http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48739
--- Comment #4 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-08-19
23:04:36 UTC ---
(In reply to comment #1)
> Simplified testcase:
>
> /* PR tree-optimization/48739 */
> /* { dg-do compile } */
> /* { dg-options "-O1 -ftree-parallelize-loops=2 -fno-tree-dominator-opts" } */
>
> extern int g;
> extern void bar (void);
>
> int
> foo (int x)
> {
> int a, b, *c = (int *) 0;
> for (a = 0; a < 10; ++a)
> {
> bar ();
> for (b = 0; b < 5; ++b)
> {
> x = 0;
> c = &x;
> g = 1;
> }
> }
> *c = x;
> for (x = 0; x != 10; x++)
> ;
> return g;
> }
>
> Started with http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=149206
> (before that the empty loop would be just removed).
>
> The bug seems to be that parloops pass requests TODO_rebuild_alias which
> changes x from TREE_ADDRESSABLE to a gimple var, but we are in loop closed SSA
> form at that point and nothing ensures that the optimized variable is put into
> loop closed SSA form.
The bug would be that TODO_rebuild_alias does a TODO_update_address_taken.
But really parloops should be re-architected to not require a
TODO_rebuild_alias ... (it isn't required anymore for correctness anyway)
