Consider property access a form of type guards

This is a proposal for using property access as another form of type guards (see #900) to narrow union types. While we’re investigating expanding the power of type guards (#1007) this feature would support the natural style that JavaScript programmers have in their code today.

Using property access to narrow union types

var x: string|number;
if (x.length) { // no error even though number lacks a length property
    var r = x.charAt(3); // x is of type string inside this block
}

var r2 = x.length || x * x; // no error, x narrowed to number on right hand side, r3 is number

var y: string[]|number[]|string;
if (y.length && y.push) {
    // y is string[]|number[] now
    var first = y[0]; // first is string|number
    if (first.length) {
        // first is string in here
    } else {
        // first is number in here
    }
}

We do not expand the situations in which types are narrowed, but we do expand the known type guard patterns to include basic property access. In these narrowing contexts it would be an error to access a property that does not exist in at least one of the constituent types of a union type (as it is today). However, it would now be valid to access a property that exists in at least one, but not all, constituent types. Any such property access will then narrow the type of the operand to only those constituent types in the union which do contain the accessed property. In any other context property access is unchanged.

Invalid property access

var x: string|number;

var r = x.length; // error, not a type guard, normal property access rules apply

if (x.len) { } // error, len does not exist on type string|number 

var r3 = !x.len && x.length; // error, len does not exist on type string|number 

var r4 = x.length && x.len; // error, len does not exist on type string

Issues/Questions

• Should the language service behave differently in these type guard contexts? Should dotting off a union type in a type guard list all members on all types rather than only those that exist on all of them?
• Need to understand performance implications

I have a sample implementation and tests in a branch here: https://github.com/Microsoft/TypeScript/tree/typeGuardsViaPropertyAccess. There’re a couple bugs remaining but examples like the above all work and no existing behavior was changed.

[AlCalzone]

Since this is tagged “Awaiting more feedback”, I’ll just link to https://github.com/microsoft/TypeScript/issues/36194 as a potential use case.

[thorn0]

Feedback for you. Prettier is a formatter that works with ASTs that come from different parsers. In the case of JS-based languages, these AST formats are all based on the same ESTree base format, but also they have differences. It’d be great to be able to use a union type for this, and it looks like it might work (simplified code for two parsers only):

import * as Babel from "@babel/types";
import { TSESTree } from "@typescript-eslint/types";
type Node = Babel.Node | TSESTree.Node;

but in practice the resulting Node type is a pain to work with.

E.g., this already quite complex condition

    node.type === "MethodDefinition" &&
    node.value &&
    node.value.type === "FunctionExpression" &&
    getFunctionParameters(node.value).length === 0 &&
    !node.value.returnType &&

needs to be rewritten this way to make TS happy:

    node.type === "MethodDefinition" &&
    node.value &&
    node.value.type === "FunctionExpression" &&
    getFunctionParameters(node.value).length === 0 &&
    (!("returnType" in node.value) || !node.value.returnType)

The added in check doesn’t make code any clearer and only bloats it. The entire code base needs adding a lot of such checks, which is why any is used almost everywhere instead of Node.