Optimize some application-related operations

[odersky]

No description provided.

[odersky]

test performance please

[dottybot]

performance test scheduled: 3 job(s) in queue, 1 running.

[dottybot]

Performance test finished successfully:

Visit http://dotty-bench.epfl.ch/9875/ to see the changes.

Benchmarks is based on merging with master (f4528ce)

[odersky]

test performance please

[dottybot]

performance test scheduled: 3 job(s) in queue, 1 running.

[dottybot]

Performance test finished successfully:

Visit http://dotty-bench.epfl.ch/9875/ to see the changes.

Benchmarks is based on merging with master (f4528ce)

[som-snytt]

In the old days, we would move the condition to pattern guard in order to avoid parens. Now we will do it to avoid then.

[liufengyun]

LGTM 👍

[odersky]

This was surprisingly effective! Still not quite sure why.

widen is certainly a hotspot, called 3.4M times for typer/*.scala, which makes it the hottest method except for getter-like things. The new implementation gets common cases out of the way before doing essentially what the old one did. The old implementation did

• two virtual calls which should be predictable by the JIT and that resolve to just returning this in the common case
• a type test against a trait

I have the impression that type tests against traits are quite inefficient. Can somebody confirm or refute this?

[liufengyun]

I did some benchmark, which seems to validate the following claim:

I have the impression that type tests against traits are quite inefficient. Can somebody confirm or refute this?

  trait TA
  trait TB extends TA
  trait TC
  class A
  class B extends A with TC with TB
  class C extends A with TC with TB
  val b = new B
  val c = new C

  var objs: Array[A] = Array(b, c)

  @Benchmark
  def typeTestTrait(): Boolean = {
    var res = false
    var count = 0
    while count < 1000 do
      res = objs(count % 2).isInstanceOf[TA]
      objs(0) = b
      count += 1
    res
  }

  @Benchmark
  def typeTestClass(): Boolean = {
    var res = false
    var count = 0
    while count < 1000 do
      res = objs(count % 2).isInstanceOf[B]
      objs(0) = b
      count += 1
    res
  }

• typeTestClass : 1109.409 ns/op
• typeTestTrait: 1829.782 ns/op

[odersky]

@liufengyun That's actually not so bad for the trait type tests. But there could be other variables to consider, like how many traits are inherited and in what order. I believe trait type test comes down to a linear scan through implemented traits.

[odersky]

test performance please

[dottybot]

performance test scheduled: 1 job(s) in queue, 0 running.

[odersky]

I am going to try to avoid having to widen before typeSymbol. This removes a common pitfall and should also help performance.

[dottybot]

Performance test finished successfully:

Visit http://dotty-bench.epfl.ch/9875/ to see the changes.

Benchmarks is based on merging with master (f4528ce)

[odersky]

test performance please

[dottybot]

performance test scheduled: 1 job(s) in queue, 0 running.

[odersky]

The last performance test reverted all of the gains made for dotc (it kept some of the smaller gains in the other tests). So was this just noise? Reverting to original state and testing again.

[michelou]

@liufengyun In your benchmark shouldn't you use "isInstanceOf[TA]" against "isInstanceOf[A]" instead of "isInstanceOf[TA]" against "isInstanceOf[B]" (same chain length) ?!

[dottybot]

Performance test finished successfully:

Visit http://dotty-bench.epfl.ch/9875/ to see the changes.

Benchmarks is based on merging with master (f4528ce)

[odersky]

test performance please

[dottybot]

performance test scheduled: 1 job(s) in queue, 0 running.

[liufengyun]

@liufengyun In your benchmark shouldn't you use "isInstanceOf[TA]" against "isInstanceOf[A]" instead of "isInstanceOf[TA]" against "isInstanceOf[B]" (same chain length) ?!

@michelou That does not make much difference in the performance numbers. I guess JVM uses clever encoding of class tags to make class testing faster.

[dottybot]

Performance test finished successfully:

Visit http://dotty-bench.epfl.ch/9875/ to see the changes.

Benchmarks is based on merging with master (f4528ce)

[odersky]

test performance please

[dottybot]

performance test scheduled: 1 job(s) in queue, 0 running.

[dottybot]

Performance test finished successfully:

Visit http://dotty-bench.epfl.ch/9875/ to see the changes.

Benchmarks is based on merging with master (f4528ce)

[odersky]

After trying many variants I conclude that the initial gain was mostly noise; but there are some small wins nevertheless