[AArch64] Missed FCCMP opportunity

[SamTebbs33]

The code

bool foo2 (float a, float b)
{
  return __builtin_isnan (a) && __builtin_isnan (b);
}

with LLVM -O2 produces

foo2(float, float):                              // @foo2(float, float)
        fcmp    s0, s0
        cset    w8, vs
        fcmp    s1, s1
        cset    w9, vs
        and     w0, w8, w9
        ret

but GCC can produce:

foo2(float, float):
        fcmp    s1, s1
        fccmp   s0, s0, 0, vs
        cset    w0, vs
        ret

https://godbolt.org/z/e58hM4Pdq

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

[efriedma-quic]

For comparison, the following is optimized:

float foo2 (float a, float b, void f())
{
  return __builtin_isnan (a) && __builtin_isnan (b) ? a : b;
}

The relevant code is emitConditionalComparison; not sure why it doesn't trigger for the original testcase.

[SamTebbs33]

Thanks efriedma, that's very useful. I have someone working on this at the moment.