Member Patterns -- the bikeshed
Brian Goetz
brian.goetz at oracle.com
Mon Apr 1 16:16:45 UTC 2024
I've received a number of private mails, each trying to "rescue" the
imperative approach somehow. First, an analogy to set the stage (at the
risk of derailing the actual discussion.)
BEGIN ANALOGY
Elsewhere in Amber, we are working on reconstruction expressions for
immutable objects:
point with { x = 2*x; y = 2*y; }
A number of people have seized on this as the backdoor through which
they'll get by-name constructor invocation, asking for some variant of
Point with { x = 3; y = 4; }
as a constructor syntax. While there's nothing *locally* wrong with
this as a linguistic form, I think there's something very wrong with the
motivation.
The motivation is that some people really, really want by-name
invocation for everything: constructors, methods, patterns. And I
totally get that; they've seen it in other languages and like it, it
makes code more readable and less error-prone, etc etc. Unfortunately,
this feature is just much harder to add to Java than it seems to 99.999%
of Java developers. (If it were as easy and harmless as everyone thinks
it is, we would have done it decades ago. It's not impossible, but it
is significant, and we'd rather invest that effort elsewhere.)
When confronted with the perceived hostility towards a feature they
want, people start bargaining ("I realize I can't have it in general,
but maybe I can have it for records?") The above pseudo-constructor
syntax for records is exactly this sort of bargaining -- but its a
terrible bargain.
Its a terrible bargain because there's no chance of extending this to
other contexts where people want it just as much (e.g., methods), and,
to make it worse, if we ever did decide to do by-name invocation in
general, the reconstructor-inspired syntax would be a white elephant next to
new Point(x: 1, y: 2)
which is more direct and concise. So its a hard no on the
reconstructor-inspired constructor syntax, because of one or both of:
- Having it for records only will quickly feel like "glass 90%" empty,
and
- If we eventually fill the glass, there'll be some sludgy residue in it.
But, the bargaining instinct is strong. First, another digression.
BEGIN NESTED DIGRESSION
I was at an AMA (Ask Me Anything) at a conference about eight years
ago. The audience peppered me with the usual "Will Java ever..."
questions. We had a great time. One of the questions was "what about
making semicolons optional." I went through some examples where
removing a semicolon in existing code could change the meaning in
surprising ways. But, the questioner was not to be deterred; he shot
back with "OK, since records are a new construct, how about you make
semicolons optional inside record declarations?"
Serious props to this dude for dogged persistence towards his goals, but
I almost fell off my chair laughing. What a stupid language that would
result in! Asking developers to reason about "in which constructs are
semicolons optional" would be worse than either extreme.
The moral of this story is that when we want something badly enough, we
become blind to the obvious negative side-effects. We convince
ourselves that getting it in a limited context, in a stilted way, is
better than nothing. But it's usually worse than nothing, because it
creates new edge cases, and new constraints for making things globally
better. When it comes to language design, "local optimizations" usually
aren't.
END NESTED DIGRESSION
The point of the nested digression was to illustrate the danger of
trying to locally re-invent a feature that is better handled globally.
The wither-inspired constructor syntax is clearly dominated by a more
general by-name invocation mechanism, but in our depression that we
might never get one, we embrace clearly suboptimal choices.
END ANALOGY
So what was the point of the analogy? AFAICS, the "imperative" syntax
has two motivations:
- To look like the syntactic dual of a constructor, and
- As a way of providing "by name" matcher completion, rather than only
"positional" matcher completion.
To the first, I think the document does a good job at dismissing; this
is a local maxima, but it falls apart in the bigger context of patterns
in the object model.
The point of the analogy is to highlight that the second is a similar
form of bargaining over the lack of positional invocation. But if we
were ever to have by-name invocation for methods, there's an obvious
syntax that extends to binding production:
matches Point(x: 1, y: 2)
as well as pattern use sites:
case Point(x: var first, y: var second): ...
which suggests that we should avoid the temptation to use the imperative
style as a "consolation prize" for the lack of general by-name invocation.
On 3/29/2024 5:58 PM, Brian Goetz wrote:
> We now come to the long-awaited bikeshed discussion on what member
> patterns should look like.
>
> Bikeshed disclaimer for EG:
> - This is likely to evoke strong opinions, so please take pains to
> be especially constructive
> - Long reply-to-reply threads should be avoided even more than usual
> - Holistic, considered replies preferred
> - Please change subject line if commenting on a sub-topic or tangential
> concern
>
> Special reminders for Remi:
> - Use of words like "should", "must", "shouldn't", "mistake",
> "wrong", "broken"
> are strictly forbidden.
> - If in doubt, ask questions first.
>
> Notes for external observers:
> - This is a working document for the EG; the discussion may continue
> for a
> while before there is an official proposal. Please be patient.
>
>
> # Pattern declaration: the bikeshed
>
> We've largely identified the model for what kinds of patterns we need to
> express, but there are still several degrees of freedom in the syntax.
>
> As the model has simplified during the design process, the space of syntax
> choices has been pruned back, which is a good thing. However, there
> are still
> quite a few smaller decisions to be made. Not all of the
> considerations are
> orthogonal, so while they are presented individually, this is not a
> "pick one
> from each column" menu.
