Skip to main content

Nullness Specification (draft)

This document is a draft specification for the precise semantics of our set of annotations for nullness analysis.

Advice to readers (non-normative)

The primary audience for this document is the authors of analysis tools. Some very advanced users might find it interesting. But it would make a very poor introduction for anyone else; instead see our Start Here page.

Status of this specification

This document is current as of JSpecify 0.2.0, but does not reflect several design changes between then and 0.3.0.

The word "nullable"​

In this doc, we aim not to refer to whether a type "is nullable." Instead, we draw some distinctions, creating at least four kinds of "Is it nullable?" questions we can ask for any given type usage:

  1. Does @Nullable appear directly on that type usage?
  2. What is the nullness operator of that type usage?
  3. Is it reasonable to assume that null won't come "out" of it?
  4. Is it reasonable to assume that null can't be put "in" to it?

The scope of this spec​

Currently, this spec does not address when tools must apply any part of the spec. For example, it does not state when tools must check that the subtyping relation holds.

We anticipate that tools will typically apply parts of this spec in the same cases that they (or javac) already apply the corresponding parts of the Java Language Specification. For example, if code contains the parameterized type List<@Nullable Foo>, we anticipate that tools will check that @Nullable Foo is a subtype of the bound of the type parameter of List.

However, this is up to tool authors, who may have reasons to take a different approach. For example:

  • Java places some restrictions that aren't necessary for soundness, and it is lenient in at least one way that can lead to runtime errors.

  • JSpecify annotations can be used even by tools that are not "nullness checkers" at all. For example, a tool that lists the members of an API could show the nullness of each type in the API, without any checking that those types are "correct."

  • Even when a tool is a "nullness checker," it might be written for another language, like Kotlin, with its own rules for when to perform type checks. Or the tool might target a future version of Java whose language features would not be covered by this version of this spec.

Note also that this spec covers only nullness information from JSpecify annotations. Tools may have additional sources of information. For example, a tool may recognize additional annotations. Or a tool may omit the concept of UNSPECIFIED and apply a policy that type usages like Object are always non-nullable.

That's all!​

On to the spec.

Normative and non-normative sections​

This document contains some non-normative comments to emphasize points or to anticipate likely questions. Those comments are set off as block quotes.

This is an example of a non-normative comment.

This document also links to other documents. Those documents are non-normative, except for when we link to the Java Language Specification to defer to its rules.

Relationship between this spec and JLS​

When a rule in this spec refers to any concept that is defined in this spec (for example, substitution or containment), apply this spec's definition (as opposed to other definitions, such as the ones in the Java Language Specification (JLS)).

Additionally, when a rule in this spec refers to a JLS rule that in turn refers to a concept that is defined in this spec, likewise apply this spec's definition.

In particular, when a JLS rule refers to types, apply this spec's definition of augmented types (as opposed to base types).

Base type​

A base type is a type as defined in JLS 4.

JLS 4 does not consider a type-use annotation to be a part of the type it annotates, so neither does our concept of "base type."

Type components​

A type component of a given type is a type that transitively forms some part of that type. Specifically, a type component is one of the following:

  • a non-wildcard type argument
  • a wildcard bound
  • an array component type
  • an enclosing type
  • an element of an intersection type
  • the entire type

Nullness operator​

A nullness operator is one of 4 values:

  • UNION_NULL
  • NO_CHANGE
  • UNSPECIFIED
  • MINUS_NULL

The informal meaning of the operators is:

