Use type parameter constraints for computed property types #17404

sandersn · 2017-07-25T16:43:56Z

Note this code calls getBaseConstraintOfType the same way that getApparentType does. Calling getApparentType does not work here because it converts string to String and so on.

weswigham

@sandersn Is there a test for T extends number as well? It's probably the only thing that'd be nice to have if not present, since I think that would produce one fewer errors now? (like the changed U extends string test)

Aside from that remark, this looks like a relatively neat change. There might be a more general way to solve this and some of our other suggestions about things like intersection types and computed property names, but I think this is probably fine for this issue, and can be expanded upon later if need be.

ahejlsberg · 2017-07-25T22:46:14Z

Instead of this targeted change, it might make more sense to review all the uses of isTypeAnyOrAllConstituentTypesHaveKind function and replace them with calls to isTypeAssignableTo as appropriate. We have other similar issues, such as not being able to use arithmetic operators with T extends number values, and we might as well fix them all in one go.

This improves a number of error messages to do with adding 'null' or 'undefined' with other types. It also allows you to add two type parameters that both extend `number`.

sandersn · 2017-07-26T23:29:54Z

I replaced the body of isTypeOfKind (which isTypeAnyOrAllConstituentsTypesHaveKind immediately delegates to). The tests are passing but I still want to improve the any/null/undefined/void exclusion code since it's really the caller's responsibility.

sandersn · 2017-07-26T23:32:20Z

Note that this most recent commit improves the error message for expressions like null + undefined or null + 1 considerably, in addition to allowing this:

function f<T extends number>(t: T) {
  return t + t
}

Here, the return type of f is number. That's still hard-coded in the checker code for binary expressions.

ahejlsberg · 2017-07-28T20:28:21Z

src/compiler/checker.ts

-                        return true;
-                    }
-                }
+            const targets = [];


Why are you building up a union type here? It causes allocations and ends up being broken down to discrete checks for each contained type anyway. I think it would be better to just have a sequence of conditional isTypeAssignableTo calls here, e.g.

return kind & TypeFlags.NumberLike && isTypeAssignableTo(source, numberType) || kind & TypeFlags.StringLike && isTypeAssignableTo(source, stringType) || kind & TypeFlags.BooleanLike && isTypeAssignableTo(source, BooleanType) || ...

var p1: number | string var o = { [p1]() { } //~~~~ A computed property name must be of type 'string', 'number', ... }

But I think I can fix that by adding a fallback to isTypeAssignableTo the union type in just the case of computed properties.

Fixed. The computed property name type check now has a fallback call to isTypeAssignableTo(links.resolvedType, getUnionType([numberType, stringType, esSymbolType]). Everything else is fine with a sequence of isTypeAssignableTo calls.

ahejlsberg · 2017-07-28T20:28:48Z

src/compiler/checker.ts

-        // of the given kind.
-        function isTypeOfKind(type: Type, kind: TypeFlags): boolean {
-            if (type.flags & kind) {
+        function isTypeOfKind(source: Type, kind: TypeFlags, excludeAny?: boolean) {


I'm not sure I get why we need this excludeAny parameter?

checkBinaryLikeExpression has a complicated way of checking legal types for +:

It disallows null + number and number + undefined, even with strictNullChecks off. I like this check and want to keep it.

It gives a different types for string + any ==> string vs T + any ==> any and T + unknown ==> unknown.

It does this with the following (simplified) structure (using the abbreviation tak for isTypeAssignableToKind):

if (tak(left, NumberLike) && tak(right, NumberLike)) { return numberType; } else if (tak(left, StringLike) || tak(right, StringLike)) { return stringType; } else if (isTypeAny(left) || isTypeAny(right)) { return anyType; } else { returned undefined; // signal an error }

Without excludeAny (now called strict), isTypeAssignableTo(number, undefined) is true, and number + undefined/null is now legal. Also, isTypeAssignableTo(number, any) is true, and number + any ==> number , and any + any ==> number.

