Originally reported here:
 https://twitter.com/junkdogAP/status/510414300587261952

Remeasuring with:
 http://cr.openjdk.java.net/~shade/8058319/ConstructorBenchmark.java
 http://cr.openjdk.java.net/~shade/8058319/benchmarks.jar

...and current jdk9-dev (with LF caching already here) yields:

 - create_handle 25.301 +- 3.162 ops/us
 - create_handle_unreflect 25.424 +- 0.967 ops/us
 - create_plain 203.557 +- 33.607 ops/us
 - create_reflect 80.443 +- 6.504 ops/us

Perfasm profiles for plain/reflect cases are relatively clean:
 http://cr.openjdk.java.net/~shade/8058319/create_plain.perfasm
 http://cr.openjdk.java.net/~shade/8058319/create_reflect.perfasm

We seem to blow inlining in MH cases:
 http://cr.openjdk.java.net/~shade/8058319/create_handle.perfasm
 http://cr.openjdk.java.net/~shade/8058319/create_handle_unreflect.perfasm