703 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefix=AVX512F3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw | FileCheck %s --check-prefix=AVX512BW4; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vpopcntdq | FileCheck %s --check-prefixes=AVX512VPOPCNTDQ,AVX512VPOPCNTDQ-NOBW5; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vpopcntdq,+avx512bw | FileCheck %s --check-prefixes=AVX512VPOPCNTDQ,AVX512VPOPCNTDQ-BW6; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bitalg | FileCheck %s --check-prefix=BITALG7 8define <8 x i64> @testv8i64(<8 x i64> %in) nounwind {9; AVX512F-LABEL: testv8i64:10; AVX512F: # %bb.0:11; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm112; AVX512F-NEXT: vpbroadcastb {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]13; AVX512F-NEXT: vpand %ymm2, %ymm1, %ymm314; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]15; AVX512F-NEXT: # ymm4 = mem[0,1,0,1]16; AVX512F-NEXT: vpshufb %ymm3, %ymm4, %ymm317; AVX512F-NEXT: vpsrlw $4, %ymm1, %ymm118; AVX512F-NEXT: vpand %ymm2, %ymm1, %ymm119; AVX512F-NEXT: vpshufb %ymm1, %ymm4, %ymm120; AVX512F-NEXT: vpaddb %ymm3, %ymm1, %ymm121; AVX512F-NEXT: vpxor %xmm3, %xmm3, %xmm322; AVX512F-NEXT: vpsadbw %ymm3, %ymm1, %ymm123; AVX512F-NEXT: vpand %ymm2, %ymm0, %ymm524; AVX512F-NEXT: vpshufb %ymm5, %ymm4, %ymm525; AVX512F-NEXT: vpsrlw $4, %ymm0, %ymm026; AVX512F-NEXT: vpand %ymm2, %ymm0, %ymm027; AVX512F-NEXT: vpshufb %ymm0, %ymm4, %ymm028; AVX512F-NEXT: vpaddb %ymm5, %ymm0, %ymm029; AVX512F-NEXT: vpsadbw %ymm3, %ymm0, %ymm030; AVX512F-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm031; AVX512F-NEXT: retq32;33; AVX512BW-LABEL: testv8i64:34; AVX512BW: # %bb.0:35; AVX512BW-NEXT: vpbroadcastb {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]36; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm237; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]38; AVX512BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]39; AVX512BW-NEXT: vpshufb %zmm2, %zmm3, %zmm240; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm041; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm042; AVX512BW-NEXT: vpshufb %zmm0, %zmm3, %zmm043; AVX512BW-NEXT: vpaddb %zmm2, %zmm0, %zmm044; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm145; AVX512BW-NEXT: vpsadbw %zmm1, %zmm0, %zmm046; AVX512BW-NEXT: retq47;48; AVX512VPOPCNTDQ-LABEL: testv8i64:49; AVX512VPOPCNTDQ: # %bb.0:50; AVX512VPOPCNTDQ-NEXT: vpopcntq %zmm0, %zmm051; AVX512VPOPCNTDQ-NEXT: retq52;53; BITALG-LABEL: testv8i64:54; BITALG: # %bb.0:55; BITALG-NEXT: vpopcntb %zmm0, %zmm056; BITALG-NEXT: vpxor %xmm1, %xmm1, %xmm157; BITALG-NEXT: vpsadbw %zmm1, %zmm0, %zmm058; BITALG-NEXT: retq59 %out = call <8 x i64> @llvm.ctpop.v8i64(<8 x i64> %in)60 ret <8 x i64> %out61}62 63define <16 x i32> @testv16i32(<16 x i32> %in) nounwind {64; AVX512F-LABEL: testv16i32:65; AVX512F: # %bb.0:66; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm167; AVX512F-NEXT: vpbroadcastb {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]68; AVX512F-NEXT: vpand %ymm2, %ymm1, %ymm369; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]70; AVX512F-NEXT: # ymm4 = mem[0,1,0,1]71; AVX512F-NEXT: vpshufb %ymm3, %ymm4, %ymm372; AVX512F-NEXT: vpsrlw $4, %ymm1, %ymm173; AVX512F-NEXT: vpand %ymm2, %ymm1, %ymm174; AVX512F-NEXT: vpshufb %ymm1, %ymm4, %ymm175; AVX512F-NEXT: vpaddb %ymm3, %ymm1, %ymm176; AVX512F-NEXT: vpxor %xmm3, %xmm3, %xmm377; AVX512F-NEXT: vpunpckhdq {{.*#+}} ymm5 = ymm1[2],ymm3[2],ymm1[3],ymm3[3],ymm1[6],ymm3[6],ymm1[7],ymm3[7]78; AVX512F-NEXT: vpsadbw %ymm3, %ymm5, %ymm579; AVX512F-NEXT: vpunpckldq {{.*#+}} ymm1 = ymm1[0],ymm3[0],ymm1[1],ymm3[1],ymm1[4],ymm3[4],ymm1[5],ymm3[5]80; AVX512F-NEXT: vpsadbw %ymm3, %ymm1, %ymm181; AVX512F-NEXT: vpackuswb %ymm5, %ymm1, %ymm182; AVX512F-NEXT: vpand %ymm2, %ymm0, %ymm583; AVX512F-NEXT: vpshufb %ymm5, %ymm4, %ymm584; AVX512F-NEXT: vpsrlw $4, %ymm0, %ymm085; AVX512F-NEXT: vpand %ymm2, %ymm0, %ymm086; AVX512F-NEXT: vpshufb %ymm0, %ymm4, %ymm087; AVX512F-NEXT: vpaddb %ymm5, %ymm0, %ymm088; AVX512F-NEXT: vpunpckhdq {{.