Patch: fix for PR java/6388

[Tom Tromey]

>>>>> "rth" == Richard Henderson <rth@redhat.com> writes:

rth> I don't think this is anywhere near correct for 64-bit H_W_I.
rth> In particular, check for long_suffix check just above is
rth> hopelessly 32-bit centric.

Yes, I see what you mean.

rth> Why don't you use int_fits_type_p instead?

I thought the existing lexer is written in a strange way to support
the `JC1_LITE' define, which is used by jv-scan and which can't use
trees.  (The code looks like it uses trees in both branches just below
this patch, but SET_LVAL_NODE_TYPE is empty for JC1_LITE -- don't ask
me why we bother invoking it.)

I'll write a new patch that fixes the sign extension and also gets rid
of the size dependency in lex.c without using trees.

Tom