# Lists * * * <i> Topics: * lists * recursive functions on lists * pattern matching * tail recursion </i> * * * ## Lists An OCaml list is a sequence of values all of which have the same type. They are implemented as singly-linked lists. These lists enjoy a first-class status in the language: there is special support for easily creating and working with lists. That's a characteristic that OCaml shares with many other functional languages. Mainstream imperative languages, like Python, have such support these days too. Maybe that's because programmers find it so pleasant to work directly with lists as a first-class part of the language, rather than having to go through a library (as in C and Java). ## Building lists **Syntax.** There are three syntactic forms for building lists: ``` [] e1::e2 [e1; e2; ...; en] ``` The empty list is written `[]` and is pronounced "nil", a name that comes from Lisp. Given a list `lst` and element `elt`, we can prepend `elt` to `lst` by writing `elt::lst`. The double-colon operator is pronounced "cons", a name that comes from an operator in Lisp that <u>cons</u>tructs objects in memory. "Cons" can also be used as a verb, as in "I will cons an element onto the list." The first element of a list is usually called its *head* and the rest of the elements (if any) are called its *tail*. The square bracket syntax is convenient but unnecessary. Any list `[e1; e2; ...; en]` could instead be written with the more primitive nil and cons syntax: `e1::e2::...::en::[]`. When a pleasant syntax can be defined in terms of a more primitive syntax within the language, we call the pleasant syntax *syntactic sugar*: it makes the language "sweeter". Transforming the sweet syntax into the more primitive syntax is called *desugaring*. Because the elements of the list can be arbitrary expressions, lists can be nested as deeply as we like, e.g., `[ [[]]; [[1;2;3]]]` **Dynamic semantics.** * `[]` is already a value. * if `e1` evaluates to `v1`, and if `e2` evaluates to `v2`, then `e1::e2` evaluates to `v1::v2`. As a consequence of those rules and how to desugar the square-bracket notation for lists, we have the following derived rule: * if `ei` evaluates to `vi` for all `i` in `1..n`, then `[e1; ...; en]` evaluates to `[v1; ...; vn]`. It's starting to get tedious to write "evaluates to" in all our evaluation rules. So let's introduce a shorter notation for it. We'll write `e ==> v` to mean that `e` evaluates to `v`. Note that `==>` is not a piece of OCaml syntax. Rather, it's a notation we use in our description of the language, kind of like metavariables. Using that notation, we can rewrite the latter two rules above: * if `e1 ==> v1`, and if `e2 ==> v2`, then `e1::e2 ==> v1::v2`. * if `ei ==> vi` for all `i` in `1..n`, then `[e1; ...; en] ==> [v1; ...; vn]`. **Static semantics.** All the elements of a list must have the same type. If that element type is `t`, then the type of the list is `t list`. You should read such types from right to left: `t list` is a list of `t`'s, `t list list` is a list of list of `t`'s, etc. The word `list` itself here is not a type: there is no way to build an OCaml value that has type simply `list`. Rather, `list` is a *type constructor*: given a type, it produces a new type. For example, given `int`, it produces the type `int list`. You could think of type constructors as being like functions that operate on types, instead of functions that operate on values. The type-checking rules: * `[] : 'a list` * if `e1 : t` and `e2 : t list` then `e1::e2 : t list`. In the rule for `[]`, recall that `'a` is a type variable: it stands for an unknown type. So the empty list is a list whose elements have an unknown type. If we cons an `int` onto it, say `2::[]`, then the compiler infers that for that particular list, `'a` must be `int`. But if in another place we cons a `bool` onto it, say `true::[]`, then the compiler infers that for that particular list, `'a` must be `bool`. ## Accessing lists There are really only two ways to build a list, with nil and cons. So if we want to take apart a list into its component pieces, we have to say what to do with the list if it's empty, and what to do if it's non-empty (that is, a cons of one element onto some other list). We do that with a language feature called *pattern matching*. Here's an example of using pattern matching to compute the sum of a list: ``` let rec sum lst = match lst with | [] -> 0 | h::t -> h + sum t ``` This function says to take the input `lst` and see whether it has the same shape as the empty list. If so, return 0. Otherwise, if it has the same shape as the list `h::t`, then let `h` be the first element of `lst`, and let `t` be the rest of the elements of `lst`, and return `h + sum t`. The choice of variable names here is meant to suggest "head" and "tail" and is a common idiom, but we could use other names if we wanted. Another common idiom is: ``` let rec sum xs = match lst with | [] -> 0 | x::xs' -> x + sum xs' ``` That is, the input list is a list of xs (pronounced