*#+}} ymm2 = ymm0[2],ymm3[2],ymm0[3],ymm3[3],ymm0[6],ymm3[6],ymm0[7],ymm3[7]89; AVX512F-NEXT: vpsadbw %ymm3, %ymm2, %ymm290; AVX512F-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm3[0],ymm0[1],ymm3[1],ymm0[4],ymm3[4],ymm0[5],ymm3[5]91; AVX512F-NEXT: vpsadbw %ymm3, %ymm0, %ymm092; AVX512F-NEXT: vpackuswb %ymm2, %ymm0, %ymm093; AVX512F-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm094; AVX512F-NEXT: retq95;96; AVX512BW-LABEL: testv16i32:97; AVX512BW: # %bb.0:98; AVX512BW-NEXT: vpbroadcastb {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]99; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm2100; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]101; AVX512BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]102; AVX512BW-NEXT: vpshufb %zmm2, %zmm3, %zmm2103; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0104; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0105; AVX512BW-NEXT: vpshufb %zmm0, %zmm3, %zmm0106; AVX512BW-NEXT: vpaddb %zmm2, %zmm0, %zmm0107; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1108; AVX512BW-NEXT: vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]109; AVX512BW-NEXT: vpsadbw %zmm1, %zmm2, %zmm2110; AVX512BW-NEXT: vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]111; AVX512BW-NEXT: vpsadbw %zmm1, %zmm0, %zmm0112; AVX512BW-NEXT: vpackuswb %zmm2, %zmm0, %zmm0113; AVX512BW-NEXT: retq114;115; AVX512VPOPCNTDQ-LABEL: testv16i32:116; AVX512VPOPCNTDQ: # %bb.0:117; AVX512VPOPCNTDQ-NEXT: vpopcntd %zmm0, %zmm0118; AVX512VPOPCNTDQ-NEXT: retq119;120; BITALG-LABEL: testv16i32:121; BITALG: # %bb.0:122; BITALG-NEXT: vpopcntb %zmm0, %zmm0123; BITALG-NEXT: vpxor %xmm1, %xmm1, %xmm1124; BITALG-NEXT: vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]125; BITALG-NEXT: vpsadbw %zmm1, %zmm2, %zmm2126; BITALG-NEXT: vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]127; BITALG-NEXT: vpsadbw %zmm1, %zmm0, %zmm0128; BITALG-NEXT: vpackuswb %zmm2, %zmm0, %zmm0129; BITALG-NEXT: retq130 %out = call <16 x i32> @llvm.ctpop.v16i32(<16 x i32> %in)131 ret <16 x i32> %out132}133 134define <32 x i16> @testv32i16(<32 x i16> %in) nounwind {135; AVX512F-LABEL: testv32i16:136; AVX512F: # %bb.0:137; AVX512F-NEXT: vpbroadcastb {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]138; AVX512F-NEXT: vpand %ymm1, %ymm0, %ymm2139; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]140; AVX512F-NEXT: # ymm3 = mem[0,1,0,1]141; AVX512F-NEXT: vpshufb %ymm2, %ymm3, %ymm2142; AVX512F-NEXT: vpsrlw $4, %ymm0, %ymm4143; AVX512F-NEXT: vpand %ymm1, %ymm4, %ymm4144; AVX512F-NEXT: vpshufb %ymm4, %ymm3, %ymm4145; AVX512F-NEXT: vpaddb %ymm2, %ymm4, %ymm2146; AVX512F-NEXT: vpsllw $8, %ymm2, %ymm4147; AVX512F-NEXT: vpaddb %ymm2, %ymm4, %ymm2148; AVX512F-NEXT: vpsrlw $8, %ymm2, %ymm2149; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm0150; AVX512F-NEXT: vpand %ymm1, %ymm0, %ymm4151; AVX512F-NEXT: vpshufb %ymm4, %ymm3, %ymm4152; AVX512F-NEXT: vpsrlw $4, %ymm0, %ymm0153; AVX512F-NEXT: vpand %ymm1, %ymm0, %ymm0154; AVX512F-NEXT: vpshufb %ymm0, %ymm3, %ymm0155; AVX512F-NEXT: vpaddb %ymm4, %ymm0, %ymm0156; AVX512F-NEXT: vpsllw $8, %ymm0, %ymm1157; AVX512F-NEXT: vpaddb %ymm0, %ymm1, %ymm0158; AVX512F-NEXT: vpsrlw $8, %ymm0, %ymm0159; AVX512F-NEXT: vinserti64x4 $1, %ymm0, %zmm2, %zmm0160; AVX512F-NEXT: retq161;162; AVX512BW-LABEL: testv32i16:163; AVX512BW: # %bb.0:164; AVX512BW-NEXT: vpbroadcastb {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]165; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm2166; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]167; AVX512BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]168; AVX512BW-NEXT: vpshufb %zmm2, %zmm3, %zmm2169; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0170; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0171; AVX512BW-NEXT: vpshufb %zmm0, %zmm3, %zmm0172; AVX512BW-NEXT: vpaddb %zmm2, %zmm0, %zmm0173; AVX512BW-NEXT: vpsllw $8, %zmm0, %zmm1174; AVX512BW-NEXT: vpaddb %zmm0, %zmm1, %zmm0175; AVX512BW-NEXT: vpsrlw $8, %zmm0, %zmm0176; AVX512BW-NEXT: retq177;178; AVX512VPOPCNTDQ-NOBW-LABEL: testv32i16:179; AVX512VPOPCNTDQ-NOBW: # %bb.0:180; AVX512VPOPCNTDQ-NOBW-NEXT: vpmovzxwd {{.