Bug 151641 – Use a better RNG for Math.random()

• Status: RESOLVED
• Resolution: FIXED
• Component: JavaScriptCore

[Jan de Mooij]

The GameRand-based RNG Safari/JSC uses is very weak. Furthermore, it only uses 32 bits of precision. I blogged about this here:

http://jandemooij.nl/blog/2015/11/27/math-random-and-32-bit-precision/

V8 moved to a better RNG and we're switching the one in SpiderMonkey as well. JSC should probably do the same.

[Oliver Hunt]

If we really do consider more than 32bits necessary id just move up to a 64bit variant of gamerand which should not have a significant perf impact, then value & (2^53-1) / 2^53 (I can recall if the rand includes 1 as a valid result).

Has someone done something to find it's aggressively non-uniform in some space?

[Geoffrey Garen]

Created attachment 266302 [details]
Patch

[Geoffrey Garen]

I kind of wanted to stick with GameRandom in order to troll the internet commentators who keep confusing / conflating RNG algorithm with number of bits provided by RNG interface.

But there's literally no downside to upgrading to Xorshift+, so I've done it.

[Anders Carlsson]

Comment on attachment 266302 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=266302&action=review

> Source/WTF/ChangeLog:10
> +        prevents programs that mulitply a random number by a large constant from

typo, multiply.

[Geoffrey Garen]

Committed r192855: <http://trac.webkit.org/changeset/192855>

[Jan de Mooij]

(In reply to comment #3)
> I kind of wanted to stick with GameRandom in order to troll the internet
> commentators who keep confusing / conflating RNG algorithm with number of
> bits provided by RNG interface.

It's not just that; GameRand also doesn't do well in statistical tests:

http://jandemooij.nl/blog/2015/11/30/testing-math-random-crushing-the-browser/

No big deal probably, but it shows the algorithm has other issues.

> But there's literally no downside to upgrading to Xorshift+, so I've done it.

\o/