This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: Conflicting IRA patches
Jeff Law wrote:
On 07/07/10 14:35, Bernd Schmidt wrote:
Yes. As long as we have something for which we can compute
{max,min}_nregs, it should be OK - if these numbers are both 2, then
allocate two objects.
OK. I'll put these concerns aside and trust we'll get the details
sorted out.
I was more worried about coalesce_allocnos, but it seems Vlad's patches
even remove some of the headaches I've had with that. I'm not sure yet
how to deal with the new left_conflicts_size mechanism.
I hadn't gotten this far -- I was just starting to look at the patches
and saw that there were likely significant conflicts that we were
going to need to work through and figured we needed to open a discussion.
Some of the speedup patches maybe could have gone into trunk directly
instead of a branch; that would have eliminated some sources of
divergence.
Yea. I approved one which looked like it could go in independently
right now. I looked at the others and they probably could go in with
some relatively minor massaging.
My temptation is to place priority on the double-word stuff & speedups
to reduce the amount of divergence we're dealing with, then sort
through the issues the removal of cover classes introduces.
Comments?
I checked Bernd's patch on SPEC2000 on x86 (32-bit modes) last week.
Code of few benchmarks changed (only vpr, gcc, crafty, parser, eon,
vortex, mesa, facerec, sixtrack, apsi). The codes size for SPECINT2000
was improved by 0.03% with patch and code size for SPECFP2000 was worsen
by 0.006%. There is no visible improvement/degradation on changed
benchmarks. Compilation time increase was quite insignificant (less
than 0.2%). So as I wrote Ken Zadeck two years ago who tried to solve
the problem for the old RA, the complexity of the problem solution does
not justify
the solution benefits.
As for patch priority for checking in, I think you are right especially
if we take that I have a branch for my work and Bernd does not have one
and resolution of the conflicts could be done on ira-improv branch.
By the way, there is already another big conflict with recent Richard
Sandiford's patches. Although it is a smaller conflict in comparison
with one of Bernd's and my patches.