839 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 | FileCheck %s --check-prefixes=SSE,SSE23; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefixes=SSE,SSE414; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1OR2,AVX15; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX1OR2,AVX26; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefixes=AVX,AVX512,AVX512F7; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefixes=AVX,AVX512,AVX512BW8; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw,+avx512vl | FileCheck %s --check-prefixes=AVX,AVX512,AVX512VL9 10;11; vXi6412;13 14define i1 @test_v2i64(ptr %ptr) nounwind {15; SSE2-LABEL: test_v2i64:16; SSE2: # %bb.0:17; SSE2-NEXT: pcmpeqd %xmm0, %xmm018; SSE2-NEXT: pcmpeqd (%rdi), %xmm019; SSE2-NEXT: movmskps %xmm0, %eax20; SSE2-NEXT: xorl $15, %eax21; SSE2-NEXT: sete %al22; SSE2-NEXT: retq23;24; SSE41-LABEL: test_v2i64:25; SSE41: # %bb.0:26; SSE41-NEXT: movdqa (%rdi), %xmm027; SSE41-NEXT: pcmpeqd %xmm1, %xmm128; SSE41-NEXT: ptest %xmm1, %xmm029; SSE41-NEXT: setb %al30; SSE41-NEXT: retq31;32; AVX-LABEL: test_v2i64:33; AVX: # %bb.0:34; AVX-NEXT: vmovdqa (%rdi), %xmm035; AVX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm136; AVX-NEXT: vptest %xmm1, %xmm037; AVX-NEXT: setb %al38; AVX-NEXT: retq39 %vload = load <2 x i64>, ptr %ptr40 %v0 = extractelement <2 x i64> %vload, i32 041 %v1 = extractelement <2 x i64> %vload, i32 142 %vreduce = and i64 %v0, %v143 %vcheck = icmp eq i64 %vreduce, -144 ret i1 %vcheck45}46 47define i1 @test_v4i64(ptr %ptr) nounwind {48; SSE2-LABEL: test_v4i64:49; SSE2: # %bb.0:50; SSE2-NEXT: movdqa (%rdi), %xmm051; SSE2-NEXT: pand 16(%rdi), %xmm052; SSE2-NEXT: pcmpeqd %xmm1, %xmm153; SSE2-NEXT: pcmpeqd %xmm0, %xmm154; SSE2-NEXT: movmskps %xmm1, %eax55; SSE2-NEXT: xorl $15, %eax56; SSE2-NEXT: sete %al57; SSE2-NEXT: retq58;59; SSE41-LABEL: test_v4i64:60; SSE41: # %bb.0:61; SSE41-NEXT: movdqa (%rdi), %xmm062; SSE41-NEXT: pand 16(%rdi), %xmm063; SSE41-NEXT: pcmpeqd %xmm1, %xmm164; SSE41-NEXT: ptest %xmm1, %xmm065; SSE41-NEXT: setb %al66; SSE41-NEXT: retq67;68; AVX1-LABEL: test_v4i64:69; AVX1: # %bb.0:70; AVX1-NEXT: vmovdqa (%rdi), %ymm071; AVX1-NEXT: vxorps %xmm1, %xmm1, %xmm172; AVX1-NEXT: vcmptrueps %ymm1, %ymm1, %ymm173; AVX1-NEXT: vptest %ymm1, %ymm074; AVX1-NEXT: setb %al75; AVX1-NEXT: vzeroupper76; AVX1-NEXT: retq77;78; AVX2-LABEL: test_v4i64:79; AVX2: # %bb.0:80; AVX2-NEXT: vmovdqa (%rdi), %ymm081; AVX2-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm182; AVX2-NEXT: vptest %ymm1, %ymm083; AVX2-NEXT: setb %al84; AVX2-NEXT: vzeroupper85; AVX2-NEXT: retq86;87; AVX512-LABEL: test_v4i64:88; AVX512: # %bb.0:89; AVX512-NEXT: vmovdqa (%rdi), %ymm090; AVX512-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm191; AVX512-NEXT: vptest %ymm1, %ymm092; AVX512-NEXT: setb %al93; AVX512-NEXT: vzeroupper94; AVX512-NEXT: retq95 %vload = load <4 x i64>, ptr %ptr96 %v0 = extractelement <4 x i64> %vload, i32 097 %v1 = extractelement <4 x i64> %vload, i32 198 %v2 = extractelement <4 x i64> %vload, i32 299 %v3 = extractelement <4 x i64> %vload, i32 3100 %vreduce01 = and i64 %v0, %v1101 %vreduce23 = and i64 %v2, %v3102 %vreduce = and i64 %vreduce01, %vreduce23103 %vcheck = icmp eq i64 %vreduce, -1104 ret i1 %vcheck105}106 107define i1 @test_v8i64(ptr %ptr) nounwind {108; SSE2-LABEL: test_v8i64:109; SSE2: # %bb.0:110; SSE2-NEXT: movdqa (%rdi), %xmm0111; SSE2-NEXT: