Micro optimise @volatile lazy vals

[allanrenucci]

Given a lazy field of the form:

class Foo {
  @volatile lazy val bar: Int = <RHS>
}

The compiler will now generate code equivalent to:

class Foo {
  import dotty.runtime.LazyVals
  var value_0: Int = _
  var bitmap: Long = 0L
  val bitmap_offset: Long = LazyVals.getOffset(classOf[LazyCell], "bitmap")

  def bar(): Int = {
    while (true) {
      val flag = LazyVals.get(this, bitmap_offset)
      val state = LazyVals.STATE(flag, <field-id>)

      if (state == <state-3>) {
        return value_0
      } else if (state == <state-0>) {
        if (LazyVals.CAS(this, bitmap_offset, flag, <state-1>, <field-id>)) {
          try {
            val result = <RHS>
            value_0 = result
            LazyVals.setFlag(this, bitmap_offset, <state-3>, <field-id>)
            return result
          }
          catch {
            case ex =>
              LazyVals.setFlag(this, bitmap_offset, <state-0>, <field-id>)
              throw ex
          }
        }
      } else /* if (state == <state-1> || state == <state-2>) */ {
        LazyVals.wait4Notification(this, bitmap_offset, flag, <field-id>)
      }
    }
  }
}

[allanrenucci]

This reduces the gap for initialised access performance:

 Benchmark        Mode  Cnt  Score   Error  Units
 measureBaseline  avgt   10  2.361 ± 0.069  ns/op
 measureScala2    avgt   10  2.385 ± 0.053  ns/op
-measureVolatile  avgt   10  4.155 ± 0.192  ns/op
+measureVolatile  avgt   10  3.160 ± 0.107  ns/op

Slightly improve uncontended initialisation performance:

 Benchmark        Mode  Cnt   Score   Error  Units
measureBaseline  avgt   10   5.802 ± 0.133  ns/op
measureScala2    avgt   10  23.221 ± 0.495  ns/op
-measureVolatile  avgt   10  24.082 ± 0.376  ns/op
+measureVolatile  avgt   10  22.238 ± 0.853  ns/op

And it doesn't hurt multi-threaded contended initialisation

[sjrd]

Have you tried the variant where the slow path is moved to a separate method?

[allanrenucci]

Have you tried the variant where the slow path is moved to a separate method?

Not yet. But this PR simplifies the code generated (and the implementation). So I would say, let's get this in and then we can experiment in separate PRs.

[sjrd]

Two tiny details.

Comments addressed