>
> Some of these simplifications include:
>
> - Patterns with "input arguments" have been removed; another way to
> get to what
> this gave us may come back in another form.
> - I have grown increasingly skeptical of the value of the imperative
> `match`
> statement. With better totality analysis, I think it can be
> eliminated.
>
> We can discuss these separately but I would like to sync first on the
> broad
> strokes for how patterns are expressed.
>
> ## Object model requirements
>
> As outlined in "Towards Member Patterns", the basic model is that
> patterns are
> the dual of other executable members (constructors, static methods,
> instance
> methods.) While they are like methods in that they have inputs,
> outputs, names,
> and an imperative body, they have additional degrees of freedom that
> constructors and methods lack:
>
> - Patterns are, in general, _conditional_ (they can succeed or fail),
> and only
> produce bindings (outputs) when they succeed. This conditionality is
> understood by the language's flow analysis, and is used for
> computing scoping
> and definite assignment.
> - Methods can return at most one value; when a pattern completes
> successfully,
> it may bind multiple values.
> - All patterns have a _match candidate_, which is a distinguished,
> possibly-implicit parameter. Some patterns also have a receiver,
> which is
> also a distinguished, possibly-implicit parameter. In some such
> cases the
> receiver and match candidate are aliased, but in others these may
> refer to
> different objects.
>
> So a pattern is a named executable member that takes a _match
> candidate_ as a
> possibly-implicit parameter, maybe takes a receiver as an implicit
> parameter,
> and has zero or more conditional _bindings_. Its body can perform
> imperative
> computation, and can terminate either with match failure or success.
> In the
> success case, it must provide a value for each binding.
>
> Deconstruction patterns are special in many of the same ways
> constructors are:
> they are constrained in their name, inheritance, and probably their
> conditionality (they should probably always succeed). Just as the
> syntax for
> constructors differs slightly from that of instance methods, the
> syntax for
> deconstructors may differ slightly from that of instance patterns. Static
> patterns, like static methods, have no receiver and do not have access
> to the
> type parameters of the enclosing class.
>
> Like constructors and methods, patterns can be overloaded, but in
> accordance
> with their duality to constructors and methods, the overloading
> happens on the
> _bindings_, not the inputs.
>
> ## Use-site syntax
>
> There are several kinds of type-driven patterns built into the
> language: type
> patterns and record patterns. A type pattern in a `switch` looks like:
>
> case String s: ...
>
> And a record pattern looks like:
>
> case MyRecord(P1, P2, ...): ...
>
> where `P1..Pn` are nested patterns that are recursively matched to the
> components of the record. This use-site syntax for record patterns
> was chosen
> for its similarity to the construction syntax, to highlight that a record
> pattern is the dual of record construction.
>
> **Deconstruction patterns.** The simplest kind of member pattern, a
> deconstruction pattern, will have the same use-site syntax as a record
> pattern;
> record patterns can be thought of as a deconstruction pattern
> "acquired for
> free" by records, just as records do with constructors, accessors, object
> methods, etc. So the use of a deconstruction pattern for `Point`
> looks like:
>
> case Point(var x, var y): ...
>
> whether `Point` is a record or an ordinary class equipped with a suitable
> deconstruction pattern.
>
> **Static patterns.** Continuing with the idea that the destructuring
> syntax
> should evoke the aggregation syntax, there is an obvious candidate for the
> use-site syntax for static patterns:
>
> case Optional.of(var e): ...
> case Optional.empty(): ...
>
> **Instance patterns.** Uses of instance patterns will likely come in
> two forms,
> analogous to bound and unbound instance method references, depending
> on whether
> the receiver and the match candidate are the same object. In the
> unbound form,
> used when the receiver is the same object as the match candidate, the
> pattern
> name is qualified by a _type_:
>
> ```
> Class<?> k = ...
> switch (k) {
> // Qualified by type
> case Class.arrayClass(var componentType): ...
> }
> ```
>
> This means that we _resolve_ the pattern `arrayClass` starting at
> `Class` and
> _select_ the pattern using the receiver, `k`. We may also be able to
> omit the
> class qualifier if the static type of the match candidate is sufficient to
> resolve the desired pattern.
>
> In the bound form, used when the receiver is distinct from the match
> candidate,
> the pattern name is qualified with an explicit _receiver expression_.
> As an
> example, consider an interface that captures primitive widening and
> narrowing
> conversions, such as those between `int` and `long`. In the widening
> direction,
> conversion is unconditional, so this can be modeled as a method from
> `int` to
> `long`. In the other direction, conversion is conditional, so this is
> better
> modeled as a _pattern_ whose match candidate is `long` and which binds
> an `int`
> on success. Since these are instance methods of some class (say,
> `NumericConversion<T,U>`), we need to provide the receiver instance in
> order to
> resolve the pattern:
>
> ```
> NumericConversion<int, long> nc = ...
>
> switch (aLong) {
> case nc.narrowed(int i):
> ...
> }
> ```
>
> The explicit receiver syntax would also be used if we exposed regular
> expression
> matching as a pattern on the `j.u.r.Pattern` object (the name collision on
> `Pattern` is unfortunate). Imagine we added a `matching` instance
> pattern to
> `j.u.r.Pattern`; then we could use it in `instanceof` as follows:
>
> ```
> static final java.util.regex.Pattern P = Pattern.compile("(a*)(b*)");
> ...
