Re: RFA: patch for register migration

[Andrew MacLeod <amacleod at redhat dot com>]

On Wed, 2004-04-21 at 10:52, Richard Kenner wrote:
>     I personally like/prefer the model of global colouring allocator
>     running, then a specialized local allocator doing the spill code. You
>     can then perform global allocation in a *single* interation, and leave
>     the spill code to be taken care of by a local allocator that understands
>     register usage in the local area of the spills.
> 
> I'm not arguing that the way we do things is ideal.  And the standard for
> a patch shouldn't be whether or not it brings things into an ideal mode.
> The standard ought to be whether the code being added makes the compiler
> better or worse.  I haven't studied the code in detail, but from a
> conceptual point of view, I think it makes the compiler better, despite
> it not converting things into an "ideal" design.

I was agreeing with you :-).  Reload is performing what I consider the
final "spilling allocator" role in many ways now. I also think reload
needs a new approach, but Im in favour of changes like this which
improve what it does in the here and now. Especially when they are
practical. 

Vlad is performing the exercise of discovering where our real weak
points are, and that will certainly be useful in the future too.
Anything he patches up now is a good thing. If we can fix the problems
he discovering in the correct earlier phase, great, then that part of
reload that he implemented could be disabled/removed at such time as
someone actually fixes it. 

Andrew