EX-uhs), the head element is an x, and the tail is xs' (pronounced EX-uhs prime). Here's another example of using pattern matching to compute the length of a list: ``` let rec length lst = match lst with | [] -> 0 | h::t -> 1 + length t ``` Note how we didn't actually need the variable `h` in the right-hand side of the pattern match. When we want to indicate the presence of some value in a pattern without actually giving it a name, we can write `_` (the underscore character): ``` let rec length lst = match lst with | [] -> 0 | _::t -> 1 + length t ``` And here's a third example that appends one list onto the beginning of another list: ``` let rec append lst1 lst2 = match lst1 with | [] -> lst2 | h::t -> h::(append t lst2) ``` For example, `append [1;2] [3;4]` is `[1;2;3;4]`. That function is actually available as a built-in operator `@`, so we could instead write `[1;2] @ [3;4]`. As a final example, we could write a function to determine whether a list is empty: ``` let empty lst = match lst with | [] -> true | h::t -> false ``` But there a much easier way to write the same function without pattern matching: ``` let empty lst = lst = [] ``` Note how all the recursive functions above are similar to doing proofs by induction on the natural numbers: every natural number is either 0 or is 1 greater than some other natural number \\(n\\), and so a proof by induction has a base case for 0 and an inductive case for \\(n+1\\). Likewise all our functions have a base case for the empty list and a recursive case for the list that has one more element than another list. This similarity is no accident. There is a deep relationship between induction and recursion; we'll explore that relationship in more detail later in the course. By the way, there are two library functions `List.hd` and `List.tl` that return the head and tail of a list. It is not good, idiomatic OCaml to apply these directly to a list. The problem is that they will raise an exception when applied to the empty list, and you will have to remember to handle that exception. Instead, you should use pattern matching: you'll then be forced to match against both the empty list and the non-empty list (at least), which will prevent exceptions from being raised, thus making your program more robust. ## Mutating lists Lists are immutable. There's no way to change an element of a list from one value to another. Instead, OCaml programmers create new lists out of old lists. For example, suppose we wanted to write a function that returned the same list as its input list, but with the first element (if there is one) incremented by 1. We could do that: ``` let inc_first lst = match lst with | [] -> [] | h::t -> (h+1)::t ``` Now you might be concerned about whether we're being wasteful of space. After all, there are at least two ways the compiler could implement the above code: 1. Copy the entire tail list `t` when the new list is created in the pattern match with cons, such that the amount of memory in use just increased by an amount proportionate to the length of `t`. 2. Share the tail list `t` between the old list and the new list, such that the amount of memory in use does not increase (beyond the one extra piece of memory needed to store `h+1`). In fact, the compiler does the latter. So there's no need for concern. The reason that it's quite safe for the compiler to implement sharing is exactly that list elements are immutable. If they were instead mutable, then we'd start having to worry about whether the list I have is shared with the list you have, and whether changes I make will be visible in your list. So immutability makes it easier to reason about the code, and makes it safe for the compiler to perform an optimization. ## Pattern matching We saw above how to access lists using pattern matching. Let's look more carefully at this feature. **Syntax.** ``` match e with | p1 -> e1 | p2 -> e2 | ... | pn -> en ``` Each of the clauses `pi -> ei` is called a *branch* or a *case* of the pattern match. The first vertical bar in the entire pattern match is optional. The `p`'s here are a new syntactic form called a *pattern*. For now, a pattern may be: * a variable name, e.g. `x` * the underscore character `_`, which is called the *wildcard* * the empty list `[]` * `p1::p2` * `[p1; ...; pn]` No variable name may appear more than once in a pattern. For example, the pattern `x::x` is illegal. The wildcard may occur any number of times. As we learn more of data structures available in OCaml, we'll expand the possibilities for what a pattern may be. **Dynamic semantics.