Re: IRA for GCC 4.4

[Vladimir Makarov <vmakarov at redhat dot com>]

Steven Bosscher wrote:
> On Tue, Apr 29, 2008 at 6:22 PM, Mark Mitchell <mark@codesourcery.com> wrote:
>   
> >  Vladimir, if you feel that Peter's code cannot be used directly in IRA,
> > would you please explain why not?
> >     
>
> I think he already has explained, see
> http://gcc.gnu.org/ml/gcc/2008-04/msg00730.html
>
> Having looked at IRA a bit, I think I have to agree with Vlad that
> Peter's code is not easily adapted to IRA.  Peter's code works for a
> single, immutable conflict graph for the whole function.  IRA works
> with inter-region and intra-region conflicts (as far as I understand,
> documentation in ira-conflict.c would be welcome), so the sorting
> trick that Peter uses, doesn't translate one-to-one to Vlad's
> allocator.
>
>   
Thanks, Steven.  It is really harder to incorporate.  Also, using live 
ranges potentially results in more compact bit vectors.  I forgot to say 
that conflict bit vectors and adjacency lists (vector of pointers to 
conflict allocnos) in IRA contains only allocno of the same cover 
classes.  I can use the fact and compress conflict vectors even more.
> Having said that, I think the "square" approach with
> mirror_conflicts() that IRA currently has, is a big step backward from
> what we have now.  IRA should at least have a representation for
> conflicts that does not duplicate information unnecessary.  The bits
> that seem to be bad in this respect are build_conflict_bit_table() and
> mirror_conflicts().  It's not clear to me how these are used, but it
> looks like you can end up building a square conflict graph for the
> whole function, like GCC did before Peter's work.  This could be a
> huge memory and speed regression if it isn't addressed.
>
>   
It will be addressed (I am working on this).  There will be no mirror 
conflicts anymore.  But I don't use and I am not going to use triangular 
matrix.  It means more memory from one point of view.  On other had, It 
permits me to use conflict bit vector for implementation of adjacency 
lists (and not allocate vector of pointers to conflict allocnos saving 
memory) if the conflict bit vector is smaller than vectors of the pointers.
> Another note: IRA uses VARRAYs, and I was under the impression we are
> trying to move away from those.  Vlad, please use VECs instead.
>   
Steven, thanks for pointing this.  I'll remove VARRAYs.  It is probably 
code which was written long ago.  I tried to use VECs in more fresh code.