On Wed, 22 Oct 2025 15:46:08 GMT, Joe Darcy <[email protected]> wrote:
> I'll do some additional double-checking that the recorded values meet the > assumptions of the WorstCaseTests. > > There is some algorithmic divergence between the FDLIBM derivative used in > Java and the OpenLibm used in Julia. Therefore, it is not necessarily > problematic that the reference values are different in the StrictMath tests. Pushed an update with a few changes: For the less-than-one-ulp error for methods already covered in WorstCaseTests, I verified the lesser magnitude number of the values bracketing the exact result was listed as the expected value. (If the wrong endpoint is chosen, the test could spuriously fail on some implementations.) For methods not covered in WorstCaseTests, I added test cases for the StrictMath tests of the method in question. ------------- PR Comment: https://git.openjdk.org/jdk/pull/27939#issuecomment-3434835055
