[Bug c++/78373] [7 Regression] error: constant not recomputed when ADDR_EXPR changed

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78373

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
So, what happens is that we have:
&self->D.2295
and on that ADDR_EXPR we have TREE_CONSTANT set, because the field offset is
constant, and the base is INDIRECT_REF/MEM_REF with the self VAR_DECL as
operand, and that VAR_DECL has TREE_CONSTANT set.
But later on we rewrite it into SSA form and it becomes &self_2->D.2295
instead,
and the MEM_REF operand in that case is SSA_NAME.  And that SSA_NAME self_2
doesn't have TREE_CONSTANT set on it.

So the questions are:
1) is it ok to have the TREE_CONSTANT on such VAR_DECLs (in the FEs, and/or in
the middle-end; e.g. if it is ok to have it in the FEs and not in ME, we need
to clear it somewhere and recompute the ADDR invariant stuff)
2) shall SSA_NAME for TREE_CONSTANT VAR_DECL have TREE_CONSTANT flag set or
not?
3) shall recompute_tree_invariant_for_addr_expr look through SSA_NAMEs to their
SSA_NAME_VAR if non-NULL for flags updating?