Implement recursive top-level inference

[scheglov]

Currently in the analyzer we do inference in two phases - first we resolve the initializer presuming all not yet inferred types of fields are dynamic. Then we get dependencies as elements referenced by the initializer, and infer in the topological order.

Type inference in the following code in the analyzer is done incorrectly.

class A {
  static final a = A.b.c;
  static final b = A();
  final c = a;
}

When we resolve the initializer of a, we see that it references b, but its type is presumed to be dynamic, so we don't know that it also references c. So, we miss the cycle a -> c -> a.

We should do something like CFE does, and perform recursive inference.

[stereotype441]

Note: I've just made an update to dart-lang/language#1650 describing in more detail how the analyzer and CFE currently handle top level inference, and discussing options. We may want to wait until a decision has been made on that issue before proceeding with this fix.

[scheglov]

https://dart-review.googlesource.com/c/sdk/+/206421 passes shared tests and smoke tests in google3.
I have not tested it yet on the whole google3 code base.

[scheglov]

Global presubmit looks green.
https://test.corp.google.com/ui#id=OCL:383979626:BASE:384026798:1625939487362:700fb7e6

[srawlins]

Is this still in the works, @scheglov ?

[scheglov]

It is up to the language team to decide here, whether they want to go with this change to inference.
This is tracked in dart-lang/language#1650.

We probably should keep this open.

[scheglov]

@leafpetersen This is the issue about type inference that I mentioned as related to macro type inference.

[eernstg]

@scheglov wrote:

It is up to the language team to decide here, whether they want to go with this change to inference

I just added a comment to dart-lang/language#1650, noting that there might not be a need to change anything in the spec.

[scheglov]

The CL landed, so I guess this can be marked fixed.