luau/rfcs/syntax-nil-forgiving-operator.md

nil-forgiving postfix operator '!'

Summary

Introduce syntax to suppress typechecking errors for nilable types by ascribing them into a non-nil type.

Motivation

Typechecking might not be able to figure out that a certain expression is a non-nil type, but the user might know additional context of the expression.

Using '::' ascriptions is the current work-around for this issue, but it's much more verbose and requires providing the full type name when you only want to ascribe T? to T.

The nil-forgiving operator will also allow chaining to be written in a very terse manner:

local p = a!.b!.c

instead of

local ((p :: Part).b :: Folder).c

Note that nil-forgiving operator is not a part of member access operator, it can be used in standalone expressions, indexing and other places:

local p = f(a!)!
local q = b!['X']

Nil-forgiving operator can be found in some programming languages such as C# (null-forgiving or null-suppression operator) and TypeScript (non-null assertion operator).

Design

To implement this, we will change the syntax of the primaryexp.

Before:

primaryexp = prefixexp { '.' NAME | '[' exp ']' | ':' NAME funcargs | funcargs }

After:

postfixeexp = ('.' NAME | '[' exp ']' | ':' NAME funcargs | funcargs) ['!']
primaryexp = prefixexp ['!'] { postfixeexp }

When we get the '!' token, we will wrap the expression that we have into a new AstExprNonNilAssertion node.

An error is generated when the type of expression node is one of the following:

• AstExprConstantBool
• AstExprConstantNumber
• AstExprConstantString
• AstExprFunction
• AstExprTable

This operator doesn't have any impact on the run-time behavior of the program, it will only affect the type of the expression in the typechecker.

Parsing of assignments

An assignment expression starts with a primaryexp and we perform a check that it is a valid l-value based on a fixed set of AstNode types.

Since using '!' on an l-value has no effect, we don't extend this list with the new node and will generate a specialized parse error for code like:

p.a! = b

Handling of union types

When operator is used on expression of a union type with a nil option, it removes that option from the set. If only one option remains, the union type is replaced with the type of a single option.

If the type is nil, typechecker will generate a warning that the operator cannot be applied to the expression.

For any other type, it has no effect and doesn't generate additional warnings.

The reason for the last rule is to simplify movement of existing code where context in each location is slightly different.

As an example from Roblox, instance path could dynamically change from being known to exist to be missing when script is changed in edit mode.

Handling of type packs

When operator is used on expression that is a type pack ('...' or a function call in tail position), we will perform nil option filtering for each member of the type pack based on the rules for types.

If the type pack is variadic or has a variadic tail, we will perform nil option filtering for the type inside VariadicTypePack.

Drawbacks

Unnecessary operator use

It might be useful to warn about unnecessary uses of this operator when the value cannot be nil, but we have no way of enabling this behavior.

Bad practice

The operator might be placed by users to ignore/silence correct warnings and lower the strength of type checking that Luau provides.

Alternatives

Test with assert call

Aside from type assertion operator '::', it should be possible to place assert function calls before the operation. Type refinement/constraints should handle that statement and avoid warning in the following expressions.

But assert call will introduce runtime overhead without adding extra safety to the case when the type is nil at run time, in both cases an error will be thrown.

Extending function definition syntax

In a function definition like function x.f() we could've extended the syntax to allow function x!.f() to silence the analysis warning in case x is optional.

But this would complicate the parsing and analysis of function definition statements and the feature complexity doesn't seem to be worth it in this particular case.