640 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=SSE3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX14; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX25; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl | FileCheck %s --check-prefixes=AVX,AVX512VL6; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,+avx512dq | FileCheck %s --check-prefixes=AVX,AVX512DQVL7 8define <2 x i64> @combine_shuffle_sext_pmuldq(<4 x i32> %a0, <4 x i32> %a1) {9; SSE-LABEL: combine_shuffle_sext_pmuldq:10; SSE: # %bb.0:11; SSE-NEXT: pmuldq %xmm1, %xmm012; SSE-NEXT: retq13;14; AVX-LABEL: combine_shuffle_sext_pmuldq:15; AVX: # %bb.0:16; AVX-NEXT: vpmuldq %xmm1, %xmm0, %xmm017; AVX-NEXT: retq18 %1 = shufflevector <4 x i32> %a0, <4 x i32> undef, <2 x i32> <i32 0, i32 2>19 %2 = shufflevector <4 x i32> %a1, <4 x i32> undef, <2 x i32> <i32 0, i32 2>20 %3 = sext <2 x i32> %1 to <2 x i64>21 %4 = sext <2 x i32> %2 to <2 x i64>22 %5 = mul nuw <2 x i64> %3, %423 ret <2 x i64> %524}25 26define <2 x i64> @combine_shuffle_zext_pmuludq(<4 x i32> %a0, <4 x i32> %a1) {27; SSE-LABEL: combine_shuffle_zext_pmuludq:28; SSE: # %bb.0:29; SSE-NEXT: pmuludq %xmm1, %xmm030; SSE-NEXT: retq31;32; AVX-LABEL: combine_shuffle_zext_pmuludq:33; AVX: # %bb.0:34; AVX-NEXT: vpmuludq %xmm1, %xmm0, %xmm035; AVX-NEXT: retq36 %1 = shufflevector <4 x i32> %a0, <4 x i32> undef, <2 x i32> <i32 0, i32 2>37 %2 = shufflevector <4 x i32> %a1, <4 x i32> undef, <2 x i32> <i32 0, i32 2>38 %3 = zext <2 x i32> %1 to <2 x i64>39 %4 = zext <2 x i32> %2 to <2 x i64>40 %5 = mul nuw <2 x i64> %3, %441 ret <2 x i64> %542}43 44define <2 x i64> @combine_shuffle_zero_pmuludq(<4 x i32> %a0, <4 x i32> %a1) {45; SSE-LABEL: combine_shuffle_zero_pmuludq:46; SSE: # %bb.0:47; SSE-NEXT: pmuludq %xmm1, %xmm048; SSE-NEXT: retq49;50; AVX-LABEL: combine_shuffle_zero_pmuludq:51; AVX: # %bb.0:52; AVX-NEXT: vpmuludq %xmm1, %xmm0, %xmm053; AVX-NEXT: retq54 %1 = shufflevector <4 x i32> %a0, <4 x i32> zeroinitializer, <4 x i32> <i32 0, i32 5, i32 2, i32 7>55 %2 = shufflevector <4 x i32> %a1, <4 x i32> zeroinitializer, <4 x i32> <i32 0, i32 5, i32 2, i32 7>56 %3 = bitcast <4 x i32> %1 to <2 x i64>57 %4 = bitcast <4 x i32> %2 to <2 x i64>58 %5 = mul <2 x i64> %3, %459 ret <2 x i64> %560}61 62define <4 x i64> @combine_shuffle_zero_pmuludq_256(<8 x i32> %a0, <8 x i32> %a1) {63; SSE-LABEL: combine_shuffle_zero_pmuludq_256:64; SSE: # %bb.0:65; SSE-NEXT: pmuludq %xmm2, %xmm066; SSE-NEXT: pmuludq %xmm3, %xmm167; SSE-NEXT: retq68;69; AVX1-LABEL: combine_shuffle_zero_pmuludq_256:70; AVX1: # %bb.0:71; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm272; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm373; AVX1-NEXT: vpmuludq %xmm2, %xmm3, %xmm274; AVX1-NEXT: vpmuludq %xmm1, %xmm0, %xmm075; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm076; AVX1-NEXT: retq77;78; AVX2-LABEL: combine_shuffle_zero_pmuludq_256:79; AVX2: # %bb.0:80; AVX2-NEXT: vpmuludq %ymm1, %ymm0, %ymm081; AVX2-NEXT: retq82;83; AVX512VL-LABEL: combine_shuffle_zero_pmuludq_256:84; AVX512VL: # %bb.0:85; AVX512VL-NEXT: vpmuludq %ymm1, %ymm0, %ymm086; AVX512VL-NEXT: retq87;88; AVX512DQVL-LABEL: combine_shuffle_zero_pmuludq_256:89; AVX512DQVL: # %bb.0:90; AVX512DQVL-NEXT: vpmuludq %ymm1, %ymm0, %ymm091; AVX512DQVL-NEXT: retq92 %1 = shufflevector <8 x i32> %a0, <8 x i32> zeroinitializer, <8 x i32> <i32 0, i32 9, i32 2, i32 11, i32 4, i32 13, i32 6, i32 15>93 %2 = shufflevector <8 x i32> %a1, <8 x i32> zeroinitializer, <8 x i32> <i32 0, i32 9, i32 2, i32 11, i32 4, i32 13, i32 6, i32 15>94 %3 = bitcast <8 x i32> %1 to <4 x i64>95 %4 = bitcast <8 x i32> %2 to <4 x i64>96 %5 = mul <4 x i64> %3, %497 ret <4 x i64> %598}99 100define <8 x i64> @combine_zext_pmuludq_256(<8 x i32> %a) {101; SSE-LABEL: combine_zext_pmuludq_256:102; SSE: # %bb.0:103; SSE-NEXT: pshufd {{.