Re: Do less generous pointer globbing in alias.c

["H.J. Lu" <hjl dot tools at gmail dot com>]

On Thu, May 28, 2015 at 1:12 PM, Jan Hubicka <hubicka@ucw.cz> wrote:
> hello,
> only providing you the testcase why I need transitive closure of "contains
> pointer" via the extra child I noticed that there is extra symmetry to handle:
>
>      struct a {void *ptr;}
>      char **ptr = (char **)&a.ptr;
>      ptr = ...
>
> This one doesn't really fly with my extra subset code, because ptr is not
> universal pointer, but struct a contains one and thus should conflict with
> every pointer.  Adding every pointer as subset of every structure with
> universal pointer is impractical (childs of those structures would be appearing
> as new pointer types get alias sets) and thus indeed it is better to handle it
> same way as alias set 0 - by a special case in alias_set_subset_of
> and alias_sets_conflict_p.
>
> So I added the second flag - has_pointer that is transitive closure of
> is_pointer and added the special case to alias_sets_conflict_p instead of
> adding the extra subset relation into the DAG.
>
> I also added statistics and made changes you suggested (making child
> hash to be possibly NULL and clenaing up alias set conflict construction)
>
> I also constructed a testcase that covers all the new code paths.
>
> The patch bootstrapped/regtested ppc64-linux.  I am not bound to teaching
> next week, so if I hear no negative comments, I will schedule commiting the
> patch for weekend to deal with possible fallout.
>
> There are few cleanups possible incrementally - i.e. the hash set seems
> irrationaly large for average type, we could avoid some pointer travelling
> overhead and we could also do better at alias_sets_must_conflict_p.
>
> Honza
>
>         * alias.c (alias_set_entry_d): Add is_pointer and has_pointer.
>         (alias_stats): Add num_universal.
>         (alias_set_subset_of): Special case pointers; be ready for NULL
>         children.
>         (alias_sets_conflict_p): Special case pointers; be ready for NULL
>         children.
>         (init_alias_set_entry): Break out from ...
>         (record_alias_subset): ... here; propagate new fields;
>         allocate children only when really needed.
>         (get_alias_set): Do less generous pointer globbing.
>         (dump_alias_stats_in_alias_c): Update statistics.
>         * gcc.dg/alias-8.c: Do not xfail.
>         * gcc.dg/pr62167.c: Prevent FRE.
>         * gcc.dg/alias-14.c: New testcase.
==========================================
> --- testsuite/gcc.dg/alias-8.c  (revision 223772)
> +++ testsuite/gcc.dg/alias-8.c  (working copy)
> @@ -8,5 +8,5 @@ struct s {
>  void
>  func(struct s *ptr)
>  {
> -  *(void **)&ptr->p = 0; /* { dg-warning "type-punned pointer" "" { xfail *-*-* } } */
> +  *(void **)&ptr->p = 0; /* { dg-warning "type-punned pointer" "" { } } */
>  }

This caused:

ERROR: gcc.dg/alias-8.c: syntax error in target selector "" for "
dg-warning 11 "type-punned pointer" "" { } "

I checked in this fix.

H.J.
---
Index: ChangeLog
===================================================================
--- ChangeLog (revision 223886)
+++ ChangeLog (working copy)
@@ -1,3 +1,7 @@
+2015-05-30  H.J. Lu  <hongjiu.lu@intel.com>
+
+ * gcc.dg/alias-8.c: Fix dg-warning.
+
 2015-05-30  Jan Hubicka  <hubicka@ucw.cz>

  * gcc.dg/alias-8.c: Do not xfail.
Index: gcc.dg/alias-8.c
===================================================================
--- gcc.dg/alias-8.c (revision 223886)
+++ gcc.dg/alias-8.c (working copy)
@@ -8,5 +8,5 @@
 void
 func(struct s *ptr)
 {
-  *(void **)&ptr->p = 0; /* { dg-warning "type-punned pointer" "" { } } */
+  *(void **)&ptr->p = 0; /* { dg-warning "type-punned pointer" } */
 }