Can you hear the drums, Erlando?

Most of us at RabbitMQ HQ have spend time working in a number of functional languages in addition to Erlang, such as Haskell, Scheme, Lisp, OCaml or others. Whilst there is lots to like about Erlang, such as its VM/Emulator, there are inevitably features that we all miss from other languages. In my case, having spent a couple of years working in Haskell before returning to the RabbitMQ fold, all sorts of features are "missing", such as laziness, type classes, additional infix operators, the ability to specify precedence of functions, fewer parenthesis, partial application, more consistent standard libraries and do-notation. That's a fair list, and it'll take me a while to get around to implementing them all in Erlang, but here are two for starters.

Introduction

Erlando is a set of syntax extensions for Erlang. Currently it consists of two syntax extensions, both of which take the form of parse transformers.

• Cut: This adds support for cuts to Erlang. These are inspired by the Scheme form of cuts. Cuts can be thought of as a light-weight form of abstraction, with similarities to partial application (or currying).
• Do: This adds support for do-syntax and monads to Erlang. These are heavily inspired by Haskell, and the monads and libraries are near-mechanical translations from the Haskell GHC libraries.

Use

To use any of these parse transformers, you must add the necessary -compile attributes to your Erlang source files. For example:

-module(test).
-compile({parse_transform, cut}).
-compile({parse_transform, do}).

Then, when compiling test.erl, you must ensure erlc can locate cut.beam and/or do.beam by passing the suitable path to erlc with a -pa or -pz argument. For example:

erlc -Wall +debug_info -I ./include -pa ebin -o ebin  src/cut.erl
erlc -Wall +debug_info -I ./include -pa ebin -o ebin  src/do.erl
erlc -Wall +debug_info -I ./include -pa test/ebin -pa ./ebin -o test/ebin test/src/test.erl

Note, if you're using QLC, you may find you need to be careful as to the order of the parse transforms: I've found that the -compile({parse_transform, cut}). must occur before the -include_lib("stdlib/include/qlc.hrl").

Cut

Motivation

Cut is motivated by the frequency with which simple abstractions (in a lambda-calculus sense) are used in Erlang, and the relatively noisy nature of declaring funs. For example, it's quite common to see code like:

with_resource(Resource, Fun) ->
    case lookup_resource(Resource) of
        {ok, R}          -> Fun(R);
        {error, _} = Err -> Err
    end.

my_fun(A, B, C) ->
    with_resource(A, fun (Resource) ->
                            my_resource_modification(Resource, B, C)
                        end).

I.e. a fun is very simply created in order to perform variable capture from the its surrounding scope but to leave holes for further arguments to be provided. Using a cut, the function my_fun can be rewritten as:

my_fun(A, B, C) ->
    with_resource(A, my_resource_modification(_, B, C)).

Definition

Normally, the variable _ can only occur in patterns: i.e. where match occurs. This can be in assignment, in cases, and in function heads. For example:

{_, bar} = {foo, bar}.

Cut uses _ in expressions to indicate where abstraction should occur. Abstraction from cuts is always performed on the shallowest enclosing expression. For example:

list_to_binary([1, 2, math:pow(2, _)]).

will create the expression

list_to_binary([1, 2, fun (X) -> math:pow(2, X) end]).

and not

fun (X) -> list_to_binary([1, 2, math:pow(2, X)]) end.

It is fine to use multiple cuts in the same expression, and the arguments to the created abstraction will match the order in which the _ var is found in the expression. For example:

assert_sum_3(X, Y, Z, Sum) when X + Y + Z == Sum -> ok;
assert_sum_3(_X, _Y, _Z, _Sum) -> {error, not_sum}.

test() ->
    Equals12 = assert_sum_3(_, _, _, 12),
    ok = Equals12(9, 2, 1).

It is perfectly legal to take cuts of cuts as the abstraction created by the cut is a normal fun expression and thus can be re-cut as necessary:

test() ->
    Equals12 = assert_sum_3(_, _, _, 12),
    Equals5 = Equals12(_, _, 7),
    ok = Equals5(2, 3).

Note that because a simple fun is being constructed by the cut, the arguments are evaluated prior to the cut function. For example:

f1(_, _) -> io:format("in f1~n").

test() ->
    F = f1(io:format("test line 1~n"), _),
    F(io:format("test line 2~n")).

will print out

test line 2
test line 1
in f1

This is because the cut creates fun (X) -> f1(io:format("test line 1~n"), X) end. Thus it is clear that X must be evaluated first, before the fun can be invoked.

