RFR: 8278114: New addnode ideal optimization: converting "x + x" into "x << 1" [v3]

[Zhiqiang Zang]

On Sat, 18 Dec 2021 03:43:08 GMT, Claes Redestad <redestad at openjdk.org> wrote:

> No other code changes, just added in those two simplified benchmarks and commented out the existing ones. I did tweak the command line to print using `ns` (you can change the annotation to get the same result) and picked the smallest number of iterations that still gave stable results on my machine while trying things out: `-i 15 -wi 15 -tu ns`.

Is there special configuration you were using? I was not able to reproduce the similar results.

This is when I used 20 iterations.
Baseline:

Benchmark                         Mode  Cnt  Score   Error  Units
AddIdeal_XPlusX_LShiftC.testInt   avgt   60  3.846 ? 0.164  ns/op
AddIdeal_XPlusX_LShiftC.testLong  avgt   60  3.332 ? 0.008  ns/op


Patch:

Benchmark                         Mode  Cnt  Score   Error  Units
AddIdeal_XPlusX_LShiftC.testInt   avgt   60  3.743 ? 0.245  ns/op
AddIdeal_XPlusX_LShiftC.testLong  avgt   60  3.664 ? 0.224  ns/op


Then I tried 100 interations to reduce noise.
Baseline:

Benchmark                         Mode  Cnt  Score   Error  Units
AddIdeal_XPlusX_LShiftC.testInt   avgt  300  3.566 ? 0.068  ns/op
AddIdeal_XPlusX_LShiftC.testLong  avgt  300  3.310 ? 0.004  ns/op


Patch:

Benchmark                         Mode  Cnt  Score   Error  Units
AddIdeal_XPlusX_LShiftC.testInt   avgt  300  3.724 ? 0.103  ns/op
AddIdeal_XPlusX_LShiftC.testLong  avgt  300  3.605 ? 0.086  ns/op


I were not able to print assembly by `-prof perfasm` on my end because of some errors from `perf` which I do not have permission to configure, but I checked native coverage and it showed this transformation did happen when the microbenchmark was running.

-------------

PR: https://git.openjdk.java.net/jdk/pull/6675