  • UNION_NULL: This is the operator produced by putting @Nullable on a type usage.
    • The type usage String UNION_NULL includes "a", "b", "ab", etc., plus null.
    • The type-variable usage T UNION_NULL includes all members of T, plus null if it wasn't already included.
  • NO_CHANGE: This is the operator produced by not putting @Nullable on a type usage (aside from the exception discussed under UNSPECIFIED below).
    • The type usage String NO_CHANGE includes "a", "b", "ab", etc., without including null.
    • The type-variable usage T NO_CHANGE includes exactly the members of T: If null was a member of T, then it's a member of T NO_CHANGE. If it was not a member of T, then it is not a member of T NO_CHANGE.
    • One way to conceptualize this is that String NO_CHANGE means "non-null String" but that T NO_CHANGE means "nullness comes from the value of T."
  • UNSPECIFIED: This is the operator produced by not putting @Nullable on a type usage in code that is outside a null-marked scope. Roughly, it is the operator assigned to "completely unannotated code."
    • The type usage String UNSPECIFIED includes "a", "b", "ab", etc., but whether null should be included is not specified.
    • The type-variable usage T UNSPECIFIED includes all members of T. But whether null should be added to the set (if it isn't already) is not specified.
  • MINUS_NULL: This operator not only does not add null but also actively removes it from a type-variable usage that might otherwise include it.
    • The type usage String MINUS_NULL includes "a", "b", "ab", etc., without including null. (This is equivalent to String NO_CHANGE.)
    • The type-variable usage T MINUS_NULL includes all members of Texcept for null. (This is equivalent to T NO_CHANGE unless null was a member of T.)

Augmented type​

An augmented type consists of a base type and a nullness operator corresponding to each of its type components.

Arguably, an augmented type with nullness operator UNSPECIFIED is better understood not as representing "a type" but as representing a lack of the nullness portion of the type.

For our purposes, base types (and thus augmented types) include not just class and interface types, array types, and type variables but also intersection types and the null type.

This spec aims to define rules for augmented types compatible with those that the JLS defines for base types.

Accordingly, in almost all cases, this spec agrees with the JLS's rules when specifying what base types appear in a piece of code. It makes an exception for "Bound of an unbounded wildcard," for which it specifies a bound of Object that the JLS does not specify.

When this spec uses capital letters, they refer to augmented types (unless otherwise noted). This is in contrast to the JLS, which typically uses them to refer to base types.

When this spec refers to "the nullness operator of" a type T, it refers specifically to the nullness operator of the type component that is the entire type T, without reference to the nullness operator of any other type that is a component of T or has T as a component.

For example, "the nullness operator of List<Object>" refers to whether the list itself may be null, not whether its elements may be.

Details common to all annotations​

For all named annotations referred to by this spec:

  • The package name is org.jspecify.nullness. [#260]
  • The Java module name is org.jspecify. [#181]
  • The Maven artifact is org.jspecify:jspecify. [#181]

All annotations have runtime retention. [#28] None of the annotations are marked repeatable.

The type-use annotation​

We provide a parameterless type-use annotation called @Nullable.

Recognized locations for type-use annotations​

A location is a recognized location for our type-use annotation in the circumstances detailed below. This spec does not define semantics for annotations in other locations.

For now, we've chosen to restrict ourselves to API locations for which tools mostly agree on what it means for a type in that location to be @Nullable.

When analyzing source code, tools are encouraged to offer an option to issue an error for an annotation in an unrecognized location (unless they define semantics for that location). Tools are especially encouraged to issue an error for an annotation in a location that is "intrinsically non-nullable" (defined below).

When reading bytecode, however, tools may be best off ignoring an annotation in an unrecognized location (again, unless they define semantics for that location).

The following locations are recognized except when overruled by one of the exceptions in the subsequent sections: [#17]

  • return type of a method

  • formal parameter type of a method or constructor, as defined in JLS 8.4.1

    This excludes the receiver parameter.

  • field type

  • type parameter upper bound [#60]

  • non-wildcard type argument

  • wildcard bound

  • array component type

  • type used in a variadic parameter declaration

However, any location above is unrecognized if it matches either of the following cases: [#17]

We refer to these cases (and some other cases below) as "intrinsically non-nullable."

  • a type usage of a value type (currently, the 8 predefined primitive types)

  • the outer type that qualifies an inner type

    For example, the annotation in @Nullable Foo.Bar is in an unrecognized location: Java syntax attaches it to the outer type Foo.