*#+}} zmm1 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero181; AVX512VPOPCNTDQ-NOBW-NEXT: vpopcntd %zmm1, %zmm1182; AVX512VPOPCNTDQ-NOBW-NEXT: vpmovdw %zmm1, %ymm1183; AVX512VPOPCNTDQ-NOBW-NEXT: vextracti64x4 $1, %zmm0, %ymm0184; AVX512VPOPCNTDQ-NOBW-NEXT: vpmovzxwd {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero185; AVX512VPOPCNTDQ-NOBW-NEXT: vpopcntd %zmm0, %zmm0186; AVX512VPOPCNTDQ-NOBW-NEXT: vpmovdw %zmm0, %ymm0187; AVX512VPOPCNTDQ-NOBW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0188; AVX512VPOPCNTDQ-NOBW-NEXT: retq189;190; AVX512VPOPCNTDQ-BW-LABEL: testv32i16:191; AVX512VPOPCNTDQ-BW: # %bb.0:192; AVX512VPOPCNTDQ-BW-NEXT: vpbroadcastb {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]193; AVX512VPOPCNTDQ-BW-NEXT: vpandq %zmm1, %zmm0, %zmm2194; AVX512VPOPCNTDQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]195; AVX512VPOPCNTDQ-BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]196; AVX512VPOPCNTDQ-BW-NEXT: vpshufb %zmm2, %zmm3, %zmm2197; AVX512VPOPCNTDQ-BW-NEXT: vpsrlw $4, %zmm0, %zmm0198; AVX512VPOPCNTDQ-BW-NEXT: vpandq %zmm1, %zmm0, %zmm0199; AVX512VPOPCNTDQ-BW-NEXT: vpshufb %zmm0, %zmm3, %zmm0200; AVX512VPOPCNTDQ-BW-NEXT: vpaddb %zmm2, %zmm0, %zmm0201; AVX512VPOPCNTDQ-BW-NEXT: vpsllw $8, %zmm0, %zmm1202; AVX512VPOPCNTDQ-BW-NEXT: vpaddb %zmm0, %zmm1, %zmm0203; AVX512VPOPCNTDQ-BW-NEXT: vpsrlw $8, %zmm0, %zmm0204; AVX512VPOPCNTDQ-BW-NEXT: retq205;206; BITALG-LABEL: testv32i16:207; BITALG: # %bb.0:208; BITALG-NEXT: vpopcntw %zmm0, %zmm0209; BITALG-NEXT: retq210 %out = call <32 x i16> @llvm.ctpop.v32i16(<32 x i16> %in)211 ret <32 x i16> %out212}213 214define <64 x i8> @testv64i8(<64 x i8> %in) nounwind {215; AVX512F-LABEL: testv64i8:216; AVX512F: # %bb.0:217; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm1218; AVX512F-NEXT: vpbroadcastb {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]219; AVX512F-NEXT: vpand %ymm2, %ymm1, %ymm3220; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]221; AVX512F-NEXT: # ymm4 = mem[0,1,0,1]222; AVX512F-NEXT: vpshufb %ymm3, %ymm4, %ymm3223; AVX512F-NEXT: vpsrlw $4, %ymm1, %ymm1224; AVX512F-NEXT: vpand %ymm2, %ymm1, %ymm1225; AVX512F-NEXT: vpshufb %ymm1, %ymm4, %ymm1226; AVX512F-NEXT: vpaddb %ymm3, %ymm1, %ymm1227; AVX512F-NEXT: vpand %ymm2, %ymm0, %ymm3228; AVX512F-NEXT: vpshufb %ymm3, %ymm4, %ymm3229; AVX512F-NEXT: vpsrlw $4, %ymm0, %ymm0230; AVX512F-NEXT: vpand %ymm2, %ymm0, %ymm0231; AVX512F-NEXT: vpshufb %ymm0, %ymm4, %ymm0232; AVX512F-NEXT: vpaddb %ymm3, %ymm0, %ymm0233; AVX512F-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0234; AVX512F-NEXT: retq235;236; AVX512BW-LABEL: testv64i8:237; AVX512BW: # %bb.0:238; AVX512BW-NEXT: vpbroadcastb {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]239; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm2240; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]241; AVX512BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]242; AVX512BW-NEXT: vpshufb %zmm2, %zmm3, %zmm2243; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0244; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0245; AVX512BW-NEXT: vpshufb %zmm0, %zmm3, %zmm0246; AVX512BW-NEXT: vpaddb %zmm2, %zmm0, %zmm0247; AVX512BW-NEXT: retq248;249; AVX512VPOPCNTDQ-NOBW-LABEL: testv64i8:250; AVX512VPOPCNTDQ-NOBW: # %bb.0:251; AVX512VPOPCNTDQ-NOBW-NEXT: vextracti64x4 $1, %zmm0, %ymm1252; AVX512VPOPCNTDQ-NOBW-NEXT: vpbroadcastb {{.