*#+}} xmm3 = xmm1[2,1,3,3]104; SSE-NEXT: pmovzxdq {{.*#+}} xmm2 = xmm1[0],zero,xmm1[1],zero105; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[2,1,3,3]106; SSE-NEXT: pmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero107; SSE-NEXT: pmovsxdq {{.*#+}} xmm4 = [715827883,715827883]108; SSE-NEXT: pmuludq %xmm4, %xmm0109; SSE-NEXT: pmuludq %xmm4, %xmm1110; SSE-NEXT: pmuludq %xmm4, %xmm2111; SSE-NEXT: pmuludq %xmm4, %xmm3112; SSE-NEXT: retq113;114; AVX1-LABEL: combine_zext_pmuludq_256:115; AVX1: # %bb.0:116; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1117; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm1[2,2,3,3]118; AVX1-NEXT: vpshufd {{.*#+}} xmm3 = xmm0[2,2,3,3]119; AVX1-NEXT: vmovddup {{.*#+}} xmm4 = [715827883,715827883]120; AVX1-NEXT: # xmm4 = mem[0,0]121; AVX1-NEXT: vpmuludq %xmm4, %xmm3, %xmm3122; AVX1-NEXT: vpmovzxdq {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero123; AVX1-NEXT: vpmuludq %xmm4, %xmm0, %xmm0124; AVX1-NEXT: vinsertf128 $1, %xmm3, %ymm0, %ymm0125; AVX1-NEXT: vpmuludq %xmm4, %xmm2, %xmm2126; AVX1-NEXT: vpmovzxdq {{.*#+}} xmm1 = xmm1[0],zero,xmm1[1],zero127; AVX1-NEXT: vpmuludq %xmm4, %xmm1, %xmm1128; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm1129; AVX1-NEXT: retq130;131; AVX2-LABEL: combine_zext_pmuludq_256:132; AVX2: # %bb.0:133; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1134; AVX2-NEXT: vpmovzxdq {{.*#+}} ymm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero135; AVX2-NEXT: vpmovzxdq {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero136; AVX2-NEXT: vpbroadcastq {{.*#+}} ymm2 = [715827883,715827883,715827883,715827883]137; AVX2-NEXT: vpmuludq %ymm2, %ymm0, %ymm0138; AVX2-NEXT: vpmuludq %ymm2, %ymm1, %ymm1139; AVX2-NEXT: retq140;141; AVX512VL-LABEL: combine_zext_pmuludq_256:142; AVX512VL: # %bb.0:143; AVX512VL-NEXT: vpmovzxdq {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero144; AVX512VL-NEXT: vpmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to8}, %zmm0, %zmm0145; AVX512VL-NEXT: retq146;147; AVX512DQVL-LABEL: combine_zext_pmuludq_256:148; AVX512DQVL: # %bb.0:149; AVX512DQVL-NEXT: vpmovzxdq {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero150; AVX512DQVL-NEXT: vpmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to8}, %zmm0, %zmm0151; AVX512DQVL-NEXT: retq152 %1 = zext <8 x i32> %a to <8 x i64>153 %2 = mul nuw nsw <8 x i64> %1, <i64 715827883, i64 715827883, i64 715827883, i64 715827883, i64 715827883, i64 715827883, i64 715827883, i64 715827883>154 ret <8 x i64> %2155}156 157define void @PR39398(i32 %a0) {158; SSE-LABEL: PR39398:159; SSE: # %bb.0: # %bb160; SSE-NEXT: .p2align 4161; SSE-NEXT: .LBB5_1: # %bb10162; SSE-NEXT: # =>This Inner Loop Header: Depth=1163; SSE-NEXT: cmpl $232, %edi164; SSE-NEXT: jne .LBB5_1165; SSE-NEXT: # %bb.2: # %bb34166; SSE-NEXT: retq167;168; AVX-LABEL: PR39398:169; AVX: # %bb.0: # %bb170; AVX-NEXT: .p2align 4171; AVX-NEXT: .LBB5_1: # %bb10172; AVX-NEXT: # =>This Inner Loop Header: Depth=1173; AVX-NEXT: cmpl $232, %edi174; AVX-NEXT: jne .LBB5_1175; AVX-NEXT: # %bb.2: # %bb34176; AVX-NEXT: retq177bb:178 %tmp9 = shufflevector <4 x i64> undef, <4 x i64> undef, <4 x i32> zeroinitializer179 br label %bb10180 181bb10: ; preds = %bb10, %bb182 %tmp12 = phi <4 x i32> [ <i32 9, i32 8, i32 7, i32 6>, %bb ], [ zeroinitializer, %bb10 ]183 %tmp16 = add <4 x i32> %tmp12, <i32 -4, i32 -4, i32 -4, i32 -4>184 %tmp18 = zext <4 x i32> %tmp12 to <4 x i64>185 %tmp19 = zext <4 x i32> %tmp16 to <4 x i64>186 %tmp20 = xor <4 x i64> %tmp18, <i64 -1, i64 -1, i64 -1, i64 -1>187 %tmp21 = xor <4 x i64> %tmp19, <i64 -1, i64 -1, i64 -1, i64 -1>188 %tmp24 = mul <4 x i64> %tmp9, %tmp20189 %tmp25 = mul <4 x i64> %tmp9, %tmp21190 %tmp26 = select <4 x i1> undef, <4 x i64> zeroinitializer, <4 x i64> %tmp24191 %tmp27 = select <4 x i1> undef, <4 x i64> zeroinitializer, <4 x i64> %tmp25192 %tmp28 = add <4 x i64> zeroinitializer, %tmp26193 %tmp29 = add <4 x i64> zeroinitializer, %tmp27194 %tmp33 = icmp eq i32 %a0, 232195 br i1 %tmp33, label %bb34, label %bb10196 197bb34: ; preds = %bb10198 %tmp35 = add <4 x i64> %tmp29, %tmp28199 ret void200}201 202define i32 @PR43159(ptr %a0) {203; SSE-LABEL: PR43159:204; SSE: # %bb.0: # %entry205; SSE-NEXT: movdqa (%rdi), %xmm0206; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]207; SSE-NEXT: pmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1 # [344322273,344322273,1916962805,1916962805]208; SSE-NEXT: movdqa %xmm0, %xmm2209; SSE-NEXT: psrld $1, %xmm2210; SSE-NEXT: pblendw {{.*#+}} xmm2 = xmm0[0,1,2,3],xmm2[4,5],xmm0[6,7]211; SSE-NEXT: psubd %xmm1, %xmm0212; SSE-NEXT: pshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]213; SSE-NEXT: pmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 # [2147483648,2147483648,2147483648,2147483648]214; SSE-NEXT: paddd %xmm1, %xmm0215; SSE-NEXT: psrld $7, %xmm0216; SSE-NEXT: pmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2 # [1645975491,344322273,2164392969,1916962805]217; SSE-NEXT: pshufd {{.*#+}} xmm1 = xmm2[1,1,3,3]218; SSE-NEXT: psrld $6, %xmm1219; SSE-NEXT: movd %xmm1, %edi220; SSE-NEXT: pextrd $1, %xmm0, %esi221; SSE-NEXT: pextrd $2, %xmm1, %edx222; SSE-NEXT: pextrd $3, %xmm0, %ecx223; SSE-NEXT: jmp foo # TAILCALL224;225; AVX1-LABEL: PR43159:226; AVX1: # %bb.0: # %entry227; AVX1-NEXT: vmovdqa (%rdi), %xmm0228; AVX1-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]229; AVX1-NEXT: vpmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1 # [344322273,344322273,1916962805,1916962805]230; AVX1-NEXT: vpsubd %xmm1, %xmm0, %xmm2231; AVX1-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]232; AVX1-NEXT: vpmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2, %xmm2 # [2147483648,2147483648,2147483648,2147483648]233; AVX1-NEXT: vpaddd %xmm1, %xmm2, %xmm1234; AVX1-NEXT: vpsrld $7, %xmm1, %xmm1235; AVX1-NEXT: vpsrld $1, %xmm0, %xmm2236; AVX1-NEXT: vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm2[4,5],xmm0[6,7]237; AVX1-NEXT: vpmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0 # [1645975491,344322273,2164392969,1916962805]238; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]239; AVX1-NEXT: vpsrld $6, %xmm0, %xmm0240; AVX1-NEXT: vmovd %xmm0, %edi241; AVX1-NEXT: vpextrd $1, %xmm1, %esi242; AVX1-NEXT: vpextrd $2, %xmm0, %edx243; AVX1-NEXT: vpextrd $3, %xmm1, %ecx244; AVX1-NEXT: jmp foo # TAILCALL245;246; AVX2-LABEL: PR43159:247; AVX2: # %bb.0: # %entry248; AVX2-NEXT: vmovdqa (%rdi), %xmm0249; AVX2-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]250; AVX2-NEXT: vpmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1 # [344322273,u,1916962805,u]251; AVX2-NEXT: vpsrlvd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm2252; AVX2-NEXT: vpmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2, %xmm2 # [1645975491,344322273,2164392969,1916962805]253; AVX2-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]254; AVX2-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0],xmm1[1],xmm2[2],xmm1[3]255; AVX2-NEXT: vpsubd %xmm1, %xmm0, %xmm0256; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]257; AVX2-NEXT: vpbroadcastd {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648]258; AVX2-NEXT: vpmuludq %xmm1, %xmm0, %xmm0259; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1260; AVX2-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2],xmm0[3]261; AVX2-NEXT: vpaddd %xmm2, %xmm0, %xmm0262; AVX2-NEXT: vpsrlvd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0263; AVX2-NEXT: vmovd %xmm0, %edi264; AVX2-NEXT: vpextrd $1, %xmm0, %esi265; AVX2-NEXT: vpextrd $2, %xmm0, %edx266; AVX2-NEXT: vpextrd $3, %xmm0, %ecx267; AVX2-NEXT: jmp foo # TAILCALL268;269; AVX512VL-LABEL: PR43159:270; AVX512VL: # %bb.0: # %entry271; AVX512VL-NEXT: vmovdqa (%rdi), %xmm0272; AVX512VL-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]273; AVX512VL-NEXT: vpmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1 # [344322273,u,1916962805,u]274; AVX512VL-NEXT: vpsrlvd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm2275; AVX512VL-NEXT: vpmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2, %xmm2 # [1645975491,344322273,2164392969,1916962805]276; AVX512VL-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]277; AVX512VL-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0],xmm1[1],xmm2[2],xmm1[3]278; AVX512VL-NEXT: vpsubd %xmm1, %xmm0, %xmm0279; AVX512VL-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]280; AVX512VL-NEXT: vpbroadcastd {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648]281; AVX512VL-NEXT: vpmuludq %xmm1, %xmm0, %xmm0282; AVX512VL-NEXT: vpxor %xmm1, %xmm1, %xmm1283; AVX512VL-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2],xmm0[3]284; AVX512VL-NEXT: vpaddd %xmm2, %xmm0, %xmm0285; AVX512VL-NEXT: vpsrlvd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0286; AVX512VL-NEXT: vmovd %xmm0, %edi287; AVX512VL-NEXT: vpextrd $1, %xmm0, %esi288; AVX512VL-NEXT: vpextrd $2, %xmm0, %edx289; AVX512VL-NEXT: vpextrd $3, %xmm0, %ecx290; AVX512VL-NEXT: jmp foo # TAILCALL291;292; AVX512DQVL-LABEL: PR43159:293; AVX512DQVL: # %bb.0: # %entry294; AVX512DQVL-NEXT: vmovdqa (%rdi), %xmm0295; AVX512DQVL-NEXT: vpshufd {{.*#+}} xmm1 = xmm0[1,1,3,3]296; AVX512DQVL-NEXT: vpmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1, %xmm1 # [344322273,u,1916962805,u]297; AVX512DQVL-NEXT: vpsrlvd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm2298; AVX512DQVL-NEXT: vpmuludq {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2, %xmm2 # [1645975491,344322273,2164392969,1916962805]299; AVX512DQVL-NEXT: vpshufd {{.*#+}} xmm2 = xmm2[1,1,3,3]300; AVX512DQVL-NEXT: vpblendd {{.