1196 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-prefix=SSE3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+xop | FileCheck %s --check-prefix=XOP4; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX15; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX26; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefixes=AVX512,AVX512F7; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl | FileCheck %s --check-prefixes=AVX512,AVX512VL8 9;10; 128-bit vectors11;12 13define <2 x i64> @bitselect_v2i64_rr(<2 x i64>, <2 x i64>) {14; SSE-LABEL: bitselect_v2i64_rr:15; SSE: # %bb.0:16; SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm017; SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm118; SSE-NEXT: orps %xmm1, %xmm019; SSE-NEXT: retq20;21; XOP-LABEL: bitselect_v2i64_rr:22; XOP: # %bb.0:23; XOP-NEXT: vpcmov {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1, %xmm024; XOP-NEXT: retq25;26; AVX-LABEL: bitselect_v2i64_rr:27; AVX: # %bb.0:28; AVX-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm029; AVX-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm130; AVX-NEXT: vorps %xmm0, %xmm1, %xmm031; AVX-NEXT: retq32;33; AVX512F-LABEL: bitselect_v2i64_rr:34; AVX512F: # %bb.0:35; AVX512F-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm136; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm037; AVX512F-NEXT: vpmovsxbd {{.*#+}} xmm2 = [4294967295,4294967294,4294967293,4294967292]38; AVX512F-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm2 & (zmm0 ^ zmm1))39; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm040; AVX512F-NEXT: vzeroupper41; AVX512F-NEXT: retq42;43; AVX512VL-LABEL: bitselect_v2i64_rr:44; AVX512VL: # %bb.0:45; AVX512VL-NEXT: vpternlogq {{.*#+}} xmm0 = xmm0 ^ (mem & (xmm0 ^ xmm1))46; AVX512VL-NEXT: retq47 %3 = and <2 x i64> %0, <i64 4294967296, i64 12884901890>48 %4 = and <2 x i64> %1, <i64 -4294967297, i64 -12884901891>49 %5 = or <2 x i64> %4, %350 ret <2 x i64> %551}52 53define <2 x i64> @bitselect_v2i64_rm(<2 x i64>, ptr nocapture readonly) {54; SSE-LABEL: bitselect_v2i64_rm:55; SSE: # %bb.0:56; SSE-NEXT: movaps (%rdi), %xmm157; SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm058; SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm159; SSE-NEXT: orps %xmm1, %xmm060; SSE-NEXT: retq61;62; XOP-LABEL: bitselect_v2i64_rm:63; XOP: # %bb.0:64; XOP-NEXT: vmovdqa (%rdi), %xmm165; XOP-NEXT: vpcmov {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1, %xmm066; XOP-NEXT: retq67;68; AVX-LABEL: bitselect_v2i64_rm:69; AVX: # %bb.0:70; AVX-NEXT: vmovaps (%rdi), %xmm171; AVX-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm072; AVX-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm173; AVX-NEXT: vorps %xmm0, %xmm1, %xmm074; AVX-NEXT: retq75;76; AVX512F-LABEL: bitselect_v2i64_rm:77; AVX512F: # %bb.0:78; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm079; AVX512F-NEXT: vmovdqa (%rdi), %xmm180; AVX512F-NEXT: vpmovsxbd {{.*#+}} xmm2 = [4294967294,4294967293,4294967292,4294967295]81; AVX512F-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm2 & (zmm0 ^ zmm1))82; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm083; AVX512F-NEXT: vzeroupper84; AVX512F-NEXT: retq85;86; AVX512VL-LABEL: bitselect_v2i64_rm:87; AVX512VL: # %bb.0:88; AVX512VL-NEXT: vmovdqa (%rdi), %xmm189; AVX512VL-NEXT: vpternlogq {{.*#+}} xmm0 = xmm0 ^ (mem & (xmm0 ^ xmm1))90; AVX512VL-NEXT: retq91 %3 = load <2 x i64>, ptr %192 %4 = and <2 x i64> %0, <i64 8589934593, i64 3>93 %5 = and <2 x i64> %3, <i64 -8589934594, i64 -4>94 %6 = or <2 x i64> %5, %495 ret <2 x i64> %696}97 98define <2 x i64> @bitselect_v2i64_mr(ptr nocapture readonly, <2 x i64>) {99; SSE-LABEL: bitselect_v2i64_mr:100; SSE: # %bb.0:101; SSE-NEXT: movaps (%rdi), %xmm1102; SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1103; SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0104; SSE-NEXT: orps %xmm1, %xmm0105; SSE-NEXT: retq106;107; XOP-LABEL: bitselect_v2i64_mr:108; XOP: # %bb.0:109; XOP-NEXT: vmovdqa (%rdi), %xmm1110; XOP-NEXT: vpcmov {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1, %xmm0111; XOP-NEXT: retq112;113; AVX-LABEL: bitselect_v2i64_mr:114; AVX: # %bb.0:115; AVX-NEXT: vmovaps (%rdi), %xmm1116; AVX-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1117; AVX-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0118; AVX-NEXT: vorps %xmm0, %xmm1, %xmm0119; AVX-NEXT: retq120;121; AVX512F-LABEL: bitselect_v2i64_mr:122; AVX512F: # %bb.0:123; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0124; AVX512F-NEXT: vmovdqa (%rdi), %xmm1125; AVX512F-NEXT: vpmovsxbd {{.*#+}} xmm2 = [2,3,0,1]126; AVX512F-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm2 & (zmm0 ^ zmm1))127; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0128; AVX512F-NEXT: vzeroupper129; AVX512F-NEXT: retq130;131; AVX512VL-LABEL: bitselect_v2i64_mr:132; AVX512VL: # %bb.0:133; AVX512VL-NEXT: vmovdqa (%rdi), %xmm1134; AVX512VL-NEXT: vpternlogq {{.*#+}} xmm0 = xmm0 ^ (mem & (xmm0 ^ xmm1))135; AVX512VL-NEXT: retq136 %3 = load <2 x i64>, ptr %0137 %4 = and <2 x i64> %3, <i64 12884901890, i64 4294967296>138 %5 = and <2 x i64> %1, <i64 -12884901891, i64 -4294967297>139 %6 = or <2 x i64> %4, %5140 ret <2 x i64> %6141}142 143define <2 x i64> @bitselect_v2i64_mm(ptr nocapture readonly, ptr nocapture readonly) {144; SSE-LABEL: bitselect_v2i64_mm:145; SSE: # %bb.0:146; SSE-NEXT: movaps (%rdi), %xmm1147; SSE-NEXT: movaps (%rsi), %xmm0148; SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1149; SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0150; SSE-NEXT: orps %xmm1, %xmm0151; SSE-NEXT: retq152;153; XOP-LABEL: bitselect_v2i64_mm:154; XOP: # %bb.0:155; XOP-NEXT: vmovdqa (%rsi), %xmm0156; XOP-NEXT: vpmovsxbd {{.