*#+}} ymm2 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]253; AVX512VPOPCNTDQ-NOBW-NEXT: vpand %ymm2, %ymm1, %ymm3254; AVX512VPOPCNTDQ-NOBW-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]255; AVX512VPOPCNTDQ-NOBW-NEXT: # ymm4 = mem[0,1,0,1]256; AVX512VPOPCNTDQ-NOBW-NEXT: vpshufb %ymm3, %ymm4, %ymm3257; AVX512VPOPCNTDQ-NOBW-NEXT: vpsrlw $4, %ymm1, %ymm1258; AVX512VPOPCNTDQ-NOBW-NEXT: vpand %ymm2, %ymm1, %ymm1259; AVX512VPOPCNTDQ-NOBW-NEXT: vpshufb %ymm1, %ymm4, %ymm1260; AVX512VPOPCNTDQ-NOBW-NEXT: vpaddb %ymm3, %ymm1, %ymm1261; AVX512VPOPCNTDQ-NOBW-NEXT: vpand %ymm2, %ymm0, %ymm3262; AVX512VPOPCNTDQ-NOBW-NEXT: vpshufb %ymm3, %ymm4, %ymm3263; AVX512VPOPCNTDQ-NOBW-NEXT: vpsrlw $4, %ymm0, %ymm0264; AVX512VPOPCNTDQ-NOBW-NEXT: vpand %ymm2, %ymm0, %ymm0265; AVX512VPOPCNTDQ-NOBW-NEXT: vpshufb %ymm0, %ymm4, %ymm0266; AVX512VPOPCNTDQ-NOBW-NEXT: vpaddb %ymm3, %ymm0, %ymm0267; AVX512VPOPCNTDQ-NOBW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0268; AVX512VPOPCNTDQ-NOBW-NEXT: retq269;270; AVX512VPOPCNTDQ-BW-LABEL: testv64i8:271; AVX512VPOPCNTDQ-BW: # %bb.0:272; AVX512VPOPCNTDQ-BW-NEXT: vpbroadcastb {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]273; AVX512VPOPCNTDQ-BW-NEXT: vpandq %zmm1, %zmm0, %zmm2274; AVX512VPOPCNTDQ-BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]275; AVX512VPOPCNTDQ-BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]276; AVX512VPOPCNTDQ-BW-NEXT: vpshufb %zmm2, %zmm3, %zmm2277; AVX512VPOPCNTDQ-BW-NEXT: vpsrlw $4, %zmm0, %zmm0278; AVX512VPOPCNTDQ-BW-NEXT: vpandq %zmm1, %zmm0, %zmm0279; AVX512VPOPCNTDQ-BW-NEXT: vpshufb %zmm0, %zmm3, %zmm0280; AVX512VPOPCNTDQ-BW-NEXT: vpaddb %zmm2, %zmm0, %zmm0281; AVX512VPOPCNTDQ-BW-NEXT: retq282;283; BITALG-LABEL: testv64i8:284; BITALG: # %bb.0:285; BITALG-NEXT: vpopcntb %zmm0, %zmm0286; BITALG-NEXT: retq287 %out = call <64 x i8> @llvm.ctpop.v64i8(<64 x i8> %in)288 ret <64 x i8> %out289}290 291define <8 x i64> @eq_1_v8i64(<8 x i64> %0) {292; AVX512F-LABEL: eq_1_v8i64:293; AVX512F: # %bb.0:294; AVX512F-NEXT: vpternlogd {{.*#+}} zmm1 = -1295; AVX512F-NEXT: vpaddq %zmm1, %zmm0, %zmm1296; AVX512F-NEXT: vpxorq %zmm1, %zmm0, %zmm0297; AVX512F-NEXT: vpcmpnleuq %zmm1, %zmm0, %k1298; AVX512F-NEXT: vpternlogq {{.*#+}} zmm0 {%k1} {z} = -1299; AVX512F-NEXT: retq300;301; AVX512BW-LABEL: eq_1_v8i64:302; AVX512BW: # %bb.0:303; AVX512BW-NEXT: vpternlogd {{.*#+}} zmm1 = -1304; AVX512BW-NEXT: vpaddq %zmm1, %zmm0, %zmm1305; AVX512BW-NEXT: vpxorq %zmm1, %zmm0, %zmm0306; AVX512BW-NEXT: vpcmpnleuq %zmm1, %zmm0, %k1307; AVX512BW-NEXT: vpternlogq {{.*#+}} zmm0 {%k1} {z} = -1308; AVX512BW-NEXT: retq309;310; AVX512VPOPCNTDQ-LABEL: eq_1_v8i64:311; AVX512VPOPCNTDQ: # %bb.0:312; AVX512VPOPCNTDQ-NEXT: vpopcntq %zmm0, %zmm0313; AVX512VPOPCNTDQ-NEXT: vpcmpeqq {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to8}, %zmm0, %k1314; AVX512VPOPCNTDQ-NEXT: vpternlogq {{.*#+}} zmm0 {%k1} {z} = -1315; AVX512VPOPCNTDQ-NEXT: retq316;317; BITALG-LABEL: eq_1_v8i64:318; BITALG: # %bb.0:319; BITALG-NEXT: vpternlogd {{.*#+}} zmm1 = -1320; BITALG-NEXT: vpaddq %zmm1, %zmm0, %zmm1321; BITALG-NEXT: vpxorq %zmm1, %zmm0, %zmm0322; BITALG-NEXT: vpcmpnleuq %zmm1, %zmm0, %k1323; BITALG-NEXT: vpternlogq {{.*#+}} zmm0 {%k1} {z} = -1324; BITALG-NEXT: retq325 %2 = tail call <8 x i64> @llvm.ctpop.v8i64(<8 x i64> %0)326 %3 = icmp eq <8 x i64> %2, <i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1>327 %4 = sext <8 x i1> %3 to <8 x i64>328 ret <8 x i64> %4329}330 331define <8 x i64> @ne_1_v8i64(<8 x i64> %0) {332; AVX512F-LABEL: ne_1_v8i64:333; AVX512F: # %bb.0:334; AVX512F-NEXT: vpternlogd {{.*#+}} zmm1 = -1335; AVX512F-NEXT: vpaddq %zmm1, %zmm0, %zmm1336; AVX512F-NEXT: vpxorq %zmm1, %zmm0, %zmm0337; AVX512F-NEXT: vpcmpleuq %zmm1, %zmm0, %k1338; AVX512F-NEXT: vpternlogq {{.