    (Note that @Nullable Foo.Bar is a Java syntax error when Bar is a static type. If Bar is a non-static type, then Java permits the code. So JSpecify tools have the oppotunity to reject it, given that the author probably intended Foo.@Nullable Bar.)

    Every outer type is intrinsically non-nullable because every instance of an inner class has an associated instance of the outer class.

Additionally, any location above is unrecognized if it makes up any type component of a type in the following locations: [#17]

These locations all fit under the umbrella of "implementation code." Implementation code may use types that contain type arguments, wildcard bounds, and array component types, which would be recognized locations if not for the exceptions defined by this section.

  • a local variable type
  • an exception parameter
  • the type in a cast or instanceof expression
  • an array or object creation expression (including via a member reference)
  • an explicit type argument supplied to a generic method or constructor (including via a member reference) or to an instance creation expression for a generic class

In practice, we anticipate that tools will treat types (and their annotations) in most of the above locations much like they treat types in other locations. Still, this spec does not concern itself with implementation code: We believe that the most important domain for us to focus on is that of APIs.

All locations that are not explicitly listed as recognized are unrecognized.

Other notable unrecognized annotations include: [#17]

Some additional intrinsically non-nullable locations:

  • supertype in a class declaration
  • thrown exception type
  • enum constant declaration
  • receiver parameter type

Some other locations that individual tools are more likely to assign semantics to:

  • a class declaration [#7]: For example, the annotation in public @Nullable class Foo {} is in an unrecognized location.
  • a type-parameter declaration or a wildcard itself [#19, #31]
  • any type component of a receiver parameter type [#157]

But note that types "inside" some of these locations can still be recognized, such as a type argument of a supertype.

The declaration annotation​

We provide a single parameterless declaration annotation called @NullMarked. [#5, #87]

Recognized locations for declaration annotations​

Our declaration annotation is specified to be recognized when applied to the locations listed below:

  • A named class.
  • A package. [#34]
  • A module. [#34]

Not a method [#43], constructor [#43], or field [#50].

Null-marked scope​

To determine whether a type usage appears in a null-marked scope:

Look for a @NullMarked annotation on any of the scopes enclosing the type usage.

Class members are enclosed by classes, which may be enclosed by other class members or classes. and top-level classes are enclosed by packages, which may be enclosed by modules.

Packages are not enclosed by "parent" packages.

This definition of "enclosing" likely matches the definition in the Java compiler API.

If one of those scopes is directly annotated with @NullMarked, then the type usage is in a null-marked scope. Otherwise, it is not.

Augmented type of a type usage appearing in code​

For most type usages in source code or bytecode on which JSpecify nullness annotations are recognized, this section defines how to determine their augmented types. Note, however, that rules for specific cases below take precedence over the general rule here.

Because the JLS already has rules for determining the base type for a type usage, this section covers only how to determine its nullness operator.

To determine the nullness operator, apply the following rules in order. Once one condition is met, skip the remaining conditions.

  • If the type usage is annotated with @Nullable, its nullness operator is UNION_NULL.
  • If the type usage appears in a null-marked scope, its nullness operator is NO_CHANGE.
  • Its nullness operator is UNSPECIFIED.

The choice of nullness operator is not affected by any nullness operator that appears in a corresponding location in a supertype. For example, if one type declares a method whose return type is annotated @Nullable, and if another type overrides that method but does not declare the return type as @Nullable, then the override's return type will not have nullness operator UNION_NULL.

The rules here never produce the fourth nullness operator, MINUS_NULL. However, if tool authors prefer, they can safely produce MINUS_NULL in any case in which it is equivalent to NO_CHANGE. For example, there is no difference between String NO_CHANGE and String MINUS_NULL.

So why does MINUS_NULL exist at all? It does appear later in this spec in the section on substitution. However, its main purpose is to provide tools with a way to represent the nullness of certain expressions in implementation code: Consider ArrayList<E>. ArrayList supports null elements, so the class has to handle the possibility that any expression of type E may be null. However, if implementation code contains the statement if (e != null) { ... }, then tools can assume that e is non-null inside. The purpose of MINUS_NULL is to represent that such an expression is known not to be null, even though its base type E suggests otherwise.

