Uploaded image for project: 'JDK'
  1. JDK
  2. JDK-8335876

JEP 488: Primitive Types in Patterns, instanceof, and switch (Second Preview)

XMLWordPrintable

    • Icon: JEP JEP
    • Resolution: Unresolved
    • Icon: P4 P4
    • None
    • specification
    • None
    • Angelos Bimpoudis
    • Feature
    • Open
    • SE
    • amber dash dev at openjdk dot org
    • M
    • M
    • 488

      Summary

      Enhance pattern matching by allowing primitive types in all pattern contexts, and extend instanceof and switch to work with all primitive types. This is a preview language feature.

      History

      This feature was originally proposed by JEP 455, and delivered as a preview feature in JDK 23. We here propose to preview it for a second time, without change.

      Goals

      • Enable uniform data exploration by allowing type patterns for all types, whether primitive or reference.

      • Align type patterns with instanceof, and align instanceof with safe casting.

      • Allow pattern matching to use primitive types in both nested and top-level pattern contexts.

      • Provide easy-to-use constructs that eliminate the risk of losing information due to unsafe casts.

      • Following the enhancements to switch in Java 5 (enum switch) and Java 7 (string switch), allow switch to process values of any primitive type.

      Non-Goals

      • It is not a goal to add new kinds of conversions to the Java language.

      Motivation

      Multiple restrictions pertaining to primitive types impose friction when using pattern matching, instanceof, and switch. Eliminating these restrictions would make the Java language more uniform and more expressive.

      Pattern matching for switch does not support primitive type patterns

      The first restriction is that pattern matching for switch (JEP 441) does not support primitive type patterns, i.e., type patterns that specify a primitive type. Only type patterns that specify a reference type are supported, such as case Integer i or case String s. (Since Java 21, record patterns (JEP 440) are also supported for switch.)

      With support for primitive type patterns in switch, we could improve the switch expression

      switch (x.getStatus()) {
          case 0 -> "okay";
          case 1 -> "warning";
          case 2 -> "error";
          default -> "unknown status: " + x.getStatus();
      }

      by turning the default clause into a case clause with a primitive type pattern that exposes the matched value:

      switch (x.getStatus()) {
          case 0 -> "okay";
          case 1 -> "warning";
          case 2 -> "error";
          case int i -> "unknown status: " + i;
      }

      Supporting primitive type patterns would also allow guards to inspect the matched value:

      switch (x.getYearlyFlights()) {
          case 0 -> ...;
          case 1 -> ...;
          case 2 -> issueDiscount();
          case int i when i >= 100 -> issueGoldCard();
          case int i -> ... appropriate action when i > 2 && i < 100 ...
      }

      Record patterns have limited support for primitive types

      Another restriction is that record patterns have limited support for primitive types. Record patterns streamline data processing by decomposing a record into its individual components. When a component is a primitive value, the record pattern must be precise about the type of the value. This is inconvenient for developers and inconsistent with the presence of helpful automatic conversions in the rest of the Java language.

      For example, suppose we wish to process JSON data represented via these record classes:

      sealed interface JsonValue {
          record JsonString(String s) implements JsonValue { }
          record JsonNumber(double d) implements JsonValue { }
          record JsonObject(Map<String, JsonValue> map) implements JsonValue { }
      }

      JSON does not distinguish integers from non-integers, so JsonNumber represents a number with a double component for maximum flexibility. However, we do not need to pass a double when creating a JsonNumber record; we can pass an int such as 30, and the Java compiler automatically widens the int to double:

      var json = new JsonObject(Map.of("name", new JsonString("John"),
                                       "age",  new JsonNumber(30)));

      Unfortunately, the Java compiler is not so obliging if we wish to decompose a JsonNumber with a record pattern. Since JsonNumber is declared with a double component, we must decompose a JsonNumber with respect to double, and convert to int manually:

      if (json instanceof JsonObject(var map)
          && map.get("name") instanceof JsonString(String n)
          && map.get("age")  instanceof JsonNumber(double a)) {
          int age = (int)a;  // unavoidable (and potentially lossy!) cast
      }

      In other words, primitive type patterns can be nested inside record patterns but are invariant: The primitive type in the pattern must be identical to the primitive type of the record component. It is not possible to decompose a JsonNumber via instanceof JsonNumber(int age) and have the compiler automatically narrow the double component to int.

