Unicode mangling (was Re: [PATCH] Java: New C++ ABI compatibility changes.)

[Jason Merrill]

It looks like you're still using the same scheme for mangling Unicode
strings.  I'd like to reexamine that, since we're stabilizing the ABI.

First, we need to remember that C99 and C++ (will) also need this
functionality.  The frontend work to support extended characters in
identifiers remains to be done, of course.

As far as I can tell, the current scheme affects individual identifiers.
If it contains extended characters, you prepend 'U' to the length and
replace each extended character with _NNNN (the 16-bit hex encoding of the
UCS2 value).  This currently has several flaws:

1) It doesn't allow for C-like symbols, which have no length specifier.
   This could be fixed by defining some encoding starting with, say, '_U'.
2) It doesn't accommodate 32-bit extended characters in C++/C99
   (\UNNNNNNNN).  This could be fixed by escaping them with, say, '_L'.
3) _NNNN is a valid component of an identifier, complicating the
   demangler intelligence.  This could be fixed by also escaping the '_'
   character in affected names.  Hmm...it looks like you intend to do
   so in unicode_mangling_length, but don't actually do so in
   append_unicode_mangled_name.  We could also just use '__'.

With these fixes, I think the current scheme is OK.  But for targets with
8-bit clean binutils, I think it makes a lot of sense to just use the UTF8
encoding in the symbol.

Thoughts?

Jason