I tried a few things to pull the any checks before the calls to isTypeAssignableToKind, but was not successful. If you see a way, let me know.

ahejlsberg · 2017-07-28T20:35:11Z

@sandersn With the latest changes this also fixes #15501 and #17069.

simonbuchan · 2017-08-01T02:23:07Z

The errors also sound similar to the string enums as keys errors, eg #16687, #16760?

Instead of building up a list and creating a union type.

sandersn · 2017-08-02T17:30:36Z

@simonbuchan Your example #16687 has the computed property name error as a side effect, but this fix doesn't provide the type that you're looking for there.

…op-types

sandersn · 2017-08-02T17:40:48Z

This PR is ready to look at again.

rbuckton · 2017-08-08T20:21:29Z

src/compiler/checker.ts

            }
-            return false;
+            return kind & TypeFlags.NumberLike && isTypeAssignableTo(source, numberType) ||


Calling isTypeAssignableTo so many times seems fairly expensive if source is not a simple type. While the assignability check has a fast path for simple types via isSimpleTypeRelatedTo, if the relation doesn't hold we fall back to the more complex checkTypeRelatedTo function. I would recommend you determine if there is any additional check time overhead with this approach.

Also note that calls to isTypeAssignableToKind only ever pass 1-3 flags in, with 1 being the most common. So the expected perf hit would be just from calling isTypeAssignableTo once compared to the old logic.

sandersn · 2017-08-08T20:25:24Z

Sure, I'll test it, although @ahejlsberg was confident that his changes to isTypeAssignableTo would make this cheap enough.

…op-types

sandersn · 2017-08-08T21:15:29Z

Here are the ts-perf results (angular doesn't build on my installation of ts-perf). Total time goes down a tiny amount. Check time went up by 0.01s for TFS, and down 0.01s for Monaco. So I don't think there's any performance impact with this change.

┌─────────────┬─────────────────────┬─────────────────────┬───────────────────┬───────────┬────────────┐
│ Project     │            Baseline │             Current │             Delta │      Best │      Worst │
╞═════════════╧═════════════════════╧═════════════════════╧═══════════════════╧═══════════╧════════════╡
│ Monaco - node (v8.1.2, x64)                                                                          │
├─────────────┬─────────────────────┬─────────────────────┬───────────────────┬───────────┬────────────┤
│ Memory used │ 366,008k (±  0.00%) │ 366,041k (±  0.01%) │   +33k (+  0.01%) │  365,987k │   366,174k │
│ Parse Time  │    1.25s (±  0.70%) │    1.24s (±  0.28%) │ -0.00s (-  0.28%) │     1.23s │      1.25s │
│ Bind Time   │    0.60s (±  0.40%) │    0.60s (±  0.81%) │ +0.00s (+  0.75%) │     0.59s │      0.64s │
│ Check Time  │    3.10s (±  1.53%) │    3.09s (±  0.26%) │ -0.01s (-  0.48%) │     3.06s │      3.13s │
│ Emit Time   │    1.87s (±  0.85%) │    1.87s (±  0.48%) │ -0.00s (-  0.03%) │     1.84s │      1.91s │
├─────────────┼─────────────────────┼─────────────────────┼───────────────────┼───────────┼────────────┤
│ Total Time  │    6.82s (±  0.89%) │    6.80s (±  0.23%) │ -0.02s (-  0.25%) │     6.77s │      6.89s │
╞═════════════╧═════════════════════╧═════════════════════╧═══════════════════╧═══════════╧════════════╡
│ TFS - node (v8.1.2, x64)                                                                             │
├─────────────┬─────────────────────┬─────────────────────┬───────────────────┬───────────┬────────────┤
│ Memory used │ 319,284k (±  0.00%) │ 319,305k (±  0.00%) │   +21k (+  0.01%) │  319,282k │   319,363k │
│ Parse Time  │    0.91s (±  0.74%) │    0.91s (±  0.34%) │ -0.00s (-  0.16%) │     0.90s │      0.92s │
│ Bind Time   │    0.49s (±  3.23%) │    0.49s (±  2.17%) │ -0.01s (-  1.22%) │     0.47s │      0.58s │
│ Check Time  │    2.75s (±  0.94%) │    2.76s (±  0.81%) │ +0.01s (+  0.44%) │     2.58s │      2.83s │
│ Emit Time   │    1.77s (±  1.12%) │    1.76s (±  0.27%) │ -0.01s (-  0.82%) │     1.74s │      1.77s │
├─────────────┼─────────────────────┼─────────────────────┼───────────────────┼───────────┼────────────┤
│ Total Time  │    5.92s (±  0.36%) │    5.91s (±  0.22%) │ -0.01s (-  0.19%) │     5.84s │      5.95s │
└─────────────┴─────────────────────┴─────────────────────┴───────────────────┴───────────┴────────────┘

