On Wed, 15 Oct 2025 10:10:43 GMT, Raffaello Giulietti <[email protected]> wrote:
>> @rgiulietti Here is the PR. > > @fabioromano1 In the comments there are claims about the relative errors. Two > questions arise: > * How were the bounds determined? There seems to be no indication in the > comments provided so far. > * How much code would be needed to attain correct rounding in all modes? @rgiulietti > * How were the bounds determined? There seems to be no indication in the > comments provided so far. In case of `n < 0` and no halfway rounding, the root's inversion gives an error of 1 ulp, that's why we add one more digit to the integer root, so that the integer root has an error of 0.1 ulps, and therefore the error of the inverted root is 1+0.1 ulps. Same reasoning for halfway rounding gives half of the error just found. > * How much code would be needed to attain correct rounding in all modes? If `n > 0`, the PR code already gives correct rounding in all modes, but in case of `n < 0`, I'm afraid there's no way to have always correct rounding, due to table maker's dilemma. ------------- PR Comment: https://git.openjdk.org/jdk/pull/27148#issuecomment-3405797048
