[PATCH] Fix flow-insensitive alias computation

[Daniel Berlin]

On Fri, Apr 18, 2008 at 6:35 PM, Richard Guenther <rguenther@suse.de> wrote:
> On Fri, 18 Apr 2008, Richard Guenther wrote:
>
>  > On Fri, 18 Apr 2008, Daniel Berlin wrote:
>  >
>
> > > >  So I believe the following should be in principle correct modulo
>  > > >  latent bugs elsewhere...
>  > >
>  > > This looks fine to me, but be careful.
>  > > As you've discovered, this part of the code is very very fragile.
>  >
>  > It actually works.  I am thinking to force PTA off and see if
>  > anything breaks that way (that basically forces the operand scanner
>  > to fall back to the pointers SMT all the time).
>
>  So trying the following independent of my patches
>
>  Index: gcc/tree-ssa-alias.c
>  ===================================================================
>  --- gcc/tree-ssa-alias.c        (revision 134452)
>  +++ gcc/tree-ssa-alias.c        (working copy)
>  @@ -2345,10 +2345,11 @@ compute_flow_sensitive_aliasing (struct
>
>
>    for (i = 0; VEC_iterate (tree, ai->processed_ptrs, i, ptr); i++)
>      {
>  -      if (!find_what_p_points_to (ptr))
>  +      if (1 || !find_what_p_points_to (ptr))
>         set_pt_anything (ptr);
>      }
>
>  +  return;
>    create_name_tags ();
>
>
>    for (i = 0; VEC_iterate (tree, ai->processed_ptrs, i, ptr); i++)
>
>  shows that call-clobbering is broken (gcc.c-torture/execute/pr28778.c
>  will fail) - I will have a look here.
>
>  It also shows that (with my patches) char * pointers are special, so
>  may_alias_p needs
>
>  @@ -2868,6 +2869,16 @@ may_alias_p (tree ptr, alias_set_type me
>      {
>        alias_stats.tbaa_queries++;
>
>  +      /* If the pointed to memory has alias set zero or the pointer
>  +        is ref-all, the MEM can alias VAR.  */
>  +      if (mem_alias_set == 0
>  +         || PTR_IS_REF_ALL (ptr))
>  +       {
>  +         alias_stats.alias_mayalias++;
>  +         alias_stats.tbaa_resolved++;
>  +         return true;
>  +       }
>

I can believe the PTR_IS_REF_ALL part, but alias_sets_conflict_p will
return the right answer for mem_alias_set == 0 already, and then we
only perform further disambiguation iff mem_alias_set != 0.

If you have to use the above with mem_alias_set == 0 to get the code
to work, something else is broken.