*#+}} xmm2 = xmm2[0],xmm1[1],xmm2[2],xmm1[3]301; AVX512DQVL-NEXT: vpsubd %xmm1, %xmm0, %xmm0302; AVX512DQVL-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[1,1,3,3]303; AVX512DQVL-NEXT: vpbroadcastd {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648]304; AVX512DQVL-NEXT: vpmuludq %xmm1, %xmm0, %xmm0305; AVX512DQVL-NEXT: vpxor %xmm1, %xmm1, %xmm1306; AVX512DQVL-NEXT: vpblendd {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2],xmm0[3]307; AVX512DQVL-NEXT: vpaddd %xmm2, %xmm0, %xmm0308; AVX512DQVL-NEXT: vpsrlvd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0309; AVX512DQVL-NEXT: vmovd %xmm0, %edi310; AVX512DQVL-NEXT: vpextrd $1, %xmm0, %esi311; AVX512DQVL-NEXT: vpextrd $2, %xmm0, %edx312; AVX512DQVL-NEXT: vpextrd $3, %xmm0, %ecx313; AVX512DQVL-NEXT: jmp foo # TAILCALL314entry:315 %0 = load <4 x i32>, ptr %a0, align 16316 %div = udiv <4 x i32> %0, <i32 167, i32 237, i32 254, i32 177>317 %ext0 = extractelement <4 x i32> %div, i32 0318 %ext1 = extractelement <4 x i32> %div, i32 1319 %ext2 = extractelement <4 x i32> %div, i32 2320 %ext3 = extractelement <4 x i32> %div, i32 3321 %call = tail call i32 @foo(i32 %ext0, i32 %ext1, i32 %ext2, i32 %ext3)322 ret i32 %call323}324declare dso_local i32 @foo(i32, i32, i32, i32)325 326define <8 x i32> @PR49658_zext(ptr %ptr, i32 %mul) {327; SSE-LABEL: PR49658_zext:328; SSE: # %bb.0: # %start329; SSE-NEXT: movd %esi, %xmm0330; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm0[0,1,0,1]331; SSE-NEXT: pxor %xmm0, %xmm0332; SSE-NEXT: movq $-2097152, %rax # imm = 0xFFE00000333; SSE-NEXT: pxor %xmm1, %xmm1334; SSE-NEXT: .p2align 4335; SSE-NEXT: .LBB7_1: # %loop336; SSE-NEXT: # =>This Inner Loop Header: Depth=1337; SSE-NEXT: pmovzxdq {{.*#+}} xmm3 = mem[0],zero,mem[1],zero338; SSE-NEXT: pmovzxdq {{.*#+}} xmm4 = mem[0],zero,mem[1],zero339; SSE-NEXT: pmovzxdq {{.*#+}} xmm5 = mem[0],zero,mem[1],zero340; SSE-NEXT: pmovzxdq {{.*#+}} xmm6 = mem[0],zero,mem[1],zero341; SSE-NEXT: pmuludq %xmm2, %xmm6342; SSE-NEXT: pmuludq %xmm2, %xmm5343; SSE-NEXT: shufps {{.*#+}} xmm5 = xmm5[1,3],xmm6[1,3]344; SSE-NEXT: paddd %xmm5, %xmm0345; SSE-NEXT: pmuludq %xmm2, %xmm4346; SSE-NEXT: pmuludq %xmm2, %xmm3347; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,3],xmm3[1,3]348; SSE-NEXT: paddd %xmm4, %xmm1349; SSE-NEXT: subq $-128, %rax350; SSE-NEXT: jne .LBB7_1351; SSE-NEXT: # %bb.2: # %end352; SSE-NEXT: retq353;354; AVX1-LABEL: PR49658_zext:355; AVX1: # %bb.0: # %start356; AVX1-NEXT: movl %esi, %eax357; AVX1-NEXT: vmovq %rax, %xmm0358; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]359; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm1360; AVX1-NEXT: vpxor %xmm0, %xmm0, %xmm0361; AVX1-NEXT: movq $-2097152, %rax # imm = 0xFFE00000362; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2363; AVX1-NEXT: .p2align 4364; AVX1-NEXT: .LBB7_1: # %loop365; AVX1-NEXT: # =>This Inner Loop Header: Depth=1366; AVX1-NEXT: vpmovzxdq {{.*#+}} xmm3 = mem[0],zero,mem[1],zero367; AVX1-NEXT: vpmovzxdq {{.*#+}} xmm4 = mem[0],zero,mem[1],zero368; AVX1-NEXT: vpmovzxdq {{.*#+}} xmm5 = mem[0],zero,mem[1],zero369; AVX1-NEXT: vpmovzxdq {{.*#+}} xmm6 = mem[0],zero,mem[1],zero370; AVX1-NEXT: vpmuludq %xmm6, %xmm2, %xmm6371; AVX1-NEXT: vpmuludq %xmm5, %xmm1, %xmm5372; AVX1-NEXT: vshufps {{.*#+}} xmm5 = xmm5[1,3],xmm6[1,3]373; AVX1-NEXT: vpmuludq %xmm4, %xmm2, %xmm4374; AVX1-NEXT: vpmuludq %xmm3, %xmm1, %xmm3375; AVX1-NEXT: vshufps {{.