Augmented type of an intersection type​

Technically speaking, the JLS does not define syntax for an intersection type. Instead, it defines a syntax for type parameters and casts that supports multiple types. Then the intersection type is derived from those. Intersection types can also arise from operations like capture conversion. See JLS 4.9.

One result of this is that it's never possible for a programmer to write an annotation "on an intersection type."

This spec assigns a nullness operator to each individual element of an intersection type, following our normal rules for type usages. It also assigns a nullness operator to the intersection type as a whole. The nullness operator of the type as a whole is always NO_CHANGE.

This lets us provide, for every base type, a rule for computing its augmented type. But we require NO_CHANGE so as to avoid questions like whether "a UNION_NULL intersection type whose members are Foo UNION_NULL and Bar UNION_NULL" is a subtype of "a NO_CHANGE intersection type with those same members." Plus, it would be difficult for tools to output the nullness operator of an intersection type in a human-readable way.

To avoid ever creating an intersection type with a nullness operator other than NO_CHANGE, we define special handling for intersection types under "Applying a nullness operator to an augmented type."

Bound of an "unbounded" wildcard​

In source, an unbounded wildcard is written as <?>. This section does not apply to <? extends Object>, even though that is often equivalent to <?>.

See JLS 4.5.1.

In bytecode, such a wildcard is represented as a wildcard type with an empty list of upper bounds and an empty list of lower bounds. This section does not apply to a wildcard with any bounds in either list, even a sole upper bound of Object.

For a wildcard with an explicit bound of Object (that is, <? extends Object>, perhaps with an annotation on Object), instead apply the normal rules for the explicit bound type.

If an unbounded wildcard appears in a null-marked scope, then it has a single upper bound whose base type is Object and whose nullness operator is UNION_NULL.

If an unbounded wildcard appears outside a null-marked scope, then it has a single upper bound whose base type is Object and whose nullness operator is UNSPECIFIED.

In both cases, we specify a bound that does not exist in the source or bytecode, deviating from the JLS. Because the base type of the bound is Object, this should produce no user-visible differences except to tools that implement JSpecify nullness analysis.

Whenever a JLS rule refers specifically to <?>, disregard it, and instead apply the rules for <? extends T>, where T has a base type of Object and the nullness operator defined by this section.

Bound of an Object-bounded type parameter​

In source, an Object-bounded type parameter can be writen in either of 2 ways:

  • <T>
  • <T extends Object> with no JSpecify nullness type annotations on the bound

See JLS 4.4.

In bytecode, <T> and <T extends Object> are both represented as a type parameter with a single upper bound, Object, and no JSpecify nullness type annotations on the bound.

If an Object-bounded type parameter appears in a null-marked scope, then its bound has a base type of Object and a nullness operator of NO_CHANGE.

Note that this gives <T> a different bound than <?> (though only in a null-marked scope).

If an Object-bounded type parameter appears outside a null-marked scope, then its bound has a base type of Object and a nullness operator of UNSPECIFIED.

All these rules match the behavior of our normal rules for determining the augmented type of the bound Object. The only "special" part is that we consider the source code <T> to have a bound of Object, just as it does when compiled to bytecode.

Augmented null types​

The JLS refers to "the null type." In this spec, we assign a nullness operator to all types, including the null type. This produces multiple null types:

  • the null base type with nullness operator NO_CHANGE: the "bottom"/"nothing" type used in capture conversion

    No value has this type, not even null itself.

  • the null base type with nullness operator MINUS_NULL

    This is equivalent to the previous type. Tools may use the 2 interchangeably.

  • the null base type with nullness operator UNION_NULL: the type of the null reference

  • the null base type with nullness operator UNSPECIFIED

    This may be relevant only in implementation code.

Multiple "worlds"​

Some of the rules in this spec come in 2 versions: One version requires a property to hold "in all worlds," and the other requires it to hold only "in some world."

Tool authors may choose to implement neither, either, or both versions of the rules.

