1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512f | FileCheck %s --check-prefix=ALL --check-prefix=KNL 3; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512f,+avx512vl,+avx512bw,+avx512dq,+fast-variable-shuffle | FileCheck %s --check-prefixes=ALL,SKX 4 5 attributes #0 = { nounwind } 6 7define <16 x i8> @trunc_16x32_to_16x8(<16 x i32> %i) #0 { 8; ALL-LABEL: trunc_16x32_to_16x8: 9; ALL: ## %bb.0: 10; ALL-NEXT: vpmovdb %zmm0, %xmm0 11; ALL-NEXT: vzeroupper 12; ALL-NEXT: retq 13 %x = trunc <16 x i32> %i to <16 x i8> 14 ret <16 x i8> %x 15} 16 17define <8 x i16> @trunc_8x64_to_8x16(<8 x i64> %i) #0 { 18; ALL-LABEL: trunc_8x64_to_8x16: 19; ALL: ## %bb.0: 20; ALL-NEXT: vpmovqw %zmm0, %xmm0 21; ALL-NEXT: vzeroupper 22; ALL-NEXT: retq 23 %x = trunc <8 x i64> %i to <8 x i16> 24 ret <8 x i16> %x 25} 26 27define <16 x i16> @trunc_v16i32_to_v16i16(<16 x i32> %x) #0 { 28; ALL-LABEL: trunc_v16i32_to_v16i16: 29; ALL: ## %bb.0: 30; ALL-NEXT: vpmovdw %zmm0, %ymm0 31; ALL-NEXT: retq 32 %1 = trunc <16 x i32> %x to <16 x i16> 33 ret <16 x i16> %1 34} 35 36define <8 x i8> @trunc_qb_512(<8 x i64> %i) #0 { 37; ALL-LABEL: trunc_qb_512: 38; ALL: ## %bb.0: 39; ALL-NEXT: vpmovqw %zmm0, %xmm0 40; ALL-NEXT: vzeroupper 41; ALL-NEXT: retq 42 %x = trunc <8 x i64> %i to <8 x i8> 43 ret <8 x i8> %x 44} 45 46define void @trunc_qb_512_mem(<8 x i64> %i, <8 x i8>* %res) #0 { 47; ALL-LABEL: trunc_qb_512_mem: 48; ALL: ## %bb.0: 49; ALL-NEXT: vpmovqb %zmm0, (%rdi) 50; ALL-NEXT: vzeroupper 51; ALL-NEXT: retq 52 %x = trunc <8 x i64> %i to <8 x i8> 53 store <8 x i8> %x, <8 x i8>* %res 54 ret void 55} 56 57define <4 x i8> @trunc_qb_256(<4 x i64> %i) #0 { 58; KNL-LABEL: trunc_qb_256: 59; KNL: ## %bb.0: 60; KNL-NEXT: ## kill: def $ymm0 killed $ymm0 def $zmm0 61; KNL-NEXT: vpmovqd %zmm0, %ymm0 62; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 killed $ymm0 63; KNL-NEXT: vzeroupper 64; KNL-NEXT: retq 65; 66; SKX-LABEL: trunc_qb_256: 67; SKX: ## %bb.0: 68; SKX-NEXT: vpmovqd %ymm0, %xmm0 69; SKX-NEXT: vzeroupper 70; SKX-NEXT: retq 71 %x = trunc <4 x i64> %i to <4 x i8> 72 ret <4 x i8> %x 73} 74 75define void @trunc_qb_256_mem(<4 x i64> %i, <4 x i8>* %res) #0 { 76; KNL-LABEL: trunc_qb_256_mem: 77; KNL: ## %bb.0: 78; KNL-NEXT: ## kill: def $ymm0 killed $ymm0 def $zmm0 79; KNL-NEXT: vpmovqd %zmm0, %ymm0 80; KNL-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u] 81; KNL-NEXT: vmovd %xmm0, (%rdi) 82; KNL-NEXT: vzeroupper 83; KNL-NEXT: retq 84; 85; SKX-LABEL: trunc_qb_256_mem: 86; SKX: ## %bb.0: 87; SKX-NEXT: vpmovqb %ymm0, (%rdi) 88; SKX-NEXT: vzeroupper 89; SKX-NEXT: retq 90 %x = trunc <4 x i64> %i to <4 x i8> 91 store <4 x i8> %x, <4 x i8>* %res 92 ret void 93} 94 95define <2 x i8> @trunc_qb_128(<2 x i64> %i) #0 { 96; ALL-LABEL: trunc_qb_128: 97; ALL: ## %bb.0: 98; ALL-NEXT: retq 99 %x = trunc <2 x i64> %i to <2 x i8> 100 ret <2 x i8> %x 101} 102 103define void @trunc_qb_128_mem(<2 x i64> %i, <2 x i8>* %res) #0 { 104; KNL-LABEL: trunc_qb_128_mem: 105; KNL: ## %bb.0: 106; KNL-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,8,u,u,u,u,u,u,u,u,u,u,u,u,u,u] 107; KNL-NEXT: vpextrw $0, %xmm0, (%rdi) 108; KNL-NEXT: retq 109; 110; SKX-LABEL: trunc_qb_128_mem: 111; SKX: ## %bb.0: 112; SKX-NEXT: vpmovqb %xmm0, (%rdi) 113; SKX-NEXT: retq 114 %x = trunc <2 x i64> %i to <2 x i8> 115 store <2 x i8> %x, <2 x i8>* %res 116 ret void 117} 118 119define <8 x i16> @trunc_qw_512(<8 x i64> %i) #0 { 120; ALL-LABEL: trunc_qw_512: 121; ALL: ## %bb.0: 122; ALL-NEXT: vpmovqw %zmm0, %xmm0 123; ALL-NEXT: vzeroupper 124; ALL-NEXT: retq 125 %x = trunc <8 x i64> %i to <8 x i16> 126 ret <8 x i16> %x 127} 128 129define void @trunc_qw_512_mem(<8 x i64> %i, <8 x i16>* %res) #0 { 130; ALL-LABEL: trunc_qw_512_mem: 131; ALL: ## %bb.0: 132; ALL-NEXT: vpmovqw %zmm0, (%rdi) 133; ALL-NEXT: vzeroupper 134; ALL-NEXT: retq 135 %x = trunc <8 x i64> %i to <8 x i16> 136 store <8 x i16> %x, <8 x i16>* %res 137 ret void 138} 139 140define <4 x i16> @trunc_qw_256(<4 x i64> %i) #0 { 141; KNL-LABEL: trunc_qw_256: 142; KNL: ## %bb.0: 143; KNL-NEXT: ## kill: def $ymm0 killed $ymm0 def $zmm0 144; KNL-NEXT: vpmovqd %zmm0, %ymm0 145; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 killed $ymm0 146; KNL-NEXT: vzeroupper 147; KNL-NEXT: retq 148; 149; SKX-LABEL: trunc_qw_256: 150; SKX: ## %bb.0: 151; SKX-NEXT: vpmovqd %ymm0, %xmm0 152; SKX-NEXT: vzeroupper 153; SKX-NEXT: retq 154 %x = trunc <4 x i64> %i to <4 x i16> 155 ret <4 x i16> %x 156} 157 158define void @trunc_qw_256_mem(<4 x i64> %i, <4 x i16>* %res) #0 { 159; KNL-LABEL: trunc_qw_256_mem: 160; KNL: ## %bb.0: 161; KNL-NEXT: ## kill: def $ymm0 killed $ymm0 def $zmm0 162; KNL-NEXT: vpmovqd %zmm0, %ymm0 163; KNL-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15] 164; KNL-NEXT: vmovq %xmm0, (%rdi) 165; KNL-NEXT: vzeroupper 166; KNL-NEXT: retq 167; 168; SKX-LABEL: trunc_qw_256_mem: 169; SKX: ## %bb.0: 170; SKX-NEXT: vpmovqw %ymm0, (%rdi) 171; SKX-NEXT: vzeroupper 172; SKX-NEXT: retq 173 %x = trunc <4 x i64> %i to <4 x i16> 174 store <4 x i16> %x, <4 x i16>* %res 175 ret void 176} 177 178define <2 x i16> @trunc_qw_128(<2 x i64> %i) #0 { 179; ALL-LABEL: trunc_qw_128: 180; ALL: ## %bb.0: 181; ALL-NEXT: retq 182 %x = trunc <2 x i64> %i to <2 x i16> 183 ret <2 x i16> %x 184} 185 186define void @trunc_qw_128_mem(<2 x i64> %i, <2 x i16>* %res) #0 { 187; KNL-LABEL: trunc_qw_128_mem: 188; KNL: ## %bb.0: 189; KNL-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,2,3] 190; KNL-NEXT: vpshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7] 191; KNL-NEXT: vmovd %xmm0, (%rdi) 192; KNL-NEXT: retq 193; 194; SKX-LABEL: trunc_qw_128_mem: 195; SKX: ## %bb.0: 196; SKX-NEXT: vpmovqw %xmm0, (%rdi) 197; SKX-NEXT: retq 198 %x = trunc <2 x i64> %i to <2 x i16> 199 store <2 x i16> %x, <2 x i16>* %res 200 ret void 201} 202 203define <8 x i32> @trunc_qd_512(<8 x i64> %i) #0 { 204; ALL-LABEL: trunc_qd_512: 205; ALL: ## %bb.0: 206; ALL-NEXT: vpmovqd %zmm0, %ymm0 207; ALL-NEXT: retq 208 %x = trunc <8 x i64> %i to <8 x i32> 209 ret <8 x i32> %x 210} 211 212define void @trunc_qd_512_mem(<8 x i64> %i, <8 x i32>* %res) #0 { 213; ALL-LABEL: trunc_qd_512_mem: 214; ALL: ## %bb.0: 215; ALL-NEXT: vpmovqd %zmm0, (%rdi) 216; ALL-NEXT: vzeroupper 217; ALL-NEXT: retq 218 %x = trunc <8 x i64> %i to <8 x i32> 219 store <8 x i32> %x, <8 x i32>* %res 220 ret void 221} 222 223define <4 x i32> @trunc_qd_256(<4 x i64> %i) #0 { 224; KNL-LABEL: trunc_qd_256: 225; KNL: ## %bb.0: 226; KNL-NEXT: ## kill: def $ymm0 killed $ymm0 def $zmm0 227; KNL-NEXT: vpmovqd %zmm0, %ymm0 228; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 killed $ymm0 229; KNL-NEXT: vzeroupper 230; KNL-NEXT: retq 231; 232; SKX-LABEL: trunc_qd_256: 233; SKX: ## %bb.0: 234; SKX-NEXT: vpmovqd %ymm0, %xmm0 235; SKX-NEXT: vzeroupper 236; SKX-NEXT: retq 237 %x = trunc <4 x i64> %i to <4 x i32> 238 ret <4 x i32> %x 239} 240 241define void @trunc_qd_256_mem(<4 x i64> %i, <4 x i32>* %res) #0 { 242; KNL-LABEL: trunc_qd_256_mem: 243; KNL: ## %bb.0: 244; KNL-NEXT: ## kill: def $ymm0 killed $ymm0 def $zmm0 245; KNL-NEXT: vpmovqd %zmm0, %ymm0 246; KNL-NEXT: vmovdqa %xmm0, (%rdi) 247; KNL-NEXT: vzeroupper 248; KNL-NEXT: retq 249; 250; SKX-LABEL: trunc_qd_256_mem: 251; SKX: ## %bb.0: 252; SKX-NEXT: vpmovqd %ymm0, (%rdi) 253; SKX-NEXT: vzeroupper 254; SKX-NEXT: retq 255 %x = trunc <4 x i64> %i to <4 x i32> 256 store <4 x i32> %x, <4 x i32>* %res 257 ret void 258} 259 260define <2 x i32> @trunc_qd_128(<2 x i64> %i) #0 { 261; ALL-LABEL: trunc_qd_128: 262; ALL: ## %bb.0: 263; ALL-NEXT: retq 264 %x = trunc <2 x i64> %i to <2 x i32> 265 ret <2 x i32> %x 266} 267 268define void @trunc_qd_128_mem(<2 x i64> %i, <2 x i32>* %res) #0 { 269; KNL-LABEL: trunc_qd_128_mem: 270; KNL: ## %bb.0: 271; KNL-NEXT: vpermilps {{.*#+}} xmm0 = xmm0[0,2,2,3] 272; KNL-NEXT: vmovlps %xmm0, (%rdi) 273; KNL-NEXT: retq 274; 275; SKX-LABEL: trunc_qd_128_mem: 276; SKX: ## %bb.0: 277; SKX-NEXT: vpmovqd %xmm0, (%rdi) 278; SKX-NEXT: retq 279 %x = trunc <2 x i64> %i to <2 x i32> 280 store <2 x i32> %x, <2 x i32>* %res 281 ret void 282} 283 284define <16 x i8> @trunc_db_512(<16 x i32> %i) #0 { 285; ALL-LABEL: trunc_db_512: 286; ALL: ## %bb.0: 287; ALL-NEXT: vpmovdb %zmm0, %xmm0 288; ALL-NEXT: vzeroupper 289; ALL-NEXT: retq 290 %x = trunc <16 x i32> %i to <16 x i8> 291 ret <16 x i8> %x 292} 293 294define void @trunc_db_512_mem(<16 x i32> %i, <16 x i8>* %res) #0 { 295; ALL-LABEL: trunc_db_512_mem: 296; ALL: ## %bb.0: 297; ALL-NEXT: vpmovdb %zmm0, (%rdi) 298; ALL-NEXT: vzeroupper 299; ALL-NEXT: retq 300 %x = trunc <16 x i32> %i to <16 x i8> 301 store <16 x i8> %x, <16 x i8>* %res 302 ret void 303} 304 305define <8 x i8> @trunc_db_256(<8 x i32> %i) #0 { 306; KNL-LABEL: trunc_db_256: 307; KNL: ## %bb.0: 308; KNL-NEXT: ## kill: def $ymm0 killed $ymm0 def $zmm0 309; KNL-NEXT: vpmovdw %zmm0, %ymm0 310; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 killed $ymm0 311; KNL-NEXT: vzeroupper 312; KNL-NEXT: retq 313; 314; SKX-LABEL: trunc_db_256: 315; SKX: ## %bb.0: 316; SKX-NEXT: vpmovdw %ymm0, %xmm0 317; SKX-NEXT: vzeroupper 318; SKX-NEXT: retq 319 %x = trunc <8 x i32> %i to <8 x i8> 320 ret <8 x i8> %x 321} 322 323define void @trunc_db_256_mem(<8 x i32> %i, <8 x i8>* %res) #0 { 324; KNL-LABEL: trunc_db_256_mem: 325; KNL: ## %bb.0: 326; KNL-NEXT: ## kill: def $ymm0 killed $ymm0 def $zmm0 327; KNL-NEXT: vpmovdw %zmm0, %ymm0 328; KNL-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u] 329; KNL-NEXT: vmovq %xmm0, (%rdi) 330; KNL-NEXT: vzeroupper 331; KNL-NEXT: retq 332; 333; SKX-LABEL: trunc_db_256_mem: 334; SKX: ## %bb.0: 335; SKX-NEXT: vpmovdb %ymm0, (%rdi) 336; SKX-NEXT: vzeroupper 337; SKX-NEXT: retq 338 %x = trunc <8 x i32> %i to <8 x i8> 339 store <8 x i8> %x, <8 x i8>* %res 340 ret void 341} 342 343define <4 x i8> @trunc_db_128(<4 x i32> %i) #0 { 