Effect cases in switch

[Remi Forax]

> From: "Brian Goetz" <brian.goetz at oracle.com>
> To: "amber-spec-experts" <amber-spec-experts at openjdk.java.net>
> Sent: Tuesday, December 12, 2023 10:23:09 PM
> Subject: Effect cases in switch

> Based on some inspiration from OCaml, and given that the significant upgrades to
> switch so far position it to do a lot more than it could before, we've been
> exploring a further refinement of switch to incorporate failure handling as
> well.

> (I realize that this may elicit strong reactions from some, but please give it
> some careful thought before giving voice to those reactions.)

Hello, 

> # Uniform handling of failure in switch

> ## Summary

> Enhance the `switch` construct to support `case` labels that match exceptions
> thrown during evaluation of the selector expression, providing uniform handling
> of normal and exceptional results.

> ## Background

> The purpose of the `switch` construct is to choose a single course of action
> based on evaluating a single expression (the "selector"). The `switch`
> construct is not strictly needed in the language; everything that `switch` does
> can be done by `if-else`. But the language includes `switch` because it
> embodies useful constraints which both streamline the code and enable more
> comprehensive error checking.

> The original version of `switch` was very limited: the selector expression was
> limited to a small number of primitive types, the `case` labels were limited to
> numeric literals, and the body of a switch was limited to operating by
> side-effects (statements only, no expressions.) Because of these limitations,
> the use of `switch` was usually limited to low-level code such as parsers and
> state machines. In Java 5 and 7, `switch` received minor upgrades to support
> primitive wrapper types, enums, and strings as selectors, but its role as "pick
> from one of these constants" did not change significantly.

> Recently, `switch` has gotten more significant upgrades, to the point where it
> can take on a much bigger role in day-to-day program logic. Switch can now be
> used as an expression in addition to a statement, enabling greater composition
> and more streamlined code. The selector expression can now be any type. The
> `case` labels in a switch block can be rich patterns, not just constants, and
> have arbitrary predicates as guards. We get much richer type checking for
> exhaustiveness when switching over selectors involving sealed types. Taken
> together, this means much more program logic can be expressed concisely and
> reliably using `switch` than previously.

> ### Bringing nulls into `switch`

> Historically, the `switch` construct was null-hostile; if the selector evaluated
> to `null`, the `switch` immediately completed abruptly with
> `NullPointerException`. This made a certain amount of sense when the only
> reference types that could be used in switch were primitive wrappers and enums,
> for which nulls were almost always indicative of an error, but as `switch`
> became more powerful, this was increasingly a mismatch for what we wanted to do
> with `switch`. Developers were forced to work around this, but the workarounds
> had undesirable consequences (such as forcing the use of statement switches
> instead of expression switches.) Previously, to handle null, one would have to
> separately evaluate the selector and compare it to `null` using `if`:

> ```
> SomeType selector = computeSelector();
> SomeOtherType result;
> if (selector == null) {
> result = handleNull();
> }
> else {
> switch (selector) {
> case X:
> result = handleX();
> break;
> case Y:
> result = handleY();
> break;
> }
> }
> ```

> Not only is this more cumbersome and less concise, but it goes against the main
> job of `switch`, which is streamline "pick one path based on a selector
> expression" decisions. Outcomes are not handled uniformly, they are not handled
> in one place, and the inability to express all of this as an expression limits
> composition with other language features.

> In Java 21, it became possible to treat `null` as just another possible value of
> the selector in a `case` clause (and even combine `null` handling with
> `default`), so that the above mess could reduce to

> ```
> SomeOtherType result = switch (computeSelector()) {
> case null -> handleNull();
> case X -> handleX();
> case Y -> handleY();
> }
> ```

> This is simpler to read, less error-prone, and interacts better with the rest of
> the language. Treating nulls uniformly as just another value, as opposed to
> treating it as an out-of-band condition, made `switch` more useful and made Java
> code simpler and better. (For compatibility, a `switch` that has no `case null`
> still throws `NullPointerException` when confronted with a null selector; we opt
> into the new behavior with `case null`.)

> ### Other switch tricks

> The accumulation of new abilities for `switch` means that it can be used in more
> situations than we might initially realize. One such use is replacing the
> ternary conditional expression with boolean switch expressions; now that
> `switch` can support boolean selectors, we can replace

> expr ? A : B

> with the switch expression

> ```
> switch (expr) {
> case true -> A;
> case false -> B;
> }
> ```

> This might not immediately seem preferable, since the ternary expression is more
> concise, but the `switch` is surely more clear. And, if we nest ternaries in
> the arms of other ternaries (possibly deeply), this can quickly become
> unreadable, whereas the corresponding nested switch remains readable even if
> nested to several levels. We don't expect people to go out and change all their
> ternaries to switches overnight, but we do expect that people will increasingly
> find uses where a boolean switch is preferable to a ternary. (If the language
> had boolean switch expressions from day 1, we might well not have had ternary
> expressions at all.)

