Re: IRA for GCC 4.4

[Vladimir Makarov <vmakarov at redhat dot com>]

J.C. Pizarro wrote:
> On Tue, 29 Apr 2008 20:25:56 +0200, "Steven Bosscher"
> <stevenb.gcc@gmail.com> 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.
> >
> > 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.
> >
> > Another note: IRA uses VARRAYs, and I was under the impression we are
> > trying to move away from those.  Vlad, please use VECs instead.
> >
> > Gr.
> > Steven
> >     
>
> Use my idea of flexible chained upper triangulars & rectangulars as
>  indicated briefly in
>
> http://gcc.gnu.org/ml/gcc/2008-04/msg00707.html
> http://gcc.gnu.org/ml/gcc/2008-04/msg00681.html
>
> You can update easily these structures through subroutines that
>  traverse the chains and update their chained local structures when
>  is needed, and is similar to the Observer pattern (when the subjects
>  are modified then it update the views of the observers too).
>
>   
Your approach was interesting but it needs some partition of allocnos.  
This partition is present in Peter's code (local allocnos in each BB and 
global allocnos).  Partition is not easy to do when  live ranges  are 
used to find potential conflicts.  Also such partition is less accurate 
and will result bigger matrices than using live ranges to find potential 
conflict for each allocno pair.

The part of what you are proposing (partitioning allocnos) is close to 
many implementation of register allocators, e.g. Chow in his 
implementation of priority coloring built conflicts only for global 
allocnos to save memory.