On Thu, 13 Jul 2023 09:27:03 GMT, Pavel Rappo <pra...@openjdk.org> wrote:
> > @pavelrappo Actually I look at the implementation of > > `ArraySupport.vectorizedHashCode` and it does branching and only uses > > vector instructions if the length is sufficiently large. Please do > > benchmarking if you are uncertain with performance characteristics of > > functions. Thanks. > > Are you sure you are talking about `vectorizedHashCode` and not > `vectorizedMismatch`? While the latter indeed branches on array length, the > former consists of simple for-loops, which might not be the actual thing > anyway because the whole method is `@IntrinsicCandidate`. > > Regardless of any performance characteristics, the call to something named > `vectorizedHashCode` might needlessly draw attention to something that does > not deserve it. > > My mental model might be wrong, but let me give you a stream analogy. Would > you not pause and think if you unexpectedly saw this? > > ``` > ... > .stream() > ... > .parallel() > ... > ``` I don't think it's a good analogy. In my understanding, it is an obvious fact that `vectorizedHashCode` will not always vectorize, as vectorized code must always handle suffixes that cannot be vectorized. I don't think judging outside `vectorizedHashCode` will make the code clear, unless `vectorizedHashCode` only handles the part that can be vectorized, and the processing of suffixes is also outside `vectorizedHashCode`. ------------- PR Comment: https://git.openjdk.org/jdk/pull/14831#issuecomment-1634378224