** In lecture we gave an abbreviated version of the dynamic semantics. Here we give the full details. Pattern matching involves two inter-related tasks: determining whether a pattern matches a value, and determining what parts of the value should be associated with which variable names in the pattern. The former task is intuitively about determining whether a pattern and a value have the same *shape*. The latter task is about determining the *variable bindings* introduced by the pattern. For example, in ``` match 1::[] with | [] -> false | h::t -> (h>=1) && (length t = 0) ``` (which evaluates to `true`) when evaluating the right-hand side of the second branch, `h=1` and `t=[]`. Let's write `h->1` to mean the variable binding saying that `h` has value `1`; this is not a piece of OCaml syntax, but rather a notation we use to reason about the language. So the variable bindings produced by the second branch would be `h->1,t->[]`. More carefully, here is a definition of when a pattern matches a value and the bindings that match produces: * The pattern `x` matches any value `v` and produces the variable binding `x->v`. * The pattern `_` matches any value and produces no bindings. * The pattern `[]` matches the value `[]` and produces no bindings. * If `p1` matches `v1` and produces a set \\(b_1\\) of bindings, and if `p2` matches `v2` and produces a set \\(b_2\\) of bindings, then `p1::p2` matches `v1::v2` and produces the set \\(b_1 \cup b_2\\) of bindings. Note that `v2` must be a list (since it's on the right-hand side of `::`) and could have any length: 0 elements, 1 element, or many elements. Note that the union \\(b_1 \cup b_2\\) of bindings will never have a problem where the same variable is bound separately in both \\(b_1\\) and \\(b_2\\) because of the syntactic restriction that no variable name may appear more than once in a pattern. * If for all `i` in `1..n`, it holds that `pi` matches `vi` and produces the set \\(b_i\\) of bindings, then `[p1; ...; pn]` matches `[v1; ...; vn]` and produces the set \\(\bigcup_i b_i\\) of bindings. Note that this pattern specifies the exact length the list must be. Now we can can say how to evaluate `match e with p1 -> e1 | ... | pn -> en`: * Evaluate `e` to a value `v`. * Match `v` against `p1`, then against `p2`, and so on, in the order they appear in the match expression. * If `v` does not match against any of the patterns, then evaluation of the match expression raises a `Match_failure` exception. We haven't yet discussed exceptions in OCaml, but you're familiar with them from CS 1110 (Python) and CS 2110 (Java). We'll come back to exceptions after we've covered some of the other built-in data structures in OCaml. * Otherwise, stop trying to match at the first time a match succeeds against a pattern. Let `pi` be that pattern and let \\(b\\) be the variable bindings produced by matching `v` against `pi`. * Substitute those bindings inside `ei`, producing a new expression `e'`. * Evaluate `e'` to a value `v'`. * The result of the entire match expression is `v'`. For example, here's how this match expression would be evaluated: ``` match 1::[] with | [] -> false | h::t -> (h=1) && (t=[]) ``` * `1::[]` is already a value * `[]` does not match ``1::[]`` * `h::t` does match `1::[]` and produces variable bindings \\(\\{\\)`h->1`,`t->[]`\\(\\}\\), because: - `h` matches `1` and produces the variable binding `h->1` - `t` matches `[]` and produces the variable binding `t=[]` * substituting \\(\\{\\)`h->1`,`t->[]`\\(\\}\\) inside `(h=1) && (t=[])` produces a new expression `(1=1) && ([]=[])` * evaluating `(1=1) && ([]=[])` yields the value `true` (we omit the justification for that fact here, but it follows from other evaluation rules for built-in operators and function application) * so the result of the entire match expression is `true`. **Static semantics.** * If `e:ta` and for all `i`, it holds that `pi:ta` and `ei:tb`, then `(match e with p1 -> e1 | ... | pn -> en) : tb`. That rule relies on being able to judge whether a pattern has a particular type. As usual, type inference comes into play here. The OCaml compiler infers the types of any pattern variables as well as all occurrences of the wildcard pattern. As for the list patterns, they have the same type-checking rules as list expressions. In addition to that type-checking rule, there are two other checks the compiler does for each match expression: * **Exhaustiveness:** the compiler checks to make sure that there are enough patterns to guarantee that at least one of them matches the expression `e`, no matter what the value of that expression is at run time. This ensures that the programmer did not forget any branches. For example, the function below will cause the compiler to emit a warning: ``` # let head lst = match lst with h::_ -> h;; Warning 8: this pattern-matching is not exhaustive. Here is an example of a value that is not matched: [] ``` By presenting that warning to the programmer, the compiler is helping the programmer to defend against the possibility of `Match_failure` exceptions at runtime. * **Unused branches:** the compiler checks to see whether any of the branches could never be matched against because one of the previous branches is guaranteed to succeed. For example, the function below will cause the compiler to emit a warning: ``` # let rec sum lst = match lst with | h::t -> h + sum t | [h] -> h | [] -> 0;; Warning 11: this match case is unused. ``` The second branch is unused because the