Decide on default value for -Xmixin-force-forwarders in 2.12.0

[lrytz]

Ticket to make sure it doesn't get forgotten.

Jason's latest benchmark shows ~10% improvement in cold performance, none when warm scala/scala#5390 (comment).

It seems that changing the value should be binary compatible, in which case we could change the default (in either direction) in a minor release.

Should think more about and test binary compatibility

• Is the signature of a forwarder always the same as the callee?
• Any special cases with bridges? Specialization?

[lrytz]

We decided to set the default value to true for 2.12.0.

Still to do: run compiler benchmarks combinations of

• -Xmixin-force-forwarders:true / false
• static super accessors yes / no (Consider eliminating static accessors for trait methods #228)

[retronym]

Benchmarking with static accessors enabled:

⚡ git log head --not origin/2.12.0 --oneline  --graph
* b68c7c2 Emit all mixin forwarders when building Scala
* c2bd533 Bump to prerelease SBT
* 01fb73c Bump sbt.version to 0.13.12, without breaking
* e08b88f Make isSeparateCompiled... robust against rogue phase time travel
* 95234e7 SD-226 Be lazier in Fields info transform for better performance
* b28bc8c Cast more pro-actively in synthetic accessor trees.
* 733aaf4 Avoid mismatched symbols in fields phase

[info] Benchmark                            (_scalaVersion)  (extraArgs)  (source)  Mode  Cnt     Score     Error  Units
[info] ColdScalacBenchmark.compile  2.12.0-b68c7c2-SNAPSHOT                 scalap    ss   16  6583.898 ± 133.473  ms/op
[info] ColdScalacBenchmark.compile  2.12.0-c2bd533-SNAPSHOT                 scalap    ss   16  7664.687 ± 578.979  ms/op

[info] Benchmark                            (_scalaVersion)  (extraArgs)                               (source)  Mode  Cnt      Score      Error  Units
[info] ColdScalacBenchmark.compile  2.12.0-b68c7c2-SNAPSHOT               @/code/scala/sandbox/args-compile.txt    ss    8  37457.935 ± 1532.913  ms/op
[info] ColdScalacBenchmark.compile  2.12.0-c2bd533-SNAPSHOT               @/code/scala/sandbox/args-compile.txt    ss    8  38863.001 ± 1388.303  ms/op

10-15% speedups for cold performance, similar to our previous measurements.

Hot performance seems very close for scalap and better-files. Using src/compiler as another benchmark, I saw an improvement mixin forwarders, but I think that I was observing a difference in warmup speed, moreso than a difference in peak performance.

I don't currently have a good way to quantify warmup speed. Would be good to gather kloc/s compile data for 0-5s, 5-10s etc and compare. But it is hard to fit large codebases (e.g. src/compiler) into these buckets.

[adriaanm]

scala/scala#5429