Allow == null to narrow unknown to null | undefined

[ericyhwang]

TypeScript Version: typescript@3.6.0-dev.20190809

Search Terms: unknown, type guard, double-equals, ==, null

Code

Minimal repro case:

export function test(v: unknown) {
  let nullOrUndefined: null | undefined;
  if (v === null || v === undefined) {
    // This works, as `v` is narrowed to `null | undefined`.
    nullOrUndefined = v;
  }
  if (v == null) {
    // Error: Type 'unknown' is not assignable to type 'null | undefined'.
    nullOrUndefined = v;
  }
}

A more practical use-case for the narrowing would be something like this:

if (typeof v === 'string' || v == null) {
  // v should be `string | null | undefined` in here.
  if (v) {
    // v is a non-empty string
    doSearch(v);
  } else {
    // v is empty string, null, or undefined
    clearSearch();
  }
}

Expected behavior:

v == null narrows an unknown type to null | undefined, same as v === null || v === undefined does.

Reading the ECMAScript spec on ==, I believe that == null will guarantee that the value is either null or undefined. == undefined would do the same thing, though linters generally prefer the shorter == null.

Actual behavior:

Compile error:

src/index.ts:11:5 - error TS2322: Type 'unknown' is not assignable to type 'null | undefined'.
  Type 'unknown' is not assignable to type 'null'.

11     nullOrUndefined = v;
       ~~~~~~~~~~~~~~~

Found 1 error.

Playground Link: https://www.typescriptlang.org/play/#code/KYDwDg9gTgLgBAMwK4DsDGMCWEVxsAZxgAoA3ALjlQGsUIB3FASjgG8AoOOAG2HhSTduAeSgBVFABNgCTCmCTKAoXAA+VKTLkKA3JziYEcMnAC85uMu5r1pMxdTTZ8ySw5cuAek9wAKgAtMAjh6aGoCADp9LitRCSdtSTM4Uj0uAF99Q2M7Cys3aLhvPwBPMGA4AHIaOkZKg2C6eABDAgJMAHMUZoAjXjwIPDKKyqs1DQSXSqiPS0ERcU1nBWTU-UzMoA

Related Issues:

• Triple-equal narrowing of unknown was added in Fix #25172: Add narrowing for unknown with triple equals #26941
  • The PR specifically noted "Does not work with == because we would need to consider coercion rules."
  • That's true for the general case. Thankfully, the behavior of == null is well-defined and can be special-cased.
• There's a general issue Allow more constructs to work as type guards for unknown #25720 open for better narrowing of unknown.

[ericyhwang]

I'd be interested in submitting a PR for this, by the way.

Looking at checker.ts, narrowTypeByEquality already has handling for == null. Without having cloned down the code for testing, my guess: Perhaps getTypeWithFacts might not be narrowing down unknown based on TypeFacts.EQUndefinedOrNull?

[jack-williams]

@ericyhwang I think looking at the changes associated with #26941 is a good place to start—that added a new code path for unknown and ===. The right fix might be to split the code path for === and ==.

[jack-williams]

@ericyhwang Are you working on this? I think it's a relatively simple fix, but it would also be nice for you to get a contributor badge.

Do you want to sketch out the solution here together and then you can open a PR?

[ericyhwang]

@jack-williams Ah, I thought bugs needed a help-wanted label first. Haven't started beyond initially looking at the code.

I'm still interested, though I may not have work-week time for a few weeks. If it's pretty simple, I might treat it like a mini weekend project and do it on my own time.

[jack-williams]

Oh yeah, forgot about that label. Well I'm sure @RyanCavanaugh would be happy to accept a PR on this given that it's on the backlog.

I'll leave it in your hands, but if you have any Q's just drop them in this issue.

[lkho]

hi everyone, I faced this same compilation error TS2322 (without strictNullChecks)

Playground:
https://www.typescriptlang.org/play/?strictNullChecks=false&ts=3.8-Beta#code/GYVwdgxgLglg9mABMAjACgB4C5EGcoBOMYA5gJQ5ggC2ARgKYGIDeAUIojMIpogLx9EVADbCyLdh0QF6UEASQYA3JIC+kmXIVCQoletahIsBMgBMmHPiKlEAHx2iKj4RI5ceGfgJfi2U6Vl5RRUOdQ5NYJd9ViA

Is it the same problem discussing in this thread?

Compilation error:
line 3:

(parameter) x: string
Type 'string' is not assignable to type 'number'.(2322)

line 10:

(parameter) x: string
Type 'string' is not assignable to type 'null'.(2322)

However, I saw from the AST viewer, both computed type of x at line 3 is never and line 10 is null, which are both correct (if strictNullChecks).