Re: [PATCH] Make alias_sets_conflict_p less conservative

[Richard Guenther <rguenther@suse.de>]

On Wed, 5 Mar 2008, Richard Kenner wrote:

> > But in the case we are trying to fix, the current alias behavior would
> > make
> > 
> >   s.size = 0;
> > 
> > conflict with the store and load from p->d, which is incorrect.
> 
> I don't think so.  In that case, we look at the alias set of s.size, not of s.

Ok, so the problem (of PR27799) is that our tree aliasing machinery
creates SMTs (structure-memory-tags) that represent all structures
with a given type.  TBAA relations here are computed by asking wheter
for example a pointer can access struct X which is done by asking for
alias_sets_conflict_p (get_alias_set (*ptr), get_alias_set (struct-type))
which in this case returns true (as struc-type has a char member).
[this all happens in may_alias_p]

Later during operand scanning we (could) ask again, but only the
same question -- does the access p->x alias a SMT?  And the answer
would be the same.  [access_can_touch_variable]

So we are lost here, as we obviously cannot ask the question this way
with the current scheme. (?)

To recap:

struct A { int i; char c; };
struct B { float f; };

the question we ask is, can you access any struct A through a pointer
to struct B?  And, can ((struct B *)p)->f touch any struct A?

For this sort of questions (with char having alias set zero and the
subsetting work as is), we can only answer yes.  With my patch
we can start to say no here.

Richard.