Fix for #64450

Dmitry,

after spending more time on this issue I came to the conclusion it
shouldn't be touched :) . The x86_64 you've mentioned is solvable on
compile time using a macros like this

defined(PHP_WIN32) || ((defined(i386) || defined(__i386)) &&
defined(STDC_IEC_559))

so either old or new variant would be activated. STDC_IEC_559 is a
macros wintessing IEEE 754 conformity, so double can hold long on 32
bit.

But another issue I've overseen is that the php_mt_rand() delivers in
uint range, that's way too few if one wants to work with double as
incoming type. There are also some other places like PHP_RAND_MAX which
are incompatible. That isn't solvable with just touching the pt_rand
function.

At the end of the day seems there is no way around to get this issue
properly working on both 64 and 32 bit platforms than implementing the
Mersenne Twister for every applicable variation, effectively double or
long. That might coexist in pecl or in core. Any other intrusion would
change the core significantly.

Thanks for your support!

Anatol

On Fri, 2013-03-22 at 11:34 +0100, Anatol Belski wrote:

Thanks, now I understood. :)

anyway I see a problem. For example on x86_64 double is not always able
to keep a long number without precision lost. May be you should receive
arguments as zvals and use old or new code depending on input types.

Damn that's right, I've just read this page
http://www.viva64.com/en/a/0004/#ID0EQ3BI

double is on both x64 and x86 64 bit, but effectively 52 bits used for the
integer part. My patch would work on 32 bit Linux and both x86 and x64
windows, but would fail on x64 Linux/unix systems. I have to come up with
a better solution to do this check on compile or (more likely) on run
time.

BTW: I'm not sure if rand() should be fixed at all.
According to http://php.net/manual/en/function.rand.php it should accept
integers and return integer.

Yep, formally it's documented. De facto users meet that issue and that
could indeed be made better, that's my motivation.

Thanks for the tips.

Anatol

Thanks. Dmitry.

Dmitry,

first of all thanks for taking a look :)

The issue in a few words

• mt_rand reads arguments as long
• user pass a float from the userspace
• zend_parse_parameters with 'l' casts to long

Here's the essential snippet
echo mt_rand(0,pow(10,12)); PHP Warning: mt_rand(): max(-727379968) is
smaller than min(0)

So reading as long would probably not work, in the first place because
the input data exceeding LONG_MAX will be corrupted by the conversion.
So what
I did

• read input as max precision possible, no corruption
• calculate as well with double
• return the old way int if it's in LONG_MAX range, otherwise return
  float

Is there a solution I don't see (or do I get your suggestion wrong)?
It's
just essential to get the input as precise as it can be across PHP.

Besides that what could be impacts changing input arg types in this
concrete case? User probably wouldn't realize that at all.

Thanks

Anatol

Hi Anatol,

To be honest, I didn't understand all the details of the patch :)

However, I see a problem:
You changed the prototype of user level rand() function to receive
double arguments instead of long. I think it's disallowed, but you may
get arguments as longs and then convert them to double. Will it work?

Thanks. Dmitry.

On Thu, Mar 21, 2013 at 9:26 PM, Dmitry Stogov dmitry@zend.com
wrote:

I'll able to look only tomorrow morning.

Thanks. Dmitry.

Hi Dmitry,

I developed a patch for this one
https://bugs.php.net/bug.php?id=64450 .
It's regarding to the long overflow in mt_rand(). The main idea is
to work with the args as double internally and then return php
float if it exceeds the LONG_MAX. I strived to let the old
behavior to be unchanged, all the older tests pass. Please take a
look.

Regards

Anatol