http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55935
--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> 2013-01-11
13:43:50 UTC ---
(In reply to comment #7)
> Created attachment 29142 [details]
> gcc48-pr55935.patch
>
> Untested fix. Although the FE perhaps could unshare_expr_without_location so
> that ADDR_EXPRs in the ctor don't have location, IMHO gimple-fold.c still has
> to at least unshare_expr those expressions it copies from the constructors,
> otherwise we'll end up with invalid sharing of ADDR_EXPRs etc. between
> different functions (or within the same function).
>
> This can be reproduced even with C:
> void foo (void);
> struct S { int i; void (*fn) (); };
> const struct S s = { 5, foo };
> void *fn1 (int x) { if (x < 0) return s.fn; if (x) return 0; return s.fn; }
> void *fn2 (int x) { if (x < 0) return s.fn; if (x) return 0; return s.fn; }
> void *fn3 (void) { return s.fn; }
> void *fn4 (void) { return s.fn; }
> at -O2, in *.copyrename1 pass all 6 ADDR_EXPRs in the IL are still shared.
> ccp1 for whatever reason unshares them all (surprisingly).
We allow those to pass verification in verify_node_sharing because:
/* Return true when the T can be shared. */
bool
tree_node_can_be_shared (tree t)
{
if (IS_TYPE_OR_DECL_P (t)
|| TREE_CODE (t) == SSA_NAME
|| is_gimple_min_invariant (t)
|| t == error_mark_node
|| TREE_CODE (t) == IDENTIFIER_NODE)
return true;
is_gimple_min_invariant is bogus here IMHO. But I guess "fixing" that
might have quite some fallout ...
Btw, get_symbol_constant_value also unshares the result from
canonicalize_constructor_val (if it's is_gimple_min_invariant).
I suppose we should try fixing tree_node_can_be_shared ...
