195 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512cd,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX512CD3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512cd,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX512CD4; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512cd,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX512CD5; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512cd,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX512CD6; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512cd,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX512CDBW7; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512cd,+avx512bw,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX512CDBW8; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512cd,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX512CDBW9; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512cd,+avx512bw,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX512CDBW10; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,avx512cd,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX512VLCDBW11; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,avx512cd,+avx512bw,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX512VLCDBW12; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512vl,avx512cd,+avx512bw,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX512VLCDBW13; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512vl,avx512cd,+avx512bw,+fast-variable-perlane-shuffle | FileCheck %s --check-prefixes=ALL,AVX512VLCDBW14 15define <2 x i64> @test_mm_epi64(<8 x i16> %a, <8 x i16> %b) {16; AVX512CD-LABEL: test_mm_epi64:17; AVX512CD: # %bb.0: # %entry18; AVX512CD-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm019; AVX512CD-NEXT: vpmovsxwq %xmm0, %zmm020; AVX512CD-NEXT: vptestmq %zmm0, %zmm0, %k021; AVX512CD-NEXT: vpbroadcastmb2q %k0, %zmm022; AVX512CD-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm023; AVX512CD-NEXT: vzeroupper24; AVX512CD-NEXT: ret{{[l|q]}}25;26; AVX512CDBW-LABEL: test_mm_epi64:27; AVX512CDBW: # %bb.0: # %entry28; AVX512CDBW-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm129; AVX512CDBW-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm030; AVX512CDBW-NEXT: vpcmpeqw %zmm1, %zmm0, %k031; AVX512CDBW-NEXT: vpbroadcastmb2q %k0, %zmm032; AVX512CDBW-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm033; AVX512CDBW-NEXT: vzeroupper34; AVX512CDBW-NEXT: ret{{[l|q]}}35;36; AVX512VLCDBW-LABEL: test_mm_epi64:37; AVX512VLCDBW: # %bb.0: # %entry38; AVX512VLCDBW-NEXT: vpcmpeqw %xmm1, %xmm0, %k039; AVX512VLCDBW-NEXT: vpbroadcastmb2q %k0, %xmm040; AVX512VLCDBW-NEXT: ret{{[l|q]}}41entry:42 %0 = icmp eq <8 x i16> %a, %b43 %1 = bitcast <8 x i1> %0 to i844 %conv.i = zext i8 %1 to i6445 %vecinit.i.i = insertelement <2 x i64> undef, i64 %conv.i, i32 046 %vecinit1.i.i = shufflevector <2 x i64> %vecinit.i.i, <2 x i64> undef, <2 x i32> zeroinitializer47 ret <2 x i64> %vecinit1.i.i48}49 50define <4 x i32> @test_mm_epi32(<16 x i8> %a, <16 x i8> %b) {51; AVX512CD-LABEL: test_mm_epi32:52; AVX512CD: # %bb.0: # %entry53; AVX512CD-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm054; AVX512CD-NEXT: vpmovmskb %xmm0, %eax55; AVX512CD-NEXT: vmovd %eax, %xmm056; AVX512CD-NEXT: vpbroadcastd %xmm0, %xmm057; AVX512CD-NEXT: ret{{[l|q]}}58;59; AVX512CDBW-LABEL: test_mm_epi32:60; AVX512CDBW: # %bb.0: # %entry61; AVX512CDBW-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm162; AVX512CDBW-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm063; AVX512CDBW-NEXT: vpcmpeqb %zmm1, %zmm0, %k064; AVX512CDBW-NEXT: vpbroadcastmw2d %k0, %zmm065; AVX512CDBW-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm066; AVX512CDBW-NEXT: vzeroupper67; AVX512CDBW-NEXT: ret{{[l|q]}}68;69; AVX512VLCDBW-LABEL: test_mm_epi32:70; AVX512VLCDBW: # %bb.0: # %entry71; AVX512VLCDBW-NEXT: vpcmpeqb %xmm1, %xmm0, %k072; AVX512VLCDBW-NEXT: vpbroadcastmw2d %k0, %xmm073; AVX512VLCDBW-NEXT: ret{{[l|q]}}74entry:75 %0 = icmp eq <16 x i8> %a, %b76 %1 = bitcast <16 x i1> %0 to i1677 %conv.i = zext i16 %1 to i3278 %vecinit.i.i = insertelement <4 x i32> undef, i32 %conv.i, i32 079 %vecinit3.i.i = shufflevector <4 x i32> %vecinit.i.i, <4 x i32> undef, <4 x i32> zeroinitializer80 ret <4 x i32> %vecinit3.i.i81}82 83define <16 x i32> @test_mm512_epi32(<16 x i32> %a, <16 x i32> %b) {84; ALL-LABEL: test_mm512_epi32:85; ALL: # %bb.0: # %entry86; ALL-NEXT: vpcmpeqd %zmm1, %zmm0, %k087; ALL-NEXT: vpbroadcastmw2d %k0, %zmm088; ALL-NEXT: ret{{[l|q]}}89entry:90 %0 = icmp eq <16 x i32> %a, %b91 %1 = bitcast <16 x i1> %0 to i1692 %conv.i = zext i16 %1 to i3293 %vecinit.i.i = insertelement <16 x i32> undef, i32 %conv.i, i32 094 %vecinit15.i.i = shufflevector <16 x i32> %vecinit.i.i, <16 x i32> undef, <16 x i32> zeroinitializer95 ret <16 x i32> %vecinit15.i.i96}97 98define <8 x i64> @test_mm512_epi64(<8 x i32> %a, <8 x i32> %b) {99; AVX512CD-LABEL: test_mm512_epi64:100; AVX512CD: # %bb.0: # %entry101; AVX512CD-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1102; AVX512CD-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0103; AVX512CD-NEXT: vpcmpeqd %zmm1, %zmm0, %k0104; AVX512CD-NEXT: vpbroadcastmb2q %k0, %zmm0105; AVX512CD-NEXT: ret{{[l|q]}}106;107; AVX512CDBW-LABEL: test_mm512_epi64:108; AVX512CDBW: # %bb.0: # %entry109; AVX512CDBW-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1110; AVX512CDBW-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0111; AVX512CDBW-NEXT: vpcmpeqd %zmm1, %zmm0, %k0112; AVX512CDBW-NEXT: vpbroadcastmb2q %k0, %zmm0113; AVX512CDBW-NEXT: ret{{[l|q]}}114;115; AVX512VLCDBW-LABEL: test_mm512_epi64:116; AVX512VLCDBW: # %bb.0: # %entry117; AVX512VLCDBW-NEXT: vpcmpeqd %ymm1, %ymm0, %k0118; AVX512VLCDBW-NEXT: vpbroadcastmb2q %k0, %zmm0119; AVX512VLCDBW-NEXT: ret{{[l|q]}}120entry:121 %0 = icmp eq <8 x i32> %a, %b122 %1 = bitcast <8 x i1> %0 to i8123 %conv.i = zext i8 %1 to i64124 %vecinit.i.i = insertelement <8 x i64> undef, i64 %conv.i, i32 0125 %vecinit7.i.i = shufflevector <8 x i64> %vecinit.i.i, <8 x i64> undef, <8 x i32> zeroinitializer126 ret <8 x i64> %vecinit7.i.i127}128 129define <4 x i64> @test_mm256_epi64(<8 x i32> %a, <8 x i32> %b) {130; AVX512CD-LABEL: test_mm256_epi64:131; AVX512CD: # %bb.0: # %entry132; AVX512CD-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1133; AVX512CD-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0134; AVX512CD-NEXT: vpcmpeqd %zmm1, %zmm0, %k0135; AVX512CD-NEXT: vpbroadcastmb2q %k0, %zmm0136; AVX512CD-NEXT: # kill: def $ymm0 killed $ymm0 killed $zmm0137; AVX512CD-NEXT: ret{{[l|q]}}138;139; AVX512CDBW-LABEL: test_mm256_epi64:140; AVX512CDBW: # %bb.0: # %entry141; AVX512CDBW-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1142; AVX512CDBW-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0143; AVX512CDBW-NEXT: vpcmpeqd %zmm1, %zmm0, %k0144; AVX512CDBW-NEXT: vpbroadcastmb2q %k0, %zmm0145; AVX512CDBW-NEXT: # kill: def $ymm0 killed $ymm0 killed $zmm0146; AVX512CDBW-NEXT: ret{{[l|q]}}147;148; AVX512VLCDBW-LABEL: test_mm256_epi64:149; AVX512VLCDBW: # %bb.0: # %entry150; AVX512VLCDBW-NEXT: vpcmpeqd %ymm1, %ymm0, %k0151; AVX512VLCDBW-NEXT: vpbroadcastmb2q %k0, %ymm0152; AVX512VLCDBW-NEXT: ret{{[l|q]}}153entry:154 %0 = icmp eq <8 x i32> %a, %b155 %1 = bitcast <8 x i1> %0 to i8156 %conv.i = zext i8 %1 to i64157 %vecinit.i.i = insertelement <4 x i64> undef, i64 %conv.i, i32 0158 %vecinit3.i.i = shufflevector <4 x i64> %vecinit.i.i, <4 x i64> undef, <4 x i32> zeroinitializer159 ret <4 x i64> %vecinit3.i.i160}161 162define <8 x i32> @test_mm256_epi32(<16 x i16> %a, <16 x i16> %b) {163; AVX512CD-LABEL: test_mm256_epi32:164; AVX512CD: # %bb.0: # %entry165; AVX512CD-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0166; AVX512CD-NEXT: vpmovsxwd %ymm0, %zmm0167; AVX512CD-NEXT: vptestmd %zmm0, %zmm0, %k0168; AVX512CD-NEXT: vpbroadcastmw2d %k0, %zmm0169; AVX512CD-NEXT: # kill: def $ymm0 killed $ymm0 killed $zmm0170; AVX512CD-NEXT: ret{{[l|q]}}171;172; AVX512CDBW-LABEL: test_mm256_epi32:173; AVX512CDBW: # %bb.0: # %entry174; AVX512CDBW-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1175; AVX512CDBW-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0176; AVX512CDBW-NEXT: vpcmpeqw %zmm1, %zmm0, %k0177; AVX512CDBW-NEXT: vpbroadcastmw2d %k0, %zmm0178; AVX512CDBW-NEXT: # kill: def $ymm0 killed $ymm0 killed $zmm0179; AVX512CDBW-NEXT: ret{{[l|q]}}180;181; AVX512VLCDBW-LABEL: test_mm256_epi32:182; AVX512VLCDBW: # %bb.0: # %entry183; AVX512VLCDBW-NEXT: vpcmpeqw %ymm1, %ymm0, %k0184; AVX512VLCDBW-NEXT: vpbroadcastmw2d %k0, %ymm0185; AVX512VLCDBW-NEXT: ret{{[l|q]}}186entry:187 %0 = icmp eq <16 x i16> %a, %b188 %1 = bitcast <16 x i1> %0 to i16189 %conv.i = zext i16 %1 to i32190 %vecinit.i.i = insertelement <8 x i32> undef, i32 %conv.i, i32 0191 %vecinit7.i.i = shufflevector <8 x i32> %vecinit.i.i, <8 x i32> undef, <8 x i32> zeroinitializer192 ret <8 x i32> %vecinit7.i.i193}194 195