*#+}} xmm1 = [4294967292,4294967295,4294967294,4294967293]157; XOP-NEXT: vpcmov %xmm1, (%rdi), %xmm0, %xmm0158; XOP-NEXT: retq159;160; AVX-LABEL: bitselect_v2i64_mm:161; AVX: # %bb.0:162; AVX-NEXT: vmovaps (%rdi), %xmm0163; AVX-NEXT: vmovaps (%rsi), %xmm1164; AVX-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0165; AVX-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1166; AVX-NEXT: vorps %xmm0, %xmm1, %xmm0167; AVX-NEXT: retq168;169; AVX512F-LABEL: bitselect_v2i64_mm:170; AVX512F: # %bb.0:171; AVX512F-NEXT: vmovdqa (%rdi), %xmm1172; AVX512F-NEXT: vmovdqa (%rsi), %xmm0173; AVX512F-NEXT: vpmovsxbd {{.*#+}} xmm2 = [4294967292,4294967295,4294967294,4294967293]174; AVX512F-NEXT: vpternlogq {{.*#+}} zmm0 = zmm1 ^ (zmm2 & (zmm0 ^ zmm1))175; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0176; AVX512F-NEXT: vzeroupper177; AVX512F-NEXT: retq178;179; AVX512VL-LABEL: bitselect_v2i64_mm:180; AVX512VL: # %bb.0:181; AVX512VL-NEXT: vmovdqa (%rsi), %xmm1182; AVX512VL-NEXT: vpmovsxbd {{.*#+}} xmm0 = [4294967292,4294967295,4294967294,4294967293]183; AVX512VL-NEXT: vpternlogq {{.*#+}} xmm0 = mem ^ (xmm0 & (xmm1 ^ mem))184; AVX512VL-NEXT: retq185 %3 = load <2 x i64>, ptr %0186 %4 = load <2 x i64>, ptr %1187 %5 = and <2 x i64> %3, <i64 3, i64 8589934593>188 %6 = and <2 x i64> %4, <i64 -4, i64 -8589934594>189 %7 = or <2 x i64> %6, %5190 ret <2 x i64> %7191}192 193define <2 x i64> @bitselect_v2i64_broadcast_rrr(<2 x i64> %a0, <2 x i64> %a1, i64 %a2) {194; SSE-LABEL: bitselect_v2i64_broadcast_rrr:195; SSE: # %bb.0:196; SSE-NEXT: movq %rdi, %xmm2197; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,0,1]198; SSE-NEXT: pand %xmm2, %xmm0199; SSE-NEXT: pandn %xmm1, %xmm2200; SSE-NEXT: por %xmm2, %xmm0201; SSE-NEXT: retq202;203; XOP-LABEL: bitselect_v2i64_broadcast_rrr:204; XOP: # %bb.0:205; XOP-NEXT: vmovq %rdi, %xmm2206; XOP-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[0,1,0,1]207; XOP-NEXT: vpcmov %xmm2, %xmm1, %xmm0, %xmm0208; XOP-NEXT: retq209;210; AVX1-LABEL: bitselect_v2i64_broadcast_rrr:211; AVX1: # %bb.0:212; AVX1-NEXT: vmovq %rdi, %xmm2213; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[0,1,0,1]214; AVX1-NEXT: vpand %xmm2, %xmm0, %xmm0215; AVX1-NEXT: vpandn %xmm1, %xmm2, %xmm1216; AVX1-NEXT: vpor %xmm1, %xmm0, %xmm0217; AVX1-NEXT: retq218;219; AVX2-LABEL: bitselect_v2i64_broadcast_rrr:220; AVX2: # %bb.0:221; AVX2-NEXT: vmovq %rdi, %xmm2222; AVX2-NEXT: vpbroadcastq %xmm2, %xmm2223; AVX2-NEXT: vpand %xmm2, %xmm0, %xmm0224; AVX2-NEXT: vpandn %xmm1, %xmm2, %xmm1225; AVX2-NEXT: vpor %xmm1, %xmm0, %xmm0226; AVX2-NEXT: retq227;228; AVX512F-LABEL: bitselect_v2i64_broadcast_rrr:229; AVX512F: # %bb.0:230; AVX512F-NEXT: vmovq %rdi, %xmm2231; AVX512F-NEXT: vpbroadcastq %xmm2, %xmm2232; AVX512F-NEXT: vpand %xmm2, %xmm0, %xmm0233; AVX512F-NEXT: vpandn %xmm1, %xmm2, %xmm1234; AVX512F-NEXT: vpor %xmm1, %xmm0, %xmm0235; AVX512F-NEXT: retq236;237; AVX512VL-LABEL: bitselect_v2i64_broadcast_rrr:238; AVX512VL: # %bb.0:239; AVX512VL-NEXT: vpbroadcastq %rdi, %xmm2240; AVX512VL-NEXT: vpternlogq {{.*#+}} xmm0 = xmm1 ^ (xmm2 & (xmm0 ^ xmm1))241; AVX512VL-NEXT: retq242 %1 = insertelement <2 x i64> undef, i64 %a2, i32 0243 %2 = shufflevector <2 x i64> %1, <2 x i64> undef, <2 x i32> zeroinitializer244 %3 = xor <2 x i64> %1, <i64 -1, i64 undef>245 %4 = shufflevector <2 x i64> %3, <2 x i64> undef, <2 x i32> zeroinitializer246 %5 = and <2 x i64> %a0, %2247 %6 = and <2 x i64> %a1, %4248 %7 = or <2 x i64> %5, %6249 ret <2 x i64> %7250}251 252define <2 x i64> @bitselect_v2i64_broadcast_rrm(<2 x i64> %a0, <2 x i64> %a1, ptr %p2) {253; SSE-LABEL: bitselect_v2i64_broadcast_rrm:254; SSE: # %bb.0:255; SSE-NEXT: movq {{.*#+}} xmm2 = mem[0],zero256; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,1,0,1]257; SSE-NEXT: pand %xmm2, %xmm0258; SSE-NEXT: pandn %xmm1, %xmm2259; SSE-NEXT: por %xmm2, %xmm0260; SSE-NEXT: retq261;262; XOP-LABEL: bitselect_v2i64_broadcast_rrm:263; XOP: # %bb.0:264; XOP-NEXT: vmovddup {{.*#+}} xmm2 = mem[0,0]265; XOP-NEXT: vpcmov %xmm2, %xmm1, %xmm0, %xmm0266; XOP-NEXT: retq267;268; AVX-LABEL: bitselect_v2i64_broadcast_rrm:269; AVX: # %bb.0:270; AVX-NEXT: vmovddup {{.*#+}} xmm2 = mem[0,0]271; AVX-NEXT: vandps %xmm2, %xmm0, %xmm0272; AVX-NEXT: vandnps %xmm1, %xmm2, %xmm1273; AVX-NEXT: vorps %xmm1, %xmm0, %xmm0274; AVX-NEXT: retq275;276; AVX512F-LABEL: bitselect_v2i64_broadcast_rrm:277; AVX512F: # %bb.0:278; AVX512F-NEXT: vmovddup {{.*#+}} xmm2 = mem[0,0]279; AVX512F-NEXT: vandps %xmm2, %xmm0, %xmm0280; AVX512F-NEXT: vandnps %xmm1, %xmm2, %xmm1281; AVX512F-NEXT: vorps %xmm1, %xmm0, %xmm0282; AVX512F-NEXT: retq283;284; AVX512VL-LABEL: bitselect_v2i64_broadcast_rrm:285; AVX512VL: # %bb.0:286; AVX512VL-NEXT: vpternlogq {{.*#+}} xmm0 = xmm1 ^ (m64bcst & (xmm0 ^ xmm1))287; AVX512VL-NEXT: retq288 %a2 = load i64, ptr %p2289 %1 = insertelement <2 x i64> undef, i64 %a2, i32 0290 %2 = shufflevector <2 x i64> %1, <2 x i64> undef, <2 x i32> zeroinitializer291 %3 = xor <2 x i64> %1, <i64 -1, i64 undef>292 %4 = shufflevector <2 x i64> %3, <2 x i64> undef, <2 x i32> zeroinitializer293 %5 = and <2 x i64> %a0, %2294 %6 = and <2 x i64> %a1, %4295 %7 = or <2 x i64> %5, %6296 ret <2 x i64> %7297}298 299;300; 256-bit vectors301;302 303define <4 x i64> @bitselect_v4i64_rr(<4 x i64>, <4 x i64>) {304; SSE-LABEL: bitselect_v4i64_rr:305; SSE: # %bb.0:306; SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1307; SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0308; SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm3309; SSE-NEXT: orps %xmm3, %xmm1310; SSE-NEXT: andps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2311; SSE-NEXT: orps %xmm2, %xmm0312; SSE-NEXT: retq313;314; XOP-LABEL: bitselect_v4i64_rr:315; XOP: # %bb.0:316; XOP-NEXT: vpcmov {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1, %ymm0317; XOP-NEXT: retq318;319; AVX-LABEL: bitselect_v4i64_rr:320; AVX: # %bb.0:321; AVX-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0322; AVX-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1323; AVX-NEXT: vorps %ymm0, %ymm1, %ymm0324; AVX-NEXT: retq325;326; AVX512F-LABEL: bitselect_v4i64_rr:327; AVX512F: # %bb.0:328; AVX512F-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1329; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0330; AVX512F-NEXT: vpmovsxbd {{.