Re: support for bitwise comparison of floats

[Bruno Haible]

Eric Blake wrote:
> You can also assume IEEE rules, and compare against signed infinity:
> 1 / +0. => +infinity
> 1 / -0. => -infinity

Nice trick :-)

But here you depend on the IEEE rules for exceptions upon division by zero.
  - On Alpha processors, division by zero (and even overflow!) leads to a SIGFPE
    signal by default. And it requires the use of a system call to change
    the FP exceptions control mask (the <fenv.h> routines in glibc >= 2.1,
    the __setfpucw function in glibc 2.0). memcmp is certainly much cheaper.
  - On all processors, the caller could also have modified the FP exceptions
    control mask, through the facilities declared in <fenv.h> or <ieeefp.h>
    (FreeBSD, IRIX) or <fptrap.h> (AIX) or <fpxcp.h> (AIX). For this case,
    we would need to deal with all possible CPU / OS / kernel combinations...

Bruno