      The reason for this limitation is that narrowing might be lossy: The value of the double component at run time might be too large, or have too much precision, for an int variable. However, a key benefit of pattern matching is that it rejects illegal values automatically, by simply not matching. If the double component of a JsonNumber is too large or too precise to narrow safely to an int, then instanceof JsonNumber(int age) could simply return false, leaving the program to handle a large double component in a different branch.

      With support for primitive type patterns, we could lift this limitation. Pattern matching could safeguard a possibly lossy narrowing conversion of a value to a primitive type, both at the top level and when nested inside record patterns. Since any double can be converted to an int, the primitive type pattern int a would be applicable to the corresponding component of JsonNumber of type double. If, and only if, the double component can be converted to an int without loss of information, then instanceof would match the pattern and the if-branch would be taken, with the local variable a in scope:

      if (json instanceof JsonObject(var map)
          && map.get("name") instanceof JsonString(String n)
          && map.get("age")  instanceof JsonNumber(int a)) {
            ... n ...
            ... a ...
      }

      This would enable nested primitive type patterns to work as smoothly as nested reference type patterns.

      Pattern matching for instanceof does not support primitive type patterns

      Yet another restriction is that pattern matching for instanceof (JEP 394) does not support primitive type patterns. Only type patterns that specify a reference type are supported. (Since Java 21, record patterns are also supported for instanceof.)

      Primitive type patterns would be just as useful in instanceof as they are in switch. The purpose of instanceof is, broadly speaking, to test whether a value can be converted safely to a given type; this is why we always see instanceof and cast operations in close proximity. This test is critical for primitive types because of the potential loss of information that can occur when converting primitive values from one type to another.

      For example, converting an int value to a float is performed automatically by an assignment statement even though it is potentially lossy — and the developer receives no warning of this:

      int getPopulation() {...}
      float pop = getPopulation();  // silent potential loss of information

      Meanwhile, converting an int value to a byte is performed with an explicit cast, but the cast is potentially lossy so it must be preceded by a laborious range check:

      if (i >= -128 && i <= 127) {
          byte b = (byte)i;
          ... b ...
      }

      Primitive type patterns in instanceof would subsume the lossy conversions built into the Java language and avoid the painstaking range checks that developers have been coding by hand for almost three decades. In other words, instanceof could check values as well as types. The two examples above could be rewritten as follows:

      if (getPopulation() instanceof float pop) {
          ... pop ...
      }
      
      if (i instanceof byte b) {
          ... b ...
      }

      The instanceof operator combines the convenience of an assignment statement with the safety of pattern matching. If the input (getPopulation() or i) can be converted safely to the type in the primitive type pattern then the pattern matches and the result of the conversion is immediately available (pop or b). But, if the conversion would lose information then the pattern does not match and the program should handle the invalid input in a different branch.

      Primitive types in instanceof and switch

      If we are going to lift restrictions around primitive type patterns then it would be helpful to lift a related restriction: When instanceof takes a type, rather than a pattern, it takes only a reference type, not a primitive type. When taking a primitive type, instanceof would check if the conversion is safe but would not actually perform it:

      if (i instanceof byte) {  // value of i fits in a byte
          ... (byte)i ...       // traditional cast required
      }

      This enhancement to instanceof restores alignment between the semantics of instanceof T and instanceof T t, which would be lost if we allowed primitive types in one context but not the other.

      Finally, it would be helpful to lift the restriction that switch can take byte, short, char, and int values but not boolean, float, double, or long values.

