patch that caused regression PR opt/9357

patch that caused regression PR opt/9357

[Janis Johnson]

The regression reported in PR optimization/9357 (which affects GCC 3.3)
showed up with this patch:

Sat Oct 27 12:12:09 CEST 2001  Jan Hubicka  <jh@suse.cz>

        * cfgcleanup.c (merge_blocks_move_predecessor_nojumps,
        merge_blocks_move_successors_nojumps): Set BB_UPDATE_LIFE.
        (merge_blocks): Properly udpate life after making edge nonfallthru.
        * cfgrtl.c (merge_blocks_nomove): Update global_live_at_end.

I used the minimized test case test case and compiler options shown in
the PR.

This information has been added to GNATS.

Janis

Re: patch that caused regression PR opt/9357

[Andreas Jaeger]

Janis Johnson <janis187@us.ibm.com> writes:

> The regression reported in PR optimization/9357 (which affects GCC 3.3)

As just mentioned on gcc-bugs: The problem is -fssa and this is not
supported yet AFAIK.  I suspect that Honza's patch just revealed
another bug in -fssa.

Shouldn't we disable -fssa for now?

Andreas

> showed up with this patch:
>
> Sat Oct 27 12:12:09 CEST 2001  Jan Hubicka  <jh@suse.cz>
>
>         * cfgcleanup.c (merge_blocks_move_predecessor_nojumps,
>         merge_blocks_move_successors_nojumps): Set BB_UPDATE_LIFE.
>         (merge_blocks): Properly udpate life after making edge nonfallthru.
>         * cfgrtl.c (merge_blocks_nomove): Update global_live_at_end.
>
> I used the minimized test case test case and compiler options shown in
> the PR.
>
> This information has been added to GNATS.
>
> Janis
>

-- 
 Andreas Jaeger
  SuSE Labs aj@suse.de
   private aj@arthur.inka.de
    http://www.suse.de/~aj

Re: patch that caused regression PR opt/9357

[Steven Bosscher]

Op ma 24-02-2003, om 17:48 schreef Andreas Jaeger:
> Janis Johnson <janis187@us.ibm.com> writes:
> 
> > The regression reported in PR optimization/9357 (which affects GCC 3.3)
> 
> As just mentioned on gcc-bugs: The problem is -fssa and this is not
> supported yet AFAIK.  I suspect that Honza's patch just revealed
> another bug in -fssa.
> 
> Shouldn't we disable -fssa for now?

There are quite a few PRs about -fssa, so apparently people use it even
if it's not supported.  So just disabling is maybe not such a great
idea.

What is the long-term plan for RTL-SSA?  Last time I read the rtlopt web
page, it said that Honza had been working on MidRTL.  What's the plan
for that now?

Even with the tree-SSA branch merged, there are still cases where
RTL-SSA could be useful.  But if the plan is to kill it altogether,
maybe it could be deprecate the feature for 3.4 (assuming the tree-SSA
will succeed and be merged before "3.5")??

Greetz
Steven

Re: patch that caused regression PR opt/9357

[Jan Hubicka]

> Op ma 24-02-2003, om 17:48 schreef Andreas Jaeger:
> > Janis Johnson <janis187@us.ibm.com> writes:
> > 
> > > The regression reported in PR optimization/9357 (which affects GCC 3.3)
> > 
> > As just mentioned on gcc-bugs: The problem is -fssa and this is not
> > supported yet AFAIK.  I suspect that Honza's patch just revealed
> > another bug in -fssa.
> > 
> > Shouldn't we disable -fssa for now?
> 
> There are quite a few PRs about -fssa, so apparently people use it even
> if it's not supported.  So just disabling is maybe not such a great
> idea.
> 
> What is the long-term plan for RTL-SSA?  Last time I read the rtlopt web
> page, it said that Honza had been working on MidRTL.  What's the plan
> for that now?

I didn't work on midRTL for a while.  I think it makes sense to first
make the tree-SSA change and then see whether yet another level of
intermediate language is needed as doing midRTL correctly is quite a lot
of effort and it will slow down the compiler somewhat.
At cfg branch there is code for kind of midrtl on i386.  I do have
patches to virtualize most of RTL interfaces so we can switch RTL
representations in more clean way too, but didn't managed to clean it up
yet.  Currently merging what we do have on rtlopt branch, unit-at-a-time
and bugfixing is my priority.

Honza
> 
> Even with the tree-SSA branch merged, there are still cases where
> RTL-SSA could be useful.  But if the plan is to kill it altogether,
> maybe it could be deprecate the feature for 3.4 (assuming the tree-SSA
> will succeed and be merged before "3.5")??
> 
> Greetz
> Steven
>