[microblaze-uclinux] Bug in software floating-point library

[Graeme Smecher <graeme.smecher@xxxxxxxxxxxxxx>]

Hi,

The softfloat implementation embedded within the Petalinux MicroBlaze 
toolchain produces incorrect answers. The following test code:

   int main(void) {
     double x;
     x = 123456789.0L; printf("__fixdfsi(%f) = %li\n", x, __fixdfsi(x));
     x = 1234567890.0L; printf("__fixdfsi(%f) = %li\n", x, __fixdfsi(x));
     x = 12345678900.0L; printf("__fixdfsi(%f) = %li\n", x, __fixdfsi(x));
     return 0;
   }


...returns:

   # ./testcase
   __fixdfsi(123456789.000000) = 123456789
   __fixdfsi(1234567890.000000) = 0
   __fixdfsi(12345678900.000000) = 2147483647


The first and third cases appear to be correct. (2147483647 corresponds 
to INT_MAX, and the third argument is out-of-range.) However, the second 
case is unexpected. It produces erroneous results in an embedded Python:

   # python 
   Could not find platform dependent libraries <exec_prefix>
   Consider setting $PYTHONHOME to <prefix>[:<exec_prefix>]
   Python 2.5.1 (r251:54863, Jan 28 2008, 09:24:22)
   [GCC 3.4.1 ( PetaLinux 0.20 Build -rc1 050607 )] on linux2
   Type "help", "copyright", "credits" or "license" for more information.
    >>> int(1234567890.0)
   0 

From the toolchain sources, the offending code lives in 
'srcs/newlib-3.4/libgloss/microblaze/floatlib/fixdfsi.c'. I've tried it 
on an x86 with the same results.

The Python expression "long(1234567890.0)" works fine as an alternative, 
but having to do so makes me nervous. (I'm actually quite surprised to 
run into this bug, since this softfloat implementation seems to be 
widely used. Perhaps it's not guaranteed to work with large arguments.)

regards,
Graeme
___________________________
microblaze-uclinux mailing list
microblaze-uclinux@xxxxxxxxxxxxxx
Project Home Page : http://www.itee.uq.edu.au/~jwilliams/mblaze-uclinux
Mailing List Archive : http://www.itee.uq.edu.au/~listarch/microblaze-uclinux/