39 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mcpu=corei7-avx -mtriple=x86_64-linux | FileCheck %s3 4; A test from pifft (after SLP-vectorization) that fails when we drop the chain on newly merged loads.5define void @cftx020(ptr nocapture %a) {6; CHECK-LABEL: cftx020:7; CHECK: # %bb.0: # %entry8; CHECK-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero9; CHECK-NEXT: vmovhpd {{.*#+}} xmm0 = xmm0[0],mem[0]10; CHECK-NEXT: vpermilpd {{.*#+}} xmm1 = mem[1,0]11; CHECK-NEXT: vaddpd %xmm1, %xmm0, %xmm012; CHECK-NEXT: vmovupd (%rdi), %xmm113; CHECK-NEXT: vmovupd %xmm0, (%rdi)14; CHECK-NEXT: vsubpd 16(%rdi), %xmm1, %xmm015; CHECK-NEXT: vmovupd %xmm0, 16(%rdi)16; CHECK-NEXT: retq17entry:18 %0 = load double, ptr %a, align 819 %arrayidx1 = getelementptr inbounds double, ptr %a, i64 220 %1 = load double, ptr %arrayidx1, align 821 %arrayidx2 = getelementptr inbounds double, ptr %a, i64 122 %2 = load double, ptr %arrayidx2, align 823 %arrayidx3 = getelementptr inbounds double, ptr %a, i64 324 %3 = load double, ptr %arrayidx3, align 825 %4 = insertelement <2 x double> undef, double %0, i32 026 %5 = insertelement <2 x double> %4, double %3, i32 127 %6 = insertelement <2 x double> undef, double %1, i32 028 %7 = insertelement <2 x double> %6, double %2, i32 129 %8 = fadd <2 x double> %5, %730 store <2 x double> %8, ptr %a, align 831 %9 = insertelement <2 x double> undef, double %0, i32 032 %10 = insertelement <2 x double> %9, double %2, i32 133 %11 = insertelement <2 x double> undef, double %1, i32 034 %12 = insertelement <2 x double> %11, double %3, i32 135 %13 = fsub <2 x double> %10, %1236 store <2 x double> %13, ptr %arrayidx1, align 837 ret void38}39