Hi,
by packing the coder and length values into a long, and combining the
length and coder mixers, we can reduce the number of MHs generated to
set up String concatenations.
Theoretically we reduce the number of MHs created by up to a third, and
real time improvements in startup tests exercising ISC typically hover
around 10%. This optimization reverts JDK-8212597[1], so expressions
that benefited from that of course sees a smaller or no added benefit.
http://cr.openjdk.java.net/~redestad/8213035/jdk.00/
https://bugs.openjdk.java.net/browse/JDK-8213035
I've attached a simple microbenchmark and performance results from
running on x64 in the bug:
- Significant reduction in work needed to bootstrap a slew of
concatenations, both on the initial bootstrap and when repeating similar
concatenation shapes.
- For C2 the peak performance is more or less the same, with very small
differences due slight differences in the generated code (both tiny
improvements and tiny regressions).
- For C1 we see some regressions on some micros, at worst 1.2x
- For interpreter we see improvements around 1.1x on several micros
Testing: tier1-2 jdk and hs tests, sanity tested overflow conditions
Thanks!
/Claes
[1] https://bugs.openjdk.java.net/browse/JDK-8212597