first branch will match anything the second branch matches. Unused match cases are usually a sign that the programmer wrote something other than what they intended. So by presenting that warning, the compiler is helping the programmer to detect latent bugs in their code. Here's an example of one of the most common bugs that causes an unused match case warning. Understanding it is also a good way to check your understanding of the dynamic semantics of match expressions: ``` let length_is lst n = match length lst with | n -> true | _ -> false ``` The programmer was thinking that if the length of `lst` is equal to `n`, then this function will return `true`, and otherwise will return `false`. But in fact this function *always* returns `true`. Why? Because the pattern variable `n` is distinct from the function argument `n`. Suppose that the length of `lst` is 5. Then the pattern match becomes: `match 5 with n -> true | _ -> false`. Does `n` match 5? Yes, according to the rules above: a variable pattern matches any value and here produces the binding `n->5`. Then evaluation applies that binding to `true`, substituting all occurrences of `n` inside of `true` with 5. Well, there are no such occurrences. So we're done, and the result of evaluation is just `true`. What the programmer really meant to write was: ``` let length_is lst n = match length lst with | m -> if m=n then true else false | _ -> false ``` or better yet: ``` let length_is lst n = match length lst with | m -> m=n | _ -> false ``` or even better yet: ``` let length_is lst n = length lst = n ``` ## Tail recursion A function is *tail recursive* if it calls itself recursively but does not perform any computation after the recursive call returns, and immediately returns to its caller the value of its recursive call. Consider these two implementations, `sum` and `sum_tr` of summing a list, where we've provided some type annotations to help you understand the code: ``` let rec sum (l : int list) : int = match l with [] -> 0 | x :: xs -> x + (sum xs) let rec sum_plus_acc (acc : int) (l : int list) : int = match l with [] -> acc | x :: xs -> sum_plus_acc (acc + x) xs let sum_tr : int list -> int = sum_plus_acc 0 ``` Observe the following difference between the `sum` and `sum_tr` functions above: In the `sum` function, which is not tail recursive, after the recursive call returned its value, we add `x` to it. In the tail recursive `sum_tr`, or rather in `sum_plus_acc`, after the recursive call returns, we immediately return the value without further computation. Why do we care about tail recursion? Actually, sometimes functional programmers fixate a bit too much upon it. If all you care about is writing the first draft of a function, you probably don't need to worry about it. But if you're going to write functions on really long lists, tail recursion becomes important for performance. Recall (from CS 1110) that there is a call stack, which is a stack (the data structure with push and pop operations) with one element for each function call that has been started but has not yet completed. Each element stores things like the value of local variables and what part of the function has not been evaluated yet. When the evaluation of one function body calls another function, a new element is pushed on the call stack and it is popped off when the called function completes. When a function makes a recursive call to itself and there is nothing more for the caller to do after the callee returns (except return the callee's result), this situation is called a tail call. Functional languages like OCaml (and even imperative languages like C++) typically include an hugely useful optimization: when a call is a tail call, the caller's stack-frame is popped before the call&mdash;the callee's stack-frame just replaces the caller's. This makes sense: the caller was just going to return the callee's result anyway. With this optimization, recursion can sometimes be as efficient as a while loop in imperative languages (such loops don't make the call-stack bigger.) The "sometimes" is exactly when calls are tail calls&mdash;something both you and the compiler can (often) figure out. With tail-call optimization, the space performance of a recursive algorithm can be reduced from \\(O(n)\\) to \\(O(1)\\), that is, from one stack frame per call to a single stack frame for all calls. So when you have a choice between using a tail-recursive vs. non-tail-recursive function, you are likely better off using the tail-recursive function on really long lists to achieve space efficiency. For that reason, the List module documents which functions are tail recursive and which are not. But that doesn't mean that a tail-recursive implementation is strictly better. For example, the tail-recursive function might be harder to read. (Consider `sum_plus_acc`.) Also, there are cases where implementing a tail-recursive function entails having to do a pre- or post-processing pass to reverse the list. On small to medium sized lists, the overhead of reversing