Re: Porting libgcj...

[scottb <scottb@netwinder.org>]

> > <scottb@netwinder.org> writes:
> 
> I'm trying to port libgcj to run on ArmLinux on a NetWinder.
> 
> It seems that when generating this file, the value for
> MAX_VALUE is interpreted as a NAN by printf..  I have been
> looking for the file generation code but have been unable to
> find it.  Could someone point me in the correct direction.

Tom Tromey wrote:
> 
> Most of the headers are generated from class files by a program named
> "gcjh".  It is part of gcc.  The source is gcc/java/gjavah.c.  You
> want to look at the function print_field_info().
> 

I think I have found the problem (at least for this case).  It is the
WORDS_TO_DOUBLE() macro in javaop.h.  The ARM has a funny floating point
memory layout for doubles.   It is the IEEE format in little endian with
the words swapped.  As an example:

MAX_DBL = 0xff 0xff 0xff 0xff 0xff 0xff 0xef 0x7f in little endian
MAX_DBL = 0xff 0xff 0xef 0x7f 0xff 0xff 0xff 0xff on the ARM

I updated my gcc-2.95 tree this morning and have a build going to test
this hypothesis.  I have modified WORDS_TO_DOUBLE to be:

static inline jdouble
WORDS_TO_DOUBLE(jword hi, jword lo)
{ union DWord wu;
#ifdef __arm__ 
  wu.l = WORDS_TO_LONG(lo,hi);
#else
  wu.l = WORDS_TO_LONG(hi,lo);
#endif
}

This is fine to test the code, but not my idea of a decent port.  There
doesn't seem to be any other way of introducing architecture dependent
code into the tree however.  Does anyone have any better suggestions?

Scott