Re: Optimize three-valued comparison between integers

[Bruno Haible]

Paul Eggert wrote:
> On 7/23/20 3:15 PM, Bruno Haible wrote:
> > +  AH_VERBATIM([micro_optimizations],
> > +[/* _GL_CMP (n1, n2) performs a three-valued comparison on n1 vs. n2.
> > +   It returns
> > +     1  if n1 > n2
> > +     0  if n1 == n2
> > +     -1 if n1 < n2
> 
> While looking into using this in Emacs I noticed that _GL_CMP (A, B) 
> currently 
> returns 0 if either A or B is a NaN.

Oh, you are applying it to floating-point numbers? Indeed, I can see that it
produces jump-free code also for 'double' arguments.

> I suggest that we change the documentation 
> comment for _GL_CMP to say specifically that the result is indeterminate if 
> either argument is a NaN, to make it clearer that callers should not expect 
> the 
> result to be 0 in this case.

I agree. Done through the patch below.

> Another possibility would be to document that the result is 0 in this case, 
> but 
> I don't see how that would be useful, as macros like _GL_CMP don't work if 
> arguments are not taken from a total order.

Right. NaNs need special treatment in the vast majority of the cases.

Bruno


2020-07-31  Bruno Haible  <bruno@clisp.org>

        _GL_CMP: Improve documentation.
        Reported by Paul Eggert in
        <https://lists.gnu.org/archive/html/bug-gnulib/2020-07/msg00188.html>.
        * m4/gnulib-common.m4 (gl_COMMON_BODY): Clarify what arguments can be
        passed.

diff --git a/m4/gnulib-common.m4 b/m4/gnulib-common.m4
index 57f3a78..50acc0a 100644
--- a/m4/gnulib-common.m4
+++ b/m4/gnulib-common.m4
@@ -1,4 +1,4 @@
-# gnulib-common.m4 serial 52
+# gnulib-common.m4 serial 53
 dnl Copyright (C) 2007-2020 Free Software Foundation, Inc.
 dnl This file is free software; the Free Software Foundation
 dnl gives unlimited permission to copy and/or distribute it,
@@ -300,7 +300,9 @@ AC_DEFUN([gl_COMMON_BODY], [
 #define _GL_ASYNC_SAFE
 ])
   AH_VERBATIM([micro_optimizations],
-[/* _GL_CMP (n1, n2) performs a three-valued comparison on n1 vs. n2.
+[/* _GL_CMP (n1, n2) performs a three-valued comparison on n1 vs. n2, where
+   n1 and n2 are expressions without side effects, that evaluate to real
+   numbers (excluding NaN).
    It returns
      1  if n1 > n2
      0  if n1 == n2