inline constructor of abstracts

[ncannasse]

We recently changed our Heaps Vector implementation to use an abstract so we can have some operator overloading, but we noticed some regression due to constructor inline not being triggered anymore.

We can reproduce with the following sample:

class P {
   public var x : Float;
   public inline function new(x=0) this.x = x;
}


@:forward
abstract PA(P) {
  public inline function new(x) this = new P(x);
}

class Test {
  static function foo(x:Int) {
    var p1 = { v : new P(x) }
    trace(p1.v.x);
    var p2 = { v : new PA(5) }
    trace(p2.v.x);
  }
  static function main() {
    foo(5);
  }
}

It seems inline new is not handled for abstracts ?

[Simn]

cc @basro

[basro]

A quick note:
Inlining abstracts is working, it seems the combination of abstracts and anonymous structs is what is cancelling the inlining.

This inlines everything:

class P {
   public var x : Float;
   public inline function new(x) this.x = x;
}


@:forward
abstract PA(P) {
  public inline function new(x) this = new P(x);
}

class Test {
  static function foo(x:Int) {
    var p1 = { v : new P(x) }
    trace(p1.v.x);
    var p2 = new PA(5);
    trace(p2.x);
  }
  static function main() {
    foo(5);
  }
}

[basro]

I have not tested with the compiler but my bet is that the cancellation is happening in this line.

It is likely that the type of the inlined object and the type of the field access are not strictly equal.

[Simn]

Good intuition:

p2<-6565>.v
  v_type: TAbstract(PA, [])
  etype: TInst(P, [])

I don't know where that goes wrong though because the type is initially correct:

 17 |   $type(p2.v);
    |         ^^^^
    | PA