bug#11310: wrong conversions of ieee754 from/to foreign endianness

bug#11310: wrong conversions of ieee754 from/to foreign endianness

[Klaus Stehle]

Hi,

There is a problem in bytevectors.c - conversions of float and double
to the foreign endianess don't work.


Here is the scheme example:
---------------------------
(use-modules (rnrs bytevectors))

(define bv (make-bytevector 4))

(bytevector-ieee-single-set! bv 0 1.0 (endianness little))
(display bv) (newline)

=> #vu8(0 0 128 63)                     correct!

(bytevector-ieee-single-set! bv 0 1.0 (endianness big))
(display bv) (newline)

=> #vu8(254 0 0 0)                      oops!

The results come from a little-endian machine and therefore the
big-endian result is wrong. We expect #vu8(63 128 0 0).



In bytevectors.c there is the following code:

#ifdef WORDS_BIGENDIAN
  /* Assuming little endian for both byte and word order.  */
  target->little_endian.negative = src.big_endian.negative;
  target->little_endian.exponent = src.big_endian.exponent;
  target->little_endian.mantissa = src.big_endian.mantissa;
#else
  target->big_endian.negative = src.little_endian.negative;
  target->big_endian.exponent = src.little_endian.exponent;
  target->big_endian.mantissa = src.little_endian.mantissa;
#endif

Using the 'FOREIGN' structure of the scm_ieee754_float union
does NOT work, because in this case the structure assumes the
wrong  *BIT* order.



example in C shows better (to run on a little endian machine):
--------------------------------------------------------------
union scm_ieee754_float target;
target.f = 0.0;
target.litte_endian.negative = 1;

=> 00 00 00 80                  correct!

target.f = 0.0;
target.big_endian.negative = 1;

=> 01 00 00 00                  oops!


Maybe this effect depends on the machine and the compiler.
I am using an amd64 with gcc.

The bug should be easy to fix, instead of using the ieee754 unions
and their weird structures, exchanging whole bytes should suffice
and this probably may be faster.

Thanks
Klaus

bug#11310: wrong conversions of ieee754 from/to foreign endianness

[Ludovic Courtès]

Hi Klaus,

Thanks for the report!  Fixed in
398446c7428b3d98d168fcc3ff170829d3e09f9a, and double-checked on HPPA.

Ludo’.