jrand48(): missing cast to int32_t?

jrand48(): missing cast to int32_t?

[Ed Schouten]

Hi there,

According to POSIX, jrand48() is supposed to return a value between
[-2^31, 2^31). Musl's version of jrand48() is implemented as follows:

long jrand48(unsigned short s[3])
{
        return __rand48_step(s, __seed48+3) >> 16;
}

Notice that the function only uses simple bit shifting and its return
type is long, meaning that on 64-bit systems, this implementation may
actually return a value between [0, 2^32).

For some reason, FreeBSD had the same bug. By adding a cast to
int32_t, its implementation now returns the same sequence of
pseudo-random numbers as macOS, Solaris, glibc, etc.

https://svnweb.freebsd.org/base?view=revision&revision=309650

Musl's version should likely be changed to this:

long jrand48(unsigned short s[3])
{
        return (int32_t)(__rand48_step(s, __seed48+3) >> 16);
}

Best regards,
-- 
Ed Schouten <ed@nuxi.nl>
Nuxi, 's-Hertogenbosch, the Netherlands
KvK-nr.: 62051717

Re: jrand48(): missing cast to int32_t?

[Szabolcs Nagy]

* Ed Schouten <ed@nuxi.nl> [2016-12-09 07:16:42 +0100]:
> According to POSIX, jrand48() is supposed to return a value between
> [-2^31, 2^31). Musl's version of jrand48() is implemented as follows:
> 
> long jrand48(unsigned short s[3])
> {
>         return __rand48_step(s, __seed48+3) >> 16;
> }

nice catch, note that posix does not guarantee that
long can represent -2^31 so this requirement is not
possible to implement portably (indeed your proposed
cast has implementation defined effect, but in musl
we already rely on it to work as expected).

i submitted a posix bug report against limits.h
http://austingroupbugs.net/view.php?id=1108