Specify function object canonicalization

[eernstg]

This PR adds a couple of paragraphs to the language specification, specifying that constant expressions whose evaluation is a function closurization or a generic function instantiation are canonicalized.

[eernstg]

@johnniwinther, @mraleph, do you think this change could be a front-end only thing? Would it potentially be a substantial effort?

[johnniwinther]

What is a "function closurization"? Can you give an example of what this change should support?

[eernstg]

A function closurization is a tearoff when the denoted function is a top-level function, a static method, or a local function, so it's a tearoff of anything-not-an-instance-method.

The best example is probably the tests: https://dart-review.googlesource.com/c/sdk/+/202243.

The current failures in those tests are mentioned in a comment from me. Here's a copy of the ones that are related to constants, for convenient access:

    103:  checkIdentical(cIntTopLevelFunction1, vIntTopLevelFunction2);
    104:  checkIdentical(cIntStaticMethod1, vIntStaticMethod2);
    105:  checkIdentical(vIntTopLevelFunction1, vIntTopLevelFunction2);
    106:  checkIdentical(vIntStaticMethod1, vIntStaticMethod2);

A complete example could be the following:

class CheckIdentical {
  const CheckIdentical(Object? o1, Object? o2): assert(identical(o1, o2));
}

X f<X>(X x) => x;

void main() {
  const int Function(int) g = f; // Generic function instantiation, "f<int>".
  const int Function(int) g2 = f; // Ditto.
  print(identical(g, g2)); // 'true' on the vm and dart2js.
  CheckIdentical(g, g2); // Succeeds on the analyzer, and elsewhere.

  // But canonicalization fails to occur when the constant
  // expression is not in a constant context.
  int Function(int) g3 = f;
  print(identical(g, g3)); // 'false' with `dart`, true with `dart2js`.
}

[johnniwinther]

Ah. We currently eagerly constantify static tearoffs - regardless of static context. We could just extend this to instantiations of those tearoffs when the type arguments are constant. We probably even need this for the constructor tearoffs.

So yes, this can be done in the cfe.

[eernstg]

Sounds good, thanks!

[eernstg]

@lrhn, given that the language team decided at the meeting June 23rd that function objects should be canonicalized, the basic question raised by this PR has been resolved. Do you have further comments on the PR?

[lrhn]

LGTM.
Do we need to address equality too?

[eernstg]

Thanks! Equality is specified, and the question is whether we'd want to ask for the specified behavior to be fully implemented, or we'd rather change the specification, cf. #1712. So we'll change the specification about equality in another PR, or we'll keep the current rules and ask for an implementation, and this PR should is separate from those efforts.