*#+}} zmm0 {%k1} {z} = -1339; AVX512F-NEXT: retq340;341; AVX512BW-LABEL: ne_1_v8i64:342; AVX512BW: # %bb.0:343; AVX512BW-NEXT: vpternlogd {{.*#+}} zmm1 = -1344; AVX512BW-NEXT: vpaddq %zmm1, %zmm0, %zmm1345; AVX512BW-NEXT: vpxorq %zmm1, %zmm0, %zmm0346; AVX512BW-NEXT: vpcmpleuq %zmm1, %zmm0, %k1347; AVX512BW-NEXT: vpternlogq {{.*#+}} zmm0 {%k1} {z} = -1348; AVX512BW-NEXT: retq349;350; AVX512VPOPCNTDQ-LABEL: ne_1_v8i64:351; AVX512VPOPCNTDQ: # %bb.0:352; AVX512VPOPCNTDQ-NEXT: vpopcntq %zmm0, %zmm0353; AVX512VPOPCNTDQ-NEXT: vpcmpneqq {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to8}, %zmm0, %k1354; AVX512VPOPCNTDQ-NEXT: vpternlogq {{.*#+}} zmm0 {%k1} {z} = -1355; AVX512VPOPCNTDQ-NEXT: retq356;357; BITALG-LABEL: ne_1_v8i64:358; BITALG: # %bb.0:359; BITALG-NEXT: vpternlogd {{.*#+}} zmm1 = -1360; BITALG-NEXT: vpaddq %zmm1, %zmm0, %zmm1361; BITALG-NEXT: vpxorq %zmm1, %zmm0, %zmm0362; BITALG-NEXT: vpcmpleuq %zmm1, %zmm0, %k1363; BITALG-NEXT: vpternlogq {{.*#+}} zmm0 {%k1} {z} = -1364; BITALG-NEXT: retq365 %2 = tail call <8 x i64> @llvm.ctpop.v8i64(<8 x i64> %0)366 %3 = icmp ne <8 x i64> %2, <i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1, i64 1>367 %4 = sext <8 x i1> %3 to <8 x i64>368 ret <8 x i64> %4369}370 371define <16 x i32> @eq_1_v16i32(<16 x i32> %0) {372; AVX512F-LABEL: eq_1_v16i32:373; AVX512F: # %bb.0:374; AVX512F-NEXT: vpternlogd {{.*#+}} zmm1 = -1375; AVX512F-NEXT: vpaddd %zmm1, %zmm0, %zmm1376; AVX512F-NEXT: vpxord %zmm1, %zmm0, %zmm0377; AVX512F-NEXT: vpcmpnleud %zmm1, %zmm0, %k1378; AVX512F-NEXT: vpternlogd {{.*#+}} zmm0 {%k1} {z} = -1379; AVX512F-NEXT: retq380;381; AVX512BW-LABEL: eq_1_v16i32:382; AVX512BW: # %bb.0:383; AVX512BW-NEXT: vpternlogd {{.*#+}} zmm1 = -1384; AVX512BW-NEXT: vpaddd %zmm1, %zmm0, %zmm1385; AVX512BW-NEXT: vpxord %zmm1, %zmm0, %zmm0386; AVX512BW-NEXT: vpcmpnleud %zmm1, %zmm0, %k1387; AVX512BW-NEXT: vpternlogd {{.*#+}} zmm0 {%k1} {z} = -1388; AVX512BW-NEXT: retq389;390; AVX512VPOPCNTDQ-LABEL: eq_1_v16i32:391; AVX512VPOPCNTDQ: # %bb.0:392; AVX512VPOPCNTDQ-NEXT: vpopcntd %zmm0, %zmm0393; AVX512VPOPCNTDQ-NEXT: vpcmpeqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm0, %k1394; AVX512VPOPCNTDQ-NEXT: vpternlogd {{.*#+}} zmm0 {%k1} {z} = -1395; AVX512VPOPCNTDQ-NEXT: retq396;397; BITALG-LABEL: eq_1_v16i32:398; BITALG: # %bb.0:399; BITALG-NEXT: vpternlogd {{.*#+}} zmm1 = -1400; BITALG-NEXT: vpaddd %zmm1, %zmm0, %zmm1401; BITALG-NEXT: vpxord %zmm1, %zmm0, %zmm0402; BITALG-NEXT: vpcmpnleud %zmm1, %zmm0, %k1403; BITALG-NEXT: vpternlogd {{.*#+}} zmm0 {%k1} {z} = -1404; BITALG-NEXT: retq405 %2 = tail call <16 x i32> @llvm.ctpop.v16i32(<16 x i32> %0)406 %3 = icmp eq <16 x i32> %2, <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>407 %4 = sext <16 x i1> %3 to <16 x i32>408 ret <16 x i32> %4409}410 411define <16 x i32> @ne_1_v16i32(<16 x i32> %0) {412; AVX512F-LABEL: ne_1_v16i32:413; AVX512F: # %bb.0:414; AVX512F-NEXT: vpternlogd {{.*#+}} zmm1 = -1415; AVX512F-NEXT: vpaddd %zmm1, %zmm0, %zmm1416; AVX512F-NEXT: vpxord %zmm1, %zmm0, %zmm0417; AVX512F-NEXT: vpcmpleud %zmm1, %zmm0, %k1418; AVX512F-NEXT: vpternlogd {{.*#+}} zmm0 {%k1} {z} = -1419; AVX512F-NEXT: retq420;421; AVX512BW-LABEL: ne_1_v16i32:422; AVX512BW: # %bb.0:423; AVX512BW-NEXT: vpternlogd {{.*#+}} zmm1 = -1424; AVX512BW-NEXT: vpaddd %zmm1, %zmm0, %zmm1425; AVX512BW-NEXT: vpxord %zmm1, %zmm0, %zmm0426; AVX512BW-NEXT: vpcmpleud %zmm1, %zmm0, %k1427; AVX512BW-NEXT: vpternlogd {{.*#+}} zmm0 {%k1} {z} = -1428; AVX512BW-NEXT: retq429;430; AVX512VPOPCNTDQ-LABEL: ne_1_v16i32:431; AVX512VPOPCNTDQ: # %bb.0:432; AVX512VPOPCNTDQ-NEXT: vpopcntd %zmm0, %zmm0433; AVX512VPOPCNTDQ-NEXT: vpcmpneqd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm0, %k1434; AVX512VPOPCNTDQ-NEXT: vpternlogd {{.*#+}} zmm0 {%k1} {z} = -1435; AVX512VPOPCNTDQ-NEXT: retq436;437; BITALG-LABEL: ne_1_v16i32:438; BITALG: # %bb.0:439; BITALG-NEXT: vpternlogd {{.