      Switching on boolean values would be a useful alternative to the ternary conditional operator (?:) because a boolean switch can contain statements as well as expressions. For example, the following code uses a boolean switch to perform some logging when false:

      startProcessing(OrderStatus.NEW, switch (user.isLoggedIn()) {
          case true  -> user.id();
          case false -> { log("Unrecognized user"); yield -1; }
      });

      Switching on long values would allow case labels to be long constants, obviating the need to handle very large constants with separate if statements:

      long v = ...;
      switch (v) {
          case 1L              -> ...;
          case 2L              -> ...;
          case 10_000_000_000L -> ...;
          case 20_000_000_000L -> ...;
          case long x          -> ... x ...;
      }

      Description

      In Java 21, primitive type patterns are permitted only as nested patterns in record patterns, as in:

      v instanceof JsonNumber(double a)

      To support the more uniform data exploration of a match candidate v with pattern matching, we will:

      1. Extend pattern matching so that primitive type patterns are applicable to a wider range of match candidate types. This will allow expressions such as v instanceof JsonNumber(int age).

      2. Enhance the instanceof and switch constructs to support primitive type patterns as top level patterns.

      3. Further enhance the instanceof construct so that, when used for type testing rather than pattern matching, it can test against all types, not just reference types. This will extend instanceof's current role, as the precondition for safe casting on reference types, to apply to all types.

        More broadly, this means that instanceof can safeguard all conversions, whether the match candidate is having its type tested (e.g., x instanceof int, or y instanceof String) or having its value matched (e.g., x instanceof int i, or y instanceof String s).

      4. Further enhance the switch construct so that it works with all primitive types, not just a subset of the

        integral primitive<br /> types

        .

      We achieve these changes by altering a small number of rules in the Java language that govern the use of primitive types, and by characterizing when a conversion from one type to another is safe — which involves knowledge of the value to be converted as well as the source and target types of the conversion.

      Safety of conversions

      A conversion is exact if no loss of information occurs. Whether a conversion is exact depends on the pair of types involved and on the input value:

      • For some pairs, it is known at compile time that conversion from the first type to the second type is guaranteed not to lose information for any value. The conversion is said to be unconditionally exact. No action is needed at run time for an unconditionally exact conversion. Examples include byte to int, int to long, and String to Object.

      • For other pairs, a run-time test is needed to check whether the value can be converted from the first type to the second type without loss of information or, if a cast were to be performed, without throwing an exception. The conversion is exact if no loss of information or exception would occur; otherwise, the conversion is not exact. Examples of conversions that might be exact are long to int and int to float, where loss of precision is detected at run time by using numerical equality (==) or

        representation<br /> equivalence

        , respectively. Converting from Object to String also needs a run-time test, and the conversion is exact or not exact depending on whether the input value is dynamically a String.

      In brief, a conversion between primitive types is unconditionally exact if it widens from one integral type to another, or from one floating-point type to another, or from byte, short, or char to a floating-point type, or from int to double. Furthermore, boxing conversions and widening reference conversions are unconditionally exact.

      The following table denotes the conversions that are permitted between primitive types. Unconditionally exact conversions are denoted with the symbol ɛ. The symbol  means the identity conversion, ω means a widening primitive conversion, η means a narrowing primitive conversion, and ωη means a widening and narrowing primitive conversion. The symbol  means no conversion is allowed.

      To → byte short char int long float double boolean
      From ↓
      byte ɛ ωη ɛ ɛ ɛ ɛ
      short η η ɛ ɛ ɛ ɛ
      char η η ɛ ɛ ɛ ɛ
      int η η η ɛ ω ɛ
      long η η η η ω ω
      float η η η η η ɛ
      double η η η η η η
      boolean

      Comparing this table to its equivalent in JLS §5.5, it can be seen that many of the conversions permitted by ω in JLS §5.5 are "upgraded" to the unconditionally exact ɛ above.

      instanceof as the precondition for safe casting

      Type tests with instanceof are traditionally limited to reference types. The classic meaning of instanceof is a precondition check that asks: Would it be safe and useful to cast this value to this type? This question is even more pertinent to primitive types than to reference types. For reference types, if the check is accidentally omitted then performing an unsafe cast will likely do no harm: A ClassCastException will be thrown and the improperly cast value will be unusable. In contrast, for primitive types, where there is no convenient way to check for safety, performing an unsafe cast will likely cause subtle bugs. Instead of throwing an exception, it can silently lose information such as magnitude, sign, or precision, allowing the improperly cast value to flow into the rest of the program.

      To enable primitive types in the instanceof type test operator, we remove the restrictions (JLS §15.20.2) that the type of the left-hand operand must be a reference type and that the right-hand operand must specify a reference type. The type test operator becomes

      InstanceofExpression:
          RelationalExpression instanceof Type
          ...

