Further optimize IODeferred

[armanbilge]

Just getting it into the record, unfortunately benchmarks showed that it made no difference at best, or slower at worst. The idea was to try to cache the get: IO[A] value in the atomic reference, instead of A, similar to what IOFiber does for join.

after

[info] Benchmark                   (count)   Mode  Cnt     Score   Error  Units
[info] DeferredBenchmark.cancel         10  thrpt   20    39.504 ± 1.192  ops/s
[info] DeferredBenchmark.cancel        100  thrpt   20     4.421 ± 0.127  ops/s
[info] DeferredBenchmark.cancel       1000  thrpt   20     0.401 ± 0.018  ops/s
[info] DeferredBenchmark.complete       10  thrpt   20    30.081 ± 3.468  ops/s
[info] DeferredBenchmark.complete      100  thrpt   20     7.963 ± 1.865  ops/s
[info] DeferredBenchmark.complete     1000  thrpt   20     1.374 ± 0.194  ops/s
[info] DeferredBenchmark.get            10  thrpt   20  1050.021 ± 9.871  ops/s
[info] DeferredBenchmark.get           100  thrpt   20   163.874 ± 1.474  ops/s
[info] DeferredBenchmark.get          1000  thrpt   20    16.155 ± 0.368  ops/s

before

[info] DeferredBenchmark.cancel         10  thrpt   20   42.230 ±  1.265  ops/s
[info] DeferredBenchmark.cancel        100  thrpt   20    4.343 ±  0.138  ops/s
[info] DeferredBenchmark.cancel       1000  thrpt   20    0.407 ±  0.016  ops/s
[info] DeferredBenchmark.complete       10  thrpt   20   31.948 ±  3.528  ops/s
[info] DeferredBenchmark.complete      100  thrpt   20    8.625 ±  1.620  ops/s
[info] DeferredBenchmark.complete     1000  thrpt   20    1.292 ±  0.267  ops/s
[info] DeferredBenchmark.get            10  thrpt   20  919.962 ± 90.952  ops/s
[info] DeferredBenchmark.get           100  thrpt   20  160.503 ±  1.369  ops/s
[info] DeferredBenchmark.get          1000  thrpt   20   20.408 ±  0.176  ops/s

[armanbilge]

Let's try this again, now that we're actually using this specialization lol. Benchmarks incoming ...

[armanbilge]

Huh 🤔 complete got faster but get got much slower and that seems bonkers. This PR replaces an IO.defer(...) with a volatile read of an IO.pure(...).

This PR

Benchmark                   (count)   Mode  Cnt    Score    Error  Units
DeferredBenchmark.cancel         10  thrpt   20   44.488 ±  2.784  ops/s
DeferredBenchmark.cancel        100  thrpt   20    5.303 ±  0.117  ops/s
DeferredBenchmark.cancel       1000  thrpt   20    0.449 ±  0.018  ops/s
DeferredBenchmark.complete       10  thrpt   20   32.755 ±  5.155  ops/s
DeferredBenchmark.complete      100  thrpt   20    8.012 ±  1.811  ops/s
DeferredBenchmark.complete     1000  thrpt   20    1.367 ±  0.044  ops/s
DeferredBenchmark.get            10  thrpt   20  702.181 ± 16.228  ops/s
DeferredBenchmark.get           100  thrpt   20   74.141 ±  3.685  ops/s
DeferredBenchmark.get          1000  thrpt   20    7.310 ±  0.139  ops/s

series/3.x

Benchmark                   (count)   Mode  Cnt     Score    Error  Units
DeferredBenchmark.cancel         10  thrpt   20    44.243 ±  1.105  ops/s
DeferredBenchmark.cancel        100  thrpt   20     4.598 ±  0.179  ops/s
DeferredBenchmark.cancel       1000  thrpt   20     0.433 ±  0.025  ops/s
DeferredBenchmark.complete       10  thrpt   20    22.119 ±  7.781  ops/s
DeferredBenchmark.complete      100  thrpt   20     7.368 ±  1.470  ops/s
DeferredBenchmark.complete     1000  thrpt   20     0.970 ±  0.284  ops/s
DeferredBenchmark.get            10  thrpt   20  1639.649 ± 60.727  ops/s
DeferredBenchmark.get           100  thrpt   20   258.978 ± 12.768  ops/s
DeferredBenchmark.get          1000  thrpt   20    28.177 ±  0.913  ops/s

[armanbilge]

There we go :)

Benchmark                    (count)   Mode  Cnt     Score     Error  Units
DeferredBenchmark.cancel          10  thrpt   20    47.890 ±   0.389  ops/s
DeferredBenchmark.cancel         100  thrpt   20     4.971 ±   0.110  ops/s
DeferredBenchmark.complete        10  thrpt   20    30.252 ±   5.486  ops/s
DeferredBenchmark.complete       100  thrpt   20     7.318 ±   2.162  ops/s
DeferredBenchmark.getAfter        10  thrpt   20  3532.888 ± 101.376  ops/s
DeferredBenchmark.getAfter       100  thrpt   20   985.239 ±  19.928  ops/s
DeferredBenchmark.getBefore       10  thrpt   20  1753.401 ±  17.655  ops/s
DeferredBenchmark.getBefore      100  thrpt   20   294.561 ±  13.756  ops/s

[durban]

Is there a reason this can't be for series/3.4.x?

[armanbilge]

Good question, technically no, but there are some existing conflicts between the branches due to asyncCheckAttempt that we've hit before, and I didn't feel like flaring them up again /shrug. Also 3.5.x is around the corner ... right? 😛

[djspiewak]

Nicely done! :-)