cross-compiled floating point constants broken

[Stephen L Moshier <moshier at mediaone dot net>]

    > Mon Dec 14 16:23:27 1998  Jim Wilson  <wilson@cygnus.com>
    >
    >    * real.c (endian): Disable last change unless
    >    HOST_BITS_PER_WIDE_INT is greater than 32.
    >
    > Thu Dec 10 16:02:06 1998  Jim Wilson  <wilson@cygnus.com>
    >    ...
    >    * real.c (endian): Sign-extend 32 bit output values on a 64 bit
    >    host.

The above change has broken cross compilations from 64-bit to 32-bit
machines.  It generates assembly language output that cannot be
assembled by many, if not most, 32-bit assemblers.  For example,
the source line

long double y = -1.3L;

compiles to

_y:
        .long 0xffffffffbfff0000,0xffffffffa6666666,0x66666666

using a current egcs alpha-to-68k cross compiler.  The 32-bit native 68k
assembler chokes on values with so many hex digits as
0xffffffffbfff0000 .

I believe that gen_lowpart_common is the only place in gcc that pays
any attention to the output from `endian', other than just to print it.
If you want those integer values sign extended, look for
REAL_VALUE_FROM_TARGET_SINGLE, REAL_VALUE_FROM_TARGET_DOUBLE in
emit-rtl.c and see if doing the sign extensions there doesn't solve
your problem.  If you still need something that does the sign
extension, we should write a new function that doesn't get used by the
printf-sensitive ASM_OUTPUT_DOUBLE etc. tm.h macros.