mmx @llvm.x86.mmx.padds.w(x86_mmx, x86_mmx) nounwind readnone 835 836 define i64 @test38(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 837 ; ALL-LABEL: @test38 838 ; ALL: paddsw 839 entry: 840 %0 = bitcast <1 x i64> %b to <4 x i16> 841 %1 = bitcast <1 x i64> %a to <4 x i16> 842 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 843 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 844 %2 = tail call x86_mmx @llvm.x86.mmx.padds.w(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 845 %3 = bitcast x86_mmx %2 to <4 x i16> 846 %4 = bitcast <4 x i16> %3 to <1 x i64> 847 %5 = extractelement <1 x i64> %4, i32 0 848 ret i64 %5 849 } 850 851 declare x86_mmx @llvm.x86.mmx.padds.b(x86_mmx, x86_mmx) nounwind readnone 852 853 define i64 @test37(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 854 ; ALL-LABEL: @test37 855 ; ALL: paddsb 856 entry: 857 %0 = bitcast <1 x i64> %b to <8 x i8> 858 %1 = bitcast <1 x i64> %a to <8 x i8> 859 %mmx_var.i = bitcast <8 x i8> %1 to x86_mmx 860 %mmx_var1.i = bitcast <8 x i8> %0 to x86_mmx 861 %2 = tail call x86_mmx @llvm.x86.mmx.padds.b(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 862 %3 = bitcast x86_mmx %2 to <8 x i8> 863 %4 = bitcast <8 x i8> %3 to <1 x i64> 864 %5 = extractelement <1 x i64> %4, i32 0 865 ret i64 %5 866 } 867 868 declare x86_mmx @llvm.x86.mmx.padd.q(x86_mmx, x86_mmx) nounwind readnone 869 870 define i64 @test36(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 871 ; ALL-LABEL: @test36 872 ; ALL: paddq 873 entry: 874 %0 = extractelement <1 x i64> %a, i32 0 875 %mmx_var = bitcast i64 %0 to x86_mmx 876 %1 = extractelement <1 x i64> %b, i32 0 877 %mmx_var1 = bitcast i64 %1 to x86_mmx 878 %2 = tail call x86_mmx @llvm.x86.mmx.padd.q(x86_mmx %mmx_var, x86_mmx %mmx_var1) 879 %3 = bitcast x86_mmx %2 to i64 880 ret i64 %3 881 } 882 883 declare x86_mmx @llvm.x86.mmx.padd.d(x86_mmx, x86_mmx) nounwind readnone 884 885 define i64 @test35(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 886 ; ALL-LABEL: @test35 887 ; ALL: paddd 888 entry: 889 %0 = bitcast <1 x i64> %b to <2 x i32> 890 %1 = bitcast <1 x i64> %a to <2 x i32> 891 %mmx_var.i = bitcast <2 x i32> %1 to x86_mmx 892 %mmx_var1.i = bitcast <2 x i32> %0 to x86_mmx 893 %2 = tail call x86_mmx @llvm.x86.mmx.padd.d(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 894 %3 = bitcast x86_mmx %2 to <2 x i32> 895 %4 = bitcast <2 x i32> %3 to <1 x i64> 896 %5 = extractelement <1 x i64> %4, i32 0 897 ret i64 %5 898 } 899 900 declare x86_mmx @llvm.x86.mmx.padd.w(x86_mmx, x86_mmx) nounwind readnone 901 902 define i64 @test34(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 903 ; ALL-LABEL: @test34 904 ; ALL: paddw 905 entry: 906 %0 = bitcast <1 x i64> %b to <4 x i16> 907 %1 = bitcast <1 x i64> %a to <4 x i16> 908 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 909 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 910 %2 = tail call