RFR: 8326227: Rounding error that may distort computeNextGaussian results [v3]

[Jaikiran Pai]

On Thu, 22 Feb 2024 05:40:10 GMT, Joe Darcy <darcy at openjdk.org> wrote:

>> Update: confirmed that the new test fails without the fix.
>
>> Update: confirmed that the new test fails without the fix.
> 
> Thanks for verifying the test checks the fix; I'll let others who have worked more directly on the random code review the actual fix.

> @jddarcy Added a regression test, and currently working on adjusting it (see https://github.com/Pr0methean/jdk/actions/runs/7984444127) to ensure we have a case that fails without the fix, passes with the fix, and is practical to run within the usual unit-test timeouts.

I gave this a try locally. It doesn't fail for me without the source code changes proposed in this PR. I see the following output from the test without the source code changes:


got 1.0 for max 1.0
got 2.0 for max 2.0
got 3.0 for max 3.0
got 4.0 for max 4.0
got 5.0 for max 5.0
got 6.0 for max 6.0
got 7.0 for max 7.0
got 11.353912041222094 for max 8.0
got 11.353912041222094 for max 9.0


With the proposed changes in this PR, the test continues to pass and I see this output:


got 7.569274694148063 for max 1.0
got 7.569274694148063 for max 2.0
got 7.569274694148063 for max 3.0
got 7.569274694148063 for max 4.0
got 7.569274694148063 for max 5.0
got 7.569274694148063 for max 6.0
got 7.569274694148063 for max 7.0
got 11.353912041222094 for max 8.0
got 11.353912041222094 for max 9.0

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

PR Comment: https://git.openjdk.org/jdk/pull/17703#issuecomment-2162623391