RFR: 8320759: Creation of random BigIntegers can be made faster [v3]

[Raffaello Giulietti]

On Tue, 5 Dec 2023 12:46:49 GMT, fabioromano1 <duke at openjdk.org> wrote:

>> Relaying comments from a colleague:
>> 
>> 1. Half of the random bits are wasted by the use of `nextInt()`, which by default invokes `nextLong()` and throws away the lower 32 bits. Not sure if complicating the code to use all the 64 bits of `nextLong()` is worthwhile, though.
>> 
>> 2. The way the random bits fill the magnitude array is different. This might break existing reproducible tests with seeded random number generators and fixed seed. Not sure if this is a real problem in practice, though.
>> 
>> 3. There seems to be no test coverage that ensures the `BigInteger` invariant has either `mag.length == 0` or `mag[0] != 0`. While the code obviously ensures it, future changes might not, so it might make sense to have this aspect covered by a test.
>
> @bplb
> About point 1. The behavior of `nextInt()` depends by the implementation of the Random object specified, which of course is not predictable. Furthermore, the claim that `nextInt()` invokes `nextLong()` by default and throws away the lower 32 bits is evidently denied by the facts: simply look at the last version of the method implementation in the class Random: https://github.com/openjdk/jdk/blob/master/src/java.base/share/classes/java/util/Random.java#L494
> So, in the light of that, I think it's not worthwhile using `nextLong()` instead of `nextInt()`, it would only complicate the code.

@fabioromano1

I made the original observations reported by @bplb.
You are right about item 1: I was erroneously looking at the implementation in interface `RandomGenerator`, not in class `Random`.

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

PR Comment: https://git.openjdk.org/jdk/pull/16817#issuecomment-1841002629