Our goal is to allow tools and their users to choose their desired level of strictness in the presence of UNSPECIFIED. The basic idea is that, every time a tool encounters a type component with the nullness operator UNSPECIFIED, it has the option to fork off 2 "worlds": 1 in which the operator is UNION_NULL and 1 in which it is NO_CHANGE.

In more detail: When tools lack a nullness specification for a type, they may choose to assume that either of the resulting worlds may be the "correct" specification. The all-worlds version of a rule, by requiring types to be compatible in all possible worlds, holds that types are incompatible unless it has enough information to prove they are compatible. The some-world version, by requiring types to be compatible only in some world, holds that types are compatible unless it has enough information to prove they are incompatible. (By behaving "optimistically," the some-world version is much like Kotlin's rules for "platform types.")

Thus, a strict tool might choose to implement the all-worlds version of rules, and a lenient tool might choose to implement the some-world version. Yet another tool might implement both and let users select which rules to apply.

Still another possibility is for a tool to implement both versions and to use that to distinguish between "errors" and "warnings." Such a tool might always first process code with the all-worlds version and then with the some-world version. If the tools detects, say, an out-of-bounds type argument in both cases, the tool would produce an error. But, if the tool detects such a problem with the all-worlds version but not with the some-world version, the tool would produce a warning. Under this scheme, a warning means roughly that "There is some way that the code could be annotated that would produce an error here."

The main body of each section of the spec describes the all-worlds rule. If the some-world rule differs, the differences are explained at the end.

A small warning: To implement the full some-world rules, a tool must also implement at least part of the all-worlds rules. Those rules are required as part of substitution.

Propagating how many worlds a relation must hold in​

When one rule in this spec refers to another, it refers to the same version of the rule. For example, when the rules for containment refer to the rules for subtyping, the some-world containment relation refers to the some-world subtyping relation, and the all-worlds containment relation refers to the all-worlds subtyping relation.

This meta-rule applies except when a rule refers explicitly to a particular version of another rule.

Same type​

S and T are the same type if S is a subtype of T and T is a subtype of S.

The same-type relation is not defined to be reflexive or transitive.

For more discussion of reflexive and transitive relations, see the comments under nullness subtyping.

Subtyping​

A is a subtype of F if both of the following conditions are met:

The first condition suffices for most cases. The second condition is necessary only for types that have subcomponents --- namely, parameterized types and arrays. And it essentially says "Check the first condition on subcomponents as appropriate."

Nullness subtyping​

A is a nullness subtype of F if any of the following conditions are met:

Nullness subtyping asks the question: If A includes null, does F also include null? There are 4 cases in which this is true, 2 easy and 2 hard:

  • F is null-inclusive under every parameterization.

    This is the first easy case: F always includes null.

  • A is null-exclusive under every parameterization.

    This is the second easy case: A never includes null.

  • A has a nullness-subtype-establishing path to any type whose base type is the same as the base type of F, and F does not have nullness operator MINUS_NULL.

    This is the first hard case: A given type-variable usage does not necessarily always include null, nor does it necessarily always exclude null. (For example, consider a usage of E inside ArrayList<E>. ArrayList may be instantiated as either an ArrayList<@Nullable String> or an ArrayList<String>.)

    Subtyping questions for type-variable usages are more complex: E is a nullness subtype of E; @Nullable E is not. Similarly, if <F extends E>, then F is a nullness subtype of E. But if <F extends @Nullable E>, it is not.

  • F is a type-variable usage that meets both of the following conditions:

    • It does not have nullness operator MINUS_NULL.

    • A is a nullness subtype of its lower bound.

    This is the second hard case: It covers type variables that are introduced by capture conversion of ? super wildcards.

    In short, whether you have a Predicate<? super String>, a Predicate<? super @Nullable String>, or unannotated code that doesn't specify the nullness operator for the bound, you can always pass its test method a String. (If you want to pass a @Nullable String, then you'll need for the bound to be null-inclusive under every parameterization. The existence of the null-inclusiveness rule frees this current rule from having to cover that case.)