x86_mmx @llvm.x86.mmx.padd.w(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 911 %3 = bitcast x86_mmx %2 to <4 x i16> 912 %4 = bitcast <4 x i16> %3 to <1 x i64> 913 %5 = extractelement <1 x i64> %4, i32 0 914 ret i64 %5 915 } 916 917 declare x86_mmx @llvm.x86.mmx.padd.b(x86_mmx, x86_mmx) nounwind readnone 918 919 define i64 @test33(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 920 ; ALL-LABEL: @test33 921 ; ALL: paddb 922 entry: 923 %0 = bitcast <1 x i64> %b to <8 x i8> 924 %1 = bitcast <1 x i64> %a to <8 x i8> 925 %mmx_var.i = bitcast <8 x i8> %1 to x86_mmx 926 %mmx_var1.i = bitcast <8 x i8> %0 to x86_mmx 927 %2 = tail call x86_mmx @llvm.x86.mmx.padd.b(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 928 %3 = bitcast x86_mmx %2 to <8 x i8> 929 %4 = bitcast <8 x i8> %3 to <1 x i64> 930 %5 = extractelement <1 x i64> %4, i32 0 931 ret i64 %5 932 } 933 934 declare x86_mmx @llvm.x86.mmx.psad.bw(x86_mmx, x86_mmx) nounwind readnone 935 936 define i64 @test32(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 937 ; ALL-LABEL: @test32 938 ; ALL: psadbw 939 entry: 940 %0 = bitcast <1 x i64> %b to <8 x i8> 941 %1 = bitcast <1 x i64> %a to <8 x i8> 942 %mmx_var.i = bitcast <8 x i8> %1 to x86_mmx 943 %mmx_var1.i = bitcast <8 x i8> %0 to x86_mmx 944 %2 = tail call x86_mmx @llvm.x86.mmx.psad.bw(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 945 %3 = bitcast x86_mmx %2 to i64 946 ret i64 %3 947 } 948 949 declare x86_mmx @llvm.x86.mmx.pmins.w(x86_mmx, x86_mmx) nounwind readnone 950 951 define i64 @test31(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 952 ; ALL-LABEL: @test31 953 ; ALL: pminsw 954 entry: 955 %0 = bitcast <1 x i64> %b to <4 x i16> 956 %1 = bitcast <1 x i64> %a to <4 x i16> 957 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 958 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 959 %2 = tail call x86_mmx @llvm.x86.mmx.pmins.w(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 960 %3 = bitcast x86_mmx %2 to <4 x i16> 961 %4 = bitcast <4 x i16> %3 to <1 x i64> 962 %5 = extractelement <1 x i64> %4, i32 0 963 ret i64 %5 964 } 965 966 declare x86_mmx @llvm.x86.mmx.pminu.b(x86_mmx, x86_mmx) nounwind readnone 967 968 define i64 @test30(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 969 ; ALL-LABEL: @test30 970 ; ALL: pminub 971 entry: 972 %0 = bitcast <1 x i64> %b to <8 x i8> 973 %1 = bitcast <1 x i64> %a to <8 x i8> 974 %mmx_var.i = bitcast <8 x i8> %1 to x86_mmx 975 %mmx_var1.i = bitcast <8 x i8> %0 to x86_mmx 976 %2 = tail call x86_mmx @llvm.x86.mmx.pminu.b(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 977 %3 = bitcast x86_mmx %2 to <8 x i8> 978 %4 = bitcast <8 x i8> %3 to <1 x i64> 979 %5 = extractelement <1 x i64> %4, i32 0 980 ret i64 %5 981 } 982 983 declare x86_mmx @llvm.x86.mmx.pmaxs.w(x86_mmx, x86_mmx) nounwind readnone 984 985 define i64 @test29(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 