> if (aString instanceof P.matching(String as, String bs)) { ... }
> ```
>
> Each of these use-site syntaxes is modeled after the use-site syntax for a
> method invocation or method reference.
>
> ## Declaration-site syntax
>
> To avoid being biased by the simpler cases, we're going to work all
> the cases
> concurrently rather than starting with the simpler cases and working
> up. (It
> might seem sensible to start with deconstructors, since they are the
> "easy"
> case, but if we did that, we would likely be biased by their
> simplicity and then
> find ourselves painted into a corner.) As our example gallery, we
> will consider:
>
> - Deconstruction pattern for `Point`;
> - Static patterns for `Optional::of` and `Optional::empty`;
> - Static pattern for "power of two" (illustrating a computations
> where success
> or failure, and computation of bindings, cannot easily be separated);
> - Instance pattern for `Class::arrayClass` (used unbound);
> - Instance pattern for `Pattern::matching` on regular expressions
> (used bound).
>
> Member patterns, like methods, have _names_. (We can think of
> constructors as
> being named for their enclosing classes, and the same for
> deconstructors.) All
> member patterns have a (possibly empty) ordered list of _bindings_,
> which are
> the dual of constructor or method parameters. Bindings, in turn, have
> names and
> types. And like constructors and methods, member patterns have a
> _body_ which
> is a block statement. Member patterns also have a _match candidate_,
> which is a
> likely-implicit method parameter.
>
> ### Member patterns as inverse methods and constructors
>
> Regardless of syntax, let us remind ourselves that that deconstructors
> are the
> categorical dual to constructors (coconstructors), and pattern methods
> are the
> categorical dual to methods (comethods). They are dual in their
> structure: a
> constructor or method takes N arguments and produces a result, the
> corresponding
> member pattern consumes a match candidate and (conditionally) produces N
> bindings.
>
> Moreover, they are semantically dual: the return value produced by
> construction
> or factory invocation is the match candidate for the corresponding member
> pattern, and the bindings produced by a member pattern are the answers
> to the
> _Pattern Question_ -- "could this object have come from an invocation
> of my
> dual, and if so, with what arguments."
>
> ### What do we call them?
>
> Given the significant overlap between methods and patterns, the first
> question
> about the declaration we need to settle is how to identify a member
> pattern
> declaration as distinct from a method or constructor declaration.
> _Towards
> Member Patterns_ tried out a syntax that recognized these as _inverse_
> methods
> and constructors:
>
> public Point(int x, int y) { ... }
> public inverse Point(int x, int y) { ... }
>
> While this is a principled choice which clearly highlights the
> duality, and one
> that might be good for specification and verbal description, it is
> questionable
> whether this would be a great syntax for reading and writing programs.
>
> A more traditional option is to choose a "noun" (conditional) keyword,
> such as
> `pattern`, `matcher`, `extractor`, `view`, etc:
>
> public pattern Point(int x, int y) { ... }
>
> If we are using a noun keyword to identify pattern declarations, we
> could use
> the same noun for all of them, or we could choose a different one for
> deconstruction patterns:
>
> public deconstructor Point(int x, int y) { ... }
>
> Alternately, we could reach for a symbol to indicate that we are
> talking about
> an inverted member. C++ fans might suggest
>
> public ~Point(int x, int y) { ... }
>
> but this is too cryptic (it's evocative once you see it, but then it
> becomes
> less evocative as we move away from deconstructors towards instance
> patterns.)
>
> If we wish to offer finer-grained control over conditionality, we might
> additionally need a `total` / `partial` modifier, though I would
> prefer to avoid
> that.
>
> Of the keyword candidates, there is one that stands out (for good and bad)
> because it connects to something that is already in the language:
> `pattern`. On
> the one hand, using the term `pattern` for the declaration is a slight
> abuse; on
> the other, users will immediately connect it with "ah, so that's how I
> make a
> new pattern" or "so that's what happens when I match against this
> pattern."
> (Lisps would resolve this tension by calling it `defpattern`.)
>
> The others (`matcher`, `view`, `extractor`, etc) are all made-up terms
> that
> don't connect to anything else in the language, for better or worse.
> If we pick
> one of these, we are asking users to sort out _three_ separate new
> things in
> their heads: (use-site) patterns, (declaration-site) matchers, and the
> rules of
> how patterns and matchers are connected. Calling them both
> "patterns", despite
> the mild abuse of terminology, ties them together in a way that
> recognizes their
> connection.
>
> My personal position: `pattern` is the strongest candidate here,
> despite some
> flaws.
>
> ### Binding lists and match candidates
>
> There are two obvious alternatives for describing the binding list and
> match
> candidate of a pattern declaration, both with their roots in the
> constructor and
> method syntax:
>
> - Pretend that a pattern declaration is like a method with multiple
> return, and
> put the binding list in the "return position", and make the match
> candidate
> an ordinary parameter;
> - Lean into the inverse relationship between constructors and methods
> (and
> consistency with the use-site syntax), and put the binding list in the
> "parameter list position". For static patterns and some instance
> patterns,
> which need to explicitly identify the match candidate type, there
> are several
> sub-options:
> - Lean further into the duality, putting the match candidate type
> in the
> "return position";
> - Put the match candidate type somewhere else, where it is less
> likely to be
> confused for a method return.
