Re: support for bitwise comparison of floats

[Bruno Haible]

Paul Eggert wrote:
> It is referring to the behavior of the VAX floating point unit, where the
> hardware representation that one thinks would represent -0.0 behaves
> sort of like a NaN.  To avoid this problem, on a VAX copysign(0, -1)
> returns +0.0, not the usual -0.0.  That is why the C and POSIX
> standards are worded the way they are.

Thanks for explaining. I have submitted the attached change request to the
Linux man pages maintainer.

> I suggested copysign instead of the signbit macro because copysign has
> been around a lot longer -- it was in the 1985 IEEE floating point
> standard -- and is optimized better.  But if you want to play it safer
> (and slower :-) you can use signbit.

Ah, there is also signbit()! Thanks for mentioning it. It's optimized better
(here with gcc-4.3 on i386):

-----------------------------------------------------------------------
#include <math.h>
int using_signbit (long double x) { return signbit (x); }
int using_copysign (long double x) { return copysign (1.0L, x) < 0; }
-----------------------------------------------------------------------

is compiled into

using_signbit:
        movl    12(%esp), %eax
        andl    $32768, %eax
        ret

using_copysign:
        subl    $12, %esp
        fldt    16(%esp)
        fstpl   (%esp)
        movl    4(%esp), %edx
        fld1
        testl   %edx, %edx
        jns     .L2
        fstp    %st(0)
        fld1
        fchs
.L2:
        fldz
        fucompp
        fnstsw  %ax
        testb   $69, %ah
        sete    %al
        addl    $12, %esp
        movzbl  %al, %eax
        ret


--- man-pages-2.43/man3/copysign.3.bak  2005-10-19 09:07:02.000000000 +0200
+++ man-pages-2.43/man3/copysign.3      2007-04-01 01:02:21.000000000 +0200
@@ -26,7 +26,8 @@
 .\"     386BSD man pages
 .\" Modified 1993-07-24 by Rik Faith (address@hidden)
 .\" Modified 2002-08-10 by Walter Harms (address@hidden)
-.TH COPYSIGN 3  2002-08-10 "GNU" "Linux Programmer's Manual"
+.\" Modified 2007-03-31 by Bruno Haible <address@hidden>
+.TH COPYSIGN 3  2007-03-31 "GNU" "Linux Programmer's Manual"
 .SH NAME
 copysign, copysignf, copysignl \- copy sign of a number
 .SH SYNOPSIS
@@ -46,7 +47,8 @@
 that of \fIx\fP, but whose sign matches that of \fIy\fP.
 If \fIx\fP is a NaN, then a NaN with the sign of \fIy\fP is returned.
 .SH NOTES
-The \fBcopysign\fP() functions may treat a negative zero as positive.
+On architectures where the floating-point formats are not IEEE 754 compliant,
+the \fBcopysign\fP() functions may treat a negative zero as positive.
 .SH "CONFORMING TO"
 C99, 4.3BSD.
 This function is defined in IEC 559 (and the appendix with