On Fri, 7 Jun 2024 12:12:44 GMT, Claes Redestad <redes...@openjdk.org> wrote:
> This PR refactors type matching in BootstrapMethodInvoker and adds a few > types, seeking to improve bootstrap overheads of some ConstantBootstraps and > in particular the ProxyGenerator condys generated for e.g. annotation proxies > since [JDK-8332457](https://bugs.openjdk.org/browse/JDK-8332457) > > I've adjusted the micro-benchmark added by JDK-8332457 to not only generate a > proxy but also call into one of the proxied methodt (`Object::hashCode`). > > Running org.openjdk.bench.java.lang.reflect.ProxyGenBench as a one-off > startup benchmark sees significant improvement (-9% instructions, -6% cycles): > > Name Cnt Base Error Test > Error Unit Change > Perfstartup-JMH 20 154,000 ± 8,165 148,000 ± > 23,164 ms/op 1,04x (p = 0,352 ) > :.cycles 925335973,200 ± 47147600,262 842221278,800 ± > 46836254,964 cycles 0,91x (p = 0,000*) > :.instructions 2101588857,600 ± 81105850,361 1966307798,400 ± > 22011043,908 instructions 0,94x (p = 0,000*) > :.taskclock 291,500 ± 16,494 262,000 ± > 15,328 ms 0,90x (p = 0,000*) > * = significant > > Number of classes loaded drops from 1096 to 1092 > > Running the micro regularly shows no significant difference: > > Name Cnt Base Error Test Error Unit > Change > ProxyGenBench.generateAndProxy100 10 26,827 ± 8,954 26,855 ± 7,531 ms/op > 1,00x (p = 0,991 ) > * = significant As an aside: I suppose we pre-generate all the linkers that get spun up for these `invokeExact`calls already. But, we might still get a little bit of mileage out of switching to `invokeBasic`, which bypasses the linkers altogether. With the caveat that we have to be very careful that the call site type is correct. ------------- PR Comment: https://git.openjdk.org/jdk/pull/19598#issuecomment-2154881850
