Re: New Regalloc development stalled?

[Daniel Berlin <dberlin at dberlin dot org>]

On Thursday, December 5, 2002, at 01:13  PM, Nathanael Nerode wrote:

> Michael Matz said
> >On 21 Nov 2002, Gabriel Dos Reis wrote:
> >
> >> | Known problem.  Some sources have even higher additional costs. 
> >This will
> >> | be fixed hopefully before 3.4.
> >>
> >> Is there any actual plan for that? I know there is a branch, but I
> >> didn't see any actual plan.
> >
> >Unfortunately it's just in my head.  The current representation of the
> >conflict graph is highly suboptimal for splitting coalesced nodes. 
> Due >to
> >that there are some hot fixes in it (to keep it correct), which 
> >basically
> >needs something like O(n^3) (worst case) per splitted node.  Those
> This is not good. :-)
But it's also something plenty of  other implementations that are quite 
fast share as a problem (in terms of worst case).


> >won't
> >be needed in the final version.  But even without that the underlying
> >representation is not very well fit for that kind of things, so I 
> plan >to
> >rewrite it into something using the usual conflict bitmaps, but 
> helped >by
> >some hash table for "difficult" edges (those which would need a 
> >reference
> >count).  Another problem is the scattering of information over 
> >different
> >data structures, which sometimes needs translation from one to the 
> >other
> >data set.  That also needs cleanup.
> Have you made any progress on this data structure rewrite?  It seems 
> that new-ra isn't really a practical replacement for the existing one 
> until these issues are solved.
No, it's not a practical replacement until it's as fast, regardless of 
whether data structures are rewritten or not.



> >That having said some other people already asked for something like a 
> >web
> >page listing things we tried already, and things planned and such, 
> for >the
> >allocator.  I promised to write something over the next days ;-)
> But you didn't, did you. ;-)
This is how things are when people lack time.
We don't use Bugzilla right now because the new sources.redhat.com 
server redhat has been promising (Estimate was september) has been held 
up because Chris Faylor hasn't had time to work on it.
I don't blame him, that's just how things go sometimes.
I'm sure, eventually, it'll happen.
If you want stuff done on new-ra, *help*, don't just complain.

> How about doing that now?  At any rate it will give other people a 
> better chance of working on new-ra.
I hope this means you plan on working on it as well, rather than simply 
ragging on Michael for being a volunteer.
I had a hand in new-ra too.
If you have questions you need answers to that are preventing you from 
working on it, or speeding it up, feel free to ask, and i'll answer 
them if michael is too busy.
Otherwise, why don't you help profile and speed it up, or fix bugs?


> --Nathanael
>
> Ciao,
> Michael.