*#+}} ymm2 = [4294967295,4294967294,4294967293,4294967292,4294967293,4294967292,4294967293,4294967292]331; AVX512F-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm2 & (zmm0 ^ zmm1))332; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 killed $zmm0333; AVX512F-NEXT: retq334;335; AVX512VL-LABEL: bitselect_v4i64_rr:336; AVX512VL: # %bb.0:337; AVX512VL-NEXT: vpternlogq {{.*#+}} ymm0 = ymm0 ^ (mem & (ymm0 ^ ymm1))338; AVX512VL-NEXT: retq339 %3 = and <4 x i64> %0, <i64 4294967296, i64 12884901890, i64 12884901890, i64 12884901890>340 %4 = and <4 x i64> %1, <i64 -4294967297, i64 -12884901891, i64 -12884901891, i64 -12884901891>341 %5 = or <4 x i64> %4, %3342 ret <4 x i64> %5343}344 345define <4 x i64> @bitselect_v4i64_rm(<4 x i64>, ptr nocapture readonly) {346; SSE-LABEL: bitselect_v4i64_rm:347; SSE: # %bb.0:348; SSE-NEXT: movaps {{.*#+}} xmm2 = [18446744065119617022,18446744073709551612]349; SSE-NEXT: movaps 16(%rdi), %xmm4350; SSE-NEXT: andps %xmm2, %xmm4351; SSE-NEXT: movaps (%rdi), %xmm5352; SSE-NEXT: andps %xmm2, %xmm5353; SSE-NEXT: movaps %xmm2, %xmm3354; SSE-NEXT: andnps %xmm0, %xmm3355; SSE-NEXT: orps %xmm5, %xmm3356; SSE-NEXT: andnps %xmm1, %xmm2357; SSE-NEXT: orps %xmm4, %xmm2358; SSE-NEXT: movaps %xmm3, %xmm0359; SSE-NEXT: movaps %xmm2, %xmm1360; SSE-NEXT: retq361;362; XOP-LABEL: bitselect_v4i64_rm:363; XOP: # %bb.0:364; XOP-NEXT: vmovdqa (%rdi), %ymm1365; XOP-NEXT: vpcmov {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1, %ymm0366; XOP-NEXT: retq367;368; AVX-LABEL: bitselect_v4i64_rm:369; AVX: # %bb.0:370; AVX-NEXT: vmovaps (%rdi), %ymm1371; AVX-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0372; AVX-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1373; AVX-NEXT: vorps %ymm0, %ymm1, %ymm0374; AVX-NEXT: retq375;376; AVX512F-LABEL: bitselect_v4i64_rm:377; AVX512F: # %bb.0:378; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0379; AVX512F-NEXT: vmovdqa (%rdi), %ymm1380; AVX512F-NEXT: vpmovsxbd {{.*#+}} ymm2 = [4294967294,4294967293,4294967292,4294967295,4294967294,4294967293,4294967292,4294967295]381; AVX512F-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm2 & (zmm0 ^ zmm1))382; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 killed $zmm0383; AVX512F-NEXT: retq384;385; AVX512VL-LABEL: bitselect_v4i64_rm:386; AVX512VL: # %bb.0:387; AVX512VL-NEXT: vmovdqa (%rdi), %ymm1388; AVX512VL-NEXT: vpternlogq {{.*#+}} ymm0 = ymm0 ^ (mem & (ymm0 ^ ymm1))389; AVX512VL-NEXT: retq390 %3 = load <4 x i64>, ptr %1391 %4 = and <4 x i64> %0, <i64 8589934593, i64 3, i64 8589934593, i64 3>392 %5 = and <4 x i64> %3, <i64 -8589934594, i64 -4, i64 -8589934594, i64 -4>393 %6 = or <4 x i64> %5, %4394 ret <4 x i64> %6395}396 397define <4 x i64> @bitselect_v4i64_mr(ptr nocapture readonly, <4 x i64>) {398; SSE-LABEL: bitselect_v4i64_mr:399; SSE: # %bb.0:400; SSE-NEXT: movaps {{.*#+}} xmm2 = [12884901890,4294967296]401; SSE-NEXT: movaps 16(%rdi), %xmm4402; SSE-NEXT: andps %xmm2, %xmm4403; SSE-NEXT: movaps (%rdi), %xmm5404; SSE-NEXT: andps %xmm2, %xmm5405; SSE-NEXT: movaps %xmm2, %xmm3406; SSE-NEXT: andnps %xmm0, %xmm3407; SSE-NEXT: orps %xmm5, %xmm3408; SSE-NEXT: andnps %xmm1, %xmm2409; SSE-NEXT: orps %xmm4, %xmm2410; SSE-NEXT: movaps %xmm3, %xmm0411; SSE-NEXT: movaps %xmm2, %xmm1412; SSE-NEXT: retq413;414; XOP-LABEL: bitselect_v4i64_mr:415; XOP: # %bb.0:416; XOP-NEXT: vmovdqa (%rdi), %ymm1417; XOP-NEXT: vpcmov {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm1, %ymm0418; XOP-NEXT: retq419;420; AVX-LABEL: bitselect_v4i64_mr:421; AVX: # %bb.0:422; AVX-NEXT: vmovaps (%rdi), %ymm1423; AVX-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1424; AVX-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0425; AVX-NEXT: vorps %ymm0, %ymm1, %ymm0426; AVX-NEXT: retq427;428; AVX512F-LABEL: bitselect_v4i64_mr:429; AVX512F: # %bb.0:430; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0431; AVX512F-NEXT: vmovdqa (%rdi), %ymm1432; AVX512F-NEXT: vpmovsxbd {{.*#+}} ymm2 = [2,3,0,1,2,3,0,1]433; AVX512F-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm2 & (zmm0 ^ zmm1))434; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 killed $zmm0435; AVX512F-NEXT: retq436;437; AVX512VL-LABEL: bitselect_v4i64_mr:438; AVX512VL: # %bb.0:439; AVX512VL-NEXT: vmovdqa (%rdi), %ymm1440; AVX512VL-NEXT: vpternlogq {{.*#+}} ymm0 = ymm0 ^ (mem & (ymm0 ^ ymm1))441; AVX512VL-NEXT: retq442 %3 = load <4 x i64>, ptr %0443 %4 = and <4 x i64> %3, <i64 12884901890, i64 4294967296, i64 12884901890, i64 4294967296>444 %5 = and <4 x i64> %1, <i64 -12884901891, i64 -4294967297, i64 -12884901891, i64 -4294967297>445 %6 = or <4 x i64> %4, %5446 ret <4 x i64> %6447}448 449define <4 x i64> @bitselect_v4i64_mm(ptr nocapture readonly, ptr nocapture readonly) {450; SSE-LABEL: bitselect_v4i64_mm:451; SSE: # %bb.0:452; SSE-NEXT: movaps {{.*#+}} xmm1 = [18446744073709551612,18446744065119617022]453; SSE-NEXT: movaps 16(%rsi), %xmm2454; SSE-NEXT: andps %xmm1, %xmm2455; SSE-NEXT: movaps (%rsi), %xmm3456; SSE-NEXT: andps %xmm1, %xmm3457; SSE-NEXT: movaps %xmm1, %xmm0458; SSE-NEXT: andnps (%rdi), %xmm0459; SSE-NEXT: orps %xmm3, %xmm0460; SSE-NEXT: andnps 16(%rdi), %xmm1461; SSE-NEXT: orps %xmm2, %xmm1462; SSE-NEXT: retq463;464; XOP-LABEL: bitselect_v4i64_mm:465; XOP: # %bb.0:466; XOP-NEXT: vmovdqa (%rsi), %ymm0467; XOP-NEXT: vbroadcastf128 {{.