patch for GCC floating-point nonsense bug

[Ulrich Drepper]

Paul Eggert <eggert@twinsun.com> writes:

> Here's a proposed patch for this GCC bug.  While I was fixing it
> I found what appear to be other instances of the bug; this patch fixes
> all the instances that I found.

Are you sure this is correct?  I compiled egcs using this patch and I
get bogus results.  The bit patterns generated for the constants are
somewhat random.  E.g., here are the results for the function I sent
in the last mail for a return value of `long double' (again on ix86):

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
        .align 4
.LC2:
        .long 0xbe000000,0xf76080de,0xbbc8
        .align 4
.LC3:
        .long 0x387c,0xeb7cbfff,0x403e
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Even when using `double' the -0.0 constant is wrong (0.0 is ok).

> + #define REAL_VALUES_IDENTICAL(x, y) \
> +   (!bcmp ((char *) &(x), (char *) &(y), sizeof (REAL_VALUE_TYPE)))

This makes me wonder since there are no separate constants for float
and long double.  I.e., a bitwise compare will fail.

I saw that the patch removed some similar bcmp call but this probably
is also wrong.

-- Uli
---------------.      drepper at gnu.org  ,-.   Rubensstrasse 5
Ulrich Drepper  \    ,-------------------'   \  76149 Karlsruhe/Germany
Cygnus Solutions `--' drepper at cygnus.com   `------------------------