[java/rfc] cleanup emit_register_classes and friends

[Andrew Haley]

Richard Henderson writes:
 > This is patch 1 of 3 trying to fix the CNI plt problem currently
 > causing lots of Java failures.
 > 
 > There are two things in here that seemed unnecessary that warrent review.
 > 
 > First, the decl duplication in register_class.  Why?

I have no idea.  This has been in every version of gcj since its
creation.

 > Duplicating decls is always extremely risky buisiness.  In this
 > case I'm particularly interested in removing the DECL_RTL
 > invocation here, since I want all DECL_RTL creation to be delayed
 > until after java_mark_class_local has run.

OK.

 > Second, the call to mark_decl_referenced in emit_register_classes
 > would seem to be redundant with the one in output_addr_const, which
 > is eventually called by assemble_integer.  Here, for the sake of 
 > eventually not referencing rtl anywhere within the java front end,
 > I've used an output routine that accepts trees instead of playing
 > with DECL_RTL ourselves.

That looks much better.

 > 
 > Tested alone on i686-linux, together with other patches on ia64 
 > and alpha linux.
 > 
 > Ok?

Yes.

Andrew.