*#+}} ymm1 = [18446744073709551612,18446744065119617022,18446744073709551612,18446744065119617022]468; XOP-NEXT: # ymm1 = mem[0,1,0,1]469; XOP-NEXT: vpcmov %ymm1, (%rdi), %ymm0, %ymm0470; XOP-NEXT: retq471;472; AVX-LABEL: bitselect_v4i64_mm:473; AVX: # %bb.0:474; AVX-NEXT: vmovaps (%rdi), %ymm0475; AVX-NEXT: vmovaps (%rsi), %ymm1476; AVX-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0477; AVX-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm1, %ymm1478; AVX-NEXT: vorps %ymm0, %ymm1, %ymm0479; AVX-NEXT: retq480;481; AVX512F-LABEL: bitselect_v4i64_mm:482; AVX512F: # %bb.0:483; AVX512F-NEXT: vmovdqa (%rdi), %ymm1484; AVX512F-NEXT: vmovdqa (%rsi), %ymm0485; AVX512F-NEXT: vpmovsxbd {{.*#+}} ymm2 = [4294967292,4294967295,4294967294,4294967293,4294967292,4294967295,4294967294,4294967293]486; AVX512F-NEXT: vpternlogq {{.*#+}} zmm0 = zmm1 ^ (zmm2 & (zmm0 ^ zmm1))487; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 killed $zmm0488; AVX512F-NEXT: retq489;490; AVX512VL-LABEL: bitselect_v4i64_mm:491; AVX512VL: # %bb.0:492; AVX512VL-NEXT: vmovdqa (%rsi), %ymm1493; AVX512VL-NEXT: vpmovsxbd {{.*#+}} ymm0 = [4294967292,4294967295,4294967294,4294967293,4294967292,4294967295,4294967294,4294967293]494; AVX512VL-NEXT: vpternlogq {{.*#+}} ymm0 = mem ^ (ymm0 & (ymm1 ^ mem))495; AVX512VL-NEXT: retq496 %3 = load <4 x i64>, ptr %0497 %4 = load <4 x i64>, ptr %1498 %5 = and <4 x i64> %3, <i64 3, i64 8589934593, i64 3, i64 8589934593>499 %6 = and <4 x i64> %4, <i64 -4, i64 -8589934594, i64 -4, i64 -8589934594>500 %7 = or <4 x i64> %6, %5501 ret <4 x i64> %7502}503 504define <4 x i64> @bitselect_v4i64_broadcast_rrr(<4 x i64> %a0, <4 x i64> %a1, i64 %a2) {505; SSE-LABEL: bitselect_v4i64_broadcast_rrr:506; SSE: # %bb.0:507; SSE-NEXT: movq %rdi, %xmm4508; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,1,0,1]509; SSE-NEXT: pand %xmm4, %xmm1510; SSE-NEXT: pand %xmm4, %xmm0511; SSE-NEXT: movdqa %xmm4, %xmm5512; SSE-NEXT: pandn %xmm3, %xmm5513; SSE-NEXT: por %xmm5, %xmm1514; SSE-NEXT: pandn %xmm2, %xmm4515; SSE-NEXT: por %xmm4, %xmm0516; SSE-NEXT: retq517;518; XOP-LABEL: bitselect_v4i64_broadcast_rrr:519; XOP: # %bb.0:520; XOP-NEXT: vmovq %rdi, %xmm2521; XOP-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[0,1,0,1]522; XOP-NEXT: vinsertf128 $1, %xmm2, %ymm2, %ymm2523; XOP-NEXT: vpcmov %ymm2, %ymm1, %ymm0, %ymm0524; XOP-NEXT: retq525;526; AVX1-LABEL: bitselect_v4i64_broadcast_rrr:527; AVX1: # %bb.0:528; AVX1-NEXT: vmovq %rdi, %xmm2529; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[0,1,0,1]530; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm2, %ymm2531; AVX1-NEXT: vandps %ymm2, %ymm0, %ymm0532; AVX1-NEXT: vandnps %ymm1, %ymm2, %ymm1533; AVX1-NEXT: vorps %ymm1, %ymm0, %ymm0534; AVX1-NEXT: retq535;536; AVX2-LABEL: bitselect_v4i64_broadcast_rrr:537; AVX2: # %bb.0:538; AVX2-NEXT: vmovq %rdi, %xmm2539; AVX2-NEXT: vpbroadcastq %xmm2, %ymm2540; AVX2-NEXT: vpand %ymm2, %ymm0, %ymm0541; AVX2-NEXT: vpandn %ymm1, %ymm2, %ymm1542; AVX2-NEXT: vpor %ymm1, %ymm0, %ymm0543; AVX2-NEXT: retq544;545; AVX512F-LABEL: bitselect_v4i64_broadcast_rrr:546; AVX512F: # %bb.0:547; AVX512F-NEXT: vmovq %rdi, %xmm2548; AVX512F-NEXT: vpbroadcastq %xmm2, %ymm2549; AVX512F-NEXT: vpand %ymm2, %ymm0, %ymm0550; AVX512F-NEXT: vpandn %ymm1, %ymm2, %ymm1551; AVX512F-NEXT: vpor %ymm1, %ymm0, %ymm0552; AVX512F-NEXT: retq553;554; AVX512VL-LABEL: bitselect_v4i64_broadcast_rrr:555; AVX512VL: # %bb.0:556; AVX512VL-NEXT: vpbroadcastq %rdi, %ymm2557; AVX512VL-NEXT: vpternlogq {{.*#+}} ymm0 = ymm1 ^ (ymm2 & (ymm0 ^ ymm1))558; AVX512VL-NEXT: retq559 %1 = insertelement <4 x i64> undef, i64 %a2, i32 0560 %2 = shufflevector <4 x i64> %1, <4 x i64> undef, <4 x i32> zeroinitializer561 %3 = xor <4 x i64> %1, <i64 -1, i64 undef, i64 undef, i64 undef>562 %4 = shufflevector <4 x i64> %3, <4 x i64> undef, <4 x i32> zeroinitializer563 %5 = and <4 x i64> %a0, %2564 %6 = and <4 x i64> %a1, %4565 %7 = or <4 x i64> %5, %6566 ret <4 x i64> %7567}568 569define <4 x i64> @bitselect_v4i64_broadcast_rrm(<4 x i64> %a0, <4 x i64> %a1, ptr %p2) {570; SSE-LABEL: bitselect_v4i64_broadcast_rrm:571; SSE: # %bb.0:572; SSE-NEXT: movq {{.*#+}} xmm4 = mem[0],zero573; SSE-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,1,0,1]574; SSE-NEXT: pand %xmm4, %xmm1575; SSE-NEXT: pand %xmm4, %xmm0576; SSE-NEXT: movdqa %xmm4, %xmm5577; SSE-NEXT: pandn %xmm3, %xmm5578; SSE-NEXT: por %xmm5, %xmm1579; SSE-NEXT: pandn %xmm2, %xmm4580; SSE-NEXT: por %xmm4, %xmm0581; SSE-NEXT: retq582;583; XOP-LABEL: bitselect_v4i64_broadcast_rrm:584; XOP: # %bb.0:585; XOP-NEXT: vbroadcastsd (%rdi), %ymm2586; XOP-NEXT: vpcmov %ymm2, %ymm1, %ymm0, %ymm0587; XOP-NEXT: retq588;589; AVX-LABEL: bitselect_v4i64_broadcast_rrm:590; AVX: # %bb.0:591; AVX-NEXT: vbroadcastsd (%rdi), %ymm2592; AVX-NEXT: vandps %ymm2, %ymm0, %ymm0593; AVX-NEXT: vandnps %ymm1, %ymm2, %ymm1594; AVX-NEXT: vorps %ymm1, %ymm0, %ymm0595; AVX-NEXT: retq596;597; AVX512F-LABEL: bitselect_v4i64_broadcast_rrm:598; AVX512F: # %bb.0:599; AVX512F-NEXT: vbroadcastsd (%rdi), %ymm2600; AVX512F-NEXT: vandps %ymm2, %ymm0, %ymm0601; AVX512F-NEXT: vandnps %ymm1, %ymm2, %ymm1602; AVX512F-NEXT: vorps %ymm1, %ymm0, %ymm0603; AVX512F-NEXT: retq604;605; AVX512VL-LABEL: bitselect_v4i64_broadcast_rrm:606; AVX512VL: # %bb.0:607; AVX512VL-NEXT: vpternlogq {{.*#+}} ymm0 = ymm1 ^ (m64bcst & (ymm0 ^ ymm1))608; AVX512VL-NEXT: retq609 %a2 = load i64, ptr %p2610 %1 = insertelement <4 x i64> undef, i64 %a2, i32 0611 %2 = shufflevector <4 x i64> %1, <4 x i64> undef, <4 x i32> zeroinitializer612 %3 = xor <4 x i64> %1, <i64 -1, i64 undef, i64 undef, i64 undef>613 %4 = shufflevector <4 x i64> %3, <4 x i64> undef, <4 x i32> zeroinitializer614 %5 = and <4 x i64> %a0, %2615 %6 = and <4 x i64> %a1, %4616 %7 = or <4 x i64> %5, %6617 ret <4 x i64> %7618}619 620;621; 512-bit vectors622;623 624define <8 x i64> @bitselect_v8i64_rr(<8 x i64>, <8 x i64>) {625; SSE-LABEL: bitselect_v8i64_rr:626; SSE: # %bb.0:627; SSE-NEXT: movaps {{.