> Another less-obvious example is using guards to do the selection, within the
> bounds of the "pick one path" that `switch` is designed for. For example, we
> can write the classic "FizzBuzz" exercise as:

> ```
> String result = switch (getNumber()) {
> case int i when i % 15 == 0 -> "FizzBuzz";
> case int i when i % 5 == 0 -> "Fizz";
> case int i when i % 3 == 0 -> "Buzz";
> case int i -> Integer.toString(i);
> }
> ```

> A more controversial use of the new-and-improved switch is as a replacement for
> block expressions. Sometimes we want to use an expression (such as when passing
> a parameter to a method), but the value can only be constructed using
> statements:

> ```
> String[] choices = new String[2];
> choices[0] = f(0);
> choices[1] = f(1);
> m(choices);
> ```

> While it is somewhat "off label", we can replace this with a switch expression:

> ```
> m(switch (0) {
> default -> {
> String[] choices = new String[2];
> choices[0] = f(0);
> choices[1] = f(1);
> yield choices;
> }
> })
> ```

> While these were not the primary use cases we had in mind when upgrading
> `switch`, it illustrates how the combination of improvements to `switch` have
> made it a sort of "swiss army knife".

> ## Handling failure uniformly

> Previously, null selector values were treated as out-of-band events, requiring
> that users handle null selectors in a non-uniform way. The improvements to
> `switch` in Java 21 enable null to be handled uniformly as a selector value, as
> just another value.

> A similar source of out-of-band events in `switch` is exceptions; if evaluating
> the selector throws an exception, the switch immediately completes with that
> exception. This is an entirely justifiable design choice, but it forces users
> to handle exceptions using a separate mechanism, often a cumbersome one, just as
> we did with null selectors:

> ```
> Number parseNumber(String s) throws NumberFormatException() { ... }

> try {
> switch (parseNumber(input)) {
> case Integer i -> handleInt(i);
> case Float f -> handleFloat(f);
> ...
> }
> }
> catch (NumberFormatException e) {
> ... handle exception ...
> }
> ```

> This is already unfortunate, as switch is designed to handle "choose one path
> based on evaluating the selector", and "parse error" is one of the possible
> consequences of evaluating the selector. It would be nice to be able to handle
> error cases uniformly with success cases, as we did with null. Worse, this code
> doesn't even mean what we want: the `catch` block catches not only exceptions
> thrown by evaluating the selector, but also by the body of the switch. To say
> what we mean, we need the even more unfortunate

> ```
> var answer = null;
> try {
> answer = parseNumber(input);
> }
> catch (NumberFormatException e) {
> ... handle exception ...
> }

> if (answer != null) {
> switch (answer) {
> case Integer i -> handleInt(i);
> case Float f -> handleFloat(f);
> ...
> }
> }
> ```

> Just as it was an improvement to handle `null` uniformly as just another
> potential value of the selector expression, we can get a similar improvement by
> handling normal and exceptional completion uniformly as well. Normal and
> exceptional completion are mutually exclusive, and the handling of exceptions in
> `try-catch` already has a great deal in common with handling normal values in
> `switch` statements (a catch clause is effectively matching to a type pattern.)
> For activities with anticipated failure modes, handling successful completion
> via one mechanism and failed completion through another makes code harder to
> read and maintain.

> ## Proposal

> We can extend `switch` to handle exceptions more uniformly in a similar was as
> we extended it to handle nulls by introducing `throws` cases, which match when
> evaluating the selector expression completes abruptly with a compatible
> exception:

> ```
> String allTheLines = switch (Files.readAllLines(path)) {
> case List<String> lines -> lines.stream().collect(Collectors.joining("\n"));
> case throws IOException e -> "";
> }
> ```

> This captures the programmer's intent much more clearly, because the expected
> success case and the expected failure case are handled uniformly and in the same
> place, and their results can flow into the result of the switch expression.

> The grammar of `case` labels is extended to include a new form, `case throws`,
> which is followed by a type pattern:

> case throws IOException e:

> Exception cases can be used in all forms of `switch`: expression and statement
> switches, switches that use traditional (colon) or single-consequence (arrow)
> case labels. Exception cases can have guards like any other pattern case.
I think I would prefer "case throws" to be spell "catch" even if we have to have a discussion about catch(Throwable t) vs catch Throwable t. 

> Exception cases have the obvious dominance order with other exception cases (the
> same one used to validate order of `catch` clauses in `try-catch`), and do not
> participate in dominance ordering with non-exceptional cases. It is a
> compile-time error if an exception case specifies an exception type that cannot
> be thrown by the selector expression, or a type that does not extend
> `Throwable`. For clarity, exception cases should probably come after all other
> non-exceptional cases.

> When evaluating a `switch` statement or expression, the selector expression is
> evaluated. If evaluation of the selector expression throws an exception, and
> one of the exception cases in the `switch` matches the exception, then control
> is transferred to the first exception case matching the exception. If no
> exception case matches the exception, then the switch completes abruptly with
> that same exception.
I don't want to be the guy implementing this :) 
Your proposal is fighting against the physics of the VM, when you enter in an exception handler (a catch block), the whole stack disapear so you need some compiler magic to store everything which is on stack into locals before calling a switch expression so you can restore it inside the exception handler. I've implemented a very similar kind of transformation 10 years ago, it severly bloats the produced bytecode. Or maybe you are suggesting that this new switch will use a new bytecode construct (inline exception handler ?), it's more or less what JSR/RET was, but perhaps there is a better way. 

I understand the appeal of such construct, it's a way to switch (eheh) from a world with exceptions to a more functional world where errors are values. 

regards, 
Rémi 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/amber-spec-observers/attachments/20231212/af388f41/attachment-0001.htm>