This is the mail archive of the
mailing list for the GCC project.
Re: Conflicting IRA patches
On 07/13/2010 09:37 PM, Jeff Law wrote:
>> I can look at what you did with left_conflict_size mechanism. I hope
>> to do this on this week. Currently I am a bit busy because I am
>> working on last AMD scheduling patch submitted last week.
> Reality is left_conflict_size has always been a little odd and it's
> going to get worse. Luckily I believe getting it wrong merely results
> in worse code and not incorrect code.
Heh. We'd still want to avoid that.
With the current version of my patch, I think I made this code slightly
more accurate. There are three possible types of conflict for two
2-word registers (four if you count no conflict), each of which removes
one possible starting register. So you can end up with any number in
[0..3] for the number of registers removed from the size of the
available set, rather than choosing between 0 and 2 (i.e. nregs). In
theory this should be possible with Vlad's new code, but I don't
understand what it does yet.
Vlad, whatever Jeff decides, I'll give you time to look at this issue.
I don't think I urgently need to check this in before end of next week
or so. Thanks for agreeing to let the DImode stuff have priority.