*#+}} xmm8 = [18446744060824649725,18446744060824649725]628; SSE-NEXT: andps %xmm8, %xmm7629; SSE-NEXT: movaps {{.*#+}} xmm9 = [18446744069414584319,18446744060824649725]630; SSE-NEXT: andps %xmm9, %xmm6631; SSE-NEXT: andps %xmm8, %xmm5632; SSE-NEXT: andps %xmm9, %xmm4633; SSE-NEXT: movaps %xmm9, %xmm10634; SSE-NEXT: andnps %xmm0, %xmm10635; SSE-NEXT: orps %xmm4, %xmm10636; SSE-NEXT: movaps %xmm8, %xmm4637; SSE-NEXT: andnps %xmm1, %xmm4638; SSE-NEXT: orps %xmm5, %xmm4639; SSE-NEXT: andnps %xmm2, %xmm9640; SSE-NEXT: orps %xmm6, %xmm9641; SSE-NEXT: andnps %xmm3, %xmm8642; SSE-NEXT: orps %xmm7, %xmm8643; SSE-NEXT: movaps %xmm10, %xmm0644; SSE-NEXT: movaps %xmm4, %xmm1645; SSE-NEXT: movaps %xmm9, %xmm2646; SSE-NEXT: movaps %xmm8, %xmm3647; SSE-NEXT: retq648;649; XOP-LABEL: bitselect_v8i64_rr:650; XOP: # %bb.0:651; XOP-NEXT: vmovdqa {{.*#+}} ymm4 = [18446744069414584319,18446744060824649725,18446744060824649725,18446744060824649725]652; XOP-NEXT: vpcmov %ymm4, %ymm0, %ymm2, %ymm0653; XOP-NEXT: vpcmov %ymm4, %ymm1, %ymm3, %ymm1654; XOP-NEXT: retq655;656; AVX-LABEL: bitselect_v8i64_rr:657; AVX: # %bb.0:658; AVX-NEXT: vmovaps {{.*#+}} ymm4 = [18446744069414584319,18446744060824649725,18446744060824649725,18446744060824649725]659; AVX-NEXT: vandps %ymm4, %ymm3, %ymm3660; AVX-NEXT: vandps %ymm4, %ymm2, %ymm2661; AVX-NEXT: vandnps %ymm0, %ymm4, %ymm0662; AVX-NEXT: vorps %ymm0, %ymm2, %ymm0663; AVX-NEXT: vandnps %ymm1, %ymm4, %ymm1664; AVX-NEXT: vorps %ymm1, %ymm3, %ymm1665; AVX-NEXT: retq666;667; AVX512-LABEL: bitselect_v8i64_rr:668; AVX512: # %bb.0:669; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm1))670; AVX512-NEXT: retq671 %3 = and <8 x i64> %0, <i64 4294967296, i64 12884901890, i64 12884901890, i64 12884901890, i64 4294967296, i64 12884901890, i64 12884901890, i64 12884901890>672 %4 = and <8 x i64> %1, <i64 -4294967297, i64 -12884901891, i64 -12884901891, i64 -12884901891, i64 -4294967297, i64 -12884901891, i64 -12884901891, i64 -12884901891>673 %5 = or <8 x i64> %4, %3674 ret <8 x i64> %5675}676 677define <8 x i64> @bitselect_v8i64_rm(<8 x i64>, ptr nocapture readonly) {678; SSE-LABEL: bitselect_v8i64_rm:679; SSE: # %bb.0:680; SSE-NEXT: movaps {{.*#+}} xmm4 = [18446744065119617022,18446744073709551612]681; SSE-NEXT: movaps 48(%rdi), %xmm8682; SSE-NEXT: andps %xmm4, %xmm8683; SSE-NEXT: movaps 32(%rdi), %xmm9684; SSE-NEXT: andps %xmm4, %xmm9685; SSE-NEXT: movaps 16(%rdi), %xmm7686; SSE-NEXT: andps %xmm4, %xmm7687; SSE-NEXT: movaps (%rdi), %xmm6688; SSE-NEXT: andps %xmm4, %xmm6689; SSE-NEXT: movaps %xmm4, %xmm5690; SSE-NEXT: andnps %xmm0, %xmm5691; SSE-NEXT: orps %xmm6, %xmm5692; SSE-NEXT: movaps %xmm4, %xmm6693; SSE-NEXT: andnps %xmm1, %xmm6694; SSE-NEXT: orps %xmm7, %xmm6695; SSE-NEXT: movaps %xmm4, %xmm7696; SSE-NEXT: andnps %xmm2, %xmm7697; SSE-NEXT: orps %xmm9, %xmm7698; SSE-NEXT: andnps %xmm3, %xmm4699; SSE-NEXT: orps %xmm8, %xmm4700; SSE-NEXT: movaps %xmm5, %xmm0701; SSE-NEXT: movaps %xmm6, %xmm1702; SSE-NEXT: movaps %xmm7, %xmm2703; SSE-NEXT: movaps %xmm4, %xmm3704; SSE-NEXT: retq705;706; XOP-LABEL: bitselect_v8i64_rm:707; XOP: # %bb.0:708; XOP-NEXT: vmovdqa (%rdi), %ymm2709; XOP-NEXT: vmovdqa 32(%rdi), %ymm3710; XOP-NEXT: vbroadcastf128 {{.*#+}} ymm4 = [18446744065119617022,18446744073709551612,18446744065119617022,18446744073709551612]711; XOP-NEXT: # ymm4 = mem[0,1,0,1]712; XOP-NEXT: vpcmov %ymm4, %ymm0, %ymm2, %ymm0713; XOP-NEXT: vpcmov %ymm4, %ymm1, %ymm3, %ymm1714; XOP-NEXT: retq715;716; AVX-LABEL: bitselect_v8i64_rm:717; AVX: # %bb.0:718; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm2 = [18446744065119617022,18446744073709551612,18446744065119617022,18446744073709551612]719; AVX-NEXT: # ymm2 = mem[0,1,0,1]720; AVX-NEXT: vandps 32(%rdi), %ymm2, %ymm3721; AVX-NEXT: vandps (%rdi), %ymm2, %ymm4722; AVX-NEXT: vandnps %ymm0, %ymm2, %ymm0723; AVX-NEXT: vorps %ymm0, %ymm4, %ymm0724; AVX-NEXT: vandnps %ymm1, %ymm2, %ymm1725; AVX-NEXT: vorps %ymm1, %ymm3, %ymm1726; AVX-NEXT: retq727;728; AVX512-LABEL: bitselect_v8i64_rm:729; AVX512: # %bb.0:730; AVX512-NEXT: vmovdqa64 (%rdi), %zmm1731; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm1))732; AVX512-NEXT: retq733 %3 = load <8 x i64>, ptr %1734 %4 = and <8 x i64> %0, <i64 8589934593, i64 3, i64 8589934593, i64 3, i64 8589934593, i64 3, i64 8589934593, i64 3>735 %5 = and <8 x i64> %3, <i64 -8589934594, i64 -4, i64 -8589934594, i64 -4, i64 -8589934594, i64 -4, i64 -8589934594, i64 -4>736 %6 = or <8 x i64> %5, %4737 ret <8 x i64> %6738}739 740define <8 x i64> @bitselect_v8i64_mr(ptr nocapture readonly, <8 x i64>) {741; SSE-LABEL: bitselect_v8i64_mr:742; SSE: # %bb.0:743; SSE-NEXT: movaps {{.*#+}} xmm4 = [12884901890,4294967296]744; SSE-NEXT: movaps 48(%rdi), %xmm8745; SSE-NEXT: andps %xmm4, %xmm8746; SSE-NEXT: movaps 32(%rdi), %xmm9747; SSE-NEXT: andps %xmm4, %xmm9748; SSE-NEXT: movaps 16(%rdi), %xmm7749; SSE-NEXT: andps %xmm4, %xmm7750; SSE-NEXT: movaps (%rdi), %xmm6751; SSE-NEXT: andps %xmm4, %xmm6752; SSE-NEXT: movaps %xmm4, %xmm5753; SSE-NEXT: andnps %xmm0, %xmm5754; SSE-NEXT: orps %xmm6, %xmm5755; SSE-NEXT: movaps %xmm4, %xmm6756; SSE-NEXT: andnps %xmm1, %xmm6757; SSE-NEXT: orps %xmm7, %xmm6758; SSE-NEXT: movaps %xmm4, %xmm7759; SSE-NEXT: andnps %xmm2, %xmm7760; SSE-NEXT: orps %xmm9, %xmm7761; SSE-NEXT: andnps %xmm3, %xmm4762; SSE-NEXT: orps %xmm8, %xmm4763; SSE-NEXT: movaps %xmm5, %xmm0764; SSE-NEXT: movaps %xmm6, %xmm1765; SSE-NEXT: movaps %xmm7, %xmm2766; SSE-NEXT: movaps %xmm4, %xmm3767; SSE-NEXT: retq768;769; XOP-LABEL: bitselect_v8i64_mr:770; XOP: # %bb.0:771; XOP-NEXT: vmovdqa (%rdi), %ymm2772; XOP-NEXT: vmovdqa 32(%rdi), %ymm3773; XOP-NEXT: vbroadcastf128 {{.