This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: bug in dominance algorithm
- To: Jan Hubicka <jh at suse dot cz>
- Subject: Re: bug in dominance algorithm
- From: Michael Matz <matzmich at cs dot tu-berlin dot de>
- Date: Fri, 15 Jun 2001 15:39:54 +0200 (MET DST)
- cc: Richard Henderson <rth at redhat dot com>, <gcc at gcc dot gnu dot org>
Hi,
On Fri, 15 Jun 2001, Jan Hubicka wrote:
>
> Do we need the old behaviour somewhere? I think it should be safe to change
Don't know. I guess not.
> it by default.
Yep.
> Only DCE uses the postdominance tree and it is broken on the moment.
The postdom _tree_ (i.e. the immediate postdom array) should already be
correct, and only the bitmap information about the whole set of postdoms
is screwed. At least if I wasn't that clever to also cripple the postdom
tree to be consistent ;-) I can't look at the code right now.
> > The interesting edge: A--e-->B
> > 1) B itself is already the noreturn block, for this case no postdominance
> > information can be used.
> It can, B is postdominator of B,
Yep. I was thinking about the immediate postdom. I usually have the tree
in mind, not the flat bitmaps, which are often useless (i.e. other
algorithms based on the dom tree can be made up), slow and big. I guess
if I'm touching dominance.c anyway, I provide a nicer interface for
dominance queries (like computing the bitmaps on demand only).
Ciao,
Michael.