A further level of complexity in all this is UNSPECIFIED. For example, in the all-worlds version of the following rules, a type with nullness operator UNSPECIFIED can be both null-inclusive under every parameterization and null-exclusive under every parameterization.

Nullness subtyping (and thus subtyping itself) is not defined to be reflexive or transitive.

If we defined nullness subtyping to be reflexive, then String UNSPECIFIED would be a subtype of String UNSPECIFIED, even under the all-worlds rules. In other words, we'd be saying that unannotated code is always free from nullness errors. That is clearly false. (Nevertheless, lenient tools will choose not to issue errors for such code. They can do this by implementing the some-world rules.)

If we defined nullness subtyping to be transitive, then String UNION_NULL would be a subtype of String NO_CHANGE under the some-world rules. That would happen because of a chain of subtyping rules:

  • String UNION_NULL is a subtype of String UNSPECIFIED.

  • String UNSPECIFIED is a subtype of String NO_CHANGE.

Therefore, String UNION_NULL is a subtype of String NO_CHANGE.

Yes, it's pretty terrible for something called "subtyping" not to be reflexive or transitive. A more accurate name for this concept would be "consistent," a term used in gradual typing. However, we use "subtyping" anyway. In our defense, we need to name multiple concepts, including not just subtyping but also the same-type relation and containment. If we were to coin a new term for each, tool authors would need to mentally map between those terms and the analogous Java terms. (Still, yes: Feel free to read terms like "subtyping" as if they hvae scare quotes around them.)

Subtyping does end up being transitive when the relation is required to hold in all worlds. And it does end up being reflexive when the relation is required to hold only in some world. We don't state those properties as rules for 2 reasons: First, they arise naturally from the definitions. Second, we don't want to suggest that subtyping is reflexive and transitive under both versions of the rule.

Contrast this with our nullness-delegating subtyping rules and containment rules: Each of those is defined as a transitive closure. However, this is incorrect, and we should fix it: Transitivity causes the same problem there as it does here: List<? extends @Nullable String> ends up as a subtype of List<? extends String> because of a chain of subtyping rules that uses String UNSPECIFIED as part of the intermediate step. Luckily, tool authors that set out to implement transitivity for these two rules are very unlikely to write code that "notices" this chain. So, in practice, users are likely to see the "mostly transitive" behavior that we intend, even if we haven't found a way to formally specify it yet.

Null-inclusive under every parameterization​

A type is null-inclusive under every parameterization if it meets any of the following conditions:

  • Its nullness operator is UNION_NULL.

    This is the simplest part of the simplest case: A type usage always includes null if it's annotated with @Nullable.

  • It is an intersection type whose elements all are null-inclusive under every parameterization.

  • It is a type variable that meets both of the following conditions:

    • It does not have nullness operator MINUS_NULL.

    • Its lower bound is null-inclusive under every parameterization.

    This third case is probably irrelevant in practice: It covers ? super @Nullable Foo, which is already covered by the rules for nullness subtyping. It's included here in case some tool has reason to check whether a type is null-inclusive under every parameterization outside of a check for nullness subtyping.

Some-world version: The rule is the same except that the requirement for "UNION_NULL" is loosened to "UNION_NULL or UNSPECIFIED."

That is: It's possible that any type usage in unannotated code "ought to be" annotated with @Nullable.

Null-exclusive under every parameterization​

This is a straightforward concept ("never includes null"), but it's not as simple to implement as the null-inclusive rule was. This null-exclusive rule has to cover cases like String, E (where <E extends Object>), and E (where <E extends @Nullable Object> but nearby code has performed a null check on the expression). The case of <E extends Object> is an example of why the following rule requires looking for a "path."

A type is null-exclusive under every parameterization if it has a nullness-subtype-establishing path to either of the following:

  • any type whose nullness operator is MINUS_NULL

  • any augmented class or array type

    This rule refers specifically to a "class or array type," as distinct from other types like type variables and intersection types.

When code dereferences an expression, we anticipate that tools will check whether the expression is null-exclusive under every parameterization.

