https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92412
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
The DECL_EXTERNAL is what should make the difference but IIRC a errno.h
with a tentative definition would be conforming, no? And that woudln't
be DECL_EXTERNAL ...
TREE_READONLY should be handled more generally,
/* If the reference is based on a decl that is not aliased the call
cannot possibly clobber it. */
if (DECL_P (base)
&& !may_be_aliased (base)
/* But local non-readonly statics can be modified through recursion
or the call may implement a threading barrier which we must
treat as may-def. */
&& (TREE_READONLY (base)
|| !is_global_var (base)))
return false;
gives a hint (but doesn't apply here due to the var being address-taken).
I think we could split out
if (DECL_P (base)
&& TREE_READONLY (base))
return false;
? I know I'm usually overly cautionous with const vs non-const, but on
decls it should work. OTOH...
int var;
const int const_var __attribute__((alias("var")));
is likely a valid way to "document" that some parts cannot modify 'var'
but for
int foo()
{
int tem = const_var;
bar ();
return const_var;
}
we cannot elide the second load because bar() might modify it via a store
to 'var'.
Likewise interposing a const int def with a non-const int def is probably
OK.
Thus we'd need to ask the varpool if 'base' is prevailing, look at the
ultimate alias target and only take that vars const-ness into account...
(ick).
