Add support for treating generic method parameters as dynamic in all contexts

[leafpetersen]

The current implementation of generic method syntax in the VM and dart2js treats generic type parameters as malformed types, which get mapped to dynamic in some (but not all) contexts. In particular, they are not usable in is or as checks. In order to remove the comment syntax fully from strong mode, we will need an option in the VM and dart2js to enable a mode in which all uses of generic method parameters are mapped to dynamic.

[crelier]

I have sent a cl for review addressing this in the VM. I did not add another mode, and the new behavior is what you get with --generic-method-syntax.

[eernstg]

Actually, I recommended a change to the vm implementation in #27437 such that it would treat a method type parameter as dynamic when performing checked mode checks. The rationale was that the intention behind type annotations is that the dynamic behavior of the program should satisfy the implied constraints. Hence, in a correct program there should never be a failure during such a check, and silently succeeding will not introduce bugs into correct programs.

An is check is different: It is perfectly plausible that a correct program uses an is check in such a way that it should evaluate to false in certain situations. This means that it would be misleading for an implementation to silently evaluate x is T to true in all cases, it would silently introduce logical bugs into correct programs. That's the reason why dart2js raises a runtime error on is and is! checks with method type parameters.

The as check falls between these two extremes. We could put it into any of the buckets ("it's just like a checked/strong mode check on a type annotation" vs. "it's a dynamic check, just like is"), and it will currently raise an error in dart2js.

In short, I'm not convinced that we should let x is T return true in all cases.

[eernstg]

The CL https://codereview.chromium.org/2388843002/ implements a new semantics for o as T in dart2js: It is checked as if T had been dynamic. I haven't changed the semantics of o is T, based on the worries I described in the previous comment.

[eernstg]

Note that we haven't actually settled what the semantics should be.

[floitschG]

I'm not sure allowing is checks on generic method arguments is a good idea. It won't work in dart2js and in the VM and there isn't even a good default way of dealing with it. If it's really much easier to deal with in the VM, then we can/should go this route, but if it doesn't make a big difference, then I would rather discourage users from doing is/as checks on these arguments until the language actually handles them.

Independently reopening since only the VM would satisfy the original request (and this issue has also been filed for dart2js and the language).

[leafpetersen]

Yes, it's really not ideal that we will get different behavior on the VM/dart2js vs DDC for these checks, but unfortunately we're already in that state because of the comment syntax. I did some code searching, and I can't find any uses of the comment syntax for is checks (thankfully!). There are many uses of the comment syntax for as checks. So I think we could get away with the compromise position of treating uses of generic parameters in is checks as an error, but allowing them everywhere else.

We may want to consider issuing a static warning for is checks on generic method parameters for the time being as well (i.e. in the analyzer).