Type guard not applied based on indexed access with string literal

[dalen]

TypeScript Version: 3.1.0-dev.20180804

Search Terms:
dash, bracket

Code

Very similar to documentation example code, but with a discriminant containing a dash. Likely not because of the dash itself, but because the lookup is using s['kind-of'] instead of s.kind.

interface Square {
  'kind-of': 'square';
  size: number;
}

interface Rectangle {
  'kind-of': 'rectangle';
  width: number;
  height: number;
}
type Shape = Square | Rectangle;

function area(s: Shape): number {
  if (s['kind-of'] === 'square') {
    // Now TypeScript *knows* that `s` must be a square ;)
    // So you can use its members safely :)
    return s.size * s.size;
  } else {
    // Wasn't a square? So TypeScript will figure out that it must be a Rectangle ;)
    // So you can use its members safely :)
    return s.width * s.height;
  }
}

Expected behavior:

No errors

Actual behavior:

test.ts:17:14 - error TS2339: Property 'size' does not exist on type 'Shape'.
  Property 'size' does not exist on type 'Rectangle'.

17     return s.size * s.size;
                ~~~~

test.ts:17:23 - error TS2339: Property 'size' does not exist on type 'Shape'.
  Property 'size' does not exist on type 'Rectangle'.

17     return s.size * s.size;
                         ~~~~

test.ts:21:14 - error TS2339: Property 'width' does not exist on type 'Shape'.
  Property 'width' does not exist on type 'Square'.

21     return s.width * s.height;
                ~~~~~

test.ts:21:24 - error TS2339: Property 'height' does not exist on type 'Shape'.
  Property 'height' does not exist on type 'Square'.

21     return s.width * s.height;
                          ~~~~~~

My real world use case is to check type of AWS Cloudwatch event using the detail-type field on the event, so can't just change the name to get around the restriction easily.

[RyanCavanaugh]

It's the use of bracket notation (x["a"]) rather than dot (x.a) that causes this

[ahejlsberg]

@sandersn already looked at this, I think there is a PR still open. The issue is that supporting indexed access (x["foo"]) as well as property access (x.foo) in control flow analysis has significant negative performance impact because we need to include many more nodes (i.e. all indexed access operations) in the control flow graph.

[sandersn]

It was #10565. It might not be too hard to revive this PR and see how bad the slowdown is now, but I doubt it will be much different.

[RyanCavanaugh]

I'm surprised a syntactic check could be that expensive. There doesn't seem to be much going on in that PR that wasn't happening before and the number of indexed-by-string-literal property accesses in a given program shouldn't be that large?

[benneq]

Could you at least add this feature with some feature flag in tsconfig?

I don't care if compilation becomes slower. But I do care about type errors. That's what this language is all about.

[sandersn]

This should be fixed now.