>
> The "method-like" approach might look like this:
>
> ```
> class Point {
> // Constructor and deconstructor
> public Point(int x, int y) { ... }
> public pattern (int x, int y) Point(Point target) { ... }
> ...
> }
>
> class Optional<T> {
> // Static factory and pattern
> public static<T> Optional<T> of(T t) { ... }
> public static<T> pattern (T t) of(Optional<T> target) { ... }
> ...
> }
> ```
>
> The "inverse" approach might look like:
>
> ```
> class Point {
> // Constructor and deconstructor
> public Point(int x, int y) { ... }
> public pattern Point(int x, int y) { ... }
> ...
> }
>
> class Optional<T> {
> // Static factory and pattern (using the first sub-option)
> public static<T> Optional<T> of(T t) { ... }
> public static<T> pattern Optional<T> of(T t) { ... }
> ...
> }
> ```
>
> With the "method-like" approach, the match candidate gets an explicit name
> selected by the author; with the inverse approach, we can go with a
> predefined
> name such as `that`. (Because deconstructors do not have receivers,
> we could by
> abuse of notation arrange for the keyword `this` to refer instead to
> the match
> candidate within the body of a deconstructor. While this might seem
> to lead to
> a more familiar notation for writing deconstructors, it would create a
> gratuitous asymmetry between the bodies of deconstruction patterns and
> those of
> other patterns.)
>
> Between these choices, nearly all the considerations favor the "inverse"
> approach:
>
> - The "inverse" approach makes the declaration look like the use
> site. This
> highlights that `pattern Point(int x, int y)` is what gets invoked
> when you
> match against the pattern use `Point(int x, int y)`. (This point is so
> strong that we should probably just stop here.)
> - The "inverse" members also look like their duals; the only
> difference is the
> `pattern` keyword (and possibly the placement of the match
> candidate type).
> This makes matched pairs much more obvious, and such matched pairs
> will be
> critical both for future language features and for library idioms.
> - The method-like approach is suggestive of multiple return or
> tuples, which is
> probably helpful for the first few minutes but actually harmful in
> the long
> term. This feature is _not_ (much as some people would like to
> believe) about
> multiple return or tuples, and playing into this misperception will
> only make
> it harder to truly understand. So this suggestion ends up propping
> up the
> wrong mental model.
>
> The main downside of the "inverse" approach is the one-time speed bump
> of the
> unfamiliarity of the inverted syntax. (The "method-like" syntax also
> has its
> own speed bumps, it is just unfamiliar in different ways.) But unlike the
> advantages of the inverse approach, which continue to add value
> forever, this
> speed bump is a one-time hurdle to get over.
>
> To smooth out the speed bumps of the inverse approach, we can consider
> moving
> the position of the match candidate for static and (suitable) instance
> pattern
> declarations, such as:
>
> ```
> class Optional<T> {
> // the usual static factory
> public static<T> Optional<T> of(T t) { ... }
>
> // Various ways of writing the corresponding pattern
> public static<T> pattern of(T t) for Optional<T> { ... }
> // or ...
> public static<T> pattern(Optional<T>) of(T t) { ... }
> // or ...
> public static<T> pattern(Optional<T> that) of(T t) { ... }
> // or ...
> public static<T> pattern<Optional<T>> of(T t) { ... }
> ...
> }
> ```
>
> (The deconstructor example looks the same with either variant.) Of these,
> treating the match candidate like a "parameter" of "pattern" is
> probably the
> most evocative:
>
> ```
> public static<T> pattern(Optional<T> that) of(T t) { ... }
> ```
>
> as it can be read as "pattern taking the parameter `Optional<T> that`
> called
> `of`, binding `T`, and is a short departure from the inverse syntax.
>
> The main value of the various rearrangements is that users don't need
> to think
> about things operating in reverse to parse the syntax. This trades
> some of the
> secondary point (patterns looking almost exactly like their inverses)
> for a
> certain amount of cognitive load, while maintaining the most important
> consideration: that the declaration site look like the use site.
>
> For instance pattern declarations, if the match candidate type is the
> same as
> the receiver type, the match candidate type can be elided as it is with
> deconstructors.
>
> My personal position: the "multiple return" version is terrible; all the
> sub-variants of the inverse version are probably workable.
>
> ### Naming the match candidate
>
> We've been assuming so far that the match candidate always has a fixed
> name,
> such as `that`; this is an entirely workable approach. Some of the
> variants are
> also amenable to allowing authors to explicitly select a name for the
> match
> candidate. For example, if we put the match candidate as a
> "parameter" to the `pattern` keyword, there is an obvious place to put
> the name:
>
> ```
> static<T> pattern(Optional<T> target) of(T t) { ... }
> ```
>
> My personal opinion: I don't think this degree of freedom buys us
> much, and in
> the long run readability probably benefits by picking a fixed name
> like `that`
> and sticking with it. Even with a fixed name, if there is a sensible
> position
> for the name, allowing users to type `that` for explicitness is fine
> (as we do
> with instance methods, though many people don't know this.) We may
> even want to
> require it.
>
> ## Body types
>
> Just as there are two obvious approaches for the declaration, there
> are two
> obvious approaches we could take for the body (though there is some
> coupling
> between them.) We'll call the two body approaches _imperative_ and
> _functional_.
