1; RUN: llc -fp-contract=fast -mattr=+fma -disable-cgp < %s -o - | FileCheck %s
2; Check that the 2nd and 3rd arguments of fmaXXX231 reg1, reg2, mem3 are not commuted.
3; <rdar://problem/16800495>
4target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
5target triple = "x86_64-apple-macosx"
6
7; CHECK-LABEL: test1:
8; %arg lives in xmm0 and it shouldn't be redefined until it is used in the FMA.
9; CHECK-NOT: {{.*}}, %xmm0
10; %addr lives in rdi.
11; %addr2 lives in rsi.
12; CHECK: vmovss (%rsi), [[ADDR2:%xmm[0-9]+]]
13; The assembly syntax is in the reverse order.
14; CHECK: vfmadd231ss (%rdi), [[ADDR2]], %xmm0
15define void @test1(float* %addr, float* %addr2, float %arg) {
16entry:
17  br label %loop
18
19loop:
20  %sum0 = phi float [ %fma, %loop ], [ %arg, %entry ]
21  %addrVal = load float, float* %addr, align 4
22  %addr2Val = load float, float* %addr2, align 4
23  %fmul = fmul float %addrVal, %addr2Val
24  %fma = fadd float %sum0, %fmul
25  br i1 true, label %exit, label %loop
26
27exit:
28  store float %fma, float* %addr, align 4
29  ret void
30}
31