*#+}} zmm1 = -1440; BITALG-NEXT: vpaddd %zmm1, %zmm0, %zmm1441; BITALG-NEXT: vpxord %zmm1, %zmm0, %zmm0442; BITALG-NEXT: vpcmpleud %zmm1, %zmm0, %k1443; BITALG-NEXT: vpternlogd {{.*#+}} zmm0 {%k1} {z} = -1444; BITALG-NEXT: retq445 %2 = tail call <16 x i32> @llvm.ctpop.v16i32(<16 x i32> %0)446 %3 = icmp ne <16 x i32> %2, <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>447 %4 = sext <16 x i1> %3 to <16 x i32>448 ret <16 x i32> %4449}450 451define <32 x i16> @eq_1_v32i16(<32 x i16> %0) {452; AVX512F-LABEL: eq_1_v32i16:453; AVX512F: # %bb.0:454; AVX512F-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1455; AVX512F-NEXT: vpaddw %ymm1, %ymm0, %ymm2456; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm3457; AVX512F-NEXT: vpaddw %ymm1, %ymm3, %ymm1458; AVX512F-NEXT: vpxor %ymm1, %ymm3, %ymm3459; AVX512F-NEXT: vpminuw %ymm1, %ymm3, %ymm1460; AVX512F-NEXT: vpcmpeqw %ymm1, %ymm3, %ymm1461; AVX512F-NEXT: vpxor %ymm2, %ymm0, %ymm0462; AVX512F-NEXT: vpminuw %ymm2, %ymm0, %ymm2463; AVX512F-NEXT: vpcmpeqw %ymm2, %ymm0, %ymm0464; AVX512F-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0465; AVX512F-NEXT: vpternlogq {{.*#+}} zmm0 = ~zmm0466; AVX512F-NEXT: retq467;468; AVX512BW-LABEL: eq_1_v32i16:469; AVX512BW: # %bb.0:470; AVX512BW-NEXT: vpternlogd {{.*#+}} zmm1 = -1471; AVX512BW-NEXT: vpaddw %zmm1, %zmm0, %zmm1472; AVX512BW-NEXT: vpxorq %zmm1, %zmm0, %zmm0473; AVX512BW-NEXT: vpcmpnleuw %zmm1, %zmm0, %k0474; AVX512BW-NEXT: vpmovm2w %k0, %zmm0475; AVX512BW-NEXT: retq476;477; AVX512VPOPCNTDQ-NOBW-LABEL: eq_1_v32i16:478; AVX512VPOPCNTDQ-NOBW: # %bb.0:479; AVX512VPOPCNTDQ-NOBW-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1480; AVX512VPOPCNTDQ-NOBW-NEXT: vpaddw %ymm1, %ymm0, %ymm2481; AVX512VPOPCNTDQ-NOBW-NEXT: vextracti64x4 $1, %zmm0, %ymm3482; AVX512VPOPCNTDQ-NOBW-NEXT: vpaddw %ymm1, %ymm3, %ymm1483; AVX512VPOPCNTDQ-NOBW-NEXT: vpxor %ymm1, %ymm3, %ymm3484; AVX512VPOPCNTDQ-NOBW-NEXT: vpminuw %ymm1, %ymm3, %ymm1485; AVX512VPOPCNTDQ-NOBW-NEXT: vpcmpeqw %ymm1, %ymm3, %ymm1486; AVX512VPOPCNTDQ-NOBW-NEXT: vpxor %ymm2, %ymm0, %ymm0487; AVX512VPOPCNTDQ-NOBW-NEXT: vpminuw %ymm2, %ymm0, %ymm2488; AVX512VPOPCNTDQ-NOBW-NEXT: vpcmpeqw %ymm2, %ymm0, %ymm0489; AVX512VPOPCNTDQ-NOBW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0490; AVX512VPOPCNTDQ-NOBW-NEXT: vpternlogq {{.*#+}} zmm0 = ~zmm0491; AVX512VPOPCNTDQ-NOBW-NEXT: retq492;493; AVX512VPOPCNTDQ-BW-LABEL: eq_1_v32i16:494; AVX512VPOPCNTDQ-BW: # %bb.0:495; AVX512VPOPCNTDQ-BW-NEXT: vpternlogd {{.*#+}} zmm1 = -1496; AVX512VPOPCNTDQ-BW-NEXT: vpaddw %zmm1, %zmm0, %zmm1497; AVX512VPOPCNTDQ-BW-NEXT: vpxorq %zmm1, %zmm0, %zmm0498; AVX512VPOPCNTDQ-BW-NEXT: vpcmpnleuw %zmm1, %zmm0, %k0499; AVX512VPOPCNTDQ-BW-NEXT: vpmovm2w %k0, %zmm0500; AVX512VPOPCNTDQ-BW-NEXT: retq501;502; BITALG-LABEL: eq_1_v32i16:503; BITALG: # %bb.0:504; BITALG-NEXT: vpopcntw %zmm0, %zmm0505; BITALG-NEXT: vpcmpeqw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %k0506; BITALG-NEXT: vpmovm2w %k0, %zmm0507; BITALG-NEXT: retq508 %2 = tail call <32 x i16> @llvm.ctpop.v32i16(<32 x i16> %0)509 %3 = icmp eq <32 x i16> %2, <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, 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>510 %4 = sext <32 x i1> %3 to <32 x i16>511 ret <32 x i16> %4512}513 514define <32 x i16> @ne_1_v32i16(<32 x i16> %0) {515; AVX512F-LABEL: ne_1_v32i16:516; AVX512F: # %bb.0:517; AVX512F-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1518; AVX512F-NEXT: vpaddw %ymm1, %ymm0, %ymm2519; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm3520; AVX512F-NEXT: vpaddw %ymm1, %ymm3, %ymm1521; AVX512F-NEXT: vpxor %ymm1, %ymm3, %ymm3522; AVX512F-NEXT: vpminuw %ymm1, %ymm3, %ymm1523; AVX512F-NEXT: vpcmpeqw %ymm1, %ymm3, %ymm1524; AVX512F-NEXT: vpxor %ymm2, %ymm0, %ymm0525; AVX512F-NEXT: vpminuw %ymm2, %ymm0, %ymm2526; AVX512F-NEXT: vpcmpeqw %ymm2, %ymm0, %ymm0527; AVX512F-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0528; AVX512F-NEXT: retq529;530; AVX512BW-LABEL: ne_1_v32i16:531; AVX512BW: # %bb.0:532; AVX512BW-NEXT: vpternlogd {{.