[Bug target/67032] [4.9/5/6 Regression] Geode optimizations incorrectly return -NaN

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67032

--- Comment #12 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Richard Biener from comment #11)
> I thought using %xmm registers is _always_ wrong unless the user properly
> saved/restored the FP stack state.
> 
> So not sure why we still have MMX support at all (apart from for intrinsics
> which probably need to expose things to the RA again... sigh).

In the above testcase, the %mm register was allocated automatically, without
using intrinsics. There is a penalty for allocation of %mm (see Comment #6),
but it is ineffective for recent gcc releases.