*#+}} xmm3 = xmm3[1,3],xmm4[1,3]376; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm4377; AVX1-NEXT: vpaddd %xmm4, %xmm5, %xmm4378; AVX1-NEXT: vpaddd %xmm0, %xmm3, %xmm0379; AVX1-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm0380; AVX1-NEXT: subq $-128, %rax381; AVX1-NEXT: jne .LBB7_1382; AVX1-NEXT: # %bb.2: # %end383; AVX1-NEXT: retq384;385; AVX2-LABEL: PR49658_zext:386; AVX2: # %bb.0: # %start387; AVX2-NEXT: movl %esi, %eax388; AVX2-NEXT: vmovq %rax, %xmm0389; AVX2-NEXT: vpbroadcastq %xmm0, %ymm1390; AVX2-NEXT: vpxor %xmm0, %xmm0, %xmm0391; AVX2-NEXT: movq $-2097152, %rax # imm = 0xFFE00000392; AVX2-NEXT: .p2align 4393; AVX2-NEXT: .LBB7_1: # %loop394; AVX2-NEXT: # =>This Inner Loop Header: Depth=1395; AVX2-NEXT: vpmovzxdq {{.*#+}} ymm2 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero396; AVX2-NEXT: vpmovzxdq {{.*#+}} ymm3 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero397; AVX2-NEXT: vpmuludq %ymm3, %ymm1, %ymm3398; AVX2-NEXT: vpmuludq %ymm2, %ymm1, %ymm2399; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[1,3],ymm3[1,3],ymm2[5,7],ymm3[5,7]400; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]401; AVX2-NEXT: vpaddd %ymm0, %ymm2, %ymm0402; AVX2-NEXT: subq $-128, %rax403; AVX2-NEXT: jne .LBB7_1404; AVX2-NEXT: # %bb.2: # %end405; AVX2-NEXT: retq406;407; AVX512VL-LABEL: PR49658_zext:408; AVX512VL: # %bb.0: # %start409; AVX512VL-NEXT: movl %esi, %eax410; AVX512VL-NEXT: vpbroadcastq %rax, %zmm1411; AVX512VL-NEXT: vpxor %xmm0, %xmm0, %xmm0412; AVX512VL-NEXT: movq $-2097152, %rax # imm = 0xFFE00000413; AVX512VL-NEXT: .p2align 4414; AVX512VL-NEXT: .LBB7_1: # %loop415; AVX512VL-NEXT: # =>This Inner Loop Header: Depth=1416; AVX512VL-NEXT: vpmovzxdq {{.*#+}} zmm2 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero417; AVX512VL-NEXT: vpmuludq %zmm2, %zmm1, %zmm2418; AVX512VL-NEXT: vpsrlq $32, %zmm2, %zmm2419; AVX512VL-NEXT: vpmovqd %zmm2, %ymm2420; AVX512VL-NEXT: vpaddd %ymm0, %ymm2, %ymm0421; AVX512VL-NEXT: subq $-128, %rax422; AVX512VL-NEXT: jne .LBB7_1423; AVX512VL-NEXT: # %bb.2: # %end424; AVX512VL-NEXT: retq425;426; AVX512DQVL-LABEL: PR49658_zext:427; AVX512DQVL: # %bb.0: # %start428; AVX512DQVL-NEXT: movl %esi, %eax429; AVX512DQVL-NEXT: vpbroadcastq %rax, %zmm1430; AVX512DQVL-NEXT: vpxor %xmm0, %xmm0, %xmm0431; AVX512DQVL-NEXT: movq $-2097152, %rax # imm = 0xFFE00000432; AVX512DQVL-NEXT: .p2align 4433; AVX512DQVL-NEXT: .LBB7_1: # %loop434; AVX512DQVL-NEXT: # =>This Inner Loop Header: Depth=1435; AVX512DQVL-NEXT: vpmovzxdq {{.*#+}} zmm2 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero436; AVX512DQVL-NEXT: vpmuludq %zmm2, %zmm1, %zmm2437; AVX512DQVL-NEXT: vpsrlq $32, %zmm2, %zmm2438; AVX512DQVL-NEXT: vpmovqd %zmm2, %ymm2439; AVX512DQVL-NEXT: vpaddd %ymm0, %ymm2, %ymm0440; AVX512DQVL-NEXT: subq $-128, %rax441; AVX512DQVL-NEXT: jne .LBB7_1442; AVX512DQVL-NEXT: # %bb.2: # %end443; AVX512DQVL-NEXT: retq444start:445 %t1 = zext i32 %mul to i64446 %t2 = insertelement <8 x i64> undef, i64 %t1, i32 0447 %mulvec = shufflevector <8 x i64> %t2, <8 x i64> undef, <8 x i32> zeroinitializer448 br label %loop449loop:450 %loopcnt = phi i64 [ 0, %start ], [ %nextcnt, %loop ]451 %sum = phi <8 x i32> [ zeroinitializer, %start ], [ %nextsum, %loop ]452 %ptroff = getelementptr inbounds i32, ptr %ptr, i64 %loopcnt453 %v = load <8 x i32>, ptr %ptroff, align 4454 %v64 = zext <8 x i32> %v to <8 x i64>455 %vmul = mul nuw <8 x i64> %mulvec, %v64456 %vmulhi = lshr <8 x i64> %vmul, <i64 32, i64 32, i64 32, i64 32, i64 32, i64 32, i64 32, i64 32>457 %vtrunc = trunc <8 x i64> %vmulhi to <8 x i32>458 %nextsum = add <8 x i32> %vtrunc, %sum459 %nextcnt = add i64 %loopcnt, 32460 %isdone = icmp eq i64 %nextcnt, 524288461 br i1 %isdone, label %end, label %loop462end:463 ret <8 x i32> %nextsum464}465 466define <8 x i32> @PR49658_sext(ptr %ptr, i32 %mul) {467; SSE-LABEL: PR49658_sext:468; SSE: # %bb.0: # %start469; SSE-NEXT: movslq %esi, %rax470; SSE-NEXT: movq %rax, %xmm0471; SSE-NEXT: pshufd {{.