Nullness-subtype-establishing path​

Note that this definition is used both by the definition of null-inclusive under every parameterization and by the third condition in the definition nullness subtyping itself (the "type-variable case").

A has a nullness-subtype-establishing path to F if both of the following hold:

Some-world version: The rules are the same except that the requirement for "NO_CHANGE or MINUS_NULL" is loosened to "NO_CHANGE, MINUS_NULL, or UNSPECIFIED."

Nullness-subtype-establishing direct-supertype edges​

This section defines the supertypes for a given type --- but limited to those that fill the gaps in our nullness checking of "top-level" types. For example, there's no need for the rules to reflect that String NO_CHANGE extends Object NO_CHANGE: If we've established that a type has a path to String NO_CHANGE, then we already know that it's null-exclusive under every parameterization, based on the rules above, and that's enough to prove subtyping. And if we haven't established that, then the String-Object edge isn't going to change that.

Thus, the rules here are restricted to type variables and intersection types, whose supertypes may have nullness annotations.

T has nullness-subtype-establishing direct-supertype edges to the following:

  • if T is an augmented intersection type: all the intersection type's elements whose nullness operator is NO_CHANGE or MINUS_NULL

  • if T is an augmented type variable: all the corresponding type parameter's upper bounds whose nullness operator is NO_CHANGE or MINUS_NULL

  • otherwise: no nodes

Some-world version: The rules are the same except that the requirements for "NO_CHANGE or MINUS_NULL" are loosened to "NO_CHANGE, MINUS_NULL, or UNSPECIFIED."

Nullness-delegating subtyping rules for Java​

Recall that this rule exists to handle subcomponents of types --- namely, type arguments and array component types. It essentially says "Check nullness subtyping for subcomponents as appropriate."

The Java subtyping rules are defined in JLS 4.10. (Each rule takes a type as input and produces zero or more direct supertypes as outputs.) We add to them as follows:

  • As always, interpret the Java rules as operating on augmented types, not base types. This raises the question of how to extend these particular rules to operate on augmented types. The answer has two parts:

    • The first part of the answer applies only to the nullness operator "of the type." (As always, this means the nullness operator of the type component that is the entire type.)

      And the first part of the answer is: No matter what nullness operator the input augmented type has, the rules still apply, and they still produce the same direct supertypes.

      Thanks to that rule, the nullness operator of any output type is never read by the subtyping rules.

      So, when computing output types, tools may produce them with any nullness operator.

      Essentially, this rule says that the top-level types do no matter: They have already been checked by the nullness subtyping check.

      For simplicity, we recommend producing a nullness operator of NO_CHANGE: That operator is valid for all types, including intersection types.

    • The nullness operators of subcomponents of the augmented types do matter. For example, some Java rules produce subtypes only if subcomponents meet certain requirements. As always, check those requirements by applying this spec's definitions.

      Those definitions (like containment) refer back to definitions (like nullness subtyping) that use the nullness operators of the subcomponents in question.

  • When the Java array rules require one type to be a direct supertype of another, consider the direct supertypes of T to be every type that T is a subtype of.

Containment​

The Java rules are defined in JLS 4.5.1. We add to them as follows:

  • Disregard the 2 rules that refer to a bare ?. Instead, treat ? like ? extends Object, where the nullness operator of the Object bound is specified by "Bound of an unbounded wildcard."

    This is just a part of our universal rule to treat a bare ? like ? extends Object.

  • The rule written specifically for ? extends Object applies only if the nullness operator of the Object bound is UNION_NULL.

  • When the JLS refers to the same type T on both sides of a rule, the rule applies if and only if this spec defines the 2 types to be the same type.

Some-world version: The rules are the same except that the requirement for "UNION_NULL" is loosened to "UNION_NULL or UNSPECIFIED."

Substitution​

Substitution on Java base types barely requires an explanation: See JLS 1.3. Substitution on augmented types, however, is trickier: If Map.get returns V UNION_NULL, and if a user has a map whose value type is String UNSPECIFIED, then what does its get method return? Naive substitution would produce String UNSPECIFIED UNION_NULL. To reduce that to a proper augmented type with a single nullness operator, we define this process.