*#+}} ymm4 = [12884901890,4294967296,12884901890,4294967296]774; XOP-NEXT: # ymm4 = mem[0,1,0,1]775; XOP-NEXT: vpcmov %ymm4, %ymm0, %ymm2, %ymm0776; XOP-NEXT: vpcmov %ymm4, %ymm1, %ymm3, %ymm1777; XOP-NEXT: retq778;779; AVX-LABEL: bitselect_v8i64_mr:780; AVX: # %bb.0:781; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm2 = [12884901890,4294967296,12884901890,4294967296]782; AVX-NEXT: # ymm2 = mem[0,1,0,1]783; AVX-NEXT: vandps 32(%rdi), %ymm2, %ymm3784; AVX-NEXT: vandps (%rdi), %ymm2, %ymm4785; AVX-NEXT: vandnps %ymm0, %ymm2, %ymm0786; AVX-NEXT: vorps %ymm0, %ymm4, %ymm0787; AVX-NEXT: vandnps %ymm1, %ymm2, %ymm1788; AVX-NEXT: vorps %ymm1, %ymm3, %ymm1789; AVX-NEXT: retq790;791; AVX512-LABEL: bitselect_v8i64_mr:792; AVX512: # %bb.0:793; AVX512-NEXT: vmovdqa64 (%rdi), %zmm1794; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm1))795; AVX512-NEXT: retq796 %3 = load <8 x i64>, ptr %0797 %4 = and <8 x i64> %3, <i64 12884901890, i64 4294967296, i64 12884901890, i64 4294967296, i64 12884901890, i64 4294967296, i64 12884901890, i64 4294967296>798 %5 = and <8 x i64> %1, <i64 -12884901891, i64 -4294967297, i64 -12884901891, i64 -4294967297, i64 -12884901891, i64 -4294967297, i64 -12884901891, i64 -4294967297>799 %6 = or <8 x i64> %4, %5800 ret <8 x i64> %6801}802 803define <8 x i64> @bitselect_v8i64_mm(ptr nocapture readonly, ptr nocapture readonly) {804; SSE-LABEL: bitselect_v8i64_mm:805; SSE: # %bb.0:806; SSE-NEXT: movaps {{.*#+}} xmm3 = [18446744073709551612,18446744065119617022]807; SSE-NEXT: movaps 48(%rsi), %xmm4808; SSE-NEXT: andps %xmm3, %xmm4809; SSE-NEXT: movaps 32(%rsi), %xmm5810; SSE-NEXT: andps %xmm3, %xmm5811; SSE-NEXT: movaps 16(%rsi), %xmm2812; SSE-NEXT: andps %xmm3, %xmm2813; SSE-NEXT: movaps (%rsi), %xmm1814; SSE-NEXT: andps %xmm3, %xmm1815; SSE-NEXT: movaps %xmm3, %xmm0816; SSE-NEXT: andnps (%rdi), %xmm0817; SSE-NEXT: orps %xmm1, %xmm0818; SSE-NEXT: movaps %xmm3, %xmm1819; SSE-NEXT: andnps 16(%rdi), %xmm1820; SSE-NEXT: orps %xmm2, %xmm1821; SSE-NEXT: movaps %xmm3, %xmm2822; SSE-NEXT: andnps 32(%rdi), %xmm2823; SSE-NEXT: orps %xmm5, %xmm2824; SSE-NEXT: andnps 48(%rdi), %xmm3825; SSE-NEXT: orps %xmm4, %xmm3826; SSE-NEXT: retq827;828; XOP-LABEL: bitselect_v8i64_mm:829; XOP: # %bb.0:830; XOP-NEXT: vmovdqa (%rsi), %ymm0831; XOP-NEXT: vmovdqa 32(%rsi), %ymm1832; XOP-NEXT: vbroadcastf128 {{.*#+}} ymm2 = [18446744073709551612,18446744065119617022,18446744073709551612,18446744065119617022]833; XOP-NEXT: # ymm2 = mem[0,1,0,1]834; XOP-NEXT: vpcmov %ymm2, (%rdi), %ymm0, %ymm0835; XOP-NEXT: vpcmov %ymm2, 32(%rdi), %ymm1, %ymm1836; XOP-NEXT: retq837;838; AVX-LABEL: bitselect_v8i64_mm:839; AVX: # %bb.0:840; AVX-NEXT: vbroadcastf128 {{.*#+}} ymm1 = [18446744073709551612,18446744065119617022,18446744073709551612,18446744065119617022]841; AVX-NEXT: # ymm1 = mem[0,1,0,1]842; AVX-NEXT: vandps 32(%rsi), %ymm1, %ymm2843; AVX-NEXT: vandps (%rsi), %ymm1, %ymm0844; AVX-NEXT: vandnps (%rdi), %ymm1, %ymm3845; AVX-NEXT: vorps %ymm3, %ymm0, %ymm0846; AVX-NEXT: vandnps 32(%rdi), %ymm1, %ymm1847; AVX-NEXT: vorps %ymm1, %ymm2, %ymm1848; AVX-NEXT: retq849;850; AVX512-LABEL: bitselect_v8i64_mm:851; AVX512: # %bb.0:852; AVX512-NEXT: vmovdqa64 (%rsi), %zmm1853; AVX512-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = [18446744073709551612,18446744065119617022,18446744073709551612,18446744065119617022,18446744073709551612,18446744065119617022,18446744073709551612,18446744065119617022]854; AVX512-NEXT: # zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]855; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = mem ^ (zmm0 & (zmm1 ^ mem))856; AVX512-NEXT: retq857 %3 = load <8 x i64>, ptr %0858 %4 = load <8 x i64>, ptr %1859 %5 = and <8 x i64> %3, <i64 3, i64 8589934593, i64 3, i64 8589934593, i64 3, i64 8589934593, i64 3, i64 8589934593>860 %6 = and <8 x i64> %4, <i64 -4, i64 -8589934594, i64 -4, i64 -8589934594, i64 -4, i64 -8589934594, i64 -4, i64 -8589934594>861 %7 = or <8 x i64> %6, %5862 ret <8 x i64> %7863}864 865define <8 x i64> @bitselect_v8i64_broadcast_rrr(<8 x i64> %a0, <8 x i64> %a1, i64 %a2) {866; SSE-LABEL: bitselect_v8i64_broadcast_rrr:867; SSE: # %bb.0:868; SSE-NEXT: movq %rdi, %xmm8869; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[0,1,0,1]870; SSE-NEXT: pand %xmm8, %xmm3871; SSE-NEXT: pand %xmm8, %xmm2872; SSE-NEXT: pand %xmm8, %xmm1873; SSE-NEXT: pand %xmm8, %xmm0874; SSE-NEXT: movdqa %xmm8, %xmm9875; SSE-NEXT: pandn %xmm7, %xmm9876; SSE-NEXT: por %xmm9, %xmm3877; SSE-NEXT: movdqa %xmm8, %xmm7878; SSE-NEXT: pandn %xmm6, %xmm7879; SSE-NEXT: por %xmm7, %xmm2880; SSE-NEXT: movdqa %xmm8, %xmm6881; SSE-NEXT: pandn %xmm5, %xmm6882; SSE-NEXT: por %xmm6, %xmm1883; SSE-NEXT: pandn %xmm4, %xmm8884; SSE-NEXT: por %xmm8, %xmm0885; SSE-NEXT: retq886;887; XOP-LABEL: bitselect_v8i64_broadcast_rrr:888; XOP: # %bb.0:889; XOP-NEXT: vmovq %rdi, %xmm4890; XOP-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[0,1,0,1]891; XOP-NEXT: vinsertf128 $1, %xmm4, %ymm4, %ymm4892; XOP-NEXT: vpcmov %ymm4, %ymm2, %ymm0, %ymm0893; XOP-NEXT: vpcmov %ymm4, %ymm3, %ymm1, %ymm1894; XOP-NEXT: retq895;896; AVX1-LABEL: bitselect_v8i64_broadcast_rrr:897; AVX1: # %bb.0:898; AVX1-NEXT: vmovq %rdi, %xmm4899; AVX1-NEXT: vpshufd {{.*#+}} xmm4 = xmm4[0,1,0,1]900; AVX1-NEXT: vinsertf128 $1, %xmm4, %ymm4, %ymm4901; AVX1-NEXT: vandps %ymm4, %ymm1, %ymm1902; AVX1-NEXT: vandps %ymm4, %ymm0, %ymm0903; AVX1-NEXT: vandnps %ymm3, %ymm4, %ymm3904; AVX1-NEXT: vorps %ymm3, %ymm1, %ymm1905; AVX1-NEXT: vandnps %ymm2, %ymm4, %ymm2906; AVX1-NEXT: vorps %ymm2, %ymm0, %ymm0907; AVX1-NEXT: retq908;909; AVX2-LABEL: bitselect_v8i64_broadcast_rrr:910; AVX2: # %bb.0:911; AVX2-NEXT: vmovq %rdi, %xmm4912; AVX2-NEXT: vpbroadcastq %xmm4, %ymm4913; AVX2-NEXT: vpand %ymm4, %ymm1, %ymm1914; AVX2-NEXT: vpand %ymm4, %ymm0, %ymm0915; AVX2-NEXT: vpandn %ymm3, %ymm4, %ymm3916; AVX2-NEXT: vpor %ymm3, %ymm1, %ymm1917; AVX2-NEXT: vpandn %ymm2, %ymm4, %ymm2918; AVX2-NEXT: vpor %ymm2, %ymm0, %ymm0919; AVX2-NEXT: retq920;921; AVX512-LABEL: bitselect_v8i64_broadcast_rrr:922; AVX512: # %bb.0:923; AVX512-NEXT: vpbroadcastq %rdi, %zmm2924; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm1 ^ (zmm2 & (zmm0 ^ zmm1))925; AVX512-NEXT: retq926 %1 = insertelement <8 x i64> undef, i64 %a2, i32 0927 %2 = shufflevector <8 x i64> %1, <8 x i64> undef, <8 x i32> zeroinitializer928 %3 = xor <8 x i64> %1, <i64 -1, i64 undef, i64 undef, i64 undef, i64 undef, i64 undef, i64 undef, i64 undef>929 %4 = shufflevector <8 x i64> %3, <8 x i64> undef, <8 x i32> zeroinitializer930 %5 = and <8 x i64> %a0, %2931 %6 = and <8 x i64> %a1, %4932 %7 = or <8 x i64> %5, %6933 ret <8 x i64> %7934}935 936define <8 x i64> @bitselect_v8i64_broadcast_rrm(<8 x i64> %a0, <8 x i64> %a1, ptr %p2) {937; SSE-LABEL: bitselect_v8i64_broadcast_rrm:938; SSE: # %bb.0:939; SSE-NEXT: movq {{.