Of course, no one would be crazy enough to have side-effects in function argument expressions, so this will never cause any issues!

Cuts are not limited to function calls. They can be used in any expression where they make sense:

Tuples

F = {_, 3},
{a, 3} = F(a).

Lists

dbl_cons(List) -> [_, _ | List].

test() ->
    F = dbl_cons([33]),
    [7, 8, 33] = F(7, 8).

Note that if you nest a list as a list tail in Erlang, it's still treated as one expression. For example:

A = [a, b | [c, d | [e]]]

is exactly the same (right from the Erlang parser onwards) as:

A = [a, b, c, d, e]

I.e. those sub-lists, when they're in the tail position do not form sub-expressions. Thus:

F = [1, _, _, [_], 5 | [6, [_] | [_]]],
%% This is the same as:
%%  [1, _, _, [_], 5, 6, [_], _]
[1, 2, 3, G, 5, 6, H, 8] = F(2, 3, 8),
[4] = G(4),
[7] = H(7).

However, be very clear about the difference between , and |: the tail of a list is only defined following a |. Following a ,, you're just defining another list element.

F = [_, [_]],
%% This is **not** the same as [_, _] or its synonym: [_ | [_]]
[a, G] = F(a),
[b] = G(b).

Records

-record(vector, { x, y, z }).

test() ->
    GetZ = _#vector.z,
    7 = GetZ(#vector { z = 7 }),
    SetX = _#vector{x = _},
    V = #vector{ x = 5, y = 4 } = SetX(#vector{ y = 4 }, 5).

Case

F = case _ of
        N when is_integer(N) -> N + N;
        N                    -> N
    end,
10 = F(5),
ok = F(ok).

See test_cut.erl for more examples, including use of cuts in list comprehensions and binary construction.

Note that cuts are not allowed where the result of the cut can only be useful by interacting with the evaluation scope. For example:

F = begin _, _, _ end.

This is not allowed, because the arguments to F would have to be evaluated before the invocation of its body, which would then have no effect, as they're already fully evaluated by that point.

Do

The Do parse transformer permits Haskell-style do-notation in Erlang, which makes using monads, and monad transformers possible and easy. Without do-notation, monads tend to look like a lot of line noise.

The Inevitable Monad Tutorial

The Mechanics of a Comma

What follows is a brief and mechanical introduction to monads. It differs from a lot of the Haskell monad tutorials, because they tend to view monads as a means of achieving sequencing of operations in Haskell, which is challenging because Haskell is a lazy language. Erlang is not a lazy language, but the powerful abstractions possible from using monads are still very worthwhile. Whilst this is a very mechanical tutorial, it should be possible to see the more advanced abstractions possible.

Let's say we have the three lines of code:

A = foo(),
B = bar(A, dog),
ok.

They are three, simple statements, which are evaluated consecutively. What a monad gives you is control over what happens between the statements: in Erlang, it is a programmatic comma.

If you wanted to implement a programmatic comma, how would you do it? You might start with something like:

A = foo(),
comma(),
B = bar(A, dog),
comma(),
ok.

But that's not quite powerful enough, because unless comma/0 throws some sort of exception, it can't actually stop the subsequent expression from being evaluated. Most of the time we'd probably like the comma/0 function to be able to act on some variables which are currently in scope, and that's not possible here either. So we should extend the function comma/0 so that it takes the result of the preceding expression, and can choose whether or not the subsequent expressions should be evaluated:

comma(foo(),
        fun (A) -> comma(bar(A, dog),
                        fun (B) -> ok end)).

Thus the function comma/2 takes all results from the previous expression, and controls how and whether they are passed to the next expression.

As defined, the comma/2 function is the monadic function >>=/2.

Now it's pretty difficult to read the program with the comma/2 function (especially as Erlang annoyingly doesn't allow us to define new infix functions), which is why some special syntax is necessary. Haskell has it's do-notation, and so we've borrowed from that and abused Erlang's list comprehensions. Haskell also has lovely type-classes, which we've sort of faked specifically for monads. So, with the Do parse transformer, you can write in Erlang:

do([Monad ||
    A <- foo(),
    B <- bar(A, dog),
    ok]).

which is readable and straightforward, but is transformed into:

Monad:'>>='(foo(),
            fun (A) -> Monad:'>>='(bar(A, dog),
                                    fun (B) -> ok end)).