movdqa 16(%rdi), %xmm1112; SSE2-NEXT: pand 48(%rdi), %xmm1113; SSE2-NEXT: pand 32(%rdi), %xmm0114; SSE2-NEXT: pand %xmm1, %xmm0115; SSE2-NEXT: pcmpeqd %xmm1, %xmm1116; SSE2-NEXT: pcmpeqd %xmm0, %xmm1117; SSE2-NEXT: movmskps %xmm1, %eax118; SSE2-NEXT: xorl $15, %eax119; SSE2-NEXT: sete %al120; SSE2-NEXT: retq121;122; SSE41-LABEL: test_v8i64:123; SSE41: # %bb.0:124; SSE41-NEXT: movdqa (%rdi), %xmm0125; SSE41-NEXT: movdqa 16(%rdi), %xmm1126; SSE41-NEXT: pand 48(%rdi), %xmm1127; SSE41-NEXT: pand 32(%rdi), %xmm0128; SSE41-NEXT: pand %xmm1, %xmm0129; SSE41-NEXT: pcmpeqd %xmm1, %xmm1130; SSE41-NEXT: ptest %xmm1, %xmm0131; SSE41-NEXT: setb %al132; SSE41-NEXT: retq133;134; AVX1-LABEL: test_v8i64:135; AVX1: # %bb.0:136; AVX1-NEXT: vmovaps (%rdi), %ymm0137; AVX1-NEXT: vandps 32(%rdi), %ymm0, %ymm0138; AVX1-NEXT: vxorps %xmm1, %xmm1, %xmm1139; AVX1-NEXT: vcmptrueps %ymm1, %ymm1, %ymm1140; AVX1-NEXT: vptest %ymm1, %ymm0141; AVX1-NEXT: setb %al142; AVX1-NEXT: vzeroupper143; AVX1-NEXT: retq144;145; AVX2-LABEL: test_v8i64:146; AVX2: # %bb.0:147; AVX2-NEXT: vmovdqa (%rdi), %ymm0148; AVX2-NEXT: vpand 32(%rdi), %ymm0, %ymm0149; AVX2-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1150; AVX2-NEXT: vptest %ymm1, %ymm0151; AVX2-NEXT: setb %al152; AVX2-NEXT: vzeroupper153; AVX2-NEXT: retq154;155; AVX512-LABEL: test_v8i64:156; AVX512: # %bb.0:157; AVX512-NEXT: vpternlogd {{.*#+}} zmm0 = -1158; AVX512-NEXT: vpcmpneqd (%rdi), %zmm0, %k0159; AVX512-NEXT: kortestw %k0, %k0160; AVX512-NEXT: sete %al161; AVX512-NEXT: vzeroupper162; AVX512-NEXT: retq163 %vload = load <8 x i64>, ptr %ptr164 %v0 = extractelement <8 x i64> %vload, i32 0165 %v1 = extractelement <8 x i64> %vload, i32 1166 %v2 = extractelement <8 x i64> %vload, i32 2167 %v3 = extractelement <8 x i64> %vload, i32 3168 %v4 = extractelement <8 x i64> %vload, i32 4169 %v5 = extractelement <8 x i64> %vload, i32 5170 %v6 = extractelement <8 x i64> %vload, i32 6171 %v7 = extractelement <8 x i64> %vload, i32 7172 %vreduce01 = and i64 %v0, %v1173 %vreduce23 = and i64 %v2, %v3174 %vreduce45 = and i64 %v4, %v5175 %vreduce67 = and i64 %v6, %v7176 %vreduce0123 = and i64 %vreduce01, %vreduce23177 %vreduce4567 = and i64 %vreduce45, %vreduce67178 %vreduce = and i64 %vreduce0123, %vreduce4567179 %vcheck = icmp eq i64 %vreduce, -1180 ret i1 %vcheck181}182 183define i1 @test_v16i64(ptr %ptr) nounwind {184; SSE2-LABEL: test_v16i64:185; SSE2: # %bb.0:186; SSE2-NEXT: movdqa (%rdi), %xmm0187; SSE2-NEXT: movdqa 16(%rdi), %xmm1188; SSE2-NEXT: movdqa 32(%rdi), %xmm2189; SSE2-NEXT: movdqa 48(%rdi), %xmm3190; SSE2-NEXT: pand 112(%rdi), %xmm3191; SSE2-NEXT: pand 80(%rdi), %xmm1192; SSE2-NEXT: pand %xmm3, %xmm1193; SSE2-NEXT: pand 96(%rdi), %xmm2194; SSE2-NEXT: pand 64(%rdi), %xmm0195; SSE2-NEXT: pand %xmm2, %xmm0196; SSE2-NEXT: pand %xmm1, %xmm0197; SSE2-NEXT: pcmpeqd %xmm1, %xmm1198; SSE2-NEXT: pcmpeqd %xmm0, %xmm1199; SSE2-NEXT: movmskps %xmm1, %eax200; SSE2-NEXT: xorl $15, %eax201; SSE2-NEXT: sete %al202; SSE2-NEXT: retq203;204; SSE41-LABEL: test_v16i64:205; SSE41: # %bb.0:206; SSE41-NEXT: movdqa (%rdi), %xmm0207; SSE41-NEXT: movdqa 16(%rdi), %xmm1208; SSE41-NEXT: movdqa 32(%rdi), %xmm2209; SSE41-NEXT: movdqa 48(%rdi), %xmm3210; SSE41-NEXT: pand 112(%rdi), %xmm3211; SSE41-NEXT: pand 80(%rdi), %xmm1212; SSE41-NEXT: pand %xmm3, %xmm1213; SSE41-NEXT: pand 96(%rdi), %xmm2214; SSE41-NEXT: pand 64(%rdi), %xmm0215; SSE41-NEXT: pand %xmm2, %xmm0216; SSE41-NEXT: pand %xmm1, %xmm0217; SSE41-NEXT: pcmpeqd %xmm1, %xmm1218; SSE41-NEXT: ptest %xmm1, %xmm0219; SSE41-NEXT: setb %al220; SSE41-NEXT: retq221;222; AVX1-LABEL: test_v16i64:223; AVX1: # %bb.0:224; AVX1-NEXT: vmovaps (%rdi), %ymm0225; AVX1-NEXT: vmovaps 32(%rdi), %ymm1226; AVX1-NEXT: vandps 96(%rdi), %ymm1, %ymm1227; AVX1-NEXT: vandps 64(%rdi), %ymm0, %ymm0228; AVX1-NEXT: vandps %ymm1, %ymm0, %ymm0229; AVX1-NEXT: vxorps %xmm1, %xmm1, %xmm1230; AVX1-NEXT: vcmptrueps %ymm1, %ymm1, %ymm1231; AVX1-NEXT: vptest %ymm1, %ymm0232; AVX1-NEXT: setb %al233; AVX1-NEXT: vzeroupper234; AVX1-NEXT: retq235;236; AVX2-LABEL: test_v16i64:237; AVX2: # %bb.0:238; AVX2-NEXT: vmovdqa (%rdi), %ymm0239; AVX2-NEXT: vmovdqa 32(%rdi), %ymm1240; AVX2-NEXT: vpand 96(%rdi), %ymm1, %ymm1241; AVX2-NEXT: vpand 64(%rdi), %ymm0, %ymm0242; AVX2-NEXT: vpand %ymm1, %ymm0, %ymm0243; AVX2-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1244; AVX2-NEXT: vptest %ymm1, %ymm0245; AVX2-NEXT: setb %al246; AVX2-NEXT: vzeroupper247; AVX2-NEXT: retq248;249; AVX512-LABEL: test_v16i64:250; AVX512: # %bb.0:251; AVX512-NEXT: vmovdqa64 (%rdi), %zmm0252; AVX512-NEXT: vpandq 64(%rdi), %zmm0, %zmm0253; AVX512-NEXT: vpternlogd {{.*#+}} zmm1 = -1254; AVX512-NEXT: vpcmpneqd %zmm1, %zmm0, %k0255; AVX512-NEXT: kortestw %k0, %k0256; AVX512-NEXT: sete %al257; AVX512-NEXT: vzeroupper258; AVX512-NEXT: retq259 %vload = load <16 x i64>, ptr %ptr260 %v0 = extractelement <16 x i64> %vload, i32 0261 %v1 = extractelement <16 x i64> %vload, i32 1262 %v2 = extractelement <16 x i64> %vload, i32 2263 %v3 = extractelement <16 x i64> %vload, i32 3264 %v4 = extractelement <16 x i64> %vload, i32 4265 %v5 = extractelement <16 x i64> %vload, i32 5266 %v6 = extractelement <16 x i64> %vload, i32 6267 %v7 = extractelement <16 x i64> %vload, i32 7268 %v8 = extractelement <16 x i64> %vload, i32 8269 %v9 = extractelement <16 x i64> %vload, i32 9270 %v10 = extractelement <16 x i64> %vload, i32 10271 %v11 = extractelement <16 x i64> %vload, i32 11272 %v12 = extractelement <16 x i64> %vload, i32 12273 %v13 = extractelement <16 x i64> %vload, i32 13274 %v14 = extractelement <16 x i64> %vload, i32 14275 %v15 = extractelement <16 x i64> %vload, i32 15276 %vreduce01 = and i64 %v0, %v1277 %vreduce23 = and i64 %v2, %v3278 %vreduce45 = and i64 %v4, %v5279 %vreduce67 = and i64 %v6, %v7280 %vreduce89 = and i64 %v8, %v9281 %vreduce1011 = and i64 %v10, %v11282 %vreduce1213 = and i64 %v12, %v13283 %vreduce1415 = and i64 %v14, %v15284 %vreduce0123 = and i64 %vreduce01, %vreduce23285 %vreduce4567 = and i64 %vreduce45, %vreduce67286 %vreduce891011 = and i64 %vreduce89, %vreduce1011287 %vreduce12131415 = and i64 %vreduce1213, %vreduce1415288 %vreduce01234567 = and i64 %vreduce0123, %vreduce4567289 %vreduce89101112131415 = and i64 %vreduce891011, %vreduce12131415290 %vreduce = and i64 %vreduce01234567, %vreduce89101112131415291 %vcheck = icmp eq i64 %vreduce, -1292 ret i1 %vcheck293}294 295;296; vXi32297;298 299define i1 @test_v2i32(ptr %ptr) nounwind {300; SSE-LABEL: test_v2i32:301; SSE: # %bb.0:302; SSE-NEXT: cmpq $-1, (%rdi)303; SSE-NEXT: sete %al304; SSE-NEXT: retq305;306; AVX-LABEL: test_v2i32:307; AVX: # %bb.0:308; AVX-NEXT: cmpq $-1, (%rdi)309; AVX-NEXT: sete %al310; AVX-NEXT: retq311 %vload = load <2 x i32>, ptr %ptr312 %v0 = extractelement <2 x i32> %vload, i32 0313 %v1 = extractelement <2 x i32> %vload, i32 