1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc < %s -mtriple=x86_64-apple-darwin9 -mcpu=knl | FileCheck %s --check-prefix=ALL_X64 --check-prefix=KNL 3; RUN: llc < %s -mtriple=x86_64-apple-darwin9 -mcpu=skx | FileCheck %s --check-prefix=ALL_X64 --check-prefix=SKX 4; RUN: llc < %s -mtriple=i686-apple-darwin9 -mcpu=knl | FileCheck %s --check-prefix=KNL_X32 5 6define <16 x i1> @test1() { 7; ALL_X64-LABEL: test1: 8; ALL_X64: ## %bb.0: 9; ALL_X64-NEXT: vxorps %xmm0, %xmm0, %xmm0 10; ALL_X64-NEXT: retq 11; 12; KNL_X32-LABEL: test1: 13; KNL_X32: ## %bb.0: 14; KNL_X32-NEXT: vxorps %xmm0, %xmm0, %xmm0 15; KNL_X32-NEXT: retl 16 ret <16 x i1> zeroinitializer 17} 18 19define <16 x i1> @test2(<16 x i1>%a, <16 x i1>%b) { 20; ALL_X64-LABEL: test2: 21; ALL_X64: ## %bb.0: 22; ALL_X64-NEXT: vandps %xmm1, %xmm0, %xmm0 23; ALL_X64-NEXT: retq 24; 25; KNL_X32-LABEL: test2: 26; KNL_X32: ## %bb.0: 27; KNL_X32-NEXT: vandps %xmm1, %xmm0, %xmm0 28; KNL_X32-NEXT: retl 29 %c = and <16 x i1>%a, %b 30 ret <16 x i1> %c 31} 32 33define <8 x i1> @test3(<8 x i1>%a, <8 x i1>%b) { 34; ALL_X64-LABEL: test3: 35; ALL_X64: ## %bb.0: 36; ALL_X64-NEXT: vandps %xmm1, %xmm0, %xmm0 37; ALL_X64-NEXT: retq 38; 39; KNL_X32-LABEL: test3: 40; KNL_X32: ## %bb.0: 41; KNL_X32-NEXT: vandps %xmm1, %xmm0, %xmm0 42; KNL_X32-NEXT: retl 43 %c = and <8 x i1>%a, %b 44 ret <8 x i1> %c 45} 46 47define <4 x i1> @test4(<4 x i1>%a, <4 x i1>%b) { 48; ALL_X64-LABEL: test4: 49; ALL_X64: ## %bb.0: 50; ALL_X64-NEXT: vandps %xmm1, %xmm0, %xmm0 51; ALL_X64-NEXT: retq 52; 53; KNL_X32-LABEL: test4: 54; KNL_X32: ## %bb.0: 55; KNL_X32-NEXT: vandps %xmm1, %xmm0, %xmm0 56; KNL_X32-NEXT: retl 57 %c = and <4 x i1>%a, %b 58 ret <4 x i1> %c 59} 60 61declare <8 x i1> @func8xi1(<8 x i1> %a) 62 63define <8 x i32> @test5(<8 x i32>%a, <8 x i32>%b) { 64; KNL-LABEL: test5: 65; KNL: ## %bb.0: 66; KNL-NEXT: pushq %rax 67; KNL-NEXT: .cfi_def_cfa_offset 16 68; KNL-NEXT: vpcmpgtd %ymm1, %ymm0, %ymm0 69; KNL-NEXT: vpmovdw %zmm0, %ymm0 70; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 killed $ymm0 71; KNL-NEXT: callq _func8xi1 72; KNL-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero 73; KNL-NEXT: vpslld $31, %ymm0, %ymm0 74; KNL-NEXT: vpsrad $31, %ymm0, %ymm0 75; KNL-NEXT: popq %rax 76; KNL-NEXT: retq 77; 78; SKX-LABEL: test5: 79; SKX: ## %bb.0: 80; SKX-NEXT: pushq %rax 81; SKX-NEXT: .cfi_def_cfa_offset 16 82; SKX-NEXT: vpcmpgtd %ymm1, %ymm0, %k0 83; SKX-NEXT: vpmovm2w %k0, %xmm0 84; SKX-NEXT: vzeroupper 85; SKX-NEXT: callq _func8xi1 86; SKX-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero 87; SKX-NEXT: vpslld $31, %ymm0, %ymm0 88; SKX-NEXT: vpsrad $31, %ymm0, %ymm0 89; SKX-NEXT: popq %rax 90; SKX-NEXT: retq 91; 92; KNL_X32-LABEL: test5: 93; KNL_X32: ## %bb.0: 94; KNL_X32-NEXT: subl $12, %esp 95; KNL_X32-NEXT: .cfi_def_cfa_offset 16 96; KNL_X32-NEXT: vpcmpgtd %ymm1, %ymm0, %ymm0 97; KNL_X32-NEXT: vpmovdw %zmm0, %ymm0 98; KNL_X32-NEXT: ## kill: def $xmm0 killed $xmm0 killed $ymm0 99; KNL_X32-NEXT: calll _func8xi1 100; KNL_X32-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero 101; KNL_X32-NEXT: vpslld $31, %ymm0, %ymm0 102; KNL_X32-NEXT: vpsrad $31, %ymm0, %ymm0 103; KNL_X32-NEXT: addl $12, %esp 104; KNL_X32-NEXT: retl 105 %cmpRes = icmp sgt <8 x i32>%a, %b 106 %resi = call <8 x i1> @func8xi1(<8 x i1> %cmpRes) 107 %res = sext <8 x i1>%resi to <8 x i32> 108 ret <8 x i32> %res 109} 110 111declare <16 x i1> @func16xi1(<16 x i1> %a) 112 113define <16 x i32> @test6(<16 x i32>%a, <16 x i32>%b) { 114; KNL-LABEL: test6: 115; KNL: ## %bb.0: 116; KNL-NEXT: pushq %rax 117; KNL-NEXT: .cfi_def_cfa_offset 16 118; KNL-NEXT: vpcmpgtd %zmm1, %zmm0, %k1 119; KNL-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z} 120; KNL-NEXT: vpmovdb %zmm0, %xmm0 121; KNL-NEXT: callq _func16xi1 122; KNL-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero 123; KNL-NEXT: vpslld $31, %zmm0, %zmm0 124; KNL-NEXT: vpsrad $31, %zmm0, %zmm0 125; KNL-NEXT: popq %rax 126; KNL-NEXT: retq 127; 128; SKX-LABEL: test6: 129; SKX: ## %bb.0: 130; SKX-NEXT: pushq %rax 131; SKX-NEXT: .cfi_def_cfa_offset 16 132; SKX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 133; SKX-NEXT: vpmovm2b %k0, %xmm0 134; SKX-NEXT: vzeroupper 135; SKX-NEXT: callq _func16xi1 136; SKX-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero 137; SKX-NEXT: vpslld $31, %zmm0, %zmm0 138; SKX-NEXT: vpsrad $31, %zmm0, %zmm0 139; SKX-NEXT: popq %rax 140; SKX-NEXT: retq 141; 142; KNL_X32-LABEL: test6: 143; KNL_X32: ## %bb.0: 144; KNL_X32-NEXT: subl $12, %esp 145; KNL_X32-NEXT: .cfi_def_cfa_offset 16 146; KNL_X32-NEXT: vpcmpgtd %zmm1, %zmm0, %k1 147; KNL_X32-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z} 148; KNL_X32-NEXT: vpmovdb %zmm0, %xmm0 149; KNL_X32-NEXT: calll _func16xi1 150; KNL_X32-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero 151; KNL_X32-NEXT: vpslld $31, %zmm0, %zmm0 152; KNL_X32-NEXT: vpsrad $31, %zmm0, %zmm0 153; KNL_X32-NEXT: addl $12, %esp 154; KNL_X32-NEXT: retl 155 %cmpRes = icmp sgt <16 x i32>%a, %b 156 %resi = call <16 x i1> @func16xi1(<16 x i1> %cmpRes) 157 %res = sext <16 x i1>%resi