Hi,
the LambdaFormEditor uses a SoftReference-based cache to avoid
generating LambdaForms we have already generated. By introducing a
lookup key that is not a SoftReference we can improve this by allocating
less up front and by inducing less work for the GC later on.
I also refactored a few cases where I had neglected to add a static
factory method to create the Transform (now TransformKey). For the
transform types affected this helps further reduce allocations.
Bug: https://bugs.openjdk.java.net/browse/JDK-8246241
Webrev: http://cr.openjdk.java.net/~redestad/8246241/open.00/
Testing: tier1-2, java/lang/invoke locally with STRESS_TEST=true
Verified a speed-up on the micro added by JDK-8246152)[1], and
on related startup tests[2].
Thanks!
/Claes
[1]
Before (JDK-8246152):
Benchmark Mode Cnt Score Error Units
StringConcatFactoryBootstraps.makeConcatWithConstants avgt 10 2764.634 ±
56.244 ns/op
StringConcatFactoryBootstraps.makeConcatWithConstants:·gc.alloc.rate.norm
avgt 10 4136.187 ± 0.013 B/op
After:
Benchmark
Mode Cnt Score Error Units
StringConcatFactoryBootstraps.makeConcatWithConstants avgt 10 2299.919 ±
58.003 ns/op
StringConcatFactoryBootstraps.makeConcatWithConstants:·gc.alloc.rate.norm
avgt 10 3308.145 ± 6.378 B/op
[2]
http://cr.openjdk.java.net/~redestad/scratch/erase_scf_types.00/ObjStringCombos.java
Before (JDK-8246152):
3,485,679,714 instructions # 1.09 insn per
cycle ( +- 0.90% )
692,247,529 branches # 491.822 M/sec
( +- 0.92% )
28,939,534 branch-misses # 4.18% of all
branches ( +- 0.90% )
0.291642667 seconds time elapsed ( +- 0.93% )
After:
3,411,619,586 instructions # 1.09 insn per
cycle ( +- 0.63% )
677,767,258 branches # 493.911 M/sec
( +- 0.64% )
28,337,022 branch-misses # 4.18% of all
branches ( +- 0.63% )
0.288959534 seconds time elapsed
( +- 0.86% )
