Unexpected ClassCastException upon atom field access when intersection types and warnings are combined

[radeusgd]

I was hitting an unexpected ClassCastException when running Table tests during development of my #12165 PR.

I was able to distill the issue into a small repro:

from Standard.Base import all

type My_Atom
    Value my_field

type My_Refinement
    Value x

My_Refinement.from (that : My_Atom) -> My_Refinement =
    My_Refinement.Value that

prepare_atom =
    x = Warning.attach "warn" <|
        My_Atom.Value 42
    x : My_Atom & My_Refinement

main =
    x = prepare_atom
    IO.println x.my_field

Expected behaviour

The code should just print:

42

Actual behaviour

The interpreter crashes with ClassCastException:

Execution finished with an error: java.lang.ClassCastException: class org.enso.interpreter.runtime.warning.WithWarnings cannot be cast to class org.enso.interpreter.runtime.data.atom.Atom (org.enso.interpreter.runtime.warning.WithWarnings and org.enso.interpreter.runtime.data.atom.Atom are in module org.enso.runtime of loader 'app')
        at <java> org.enso.runtime/org.enso.interpreter.runtime.data.atom.GetFieldNode.execute(GetFieldNode.java:38)
        at <enso> my_field(Internal)
        at <enso> err.main<arg-1>(err.enso:19:16-25)
        at <enso> err.main(err.enso:19:5-25)

[radeusgd]

Context that I had to find when trying to figure out a repro smaller than running the whole Table_Tests suite with my changes.

enso/engine/runtime/src/main/java/org/enso/interpreter/runtime/data/atom/GetFieldNode.java

Lines 36 to 40 in 4cae524

	public Object execute(VirtualFrame frame) {
	// this is safe, as only Atoms will ever get here through method dispatch.
	Atom atom = (Atom) Function.ArgumentsHelper.getPositionalArguments(frame.getArguments())[0];
	return structs.getField(atom, index);
	}

The (Atom) cast in GetFieldNode assumes that the thing it got is an Atom but in this case it is a WithWarnings instance.

The assumption seems to be broken because InvokeMethodNode::doMultiValue delegates to InvokeFunctionNode, extracting the first value from the multi value, but it does not extract warnings. If this was not a multi value, the doWarning specialization would have removed the warnings. Perhaps we need doMultiValue to somehow delegate to doWarning?

[JaroslavTulach]

both EnsoMultiValue and WithWarnings are a kind of 'enhance a value' wrapper and while a single wrapper like this is ok having two starts to show possibly unexpected interactions between them

• Yes. That's exactly our problem.
• I was thinking just of defining order among the values.
• Just another assert at

enso/engine/runtime/src/main/java/org/enso/interpreter/runtime/data/EnsoMultiValue.java

Line 94 in fe49064

assert !Stream.of(values).anyMatch(v -> v instanceof EnsoMultiValue)

• to prevent WithWarnings to be wrapped by EnsoMultiValue
• remove warnings from the values and keep the warnings only for the whole multivalue
• one has to make sure warnings are reattached when any of the multi values is "extracted from it"