>
> The imperative approach treats bindings as initially-DU variables that
> must be
> DA on successful completion, getting their value through ordinary
> assignment;
> the functional approach sets all the bindings at once, positionally.
> Either
> way, member patterns (except maybe deconstructors) also need a way to
> differentiate a successful match from a failed match.
>
> Here is the `Point` deconstructor with both imperative and functional
> style. The
> functional style uses a placeholder `match` statement to indicate a
> successful
> match and provision of bindings:
>
> ```
> class Point {
> int x, y;
>
> Point(int x, int y) {
> this.x = x;
> this.y = y;
> }
>
> // Imperative style, deconstructor always succeeds
> pattern Point(int x, int y) {
> x = that.x;
> y = that.y;
> }
>
> // Functional style
> pattern Point(int x, int y) {
> match(that.x, that.y);
> }
> }
> ```
>
> There are some obvious differences here. In the imperative style, the
> dtor body
> looks much more like the reverse of the ctor body. The functional
> style is more
> concise (and amenable to further concision via the "concise method bodies"
> mechanism in the future), as well as a number of less obvious
> differences. For
> deconstructors, the imperative approach is likely to feel more natural
> because
> of the obvious symmetry with constructors.
>
> In reality, it is _premature at this point to have an opinion_, because we
> haven't yet seen the full scope of the problem; deconstructors are a
> special
> case in many ways, which almost surely is distorting our initial
> opinion. As we
> move towards conditional patterns (and pattern lambdas), our opinions
> may flip.
>
> Regardless of which we pick, there are some additional syntactic
> choices to be
> made -- what syntax to use to indicate success (we used `match` in the
> above
> example) or failure. (We should be especially careful around trying
> to reuse
> words like `return`, `break`, or `yield` because, in the case where
> there are
> zero bindings (which is allowable), it becomes unclear whether they
> mean "fail"
> or "succeed with zero bindings".)
>
> ### Success and failure
>
> Except for possibly deconstructors, which we may require to be total,
> a pattern
> declaration needs a way to indicate success and failure. In the
> examples above,
> we posited a `match` statement to indicate success in the functional
> approach,
> and in both examples leaned on the "implicit success" of
> deconstructors (under
> the assumption they always succeed). Now let's look at the more
> general case to
> figure out what else is needed.
>
> For a static pattern like `Optional::of`, success is conditional. Using
> `match-fail` as a placeholder for "the match failed", this might look like
> (functional version):
>
> ```
> public static<T> pattern(Optional<T> that) of(T t) {
> if (that.isPresent())
> match (that.get());
> else
> match-fail;
> }
> ```
>
> The imperative version is less pretty, though. Using `match-success` as a
> placeholder:
>
> ```
> public static<T> pattern(Optional<T> that) of(T t) {
> if (that.isPresent()) {
> t = that.get();
> match-success;
> }
> else
> match-fail;
> }
> ```
>
> Both arms of the `if` feel excessively ceremonial here. And if we
> chose to not
> make all deconstruction patterns unconditional, deconstructors would
> likely need
> some explicit success as well:
>
> ```
> pattern Point(int x, int y) {
> x = that.x;
> y = that.y;
> match-success;
> }
> ```
>
> It might be tempting to try and eliminate the need for explicit success by
> inferring it from whether or not the bindings are DA or not, but this is
> error-prone, is less type-checkable, and falls apart completely for
> patterns
> with no bindings.
>
> ### Implicit failure in the functional approach
>
> One of the ceremonial-seeming aspects of `Optional::of` above is
> having to say
> `else match-fail`, which doesn't feel like it adds a lot of value.
> Perhaps we
> can be more concise without losing clarity.
>
> Most conditional patterns will have a predicate to determine matching,
> and then
> some conditional code to compute the bindings and claim success.
> Having to say
> "and if the predicate didn't hold, then I fail" seems like ceremony
> for the
> author and noise for the reader. Instead, if a conditional pattern
> falls off
> the end without matching, we could treat that as simply not matching:
>
> ```
> public static<T> pattern(Optional<T> that) of(T t) {
> if (that.isPresent())
> match (that.get());
> }
> ```
>
> This says what we mean: if the optional is present, then this pattern
> succeeds
> and bind the contents of the `Optional`. As long as our "succeed"
> construct
> strongly enough connotes that we are terminating abruptly and
> successfully, this
> code is perfectly clear. And most conditional patterns will look a
> lot like
> `Optional::of`; do some sort of test and if it succeeds, extract the
> state and
> bind it.
>
> At first glance, this "implicit fail" idiom may seem error-prone or
> sloppy. But
> after writing a few dozen patterns, one quickly tires of saying "else
> match-fail" -- and the reader doesn't necessarily appreciate reading
> it either.
>
> Implicit failure also simplifies the selection of how we explicitly
> indicate
> failure; using `return` in a pattern for "no match" becomes pretty
> much a forced
> move. We observe that (in a void method), "return" and "falling off
> the end"
> are equivalent; if "falling off the end" means "no match", then so
> should an
> explicit `return`. So in those few cases where we need to explicitly
> signal "no
> match", we can just use `return`. It won't come up that often, but
> here's an
> example where it does:
>
> ```
> static pattern(int that) powerOfTwo(int exp) {
> int exp = 0;
>
> if (that < 1)
> return; // explicit fail
>
> while (that > 1) {
> if (that % 2 == 0) {
> that /= 2;
> ++exp;
> }
> else
> return; // explicit fail
> }
> match (exp);
> }
> ```
>
> As a bonus, if `return` as match failure is a forced move, we need
> only select a
> term for "successful match" (which obviously can't be `return`). We
> could use
> `match` as we have in the examples, or a variant like `matched` or
> `matches`.