986 ; ALL-LABEL: @test29 987 ; ALL: pmaxsw 988 entry: 989 %0 = bitcast <1 x i64> %b to <4 x i16> 990 %1 = bitcast <1 x i64> %a to <4 x i16> 991 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 992 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 993 %2 = tail call x86_mmx @llvm.x86.mmx.pmaxs.w(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 994 %3 = bitcast x86_mmx %2 to <4 x i16> 995 %4 = bitcast <4 x i16> %3 to <1 x i64> 996 %5 = extractelement <1 x i64> %4, i32 0 997 ret i64 %5 998 } 999 1000 declare x86_mmx @llvm.x86.mmx.pmaxu.b(x86_mmx, x86_mmx) nounwind readnone 1001 1002 define i64 @test28(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 1003 ; ALL-LABEL: @test28 1004 ; ALL: pmaxub 1005 entry: 1006 %0 = bitcast <1 x i64> %b to <8 x i8> 1007 %1 = bitcast <1 x i64> %a to <8 x i8> 1008 %mmx_var.i = bitcast <8 x i8> %1 to x86_mmx 1009 %mmx_var1.i = bitcast <8 x i8> %0 to x86_mmx 1010 %2 = tail call x86_mmx @llvm.x86.mmx.pmaxu.b(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 1011 %3 = bitcast x86_mmx %2 to <8 x i8> 1012 %4 = bitcast <8 x i8> %3 to <1 x i64> 1013 %5 = extractelement <1 x i64> %4, i32 0 1014 ret i64 %5 1015 } 1016 1017 declare x86_mmx @llvm.x86.mmx.pavg.w(x86_mmx, x86_mmx) nounwind readnone 1018 1019 define i64 @test27(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 1020 ; ALL-LABEL: @test27 1021 ; ALL: pavgw 1022 entry: 1023 %0 = bitcast <1 x i64> %b to <4 x i16> 1024 %1 = bitcast <1 x i64> %a to <4 x i16> 1025 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 1026 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 1027 %2 = tail call x86_mmx @llvm.x86.mmx.pavg.w(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 1028 %3 = bitcast x86_mmx %2 to <4 x i16> 1029 %4 = bitcast <4 x i16> %3 to <1 x i64> 1030 %5 = extractelement <1 x i64> %4, i32 0 1031 ret i64 %5 1032 } 1033 1034 declare x86_mmx @llvm.x86.mmx.pavg.b(x86_mmx, x86_mmx) nounwind readnone 1035 1036 define i64 @test26(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 1037 ; ALL-LABEL: @test26 1038 ; ALL: pavgb 1039 entry: 1040 %0 = bitcast <1 x i64> %b to <8 x i8> 1041 %1 = bitcast <1 x i64> %a to <8 x i8> 1042 %mmx_var.i = bitcast <8 x i8> %1 to x86_mmx 1043 %mmx_var1.i = bitcast <8 x i8> %0 to x86_mmx 1044 %2 = tail call x86_mmx @llvm.x86.mmx.pavg.b(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 1045 %3 = bitcast x86_mmx %2 to <8 x i8> 1046 %4 = bitcast <8 x i8> %3 to <1 x i64> 1047 %5 = extractelement <1 x i64> %4, i32 0 1048 ret i64 %5 1049 } 1050 1051 declare void @llvm.x86.mmx.movnt.dq(x86_mmx*, x86_mmx) nounwind 1052 1053 define void @test25(<1 x i64>* %p, <1 x i64> %a) nounwind optsize ssp { 1054 ; ALL-LABEL: @test25 1055 ; ALL: movntq 1056 entry: 1057 %mmx_ptr_var.i = bitcast <1 x i64>* %p to x86_mmx* 1058 %0 = extractelement <1 x i64> %a, i32 0 1059 %mmx_var.i = bitcast i64 %0 to x86_mmx 1060 tail call void @llvm.x86.mmx.movnt.dq(x86_mmx* %mmx_ptr_var.i, x86_mmx %mmx_var.i) nounwind 1061 ret void 1062 } 1063 1064 declare i32 @llvm.x86.mmx.pmovmskb(x86_mmx) nounwind readnone 1065 1066 define i32 @test24(<1 x i64> %a) nounwind readnone optsize ssp { 1067 ; ALL-LABEL: @test24 1068 ; ALL: pmovmskb 1069 entry: 1070 %0 = bitcast <1 x i64> %a to <8 x i8> 1071 %mmx_var.i = bitcast <8 x i8> %0 to x86_mmx 1072 %1 = tail call i32 @llvm.x86.mmx.pmovmskb(x86_mmx %mmx_var.i) nounwind 1073 ret i32 %1 1074 } 1075 1076 declare void @llvm.x86.mmx.maskmovq(x86_mmx, x86_mmx, i8*) nounwind 1077 1078 define void @test23(<1 x i64> %d, <1 x i64> %n, i8* %p) nounwind optsize ssp { 1079 ; ALL-LABEL: @test23 1080 ; ALL: maskmovq 1081 entry: 1082 %0 = bitcast <1 x i64> %n to <8 x i8> 1083 %1 = bitcast <1 x i64> %d to <8 x i8> 1084 %mmx_var.i = bitcast <8 x i8> %1 to x86_mmx 1085 %mmx_var1.i = bitcast <8 x i8> %0 to x86_mmx 1086 tail call void @llvm.x86.mmx.maskmovq(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i, i8* %p) nounwind 1087 ret void 1088 } 1089 1090 declare x86_mmx @llvm.x86.mmx.pmulhu.w(x86_mmx, x86_mmx) nounwind readnone 1091 1092 define i64 @test22(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 1093 ; ALL-LABEL: @test22 1094 ; ALL: pmulhuw 1095 entry: 1096 %0 = bitcast <1 x i64> %b to <4 x i16> 1097 %1 = bitcast <1 x i64> %a to <4 x i16> 1098 %mmx_var.i = bitcast <4 x i16> %1 to x86_mmx 1099 %mmx_var1.i = bitcast <4 x i16> %0 to x86_mmx 1100 %2 = tail call x86_mmx @llvm.x86.mmx.pmulhu.w(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 1101 %3 = bitcast x86_mmx %2 to <4 x i16> 1102 %4 = bitcast <4 x i16> %3 to <1 x i64> 1103 %5 = extractelement <1 x i64> %4, i32 0 1104 ret i64 %5 1105 } 1106 1107 declare x86_mmx @llvm.x86.sse.pshuf.w(x86_mmx, i8) nounwind readnone 1108 1109 define i64 @test21(<1 x i64> %a) nounwind readnone optsize ssp { 1110 ; ALL-LABEL: @test21 1111 ; X86: pshufw {{.*#+}} mm0 = mem[3,0,0,0] 1112 ; X64: pshufw {{.*#+}} mm0 = mm0[3,0,0,0] 1113 entry: 1114 %0 = bitcast <1 x i64> %a to <4 x i16> 1115 %1 = bitcast <4 x i16> %0 to x86_mmx 1116 %2 = tail call x86_mmx @llvm.x86.sse.pshuf.w(x86_mmx %1, i8 3) nounwind readnone 1117 %3 = bitcast x86_mmx %2 to <4 x i16> 1118 %4 = bitcast <4 x i16> %3 to <1 x i64> 1119 %5 = extractelement <1 x i64> %4, i32 0 1120 ret i64 %5 1121 } 1122 1123 define i32 @test21_2(<1 x i64> %a) nounwind readnone optsize ssp { 1124 ; ALL-LABEL: @test21_2 1125 ; X86: pshufw {{.*#+}} mm0 = mem[3,0,0,0] 1126 ; X64: pshufw {{.*#+}} mm0 = mm0[3,0,0,0] 1127 ; ALL: movd 1128 entry: 1129 %0 = bitcast <1 x i64> %a to <4 x i16> 1130 %1 = bitcast <4 x i16> %0 to x86_mmx 1131 %2 = tail call x86_mmx @llvm.x86.sse.pshuf.w(x86_mmx %1, i8 3) nounwind readnone 1132 %3 = bitcast x86_mmx %2 to <4 x i16> 1133 %4 = bitcast <4 x i16> %3 to <2 x i32> 1134 %5 = extractelement <2 x i32> %4, i32 0 1135 ret i32 %5 1136 } 1137 1138 declare x86_mmx @llvm.x86.mmx.pmulu.dq(x86_mmx, x86_mmx) nounwind readnone 1139 1140 define i64 @test20(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 1141 ; ALL-LABEL: @test20 1142 ; ALL: pmuludq 1143 entry: 1144 %0 = bitcast <1 x i64> %b to <2 x i32> 1145 %1 = bitcast <1 x i64> %a to <2 x i32> 1146 %mmx_var.i = bitcast <2 x i32> %1 to x86_mmx 1147 %mmx_var1.i = bitcast <2 x i32> %0 to x86_mmx 1148 %2 = tail call x86_mmx @llvm.x86.mmx.pmulu.dq(x86_mmx %mmx_var.i, x86_mmx %mmx_var1.i) nounwind 1149 %3 = bitcast x86_mmx %2 to i64 1150 ret i64 %3 1151 } 1152 1153 declare <2 x double> @llvm.x86.sse.cvtpi2pd(x86_mmx) nounwind readnone 1154 1155 define <2 x double> @test19(<1 x i64> %a) nounwind readnone optsize ssp { 1156 ; ALL-LABEL: @test19 1157 ; ALL: cvtpi2pd 1158 entry: 1159 %0 = bitcast <1 x i64> %a to <2 x i32> 1160 %1 = bitcast <2 x i32> %0 to x86_mmx 1161 %2 = tail call <2 x double> @llvm.x86.sse.cvtpi2pd(x86_mmx %1) nounwind readnone 1162 ret <2 x double> %2 1163 } 1164 1165 declare x86_mmx @llvm.x86.sse.cvttpd2pi(<2 x double>) nounwind readnone 1166 1167 define i64 @test18(<2 x double> %a) nounwind readnone optsize ssp { 1168 ; ALL-LABEL: @test18 1169 ; ALL: cvttpd2pi 1170 entry: 1171 %0 = tail call x86_mmx @llvm.x86.sse.cvttpd2pi(<2 x double> %a) nounwind readnone 1172 %1 = bitcast x86_mmx %0 to <2 x i32> 1173 %2 = bitcast <2 x i32> %1 to <1 x i64> 1174 %3 = extractelement <1 x i64> %2, i32 0 1175 ret i64 %3 1176 } 1177 1178 declare x86_mmx @llvm.x86.sse.cvtpd2pi(<2 x double>) nounwind readnone 1179 1180 define i64 @test17(<2 x double> %a) nounwind readnone optsize ssp { 1181 ; ALL-LABEL: @test17 1182 ; ALL: cvtpd2pi 1183 entry: 1184 %0 = tail call x86_mmx @llvm.x86.sse.cvtpd2pi(<2 x double> %a) nounwind readnone 1185 %1 = bitcast x86_mmx %0 to <2 x i32> 1186 %2 = bitcast <2 x i32> %1 to <1 x i64> 1187 %3 = extractelement <1 x i64> %2, i32 0 1188 ret i64 %3 1189 } 1190 1191 declare x86_mmx @llvm.x86.mmx.palignr.b(x86_mmx, x86_mmx, i8) nounwind readnone 1192 1193 define i64 @test16(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 1194 ; ALL-LABEL: @test16 1195 ; ALL: palignr 1196 entry: 1197 %0 = extractelement <1 x i64> %a, i32 0 1198 %mmx_var = bitcast i64 %0 to x86_mmx 1199 %1 = extractelement <1 x i64> %b, i32 0 1200 %mmx_var1 = bitcast i64 %1 to x86_mmx 1201 %2 = tail call x86_mmx @llvm.x86.mmx.palignr.b(x86_mmx %mmx_var, x86_mmx %mmx_var1, i8 16) 1202 %3 = bitcast x86_mmx %2 to i64 1203 ret i64 %3 1204 } 1205 1206 declare x86_mmx @llvm.x86.ssse3.pabs.d(x86_mmx) nounwind readnone 1207 1208 define i64 @test15(<1 x i64> %a) nounwind readnone optsize ssp { 1209 ; ALL-LABEL: @test15 1210 ; ALL: pabsd 1211 entry: 1212 %0 = bitcast <1 x i64> %a to <2 x i32> 1213 %1 = bitcast <2 x i32> %0 to x86_mmx 1214 %2 = tail call x86_mmx @llvm.x86.ssse3.pabs.d(x86_mmx %1) nounwind readnone 1215 %3 = bitcast x86_mmx %2 to <2 x i32> 1216 %4 = bitcast <2 x i32> %3 to <1 x i64> 1217 %5 = extractelement <1 x i64> %4, i32 0 1218 ret i64 %5 1219 } 1220 1221 declare x86_mmx @llvm.x86.ssse3.pabs.w(x86_mmx) nounwind readnone 1222 1223 define i64 @test14(<1 x i64> %a) nounwind readnone optsize ssp { 1224 ; ALL-LABEL: @test14 1225 ; ALL: pabsw 1226 entry: 1227 %0 = bitcast <1 x i64> %a to <4 x i16> 1228 %1 = bitcast <4 x i16> %0 to x86_mmx 1229 %2 = tail call x86_mmx @llvm.x86.ssse3.pabs.w(x86_mmx %1) nounwind readnone 1230 %3 = bitcast x86_mmx %2 to <4 x i16> 1231 %4 = bitcast <4 x i16> %3 to <1 x i64> 1232 %5 = extractelement <1 x i64> %4, i32 0 1233 ret i64 %5 1234 } 1235 1236 declare x86_mmx @llvm.x86.ssse3.pabs.b(x86_mmx) nounwind readnone 1237 1238 define i64 @test13(<1 x i64> %a) nounwind readnone optsize ssp { 1239 ; ALL-LABEL: @test13 1240 ; ALL: pabsb 1241 entry: 1242 %0 = bitcast <1 x i64> %a to <8 x i8> 1243 %1 = bitcast <8 x i8> %0 to x86_mmx 1244 %2 = tail call x86_mmx @llvm.x86.ssse3.pabs.b(x86_mmx %1) nounwind readnone 1245 %3 = bitcast x86_mmx %2 to <8 x i8> 1246 %4 = bitcast <8 x i8> %3 to <1 x i64> 1247 %5 = extractelement <1 x i64> %4, i32 0 1248 ret i64 %5 1249 } 1250 1251 declare x86_mmx @llvm.x86.ssse3.psign.d(x86_mmx, x86_mmx) nounwind readnone 1252 1253 define i64 @test12(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 1254 ; ALL-LABEL: @test12 1255 ; ALL: psignd 1256 entry: 1257 %0 = bitcast <1 x i64> %b to <2 x i32> 1258 %1 = bitcast <1 x i64> %a to <2 x i32> 1259 %2 = bitcast <2 x i32> %1 to x86_mmx 1260 %3 = bitcast <2 x i32> %0 to x86_mmx 1261 %4 = tail call x86_mmx @llvm.x86.ssse3.psign.d(x86_mmx %2, x86_mmx %3) nounwind readnone 1262 %5 = bitcast x86_mmx %4 to <2 x i32> 1263 %6 = bitcast <2 x i32> %5 to <1 x i64> 1264 %7 = extractelement <1 x i64> %6, i32 0 1265 ret i64 %7 1266 } 1267 1268 declare x86_mmx @llvm.x86.ssse3.psign.w(x86_mmx, x86_mmx) nounwind readnone 1269 1270 define i64 @test11(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 1271 ; ALL-LABEL: @test11 1272 ; ALL: psignw 1273 entry: 1274 %0 = bitcast <1 x i64> %b to <4 x i16> 1275 %1 = bitcast <1 x i64> %a to <4 x i16> 1276 %2 = bitcast <4 x i16> %1 to x86_mmx 1277 %3 = bitcast <4 x i16> %0 to x86_mmx 1278 %4 = tail call x86_mmx @llvm.x86.ssse3.psign.w(x86_mmx %2, x86_mmx %3) nounwind readnone 1279 %5 = bitcast x86_mmx %4 to <4 x i16> 1280 %6 = bitcast <4 x i16> %5 to <1 x i64> 1281 %7 = extractelement <1 x i64> %6, i32 0 1282 ret i64 %7 1283 } 1284 1285 declare x86_mmx @llvm.x86.ssse3.psign.b(x86_mmx, x86_mmx) nounwind readnone 1286 1287 define i64 @test10(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 1288 ; ALL-LABEL: @test10 1289 ; ALL: psignb 1290 entry: 1291 %0 = bitcast <1 x i64> %b to <8 x i8> 1292 %1 = bitcast <1 x i64> %a to <8 x i8> 1293 %2 = bitcast <8 x i8> %1 to x86_mmx 1294 %3 = bitcast <8 x i8> %0 to x86_mmx 1295 %4 = tail call x86_mmx @llvm.x86.ssse3.psign.b(x86_mmx %2, x86_mmx %3) nounwind readnone 1296 %5 = bitcast x86_mmx %4 to <8 x i8> 1297 %6 = bitcast <8 x i8> %5 to <1 x i64> 1298 %7 = extractelement <1 x i64> %6, i32 0 1299 ret i64 %7 1300 } 1301 1302 declare x86_mmx @llvm.x86.ssse3.pshuf.b(x86_mmx, x86_mmx) nounwind readnone 1303 1304 define i64 @test9(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 1305 ; ALL-LABEL: @test9 1306 ; ALL: pshufb 1307 entry: 1308 %0 = bitcast <1 x i64> %b to <8 x i8> 1309 %1 = bitcast <1 x i64> %a to <8 x i8> 1310 %2 = bitcast <8 x i8> %1 to x86_mmx 1311 %3 = bitcast <8 x i8> %0 to x86_mmx 1312 %4 = tail call x86_mmx @llvm.x86.ssse3.pshuf.b(x86_mmx %2, x86_mmx %3) nounwind readnone 1313 %5 = bitcast x86_mmx %4 to <8 x i8> 1314 %6 = bitcast <8 x i8> %5 to <1 x i64> 1315 %7 = extractelement <1 x i64> %6, i32 0 1316 ret i64 %7 1317 } 1318 1319 declare x86_mmx @llvm.x86.ssse3.pmul.hr.sw(x86_mmx, x86_mmx) nounwind readnone 1320 1321 define i64 @test8(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 1322 ; ALL-LABEL: @test8 1323 ; ALL: pmulhrsw 1324 entry: 1325 %0 = bitcast <1 x i64> %b to <4 x i16> 1326 %1 = bitcast <1 x i64> %a to <4 x i16> 1327 %2 = bitcast <4 x i16> %1 to x86_mmx 1328 %3 = bitcast <4 x i16> %0 to x86_mmx 1329 %4 = tail call x86_mmx @llvm.x86.ssse3.pmul.hr.sw(x86_mmx %2, x86_mmx %3) nounwind readnone 1330 %5 = bitcast x86_mmx %4 to <4 x i16> 1331 %6 = bitcast <4 x i16> %5 to <1 x i64> 1332 %7 = extractelement <1 x i64> %6, i32 0 1333 ret i64 %7 1334 } 1335 1336 declare x86_mmx @llvm.x86.ssse3.pmadd.ub.sw(x86_mmx, x86_mmx) nounwind readnone 1337 1338 define i64 @test7(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 1339 ; ALL-LABEL: @test7 1340 ; ALL: pmaddubsw 1341 entry: 1342 %0 = bitcast <1 x i64> %b to <8 x i8> 1343 %1 = bitcast <1 x i64> %a to <8 x i8> 1344 %2 = bitcast <8 x i8> %1 to x86_mmx 1345 %3 = bitcast <8 x i8> %0 to x86_mmx 1346 %4 = tail call x86_mmx @llvm.x86.ssse3.pmadd.ub.sw(x86_mmx %2, x86_mmx %3) nounwind readnone 1347 %5 = bitcast x86_mmx %4 to <8 x i8> 1348 %6 = bitcast <8 x i8> %5 to <1 x i64> 1349 %7 = extractelement <1 x i64> %6, i32 0 1350 ret i64 %7 1351 } 1352 1353 declare x86_mmx @llvm.x86.ssse3.phsub.sw(x86_mmx, x86_mmx) nounwind readnone 1354 1355 define i64 @test6(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 1356 ; ALL-LABEL: @test6 1357 ; ALL: phsubsw 1358 entry: 1359 %0 = bitcast <1 x i64> %b to <4 x i16> 1360 %1 = bitcast <1 x i64> %a to <4 x i16> 1361 %2 = bitcast <4 x i16> %1 to x86_mmx 1362 %3 = bitcast <4 x i16> %0 to x86_mmx 1363 %4 = tail call x86_mmx @llvm.x86.ssse3.phsub.sw(x86_mmx %2, x86_mmx %3) nounwind readnone 1364 %5 = bitcast x86_mmx %4 to <4 x i16> 1365 %6 = bitcast <4 x i16> %5 to <1 x i64> 1366 %7 = extractelement <1 x i64> %6, i32 0 1367 ret i64 %7 1368 } 1369 1370 declare x86_mmx @llvm.x86.ssse3.phsub.d(x86_mmx, x86_mmx) nounwind readnone 1371 1372 define i64 @test5(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 1373 ; ALL-LABEL: @test5 1374 ; ALL: phsubd 