*#+}} xmm2 = xmm0[0,1,0,1]472; SSE-NEXT: pxor %xmm0, %xmm0473; SSE-NEXT: movq $-2097152, %rax # imm = 0xFFE00000474; SSE-NEXT: movdqa %xmm2, %xmm3475; SSE-NEXT: psrlq $32, %xmm3476; SSE-NEXT: pxor %xmm1, %xmm1477; SSE-NEXT: .p2align 4478; SSE-NEXT: .LBB8_1: # %loop479; SSE-NEXT: # =>This Inner Loop Header: Depth=1480; SSE-NEXT: pmovsxdq 2097176(%rdi,%rax), %xmm5481; SSE-NEXT: pmovsxdq 2097168(%rdi,%rax), %xmm4482; SSE-NEXT: pmovsxdq 2097152(%rdi,%rax), %xmm6483; SSE-NEXT: pmovsxdq 2097160(%rdi,%rax), %xmm7484; SSE-NEXT: movdqa %xmm3, %xmm8485; SSE-NEXT: pmuludq %xmm7, %xmm8486; SSE-NEXT: movdqa %xmm2, %xmm9487; SSE-NEXT: pmuludq %xmm7, %xmm9488; SSE-NEXT: psrlq $32, %xmm7489; SSE-NEXT: pmuludq %xmm2, %xmm7490; SSE-NEXT: paddq %xmm8, %xmm7491; SSE-NEXT: psllq $32, %xmm7492; SSE-NEXT: paddq %xmm9, %xmm7493; SSE-NEXT: movdqa %xmm3, %xmm8494; SSE-NEXT: pmuludq %xmm6, %xmm8495; SSE-NEXT: movdqa %xmm2, %xmm9496; SSE-NEXT: pmuludq %xmm6, %xmm9497; SSE-NEXT: psrlq $32, %xmm6498; SSE-NEXT: pmuludq %xmm2, %xmm6499; SSE-NEXT: paddq %xmm8, %xmm6500; SSE-NEXT: psllq $32, %xmm6501; SSE-NEXT: paddq %xmm9, %xmm6502; SSE-NEXT: shufps {{.*#+}} xmm6 = xmm6[1,3],xmm7[1,3]503; SSE-NEXT: paddd %xmm6, %xmm0504; SSE-NEXT: movdqa %xmm4, %xmm6505; SSE-NEXT: psrlq $32, %xmm6506; SSE-NEXT: pmuludq %xmm2, %xmm6507; SSE-NEXT: movdqa %xmm3, %xmm7508; SSE-NEXT: pmuludq %xmm4, %xmm7509; SSE-NEXT: paddq %xmm6, %xmm7510; SSE-NEXT: psllq $32, %xmm7511; SSE-NEXT: pmuludq %xmm2, %xmm4512; SSE-NEXT: paddq %xmm7, %xmm4513; SSE-NEXT: movdqa %xmm5, %xmm6514; SSE-NEXT: psrlq $32, %xmm6515; SSE-NEXT: pmuludq %xmm2, %xmm6516; SSE-NEXT: movdqa %xmm3, %xmm7517; SSE-NEXT: pmuludq %xmm5, %xmm7518; SSE-NEXT: paddq %xmm6, %xmm7519; SSE-NEXT: psllq $32, %xmm7520; SSE-NEXT: pmuludq %xmm2, %xmm5521; SSE-NEXT: paddq %xmm7, %xmm5522; SSE-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,3],xmm5[1,3]523; SSE-NEXT: paddd %xmm4, %xmm1524; SSE-NEXT: subq $-128, %rax525; SSE-NEXT: jne .LBB8_1526; SSE-NEXT: # %bb.2: # %end527; SSE-NEXT: retq528;529; AVX1-LABEL: PR49658_sext:530; AVX1: # %bb.0: # %start531; AVX1-NEXT: movslq %esi, %rax532; AVX1-NEXT: vmovq %rax, %xmm0533; AVX1-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]534; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm0, %ymm1535; AVX1-NEXT: vpxor %xmm0, %xmm0, %xmm0536; AVX1-NEXT: movq $-2097152, %rax # imm = 0xFFE00000537; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2538; AVX1-NEXT: .p2align 4539; AVX1-NEXT: .LBB8_1: # %loop540; AVX1-NEXT: # =>This Inner Loop Header: Depth=1541; AVX1-NEXT: vpmovsxdq 2097152(%rdi,%rax), %xmm3542; AVX1-NEXT: vpmovsxdq 2097160(%rdi,%rax), %xmm4543; AVX1-NEXT: vpmovsxdq 2097168(%rdi,%rax), %xmm5544; AVX1-NEXT: vpmovsxdq 2097176(%rdi,%rax), %xmm6545; AVX1-NEXT: vpmuldq %xmm6, %xmm2, %xmm6546; AVX1-NEXT: vpmuldq %xmm5, %xmm1, %xmm5547; AVX1-NEXT: vshufps {{.*#+}} xmm5 = xmm5[1,3],xmm6[1,3]548; AVX1-NEXT: vpmuldq %xmm4, %xmm2, %xmm4549; AVX1-NEXT: vpmuldq %xmm3, %xmm1, %xmm3550; AVX1-NEXT: vshufps {{.