http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54224
--- Comment #5 from Tobias Burnus <burnus at gcc dot gnu.org> 2012-08-16
09:56:51 UTC ---
> c) For the program in comment 0, one enters rtx_for_function_call [...]
> However, that function is called *AFTER* check_global_declaration_1 and comes
> thus too late.
check_global_declaration_1 is called write_global_declarations via compile_file
via do_compile
While rtx_for_function_call is later also called via write_global_declarations
-> finalize_compilation_unit -> output_in_order -> expand_function ->
execute_pass_list -> execute_one_pass -> ... -> expand_call_stmt -> ...
In C, one has the same order - except that TREE_USED is already set. My
impression is that in C it gets set when the DECL is 'created' in the caller.
* * *
Hence, I think one should think of handling it on Fortran FE level by setting
USE_TREE.
For procedures, I think we need to cover procedure calls, proc-pointer
assignments and proc-pointer assignments via initializers and default
initializers. On tree level, we could either plug into
gfc_trans_pointer_assignment and conv_function_val - or make some use of
attr.referenced (though, the latter seems to be also set for uncalled
functions).
Not as important as we do not currently get a bogus warning, but still nice to
have: A warning for PRIVATE module variables. For
module m
private
integer, save :: var ! = 5
end module m
we always set TREE_USED in gfc_finish_var_decl. One could check whether simply
not doing so for private variables is sufficient to get a warning, but I am not
sure whether we won't run into the problem of getting the diagnostic in
check_global_declaration_1
sym->referenced is always set: It is set before gfc_create_module_variable and
it is unconditionally set in that function. (EQUIVALENCE with PRIVATE/PUBLIC
has to be handled as well.)
