Re: [tree-ssa] Bootstrap broken on PPC

[Jan Hubicka <jh at suse dot cz>]

> 
> On Nov 8, 2003, at 11:36 AM, Jan Hubicka wrote:
> 
> >>
> >>Jan,
> >>
> >>This breakage appears related to your changes.
> >>
> >>---------------------------------------------------------------------- 
> >>------
> >>/home/cygnus/dnovillo/perf/sbox/tree-ssa-branch/local.ppc64/src/gcc/ 
> >>builtins.c:5941: warning: 'result' might be used uninitialized in  
> >>this function
> >
> >It is not in functio I've touched.  It looks like that one of our
> >optimizations confuse CFG enought so there is new path to exit not
> >initializing result (the result is initialized in many interesting ways
> >inside switch statement).
> >I would suspect that this is result of combination of fact that edge
> >splitting is broken (it redirects CFG edges without updating  the
> >underlying program code and it also invalidate SSA form by not updating
> >PHI nodes for new edges as noticed by verify_flow_info patch I sent)
> >and Daniel's changes to make ssapre more aggressive.
> 
> Just a note, the changes below don't actually make it more aggressive  
> (those were committed a few days ago, and they were bootstrapped on  
> ppc64), they just speed it up.
> 
OK, then my theory has problem.  I will look into what is really going
on.

Anyway we should sort out the edge splitting problems as they makes me
nervous at minimum.  I got edge splitting right in most common cases
that were firing verify_flow_info by somewhat dirty hack,
Zdenek has proper sollution (rewrite of the whole thing) but it depends
on his double linking changes that depends on Richard's work.  So
perhaps I can send the dirty patch?

Honza