https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121670
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jakub at gcc dot gnu.org,
| |jason at gcc dot gnu.org
--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
The difference is in store_init_value doing
927 const_init = (reduced_constant_expression_p (value)
928 || error_operand_p (value));
929 DECL_INITIALIZED_BY_CONSTANT_EXPRESSION_P (decl) = const_init;
Now, the initializers aren't valid constant expressions due to the
reinterpret_cast, which is why we fail to simplify those early.
reduced_constant_expression_p under the hood for scalars calls the middle-end
initializer_constant_valid_p function, which as an extension will handle some
casts, POINTER_PLUS_EXPR etc.
In the q0 case the POINTER_PLUS_EXPR second argument is INTEGER_CST 4, so
initializer_constant_valid_p is happy with it and we set const_init to true.
In the q1 case the POINTER_PLUS_EXPR second argument is SIZEOF_EXPR with
NOP_EXPR to int argument (that is the representation of sizeof on types; not
folded early for the benefit of -Wsizeof-mem* warnings etc.). But SIZEOF_EXPR
is a FE tree that the middle-end has no idea about, so
reduced_constant_expression_p returns false.
Given that neither of these initializers are valid constant expressions, I
wonder if we shouldn't test TREE_CONSTANT (value) &&
reduced_constant_expression_p (value) instead of just
reduced_constant_expression_p (value).
BTW, I think we eventually in cp_fully_fold_init will fold the SIZEOF_EXPR, but
pedantically it is not initialized by constant expression. And neither is q0
or p.
