On 5/6/20 12:50 PM, Claes Redestad wrote:
On 2020-05-06 16:59, Claes Redestad wrote:
-XX:TieredStopAtLevel=1:
Benchmark Mode Cnt Score Error Units
15-b19 Capture0.lambda_01 avgt 5 6.312 ± 0.919 ns/op
15-b20 Capture0.lambda_01 avgt 5 105.761 ± 7.575 ns/op
jdk Capture0.lambda_01 avgt 5 105.228 ± 7.175 ns/op
So outside of the direct bootstrap improvement, there's a positive
effect during early warmup. Sadly this patch doesn't translate into an
improvement in C1. I can see a rather large regression in C1 on this
specific micro in 15-b20 (which is where the Hidden Classes integration
happened).
Filed: https://bugs.openjdk.java.net/browse/JDK-8244550
Thanks.
Overhead is entirely in Unsafe::allocateInstance, which will be
intrinsified by C2, but has ~100ns/op overhead when called from
interpreter and C1 code.
I've prototyped a patch to remove a few redundant VM->native transitions
and gotten this overhead down to 60ns/op.
FTR, prior to hidden classes, LMF generated a factory method to create
an instance of lambda proxy via invokespecial. With hidden classes, it
can get the constructor iva findConstructor and that causes
Unsafe::allocateInstance to be called.
Mandy