[jdk17] RFR: 8266313: (JEP-356) - RandomGenerator spec implementation requirements tightly coupled to JDK internal classes [v2]

[Jim Laskey]

On Mon, 28 Jun 2021 15:57:21 GMT, Jim Laskey <jlaskey at openjdk.org> wrote:

>> src/java.base/share/classes/java/util/random/RandomGenerator.java line 648:
>> 
>>> 646:      * power of two then limiting is a simple masking operation. Otherwise, a
>>> 647:      * new result is re-calculated by averaging the previous result and
>>> 648:      * {@code nextInt()} until the final result is greater equal zero and less
>> 
>> I don't see how 'averaging' comes from the invocation of boundedNextInt which appears to choose the first candidate that meets the criteria.
>> The comment also applies to nextLong overloads.
>
> If the first value calculated doesn't meet the criteria, then you either mask (power of 2) or go into this loop which does the averaging. The intent is not to bias the outcome.
> 
> 
>                 for (long u = r >>> 1;            // ensure nonnegative
>                      u + m - (r = u % n) < 0L;    // rejection check
>                      u = rng.nextLong() >>> 1) // retry
>                     ;

Yea you're right. I was focused on the divide by two, which appears to be to just remove the sign. Clearer interpretation:


r = rng.nextLong();
n = range;
m = n - 1;                   // mask
for (long u = r / 2;         // ensure nonnegative
     u + m < r;              // rejection check
     u = rng.nextLong() / 2) // retry
    r = u % n;

-------------

PR: https://git.openjdk.java.net/jdk17/pull/151