[Bug c/103310] null comparison with a weak symbol eliminated

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103310

Jason Merrill <jason at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1
                 CC|                            |jason at gcc dot gnu.org
   Last reconfirmed|                            |2021-11-18

--- Comment #2 from Jason Merrill <jason at gcc dot gnu.org> ---
The C++ front end rejects the testcase with "declared weak after used".  We set
refuse_visibility_changes under fold's tree_single_nonzero_warnv_p. 

Changing get_create to get in maybe_nonzero_address makes the C++ compiler
accept the testcase, and properly test whether alias is null.

This issue seems to go back to Honza's r5-3627, which changed symtab_node::get
to symtab_node::get_create in the code that became maybe_nonzero_address, so
that we decide early whether a particular function is weak or not.

This was done so that constant-evaluation could properly decide that the
address of a function is non-null.  But it's harmful when we do that for
speculative folding; we should only return a definitive answer, and set
refuse_visibility_changes, when a constant result is required.

It seems we need a way to tell fold that we really want a constant value, have
the C++ front end set that for manifestly-constant-evaluated expressions, and
only use get_create in that case.

But I also guess the issue that the C front end is both optimizing away the
test and not setting refuse_visibility_changes is a C front end issue, that
it's doing the optimization without involving fold.