      At run time, we extend instanceof to primitive types by appealing to exact conversions: If the value on the left-hand side can be converted to the type on the right-hand side via an exact conversion then it would be safe to cast the value to that type, and instanceof reports true.

      Here are some examples of how the extended instanceof can safeguard casting. Unconditionally exact conversions return true regardless of the input value; all other conversions require a run-time test whose result is shown.

      byte b = 42;
      b instanceof int;         // true (unconditionally exact)
      
      int i = 42;
      i instanceof byte;        // true (exact)
      
      int i = 1000;
      i instanceof byte;        // false (not exact)
      
      int i = 16_777_217;       // 2^24 + 1
      i instanceof float;       // false (not exact)
      i instanceof double;      // true (unconditionally exact)
      i instanceof Integer;     // true (unconditionally exact)
      i instanceof Number;      // true (unconditionally exact)
      
      float f = 1000.0f;
      f instanceof byte;        // false
      f instanceof int;         // true (exact)
      f instanceof double;      // true (unconditionally exact)
      
      double d = 1000.0d;
      d instanceof byte;        // false
      d instanceof int;         // true (exact)
      d instanceof float;       // true (exact)
      
      Integer ii = 1000;
      ii instanceof int;        // true (exact)
      ii instanceof float;      // true (exact)
      ii instanceof double;     // true (exact)
      
      Integer ii = 16_777_217;
      ii instanceof float;      // false (not exact)
      ii instanceof double;     // true (exact)

      We do not add any new conversions to the Java language, nor change existing conversions, nor change which conversions are allowed in existing contexts such as assignment. Whether instanceof is applicable to a given value and type is determined by whether a conversion is allowed in a casting context and whether it is exact. For example, b instanceof char is never allowed if b is a boolean variable, because there is no casting conversion from boolean to char.

      Primitive type patterns in instanceof and switch

      A type pattern merges a type test with a conditional conversion. This avoids the need for an explicit cast if the type test succeeds, while the uncast value can be handled in a different branch if the type test fails. When the instanceof type test operator supported only reference types, it was natural that only reference type patterns were allowed in instanceof and switch; now that the instanceof type test operator supports primitive types, it is natural to allow primitive type patterns in instanceof and switch.

      To achieve this, we drop the restriction that primitive types cannot be used in a top level type pattern. As a result, the laborious and error-prone code

      int i = 1000;
      if (i instanceof byte) {    // false -- i cannot be converted exactly to byte
          byte b = (byte)i;       // potentially lossy
          ... b ...
      }

      can be written as

      if (i instanceof byte b) {
          ... b ...               // no loss of information
      }

      because i instanceof byte b means "test if i instanceof byte and, if so, cast i to byte and bind that value to b".

      The semantics of type patterns are defined by three predicates: applicability, unconditionality, and matching. We lift restrictions on the treatment of primitive type patterns as follows:

      • Applicability is whether a pattern is legal at compile time. Previously, applicability for a primitive type pattern required that the match candidate have the exact same type as the type in the pattern. For example, switch (... an int ...) { case double d: ... } was not allowed because the pattern double was not applicable to int.

        Now, a type pattern T t is applicable to a match candidate of type U if a U could be cast to T without an unchecked warning. Since int can be cast to double, that switch is now legal.

      • Unconditionality is whether it is known at compile time that an applicable pattern will match all possible run-time values of the match candidate. An unconditional pattern requires no run-time checks.

        As we extend primitive type patterns to be applicable to more types, we must specify on which types they are unconditional. A primitive type pattern for type T is unconditional on a match candidate of type U if the conversion from U to T is unconditionally exact. This is because an unconditionally exact conversion is safe regardless of the input value.

      • Previously, a value v that is not the null reference matches a type pattern of type T if v can be cast to T without throwing a ClassCastException. This definition of matching sufficed when primitive type patterns had a limited role. Now that primitive type patterns can be used widely, matching is generalized to mean that a value can be cast exactly to T, which covers throwing a ClassCastException as well as potential loss of information.