344; ALL-LABEL: trunc_db_128: 345; ALL: ## %bb.0: 346; ALL-NEXT: retq 347 %x = trunc <4 x i32> %i to <4 x i8> 348 ret <4 x i8> %x 349} 350 351define void @trunc_db_128_mem(<4 x i32> %i, <4 x i8>* %res) #0 { 352; KNL-LABEL: trunc_db_128_mem: 353; KNL: ## %bb.0: 354; KNL-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,4,8,12,u,u,u,u,u,u,u,u,u,u,u,u] 355; KNL-NEXT: vmovd %xmm0, (%rdi) 356; KNL-NEXT: retq 357; 358; SKX-LABEL: trunc_db_128_mem: 359; SKX: ## %bb.0: 360; SKX-NEXT: vpmovdb %xmm0, (%rdi) 361; SKX-NEXT: retq 362 %x = trunc <4 x i32> %i to <4 x i8> 363 store <4 x i8> %x, <4 x i8>* %res 364 ret void 365} 366 367define <16 x i16> @trunc_dw_512(<16 x i32> %i) #0 { 368; ALL-LABEL: trunc_dw_512: 369; ALL: ## %bb.0: 370; ALL-NEXT: vpmovdw %zmm0, %ymm0 371; ALL-NEXT: retq 372 %x = trunc <16 x i32> %i to <16 x i16> 373 ret <16 x i16> %x 374} 375 376define void @trunc_dw_512_mem(<16 x i32> %i, <16 x i16>* %res) #0 { 377; ALL-LABEL: trunc_dw_512_mem: 378; ALL: ## %bb.0: 379; ALL-NEXT: vpmovdw %zmm0, (%rdi) 380; ALL-NEXT: vzeroupper 381; ALL-NEXT: retq 382 %x = trunc <16 x i32> %i to <16 x i16> 383 store <16 x i16> %x, <16 x i16>* %res 384 ret void 385} 386 387define <8 x i16> @trunc_dw_256(<8 x i32> %i) #0 { 388; KNL-LABEL: trunc_dw_256: 389; KNL: ## %bb.0: 390; KNL-NEXT: ## kill: def $ymm0 killed $ymm0 def $zmm0 391; KNL-NEXT: vpmovdw %zmm0, %ymm0 392; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 killed $ymm0 393; KNL-NEXT: vzeroupper 394; KNL-NEXT: retq 395; 396; SKX-LABEL: trunc_dw_256: 397; SKX: ## %bb.0: 398; SKX-NEXT: vpmovdw %ymm0, %xmm0 399; SKX-NEXT: vzeroupper 400; SKX-NEXT: retq 401 %x = trunc <8 x i32> %i to <8 x i16> 402 ret <8 x i16> %x 403} 404 405define void @trunc_dw_256_mem(<8 x i32> %i, <8 x i16>* %res) #0 { 406; KNL-LABEL: trunc_dw_256_mem: 407; KNL: ## %bb.0: 408; KNL-NEXT: ## kill: def $ymm0 killed $ymm0 def $zmm0 409; KNL-NEXT: vpmovdw %zmm0, %ymm0 410; KNL-NEXT: vmovdqa %xmm0, (%rdi) 411; KNL-NEXT: vzeroupper 412; KNL-NEXT: retq 413; 414; SKX-LABEL: trunc_dw_256_mem: 415; SKX: ## %bb.0: 416; SKX-NEXT: vpmovdw %ymm0, (%rdi) 417; SKX-NEXT: vzeroupper 418; SKX-NEXT: retq 419 %x = trunc <8 x i32> %i to <8 x i16> 420 store <8 x i16> %x, <8 x i16>* %res 421 ret void 422} 423 424define void @trunc_dw_128_mem(<4 x i32> %i, <4 x i16>* %res) #0 { 425; KNL-LABEL: trunc_dw_128_mem: 426; KNL: ## %bb.0: 427; KNL-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,1,4,5,8,9,12,13,8,9,12,13,12,13,14,15] 428; KNL-NEXT: vmovq %xmm0, (%rdi) 429; KNL-NEXT: retq 430; 431; SKX-LABEL: trunc_dw_128_mem: 432; SKX: ## %bb.0: 433; SKX-NEXT: vpmovdw %xmm0, (%rdi) 434; SKX-NEXT: retq 435 %x = trunc <4 x i32> %i to <4 x i16> 436 store <4 x i16> %x, <4 x i16>* %res 437 ret void 438} 439 440define <32 x i8> @trunc_wb_512(<32 x i16> %i) #0 { 441; KNL-LABEL: trunc_wb_512: 442; KNL: ## %bb.0: 443; KNL-NEXT: vpmovsxwd %ymm0, %zmm0 444; KNL-NEXT: vpmovdb %zmm0, %xmm0 445; KNL-NEXT: vpmovsxwd %ymm1, %zmm1 446; KNL-NEXT: vpmovdb %zmm1, %xmm1 447; KNL-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 448; KNL-NEXT: retq 449; 450; SKX-LABEL: trunc_wb_512: 451; SKX: ## %bb.0: 452; SKX-NEXT: vpmovwb %zmm0, %ymm0 453; SKX-NEXT: retq 454 %x = trunc <32 x i16> %i to <32 x i8> 455 ret <32 x i8> %x 456} 457 458define void @trunc_wb_512_mem(<32 x i16> %i, <32 x i8>* %res) #0 { 459; KNL-LABEL: trunc_wb_512_mem: 460; KNL: ## %bb.0: 461; KNL-NEXT: vpmovsxwd %ymm0, %zmm0 462; KNL-NEXT: vpmovdb %zmm0, %xmm0 463; KNL-NEXT: vpmovsxwd %ymm1, %zmm1 464; KNL-NEXT: vpmovdb %zmm1, %xmm1 465; KNL-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 466; KNL-NEXT: vmovdqa %ymm0, (%rdi) 467; KNL-NEXT: vzeroupper 468; KNL-NEXT: retq 469; 470; SKX-LABEL: trunc_wb_512_mem: 471; SKX: ## %bb.0: 472; SKX-NEXT: vpmovwb %zmm0, (%rdi) 473; SKX-NEXT: vzeroupper 474; SKX-NEXT: retq 475 %x = trunc <32 x i16> %i to <32 x i8> 476 store <32 x i8> %x, <32 x i8>* %res 477 ret void 478} 479 480define <16 x i8> @trunc_wb_256(<16 x i16> %i) #0 { 481; KNL-LABEL: trunc_wb_256: 482; KNL: ## %bb.0: 483; KNL-NEXT: vpmovsxwd %ymm0, %zmm0 484; KNL-NEXT: vpmovdb %zmm0, %xmm0 485; KNL-NEXT: vzeroupper 486; KNL-NEXT: retq 487; 488; SKX-LABEL: trunc_wb_256: 489; SKX: ## %bb.0: 490; SKX-NEXT: vpmovwb %ymm0, %xmm0 491; SKX-NEXT: vzeroupper 492; SKX-NEXT: retq 493 %x = trunc <16 x i16> %i to <16 x i8> 494 ret <16 x i8> %x 495} 496 497define void @trunc_wb_256_mem(<16 x i16> %i, <16 x i8>* %res) #0 { 498; KNL-LABEL: trunc_wb_256_mem: 499; KNL: ## %bb.0: 500; KNL-NEXT: vpmovsxwd %ymm0, %zmm0 501; KNL-NEXT: vpmovdb %zmm0, (%rdi) 502; KNL-NEXT: vzeroupper 503; KNL-NEXT: retq 504; 505; SKX-LABEL: trunc_wb_256_mem: 506; SKX: ## %bb.0: 507; SKX-NEXT: vpmovwb %ymm0, (%rdi) 508; SKX-NEXT: vzeroupper 509; SKX-NEXT: retq 510 %x = trunc <16 x i16> %i to <16 x i8> 511 store <16 x i8> %x, <16 x i8>* %res 512 ret void 513} 514 515define <8 x i8> @trunc_wb_128(<8 x i16> %i) #0 { 516; ALL-LABEL: trunc_wb_128: 517; ALL: ## %bb.0: 518; ALL-NEXT: retq 519 %x = trunc <8 x i16> %i to <8 x i8> 520 ret <8 x i8> %x 521} 522 523define void @trunc_wb_128_mem(<8 x i16> %i, <8 x i8>* %res) #0 { 524; KNL-LABEL: trunc_wb_128_mem: 525; KNL: ## %bb.0: 526; KNL-NEXT: vpshufb {{.*#+}} xmm0 = xmm0[0,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u] 527; KNL-NEXT: vmovq %xmm0, (%rdi) 528; KNL-NEXT: retq 529; 530; SKX-LABEL: trunc_wb_128_mem: 531; SKX: ## %bb.0: 532; SKX-NEXT: vpmovwb %xmm0, (%rdi) 533; SKX-NEXT: retq 534 %x = trunc <8 x i16> %i to <8 x i8> 535 store <8 x i8> %x, <8 x i8>* %res 536 ret void 537} 538 539 540define void @usat_trunc_wb_256_mem(<16 x i16> %i, <16 x i8>* %res) { 541; KNL-LABEL: usat_trunc_wb_256_mem: 542; KNL: ## %bb.0: 543; KNL-NEXT: vpminuw {{.*}}(%rip), %ymm0, %ymm0 544; KNL-NEXT: vpmovsxwd %ymm0, %zmm0 545; KNL-NEXT: vpmovdb %zmm0, (%rdi) 546; KNL-NEXT: vzeroupper 547; KNL-NEXT: retq 548; 549; SKX-LABEL: usat_trunc_wb_256_mem: 550; SKX: ## %bb.0: 551; SKX-NEXT: vpmovuswb %ymm0, (%rdi) 552; SKX-NEXT: vzeroupper 553; SKX-NEXT: retq 554 %x3 = icmp ult <16 x i16> %i, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> 555 %x5 = select <16 x i1> %x3, <16 x i16> %i, <16 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> 556 %x6 = trunc <16 x i16> %x5 to <16 x i8> 557 store <16 x i8> %x6, <16 x i8>* %res, align 1 558 ret void 559} 560 561define <16 x i8> @usat_trunc_wb_256(<16 x i16> %i) { 562; KNL-LABEL: usat_trunc_wb_256: 563; KNL: ## %bb.0: 564; KNL-NEXT: vpminuw {{.*}}(%rip), %ymm0, %ymm0 565; KNL-NEXT: vpmovsxwd %ymm0, %zmm0 566; KNL-NEXT: vpmovdb %zmm0, %xmm0 567; KNL-NEXT: vzeroupper 568; KNL-NEXT: retq 569; 570; SKX-LABEL: usat_trunc_wb_256: 571; SKX: ## %bb.0: 572; SKX-NEXT: vpmovuswb %ymm0, %xmm0 573; SKX-NEXT: vzeroupper 574; SKX-NEXT: retq 575 %x3 = icmp ult <16 x i16> %i, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> 576 %x5 = select <16 x i1> %x3, <16 x i16> %i, <16 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> 577 %x6 = trunc <16 x i16> %x5 to <16 x i8> 578 ret <16 x i8> %x6 579} 580 581define void @usat_trunc_wb_128_mem(<8 x i16> %i, <8 x i8>* %res) { 582; KNL-LABEL: usat_trunc_wb_128_mem: 583; KNL: ## %bb.0: 584; KNL-NEXT: vpminuw {{.*}}(%rip), %xmm0, %xmm0 585; KNL-NEXT: vpackuswb %xmm0, %xmm0, %xmm0 586; KNL-NEXT: vmovq %xmm0, (%rdi) 587; KNL-NEXT: retq 588; 589; SKX-LABEL: usat_trunc_wb_128_mem: 590; SKX: ## %bb.0: 591; SKX-NEXT: vpmovuswb %xmm0, (%rdi) 592; SKX-NEXT: retq 593 %x3 = icmp ult <8 x i16> %i, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> 594 %x5 = select <8 x i1> %x3, <8 x i16> %i, <8 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> 595 %x6 = trunc <8 x i16> %x5 to <8 x i8> 596 store <8 x i8> %x6, <8 x i8>* %res, align 1 597 ret void 598} 599 600define void @usat_trunc_db_512_mem(<16 x i32> %i, <16 x i8>* %res) { 601; ALL-LABEL: usat_trunc_db_512_mem: 602; ALL: ## %bb.0: 603; ALL-NEXT: vpmovusdb %zmm0, (%rdi) 604; ALL-NEXT: vzeroupper 605; ALL-NEXT: retq 606 %x3 = icmp ult <16 x i32> %i, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> 607 %x5 = select <16 x i1> %x3, <16 x i32> %i, <16 x i32> <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> 608 %x6 = trunc <16 x i32> %x5 to <16 x i8> 609 store <16 x i8> %x6, <16 x i8>* %res, align 1 610 ret void 611} 612 613define void @usat_trunc_qb_512_mem(<8 x i64> %i, <8 x i8>* %res) { 614; ALL-LABEL: usat_trunc_qb_512_mem: 615; ALL: ## %bb.0: 616; ALL-NEXT: vpmovusqb %zmm0, (%rdi) 617; ALL-NEXT: vzeroupper 618; ALL-NEXT: retq 619 %x3 = icmp ult <8 x i64> %i, <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255> 620 %x5 = select <8 x i1> %x3, <8 x i64> %i, <8 x i64> <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255> 621 %x6 = trunc <8 x i64> %x5 to <8 x i8> 622 store <8 x i8> %x6, <8 x i8>* %res, align 1 623 ret void 624} 625 626define void @usat_trunc_qd_512_mem(<8 x i64> %i, <8 x i32>* %res) { 627; ALL-LABEL: usat_trunc_qd_512_mem: 628; ALL: ## %bb.0: 629; ALL-NEXT: vpmovusqd %zmm0, (%rdi) 630; ALL-NEXT: vzeroupper 631; ALL-NEXT: retq 632 %x3 = icmp ult <8 x i64> %i, <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295> 633 %x5 = select <8 x i1> %x3, <8 x i64> %i, <8 x i64> <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295> 634 %x6 = trunc <8 x i64> %x5 to <8 x i32> 635 store <8 x i32> %x6, <8 x i32>* %res, align 1 636 ret void 637} 638 639define void @usat_trunc_qw_512_mem(<8 x i64> %i, <8 x i16>* %res) { 640; ALL-LABEL: usat_trunc_qw_512_mem: 641; ALL: ## %bb.0: 642; ALL-NEXT: vpmovusqw %zmm0, (%rdi) 643; ALL-NEXT: vzeroupper 644; ALL-NEXT: retq 645 %x3 = icmp ult <8 x i64> %i, <i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535> 646 %x5 = select <8 x i1> %x3, <8 x i64> %i, <8 x i64> <i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535> 647 %x6 = trunc <8 x i64> %x5 to <8 x i16> 648 store <8 x i16> %x6, <8 x i16>* %res, align 1 649 ret void 650} 651 652define <32 x i8> @usat_trunc_db_1024(<32 x i32> %i) { 653; KNL-LABEL: usat_trunc_db_1024: 654; KNL: ## %bb.0: 655; KNL-NEXT: vpmovusdb %zmm0, %xmm0 656; KNL-NEXT: vpmovusdb %zmm1, %xmm1 657; KNL-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 658; KNL-NEXT: retq 659; 660; SKX-LABEL: usat_trunc_db_1024: 661; SKX: ## %bb.0: 662; SKX-NEXT: vpbroadcastd {{.*#+}} zmm2 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255] 663; SKX-NEXT: vpminud %zmm2, %zmm1, %zmm1 664; SKX-NEXT: vpminud %zmm2, %zmm0, %zmm0 665; SKX-NEXT: vpmovdw %zmm0, %ymm0 666; SKX-NEXT: vpmovdw %zmm1, %ymm1 667; SKX-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 668; SKX-NEXT: vpmovwb %zmm0, %ymm0 669; SKX-NEXT: retq 670 %x3 = icmp ult <32 x i32> %i, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> 671 %x5 = select <32 x i1> %x3, <32 x i32> %i, <32 x i32> <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> 672 %x6 = trunc <32 x i32> %x5 to <32 x i8> 673 ret <32 x i8> %x6 674} 675 676define void @usat_trunc_db_1024_mem(<32 x i32> %i, <32 x i8>* %p) { 677; KNL-LABEL: usat_trunc_db_1024_mem: 678; KNL: ## %bb.0: 679; KNL-NEXT: vpmovusdb %zmm0, %xmm0 680; KNL-NEXT: vpmovusdb %zmm1, %xmm1 681; KNL-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 682; KNL-NEXT: vmovdqu %ymm0, (%rdi) 683; KNL-NEXT: vzeroupper 684; KNL-NEXT: retq 685; 686; SKX-LABEL: usat_trunc_db_1024_mem: 687; SKX: ## %bb.0: 688; SKX-NEXT: vpbroadcastd {{.*#+}} zmm2 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255] 689; SKX-NEXT: vpminud %zmm2, %zmm1, %zmm1 690; SKX-NEXT: vpminud %zmm2, %zmm0, %zmm0 691; SKX-NEXT: vpmovdw %zmm0, %ymm0 692; SKX-NEXT: vpmovdw %zmm1, %ymm1 693; SKX-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 694; SKX-NEXT: vpmovwb %zmm0, (%rdi) 695; SKX-NEXT: vzeroupper 696; SKX-NEXT: retq 697 %x3 = icmp ult <32 x i32> %i, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> 698 %x5 = select <32 x i1> %x3, <32 x i32> %i, <32 x i32> <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> 699 %x6 = trunc <32 x i32> %x5 to <32 x i8> 700 store <32 x i8>%x6, <32 x i8>* %p, align 1 701 ret void 702} 703 704define <16 x i16> @usat_trunc_dw_512(<16 x i32> %i) { 705; ALL-LABEL: usat_trunc_dw_512: 706; ALL: ## %bb.0: 707; ALL-NEXT: vpmovusdw %zmm0, %ymm0 708; ALL-NEXT: retq 709 %x3 = icmp ult <16 x i32> %i, <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535> 710 %x5 = select <16 x i1> %x3, <16 x i32> %i, <16 x i32> <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535> 711 %x6 = trunc <16 x i32> %x5 to <16 x i16> 712 ret <16 x i16> %x6 713} 714 715define <8 x i8> @usat_trunc_wb_128(<8 x i16> %i) { 716; ALL-LABEL: usat_trunc_wb_128: 717; ALL: ## %bb.0: 718; ALL-NEXT: vpminuw {{.*}}(%rip), %xmm0, %xmm0 719; ALL-NEXT: retq 720 %x3 = icmp ult <8 x i16> %i, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> 721 %x5 = select <8 x i1> %x3, <8 x i16> %i, <8 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> 722 %x6 = trunc <8 x i16> %x5 to <8 x i8> 723 ret <8 x i8>%x6 724} 725 726define <16 x i16> @usat_trunc_qw_1024(<16 x i64> %i) { 727; ALL-LABEL: usat_trunc_qw_1024: 728; ALL: ## %bb.0: 729; ALL-NEXT: vpbroadcastq {{.*#+}} zmm2 = [65535,65535,65535,65535,65535,65535,65535,65535] 730; ALL-NEXT: vpminuq %zmm2, %zmm1, %zmm1 731; ALL-NEXT: vpminuq %zmm2, %zmm0, %zmm0 732; ALL-NEXT: vpmovqd %zmm0, %ymm0 733; ALL-NEXT: vpmovqd %zmm1, %ymm1 734; ALL-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 735; ALL-NEXT: vpmovdw %zmm0, %ymm0 736; ALL-NEXT: retq 737 %x3 = icmp ult <16 x i64> %i, <i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535> 738 %x5 = select <16 x i1> %x3, <16 x i64> %i, <16 x i64> <i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535> 739 %x6 = trunc <16 x i64> %x5 to <16 x i16> 740 ret <16 x i16> %x6 741} 742 743define <16 x i8> @usat_trunc_db_256(<8 x i32> %x) { 744; KNL-LABEL: usat_trunc_db_256: 745; KNL: ## %bb.0: 746; KNL-NEXT: vpbroadcastd {{.*#+}} ymm1 = [255,255,255,255,255,255,255,255] 747; KNL-NEXT: vpminud %ymm1, %ymm0, %ymm0 748; KNL-NEXT: vpmovdw %zmm0, %ymm0 749; KNL-NEXT: vpackuswb %xmm0, %xmm0, %xmm0 750; KNL-NEXT: vzeroupper 751; KNL-NEXT: retq 752; 753; SKX-LABEL: usat_trunc_db_256: 754; SKX: ## %bb.0: 755; SKX-NEXT: vpminud {{.*}}(%rip){1to8}, %ymm0, %ymm0 756; SKX-NEXT: vpmovdw %ymm0, %xmm0 757; SKX-NEXT: vpackuswb %xmm0, %xmm0, %xmm0 758; SKX-NEXT: vzeroupper 759; SKX-NEXT: retq 760 %tmp1 = icmp ult <8 x i32> %x, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> 761 %tmp2 = select <8 x i1> %tmp1, <8 x i32> %x, <8 x i32> <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> 762 %tmp3 = trunc <8 x i32> %tmp2 to <8 x i8> 763 %tmp4 = shufflevector <8 x i8> %tmp3, <8 x i8> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15> 764 ret <16 x i8> %tmp4 765} 766 767 768 769; Tests for the following unsigned saturation pattern: 770 771; %a = icmp sgt %x, C1 772; %b = select %a, %x, C2 773; %c = icmp slt %b, C2 774; %d = select %c, %b, C2 775; %res = trunc %d 776 777 778define void @smax_usat_trunc_wb_256_mem1(<16 x i16> %i, <16 x i8>* %res) { 779; KNL-LABEL: smax_usat_trunc_wb_256_mem1: 780; KNL: ## %bb.0: 781; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1 782; KNL-NEXT: vpmaxsw %ymm1, %ymm0, %ymm0 783; KNL-NEXT: vpminsw {{.