[Bug middle-end/50640] [4.7 Regression] FAIL: gfortran.dg/select_type_12.f03 -O (internal compiler error)

["rguenth at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50640

--- Comment #6 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-10-07 18:53:44 UTC ---
(In reply to comment #3)
> Hmm, this is not as trivial as PR50638.  fortran frontend generates this
> static variable local to MAIN:
> 
> static struct __vtype_MAIN___T1 __vtab_MAIN___T1 = {._hash=41707971, ._size=4,
> ._extends=0B, ._def_init=&__def_init_MAIN___T1, ._copy=__copy_MAIN___T1};
> 
> (the reference to &__def_init_MAIN___T1 will later cause the segfault).
> __vtab_MAIN___T1  will be referenced from __vtab_MAIN___T2 (also static local
> in MAIN).  But __vtab_MAIN___T2 will _not_ be referenced from MAIN.
> 
> Instead references to __vtab_MAIN___T2 and __vtab_MAIN___T1 are emitted from
> function 'fun'.  So far so good.  The problem lies in walking DECL_INITIAL
> for these statics.  add_referenced_var will only walk it if
> DECL_CONTEXT==current_function_decl (sensible).  Nobody calls
> add_rerefenced_var(__vtab_MAIN___T1) for function MAIN (sensible, as there
> are no references).
> 
> We will call add_rerefenced_var(__vtab_MAIN___T1) from function 'fun' (also
> sane), which then won't walk DECL_INITIAL, i.e. nobody makes
> __def_init_MAIN___T1 referenced.
> 
> The the code in tree-ssa-live comes and walks initializers of local variables
> that are used.  It does so for function MAIN and variable __vtab_MAIN___T
> (which is in local_decls, just not in referenced_vars), because it is marked
> used (it's a non-local decl, therefore var_ann is still active from the calls
> from function 'fun').  That one then walks the initializer and now notices
> that nothing in it is marked referenced.
> 
> The non-triviality here lies in deciding who's wrong: add_referenced_var
> (i.e. should it always mark referenced all initializers, even though they
> might come from other function contexts), the tree-ssa-live code (should
> it perhaps not look into initializers or at least expect them not necessarily
> marked), or in the fortran frontend (should it perhaps call add_referenced_var
> also for the local static in MAIN).

Huh, I don't see why tree-ssa-live would need to walk variable initializers
(which generally don't exist unless for static storage duration vars).