1; RUN: llc -verify-machineinstrs -mtriple=powerpc64-linux-gnu -mcpu=pwr8 -mattr=+vsx < %s | FileCheck %s
2
3define <4 x float> @bar(float* %p, float* %q) {
4  %1 = bitcast float* %p to <12 x float>*
5  %2 = bitcast float* %q to <12 x float>*
6  %3 = load <12 x float>, <12 x float>* %1, align 16
7  %4 = load <12 x float>, <12 x float>* %2, align 16
8  %5 = fsub <12 x float> %4, %3
9  %6 = shufflevector <12 x float> %5, <12 x float> undef, <4 x i32> <i32 0, i32 3, i32 6, i32 9>
10  ret <4 x float>  %6
11
12; CHECK: xxsldwi
13; CHECK-DAG: vmrghw
14; CHECK-DAG: vmrglw
15; CHECK-NEXT: xxsldwi
16; CHECK-NEXT: xxsldwi
17; CHECK-NEXT: xxsldwi
18; CHECK-NEXT: blr
19}
20