Re: [PATCH] move superblock formation to Tree-SSA

[Robert Kidd <rkidd at crhc dot uiuc dot edu>]

On Dec 26, 2006, at 6:56 PM, Jan Hubicka wrote:

> Robert,
> thanks for bringing tracer to SSA lavel.  The implementation seems  
> sane
> to me (modulo the problem mentioned by Zdenek) and missing comments in
> the front of new functions required by coding standard.
> I wonder if you did some actual benchmarks comparing old and new  
> tracer
> behaviour?  The placement in pass queue seems a bit early to me, since
> tracer might actually confuse some other optimizations and thus it  
> might
> be profitable to move it down in the queue.

I experimented a little with moving the pass with respect to the  
first may_alias pass, but I haven't tried moving it around any  
optimization passes.  I'll do a couple runs based on Steven's  
suggestions.

>   It might be even profitable
> to retain the current RTL functionality as we do for some other code
> duplicating optimizations, such as unroller, so ther later tracer can
> duplicate stuff that compiled to small code and was missed by earlier
> tracing.

Interesting.  Is there a way for a pass to dynamically determine if  
it is operating on Tree-SSA or RTL?  The tracer pass is mostly the  
same for either, so the same code could potentially work on either.

> If it is problematic for you to run few benchmarks, I can do SPECs on
> AMD64.

I have spec set up on a few machines here.  I'll kick off a rigorous  
run.

Thanks
Robert