> But rather than just creating a new control operator, we have an
> opportunity to
> lean into the duality a little harder, by including the pattern syntax
> in the
> match:
>
> ```
> matches of(that.get());
> ```
>
> or the (optionally?) qualified (inferring type arguments, as we do at
> the use
> site):
>
> ```
> matches Optional.of(that.get());
> ```
>
> These "use the name" approaches trades a small amount of verbosity to
> gain a
> higher degree of fidelity to the pattern use site (and to evoke the
> comethod
> completion.)
>
> If we don't choose "implicit fail", we would have to invent _two_ new
> control
> flow statements to indicate "success" and "failure".
>
> My personal position: for the functional approach, implicit failure
> both makes
> the code simpler and clearer, and after you get used to it, you don't
> want to go
> back. Whether we say `match` or `matches` or `matches <pattern-name>`
> are all
> workable, though I like some variant that names the pattern.
>
> ### Implicit success in the imperative approach
>
> In the imperative approach, we can be implicit as well, but it feels more
> natural (at least, initially) to choose implicit success rather than
> failure.
> This works great for unconditional patterns:
>
> ```
> pattern Point(int x, int y) {
> x = that.x;
> y = that.y;
> // implicit success
> }
> ```
>
> but not quite as well for conditional patterns:
>
> ```
> static<T> pattern(Optional<T> that) of(T t) {
> if (that.isPresent()) {
> t = that.get();
> }
> else
> match-fail;
> // implicit success
> }
> ```
>
> We can eliminate one of the arms of the if, with the more concise (but
> convoluted) inversion:
>
> ```
> static<T> pattern(Optional<T> that) of(T t) {
> if (!that.isPresent())
> match-fail;
> t = that.get();
> // implicit success
> }
> ```
>
> Just as with the functional approach, if we choose imperative and
> "implicit
> success", using `return` to indicate success is pretty much a forced
> move.
>
> ### Imperative is a trap
>
> If we assume that functional implies implicit failure, and imperative
> implies
> implicit success, then our choices become:
>
> ```
> class Optional<T> {
> public static<T> Optional<T> of(T t) { ... }
>
> // imperative, implicit success
> public static<T> pattern(Optional<T> that) of(T t) {
> if (that.isPresent()) {
> t = that.get();
> }
> else
> match-fail;
> }
>
> // functional, implicit failure
> public static<T> pattern(Optional<T> that) of(T t) {
> if (that.isPresent())
> matches of(that.get());
> }
> }
> ```
>
> Once we get past deconstructors, the imperative approach looks worse by
> comparison because we need to assign all the bindings (which is _O(n)_
> assignments) _and also_ indicate success or failure somehow, whereas
> in the
> functional style all can be done together with a single `matches`
> statement.
>
> Looking at the alternatives, except maybe for unconditional patterns, the
> functional example above seems a lot more natural. The imperative
> approach
> works with deconstructors (assuming they are not conditional), but
> does not
> scale so well to conditionality -- which is the essence of patterns.
>
> From a theoretical perspective, the method-comethod duality also gives
> us a
> forceful nudge towards the functional approach. In a method, the method
> arguments are specified as a positional list of expressions at the use
> site:
>
> m(a, b, c)
>
> and these values are invisibly copied into the parameter slots of the
> method
> prior to frame activation. The dual to that for a comethod to
> similarly convey
> the bindings in a positional list of expressions (as they must either
> all be
> produced or none), where they are copied into the slots provided at
> the use
> site, as is indicated by `matches` in the above examples.
>
> My personal position: the imperative style feels like a trap. It seems
> "obvious" at first if we start with deconstructors, but becomes
> increasingly
> difficult when we get past this case, and gets in the way of other
> opportunities. The last gasp before acceptance is the discomfort that
> dtor and
> ctor bodies are written in different styles, but in the rear-view
> mirror, this
> feels like a non-issue.
>
> ### Derive imperative from functional?
>
> If we start with "functional with implicit failure", we can possibly
> rescue
> imperative by deriving a version of imperative from functional, by
> "overloading"
> the match-success operator.
>
> If we have a pattern whose binding names are `b1..bn` of types
> `B1..Bn`, then
> the `matches` operator must take a list of expressions `e1..en` whose
> arity and
> types are compatible with `B1..Bn`. But we could allow `matches` to
> also have a
> nilary form, which would have the effect of being shorthand for
>
> matches <pattern-name>(b1, b2, ..., bn)
>
> where each of `b1..bn` must be DA at the point of matching. This means
> that we
> could express patterns in either form:
>
> ```
> class Optional<T> {
> public static<T> Optional<T> of(T t) { ... }
>
> // imperative, derived from functional with implicit failure
> public static<T> pattern(Optional<T> that) of(T t) {
> if (that.isPresent()) {
> t = that.get();
> matches of;
> }
> }
>
> public static<T> pattern(Optional<T> that) of(T t) {
> if (that.isPresent())
> matches of(that.get());
> }
> }
> ```
>
> This flexibility allows users to select a more verbose expression in
> exchange
> for a clearer association of expressions and bindings, though as we'll
> see, it
> does come with some additional constraints.