*#+}} xmm3 = xmm3[1,3],xmm4[1,3]551; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm4552; AVX1-NEXT: vpaddd %xmm4, %xmm5, %xmm4553; AVX1-NEXT: vpaddd %xmm0, %xmm3, %xmm0554; AVX1-NEXT: vinsertf128 $1, %xmm4, %ymm0, %ymm0555; AVX1-NEXT: subq $-128, %rax556; AVX1-NEXT: jne .LBB8_1557; AVX1-NEXT: # %bb.2: # %end558; AVX1-NEXT: retq559;560; AVX2-LABEL: PR49658_sext:561; AVX2: # %bb.0: # %start562; AVX2-NEXT: movslq %esi, %rax563; AVX2-NEXT: vmovq %rax, %xmm0564; AVX2-NEXT: vpbroadcastq %xmm0, %ymm1565; AVX2-NEXT: vpxor %xmm0, %xmm0, %xmm0566; AVX2-NEXT: movq $-2097152, %rax # imm = 0xFFE00000567; AVX2-NEXT: .p2align 4568; AVX2-NEXT: .LBB8_1: # %loop569; AVX2-NEXT: # =>This Inner Loop Header: Depth=1570; AVX2-NEXT: vpmovsxdq 2097152(%rdi,%rax), %ymm2571; AVX2-NEXT: vpmovsxdq 2097168(%rdi,%rax), %ymm3572; AVX2-NEXT: vpmuldq %ymm3, %ymm1, %ymm3573; AVX2-NEXT: vpmuldq %ymm2, %ymm1, %ymm2574; AVX2-NEXT: vshufps {{.*#+}} ymm2 = ymm2[1,3],ymm3[1,3],ymm2[5,7],ymm3[5,7]575; AVX2-NEXT: vpermpd {{.*#+}} ymm2 = ymm2[0,2,1,3]576; AVX2-NEXT: vpaddd %ymm0, %ymm2, %ymm0577; AVX2-NEXT: subq $-128, %rax578; AVX2-NEXT: jne .LBB8_1579; AVX2-NEXT: # %bb.2: # %end580; AVX2-NEXT: retq581;582; AVX512VL-LABEL: PR49658_sext:583; AVX512VL: # %bb.0: # %start584; AVX512VL-NEXT: movslq %esi, %rax585; AVX512VL-NEXT: vpbroadcastq %rax, %zmm1586; AVX512VL-NEXT: vpxor %xmm0, %xmm0, %xmm0587; AVX512VL-NEXT: movq $-2097152, %rax # imm = 0xFFE00000588; AVX512VL-NEXT: .p2align 4589; AVX512VL-NEXT: .LBB8_1: # %loop590; AVX512VL-NEXT: # =>This Inner Loop Header: Depth=1591; AVX512VL-NEXT: vpmovzxdq {{.*#+}} zmm2 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero592; AVX512VL-NEXT: vpmuldq %zmm2, %zmm1, %zmm2593; AVX512VL-NEXT: vpsrlq $32, %zmm2, %zmm2594; AVX512VL-NEXT: vpmovqd %zmm2, %ymm2595; AVX512VL-NEXT: vpaddd %ymm0, %ymm2, %ymm0596; AVX512VL-NEXT: subq $-128, %rax597; AVX512VL-NEXT: jne .LBB8_1598; AVX512VL-NEXT: # %bb.2: # %end599; AVX512VL-NEXT: retq600;601; AVX512DQVL-LABEL: PR49658_sext:602; AVX512DQVL: # %bb.0: # %start603; AVX512DQVL-NEXT: movslq %esi, %rax604; AVX512DQVL-NEXT: vpbroadcastq %rax, %zmm1605; AVX512DQVL-NEXT: vpxor %xmm0, %xmm0, %xmm0606; AVX512DQVL-NEXT: movq $-2097152, %rax # imm = 0xFFE00000607; AVX512DQVL-NEXT: .p2align 4608; AVX512DQVL-NEXT: .LBB8_1: # %loop609; AVX512DQVL-NEXT: # =>This Inner Loop Header: Depth=1610; AVX512DQVL-NEXT: vpmovzxdq {{.*#+}} zmm2 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero611; AVX512DQVL-NEXT: vpmuldq %zmm2, %zmm1, %zmm2612; AVX512DQVL-NEXT: vpsrlq $32, %zmm2, %zmm2613; AVX512DQVL-NEXT: vpmovqd %zmm2, %ymm2614; AVX512DQVL-NEXT: vpaddd %ymm0, %ymm2, %ymm0615; AVX512DQVL-NEXT: subq $-128, %rax616; AVX512DQVL-NEXT: jne .LBB8_1617; AVX512DQVL-NEXT: # %bb.2: # %end618; AVX512DQVL-NEXT: retq619start:620 %t1 = sext i32 %mul to i64621 %t2 = insertelement <8 x i64> undef, i64 %t1, i32 0622 %mulvec = shufflevector <8 x i64> %t2, <8 x i64> undef, <8 x i32> zeroinitializer623 br label %loop624loop:625 %loopcnt = phi i64 [ 0, %start ], [ %nextcnt, %loop ]626 %sum = phi <8 x i32> [ zeroinitializer, %start ], [ %nextsum, %loop ]627 %ptroff = getelementptr inbounds i32, ptr %ptr, i64 %loopcnt628 %v = load <8 x i32>, ptr %ptroff, align 4629 %v64 = sext <8 x i32> %v to <8 x i64>630 %vmul = mul <8 x i64> %mulvec, %v64631 %vmulhi = ashr <8 x i64> %vmul, <i64 32, i64 32, i64 32, i64 32, i64 32, i64 32, i64 32, i64 32>632 %vtrunc = trunc <8 x i64> %vmulhi to <8 x i32>633 %nextsum = add <8 x i32> %vtrunc, %sum634 %nextcnt = add i64 %loopcnt, 32635 %isdone = icmp eq i64 %nextcnt, 524288636 br i1 %isdone, label %end, label %loop637end:638 ret <8 x i32> %nextsum639}640