byteswap.h behavior

[Collin Funk <collin.funk1@gmail.com>]

A few months ago there was a suggestion on emacs-devel to use
__builtin_bswap functions when available [1]. While I agree that GCC
can deal with the optimizations properly, I noted an important
difference between the macros in byteswap.h.in and inline functions
provided by glibc.

Using this test program to compare the real glibc header to a macro
copy-pasted from the replacement header:

======================================
#define _GNU_SOURCE 1
#include <stdio.h>
#include <inttypes.h>
#include <byteswap.h>
#define bswap_32_macro(x) ((((x) & 0x000000FF) << 24) | \
                           (((x) & 0x0000FF00) << 8) |  \
                           (((x) & 0x00FF0000) >> 8) |  \
                           (((x) & 0xFF000000) >> 24))
int
main (void)
{
  uint32_t value = 0x12345678;
  uint32_t value_macro = 0x12345678;
  printf ("1. %#" PRIX32 "\n", bswap_32 (value++));
  printf ("2. %#" PRIX32 "\n", bswap_32_macro (value_macro++));
  printf ("3. %#" PRIX32 "\n", value);
  printf ("4. %#" PRIX32 "\n", value_macro);
  return 0;
}
======================================

We get the output:

    $ ./a.out 
    1. 0X78563412
    2. 0X78563412
    3. 0X12345679
    4. 0X1234567C

I would like to deal with this concern before I implement the
replacement for <endian.h>. I think the best decision is to use
'extern inline' to match the behavior of glibc. Any other opinions on
this?

Also if we can agree upon making sure these are defined as functions,
what is the proper way to test it in a configure script? My instinct
tells me that assigning a function pointer to bswap_16, etc. would
fail if they are macros but I am not sure the "standard" way of
performing that check.

Collin

[1] https://lists.gnu.org/archive/html/emacs-devel/2024-03/msg00736.html