[PATCH] Fix PR69274, 435.gromacs performance regression due to RA
Mon Feb 8 12:06:00 GMT 2016
On 02/08/2016 10:09 AM, Richard Biener wrote:
> The gcc.target/i386/addr-sel-1.c (for PR28940) seems to just started
> working at some point past in time and thus it was added and the
> bug closed. You could say RA does a better job after the patch
> as it uses 1 less register but that restricts the followup
> postreload combine attempts. Though I wonder about what's "better"
> RA here - isn't the best allocation one that avoids spills but
> uses as many registers as possible (at least when targeting a CPU
> that cannot to register renaming)? regrename doesn't help this
> testcase either (it runs too late and does a renaming that doesn't help).
I don't think regrename's place in the pass pipeline is set in stone. If
we enable it for gcc-7. we could also experiment with putting it just
before postreload, or schedule another reload-combine afterwards.
More information about the Gcc-patches