*}}(%rip), %ymm0, %ymm0 784; KNL-NEXT: vpmovsxwd %ymm0, %zmm0 785; KNL-NEXT: vpmovdb %zmm0, (%rdi) 786; KNL-NEXT: vzeroupper 787; KNL-NEXT: retq 788; 789; SKX-LABEL: smax_usat_trunc_wb_256_mem1: 790; SKX: ## %bb.0: 791; SKX-NEXT: vpxor %xmm1, %xmm1, %xmm1 792; SKX-NEXT: vpmaxsw %ymm1, %ymm0, %ymm0 793; SKX-NEXT: vpmovuswb %ymm0, (%rdi) 794; SKX-NEXT: vzeroupper 795; SKX-NEXT: retq 796 %x1 = icmp sgt <16 x i16> %i, <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0> 797 %x2 = select <16 x i1> %x1, <16 x i16> %i, <16 x i16> <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0> 798 %x3 = icmp slt <16 x i16> %x2, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> 799 %x5 = select <16 x i1> %x3, <16 x i16> %x2, <16 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> 800 %x6 = trunc <16 x i16> %x5 to <16 x i8> 801 store <16 x i8> %x6, <16 x i8>* %res, align 1 802 ret void 803} 804 805; Test for smax(smin(x, C2), C1). 806define void @smax_usat_trunc_wb_256_mem2(<16 x i16> %i, <16 x i8>* %res) { 807; KNL-LABEL: smax_usat_trunc_wb_256_mem2: 808; KNL: ## %bb.0: 809; KNL-NEXT: vpminsw {{.*}}(%rip), %ymm0, %ymm0 810; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1 811; KNL-NEXT: vpmaxsw %ymm1, %ymm0, %ymm0 812; KNL-NEXT: vpmovsxwd %ymm0, %zmm0 813; KNL-NEXT: vpmovdb %zmm0, (%rdi) 814; KNL-NEXT: vzeroupper 815; KNL-NEXT: retq 816; 817; SKX-LABEL: smax_usat_trunc_wb_256_mem2: 818; SKX: ## %bb.0: 819; SKX-NEXT: vpxor %xmm1, %xmm1, %xmm1 820; SKX-NEXT: vpmaxsw %ymm1, %ymm0, %ymm0 821; SKX-NEXT: vpmovuswb %ymm0, (%rdi) 822; SKX-NEXT: vzeroupper 823; SKX-NEXT: retq 824 %x1 = icmp slt <16 x i16> %i, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> 825 %x2 = select <16 x i1> %x1, <16 x i16> %i, <16 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> 826 %x3 = icmp sgt <16 x i16> %x2, <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0> 827 %x5 = select <16 x i1> %x3, <16 x i16> %x2, <16 x i16> <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0> 828 %x6 = trunc <16 x i16> %x5 to <16 x i8> 829 store <16 x i8> %x6, <16 x i8>* %res, align 1 830 ret void 831} 832 833define <16 x i8> @smax_usat_trunc_wb_256(<16 x i16> %i) { 834; KNL-LABEL: smax_usat_trunc_wb_256: 835; KNL: ## %bb.0: 836; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1 837; KNL-NEXT: vpmaxsw %ymm1, %ymm0, %ymm0 838; KNL-NEXT: vpminsw {{.*}}(%rip), %ymm0, %ymm0 839; KNL-NEXT: vpmovsxwd %ymm0, %zmm0 840; KNL-NEXT: vpmovdb %zmm0, %xmm0 841; KNL-NEXT: vzeroupper 842; KNL-NEXT: retq 843; 844; SKX-LABEL: smax_usat_trunc_wb_256: 845; SKX: ## %bb.0: 846; SKX-NEXT: vpxor %xmm1, %xmm1, %xmm1 847; SKX-NEXT: vpmaxsw %ymm1, %ymm0, %ymm0 848; SKX-NEXT: vpmovuswb %ymm0, %xmm0 849; SKX-NEXT: vzeroupper 850; SKX-NEXT: retq 851 %x1 = icmp sgt <16 x i16> %i, <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0> 852 %x2 = select <16 x i1> %x1, <16 x i16> %i, <16 x i16> <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0> 853 %x3 = icmp slt <16 x i16> %x2, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> 854 %x5 = select <16 x i1> %x3, <16 x i16> %x2, <16 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> 855 %x6 = trunc <16 x i16> %x5 to <16 x i8> 856 ret <16 x i8> %x6 857 } 858 859define void @smax_usat_trunc_wb_128_mem(<8 x i16> %i, <8 x i8>* %res) { 860; KNL-LABEL: smax_usat_trunc_wb_128_mem: 861; KNL: ## %bb.0: 862; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1 863; KNL-NEXT: vpmaxsw %xmm1, %xmm0, %xmm0 864; KNL-NEXT: vpminsw {{.*}}(%rip), %xmm0, %xmm0 865; KNL-NEXT: vpackuswb %xmm0, %xmm0, %xmm0 866; KNL-NEXT: vmovq %xmm0, (%rdi) 867; KNL-NEXT: retq 868; 869; SKX-LABEL: smax_usat_trunc_wb_128_mem: 870; SKX: ## %bb.0: 871; SKX-NEXT: vpxor %xmm1, %xmm1, %xmm1 872; SKX-NEXT: vpmaxsw %xmm1, %xmm0, %xmm0 873; SKX-NEXT: vpmovuswb %xmm0, (%rdi) 874; SKX-NEXT: retq 875 %x1 = icmp sgt <8 x i16> %i, <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0> 876 %x2 = select <8 x i1> %x1, <8 x i16> %i, <8 x i16> <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0> 877 %x3 = icmp slt <8 x i16> %x2, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> 878 %x5 = select <8 x i1> %x3, <8 x i16> %x2, <8 x i16> <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> 879 %x6 = trunc <8 x i16> %x5 to <8 x i8> 880 store <8 x i8> %x6, <8 x i8>* %res, align 1 881 ret void 882} 883 884define void @smax_usat_trunc_db_512_mem(<16 x i32> %i, <16 x i8>* %res) { 885; ALL-LABEL: smax_usat_trunc_db_512_mem: 886; ALL: ## %bb.0: 887; ALL-NEXT: vpxor %xmm1, %xmm1, %xmm1 888; ALL-NEXT: vpmaxsd %zmm1, %zmm0, %zmm0 889; ALL-NEXT: vpmovusdb %zmm0, (%rdi) 890; ALL-NEXT: vzeroupper 891; ALL-NEXT: retq 892 %x1 = icmp sgt <16 x i32> %i, <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0> 893 %x2 = select <16 x i1> %x1, <16 x i32> %i, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0> 894 %x3 = icmp slt <16 x i32> %x2, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> 895 %x5 = select <16 x i1> %x3, <16 x i32> %x2, <16 x i32> <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> 896 %x6 = trunc <16 x i32> %x5 to <16 x i8> 897 store <16 x i8> %x6, <16 x i8>* %res, align 1 898 ret void 899} 900 901define void @smax_usat_trunc_qb_512_mem(<8 x i64> %i, <8 x i8>* %res) { 902; ALL-LABEL: smax_usat_trunc_qb_512_mem: 903; ALL: ## %bb.0: 904; ALL-NEXT: vpxor %xmm1, %xmm1, %xmm1 905; ALL-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0 906; ALL-NEXT: vpmovusqb %zmm0, (%rdi) 907; ALL-NEXT: vzeroupper 908; ALL-NEXT: retq 909 %x1 = icmp sgt <8 x i64> %i, <i64 0, i64 0, i64 0, i64 0, i64 0, i64 0, i64 0, i64 0> 910 %x2 = select <8 x i1> %x1, <8 x i64> %i, <8 x i64> <i64 0, i64 0, i64 0, i64 0, i64 0, i64 0, i64 0, i64 0> 911 %x3 = icmp slt <8 x i64> %x2, <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255> 912 %x5 = select <8 x i1> %x3, <8 x i64> %x2, <8 x i64> <i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255, i64 255> 913 %x6 = trunc <8 x i64> %x5 to <8 x i8> 914 store <8 x i8> %x6, <8 x i8>* %res, align 1 915 ret void 916} 917 918define void @smax_usat_trunc_qd_512_mem(<8 x i64> %i, <8 x i32>* %res) { 919; ALL-LABEL: smax_usat_trunc_qd_512_mem: 920; ALL: ## %bb.0: 921; ALL-NEXT: vpxor %xmm1, %xmm1, %xmm1 922; ALL-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0 923; ALL-NEXT: vpmovusqd %zmm0, (%rdi) 924; ALL-NEXT: vzeroupper 925; ALL-NEXT: retq 926 %x1 = icmp sgt <8 x i64> %i, <i64 0, i64 0, i64 0, i64 0, i64 0, i64 0, i64 0, i64 0> 927 %x2 = select <8 x i1> %x1, <8 x i64> %i, <8 x i64> <i64 0, i64 0, i64 0, i64 0, i64 0, i64 0, i64 0, i64 0> 928 %x3 = icmp slt <8 x i64> %x2, <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295> 929 %x5 = select <8 x i1> %x3, <8 x i64> %x2, <8 x i64> <i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295, i64 4294967295> 930 %x6 = trunc <8 x i64> %x5 to <8 x i32> 931 store <8 x i32> %x6, <8 x i32>* %res, align 1 932 ret void 933} 934 935define void @smax_usat_trunc_qw_512_mem(<8 x i64> %i, <8 x i16>* %res) { 936; ALL-LABEL: smax_usat_trunc_qw_512_mem: 937; ALL: ## %bb.0: 938; ALL-NEXT: vpxor %xmm1, %xmm1, %xmm1 939; ALL-NEXT: vpmaxsq %zmm1, %zmm0, %zmm0 940; ALL-NEXT: vpmovusqw %zmm0, (%rdi) 941; ALL-NEXT: vzeroupper 942; ALL-NEXT: retq 943 %x1 = icmp sgt <8 x i64> %i, <i64 0, i64 0, i64 0, i64 0, i64 0, i64 0, i64 0, i64 0> 944 %x2 = select <8 x i1> %x1, <8 x i64> %i, <8 x i64> <i64 0, i64 0, i64 0, i64 0, i64 0, i64 0, i64 0, i64 0> 945 %x3 = icmp slt <8 x i64> %x2, <i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535> 946 %x5 = select <8 x i1> %x3, <8 x i64> %x2, <8 x i64> <i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535, i64 65535> 947 %x6 = trunc <8 x i64> %x5 to <8 x i16> 948 store <8 x i16> %x6, <8 x i16>* %res, align 1 949 ret void 950} 951 952define <32 x i8> @smax_usat_trunc_db_1024(<32 x i32> %i) { 953; KNL-LABEL: smax_usat_trunc_db_1024: 954; KNL: ## %bb.0: 955; KNL-NEXT: vpxor %xmm2, %xmm2, %xmm2 956; KNL-NEXT: vpmaxsd %zmm2, %zmm1, %zmm1 957; KNL-NEXT: vpmaxsd %zmm2, %zmm0, %zmm0 958; KNL-NEXT: vpmovusdb %zmm0, %xmm0 959; KNL-NEXT: vpmovusdb %zmm1, %xmm1 960; KNL-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 961; KNL-NEXT: retq 962; 963; SKX-LABEL: smax_usat_trunc_db_1024: 964; SKX: ## %bb.0: 965; SKX-NEXT: vpxor %xmm2, %xmm2, %xmm2 966; SKX-NEXT: vpmaxsd %zmm2, %zmm0, %zmm0 967; SKX-NEXT: vpmaxsd %zmm2, %zmm1, %zmm1 968; SKX-NEXT: vpbroadcastd {{.*#+}} zmm2 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255] 969; SKX-NEXT: vpminsd %zmm2, %zmm1, %zmm1 970; SKX-NEXT: vpminsd %zmm2, %zmm0, %zmm0 971; SKX-NEXT: vpmovdw %zmm0, %ymm0 972; SKX-NEXT: vpmovdw %zmm1, %ymm1 973; SKX-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 974; SKX-NEXT: vpmovwb %zmm0, %ymm0 975; SKX-NEXT: retq 976 %x1 = icmp sgt <32 x i32> %i, <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0> 977 %x2 = select <32 x i1> %x1, <32 x i32> %i, <32 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0> 978 %x3 = icmp slt <32 x i32> %x2, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> 979 %x5 = select <32 x i1> %x3, <32 x i32> %x2, <32 x i32> <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> 980 %x6 = trunc <32 x i32> %x5 to <32 x i8> 981 ret <32 x i8> %x6 982} 983 984define void @smax_usat_trunc_db_1024_mem(<32 x i32> %i, <32 x i8>* %p) { 985; KNL-LABEL: smax_usat_trunc_db_1024_mem: 986; KNL: ## %bb.0: 987; KNL-NEXT: vpxor %xmm2, %xmm2, %xmm2 988; KNL-NEXT: vpmaxsd %zmm2, %zmm1, %zmm1 989; KNL-NEXT: vpmaxsd %zmm2, %zmm0, %zmm0 990; KNL-NEXT: vpmovusdb %zmm0, %xmm0 991; KNL-NEXT: vpmovusdb %zmm1, %xmm1 992; KNL-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0 993; KNL-NEXT: vmovdqu %ymm0, (%rdi) 994; KNL-NEXT: vzeroupper 995; KNL-NEXT: retq 996; 997; SKX-LABEL: smax_usat_trunc_db_1024_mem: 998; SKX: ## %bb.0: 999; SKX-NEXT: vpxor %xmm2, %xmm2, %xmm2 1000; SKX-NEXT: vpmaxsd %zmm2, %zmm0, %zmm0 1001; SKX-NEXT: vpmaxsd %zmm2, %zmm1, %zmm1 1002; SKX-NEXT: vpbroadcastd {{.