Re: Register classes

[Daniel Berlin <dberlin at dberlin dot org>]

On Aug 19, 2004, at 8:56 AM, Michael Matz wrote:

> Hi,
>
> On Wed, 18 Aug 2004, Daniel Berlin wrote:
>
> > Just so nobody has any doubts, i have no current plans to do anything 
> > to
> > the register allocator. I'm just trying to gather more information in
> > case i decide to mount another future assault on it.
> >
> > I should point out something from the below paper though:
>
> As this paper didn't yet show up in citeseer (it seems), I would be
> thankful for this paper.
I'll send it to you.

> Note that new-ra does not really use register classes for the 
> interference
> graph, but instead register sets.
Except they are taken from register classes.

> Trivially colorable nodes are still trivially colorable also in new-ra,
> _as long_ as they only require one register, so I'm not exactly sure 
> what
> you were speaking about in your first mail.

Uh, handling multi-regs is something very ugly in new-ra right now.
Also, the interference graph in new-ra is split into two graphs, which 
is also a bad idea.
(sup_igraph and igraph).

> > "One of our key insights is that coloring constraints on each
> > interference-graph node should be expressed in terms of the set
>
> Without having read the paper, at least this part I think is basically
> implemented in new-ra.

But it's not, because we don't gracefully handle multi-reg pseudos and 
whatnot simply by using the interference graph.  We go through all 
kinds of contortions.

>   Sometimes we still use the regclasses for
> preferring some regs.  And of course when looking on the constraints 
> of an
> insn you have to deal with regclasses.
>
>
> Ciao,
> Michael.