173 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=X863; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=X644 5declare <4 x float> @llvm.x86.avx2.gather.d.ps(<4 x float>, ptr,6 <4 x i32>, <4 x float>, i8) nounwind readonly7 8define <4 x float> @test_x86_avx2_gather_d_ps(ptr %a1, <4 x i32> %idx, <4 x float> %mask) {9; X86-LABEL: test_x86_avx2_gather_d_ps:10; X86: # %bb.0:11; X86-NEXT: movl {{[0-9]+}}(%esp), %eax12; X86-NEXT: vxorps %xmm2, %xmm2, %xmm213; X86-NEXT: vgatherdps %xmm1, (%eax,%xmm0,2), %xmm214; X86-NEXT: vmovaps %xmm2, %xmm015; X86-NEXT: retl16;17; X64-LABEL: test_x86_avx2_gather_d_ps:18; X64: # %bb.0:19; X64-NEXT: vxorps %xmm2, %xmm2, %xmm220; X64-NEXT: vgatherdps %xmm1, (%rdi,%xmm0,2), %xmm221; X64-NEXT: vmovaps %xmm2, %xmm022; X64-NEXT: retq23 %res = call <4 x float> @llvm.x86.avx2.gather.d.ps(<4 x float> undef,24 ptr %a1, <4 x i32> %idx, <4 x float> %mask, i8 2) ;25 ret <4 x float> %res26}27 28declare <2 x double> @llvm.x86.avx2.gather.d.pd(<2 x double>, ptr,29 <4 x i32>, <2 x double>, i8) nounwind readonly30 31define <2 x double> @test_x86_avx2_gather_d_pd(ptr %a1, <4 x i32> %idx, <2 x double> %mask) {32; X86-LABEL: test_x86_avx2_gather_d_pd:33; X86: # %bb.0:34; X86-NEXT: movl {{[0-9]+}}(%esp), %eax35; X86-NEXT: vxorpd %xmm2, %xmm2, %xmm236; X86-NEXT: vgatherdpd %xmm1, (%eax,%xmm0,2), %xmm237; X86-NEXT: vmovapd %xmm2, %xmm038; X86-NEXT: retl39;40; X64-LABEL: test_x86_avx2_gather_d_pd:41; X64: # %bb.0:42; X64-NEXT: vxorpd %xmm2, %xmm2, %xmm243; X64-NEXT: vgatherdpd %xmm1, (%rdi,%xmm0,2), %xmm244; X64-NEXT: vmovapd %xmm2, %xmm045; X64-NEXT: retq46 %res = call <2 x double> @llvm.x86.avx2.gather.d.pd(<2 x double> undef,47 ptr %a1, <4 x i32> %idx, <2 x double> %mask, i8 2) ;48 ret <2 x double> %res49}50 51declare <8 x float> @llvm.x86.avx2.gather.d.ps.256(<8 x float>, ptr,52 <8 x i32>, <8 x float>, i8) nounwind readonly53 54define <8 x float> @test_x86_avx2_gather_d_ps_256(ptr %a1, <8 x i32> %idx, <8 x float> %mask) {55; X86-LABEL: test_x86_avx2_gather_d_ps_256:56; X86: # %bb.0:57; X86-NEXT: movl {{[0-9]+}}(%esp), %eax58; X86-NEXT: vxorps %xmm2, %xmm2, %xmm259; X86-NEXT: vgatherdps %ymm1, (%eax,%ymm0,4), %ymm260; X86-NEXT: vmovaps %ymm2, %ymm061; X86-NEXT: retl62;63; X64-LABEL: test_x86_avx2_gather_d_ps_256:64; X64: # %bb.0:65; X64-NEXT: vxorps %xmm2, %xmm2, %xmm266; X64-NEXT: vgatherdps %ymm1, (%rdi,%ymm0,4), %ymm267; X64-NEXT: vmovaps %ymm2, %ymm068; X64-NEXT: retq69 %res = call <8 x float> @llvm.x86.avx2.gather.d.ps.256(<8 x float> undef,70 ptr %a1, <8 x i32> %idx, <8 x float> %mask, i8 4) ;71 ret <8 x float> %res72}73 74declare <4 x double> @llvm.x86.avx2.gather.d.pd.256(<4 x double>, ptr,75 <4 x i32>, <4 x double>, i8) nounwind readonly76 77define <4 x double> @test_x86_avx2_gather_d_pd_256(ptr %a1, <4 x i32> %idx, <4 x double> %mask) {78; X86-LABEL: test_x86_avx2_gather_d_pd_256:79; X86: # %bb.0:80; X86-NEXT: movl {{[0-9]+}}(%esp), %eax81; X86-NEXT: vxorpd %xmm2, %xmm2, %xmm282; X86-NEXT: vgatherdpd %ymm1, (%eax,%xmm0,8), %ymm283; X86-NEXT: vmovapd %ymm2, %ymm084; X86-NEXT: retl85;86; X64-LABEL: test_x86_avx2_gather_d_pd_256:87; X64: # %bb.0:88; X64-NEXT: vxorpd %xmm2, %xmm2, %xmm289; X64-NEXT: vgatherdpd %ymm1, (%rdi,%xmm0,8), %ymm290; X64-NEXT: vmovapd %ymm2, %ymm091; X64-NEXT: retq92 %res = call <4 x double> @llvm.x86.avx2.gather.d.pd.256(<4 x double> undef,93 ptr %a1, <4 x i32> %idx, <4 x double> %mask, i8 8) ;94 ret <4 x double> %res95}96 97define <2 x i64> @test_mm_i32gather_epi32(ptr%a0, <2 x i64> %a1) {98; X86-LABEL: test_mm_i32gather_epi32:99; X86: # %bb.0:100; X86-NEXT: movl {{[0-9]+}}(%esp), %eax101; X86-NEXT: vpcmpeqd %xmm2, %xmm2, %xmm2102; X86-NEXT: vpxor %xmm1, %xmm1, %xmm1103; X86-NEXT: vpgatherdd %xmm2, (%eax,%xmm0,2), %xmm1104; X86-NEXT: vmovdqa %xmm1, %xmm0105; X86-NEXT: retl106;107; X64-LABEL: test_mm_i32gather_epi32:108; X64: # %bb.0:109; X64-NEXT: vpcmpeqd %xmm2, %xmm2, %xmm2110; X64-NEXT: vpxor %xmm1, %xmm1, %xmm1111; X64-NEXT: vpgatherdd %xmm2, (%rdi,%xmm0,2), %xmm1112; X64-NEXT: vmovdqa %xmm1, %xmm0113; X64-NEXT: retq114 %arg0 = bitcast ptr%a0 to ptr115 %arg1 = bitcast <2 x i64> %a1 to <4 x i32>116 %mask = bitcast <2 x i64> <i64 -1, i64 -1> to <4 x i32>117 %call = call <4 x i32> @llvm.x86.avx2.gather.d.d(<4 x i32> zeroinitializer, ptr %arg0, <4 x i32> %arg1, <4 x i32> %mask, i8 2)118 %bc = bitcast <4 x i32> %call to <2 x i64>119 ret <2 x i64> %bc120}121declare <4 x i32> @llvm.x86.avx2.gather.d.d(<4 x i32>, ptr, <4 x i32>, <4 x i32>, i8) nounwind readonly122 123define <2 x double> @test_mm_i32gather_pd(ptr%a0, <2 x i64> %a1) {124; X86-LABEL: test_mm_i32gather_pd:125; X86: # %bb.0:126; X86-NEXT: movl {{[0-9]+}}(%esp), %eax127; X86-NEXT: vxorpd %xmm1, %xmm1, %xmm1128; X86-NEXT: vpcmpeqd %xmm2, %xmm2, %xmm2129; X86-NEXT: vgatherdpd %xmm2, (%eax,%xmm0,2), %xmm1130; X86-NEXT: vmovapd %xmm1, %xmm0131; X86-NEXT: retl132;133; X64-LABEL: test_mm_i32gather_pd:134; X64: # %bb.0:135; X64-NEXT: vxorpd %xmm1, %xmm1, %xmm1136; X64-NEXT: vpcmpeqd %xmm2, %xmm2, %xmm2137; X64-NEXT: vgatherdpd %xmm2, (%rdi,%xmm0,2), %xmm1138; X64-NEXT: vmovapd %xmm1, %xmm0139; X64-NEXT: retq140 %arg0 = bitcast ptr%a0 to ptr141 %arg1 = bitcast <2 x i64> %a1 to <4 x i32>142 %cmp = fcmp oeq <2 x double> zeroinitializer, zeroinitializer143 %sext = sext <2 x i1> %cmp to <2 x i64>144 %mask = bitcast <2 x i64> %sext to <2 x double>145 %res = call <2 x double> @llvm.x86.avx2.gather.d.pd(<2 x double> zeroinitializer, ptr %arg0, <4 x i32> %arg1, <2 x double> %mask, i8 2)146 ret <2 x double> %res147}148 149@x = dso_local global [1024 x float] zeroinitializer, align 16150 151define <4 x float> @gather_global(<4 x i64>, ptr nocapture readnone) {152; X86-LABEL: gather_global:153; X86: # %bb.0:154; X86-NEXT: vpcmpeqd %xmm2, %xmm2, %xmm2155; X86-NEXT: vxorps %xmm1, %xmm1, %xmm1156; X86-NEXT: vgatherqps %xmm2, x(,%ymm0,4), %xmm1157; X86-NEXT: vmovaps %xmm1, %xmm0158; X86-NEXT: vzeroupper159; X86-NEXT: retl160;161; X64-LABEL: gather_global:162; X64: # %bb.0:163; X64-NEXT: vpcmpeqd %xmm2, %xmm2, %xmm2164; X64-NEXT: vxorps %xmm1, %xmm1, %xmm1165; X64-NEXT: vgatherqps %xmm2, x(,%ymm0,4), %xmm1166; X64-NEXT: vmovaps %xmm1, %xmm0167; X64-NEXT: vzeroupper168; X64-NEXT: retq169 %3 = tail call <4 x float> @llvm.x86.avx2.gather.q.ps.256(<4 x float> zeroinitializer, ptr @x, <4 x i64> %0, <4 x float> <float 0xFFFFFFFFE0000000, float 0xFFFFFFFFE0000000, float 0xFFFFFFFFE0000000, float 0xFFFFFFFFE0000000>, i8 4)170 ret <4 x float> %3171}172declare <4 x float> @llvm.x86.avx2.gather.q.ps.256(<4 x float>, ptr, <4 x i64>, <4 x float>, i8)173