rbuckton

One minor comment, though I would probably wait to see if @ahejlsberg has any other further comments before merging.

rbuckton · 2017-08-08T20:23:44Z

src/compiler/checker.ts

@@ -17073,7 +17061,7 @@ namespace ts {
            // and the right operand to be of type Any, a subtype of the 'Function' interface type, or have a call or construct signature.
            // The result is always of the Boolean primitive type.
            // NOTE: do not raise error if leftType is unknown as related error was already reported
-            if (isTypeOfKind(leftType, TypeFlags.Primitive)) {
+            if (!(leftType.flags & TypeFlags.Any) && isTypeAssignableToKind(leftType, TypeFlags.Primitive)) {


We have an isTypeAny that is a bit more readable than the bitmask, e.g. !isTypeAny(leftType) && isTypeAssignableToKind(leftType, TypeFlags.Primitive)

sandersn · 2017-08-08T21:56:50Z

I'd like to get this into 2.5 RC so people can catch any lurking bugs. I'm on vacation starting Thursday, so let's leave this open until Friday to see if @ahejlsberg has a chance to look at this while he's gone too. I'll leave it up to you to merge or not at that point. @rbuckton.

I'll also try running ts-perf on a stress-test scenario that has lots of object type + string expressions.

rbuckton · 2017-08-08T21:57:22Z

I can see how isTypeAssignableTo has the same or better performance when the source type is simple enough (e.g. already a number, string, enum, or union as would generally be the case for our perf tests which have mostly correct code), but if the source type is a structured type or a type variable we end up in recursiveTypeRelatedTo which has 3 array allocations per each call to isTypeAssignableTo and will inflate the size of the assignability cache by 9 entries.

This may be acceptable overhead for cases where we are likely to fail, but something we should consider. It may be worthwhile for us to compose a performance benchmark against a scenario we tailor to be filled with incorrect code so that we can watch for performance regressions in those cases.

rbuckton · 2017-08-08T23:04:26Z

After reviewing this further with @sandersn, I'm not too concerned about the performance since we don't dig into target as it is a simple type. 👍

in isTypeAssignableToKind

sandersn added 2 commits July 25, 2017 09:38

Use type param constraints for computed prop types

43981ea

Test:use type param constraints for computed prop types

1fb6d34

msftclas added the cla-not-required label Jul 25, 2017

sandersn requested review from rbuckton and weswigham July 25, 2017 16:44

sandersn mentioned this pull request Jul 25, 2017

Generics and keyof does not work with computed properties #14473

Closed

weswigham approved these changes Jul 25, 2017

View reviewed changes

Make isTypeOfKind delegate to isTypeAssignableTo

838fbdd

This improves a number of error messages to do with adding 'null' or 'undefined' with other types. It also allows you to add two type parameters that both extend `number`.

ahejlsberg requested changes Jul 28, 2017

View reviewed changes

gcnew mentioned this pull request Aug 2, 2017

Typescript forbids safety checks that check equality of values with different generic types #17445

Closed

sandersn added 2 commits August 2, 2017 09:38

Improve name:isTypeAssignableToKind+cleanup TODOs

d07eca7

isTypeAssignableToKind:Chain isTypeAssignableTo calls

011f712

Instead of building up a list and creating a union type.