      Exhaustiveness

      A switch expression, or a switch statement whose case labels are patterns, is required to be exhaustive: All possible values of the selector expression must be handled in the switch block. A switch is exhaustive if it contains an unconditional type pattern; it can be exhaustive for other reasons as well, such as covering all possible permitted subtypes of a sealed class. In some situations, a switch can be deemed exhaustive even when there are possible run-time values that will not be matched by any case; in such situations the Java compiler inserts a synthetic default clause to handle these unanticipated inputs. Exhaustiveness is covered in greater detail in Patterns: Exhaustiveness, Unconditionality, and Remainder.

      With the introduction of primitive type patterns, we add one new rule to the determination of exhaustiveness: Given a switch whose match candidate is a wrapper type W for some primitive type P, a type pattern T t exhausts W if T is unconditionally exact on P. In that case, null becomes part of the remainder. In the following example, the match candidate is a wrapper type of the primitive type byte and the conversion from byte to int is unconditionally exact. As a result the following switch is exhaustive:

      Byte b = ...
      switch (b) {             // exhaustive switch
          case int p -> 0;
      }

      This behavior is similar to the exhaustiveness treatment of record patterns.

      Just as switch uses pattern exhaustiveness to determine if the cases cover all input values, switch uses dominance to determine if there are any cases that will match no input values.

      One pattern dominates another pattern if it matches all the values that the other pattern matches. For example, the type pattern Object o dominates the type pattern String s because everything that would match String s would also match Object o. In a switch, it is illegal for a case label with an unguarded type pattern P to precede a case label with type pattern Q if P dominates Q. The meaning of dominance is unchanged: A type pattern T t dominates a type pattern U u if T t would be unconditional on a match candidate of type U.

      Expanded primitive support in switch

      We enhance the switch construct to allow a selector expression of type long, float, double, and boolean, as well as the corresponding boxed types.

      If the selector expression has type long, float, double, or boolean, any constants used in case labels must have the same type as the selector expression, or its corresponding boxed type. For example, if the type of the selector expression is float or Float then any case constants must be floating-point literals (JLS §3.10.2) of type float. This restriction is required because mismatches between case constants and the selector expression could introduce lossy conversions, undermining programmer intent. The following switch is legal, but it would be illegal if the 0f constant were accidentally written as 0.

      float v = ...
      switch (v) {
          case 0f -> 5f;
          case float x when x == 1f -> 6f + x;
          case float x -> 7f + x;
      }

      The semantics of floating-point literals in case labels is defined in terms of representation equivalence at compile time and run time. It is a compile-time error to use two floating-point literals that are representation equivalent. For example, the following switch is illegal because the literal 0.999999999f is rounded up to 1.0f, creating a duplicate case label.

      float v = ...
      switch (v) {
          case 1.0f -> ...
          case 0.999999999f -> ...    // error: duplicate label
          default -> ...
      }

      Since the boolean type has only two distinct values, a switch that lists both the true and false cases is considered exhaustive. The following switch is legal, but it would be illegal if there were a default clause.

      boolean v = ...
      switch (v) {
          case true -> ...
          case false -> ...
          // Alternatively: case true, false -> ...
      }

      Future Work

      Having regularized the Java language’s rules around type comparisons and pattern matching, we may then consider introducing constant patterns. At present, in a switch, constants can only appear as case constants, e.g., the 42 in this code:

      short s = ...
      switch (s) {
          case 42 -> ...
          case int i -> ...
      }

      Constants cannot appear in record patterns, which limits the usefulness of pattern matching. For example, the following switch is not possible:

      record Box(short s) {}
      
      Box b = ...
      switch (b) {
          case Box(42) -> ...  // Box(42) is not a valid record pattern
          case Box(int i) -> ...
      }

      Thanks to the applicability rules defined here, constants could be allowed to appear in record patterns. In a switch, case Box(42) would mean case Box(int i) when i == 42, since 42 is a literal of type int.

            abimpoudis Angelos Bimpoudis
            abimpoudis Angelos Bimpoudis
            Angelos Bimpoudis Angelos Bimpoudis
            Alex Buckley, Maurizio Cimadamore
            Votes:
            0 Vote for this issue
            Watchers:
            5 Start watching this issue

              Created:
              Updated: