This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Bug in CFG or CD-DCE?


On Wednesday 14 July 2004 03:17, Devang Patel wrote:
> On Jul 13, 2004, at 4:02 AM, Diego Novillo wrote:
> > On Tue, 2004-07-13 at 06:53, Diego Novillo wrote:
> >> On Tue, 2004-07-13 at 03:46, Steven Bosscher wrote:
> >>> Whether there's a FALLTHRU edge or not is irrelevant.  Block L22
> >>> is not control dependent on block L41, itt is control dependent on
> >>> block L26.
> >>
> >> Exactly.  Though, I did try clearing/recomputing PDOM info.  That's
> >> where the bug may be, ultimately.
> >
> > Wait.  I don't think that's right.  This code is testing CD with PDOM.
> > In this case, block L22 *does* post-dominate block L41.
>
> I do not know if it is useful or not but it is possible that cd-dce
> receives stale post dominator graph through calculate_dominance_info
> My fix was to free post dominators at the end of the pass that 
> calculates
> post dominators

Freeing did not help in this case.  I also thought that this was the
most likely cause of the problem, but it's not.

No, we don't find a _stale_ post dominator graph, we find a _broken_
one.  Noreturn blocks are improperly handled when computing post 
dominance info.  We are working on a fix (rth has a patch posted a
patch off-list).

Gr.
Steven


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]