Re: treelang fix for rs6000 / treelang not proper front end

[Steven Bosscher <s dot bosscher at student dot tudelft dot nl>]

Gerald Pfeifer wrote:

> On Tue, 10 Jun 2003, Andrew Haley wrote:
>  
>
> > Yes.  My current plan is to convert the gcj byte compiler as part of
> > converting java to tree-ssa.  If that's going to be too late in the
> > release cycle I guess my plan needs to change.
> >    
>
> Which release?
>
> Last time I tested, tree-ssa slowed down our already slow compiler quite
> a bit, so I cannot imagine tree-ssa being part of a release in the near
> future, as much as I like that approach and project. :-(
>
> (See PR8361 for an example.)
>
> Gerald
That's not quite fair on tree-ssa.  It really isn't that slow AFAICT.

The version you tested did not even have a proper un-ssa pass, and the 
hack that was in place instead was just butt-slow.  But if you go look 
at the individual built times from Diego's SPEC testers, you'll see that 
tree-ssa is really not _that_ much slower: Something in the order of 
10%  (Don't look at the total build times because 252.eon apparently 
doesn't build with mainline for Diego).  And there is still much tuning 
to do, and figuring out which (parts of certain) RTL optimizers are 
redundant with tree-ssa enabled.

Gr.
Steven