Inlining issue because @:coreApi is annoying

[Simn]

class Main {
	static public function main() {
		var map = new StringMap();
		map.set("foo", 12);
		map.set("bar", 13);
		for (value in map) { }
	}
}

This generates:

	var value = new haxe_ds__$StringMap_StringMapIterator(map,map.arrayKeys());
	while(value.hasNext()) value.next();

The reason for that is obvious in the dump file:

[Var value(4799):Iterator<Int>]
	[Cast:Iterator<Int>]
		[New:haxe.ds._StringMap.StringMapIterator<Int>]
			haxe.ds._StringMap.StringMapIterator<Int>
			[Local map(4563):haxe.ds.StringMap<Int>:haxe.ds.StringMap<Int>]
			[Call:Array<String>]
				[Field:Void -> Array<String>]
					[Local map(4563):haxe.ds.StringMap<Int>:haxe.ds.StringMap<Int>]
					[FInstance:Void -> Array<String>]
						haxe.ds.StringMap<Int>
						arrayKeys
[While:Void]
	[Parenthesis:Bool]
		[Call:Bool]
			[Field:Void -> Bool]
				[Local value(4799):Iterator<Int>:Iterator<Int>]
				[FAnon:Void -> Bool] hasNext
	[Call:Int]
		[Field:Void -> Int]
			[Local value(4799):Iterator<Int>:Iterator<Int>]
			[FAnon:Void -> Int] next

The underlying problem here is that StringMap.iterator is typed to be Iterator<T>, which means the next/hasNext field accesses are made on the structure and are not inlined.

From the inliner's point of view, this is the correct behavior. We don't want to lose the return type of an inline function, so casting to Iterator<Int> is only consistent.

I think @:coreApi should allow covariance for return types so that StringMap.iterator can be typed to return StringMapIterator<T> instead of Iterator<T>.

[Simn]

Surprisingly, this isn't a regression. But it's pretty dumb...

[ncannasse]

covariance is fine, but requires proper docgen handling so it doesn't generate different method signatures

[Simn]

Yes docgen is also my main concern here...

[Simn]

Supporting covariance here is not an easy change. We currently use type_eq which has special semantics for the EqCoreType case. That's because we compare classes from different contexts, which means cTarget != cCoreType. Supporting covariance would require us to go through unify instead, but unify doesn't support this kind of equality checks.

[Simn]

Not a regression and very tricky, pushing back.