>
> ### Wrapping an existing API
>
> Nearly every library has methods (sometimes sets of methods) that are
> patterns
> in disguise, such as the pair of methods `isArray` and
> `getComponentType` in
> `Class`, or the `Matcher` helper type in `java.util.regex`. Library
> maintainers
> will likely want to wrap (or replace) these with real patterns, so
> these can
> participate more effectively in conditional contexts, and in some cases,
> highlight their duality with factory methods.
>
> Matching a string against a `j.u.r.Pattern` regular expression has all
> the same
> elements as a pattern, just with an ad-hoc API (and one that I have to
> look up
> every time). But we can fairly easily wrap a true pattern around the
> existing
> API. To match against a `Pattern` today, we pass the match candidate to
> `Pattern::matcher`, which returns a `Matcher` with accessors
> `Matcher::matches`
> (did it match) and `Matcher::group` (conditionally extract a
> particular capture
> group.) If we want to wrap this with a pattern called `regexMatch`:
>
> ```
> pattern(String that) regexMatch(String... groups) {
> Matcher m = this.matcher(that);
> if (m.matches())
> matches Pattern.regexMatch(IntStream.range(1, m.groupCount())
> .map(Matcher::group)
> .toArray(String[]::new));
> // whole lotta matchin' goin' on
> }
> ```
>
> This says that a `j.u.r.Pattern` has an instance pattern called
> `regex`, whose
> match candidate is `String`, and which binds a varargs of `String`
> corresponding
> to the capture groups. The implementation simply delegates to the
> existing
> `j.u.r.Matcher` API. This means that `j.u.r.Pattern` becomes a sort
> of "pattern
> object", and we can use it as a receiver at the use site:
>
> ```
> static Pattern As = Pattern.compile("(a*)");
> static Pattern Bs = Pattern.compile("(b*)");
> ...
> switch (string) {
> case As.regexMatch(var as): ...
> case Bs.regexMatch(var bs): ...
> ...
> }
> ```
>
> ### Odds and ends
>
> There are a number of loose ends here. We could choose other names
> for the
> match-success and match-fail operations, including trying to reuse
> `break` or
> `yield`. But, this reuse is tricky; it must be very clear whether a
> given form
> of abrupt completion means "success" or "failure", because in the case of
> patterns with no bindings, we will have no other syntactic cues to help
> disambiguate. (I think having a single `matches`, with implicit
> failure and
> `return` meaning failure, is the sweet spot here.)
>
> Another question is whether the binding list introduces corresponding
> variables
> into the scope of the body. For imperative, the answer is "surely
> yes"; for
> functional, the answer is "maybe" (unless we want to do the trick where we
> derive imperative from functional, in which case the answer is "yes"
> again.)
>
> If the binding list does not correspond to variables in the body, this
> may be
> initially discomforting; because they do not declare program elements,
> they may
> feel that they are left "dangling". But even if they are not declaring
> _program_ elements, they are still declaring _API_ elements (similar
> to the
> return type of a method.) We will want to provide Javadoc on the
> bindings, just
> like with parameters; we will want to match up binding names in
> deconstructors
> with parameter names in constructors; we may even someday want to support
> by-name binding at the use site (e.g., `case Foo(a: var a)`). The
> names are
> needed for all of these, just not for the body. Names still matter.
> My take
> here is that this is a transient "different is scary" reaction, one
> that we
> would get over quickly.
>
> A final question is whether we should consider unqualified names as
> implicitly
> qualified by `that` (and also `this`, for instance patterns, with some
> conflict
> resolution). Users will probably grow tired of typing `that.` all the
> time, and most of the time, the unqualified use is perfectly readable.
>
> ## Exhaustiveness
>
> There is one last syntax question in front of us: how to indicate that
> a set of
> patterns are (claimed to be) exhaustive on a given match candidate
> type. We see
> this with `Optional::of` and `Optional::empty`; it would be sad if the
> compiler
> did not realize that these two patterns together were exhaustive on
> `Optional`.
> This is not a feature that will be used often, but not having it at
> all will be
> a repeated irritant.
>
> The best I've come up with is to call these `case` patterns, where a
> set of
> `case` patterns for a given match candidate type in a given class are
> asserted
> to be an exhaustive set:
>
> ```
> class Optional<T> {
> static<T> Optional<T> of(T t) { ... }
> static<T> Optional<T> empty() { ... }
>
> static<T> case pattern of(T t) for Optional<T> { ... }
> static<T> case pattern empty() for Optional<T> { ... }
> }
> ```
>
> Because they may not be truly exhaustive, `switch` constructs will
> have to back
> up the static assumption of exhaustiveness with a dynamic check, as we
> do for
> other sets of exhaustive patterns that may have remainder.
>
> I've experimented with variants of `sealed` but it felt more forced,
> so this is
> the best I've come up with.
>
> ## Example: patterns delegating to other patterns
>
> Pattern implementations must compose. Just as a subclass constructor
> delegates
> to a superclass constructor, the same should be true for deconstructors.