*#+}} xmm8 = mem[0],zero940; SSE-NEXT: pshufd {{.*#+}} xmm8 = xmm8[0,1,0,1]941; SSE-NEXT: pand %xmm8, %xmm3942; SSE-NEXT: pand %xmm8, %xmm2943; SSE-NEXT: pand %xmm8, %xmm1944; SSE-NEXT: pand %xmm8, %xmm0945; SSE-NEXT: movdqa %xmm8, %xmm9946; SSE-NEXT: pandn %xmm7, %xmm9947; SSE-NEXT: por %xmm9, %xmm3948; SSE-NEXT: movdqa %xmm8, %xmm7949; SSE-NEXT: pandn %xmm6, %xmm7950; SSE-NEXT: por %xmm7, %xmm2951; SSE-NEXT: movdqa %xmm8, %xmm6952; SSE-NEXT: pandn %xmm5, %xmm6953; SSE-NEXT: por %xmm6, %xmm1954; SSE-NEXT: pandn %xmm4, %xmm8955; SSE-NEXT: por %xmm8, %xmm0956; SSE-NEXT: retq957;958; XOP-LABEL: bitselect_v8i64_broadcast_rrm:959; XOP: # %bb.0:960; XOP-NEXT: vbroadcastsd (%rdi), %ymm4961; XOP-NEXT: vpcmov %ymm4, %ymm2, %ymm0, %ymm0962; XOP-NEXT: vpcmov %ymm4, %ymm3, %ymm1, %ymm1963; XOP-NEXT: retq964;965; AVX-LABEL: bitselect_v8i64_broadcast_rrm:966; AVX: # %bb.0:967; AVX-NEXT: vbroadcastsd (%rdi), %ymm4968; AVX-NEXT: vandps %ymm4, %ymm1, %ymm1969; AVX-NEXT: vandps %ymm4, %ymm0, %ymm0970; AVX-NEXT: vandnps %ymm3, %ymm4, %ymm3971; AVX-NEXT: vorps %ymm3, %ymm1, %ymm1972; AVX-NEXT: vandnps %ymm2, %ymm4, %ymm2973; AVX-NEXT: vorps %ymm2, %ymm0, %ymm0974; AVX-NEXT: retq975;976; AVX512-LABEL: bitselect_v8i64_broadcast_rrm:977; AVX512: # %bb.0:978; AVX512-NEXT: vpternlogq {{.*#+}} zmm0 = zmm1 ^ (m64bcst & (zmm0 ^ zmm1))979; AVX512-NEXT: retq980 %a2 = load i64, ptr %p2981 %1 = insertelement <8 x i64> undef, i64 %a2, i32 0982 %2 = shufflevector <8 x i64> %1, <8 x i64> undef, <8 x i32> zeroinitializer983 %3 = xor <8 x i64> %1, <i64 -1, i64 undef, i64 undef, i64 undef, i64 undef, i64 undef, i64 undef, i64 undef>984 %4 = shufflevector <8 x i64> %3, <8 x i64> undef, <8 x i32> zeroinitializer985 %5 = and <8 x i64> %a0, %2986 %6 = and <8 x i64> %a1, %4987 %7 = or <8 x i64> %5, %6988 ret <8 x i64> %7989}990 991; Check that mask registers don't get canonicalized.992define <4 x i1> @bitselect_v4i1_loop(<4 x i32> %a0, <4 x i32> %a1) {993; SSE-LABEL: bitselect_v4i1_loop:994; SSE: # %bb.0: # %bb995; SSE-NEXT: pxor %xmm2, %xmm2996; SSE-NEXT: pcmpeqd %xmm2, %xmm0997; SSE-NEXT: movdqa {{.*#+}} xmm2 = [12,12,12,12]998; SSE-NEXT: pcmpeqd %xmm1, %xmm2999; SSE-NEXT: pcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm11000; SSE-NEXT: pand %xmm0, %xmm11001; SSE-NEXT: pandn %xmm2, %xmm01002; SSE-NEXT: por %xmm1, %xmm01003; SSE-NEXT: retq1004;1005; XOP-LABEL: bitselect_v4i1_loop:1006; XOP: # %bb.0: # %bb1007; XOP-NEXT: vpxor %xmm2, %xmm2, %xmm21008; XOP-NEXT: vpcomneqd %xmm2, %xmm0, %xmm01009; XOP-NEXT: vpcomeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm21010; XOP-NEXT: vpcomeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm11011; XOP-NEXT: vblendvps %xmm0, %xmm2, %xmm1, %xmm01012; XOP-NEXT: retq1013;1014; AVX1-LABEL: bitselect_v4i1_loop:1015; AVX1: # %bb.0: # %bb1016; AVX1-NEXT: vpxor %xmm2, %xmm2, %xmm21017; AVX1-NEXT: vpcmpeqd %xmm2, %xmm0, %xmm01018; AVX1-NEXT: vpcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm21019; AVX1-NEXT: vpcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm11020; AVX1-NEXT: vblendvps %xmm0, %xmm1, %xmm2, %xmm01021; AVX1-NEXT: retq1022;1023; AVX2-LABEL: bitselect_v4i1_loop:1024; AVX2: # %bb.0: # %bb1025; AVX2-NEXT: vpxor %xmm2, %xmm2, %xmm21026; AVX2-NEXT: vpcmpeqd %xmm2, %xmm0, %xmm01027; AVX2-NEXT: vpbroadcastd {{.*#+}} xmm2 = [12,12,12,12]1028; AVX2-NEXT: vpcmpeqd %xmm2, %xmm1, %xmm21029; AVX2-NEXT: vpbroadcastd {{.*#+}} xmm3 = [15,15,15,15]1030; AVX2-NEXT: vpcmpeqd %xmm3, %xmm1, %xmm11031; AVX2-NEXT: vblendvps %xmm0, %xmm1, %xmm2, %xmm01032; AVX2-NEXT: retq1033;1034; AVX512F-LABEL: bitselect_v4i1_loop:1035; AVX512F: # %bb.0: # %bb1036; AVX512F-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm11037; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm01038; AVX512F-NEXT: vpcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm1, %k11039; AVX512F-NEXT: vpcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm1, %k21040; AVX512F-NEXT: vptestnmd %zmm0, %zmm0, %k0 {%k2}1041; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k1 {%k1}1042; AVX512F-NEXT: korw %k0, %k1, %k11043; AVX512F-NEXT: vpternlogd {{.*#+}} zmm0 {%k1} {z} = -11044; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm01045; AVX512F-NEXT: vzeroupper1046; AVX512F-NEXT: retq1047;1048; AVX512VL-LABEL: bitselect_v4i1_loop:1049; AVX512VL: # %bb.0: # %bb1050; AVX512VL-NEXT: vpcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm1, %k11051; AVX512VL-NEXT: vpcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to4}, %xmm1, %k21052; AVX512VL-NEXT: vptestnmd %xmm0, %xmm0, %k0 {%k2}1053; AVX512VL-NEXT: vptestmd %xmm0, %xmm0, %k1 {%k1}1054; AVX512VL-NEXT: korw %k0, %k1, %k11055; AVX512VL-NEXT: vpcmpeqd %xmm0, %xmm0, %xmm01056; AVX512VL-NEXT: vmovdqa32 %xmm0, %xmm0 {%k1} {z}1057; AVX512VL-NEXT: retq1058bb:1059 %tmp = icmp ne <4 x i32> %a0, zeroinitializer1060 %tmp2 = icmp eq <4 x i32> %a1, <i32 12, i32 12, i32 12, i32 12>1061 %tmp3 = icmp eq <4 x i32> %a1, <i32 15, i32 15, i32 15, i32 15>1062 %tmp4 = select <4 x i1> %tmp, <4 x i1> %tmp2, <4 x i1> %tmp31063 ret <4 x i1> %tmp41064}1065 1066; Regression reported on 057db2002bb3d79429db3c5fe436c8cefc50cb251067@d = external global <2 x i64>, align 161068define void @constantfold_andn_mask() nounwind {1069; SSE-LABEL: constantfold_andn_mask:1070; SSE: # %bb.0: # %entry1071; SSE-NEXT: pushq %rax1072; SSE-NEXT: callq use@PLT1073; SSE-NEXT: movdqu (%rax), %xmm11074; SSE-NEXT: movdqa {{.