[PATCH] Fix PR69274, 435.gromacs performance regression due to RA
Mon Feb 8 15:47:00 GMT 2016
On 02/08/2016 04:09 AM, Richard Biener wrote:
> With all of the above I'm not sure what to do for GCC 6 (even though
> you just approved the patch). Going with the patch alternative (just
> revert swapping parts of the commutative operands) looks like
> completely bogus though it works for fixing the regression as well.
Yes, it is bogus. But I don't see other easy way to close P1 PR except
making it P2.
I still have plans to rewrite ira-costs.c (may be for GCC7). The
algorithm is actually adaptation of old regclass.c one. There are few
complaints about wrong costs calculations because of this (the most
recent complaint was Peter Bergner's one about a power8 test). I don't
like the algorithm, it ignores the fact that insn operands should be
from the same alternative during pseudo (allocno) cost calculations.
So this change probably will not survive after rewriting ira-costs.c.
More information about the Gcc-patches