> Here's a typical superclass-subclass pair:
>
> ```
> class A {
> private final int a;
>
> public A(int a) { this.a = a; }
> public pattern A(int a) { matches A(that.a); }
> }
>
> class B extends A {
> private final int b;
>
> public B(int a, int b) {
> super(a);
> this.b = b;
> }
>
> // Imperative style
> public pattern B(int a, int b) {
> if (that instanceof super(var aa)) {
> a = aa;
> b = that.b;
> matches B;
> }
> }
>
> // Functional style
> public pattern B(int a, int b) {
> if (that instanceof super(var a))
> matches B(a, b);
> }
> }
> ```
>
> (Ignore the flow analysis and totality for the time being; we'll come
> back to
> this in a separate document.)
>
> The first thing that jumps out at us is that, in the imperative
> version, we had
> to create a "garbage" variable `aa` to receive the binding, because
> `a` was
> already in scope, and then we have to copy the garbage variable into
> the real
> binding variable. Users will surely balk at this, and rightly so. In the
> functional version (depending on the choices from "Odds and Ends") we
> are free
> to use the more natural name and avoid the roundabout locution.
>
> We might be tempted to fix the "garbage variable" problem by inventing
> another
> sub-feature: the ability to use an existing variable as the target of
> a binding,
> such as:
>
> ```
> pattern Point(int a, int b) {
> if (this instanceof A(__bind a))
> b = this.b;
> }
> ```
>
> But, I think the language is stronger without this feature, for two
> reasons.
> First, having to reason about whether a pattern match introduces a new
> binding
> or assigns to an existing variables is additional cognitive load for
> users to
> reason about, and second, having assignment to locals happening through
> something other than assignment introduces additional complexity in
> finding
> where a variable is modified. While we can argue about the general
> utility of
> this feature, bringing it in just to solve the garbage-variable problem is
> particularly unattractive.
>
> ## Pattern lambdas
>
> One final consideration is is that patterns may also have a lambda
> form. Given
> a single-abstract-pattern (SAP) interface:
>
> ```
> interface Converter<T,U> {
> pattern(T t) convert(U u);
> }
> ```
>
> one can implement such a pattern with a lambda. Such a lambda has one
> parameter
> (the match candidate), and its body looks like the body of a declared
> pattern:
>
> ```
> Converter<Integer, Short> c =
> i -> {
> if (i >= Short.MIN_VALUE && i <= Short.MAX_VALUE)
> matches Converter.convert((short) i);
> };
> ```
>
> Because the bindings of the pattern lambda are defined in the
> interface, not in
> the lambda, this is one more reason not to like the imperative
> version: it is
> brittle, and alpha-renaming bindings in the interface would be a
> source-incompatible change.
>
> ## Example gallery
>
> Here's all the pattern examples so far, and a few more, using the
> suggested
> style (functional, implicit fail, implicit `that`-qualification):
>
> ```
> // Point dtor
> pattern Point(int x, int y) {
> matches Point(x, y);
> }
>
> // Optional -- static patterns for Optional::of, Optional::empty
> static<T> case pattern(Optional<T> that) of(T t) {
> if (isPresent())
> matches of(t);
> }
>
> static<T> case pattern(Optional<T> that) empty() {
> if (!isPresent())
> matches empty();
> }
>
> // Class -- instance pattern for arrayClass (match candidate type
> inferred)
> pattern arrayClass(Class<?> componentType) {
> if (that.isArray())
> matches arrayClass(that.getComponentType());
> }
>
> // regular expression -- instance pattern in j.u.r.Pattern
> pattern(String that) regexMatch(String... groups) {
> Matcher m = matcher(that);
> if (m.matches())
> matches Pattern.regexMatch(IntStream.range(1, m.groupCount())
> .map(Matcher::group)
> .toArray(String[]::new));
> }
>
> // power of two (somewhere)
> static pattern(int that) powerOfTwo(int exp) {
> int exp = 0;
>
> if (that < 1)
> return;
>
> while (that > 1) {
> if (that % 2 == 0) {
> that /= 2;
> exp++;
> }
> else
> return;
> }
> matches powerOfTwo(exp);
> }
> ```
>
> ## Closing thoughts
>
> I came out of this exploration with very different conclusions than I
> expected
> when going in. At first, the "inverse" syntax seemed stilted, but
> over time it
> started to seem more obvious. Similarly, I went in expecting to
> prefer the
> imperative approach for the body, but over time, started to warm to the
> functional approach, and eventually concluded it was basically a
> forced move if
> we want to support more than just deconstructors. And I started out
> skeptical
> of "implicit fail", but after writing a few dozen patterns with it,
> going back
> to fully explicit felt painful. All of this is to say, you should
> hold your
> initial opinions at arm's length, and give the alternatives a chance
> to sink in.
>
> For most _conditional_ patterns (and conditionality is at the heart of
> pattern
> matching), the functional approach cleanly highlights both the match
> predicate
> and the flow of values, and is considerably less fussy than the imperative
> approach in the same situation; `Optional::of`, `Class::arrayClass`,
> and `regex`
> look great here, much better than the would with imperative. None of these
> illustrate delegation, but in the presence of delegation, the gap gets
> even
> wider.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/amber-spec-observers/attachments/20240401/ef356a51/attachment-0001.htm>
More information about the amber-spec-observers
mailing list