To substitute each type argument Aᵢ for each corresponding type parameter Pᵢ:

For every type-variable usage V whose base type is Pᵢ, replace V with the result of the following operation:

  • If V is null-exclusive under every parameterization in all worlds, then replace it with the result of applying MINUS_NULL to Aᵢ.

    This is the one instance in which a rule specifically refers to the all-worlds version of another rule. Normally, a rule "propagates" its version to other rules. But in this instance, the null-exclusivity rule (and all rules that it in turn applies) are the all-worlds versions.

    The purpose of this special case is to improve behavior in "the ImmutableList.Builder case": Because ImmutableList.Builder.add always throws NullPointerException for a null argument, we would like for add(null) to be a compile error, even under lenient tools. Unfortunately, without this special case, lenient tools could permit add(null) in unannotated code. For an example, read on.

    Consider an unannotated user of ImmutableList.Builder<Foo> builder. Its type argument Foo will have a nullness operator of UNSPECIFIED. Without this special case, the parameter of builder.add would have a nullness operator of UNSPECIFIED, too. Then, when a lenient tool would check whether the some-world subtyping relation holds for builder.add(null), it would find that it does.

    To solve this, we need a special case for substitution for null-exclusive type parameters like the one on ImmutableList.Builder. That special case needs to produce a type with a nullness operator other than UNSPECIFIED. One valid option is to produce NO_CHANGE; we happened to choose MINUS_NULL.

    The choice between NO_CHANGE and MINUS_NULL makes little difference for the parameter types of ImmutableList.Builder, but it can matter more for other APIs' return types. For example, consider @NullMarked class Foo<E extends @Nullable Object>, which somewhere uses the type FluentIterable<E>. FluentIterable has a method Optional<E> first(). Even when E is a type like String UNION_NULL (or String UNSPECIFIED), we know that first().get() will never return null. To surface that information to tools, we need to define our substitution rule to return E MINUS_NULL: If we instead used E NO_CHANGE, then the return type would look like it might include null.

  • Otherwise, replace V with the result of applying the nullness operator of V to Aᵢ.

Applying a nullness operator to an augmented type​

The process of applying a nullness operator requires 2 inputs:

  • the nullness operator to apply
  • the augmented type (which, as always, includes a nullness operator for that type) to apply it to

The result of the process is an augmented type.

The process is as follows:

First, based on the pair of nullness operators (the one to apply and the one from the augmented type), compute a "desired nullness operator." Do so by applying the following rules in order. Once one condition is met, skip the remaining conditions.

  • If the nullness operator to apply is MINUS_NULL, the desired nullness operator is MINUS_NULL.
  • If either nullness operator is UNION_NULL, the desired nullness operator is UNION_NULL.
  • If either nullness operator is UNSPECIFIED, the desired nullness operator is UNSPECIFIED.
  • The desired nullness operator is NO_CHANGE.

Then, if the input augmented type is not an intersection type, the output is the same as the input but with its nullness operator replaced with the desired nullness operator.

Otherwise, the output is an intersection type. For every element Tᵢ of the input type, the output type has an element that is the result of applying the desired nullness operator to Tᵢ.

In this case, the desired nullness operator is always equal to the nullness operator to apply that was an input to this process. That's because the nullness operator of the intersection type itself is defined to always be NO_CHANGE.

Capture conversion​

The Java rules are defined in JLS 5.1.10. We add to them as follows:

  • The parameterized type that is the output of the conversion has the same nullness operator as the parameterized type that is the input type.

  • Disregard the JLS rule about <?>. Instead, treat ? like ? extends Object, where the nullness operator of the Object bound is specified by "Bound of an unbounded wildcard."

    This is just a part of our universal rule to treat a bare ? like ? extends Object.

  • When a rule generates a lower bound that is the null type, we specify that its nullness operator is NO_CHANGE.

    See "Augmented null types."