Re: It's time to merge new-ra branch into mainline

[Daniel Berlin <dberlin at dberlin dot org>]

On Jun 6, 2004, at 2:22 PM, Bradley Lucier wrote:

> It's time to merge the new-ra branch into mainline.
>
> We've gone through two major releases (3.3 and 3.4) with -fnew-ra 
> "experimental" and without even bugfixes being merged.  There are bugs 
> in bugzilla reported against various mainlines that have been fixed on 
> -fnew-ra.  new-ra will get more testing and will perhaps attract more 
> hacking attention if it's part of mainline.

If Vlad follows through (and i have absolutely no reason to believe he 
won't) with the improvements specified in his paper, there won't be 
much point in what is left in new-ra, except for pre-reload.
By "not much point", i mean that I believe we won't get much better 
performance using new-ra than with his patches[1]
If/when his patches go in, unless new-ra is still providing some 
performance benefit, I would suggest we remove new-ra from the mainline 
and declare it dead.

Viva la register allocation,
Dan

[1] I still believe the abstraction we currently use in 
global/local-alloc is the wrong one. Coloring the registers easier is 
not the real point of using a graph coloring allocator.  Coloring the 
registers is the easy part of any allocator.  It's that it makes it 
easier to do other things, like coalescing, etc.  I also don't 
necessarily agree that all the things he is implementing as a separate 
pass actually belong as separate passes, since in order to get the best 
performance, they need to share information.  However, since i'm not 
about to redo this stuff myself anytime soon, I'm not in a position to 
demonstrate code that does it better, and thus, simply made this a 
footnote :).