*#+}} xmm2 = [31,248,31,248,31,248,31,248,31,248,31,248,31,248,31,248]1075; SSE-NEXT: pand %xmm2, %xmm01076; SSE-NEXT: pavgb %xmm2, %xmm01077; SSE-NEXT: pandn %xmm1, %xmm01078; SSE-NEXT: pand %xmm2, %xmm11079; SSE-NEXT: pandn %xmm0, %xmm21080; SSE-NEXT: por %xmm1, %xmm21081; SSE-NEXT: movabsq $87960930222080, %rax # imm = 0x5000000000001082; SSE-NEXT: xorq d@GOTPCREL(%rip), %rax1083; SSE-NEXT: movdqa %xmm2, (%rax)1084; SSE-NEXT: popq %rax1085; SSE-NEXT: retq1086;1087; XOP-LABEL: constantfold_andn_mask:1088; XOP: # %bb.0: # %entry1089; XOP-NEXT: pushq %rax1090; XOP-NEXT: callq use@PLT1091; XOP-NEXT: vmovdqu (%rax), %xmm11092; XOP-NEXT: vbroadcastss {{.*#+}} xmm2 = [31,248,31,248,31,248,31,248,31,248,31,248,31,248,31,248]1093; XOP-NEXT: vpand %xmm2, %xmm1, %xmm31094; XOP-NEXT: vpand %xmm2, %xmm0, %xmm01095; XOP-NEXT: vpavgb %xmm2, %xmm0, %xmm01096; XOP-NEXT: vpandn %xmm1, %xmm0, %xmm01097; XOP-NEXT: vpandn %xmm0, %xmm2, %xmm01098; XOP-NEXT: vpor %xmm0, %xmm3, %xmm01099; XOP-NEXT: movabsq $87960930222080, %rax # imm = 0x5000000000001100; XOP-NEXT: xorq d@GOTPCREL(%rip), %rax1101; XOP-NEXT: vmovdqa %xmm0, (%rax)1102; XOP-NEXT: popq %rax1103; XOP-NEXT: retq1104;1105; AVX1-LABEL: constantfold_andn_mask:1106; AVX1: # %bb.0: # %entry1107; AVX1-NEXT: pushq %rax1108; AVX1-NEXT: callq use@PLT1109; AVX1-NEXT: vmovdqu (%rax), %xmm11110; AVX1-NEXT: vbroadcastss {{.*#+}} xmm2 = [31,248,31,248,31,248,31,248,31,248,31,248,31,248,31,248]1111; AVX1-NEXT: vpand %xmm2, %xmm1, %xmm31112; AVX1-NEXT: vpand %xmm2, %xmm0, %xmm01113; AVX1-NEXT: vpavgb %xmm2, %xmm0, %xmm01114; AVX1-NEXT: vpandn %xmm1, %xmm0, %xmm01115; AVX1-NEXT: vpandn %xmm0, %xmm2, %xmm01116; AVX1-NEXT: vpor %xmm0, %xmm3, %xmm01117; AVX1-NEXT: movabsq $87960930222080, %rax # imm = 0x5000000000001118; AVX1-NEXT: xorq d@GOTPCREL(%rip), %rax1119; AVX1-NEXT: vmovdqa %xmm0, (%rax)1120; AVX1-NEXT: popq %rax1121; AVX1-NEXT: retq1122;1123; AVX2-LABEL: constantfold_andn_mask:1124; AVX2: # %bb.0: # %entry1125; AVX2-NEXT: pushq %rax1126; AVX2-NEXT: callq use@PLT1127; AVX2-NEXT: vmovdqu (%rax), %xmm11128; AVX2-NEXT: vpbroadcastw {{.*#+}} xmm2 = [31,248,31,248,31,248,31,248,31,248,31,248,31,248,31,248]1129; AVX2-NEXT: vpand %xmm2, %xmm1, %xmm31130; AVX2-NEXT: vpand %xmm2, %xmm0, %xmm01131; AVX2-NEXT: vpavgb %xmm2, %xmm0, %xmm01132; AVX2-NEXT: vpandn %xmm1, %xmm0, %xmm01133; AVX2-NEXT: vpandn %xmm0, %xmm2, %xmm01134; AVX2-NEXT: vpor %xmm0, %xmm3, %xmm01135; AVX2-NEXT: movabsq $87960930222080, %rax # imm = 0x5000000000001136; AVX2-NEXT: xorq d@GOTPCREL(%rip), %rax1137; AVX2-NEXT: vmovdqa %xmm0, (%rax)1138; AVX2-NEXT: popq %rax1139; AVX2-NEXT: retq1140;1141; AVX512F-LABEL: constantfold_andn_mask:1142; AVX512F: # %bb.0: # %entry1143; AVX512F-NEXT: pushq %rax1144; AVX512F-NEXT: callq use@PLT1145; AVX512F-NEXT: vmovdqu (%rax), %xmm11146; AVX512F-NEXT: vpbroadcastw {{.*#+}} xmm2 = [31,248,31,248,31,248,31,248,31,248,31,248,31,248,31,248]1147; AVX512F-NEXT: vpand %xmm2, %xmm0, %xmm01148; AVX512F-NEXT: vpavgb %xmm2, %xmm0, %xmm01149; AVX512F-NEXT: vpandn %xmm1, %xmm0, %xmm01150; AVX512F-NEXT: vpternlogq {{.*#+}} zmm0 = zmm0 ^ (zmm2 & (zmm0 ^ zmm1))1151; AVX512F-NEXT: movabsq $87960930222080, %rax # imm = 0x5000000000001152; AVX512F-NEXT: xorq d@GOTPCREL(%rip), %rax1153; AVX512F-NEXT: vmovdqa %xmm0, (%rax)1154; AVX512F-NEXT: popq %rax1155; AVX512F-NEXT: vzeroupper1156; AVX512F-NEXT: retq1157;1158; AVX512VL-LABEL: constantfold_andn_mask:1159; AVX512VL: # %bb.0: # %entry1160; AVX512VL-NEXT: pushq %rax1161; AVX512VL-NEXT: callq use@PLT1162; AVX512VL-NEXT: vmovdqu (%rax), %xmm11163; AVX512VL-NEXT: vpbroadcastw {{.*#+}} xmm2 = [31,248,31,248,31,248,31,248,31,248,31,248,31,248,31,248]1164; AVX512VL-NEXT: vpand %xmm2, %xmm0, %xmm01165; AVX512VL-NEXT: vpavgb %xmm2, %xmm0, %xmm01166; AVX512VL-NEXT: vpandn %xmm1, %xmm0, %xmm01167; AVX512VL-NEXT: vpternlogq {{.*#+}} xmm0 = xmm0 ^ (xmm2 & (xmm0 ^ xmm1))1168; AVX512VL-NEXT: movabsq $87960930222080, %rax # imm = 0x5000000000001169; AVX512VL-NEXT: xorq d@GOTPCREL(%rip), %rax1170; AVX512VL-NEXT: vmovdqa %xmm0, (%rax)1171; AVX512VL-NEXT: popq %rax1172; AVX512VL-NEXT: retq1173entry:1174 %call = call noundef <2 x i64> @use()1175 %_msret = load <2 x i64>, ptr undef, align 81176 %i = bitcast <2 x i64> %_msret to <16 x i8>1177 %i1 = bitcast <2 x i64> %call to <16 x i8>1178 %i2 = and <16 x i8> %i, <i8 31, i8 -8, i8 31, i8 -8, i8 31, i8 -8, i8 31, i8 -8, i8 31, i8 -8, i8 31, i8 -8, i8 31, i8 -8, i8 31, i8 -8>1179 %i3 = and <16 x i8> %i1, <i8 31, i8 -8, i8 31, i8 -8, i8 31, i8 -8, i8 31, i8 -8, i8 31, i8 -8, i8 31, i8 -8, i8 31, i8 -8, i8 31, i8 -8>1180 %i4 = call <16 x i8> @llvm.x86.sse2.pavg.b(<16 x i8> <i8 31, i8 -8, i8 31, i8 -8, i8 31, i8 -8, i8 31, i8 -8, i8 31, i8 -8, i8 31, i8 -8, i8 31, i8 -8, i8 31, i8 -8>, <16 x i8> %i3)1181 %i5 = bitcast <16 x i8> %i2 to <2 x i64>1182 %i6 = bitcast <16 x i8> %i4 to <2 x i64>1183 %i7 = and <2 x i64> %_msret, <i64 567462211834873824, i64 567462211834873824>1184 %i8 = xor <2 x i64> zeroinitializer, <i64 -1, i64 -1>1185 %i9 = xor <2 x i64> %i6, <i64 -1, i64 -1>1186 %i10 = and <2 x i64> %i8, %i51187 %i11 = and <2 x i64> %i7, %i91188 %i12 = or <2 x i64> zeroinitializer, %i101189 %i13 = or <2 x i64> %i12, %i111190 store <2 x i64> %i13, ptr inttoptr (i64 xor (i64 ptrtoint (ptr @d to i64), i64 87960930222080) to ptr), align 161191 ret void1192}1193 1194declare <2 x i64> @use()1195declare <16 x i8> @llvm.x86.sse2.pavg.b(<16 x i8>, <16 x i8>)1196