https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79437
--- Comment #4 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Vladimir Makarov from comment #3)
> 17: ax:DI=[sp:SI+0x4]
> 6: {ax:DI=ax:DI 0>>0x3f;clobber flags:CC;}
> 12: ax:SI=ax:SI
> 13: use ax:SI
>
> The code is fine with RA point of view. In any case, it hardly can be fixed
> in RA. And it is practically impossible to teach RA to predict all
> subsequent split decisions.
Interesting, that -fregmove doesn't improve the code, although the
cprop_hardreg pass is placed after the post-reload split.