1314 %vreduce = and i32 %v0, %v1315 %vcheck = icmp eq i32 %vreduce, -1316 ret i1 %vcheck317}318 319define i1 @test_v4i32(ptr %ptr) nounwind {320; SSE2-LABEL: test_v4i32:321; SSE2: # %bb.0:322; SSE2-NEXT: pcmpeqd %xmm0, %xmm0323; SSE2-NEXT: pcmpeqd (%rdi), %xmm0324; SSE2-NEXT: movmskps %xmm0, %eax325; SSE2-NEXT: xorl $15, %eax326; SSE2-NEXT: sete %al327; SSE2-NEXT: retq328;329; SSE41-LABEL: test_v4i32:330; SSE41: # %bb.0:331; SSE41-NEXT: movdqa (%rdi), %xmm0332; SSE41-NEXT: pcmpeqd %xmm1, %xmm1333; SSE41-NEXT: ptest %xmm1, %xmm0334; SSE41-NEXT: setb %al335; SSE41-NEXT: retq336;337; AVX-LABEL: test_v4i32:338; AVX: # %bb.0:339; AVX-NEXT: vmovdqa (%rdi), %xmm0340; AVX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1341; AVX-NEXT: vptest %xmm1, %xmm0342; AVX-NEXT: setb %al343; AVX-NEXT: retq344 %vload = load <4 x i32>, ptr %ptr345 %v0 = extractelement <4 x i32> %vload, i32 0346 %v1 = extractelement <4 x i32> %vload, i32 1347 %v2 = extractelement <4 x i32> %vload, i32 2348 %v3 = extractelement <4 x i32> %vload, i32 3349 %vreduce01 = and i32 %v0, %v1350 %vreduce23 = and i32 %v2, %v3351 %vreduce = and i32 %vreduce01, %vreduce23352 %vcheck = icmp eq i32 %vreduce, -1353 ret i1 %vcheck354}355 356define i1 @test_v8i32(ptr %ptr) nounwind {357; SSE2-LABEL: test_v8i32:358; SSE2: # %bb.0:359; SSE2-NEXT: movdqa (%rdi), %xmm0360; SSE2-NEXT: pand 16(%rdi), %xmm0361; SSE2-NEXT: pcmpeqd %xmm1, %xmm1362; SSE2-NEXT: pcmpeqd %xmm0, %xmm1363; SSE2-NEXT: movmskps %xmm1, %eax364; SSE2-NEXT: xorl $15, %eax365; SSE2-NEXT: sete %al366; SSE2-NEXT: retq367;368; SSE41-LABEL: test_v8i32:369; SSE41: # %bb.0:370; SSE41-NEXT: movdqa (%rdi), %xmm0371; SSE41-NEXT: pand 16(%rdi), %xmm0372; SSE41-NEXT: pcmpeqd %xmm1, %xmm1373; SSE41-NEXT: ptest %xmm1, %xmm0374; SSE41-NEXT: setb %al375; SSE41-NEXT: retq376;377; AVX1-LABEL: test_v8i32:378; AVX1: # %bb.0:379; AVX1-NEXT: vmovdqa (%rdi), %ymm0380; AVX1-NEXT: vxorps %xmm1, %xmm1, %xmm1381; AVX1-NEXT: vcmptrueps %ymm1, %ymm1, %ymm1382; AVX1-NEXT: vptest %ymm1, %ymm0383; AVX1-NEXT: setb %al384; AVX1-NEXT: vzeroupper385; AVX1-NEXT: retq386;387; AVX2-LABEL: test_v8i32:388; AVX2: # %bb.0:389; AVX2-NEXT: vmovdqa (%rdi), %ymm0390; AVX2-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1391; AVX2-NEXT: vptest %ymm1, %ymm0392; AVX2-NEXT: setb %al393; AVX2-NEXT: vzeroupper394; AVX2-NEXT: retq395;396; AVX512-LABEL: test_v8i32:397; AVX512: # %bb.0:398; AVX512-NEXT: vmovdqa (%rdi), %ymm0399; AVX512-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1400; AVX512-NEXT: vptest %ymm1, %ymm0401; AVX512-NEXT: setb %al402; AVX512-NEXT: vzeroupper403; AVX512-NEXT: retq404 %vload = load <8 x i32>, ptr %ptr405 %v0 = extractelement <8 x i32> %vload, i32 0406 %v1 = extractelement <8 x i32> %vload, i32 1407 %v2 = extractelement <8 x i32> %vload, i32 2408 %v3 = extractelement <8 x i32> %vload, i32 3409 %v4 = extractelement <8 x i32> %vload, i32 4410 %v5 = extractelement <8 x i32> %vload, i32 5411 %v6 = extractelement <8 x i32> %vload, i32 6412 %v7 = extractelement <8 x i32> %vload, i32 7413 %vreduce01 = and i32 %v0, %v1414 %vreduce23 = and i32 %v2, %v3415 %vreduce45 = and i32 %v4, %v5416 %vreduce67 = and i32 %v6, %v7417 %vreduce0123 = and i32 %vreduce01, %vreduce23418 %vreduce4567 = and i32 %vreduce45, %vreduce67419 %vreduce = and i32 %vreduce0123, %vreduce4567420 %vcheck = icmp eq i32 %vreduce, -1421 ret i1 %vcheck422}423 424define i1 @test_v16i32(ptr %ptr) nounwind {425; SSE2-LABEL: test_v16i32:426; SSE2: # %bb.0:427; SSE2-NEXT: movdqa (%rdi), %xmm0428; SSE2-NEXT: movdqa 16(%rdi), %xmm1429; SSE2-NEXT: pand 48(%rdi), %xmm1430; SSE2-NEXT: pand 32(%rdi), %xmm0431; SSE2-NEXT: pand %xmm1, %xmm0432; SSE2-NEXT: pcmpeqd %xmm1, %xmm1433; SSE2-NEXT: pcmpeqd %xmm0, %xmm1434; SSE2-NEXT: movmskps %xmm1, %eax435; SSE2-NEXT: xorl $15, %eax436; SSE2-NEXT: sete %al437; SSE2-NEXT: retq438;439; SSE41-LABEL: test_v16i32:440; SSE41: # %bb.0:441; SSE41-NEXT: movdqa (%rdi), %xmm0442; SSE41-NEXT: movdqa 16(%rdi), %xmm1443; SSE41-NEXT: pand 48(%rdi), %xmm1444; SSE41-NEXT: pand 32(%rdi), %xmm0445; SSE41-NEXT: pand %xmm1, %xmm0446; SSE41-NEXT: pcmpeqd %xmm1, %xmm1447; SSE41-NEXT: ptest %xmm1, %xmm0448; SSE41-NEXT: setb %al449; SSE41-NEXT: retq450;451; AVX1-LABEL: test_v16i32:452; AVX1: # %bb.0:453; AVX1-NEXT: vmovaps (%rdi), %ymm0454; AVX1-NEXT: vandps 32(%rdi), %ymm0, %ymm0455; AVX1-NEXT: vxorps %xmm1, %xmm1, %xmm1456; AVX1-NEXT: vcmptrueps %ymm1, %ymm1, %ymm1457; AVX1-NEXT: vptest %ymm1, %ymm0458; AVX1-NEXT: setb %al459; AVX1-NEXT: vzeroupper460; AVX1-NEXT: retq461;462; AVX2-LABEL: test_v16i32:463; AVX2: # %bb.0:464; AVX2-NEXT: vmovdqa (%rdi), %ymm0465; AVX2-NEXT: vpand 32(%rdi), %ymm0, %ymm0466; AVX2-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1467; AVX2-NEXT: vptest %ymm1, %ymm0468; AVX2-NEXT: setb %al469; AVX2-NEXT: vzeroupper470; AVX2-NEXT: retq471;472; AVX512-LABEL: test_v16i32:473; AVX512: # %bb.0:474; AVX512-NEXT: vpternlogd {{.*#+}} zmm0 = -1475; AVX512-NEXT: vpcmpneqd (%rdi), %zmm0, %k0476; AVX512-NEXT: kortestw %k0, %k0477; AVX512-NEXT: sete %al478; AVX512-NEXT: vzeroupper479; AVX512-NEXT: retq480 %vload = load <16 x i32>, ptr %ptr481 %v0 = extractelement <16 x i32> %vload, i32 0482 %v1 = extractelement <16 x i32> %vload, i32 1483 %v2 = extractelement <16 x i32> %vload, i32 2484 %v3 = extractelement <16 x i32> %vload, i32 3485 %v4 = extractelement <16 x i32> %vload, i32 4486 %v5 = extractelement <16 x i32> %vload, i32 5487 %v6 = extractelement <16 x i32> %vload, i32 6488 %v7 = extractelement <16 x i32> %vload, i32 7489 %v8 = extractelement <16 x i32> %vload, i32 8490 %v9 = extractelement <16 x i32> %vload, i32 9491 %v10 = extractelement <16 x i32> %vload, i32 10492 %v11 = extractelement <16 x i32> %vload, i32 11493 %v12 = extractelement <16 x i32> %vload, i32 12494 %v13 = extractelement <16 x i32> %vload, i32 13495 %v14 = extractelement <16 x i32> %vload, i32 14496 %v15 = extractelement <16 x i32> %vload, i32 15497 %vreduce01 = and i32 %v0, %v1498 %vreduce23 = and i32 %v2, %v3499 %vreduce45 = and i32 %v4, %v5500 %vreduce67 = and i32 %v6, %v7501 %vreduce89 = and i32 %v8, %v9502 %vreduce1011 = and i32 %v10, %v11503 %vreduce1213 = and i32 %v12, %v13504 %vreduce1415 = and i32 %v14, %v15505 %vreduce0123 = and i32 %vreduce01, %vreduce23506 %vreduce4567 = and i32 %vreduce45, %vreduce67507 %vreduce891011 = and i32 %vreduce89, %vreduce1011508 %vreduce12131415 = and i32 %vreduce1213, %vreduce1415509 %vreduce01234567 = and i32 %vreduce0123, %vreduce4567510 %vreduce89101112131415 = and i32 %vreduce891011, %vreduce12131415511 %vreduce = and i32 %vreduce01234567, %vreduce89101112131415512 %vcheck = icmp eq i32 %vreduce, -1513 ret i1 %vcheck514}515 516;517; vXi16518;519 520define i1 @test_v2i16(ptr %ptr) nounwind {521; SSE-LABEL: test_v2i16:522; SSE: # %bb.0:523; SSE-NEXT: cmpl $-1, (%rdi)524; SSE-NEXT: sete %al525; SSE-NEXT: retq526;527; AVX-LABEL: test_v2i16:528; AVX: # %bb.0:529; AVX-NEXT: cmpl $-1, (%rdi)530; AVX-NEXT: sete %al531; AVX-NEXT: retq532 %vload = load <2 x i16>, ptr %ptr533 %v0 = extractelement <2 x i16> %vload, i32 0534 %v1 = extractelement <2 x i16> %vload, i32 1535 %vreduce = and i16 %v0, %v1536 %vcheck = icmp eq i16 %vreduce, -1537 ret i1 %vcheck538}539 540define i1 @test_v4i16(ptr %ptr) nounwind {541; SSE-LABEL: test_v4i16:542; SSE: # %bb.0:543; SSE-NEXT: cmpq $-1, (%rdi)544; SSE-NEXT: sete %al545; SSE-NEXT: retq546;547; AVX-LABEL: test_v4i16:548; AVX: # %bb.0:549; AVX-NEXT: cmpq $-1, (%rdi)550; AVX-NEXT: sete %al551; AVX-NEXT: retq552 %vload = load <4 x i16>, ptr %ptr553 %v0 = extractelement <4 x i16> %vload, i32 0554 %v1 = extractelement <4 x i16> %vload, i32 1555 %v2 = extractelement <4 x i16> %vload, i32 2556 %v3 = extractelement <4 x i16> %vload, i32 3557 %vreduce01 = and i16 %v0, %v1558 %vreduce23 = and i16 %v2, %v3559 %vreduce = and i16 %vreduce01, %vreduce23560 %vcheck = icmp eq i16 %vreduce, -1561 ret i1 %vcheck562}563 564define i1 @test_v8i16(ptr %ptr) nounwind {565; SSE2-LABEL: test_v8i16:566; SSE2: # %bb.0:567; SSE2-NEXT: pcmpeqd %xmm0, %xmm0568; SSE2-NEXT: pcmpeqb (%rdi), %xmm0569; SSE2-NEXT: pmovmskb %xmm0, %eax570; SSE2-NEXT: xorl $65535, %eax # imm = 0xFFFF571; SSE2-NEXT: sete %al572; SSE2-NEXT: retq573;574; SSE41-LABEL: test_v8i16:575; SSE41: # %bb.0:576; SSE41-NEXT: movdqa (%rdi), %xmm0577; SSE41-NEXT: pcmpeqd %xmm1, %xmm1578; SSE41-NEXT: ptest %xmm1, %xmm0579; SSE41-NEXT: setb %al580; SSE41-NEXT: retq581;582; AVX-LABEL: test_v8i16:583; AVX: # %bb.0:584; AVX-NEXT: vmovdqa (%rdi), %xmm0585; AVX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1586; AVX-NEXT: vptest %xmm1, %xmm0587; AVX-NEXT: setb %al588; AVX-NEXT: retq589 %vload = load <8 x i16>, ptr %ptr590 %v0 = extractelement <8 x i16> %vload, i32 0591 %v1 = extractelement <8 x i16> %vload, i32 1592 %v2 = extractelement <8 x i16> %vload, i32 2593 %v3 = extractelement <8 x i16> %vload, i32 3594 %v4 = extractelement <8 x i16> %vload, i32 4595 %v5 = extractelement <8 x i16> %vload, i32 5596 %v6 = extractelement <8 x i16> %vload, i32 6597 %v7 = extractelement <8 x i16> %vload, i32 7598 %vreduce01 = and i16 %v0, %v1599 %vreduce23 = and i16 %v2, %v3600 %vreduce45 = and i16 %v4, %v5601 %vreduce67 = and i16 %v6, %v7602 %vreduce0123 = and i16 %vreduce01, %vreduce23603 %vreduce4567 = and i16 %vreduce45, %vreduce67604 %vreduce = and i16 %vreduce0123, %vreduce4567605 %vcheck = icmp eq i16 %vreduce, -1606 ret i1 %vcheck607}608 609define i1 @test_v16i16(ptr %ptr) nounwind {610; SSE2-LABEL: test_v16i16:611; SSE2: # %bb.0:612; SSE2-NEXT: movdqa (%rdi), %xmm0613; SSE2-NEXT: pand 16(%rdi), %xmm0614; SSE2-NEXT: pcmpeqd %xmm1, %xmm1615; SSE2-NEXT: pcmpeqb %xmm0, %xmm1616; SSE2-NEXT: pmovmskb %xmm1, %eax617; SSE2-NEXT: xorl $65535, %eax # imm = 0xFFFF618; SSE2-NEXT: sete %al619; SSE2-NEXT: retq620;621; SSE41-LABEL: test_v16i16:622; SSE41: # %bb.0:623; SSE41-NEXT: movdqa (%rdi), %xmm0624; SSE41-NEXT: pand 16(%rdi), %xmm0625; SSE41-NEXT: pcmpeqd %xmm1, %xmm1626; SSE41-NEXT: ptest %xmm1, %xmm0627; SSE41-NEXT: setb %al628; SSE41-NEXT: retq629;630; AVX1-LABEL: test_v16i16:631; AVX1: # %bb.0:632; AVX1-NEXT: vmovdqa (%rdi), %ymm0633; AVX1-NEXT: vxorps %xmm1, %xmm1, %xmm1634; AVX1-NEXT: vcmptrueps %ymm1, %ymm1, %ymm1635; AVX1-NEXT: vptest %ymm1, %ymm0636; AVX1-NEXT: setb %al637; AVX1-NEXT: vzeroupper638; AVX1-NEXT: retq639;640; AVX2-LABEL: test_v16i16:641; AVX2: # %bb.0:642; AVX2-NEXT: vmovdqa (%rdi), %ymm0643; AVX2-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1644; AVX2-NEXT: vptest %ymm1, %ymm0645; AVX2-NEXT: setb %al646; AVX2-NEXT: vzeroupper647; AVX2-NEXT: retq648;649; AVX512-LABEL: test_v16i16:650; AVX512: # %bb.0:651; AVX512-NEXT: vmovdqa (%rdi), %ymm0652; AVX512-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1653; AVX512-NEXT: vptest %ymm1, %ymm0654; AVX512-NEXT: setb %al655; AVX512-NEXT: vzeroupper656; AVX512-NEXT: retq657 %vload = load <16 x i16>, ptr %ptr658 %v0 = extractelement <16 x i16> %vload, i32 0659 %v1 = extractelement <16 x i16> %vload, i32 1660 %v2 = extractelement <16 x i16> %vload, i32 2661 %v3 = extractelement <16 x i16> %vload, i32 3662 %v4 = extractelement <16 x i16> %vload, i32 4663 %v5 = extractelement <16 x i16> %vload, i32 5664 %v6 = extractelement <16 x i16> %vload, i32 6665 %v7 = extractelement <16 x i16> %vload, i32 7666 %v8 = extractelement <16 x i16> %vload, i32 8667 %v9 = extractelement <16 x i16> %vload, i32 9668 %v10 = extractelement <16 x i16> %vload, i32 10669 %v11 = extractelement <16 x i16> %vload, i32 11670 %v12 = extractelement <16 x i16> %vload, i32 12671 %v13 = extractelement <16 x i16> %vload, i32 13672 %v14 = extractelement <16 x i16> %vload, i32 14673 %v15 = extractelement <16 x i16> %vload, i32 15674 %vreduce01 = and i16 %v0, %v1675 %vreduce23 = and i16 %v2, %v3676 %vreduce45 = and i16 %v4, %v5677 %vreduce67 = and i16 %v6, %v7678 %vreduce89 = and i16 %v8, %v9679 %vreduce1011 = and i16 %v10, %v11680 %vreduce1213 = and i16 %v12, %v13681 %vreduce1415 = and i16 %v14, %v15682 %vreduce0123 = and i16 %vreduce01, %vreduce23683 %vreduce4567 = and i16 %vreduce45, %vreduce67684 %vreduce891011 = and i16 %vreduce89, %vreduce1011685 %vreduce12131415 = and i16 %vreduce1213, %vreduce1415686 %vreduce01234567 = and i16 %vreduce0123, %vreduce4567687 %vreduce89101112131415 = and i16 %vreduce891011, %vreduce12131415688 %vreduce = and i16 %vreduce01234567, %vreduce89101112131415689 %vcheck = icmp eq i16 %vreduce, -1690 ret i1 %vcheck691}692 693;694; vXi8695;696 697define i1 @test_v2i8(ptr %ptr) nounwind {698; SSE-LABEL: test_v2i8:699; SSE: # %bb.0:700; SSE-NEXT: cmpw $-1, (%rdi)701; SSE-NEXT: sete %al702; SSE-NEXT: retq703;704; AVX-LABEL: test_v2i8:705; AVX: # %bb.0:706; AVX-NEXT: cmpw $-1, (%rdi)707; AVX-NEXT: sete %al708; AVX-NEXT: retq709 %vload = load <2 x i8>, ptr %ptr710 %v0 = extractelement <2 x i8> %vload, i32 0711 %v1 = extractelement <2 x i8> %vload, i32 1712 %vreduce = and i8 %v0, %v1713 %vcheck = icmp eq i8 %vreduce, -1714 ret i1 %vcheck715}716 717define i1 @test_v4i8(ptr %ptr) nounwind {718; SSE-LABEL: test_v4i8:719; SSE: # %bb.0:720; SSE-NEXT: cmpl $-1, (%rdi)721; SSE-NEXT: sete %al722; SSE-NEXT: retq723;724; AVX-LABEL: test_v4i8:725; AVX: # %bb.0:726; AVX-NEXT: cmpl $-1, (%rdi)727; AVX-NEXT: sete %al728; AVX-NEXT: retq729 %vload = load <4 x i8>, ptr %ptr730 %v0 = extractelement <4 x i8> %vload, i32 0731 %v1 = extractelement <4 x i8> %vload, i32 1732 %v2 = extractelement <4 x i8> %vload, i32 2733 %v3 = extractelement <4 x i8> %vload, i32 3734 %vreduce01 = and i8 %v0, %v1735 %vreduce23 = and i8 %v2, %v3736 %vreduce = and i8 %vreduce01, %vreduce23737 %vcheck = icmp eq i8 %vreduce, -1738 ret i1 %vcheck739}740 741define i1 @test_v8i8(ptr %ptr) nounwind {742; SSE-LABEL: test_v8i8:743; SSE: # %bb.0:744; SSE-NEXT: cmpq $-1, (%rdi)745; SSE-NEXT: sete %al746; SSE-NEXT: retq747;748; AVX-LABEL: test_v8i8:749; AVX: # %bb.0:750; AVX-NEXT: cmpq $-1, (%rdi)751; AVX-NEXT: sete %al752; AVX-NEXT: retq753 %vload = load <8 x i8>, ptr %ptr754 %v0 = extractelement <8 x i8> %vload, i32 0755 %v1 = extractelement <8 x i8> %vload, i32 1756 %v2 = extractelement <8 x i8> %vload, i32 2757 %v3 = extractelement <8 x i8> %vload, i32 3758 %v4 = extractelement <8 x i8> %vload, i32 4759 %v5 = extractelement <8 x i8> %vload, i32 5760 %v6 = extractelement <8 x i8> %vload, i32 6761 %v7 = extractelement <8 x i8> %vload, i32 7762 %vreduce01 = and i8 %v0, %v1763 %vreduce23 = and i8 %v2, %v3764 %vreduce45 = and i8 %v4, %v5765 %vreduce67 = and i8 %v6, %v7766 %vreduce0123 = and i8 %vreduce01, %vreduce23767 %vreduce4567 = and i8 %vreduce45, %vreduce67768 %vreduce = and i8 %vreduce0123, %vreduce4567769 %vcheck = icmp eq i8 %vreduce, -1770 ret i1 %vcheck771}772 773define i1 @test_v16i8(ptr %ptr) nounwind {774; SSE2-LABEL: test_v16i8:775; SSE2: # %bb.0:776; SSE2-NEXT: pcmpeqd %xmm0, %xmm0777; SSE2-NEXT: pcmpeqb (%rdi), %xmm0778; SSE2-NEXT: pmovmskb %xmm0, %eax779; SSE2-NEXT: xorl $65535, %eax # imm = 0xFFFF780; SSE2-NEXT: sete %al781; SSE2-NEXT: retq782;783; SSE41-LABEL: test_v16i8:784; SSE41: # %bb.0:785; SSE41-NEXT: movdqa (%rdi), %xmm0786; SSE41-NEXT: pcmpeqd %xmm1, %xmm1787; SSE41-NEXT: ptest %xmm1, %xmm0788; SSE41-NEXT: setb %al789; SSE41-NEXT: retq790;791; AVX-LABEL: test_v16i8:792; AVX: # %bb.0:793; AVX-NEXT: vmovdqa (%rdi), %xmm0794; AVX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1795; AVX-NEXT: vptest %xmm1, %xmm0796; AVX-NEXT: setb %al797; AVX-NEXT: retq798 %vload = load <16 x i8>, ptr %ptr799 %v0 = extractelement <16 x i8> %vload, i32 0800 %v1 = extractelement <16 x i8> %vload, i32 1801 %v2 = extractelement <16 x i8> %vload, i32 2802 %v3 = extractelement <16 x i8> %vload, i32 3803 %v4 = extractelement <16 x i8> %vload, i32 4804 %v5 = extractelement <16 x i8> %vload, i32 5805 %v6 = extractelement <16 x i8> %vload, i32 6806 %v7 = extractelement <16 x i8> %vload, i32 7807 %v8 = extractelement <16 x i8> %vload, i32 8808 %v9 = extractelement <16 x i8> %vload, i32 9809 %v10 = extractelement <16 x i8> %vload, i32 10810 %v11 = extractelement <16 x i8> %vload, i32 11811 %v12 = extractelement <16 x i8> %vload, i32 12812 %v13 = extractelement <16 x i8> %vload, i32 13813 %v14 = extractelement <16 x i8> %vload, i32 14814 %v15 = extractelement <16 x i8> %vload, i32 15815 %vreduce01 = and i8 %v0, %v1816 %vreduce23 = and i8 %v2, %v3817 %vreduce45 = and i8 %v4, %v5818 %vreduce67 = and i8 %v6, %v7819 %vreduce89 = and i8 %v8, %v9820 %vreduce1011 = and i8 %v10, %v11821 %vreduce1213 = and i8 %v12, %v13822 %vreduce1415 = and i8 %v14, %v15823 %vreduce0123 = and i8 %vreduce01, %vreduce23824 %vreduce4567 = and i8 %vreduce45, %vreduce67825 %vreduce891011 = and i8 %vreduce89, %vreduce1011826 %vreduce12131415 = and i8 %vreduce1213, %vreduce1415827 %vreduce01234567 = and i8 %vreduce0123, %vreduce4567828 %vreduce89101112131415 = and i8 %vreduce891011, %vreduce12131415829 %vreduce = and i8 %vreduce01234567, %vreduce89101112131415830 %vcheck = icmp eq i8 %vreduce, -1831 ret i1 %vcheck832}833 834;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:835; AVX1OR2: {{.*}}836; AVX512BW: {{.*}}837; AVX512F: {{.*}}838; AVX512VL: {{.*}}839