1375 entry: 1376 %0 = bitcast <1 x i64> %b to <2 x i32> 1377 %1 = bitcast <1 x i64> %a to <2 x i32> 1378 %2 = bitcast <2 x i32> %1 to x86_mmx 1379 %3 = bitcast <2 x i32> %0 to x86_mmx 1380 %4 = tail call x86_mmx @llvm.x86.ssse3.phsub.d(x86_mmx %2, x86_mmx %3) nounwind readnone 1381 %5 = bitcast x86_mmx %4 to <2 x i32> 1382 %6 = bitcast <2 x i32> %5 to <1 x i64> 1383 %7 = extractelement <1 x i64> %6, i32 0 1384 ret i64 %7 1385 } 1386 1387 declare x86_mmx @llvm.x86.ssse3.phsub.w(x86_mmx, x86_mmx) nounwind readnone 1388 1389 define i64 @test4(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 1390 ; ALL-LABEL: @test4 1391 ; ALL: phsubw 1392 entry: 1393 %0 = bitcast <1 x i64> %b to <4 x i16> 1394 %1 = bitcast <1 x i64> %a to <4 x i16> 1395 %2 = bitcast <4 x i16> %1 to x86_mmx 1396 %3 = bitcast <4 x i16> %0 to x86_mmx 1397 %4 = tail call x86_mmx @llvm.x86.ssse3.phsub.w(x86_mmx %2, x86_mmx %3) nounwind readnone 1398 %5 = bitcast x86_mmx %4 to <4 x i16> 1399 %6 = bitcast <4 x i16> %5 to <1 x i64> 1400 %7 = extractelement <1 x i64> %6, i32 0 1401 ret i64 %7 1402 } 1403 1404 declare x86_mmx @llvm.x86.ssse3.phadd.sw(x86_mmx, x86_mmx) nounwind readnone 1405 1406 define i64 @test3(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 1407 ; ALL-LABEL: @test3 1408 ; ALL: phaddsw 1409 entry: 1410 %0 = bitcast <1 x i64> %b to <4 x i16> 1411 %1 = bitcast <1 x i64> %a to <4 x i16> 1412 %2 = bitcast <4 x i16> %1 to x86_mmx 1413 %3 = bitcast <4 x i16> %0 to x86_mmx 1414 %4 = tail call x86_mmx @llvm.x86.ssse3.phadd.sw(x86_mmx %2, x86_mmx %3) nounwind readnone 1415 %5 = bitcast x86_mmx %4 to <4 x i16> 1416 %6 = bitcast <4 x i16> %5 to <1 x i64> 1417 %7 = extractelement <1 x i64> %6, i32 0 1418 ret i64 %7 1419 } 1420 1421 declare x86_mmx @llvm.x86.ssse3.phadd.d(x86_mmx, x86_mmx) nounwind readnone 1422 1423 define i64 @test2(<1 x i64> %a, <1 x i64> %b) nounwind readnone optsize ssp { 1424 ; ALL-LABEL: @test2 1425 ; ALL: phaddd 1426 entry: 1427 %0 = bitcast <1 x i64> %b to <2 x i32> 1428 %1 = bitcast <1 x i64> %a to <2 x i32> 1429 %2 = bitcast <2 x i32> %1 to x86_mmx 1430 %3 = bitcast <2 x i32> %0 to x86_mmx 1431 %4 = tail call x86_mmx @llvm.x86.ssse3.phadd.d(x86_mmx %2, x86_mmx %3) nounwind readnone 1432 %5 = bitcast x86_mmx %4 to <2 x i32> 1433 %6 = bitcast <2 x i32> %5 to <1 x i64> 1434 %7 = extractelement <1 x i64> %6, i32 0 1435 ret i64 %7 1436 } 1437 1438 define <4 x float> @test89(<4 x float> %a, x86_mmx %b) nounwind { 1439 ; ALL-LABEL: @test89 1440 ; ALL: cvtpi2ps 1441 %c = tail call <4 x float> @llvm.x86.sse.cvtpi2ps(<4 x float> %a, x86_mmx %b) 1442 ret <4 x float> %c 1443 } 1444 1445 declare <4 x float> @llvm.x86.sse.cvtpi2ps(<4 x float>, x86_mmx) nounwind readnone 1446 1447 ; ALL-LABEL: test90 1448 define void @test90() { 1449 ; ALL-LABEL: @test90 1450 ; ALL: emms 1451 call void @llvm.x86.mmx.emms() 1452 ret void 1453 } 1454 1455 declare void @llvm.x86.mmx.emms() 1456
