-D no-inline breaks map literals in HashLink

[player-03]

Edit: See my reply below for a minimal example.

---

I'm trying to build the Lime tools for HL, but it gives me an Invalid_argument("List.map2") error. Using OCAMLRUNPARAM=b I was able to track it down to genhl.ml line 519, but that's not very helpful.

I would absolutely love to give you a minimal example, but the error message isn't specific enough for me to narrow it down.

When I have more time, I think my plan is to insert a temporary try block in genhl.ml so I can get e.epos (which I would use to narrow the error down). But I have no Ocaml experience, so I was hoping someone else could tell me how to print that data. This is my best attempt at the code:

-				TFun (List.map2 (fun ((name,opt,t) as a) ((_,_,t2) as a2) ->
+				try TFun (List.map2 (fun ((name,opt,t) as a) ((_,_,t2) as a2) ->
					match t, t2 with
					(*
						Handle function variance:
						If we have type parameters which are function types, we need to keep the functions
						because we might need to insert a cast to coerce Void->Bool to Void->Dynamic for instance.
					*)
					| TInst ({cl_kind=KTypeParameter _},_), TFun _ -> a2
					(*
						If we have a number, it is more accurate to cast it to the type parameter before wrapping it as dynamic
						Ignore dynamic method (#7166)
					*)
					| TInst ({cl_kind=KTypeParameter _},_), t when is_number (to_type ctx t) && (match f with FInstance (_,_,{ cf_kind = Var _ | Method MethDynamic }) -> false | _ -> true) ->
						(name, opt, TAbstract (fake_tnull,[t]))
					| _ ->
-						a
+						a with
+				Invalid_argument -> print_endline e.epos; None

[Simn]

Yeah, OCaml has no "trace anything" functionality. You can try something like abort "Uhm" e.epos.

[player-03]

Ok, the try ... with simply didn't work, but thanks to print_string() and print_int() I was able to narrow it down.

It turns out this is pretty easy to reproduce. Just compile the following for HashLink with -D no-inline:

class Main {
	public static function main():Void {
		var map:Map<String, String> = ["a" => "b"];
	}
}

(It's probably obvious, but e.epos will point to the characters ["a" => "b"].)

(I don't know why Lime would use no-inline, but that's what happened.)

[Apprentice-Alchemist]

When map comprehension is used (with -D no-inline), this is passed to the generator: [Field:(String, String) -> Void] [TypeExpr haxe.ds._Map.Map_Impl_:Abstract<haxe.ds.Map>] [FStatic:(String, String) -> Void] haxe.ds._Map.Map_Impl_ set:(this : haxe.IMap<haxe.ds.Map.K, haxe.ds.Map.V>, key : haxe.ds.Map.K, value : haxe.ds.Map.V) -> Void
List.map2 is unhappy here because the type in the Field node is (String, String) -> Void, but field_type gives (this : haxe.IMap<haxe.ds.Map.K, haxe.ds.Map.V>, key : haxe.ds.Map.K, value : haxe.ds.Map.V) -> Void. One has 2 arguments, one has 3, therefore an exception is thrown.

When one builds the map manually, the types match, so this seems like bad codegen in whatever's responsible for desugaring map comprehension?