On Thu, 8 Jun 2023 14:07:24 GMT, Brian Goetz <briango...@openjdk.org> wrote:
> Here are a few options: - Make the CHA cache thread-safe using, say, CHM; This is current solution for the default system CH cache. We may also use a bit less strict custom semi-synchronization (just enough to be thread safe) instead of fully-synchronized `CHM::computeIfAbsent`, for example `CHM:get … (compute if absent without lock) ... CHM::put`. CHM does not lock on get, so the only penalty would be synchronous put call. > - Make the CHA cache unshared, by using a ThreadLocal (this brings the >context back to being immutable) This is interesting option for small number of threads, however it makes big footprint for heavy-parallel systems. > - Detect when implicit use of a context happens on a thread other than the >creating thread (currently the only vector for this is Classfile::transform), >and inflate a new context with empty cache in that case; This would degrade the caching completely by triggering a new context for each transformed method in case there are two threads working synchronously. > - Always create a fresh context with an empty cache in Classfile::transform This would destroy the caching. Such cache would work only in scope of one method. ------------- PR Comment: https://git.openjdk.org/jdk/pull/14180#issuecomment-1582943385