There is no intention that this latter form is any more readable than the comma/2 form - it is not. However, it should be clear that the function Monad:'>>='/2 now has complete control over what happens: does the fun on the right hand side ever get invoked? If so, with what value?

Lots of different types of Monads

So now that we have some relatively nice syntax for using monads, what can we do with them? Also, in the code

do([Monad ||
    A <- foo(),
    B <- bar(A, dog),
    ok]).

what are the possible values of Monad?

The answer to the first question is almost anything; and to the later question, is any module name that implements the monad behaviour.

Above, we covered one of the three monadic operators, >>=/2. The others are:

• return/1: This lifts a value into the monad. We'll see examples of this shortly.

• fail/1: This takes a term describing the error encountered, and informs whichever monad currently in use that some sort of error has occurred.

Note that within do-notation, any function call to functions named return or fail, are automatically rewritten to invoke return or fail within the current monad.

Some people familiar with Haskell's monads may be expecting to see a fourth operator, >>/2. Interestingly, it turns out that you can't implement >>/2 in a strict language unless all your monad types are built on top a function. This is because in a strict language, arguments to functions are evaluated before the function is invoked. For >>=/2, the 2nd argument is only reduced to a function prior to invocation of >>=/2. But the 2nd argument to >>/2 is not a function, and so in strict languages, will be fully reduced prior to >>/2 being invoked. This is problematic because the >>/2 operator is meant to be in control of whether or not subsequent expressions are evaluated. The only solution here would be to make the basic monad type a function, which would then mean that the 2nd argument to >>=/2 would become a function to a function to a result! However, it is required that '>>'(A, B) behaves identically to '>>='(A, fun (_) -> B end), and so that is what we do: whenever we come to a do([Monad || A, B ]), we rewrite it to '>>='(A, fun (_) -> B end) rather than '>>'(A, B). The effect of this is that the >>/2 operator does not exist.

The simplest monad possible is the Identity-monad:

-module(identity_m).
-behaviour(monad).
-export(['>>='/2, return/1, fail/1]).

'>>='(X, Fun) -> Fun(X).
return(X)     -> X.
fail(X)       -> throw({error, X}).

This makes our programmatic comma behave just like Erlang's comma normally does. The bind operator (>>=/2) does not inspect the values passed to it, and always invokes the subsequent expression fun.

What could we do if we did inspect the values passed to the sequencing combinators? One possibility results in the Maybe-monad:

-module(maybe_m).
-behaviour(monad).
-export(['>>='/2, return/1, fail/1]).

'>>='({just, X}, Fun) -> Fun(X);
'>>='(nothing,  _Fun) -> nothing.

return(X) -> {just, X}.
fail(_X)  -> nothing.

Thus if the result of the preceding expression is nothing, then the subsequent expressions are not evaluated. This means that we can write very neat looking code which immediately stops should any failure be encountered.

if_safe_div_zero(X, Y, Fun) ->
    do([maybe_m ||
        Result <- case Y == 0 of
                        true  -> fail("Cannot divide by zero");
                        false -> return(X / Y)
                    end,
        return(Fun(Result))]).

If Y is equal to 0, then Fun will not be invoked, and the result of the if_safe_div_zero function call will be nothing. If Y is not equal to 0, then the result of the if_safe_div_zero function call will be {just, Fun(X / Y)}.

We see here that within the do-block, there is no mention of nothing or just: they are abstracted away by the Maybe-monad. As a result, it is possible to change the monad in use, without having to rewrite any further code.

One common place to use a monad like the Maybe-monad is where you'd otherwise have a lot of nested case statements in order to detect errors. For example:

write_file(Path, Data, Modes) ->
    Modes1 = [binary, write | (Modes -- [binary, write])],
    case make_binary(Data) of
        Bin when is_binary(Bin) ->
            case file:open(Path, Modes1) of
                {ok, Hdl} ->
                    case file:write(Hdl, Bin) of
                        ok ->
                            case file:sync(Hdl) of
                                ok ->
                                    file:close(Hdl);
                                {error, _} = E ->
                                    file:close(Hdl),
                                    E
                            end;
                        {error, _} = E ->
                            file:close(Hdl),
                            E
                    end;
                {error, _} = E -> E
            end;
        {error, _} = E -> E
    end.

