Re: [tree-ssa] Fix dominator bug

[Jan Hubicka <jh at suse dot cz>]

> Hello,
> 
> > > this patch adds a test whether we do not forget to remove the basic
> > > blocks from the dominance forest.  The Steven's patch (or at least
> > > its subset included in the patch, since I don't really care about
> > > updating postdominators) is of course neccesary to make it bootstrap.
> > 
> > Nice :)
> > What happens in the RTL world?  Do we want to remove the dominators
> > there too?  It would be nice to have the functions symmetric.
> 
> definitely.  They have unfortunately diverged somewhat (not only in
> handling of the dominators, but also in details like that split_block
> on rtl creates the new block after the original one, while the tree one
> before, etc. -- I have fixed a few of them on lno branch).

Can you please try to put such changes into tree-ssa branch itself so we
don't have too much divergence in the APIs?
This is rather important even when just cosmetic stuff. The unnecesary
divergence brings many headaches when you forget about it.
> 
> It is also neccessary to separate the ir specific stuff from the generic
> one (updating dominators, probabilities, etc.).  My idea is to change
> the macros defined in cfghooks.h into wrapper functions that would take
> care of this, and disallow any other code but these wrappers to access
> the low-level hooks that would be changed to do only the ir-specific stuff.
> 
> I think I have a program for this evening :-).

Hmm, this was my plan too, hehe :)
This is already done for verify_flow_info, so following same scheme
would be easy, if you want to take care of it, I would be happy to spend
evening by something else. (erm, fixing hammer and 3.4 bugs shall be my
priority right now)
> 
> > Also I
> > wonder why tree_remove_bb is not registered in cfghooks...
> 
> For long time it had an additional argument; and when it got removed,
> it was just somehow missed.

Also program for this evening for either of us :)

Honza
> 
> Zdenek