Different behavior of CFE and the analyzer when constant depends on itself

[sgrekhov]

The following code produces an error in the analyzer but no issues in VM.

const int? one = 2 > 1 ? null : one; // Analyzer error: The compile-time constant expression depends on itself

main() {
  print(one); // prints null
}

If to change the example to really invoke the constant then VM produces an error as well.

const int? two = 2 > 1 ? two: null; // Error: Constant evaluation error. Context: Constant expression depends on itself.

main() {
  print(two);
}

@eernstg @lrhn does in the first case the constant really depends on itself?

[dart-github-bot]

Summary: Analyzer detects self-referential constant one, but the VM doesn't, though a modified example (two) triggers VM error. Inconsistency needs investigation.

[mraleph]

cc @eernstg

[eernstg]

The language team decided that cyclic dependencies should be computed based on a simple syntactic criterion: If a non-local variable declaration D1 has an initializing expression e where the name that resolves to a declaration D2 occurs then D1 depends on D2. A compile-time error occurs if the dependency graph has a cycle.

This means that a syntactically cyclic set of declarations will be reported as a compile-time error even in the case where some analysis can prove that no infinite loop will occur during initialization.

The implementation was handled in #50383. There was no need to change the specification.

We could try to introduce special rules about constant expressions because they have a control flow which is known at compile time (or link time, depending on the treatment of some constructors whose name includes fromEnvironment), but this should be done in a new issue in the language repository.

So I'd say that the analyzer is right and the CFE should be adjusted such that all syntactic circularities are reported. I changed the labels accordingly.