make_binary(Bin) when is_binary(Bin) ->
    Bin;
make_binary(List) ->
    try
        iolist_to_binary(List)
    catch error:Reason ->
            {error, Reason}
    end.

can be transformed into the much shorter

write_file(Path, Data, Modes) ->
    Modes1 = [binary, write | (Modes -- [binary, write])],
    do([error_m ||
        Bin <- make_binary(Data),
        {ok, Hdl} <- file:open(Path, Modes1),
        {ok, Result} <- return(do([error_m ||
                                    ok <- file:write(Hdl, Bin),
                                    file:sync(Hdl)])),
        file:close(Hdl),
        Result]).

Note that we have a nested do-block so that, as with the non-monadic code, we ensure that once the file is opened, we always call file:close/1 even if an error occurs in a subsequent operation. This is achieved by wrapping the nested do-block with a return/1 call: even if the inner do-block errors, the error is lifted to a non-error value in the outer do-block, and thus execution continues to the subsequent file:close/1 call.

Here we are using an Error-monad which is remarkably similar to the Maybe-monad, but matches the typical Erlang practice of indicating errors by an {error, Reason} tuple:

-module(error_m).
-behaviour(monad).
-export(['>>='/2, return/1, fail/1]).

'>>='({error, _Err} = Error, _Fun) -> Error;
'>>='(Result,                 Fun) -> Fun(Result).

return(X) -> {ok,    X}.
fail(X)   -> {error, X}.

Monad Transformers

Monads can be nested by having do-blocks inside do-blocks, and parameterized by defining a monad as a transformation of another, inner, monad. The State Transform is a very commonly used monad transformer, and is especially relevant for Erlang. Because Erlang is a single-assignment language, it's very common to end up with a lot of code that incrementally numbers variables:

State1 = init(Dimensions),
State2 = plant_seeds(SeedCount, State1),
{DidFlood, State3} = pour_on_water(WaterVolume, State2),
State4 = apply_sunlight(Time, State3),
{DidFlood2, State5} = pour_on_water(WaterVolume, State4),
{Crop, State6} = harvest(State5),
    ...

This is doubly annoying, not only because it looks awful, but also because you have to re-number many variables and references whenever a line is added or removed. Wouldn't it be nice if we could abstract out the State? We could then have a monad encapsulate the state and provide it to (and collect it from) the functions we wish to run.

Our implementation of monad-transformers (like State) uses a "hidden feature" of the Erlang distribution called parameterized modules. These are described in Parameterized Modules in Erlang.

The State-transform can be applied to any monad. If we apply it to the Identity-monad then we get what we're looking for. The key extra functionality that the State transformer provides us with is the ability to get and set (or just plain modify) state from within the inner monad. If we use both the Do and Cut parse transformers, we can write:

StateT = state_t:new(identity_m),
SM = StateT:modify(_),
SMR = StateT:modify_and_return(_),
StateT:exec(
    do([StateT ||

        StateT:put(init(Dimensions)),
        SM(plant_seeds(SeedCount, _)),
        DidFlood <- SMR(pour_on_water(WaterVolume, _)),
        SM(apply_sunlight(Time, _)),
        DidFlood2 <- SMR(pour_on_water(WaterVolume, _)),
        Crop <- SMR(harvest(_)),
        ...

        ]), undefined).

We start by creating a State-transform over the Identity-monad.

This is the syntax for instantiating parameterized modules. StateT is a variable referencing a module instance which, in this case, is a monad.

We set up two shorthands for running functions that either just modify the state, or modify the state and return a result. Whilst there's a bit of bookkeeping to do, we achieve our goal: there are no state variables now to renumber whenever we make a change: we use cut to leave holes in the functions where State should be fed in, and we obey the protocol that if functions return both a result and state, it should be in the form of a {Result, State} tuple. The State-transform does the rest.

Beyond Monads

There are some standard monad functions such as join/2 and sequence/2 available in the monad module. We have also implemented monad_plus which works for monads where there's an obvious sense of zero and plus (currently Maybe-monad, List-monad, and Omega-monad). The associated functions guard, msum and mfilter are available in the monad_plus module.

In many cases, a fairly mechanical translation from Haskell to Erlang is possible, so in many cases converting other monads or combinators should be straightforward. However, the lack of type classes in Erlang is limiting.