On Mon, 1 May 2023 22:23:11 GMT, Chen Liang <li...@openjdk.org> wrote:
> It had an invocation performance of 2ns/op as opposed to Proxy's 6ns/op, but > the condy implementation has 0.41ns/op. Good, so let’s take the win relative to 6ns/op metric. The condy implementation devotes a whole class to a single MH, so it can go faster. But if you use more than exactly one MH, the extra classes will pollute the same type profiles that provide 0.41ns/op, if you use more than one MH. I think that’s too brittle a performance curve. Let’s be more economical with our spinning of classes. That should provide a smoother performance curve (scaling towards use cases with more receivers) and (with fewer, more regular classes, with no class-data) be more friendly to Leyden as well. ------------- PR Comment: https://git.openjdk.org/jdk/pull/13197#issuecomment-1532025066
