[java/rfc] don't call make_decl_rtl from the front end

[Andrew Haley]

Richard Henderson writes:
 > This is patch 2 of 3 trying to fix the CNT plt problem.  The main
 > goal here is to avoid creating rtl too early, and indirectly forcing 
 > DECL_ASSEMBLER_NAME to be set too early.  I very much want that to
 > be done after java_mark_class_local, which happens after all parsing.
 > 
 > A secondary cleanliness goal says that a front end shouldn't have to
 > concern itself with such low-level details as rtl.

Yes, indeed.

 > One thing for review is what build_class_ref played with DECL_SIZE?
 > It would seem that would have been done by layout_decl, which is 
 > called by build_decl when presented with a complete type.

Right.  the more of this front-end messing with back-end details we
lose, the better.

I'm sure there was a good reason for all this once upon a time, but if
it's no longer necessary that's fine.

 > 	* builtins.c (define_builtin): Don't call make_decl_rtl.
 > 	* constants.c (build_constant_data_ref): Likewise.
 > 	* class.c (build_utf8_ref): Likewise.
 > 	(build_fieldref_cache_entry, build_static_field_ref): Likewise.
 > 	(get_dispatch_table, layout_class_method): Likewise.
 > 	(build_class_ref): Likewise. Don't set DECL_SIZE or DECL_SIZE_UNIT
 > 	by hand.
 > 	(make_local_function_alias): Don't SET_DECL_ASSEMBLER_NAME.
 > 	(make_method_value): Use METHOD_ABSTRACT instead of DECL_RTL_SET_P
 > 	to determine if we need a non-zero address.
 > 	* decl.c (builtin_function): Don't call make_decl_rtl.
 > 	(give_name_to_locals): Don't SET_DECL_ASSEMBLER_NAME.
 > 	* expr.c (build_known_method_ref): Don't call make_decl_rtl.
 > 	* resource.c (compile_resource_data): Likewise.
 > 	* parse.y (resolve_field_access): Re-word comment to avoid
 > 	building DECL_RTL.
 > 

Thanks,
Andrew.