Re: [ tree-ssa] PRE problem

[Daniel Berlin <dberlin at dberlin dot org>]

On Feb 20, 2004, at 7:49 PM, law@redhat.com wrote:

> In message <1077321774.3219.22.camel@localhost.localdomain>, Diego 
> Novillo writ
> es:
> > On Fri, 2004-02-20 at 18:48, Daniel Berlin wrote:
> >
> > > Um, Diego, when do you think that code was written?
> > > Last week?
> > > It hasn't been fundamentally touched in ages (I believe someone 
> > > touched
> > > the generate_vops part when we moved to having VUSE_OPS_PTR).
> > Dan, what is happening here is that we are only *now* discovering that
> > the changes made to the copy propagation of pointers were affecting
> > PRE.
> True.  Basically I suspect this could be summed up as "we weren't aware
> that PRE did this kind of thing internally.

It says so in the comment :)

/* Generate EXPR as it would look in basic block PRED (using the phi in
   block BB).  We do this by replacing the variables with the phi
   argument definitions for block J if they are defined by a phi in
   block BB.  */

Though admittedly, you'd have to think about it for a second to realize 
that this is propagating values.

> Though this does bring up a documentation issue -- I don't think we
> discuss these issues anywhere in the documentation (source or 
> external).
>
> > When I introduced propagate_copy and may_propagate_copy, I simply 
> > didn't
> > run into this problem.  The fix is simple, just teach PRE to use the
> > value propagation API.
> Given that we've got two fixes for these issues, let's get one tested
> and installed and get on with our lives.

Neither fix is quite correct (and in fact, both would introduce silent 
but deadly bugs :P), which is why i'm testing the third fix.
--Dan