*#+}} zmm2 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255] 1003; SKX-NEXT: vpminsd %zmm2, %zmm1, %zmm1 1004; SKX-NEXT: vpminsd %zmm2, %zmm0, %zmm0 1005; SKX-NEXT: vpmovdw %zmm0, %ymm0 1006; SKX-NEXT: vpmovdw %zmm1, %ymm1 1007; SKX-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0 1008; SKX-NEXT: vpmovwb %zmm0, (%rdi) 1009; SKX-NEXT: vzeroupper 1010; SKX-NEXT: retq 1011 %x1 = icmp sgt <32 x i32> %i, <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0> 1012 %x2 = select <32 x i1> %x1, <32 x i32> %i, <32 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0> 1013 %x3 = icmp slt <32 x i32> %x2, <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> 1014 %x5 = select <32 x i1> %x3, <32 x i32> %x2, <32 x i32> <i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255, i32 255> 1015 %x6 = trunc <32 x i32> %x5 to <32 x i8> 1016 store <32 x i8>%x6, <32 x i8>* %p, align 1 1017 ret void 1018} 1019 1020define <16 x i16> @smax_usat_trunc_dw_512(<16 x i32> %i) { 1021; ALL-LABEL: smax_usat_trunc_dw_512: 1022; ALL: ## %bb.0: 1023; ALL-NEXT: vpxor %xmm1, %xmm1, %xmm1 1024; ALL-NEXT: vpmaxsd %zmm1, %zmm0, %zmm0 1025; ALL-NEXT: vpmovusdw %zmm0, %ymm0 1026; ALL-NEXT: retq 1027 %x1 = icmp sgt <16 x i32> %i, <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0> 1028 %x2 = select <16 x i1> %x1, <16 x i32> %i, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0> 1029 %x3 = icmp slt <16 x i32> %x2, <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535> 1030 %x5 = select <16 x i1> %x3, <16 x i32> %x2, <16 x i32> <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535> 1031 %x6 = trunc <16 x i32> %x5 to <16 x i16> 1032 ret <16 x i16> %x6 1033} 1034 1035define void @negative_test1_smax_usat_trunc_wb_256_mem(<16 x i16> %i, <16 x i8>* %res) { 1036; KNL-LABEL: negative_test1_smax_usat_trunc_wb_256_mem: 1037; KNL: ## %bb.0: 1038; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1 1039; KNL-NEXT: vpmaxsw %ymm1, %ymm0, %ymm0 1040; KNL-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 1041; KNL-NEXT: vpminsw %ymm1, %ymm0, %ymm0 1042; KNL-NEXT: vpmovsxwd %ymm0, %zmm0 1043; KNL-NEXT: vpmovdb %zmm0, (%rdi) 1044; KNL-NEXT: vzeroupper 1045; KNL-NEXT: retq 1046; 1047; SKX-LABEL: negative_test1_smax_usat_trunc_wb_256_mem: 1048; SKX: ## %bb.0: 1049; SKX-NEXT: vpxor %xmm1, %xmm1, %xmm1 1050; SKX-NEXT: vpmaxsw %ymm1, %ymm0, %ymm0 1051; SKX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 1052; SKX-NEXT: vpminsw %ymm1, %ymm0, %ymm0 1053; SKX-NEXT: vpmovwb %ymm0, (%rdi) 1054; SKX-NEXT: vzeroupper 1055; SKX-NEXT: retq 1056 %x1 = icmp sgt <16 x i16> %i, <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0> 1057 %x2 = select <16 x i1> %x1, <16 x i16> %i, <16 x i16> <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0> 1058 %x3 = icmp slt <16 x i16> %x2, <i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1> 1059 %x5 = select <16 x i1> %x3, <16 x i16> %x2, <16 x i16> <i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1> 1060 %x6 = trunc <16 x i16> %x5 to <16 x i8> 1061 store <16 x i8> %x6, <16 x i8>* %res, align 1 1062 ret void 1063} 1064 1065define void @negative_test2_smax_usat_trunc_wb_256_mem(<16 x i16> %i, <16 x i8>* %res) { 1066; KNL-LABEL: negative_test2_smax_usat_trunc_wb_256_mem: 1067; KNL: ## %bb.0: 1068; KNL-NEXT: vpmaxsw {{.*}}(%rip), %ymm0, %ymm0 1069; KNL-NEXT: vpminsw {{.*}}(%rip), %ymm0, %ymm0 1070; KNL-NEXT: vpmovsxwd %ymm0, %zmm0 1071; KNL-NEXT: vpmovdb %zmm0, (%rdi) 1072; KNL-NEXT: vzeroupper 1073; KNL-NEXT: retq 1074; 1075; SKX-LABEL: negative_test2_smax_usat_trunc_wb_256_mem: 1076; SKX: ## %bb.0: 1077; SKX-NEXT: vpmaxsw {{.*}}(%rip), %ymm0, %ymm0 1078; SKX-NEXT: vpminsw {{.*}}(%rip), %ymm0, %ymm0 1079; SKX-NEXT: vpmovwb %ymm0, (%rdi) 1080; SKX-NEXT: vzeroupper 1081; SKX-NEXT: retq 1082 %x1 = icmp sgt <16 x i16> %i, <i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10> 1083 %x2 = select <16 x i1> %x1, <16 x i16> %i, <16 x i16> <i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10, i16 -10> 1084 %x3 = icmp slt <16 x i16> %x2, <i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5> 1085 %x5 = select <16 x i1> %x3, <16 x i16> %x2, <16 x i16> <i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5, i16 -5> 1086 %x6 = trunc <16 x i16> %x5 to <16 x i8> 1087 store <16 x i8> %x6, <16 x i8>* %res, align 1 1088 ret void 1089} 1090