the list (both in time and in allocating memory for the reversed list) can make the tail-recursive version less time efficient. What constitutes "small" vs. "big" here? That's hard to say, but maybe 10,000 is a good estimate, according to the [standard library documentation of the `List` module][list]. [list]: http://caml.inria.fr/pub/docs/manual-ocaml/libref/List.html ## Bonus syntax Here are a couple additional pieces of syntax related to lists and pattern matching. **Immediate matches:** When you have a function that immediately pattern-matches against its final argument, there's a nice piece of syntactic sugar you can use to avoid writing extra code. Here's an example: instead of ``` let rec sum lst = match lst with | [] -> 0 | h::t -> h + sum t ``` you can write ``` let rec sum = function | [] -> 0 | h::t -> h + sum t ``` The word `function` is a keyword. Notice that we're able to leave out the line containing `match` as well as the name of the argument, which was never used anywhere else but that line. In such cases, though, it's especially important in the specification comment for the function to document what that argument is supposed to be, since the code no longer gives it a descriptive name. **OCamldoc:** OCamldoc is a documentation generator similar to Javadoc. It extracts comments from source code and produces HTML (as well as other output formats). The [standard library web documentation][std-web] for the List module is generated by OCamldoc from the [standard library source code][std-src] for that module, for example. We won't stress OCamldoc in this course, but we do bring it up here for a reason: there is a syntactic convention in it that can be confusing with respect to lists. In an OCamldoc comment, source code is surrounded by square brackets. That code will be rendered in typewriter face and syntax-highlighted in the output HTML. The square brackets in this case do not indicate a list. For example, here is the comment for `List.hd` in the source code: ``` (** Return the first element of the given list. Raise [Failure "hd"] if the list is empty. *) ``` The `[Failure "hd"]` does not mean a list containing the exception `Failure "hd"`. Rather it means to typeset the expression `Failure "hd"` as source code, as you can see [here][std-web]. This can get especially confusing when you want to talk about lists as part of the documentation. For example, here is a way we could rewrite that comment: ``` (** returns: the first element of [lst]. * raises: [Failure "hd"] if [lst = []]. *) ``` In `[lst = []]`, the outer square brackets indicate source code as part of a comment, whereas the inner square brackets indicate the empty list. [std-web]: http://caml.inria.fr/pub/docs/manual-ocaml/libref/List.html [std-src]: https://github.com/ocaml/ocaml/blob/trunk/stdlib/list.mli **List comprehensions:** Some languages, including Python and Haskell, have a syntax called *comprehension* that allows lists to be written somewhat like set comprehensions from mathematics. The earliest example of comprehensions seems to be the functional language NPL, which was designed in 1977. OCaml doesn't have built-in syntactic support for comprehensions, though that doesn't make the language any less expressive: everything you can do with comprehensions can be done with more primitive language features. Moreover, we won't be using comprehensions in this course, so it's safe for you to ignore the rest of this subsection. It is possible to get support for them with the `ppx_monadic` package. Install it with this command: ``` opam install -y ppx_monadic ``` Then in utop you can write comprehensions in a Haskell-like notation: ``` # #require "ppx_monadic";; # [%comp x+y || x <-- [1;2;3]; y <-- [1;2;3]; x<y];; - : int list = [3; 4; 5] ``` ## Summary Lists are a highly useful built-in data structure in OCaml. The language provides a lightweight syntax for building them, rather than requiring you to use a library. Accessing parts of a list makes use of pattern matching, a very powerful feature (as you might expect from its rather lengthy semantics). We'll see more uses for pattern matching as the course proceeds. These built-in lists are implemented as singly-linked lists. That's important to keep in mind when your needs go beyond small to medium sized lists. Recursive functions on long lists will take up a lot of stack space, so tail recursion becomes important. And if you're attempting to process really huge lists, you probably don't want linked lists at all, but instead a data structure that will do a better job of exploiting memory locality. ## Terms and concepts * append * binding * branch * cons * copying * desugaring * exhaustiveness * head * induction * list * nil * pattern matching * prepend * recursion * sharing * stack frame * syntactic sugar * tail * tail call * tail recursion * type constructor * wildcard ## Further reading * *Introduction to Objective Caml*, chapters 4, 5.3, 5.4 * *OCaml from the Very Beginning*, chapters 3, 4, 5 * *Real World OCaml*, chapter 3