*#+}} zmm1 = -1533; AVX512BW-NEXT: vpaddw %zmm1, %zmm0, %zmm1534; AVX512BW-NEXT: vpxorq %zmm1, %zmm0, %zmm0535; AVX512BW-NEXT: vpcmpleuw %zmm1, %zmm0, %k0536; AVX512BW-NEXT: vpmovm2w %k0, %zmm0537; AVX512BW-NEXT: retq538;539; AVX512VPOPCNTDQ-NOBW-LABEL: ne_1_v32i16:540; AVX512VPOPCNTDQ-NOBW: # %bb.0:541; AVX512VPOPCNTDQ-NOBW-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1542; AVX512VPOPCNTDQ-NOBW-NEXT: vpaddw %ymm1, %ymm0, %ymm2543; AVX512VPOPCNTDQ-NOBW-NEXT: vextracti64x4 $1, %zmm0, %ymm3544; AVX512VPOPCNTDQ-NOBW-NEXT: vpaddw %ymm1, %ymm3, %ymm1545; AVX512VPOPCNTDQ-NOBW-NEXT: vpxor %ymm1, %ymm3, %ymm3546; AVX512VPOPCNTDQ-NOBW-NEXT: vpminuw %ymm1, %ymm3, %ymm1547; AVX512VPOPCNTDQ-NOBW-NEXT: vpcmpeqw %ymm1, %ymm3, %ymm1548; AVX512VPOPCNTDQ-NOBW-NEXT: vpxor %ymm2, %ymm0, %ymm0549; AVX512VPOPCNTDQ-NOBW-NEXT: vpminuw %ymm2, %ymm0, %ymm2550; AVX512VPOPCNTDQ-NOBW-NEXT: vpcmpeqw %ymm2, %ymm0, %ymm0551; AVX512VPOPCNTDQ-NOBW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0552; AVX512VPOPCNTDQ-NOBW-NEXT: retq553;554; AVX512VPOPCNTDQ-BW-LABEL: ne_1_v32i16:555; AVX512VPOPCNTDQ-BW: # %bb.0:556; AVX512VPOPCNTDQ-BW-NEXT: vpternlogd {{.*#+}} zmm1 = -1557; AVX512VPOPCNTDQ-BW-NEXT: vpaddw %zmm1, %zmm0, %zmm1558; AVX512VPOPCNTDQ-BW-NEXT: vpxorq %zmm1, %zmm0, %zmm0559; AVX512VPOPCNTDQ-BW-NEXT: vpcmpleuw %zmm1, %zmm0, %k0560; AVX512VPOPCNTDQ-BW-NEXT: vpmovm2w %k0, %zmm0561; AVX512VPOPCNTDQ-BW-NEXT: retq562;563; BITALG-LABEL: ne_1_v32i16:564; BITALG: # %bb.0:565; BITALG-NEXT: vpopcntw %zmm0, %zmm0566; BITALG-NEXT: vpcmpneqw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %k0567; BITALG-NEXT: vpmovm2w %k0, %zmm0568; BITALG-NEXT: retq569 %2 = tail call <32 x i16> @llvm.ctpop.v32i16(<32 x i16> %0)570 %3 = icmp ne <32 x i16> %2, <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, 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>571 %4 = sext <32 x i1> %3 to <32 x i16>572 ret <32 x i16> %4573}574 575define <64 x i8> @eq_1_v64i8(<64 x i8> %0) {576; AVX512F-LABEL: eq_1_v64i8:577; AVX512F: # %bb.0:578; AVX512F-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1579; AVX512F-NEXT: vpaddb %ymm1, %ymm0, %ymm2580; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm3581; AVX512F-NEXT: vpaddb %ymm1, %ymm3, %ymm1582; AVX512F-NEXT: vpxor %ymm1, %ymm3, %ymm3583; AVX512F-NEXT: vpminub %ymm1, %ymm3, %ymm1584; AVX512F-NEXT: vpcmpeqb %ymm1, %ymm3, %ymm1585; AVX512F-NEXT: vpxor %ymm2, %ymm0, %ymm0586; AVX512F-NEXT: vpminub %ymm2, %ymm0, %ymm2587; AVX512F-NEXT: vpcmpeqb %ymm2, %ymm0, %ymm0588; AVX512F-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0589; AVX512F-NEXT: vpternlogq {{.*#+}} zmm0 = ~zmm0590; AVX512F-NEXT: retq591;592; AVX512BW-LABEL: eq_1_v64i8:593; AVX512BW: # %bb.0:594; AVX512BW-NEXT: vpternlogd {{.*#+}} zmm1 = -1595; AVX512BW-NEXT: vpaddb %zmm1, %zmm0, %zmm1596; AVX512BW-NEXT: vpxorq %zmm1, %zmm0, %zmm0597; AVX512BW-NEXT: vpcmpnleub %zmm1, %zmm0, %k0598; AVX512BW-NEXT: vpmovm2b %k0, %zmm0599; AVX512BW-NEXT: retq600;601; AVX512VPOPCNTDQ-NOBW-LABEL: eq_1_v64i8:602; AVX512VPOPCNTDQ-NOBW: # %bb.0:603; AVX512VPOPCNTDQ-NOBW-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1604; AVX512VPOPCNTDQ-NOBW-NEXT: vpaddb %ymm1, %ymm0, %ymm2605; AVX512VPOPCNTDQ-NOBW-NEXT: vextracti64x4 $1, %zmm0, %ymm3606; AVX512VPOPCNTDQ-NOBW-NEXT: vpaddb %ymm1, %ymm3, %ymm1607; AVX512VPOPCNTDQ-NOBW-NEXT: vpxor %ymm1, %ymm3, %ymm3608; AVX512VPOPCNTDQ-NOBW-NEXT: vpminub %ymm1, %ymm3, %ymm1609; AVX512VPOPCNTDQ-NOBW-NEXT: vpcmpeqb %ymm1, %ymm3, %ymm1610; AVX512VPOPCNTDQ-NOBW-NEXT: vpxor %ymm2, %ymm0, %ymm0611; AVX512VPOPCNTDQ-NOBW-NEXT: vpminub %ymm2, %ymm0, %ymm2612; AVX512VPOPCNTDQ-NOBW-NEXT: vpcmpeqb %ymm2, %ymm0, %ymm0613; AVX512VPOPCNTDQ-NOBW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0614; AVX512VPOPCNTDQ-NOBW-NEXT: vpternlogq {{.