Description:
------------
This is a follow-up on #40543 (http://bugs.php.net/bug.php?id=40543,
since
that bug is closed, I can't add comments). 
Please note : it's not identical to #4053 (other weird behaviors 
are demonstrated).

Iliaa,
Not sure why you suggest to use little endian or host conversions
routines,
but in my standpoint if you reverse twice a number's byte ordering
then you should get the original number back (assuming the number don't
overflows php's internals).

At least, that's the standard behavior for perl and python.

Beside, I can't see why php should handles those endianness conversions
differently on an i386 (32 bits) and on an x86_64 (64 bits), both
having the same byte order.

The following on a 64bit, little endian host :
x86_64$ uname -mprsv
OpenBSD 4.0 GENERIC#0 amd64 AMD Sempron(tm) Processor 3400+

x86_64$ perl -e 'print unpack("N", pack("N", 41445)) ."\n"'
41445

x86_64$ python
Python 2.4.3 (#1, Sep  6 2006, 20:33:08)
[GCC 3.3.5 (propolice)] on openbsd4
Type "help", "copyright", "credits" or "license" for more information.
>>> from struct import *
>>> unpack('>L', pack('>L', 41445))
(41445L,)

And, indeed :
#include <stdio.h>
#include <sys/types.h>
int main(void)
{
        u_int32_t x, y, z; /* 32 bits unsigned longs */
        x = 41445;
        /* conv host (little) to network (big endian) long : pack("N",
41445) */
        y = htonl(x);
        /* conv network (big endian) to host (little) long : unpack("N",
...) */
        z = ntohl(y);
        printf("Host : %li\nBig : %li\nHost : %li\n", x, y, z);
        return 0;
}

x86_64$ gcc conv.c -o conv ; ./conv
Host : 41445
Big : -442433536
Host : 41445

But still (PHP 5.2.2-dev (cli) (built: Feb 27 2007 22:10:11)) :
x86_64$ php -r 'print_r(unpack("N", pack("N", 41445)));'
Array
(
    [1] => -2147442203
)

While on a plain old x86 little endian host (PHP 4.4.0), we get
a different result :
i386_32$ uname -mprsv
OpenBSD 3.9 GENERIC#0 i386 Intel(R) Pentium(R) 4 CPU 1.70GHz
("GenuineIntel" 686-class)
i386_32$ php -r 'print_r(unpack("N", pack("N", 41445)));'
Array
(
    [1] => 41445
)

Still on the 64 bits little endian host :
x86_64$ php -r '$a = unpack("N", pack("N", 65536)); printf("$a[1]\n");'
65536           # Ok
x86_64$ php -r '$a = unpack("N", pack("N", 65535)); printf("$a[1]\n");'
-2147418113     # Weird

x86_64$ php -r '$a = unpack("N", pack("N", 0x1234)); printf("0x%x\n",
$a[1]);'
0x1234     # Ok
x86_64$ php -r '$a = unpack("L", pack("N", 0x1234)); printf("0x%x\n",
$a[1]);'
0x34120000 # Ok
x86_64$ php -r '$a = unpack("L", pack("L", 0xffff)); printf("0x%x\n",
$a[1]);'
0xffff     # Ok
x86_64$ php -r '$a = unpack("N", pack("N", 0xffff)); printf("0x%x\n",
$a[1]);'
0xffffffff8000ffff # The doc says "N" gives you "always 32 bit", and we
get
                   # 8 bytes. No wonder why we overflow.
x86_64$ php -r '$a = unpack("N", pack("N", 0xff )); printf("0x%x\n",
$a[1]);'
0xffffffff800000ff # Same. Don't tell me 0xff is too large.

And now, all the following tests are on a 64 bits _big endian_ host
(sparc64, running php-5.2.1) :
sparc64$ uname -mprsv
OpenBSD 3.8 GENERIC#607 sparc64 SUNW,UltraSPARC-IIi @ 440 MHz, version 0
FPU
sparc64$ php -r '$a = unpack("N", pack("N", 0xffff)); printf("0x%x\n",
$a[1]);'
0xffff # Ok
# The same, but prefixing 0000 to the argument :
sparc64$ php -r '$a = unpack("N", pack("N", 0x0000ffff));
printf("0x%x\n", $a[1]);'
0xffffffff8000ffff
# Weird (and with "N", we stayed on the host byte order this time).
# Shouldn't 0xffff == 0x0000ffff, even on big endian ? Apparently, yes
:
sparc64$ php -r 'printf("0x%x\n", 0x0000ffff);'
0xffff

# Also, look at this :
sparc64$ php -r '$a = unpack("N", pack("N", 41445));
printf("$a[1]\n");'
41445
# And now let's just remove the line feed (\n) from the above printf :
sparc64$ php -r '$a = unpack("N", pack("N", 41445)); printf("$a[1]");'
-2147442203

# Same for 2^16 -1 / 65535 / 0xfff :
sparc64$ php -r '$a = unpack("N", pack("N", 65535));
printf("$a[1]\n");'
65535
sparc64$ php -r '$a = unpack("N", pack("N", 65535)); printf("$a[1]");'
-2147418113

# We get the opposite (bogus with \n, correct without) when converting
# to little endian and back to host :
sparc64$ php -r '$a = unpack("L", pack("L", 0xffff)); printf( $a[1]);'
65535
sparc64$ php -r '$a = unpack("L", pack("L", 0xffff)); printf(
$a[1]."\n");'
-2147418113

This doesn't help :
SKIP Generic pack()/unpack() tests
[ext/standard/tests/strings/pack.phpt] reason: 32bit test only

Here is another example of a problem with unpack on 64bit systems. It
worked in 5.1.6, but with 5.2.1 the results are bogus.

The expected value from the file content is 200, but PHP says
-2147483448 when I echo $a['i'].

<?php
$f = fopen('test.pdf','rb');
//Read a 4-byte integer from file
$a = unpack('Ni',fread($f,4));
echo $a['i'];
fclose($f);
?>

In any case if you call it bug or a feature this is serious behavior
change for something which a lot of people could be depending on. 

It breaks in MySQL 5.2.0 -> 5.2.1  which is minor version upgrade.

It appears that the sign bit is taken from LSB instead of MSB

> php -r 'print_r( unpack('N',pack('N',127)));'
Array
(
    [1] => 127
)

> php -r 'print_r( unpack('N',pack('N',128)));'
Array
(
    [1] => -2147483520
)

The last number is 2's complement of -128 for 32 bit integers

Cheers

Please try using this CVS snapshot:

  http://snaps.php.net/php5.2-latest.tar.gz
 
For Windows:
 
  http://snaps.php.net/win32/php5.2-win32-latest.zip

Fix the problem for me, thanks. 
(ps: meanwhile, I can't verify there's no regressions, given how the
testsuite has changed).