On Fri, 18 Dec 2020 15:48:52 GMT, PROgrm_JARvis <github.com+7693005+jarviscr...@openjdk.org> wrote:
>> Might be fun to try, but it looks like rewriting to have MD5 to only use >> transient state will be a significant effort, and might just end up >> shuffling over allocations from `getInstance` to `digest`, which could >> regress code that reuse a digest instance (in a thread safe manner). > >> Might be fun to try, but it looks like rewriting to have MD5 to only use >> transient state will be a significant effort, and might just end up >> shuffling over allocations from `getInstance` to `digest`, which could >> regress code that reuse a digest instance (in a thread safe manner). > > I've thought about using `ThreadLocal` at holder level and my naive benchmark > shows significant improvement in speed compared to current implementation: > https://gist.github.com/JarvisCraft/61cd45f83aea7f33cba9510e523a5c4e > > As of the state from single thread's perspective, it seems to be safe to > reuse `MessageDigest` in this case: > >> The `digest` method can be called once for a given number of updates. After >> `digest` has been called, the MessageDigest object is reset to its >> initialized state. I've pushed the change utilizing `ThreadLocal` and my benchmark to correspond to the discussion. ------------- PR: https://git.openjdk.java.net/jdk/pull/1821