*#+}} zmm0 = ~zmm0615; AVX512VPOPCNTDQ-NOBW-NEXT: retq616;617; AVX512VPOPCNTDQ-BW-LABEL: eq_1_v64i8:618; AVX512VPOPCNTDQ-BW: # %bb.0:619; AVX512VPOPCNTDQ-BW-NEXT: vpternlogd {{.*#+}} zmm1 = -1620; AVX512VPOPCNTDQ-BW-NEXT: vpaddb %zmm1, %zmm0, %zmm1621; AVX512VPOPCNTDQ-BW-NEXT: vpxorq %zmm1, %zmm0, %zmm0622; AVX512VPOPCNTDQ-BW-NEXT: vpcmpnleub %zmm1, %zmm0, %k0623; AVX512VPOPCNTDQ-BW-NEXT: vpmovm2b %k0, %zmm0624; AVX512VPOPCNTDQ-BW-NEXT: retq625;626; BITALG-LABEL: eq_1_v64i8:627; BITALG: # %bb.0:628; BITALG-NEXT: vpopcntb %zmm0, %zmm0629; BITALG-NEXT: vpcmpeqb {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %k0630; BITALG-NEXT: vpmovm2b %k0, %zmm0631; BITALG-NEXT: retq632 %2 = tail call <64 x i8> @llvm.ctpop.v64i8(<64 x i8> %0)633 %3 = icmp eq <64 x i8> %2, <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>634 %4 = sext <64 x i1> %3 to <64 x i8>635 ret <64 x i8> %4636}637 638define <64 x i8> @ne_1_v64i8(<64 x i8> %0) {639; AVX512F-LABEL: ne_1_v64i8:640; AVX512F: # %bb.0:641; AVX512F-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1642; AVX512F-NEXT: vpaddb %ymm1, %ymm0, %ymm2643; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm3644; AVX512F-NEXT: vpaddb %ymm1, %ymm3, %ymm1645; AVX512F-NEXT: vpxor %ymm1, %ymm3, %ymm3646; AVX512F-NEXT: vpminub %ymm1, %ymm3, %ymm1647; AVX512F-NEXT: vpcmpeqb %ymm1, %ymm3, %ymm1648; AVX512F-NEXT: vpxor %ymm2, %ymm0, %ymm0649; AVX512F-NEXT: vpminub %ymm2, %ymm0, %ymm2650; AVX512F-NEXT: vpcmpeqb %ymm2, %ymm0, %ymm0651; AVX512F-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0652; AVX512F-NEXT: retq653;654; AVX512BW-LABEL: ne_1_v64i8:655; AVX512BW: # %bb.0:656; AVX512BW-NEXT: vpternlogd {{.*#+}} zmm1 = -1657; AVX512BW-NEXT: vpaddb %zmm1, %zmm0, %zmm1658; AVX512BW-NEXT: vpxorq %zmm1, %zmm0, %zmm0659; AVX512BW-NEXT: vpcmpleub %zmm1, %zmm0, %k0660; AVX512BW-NEXT: vpmovm2b %k0, %zmm0661; AVX512BW-NEXT: retq662;663; AVX512VPOPCNTDQ-NOBW-LABEL: ne_1_v64i8:664; AVX512VPOPCNTDQ-NOBW: # %bb.0:665; AVX512VPOPCNTDQ-NOBW-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1666; AVX512VPOPCNTDQ-NOBW-NEXT: vpaddb %ymm1, %ymm0, %ymm2667; AVX512VPOPCNTDQ-NOBW-NEXT: vextracti64x4 $1, %zmm0, %ymm3668; AVX512VPOPCNTDQ-NOBW-NEXT: vpaddb %ymm1, %ymm3, %ymm1669; AVX512VPOPCNTDQ-NOBW-NEXT: vpxor %ymm1, %ymm3, %ymm3670; AVX512VPOPCNTDQ-NOBW-NEXT: vpminub %ymm1, %ymm3, %ymm1671; AVX512VPOPCNTDQ-NOBW-NEXT: vpcmpeqb %ymm1, %ymm3, %ymm1672; AVX512VPOPCNTDQ-NOBW-NEXT: vpxor %ymm2, %ymm0, %ymm0673; AVX512VPOPCNTDQ-NOBW-NEXT: vpminub %ymm2, %ymm0, %ymm2674; AVX512VPOPCNTDQ-NOBW-NEXT: vpcmpeqb %ymm2, %ymm0, %ymm0675; AVX512VPOPCNTDQ-NOBW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0676; AVX512VPOPCNTDQ-NOBW-NEXT: retq677;678; AVX512VPOPCNTDQ-BW-LABEL: ne_1_v64i8:679; AVX512VPOPCNTDQ-BW: # %bb.0:680; AVX512VPOPCNTDQ-BW-NEXT: vpternlogd {{.*#+}} zmm1 = -1681; AVX512VPOPCNTDQ-BW-NEXT: vpaddb %zmm1, %zmm0, %zmm1682; AVX512VPOPCNTDQ-BW-NEXT: vpxorq %zmm1, %zmm0, %zmm0683; AVX512VPOPCNTDQ-BW-NEXT: vpcmpleub %zmm1, %zmm0, %k0684; AVX512VPOPCNTDQ-BW-NEXT: vpmovm2b %k0, %zmm0685; AVX512VPOPCNTDQ-BW-NEXT: retq686;687; BITALG-LABEL: ne_1_v64i8:688; BITALG: # %bb.0:689; BITALG-NEXT: vpopcntb %zmm0, %zmm0690; BITALG-NEXT: vpcmpneqb {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %k0691; BITALG-NEXT: vpmovm2b %k0, %zmm0692; BITALG-NEXT: retq693 %2 = tail call <64 x i8> @llvm.ctpop.v64i8(<64 x i8> %0)694 %3 = icmp ne <64 x i8> %2, <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>695 %4 = sext <64 x i1> %3 to <64 x i8>696 ret <64 x i8> %4697}698 699declare <8 x i64> @llvm.ctpop.v8i64(<8 x i64>)700declare <16 x i32> @llvm.ctpop.v16i32(<16 x i32>)701declare <32 x i16> @llvm.ctpop.v32i16(<32 x i16>)702declare <64 x i8> @llvm.ctpop.v64i8(<64 x i8>)703