sandersn added 2 commits August 2, 2017 10:33

Add regression test cases and rename test

aed386c

Merge branch 'master' into use-type-param-constraints-for-computed-pr…

a0333b5

…op-types

rbuckton reviewed Aug 8, 2017

View reviewed changes

Merge branch 'master' into use-type-param-constraints-for-computed-pr…

e494d73

…op-types

rbuckton approved these changes Aug 8, 2017

View reviewed changes

Use isTypeAny instead of checking flags directly

e47df36

Add parentheses:clarify evaluation order of &&/||

fac93a3

in isTypeAssignableToKind

sandersn merged commit 847d7fe into master Aug 8, 2017

sandersn deleted the use-type-param-constraints-for-computed-prop-types branch August 8, 2017 23:29

sandersn mentioned this pull request Aug 17, 2017

Unable to add numeric types that are not simply 'number' #17069

Closed

weswigham mentioned this pull request Sep 5, 2017

Recreate old less safe for loop behavior #18262

Closed

microsoft locked and limited conversation to collaborators Jun 14, 2018

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Use type parameter constraints for computed property types #17404

Use type parameter constraints for computed property types #17404

sandersn commented Jul 25, 2017 •

edited

Loading

weswigham left a comment

ahejlsberg commented Jul 25, 2017 •

edited

Loading

sandersn commented Jul 26, 2017

sandersn commented Jul 26, 2017

ahejlsberg Jul 28, 2017

sandersn Aug 2, 2017 •

edited

Loading

sandersn Aug 2, 2017

ahejlsberg Jul 28, 2017

sandersn Aug 2, 2017

ahejlsberg commented Jul 28, 2017

simonbuchan commented Aug 1, 2017

sandersn commented Aug 2, 2017 •

edited

Loading

sandersn commented Aug 2, 2017

rbuckton Aug 8, 2017

sandersn Aug 8, 2017

sandersn commented Aug 8, 2017

sandersn commented Aug 8, 2017

rbuckton left a comment

rbuckton Aug 8, 2017

sandersn commented Aug 8, 2017

rbuckton commented Aug 8, 2017

rbuckton commented Aug 8, 2017 •

edited

Loading

Use type parameter constraints for computed property types #17404

Use type parameter constraints for computed property types #17404

Conversation

sandersn commented Jul 25, 2017 • edited Loading

weswigham left a comment

Choose a reason for hiding this comment

ahejlsberg commented Jul 25, 2017 • edited Loading

sandersn commented Jul 26, 2017

sandersn commented Jul 26, 2017

ahejlsberg Jul 28, 2017

Choose a reason for hiding this comment

sandersn Aug 2, 2017 • edited Loading

Choose a reason for hiding this comment

sandersn Aug 2, 2017

Choose a reason for hiding this comment

ahejlsberg Jul 28, 2017

Choose a reason for hiding this comment

sandersn Aug 2, 2017

Choose a reason for hiding this comment

ahejlsberg commented Jul 28, 2017

simonbuchan commented Aug 1, 2017

sandersn commented Aug 2, 2017 • edited Loading

sandersn commented Aug 2, 2017

rbuckton Aug 8, 2017

Choose a reason for hiding this comment

sandersn Aug 8, 2017

Choose a reason for hiding this comment

sandersn commented Aug 8, 2017

sandersn commented Aug 8, 2017

rbuckton left a comment

Choose a reason for hiding this comment

rbuckton Aug 8, 2017

Choose a reason for hiding this comment

sandersn commented Aug 8, 2017

rbuckton commented Aug 8, 2017

rbuckton commented Aug 8, 2017 • edited Loading

sandersn commented Jul 25, 2017 •

edited

Loading

ahejlsberg commented Jul 25, 2017 •

edited

Loading

sandersn Aug 2, 2017 •

edited

Loading

sandersn commented Aug 2, 2017 •

edited

Loading

rbuckton commented Aug 8, 2017 •

edited

Loading