I just finished the first prototype, which can be found at
https://gist.github.com/DasBrain/7766bedfbc8b76ba6a0ee66a54ba97ed - it
contains a patch, the javap output of a generated proxy class with and
without that patch, and the code I used to dump the proxy class. "make
run-test-tier1" passes all JDK tests with the patch. I hope this will
help the discussion a bit.
Good start.
https://mail.openjdk.java.net/pipermail/core-libs-dev/2019-August/061923.html
But there were never a real discussion about the benefits yet. It's
good that we have one now.
As I am a novice JMH user, I don't trust myself to write good
benchmarks that actually measure what I want to measure, not confirm
what I would like it to be
Right, I think that's the next step. You need to design an experiment
to measure the cost of the <clinit> here. I suspect the best way is to
lift that work to an instance method (one that just overwrites the
fields), and benchmark that, for various numbers of reflective calls.
Then we can separately measure the cost of spinning the bytecode for a
proxy, and of loading the class from a cached byte[]. This will give us
a sense of the relative costs of the various pieces, to validate that
reducing the <clinit> cost is relevant.
Finally, we can benchmark the current approach against the LDC approach
on a per-Method basis. The LDC approach may well be doing more work per
Method, so it's a tradeoff to determine whether deferring that work is a
win.
