Improve checking of in operator combined with noUncheckedIndexedAccess

[Andarist]

follow up to #50666 , mostly piggy-backing on the logic introduced there

fixes #43614

cc @ahejlsberg @andreialecu

[JakobBruenker]

Thanks! This is exactly what's preventing us from using noUncheckedIndexedAccess, so I'm looking forward to seeing this merged

[gabritto]

Small thing missing, but other than that, it looks good to me. Thanks 😊

[gabritto]

I think you should only do this if assumeTrue is true (maybe have the check above be if (!compilerOptions.noUncheckedIndexedAccess || !assumeTrue) ...).

Might be nice to add that to the tests as well, something like:

function f2(obj: Record<string, string>) {
  if ("test" in obj) {
    return obj.test
  }
  else {
    obj.test // Here it should be `string | undefined` still
  }
  return "default"
}

[Andarist]

great catch! thanks. I've pushed out the fix for this

[gabritto]

@ahejlsberg could you confirm we're ok with the solution proposed here of adding extra intersections with Record<propName, propType> during narrowing when noUncheckedIndexedAccess is true?

[ahejlsberg]

First, I very much like the idea of narrowing indexed accesses guarded by in checks since we already do that for optional properties in --exactOptionalPropertyTypes mode. But, by that same reasoning, I think it should be done through the same mechanism as --exactOptionalPropertyTypes, i.e. use of the special missingType that indicates an undefined originating in a missing property. Specifically, we should use missingType as the undefined that is added when reading from an index signature and then rely on our existing CFA infrastructure to remove that missingType in code guarded by an in check. This would also get us support for narrowing through obj.hasOwnProperty("foo") checks, which is already in place for optional properties.

I realize that it is possible to get a similar effect using intersections, but it leads to more complex types and is oddly inconsistent with what we do for optional properties. And, really, index signatures are just another form of optional properties.

I will put up a PR in the next day or two to show what I mean. We may need to do a bit of optimizing of the logic that detects multiple kinds of undefined in unions and intersections, but we'll see what the performance suites reveal.

[Andarist]

Closing as superseded by #51653