Empty intersection reduction

[ahejlsberg]

With this PR we reduce empty intersections to never immediately upon construction. For example, instead of preserving the type string & number and only reducing it away in union types we now immediately reduce it to never.

Some examples of types that now immediately reduce to never:

typeT1='a'&'b';// nevertypeT2={a: string}&null;// never in strict mode, null in non-strict modetypeT3={a: string}&undefined;// never in strict mode, undefined in non-strict modetypeT4=string&number;// nevertypeT5=number&object;// nevertypeT6=symbol&string;// nevertypeT7=void&string;// never

In #18438 we chose to preserve empty intersection types (a) to make it easier to understand how they originated and (b) to partially preserve the ability to "tag" primtive types using enum types (e.g. string & Tag where Tag is an enum). In practice this doesn't seem to outweigh the confusion these types bring about, in particular because there are subtle behavioral differences between never and types like string & number even though they really should be the same. The differences include:

• never is assignable to anything whereas string & number is assignable only to string or number.
• any is not assignable to never but is assignable to string & number.
• For a type T with no index signatures, T[never] resolves to never whereas T[string & number] is an error.

These differences have no practical value. Particularly confusing is that string & boolean reduces to never (because it normalizes to string & true | string & false, which then reduces to never), whereas string & number sticks around and behaves differently.

Fixes#31663.

[ahejlsberg]

@typescript-bot test this
@typescript-bot run dt

[typescript-bot]

Heya @ahejlsberg, I've started to run the parallelized Definitely Typed test suite on this PR at 015f4c5. You can monitor the build here. It should now contribute to this PR's status checks.

[typescript-bot]

Heya @ahejlsberg, I've started to run the extended test suite on this PR at 015f4c5. You can monitor the build here. It should now contribute to this PR's status checks.

[ahejlsberg]

RWC test run is clean. DT test run has one minor difference in ember/test/detect-instance.ts. Types such as null &{foo: string } and undefined &{foo: string } now reduce to never but previously would stick around (yet disappear in union types) and allow properties to be accessed. The new behavior is definitely more consistent.

[jack-williams]

Does this still support branding via object literals?

typeTag<T>=T&{readonlybrand: unique symbol};typeTaggedNumber=Tag<number>;// not never

[ahejlsberg]

@jack-williams It does, except you can't brand null or undefined. You sort of could before, but only in a limited sense. The minute you put null &{brand: xxx } in a union type it would disappear.

[ahejlsberg]

@jack-williams The only thing that changes with this PR is when we reduce away empty intersections. We now do it upon construction but we used to do it only when an intersection was put in a union type. The actual patterns we reduce away (or don't reduce away) remain the same.

[jack-williams]

The only thing that changes with this PR is when we reduce away empty intersections.

Ah, that makes it clearer. Thank you!

[weswigham]

Counterexample:

enumA{Zero=0,One=1}typeZeroish<T>=T&0;typeShouldNotBeNever=Zeroish<A.Zero>|never;

we handle this incorrectly in master already - I'm just not a fan of propagating the logical error.

[fatcerberus]

So is the reason for that that the code highlighted considers A.Zero and 0 to be distinct? I'm just trying to ascertain where in that example above the never creeps in.

[ahejlsberg]

@weswigham I think it is debatable whether we want to consider literals and literal enum members with the same underlying value to be intersectable. But either way, it is an issue that should be covered in a separate PR. Here we're just concerned with changing when we reduce, not how.

[ahejlsberg]

@typescript-bot run dt

[typescript-bot]

Heya @ahejlsberg, I've started to run the parallelized Definitely Typed test suite on this PR at 015f4c5. You can monitor the build here. It should now contribute to this PR's status checks.

[ahejlsberg]

RWC and DT test runs are both clean now.

[weswigham]

I expect we'll get some reports about the using-an-enum-to-tag break upon release - we should remember to include this in the breaking change log.

[dharnil]

Hi I wanted to check to see if this has already been resolved as I am still facing this issue.