to <16 x i32> 158 ret <16 x i32> %res 159} 160 161declare <4 x i1> @func4xi1(<4 x i1> %a) 162 163define <4 x i32> @test7(<4 x i32>%a, <4 x i32>%b) { 164; ALL_X64-LABEL: test7: 165; ALL_X64: ## %bb.0: 166; ALL_X64-NEXT: pushq %rax 167; ALL_X64-NEXT: .cfi_def_cfa_offset 16 168; ALL_X64-NEXT: vpcmpgtd %xmm1, %xmm0, %xmm0 169; ALL_X64-NEXT: callq _func4xi1 170; ALL_X64-NEXT: vpslld $31, %xmm0, %xmm0 171; ALL_X64-NEXT: vpsrad $31, %xmm0, %xmm0 172; ALL_X64-NEXT: popq %rax 173; ALL_X64-NEXT: retq 174; 175; KNL_X32-LABEL: test7: 176; KNL_X32: ## %bb.0: 177; KNL_X32-NEXT: subl $12, %esp 178; KNL_X32-NEXT: .cfi_def_cfa_offset 16 179; KNL_X32-NEXT: vpcmpgtd %xmm1, %xmm0, %xmm0 180; KNL_X32-NEXT: calll _func4xi1 181; KNL_X32-NEXT: vpslld $31, %xmm0, %xmm0 182; KNL_X32-NEXT: vpsrad $31, %xmm0, %xmm0 183; KNL_X32-NEXT: addl $12, %esp 184; KNL_X32-NEXT: retl 185 %cmpRes = icmp sgt <4 x i32>%a, %b 186 %resi = call <4 x i1> @func4xi1(<4 x i1> %cmpRes) 187 %res = sext <4 x i1>%resi to <4 x i32> 188 ret <4 x i32> %res 189} 190 191define <8 x i1> @test7a(<8 x i32>%a, <8 x i32>%b) { 192; KNL-LABEL: test7a: 193; KNL: ## %bb.0: 194; KNL-NEXT: pushq %rax 195; KNL-NEXT: .cfi_def_cfa_offset 16 196; KNL-NEXT: vpcmpgtd %ymm1, %ymm0, %ymm0 197; KNL-NEXT: vpmovdw %zmm0, %ymm0 198; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 killed $ymm0 199; KNL-NEXT: callq _func8xi1 200; KNL-NEXT: vandps {{.*}}(%rip), %xmm0, %xmm0 201; KNL-NEXT: popq %rax 202; KNL-NEXT: retq 203; 204; SKX-LABEL: test7a: 205; SKX: ## %bb.0: 206; SKX-NEXT: pushq %rax 207; SKX-NEXT: .cfi_def_cfa_offset 16 208; SKX-NEXT: vpcmpgtd %ymm1, %ymm0, %k0 209; SKX-NEXT: vpmovm2w %k0, %xmm0 210; SKX-NEXT: vzeroupper 211; SKX-NEXT: callq _func8xi1 212; SKX-NEXT: vpand {{.*}}(%rip), %xmm0, %xmm0 213; SKX-NEXT: popq %rax 214; SKX-NEXT: retq 215; 216; KNL_X32-LABEL: test7a: 217; KNL_X32: ## %bb.0: 218; KNL_X32-NEXT: subl $12, %esp 219; KNL_X32-NEXT: .cfi_def_cfa_offset 16 220; KNL_X32-NEXT: vpcmpgtd %ymm1, %ymm0, %ymm0 221; KNL_X32-NEXT: vpmovdw %zmm0, %ymm0 222; KNL_X32-NEXT: ## kill: def $xmm0 killed $xmm0 killed $ymm0 223; KNL_X32-NEXT: calll _func8xi1 224; KNL_X32-NEXT: vandps LCPI7_0, %xmm0, %xmm0 225; KNL_X32-NEXT: addl $12, %esp 226; KNL_X32-NEXT: retl 227 %cmpRes = icmp sgt <8 x i32>%a, %b 228 %resi = call <8 x i1> @func8xi1(<8 x i1> %cmpRes) 229 %res = and <8 x i1>%resi, <i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false> 230 ret <8 x i1> %res 231} 232 233define <16 x i8> @test8(<16 x i8> %a1, <16 x i8> %a2, i1 %cond) { 234; ALL_X64-LABEL: test8: 235; ALL_X64: ## %bb.0: 236; ALL_X64-NEXT: testb $1, %dil 237; ALL_X64-NEXT: jne LBB8_2 238; ALL_X64-NEXT: ## %bb.1: 239; ALL_X64-NEXT: vmovaps %xmm1, %xmm0 240; ALL_X64-NEXT: LBB8_2: 241; ALL_X64-NEXT: retq 242; 243; KNL_X32-LABEL: test8: 244; KNL_X32: ## %bb.0: 245; KNL_X32-NEXT: testb $1, {{[0-9]+}}(%esp) 246; KNL_X32-NEXT: jne LBB8_2 247; KNL_X32-NEXT: ## %bb.1: 248; KNL_X32-NEXT: vmovaps %xmm1, %xmm0 249; KNL_X32-NEXT: LBB8_2: 250; KNL_X32-NEXT: retl 251 %res = select i1 %cond, <16 x i8> %a1, <16 x i8> %a2 252 ret <16 x i8> %res 253} 254 255define i1 @test9(double %a, double %b) { 256; ALL_X64-LABEL: test9: 257; ALL_X64: ## %bb.0: 258; ALL_X64-NEXT: vucomisd %xmm0, %xmm1 259; ALL_X64-NEXT: setb %al 260; ALL_X64-NEXT: retq 261; 262; KNL_X32-LABEL: test9: 263; KNL_X32: ## %bb.0: 264; KNL_X32-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero 265; KNL_X32-NEXT: vucomisd {{[0-9]+}}(%esp), %xmm0 266; KNL_X32-NEXT: setb %al 267; KNL_X32-NEXT: retl 268 %c = fcmp ugt double %a, %b 269 ret i1 %c 270} 271 272define i32 @test10(i32 %a, i32 %b, i1 %cond) { 273; ALL_X64-LABEL: test10: 274; ALL_X64: ## %bb.0: 275; ALL_X64-NEXT: testb $1, %dl 276; ALL_X64-NEXT: cmovel %esi, %edi 277; ALL_X64-NEXT: movl %edi, %eax 278; ALL_X64-NEXT: retq 279; 280; KNL_X32-LABEL: test10: 281; KNL_X32: ## %bb.0: 282; KNL_X32-NEXT: testb $1, {{[0-9]+}}(%esp) 283; KNL_X32-NEXT: leal {{[0-9]+}}(%esp), %eax 284; KNL_X32-NEXT: leal {{[0-9]+}}(%esp), %ecx 285; KNL_X32-NEXT: cmovnel %eax, %ecx 286; KNL_X32-NEXT: movl (%ecx), %eax 287; KNL_X32-NEXT: retl 288 %c = select i1 %cond, i32 %a, i32 %b 289 ret i32 %c 290} 291 292define i1 @test11(i32 %a, i32 %b) { 293; ALL_X64-LABEL: test11: 294; ALL_X64: ## %bb.0: 295; ALL_X64-NEXT: cmpl %esi, %edi 296; ALL_X64-NEXT: setg %al 297; ALL_X64-NEXT: retq 298; 299; KNL_X32-LABEL: test11: 300; KNL_X32: ## %bb.0: 301; KNL_X32-NEXT: movl {{[0-9]+}}(%esp), %eax 302; KNL_X32-NEXT: cmpl {{[0-9]+}}(%esp), %eax 303; KNL_X32-NEXT: setg %al 304; KNL_X32-NEXT: retl 305 %c = icmp sgt i32 %a, %b 306 ret i1 %c 307} 308 309define i32 @test12(i32 %a1, i32 %a2, i32 %b1) { 310; ALL_X64-LABEL: test12: 311; ALL_X64: ## %bb.0: 312; ALL_X64-NEXT: pushq %rbp 313; ALL_X64-NEXT: .cfi_def_cfa_offset 16 314; ALL_X64-NEXT: pushq %r14 315; ALL_X64-NEXT: .cfi_def_cfa_offset 24 316; ALL_X64-NEXT: pushq %rbx 317; ALL_X64-NEXT: .cfi_def_cfa_offset 32 318; ALL_X64-NEXT: .cfi_offset %rbx, -32 319; ALL_X64-NEXT: .cfi_offset %r14, -24 320; ALL_X64-NEXT: .cfi_offset %rbp, -16 321; ALL_X64-NEXT: movl %esi, %r14d 322; ALL_X64-NEXT: movl %edi, %ebp 323; ALL_X64-NEXT: movl %edx, %esi 324; ALL_X64-NEXT: callq _test11 325; ALL_X64-NEXT: movzbl %al, %ebx 326; ALL_X64-NEXT: movl %ebp, %edi 327; ALL_X64-NEXT: movl %r14d, %esi 328; ALL_X64-NEXT: movl %ebx, %edx 329; ALL_X64-NEXT: callq _test10 330; ALL_X64-NEXT: xorl %ecx, %ecx 331; ALL_X64-NEXT: testb $1, %bl 332; ALL_X64-NEXT: cmovel %ecx, %eax 333; ALL_X64-NEXT: popq %rbx 334; ALL_X64-NEXT: popq %r14 335; ALL_X64-NEXT: popq %rbp 336; ALL_X64-NEXT: retq 337; 338; KNL_X32-LABEL: test12: 339; KNL_X32: ## %bb.0: 340; KNL_X32-NEXT: pushl %ebx 341; KNL_X32-NEXT: .cfi_def_cfa_offset 8 342; KNL_X32-NEXT: pushl %edi 343; KNL_X32-NEXT: .cfi_def_cfa_offset 12 344; KNL_X32-NEXT: pushl %esi 345; KNL_X32-NEXT: .cfi_def_cfa_offset 16 346; KNL_X32-NEXT: subl $16, %esp 347; KNL_X32-NEXT: .cfi_def_cfa_offset 32 348; KNL_X32-NEXT: .cfi_offset %esi, -16 349; KNL_X32-NEXT: .cfi_offset %edi, -12 350; KNL_X32-NEXT: .cfi_offset %ebx, -8 351; KNL_X32-NEXT: movl {{[0-9]+}}(%esp), %esi 352; KNL_X32-NEXT: movl {{[0-9]+}}(%esp), %edi 353; KNL_X32-NEXT: movl {{[0-9]+}}(%esp), %eax 354; KNL_X32-NEXT: movl %eax, {{[0-9]+}}(%esp) 355; KNL_X32-NEXT: movl %edi, (%esp) 356; KNL_X32-NEXT: calll _test11 357; KNL_X32-NEXT: movl %eax, %ebx 358; KNL_X32-NEXT: movzbl %al, %eax 359; KNL_X32-NEXT: movl %eax, {{[0-9]+}}(%esp) 360; KNL_X32-NEXT: movl %esi, {{[0-9]+}}(%esp) 361; KNL_X32-NEXT: movl %edi, (%esp) 362; KNL_X32-NEXT: calll _test10 363; KNL_X32-NEXT: xorl %ecx, %ecx 364; KNL_X32-NEXT: testb $1, %bl 365; KNL_X32-NEXT: cmovel %ecx, %eax 366; KNL_X32-NEXT: addl $16, %esp 367; KNL_X32-NEXT: popl %esi 368; KNL_X32-NEXT: popl %edi 369; KNL_X32-NEXT: popl %ebx 370; KNL_X32-NEXT: retl 371 %cond = call i1 @test11(i32 %a1, i32 %b1) 372 %res = call i32 @test10(i32 %a1, i32 %a2, i1 %cond) 373 %res1 = select i1 %cond, i32 %res, i32 0 374 ret i32 %res1 375} 376 377define <1 x i1> @test13(<1 x i1>* %foo) { 378; KNL-LABEL: test13: 379; KNL: ## %bb.0: 380; KNL-NEXT: movzbl (%rdi), %eax 381; KNL-NEXT: ## kill: def $al killed $al killed $eax 382; KNL-NEXT: retq 383; 384; SKX-LABEL: test13: 385; SKX: ## %bb.0: 386; SKX-NEXT: kmovb (%rdi), %k0 387; SKX-NEXT: kmovd %k0, %eax 388; SKX-NEXT: ## kill: def $al killed $al killed $eax 389; SKX-NEXT: retq 390; 391; KNL_X32-LABEL: test13: 392; KNL_X32: ## %bb.0: 393; KNL_X32-NEXT: movl {{[0-9]+}}(%esp), %eax 394; KNL_X32-NEXT: movzbl (%eax), %eax 395; KNL_X32-NEXT: ## kill: def $al killed $al killed $eax 396; KNL_X32-NEXT: retl 397 %bar = load <1 x i1>, <1 x i1>* %foo 398 ret <1 x i1> %bar 399} 400