1; RUN: llc -mtriple=arm-eabi -float-abi=hard -mcpu=cortex-a15 -mattr=+neon,+neonfp %s -o - \
2; RUN:  | FileCheck %s
3
4; This test checks that the VMLxForwarting feature is disabled for A15.
5; CHECK: fun_a:
6define <4 x i32> @fun_a(<4 x i32> %x, <4 x i32> %y) nounwind{
7  %1 = add <4 x i32> %x, %y
8; CHECK-NOT: vmul
9; CHECK: vmla
10  %2 = mul <4 x i32> %1, %1
11  %3 = add <4 x i32> %y, %2
12  ret <4 x i32> %3
13}
14
15; This tests checks that VMLA FP patterns can be matched in instruction selection when targeting
16; Cortex-A15.
17; CHECK: fun_b:
18define <4 x float> @fun_b(<4 x float> %x, <4 x float> %y, <4 x float> %z) nounwind{
19; CHECK: vmla.f32
20  %t = fmul <4 x float> %x, %y
21  %r = fadd <4 x float> %t, %z
22  ret <4 x float> %r
23}
24
25; This tests checks that FP VMLA instructions are not expanded into separate multiply/addition
26; operations when targeting Cortex-A15.
27; CHECK: fun_c:
28define <4 x float> @fun_c(<4 x float> %x, <4 x float> %y, <4 x float> %z, <4 x float> %u, <4 x float> %v) nounwind{
29; CHECK: vmla.f32
30  %t1 = fmul <4 x float> %x, %y
31  %r1 = fadd <4 x float> %t1, %z
32; CHECK: vmla.f32
33  %t2 = fmul <4 x float> %u, %v
34  %r2 = fadd <4 x float> %t2, %r1
35  ret <4 x float> %r2
36}
37
38