Description

This repository implements a technique for embedding pattern matching into a Haskell EDSL, with the intention of integrating the ideas into the Haski compiler.

EDIT (2021-06-25): Precise implementation details in this repository may differ slightly from that of the Haski compiler. In general though, the technique used is the same.

Specifically, a framework is provided that allows the user to:

1. Define how to convert some input type (e.g. Int) into a user-defined ADT.

2. Use the above definition to convert into, and deconstruct using pattern matching, values of the user-defined ADT.

The technique is demonstrated in a simple expression language E along with a compiler targeting C.

Example programs can be found in the src/E/Examples directory.

Building

The library can be built with Stack:

stack build

To generate documentation:

stack haddock

To compile one of the example programs, first run GHCi:

stack repl

Then, apply the printProg function to an E expression:

# A dummy argument may need to be provided to the program; the compiled
# programs don't read arguments.
GHCi> printProg (needsWatering 0)

The expression language

The expression language E is a straightforward deep embedding, supporting simple arithmetics and some basic bitwise operations. The constructors are given by the GADT found in E.Core. Functions for compiling E programs can be found in E.Compile. The generated C does not handle input; its main() function simply runs with a hardcoded value. E is intended to represent a generic embedded language, so that techniques that work for E should also work for other embedded languages.

A worked example

This section will walk through the hypothetical example use case presented in E.Examples.Watering. The example is chosen with the target domain of Haski in mind (IoT-devices).

Input data

An example of a typical smart device is a sensor which can collect data from its environment, and send the data to some application which can process it. Imagine that we are growing plants, and that we want to hook up a sensor to collect some information about the air temperature and soil humidity, so that a connected application can determine whether the plants need watering.

For low-level devices such as this, it is common for the sent data to be encoded in a format that might be a bit inconvenient to work with. For this example, imagine the data is encoded within a 16-bit integer, and that it can take one of two forms:

If control bit C is set:

| C = 1 | 7           | 8             |
  ^       ^             ^------------ Temperature (degrees Celsius)
  |       +-------------------------- Humidity (percentage)
  +---------------------------------- Control bit

If control bit C is not set:

| C = 0 | 15                          |
  ^        ^------------------------- Error code
  +---------------------------------- Control bit

The data is encoded within the 16-bit integer, but has two intepretations (given by the specification of the sensor). Which one to use depends on whether the most significant bit, the control bit, is set:

• If the control bit is set, then the first 8 bits of the input represent the temperature in degrees Celsius as a signed 8-bit integer. The following 7 bits represent the humidity percentage of the soil as a 7-bit unsigned integer (a bit of a nonsensical representation to be fair, as it only spans the 1% - 99% range).

• If the control bit is not set, then the sensor is signals an error, and the remaining 15 bits represent some sort of error code.

Writing a program

Now, say we wanted to perform some logic based on the input sensor data. For example, we might want a function to return True if the soil humidity is sufficiently low, or if the temperature is sufficiently high.

The type signature for such a function could look something like this:

needsWater :: E Word16 -> E Bool

The problem is that we need to perform a bunch of bitwise masking and shifting to extract the data we want, and things could clutter up the function body quite a bit, which might obscure our application logic. Besides, the form of the sensor data will not change while the program runs; it is always going to be one of the two encodings.

Therefore, we might want to define a Haskell data type to model the two encodings of the sensor data.

type Temp      = E Word8   -- Unsigned 8-bit integer
type Humidity  = E Int8    -- Signed 8-bit integer
type ErrorCode = E Word16  -- Unsgined 16-bit integer

data SensorData
    = Sensor Temp Humidity
    | Error ErrorCode

toSensorData :: E Word16 -> SensorData
{- ... -}

Our data type SensorData has two constructors, corresponding to the two different encodings for our sensor data. The fields of the constructors correspond to the encoded data; the integer types used (Int8, Word8, Word16) are the smallest sizes of multiples of 8 that can fit the data. We could now define a toSensorData conversion function to contain most of the dirty bitwise operations, so that the application logic stays clean.

Now that we have our custom data type, which is more convenient to work with than the 16-bit binary encoding, we'd like to define needsWater similar to this:

needsWater :: E Word16 -> E Bool
needsWater x = case toSensorData x of
    Sensor temp humidity -> temp >. 30 ||. humidity <. 25
    Error _errorCode     -> valE False  -- Placeholder; do something sensible here

The operators >., ||., and <. are variants of the normal >, ||, and <, but operating on values in the expression language. valE :: a -> E a simply brings a value into the expression languge.

The problem with an implementation like this is that our pattern matching in the case-expression is not on the embedded level, like our temperature- and humidity-checking logic is. When we compile needsWater, the result of toSensorData x will be either a Sensor value or an Error value. What we want is to inspect the value of toSensorData x during the runtime of the generated target language program (C), but currently we instead inspect it during the runtime of our host language program (Haskell). What we need is a representation of the case-expression inside the expression language, as a constructor of the E type.

Partition and match

To work around this problem, we use the framework provided and modify our needsWater program to the following:

{-# LANGUAGE LambdaCase #-}

needsWatering :: E Word16 -> Estate (E Bool)
needsWatering x = match x $ \case
    Sensor temp humidity -> temp >. 30 ||. humidity <. 25
    Error _errorCode     -> valE False  -- Do something sensible here

We note that we call a function match in place of toSensorData and that we no longer use the built-in Haskell case of-construct, instead providing a function defined using a lambda-case (enabled by the LambdaCase languge pragma). The return type is also slightly different; Estate is simply a specialization of the State monad and is used to generate unique identifiers when compiling the E AST:

type Estate = State Env
data Env = {- ... -}

To understand how this works, we first take a look at the signature for the match function, here slightly simplified:

match :: Partition a p => E a -> (p -> E b) -> Estate (E b)

match takes value of type E a as its first argument, this is the scrutinee. The second argument is the function that can perform pattern matching, but it takes a value of type p, not E a. When match is applied, it will transform the E a value to a p value; this transformation is indicated by the Partition a p constraint and corresponds to the toSensorData function from the earlier example.

{-# LANGUAGE MultiParamTypeClasses #-}

class Partition a p where
    partition :: [E a -> (E Bool, p)]

A Partition a p instance defines how to transform values from some input type a to a sum type p. Since normal Haskell data-style types are sum types, this means that we can define instances of Partition to transform values of other types into instances of our user-defined types, such as SensorData. Because sum types can have multiple constructors, an instance of Partition a p must also define conditions to determine which constructor to use, depending on the value of the input (of type a).

Note that there are no actual constraints on the type variables a and p; they can be any types, in theory. However, seeing as the point of Partition is to define transformation and construction of data, it might be hard though to implement anything useful for uninhabited types (such as Void), for example.

partition :: [E a -> (E Bool, Estate p)] takes no arguments and returns a list of functions where each function represents one constructor of the sum type p, or one "branch" of a case of-construct. Each function takes a value of the input type a and returns a pair:

• The first half of the pair is a predicate on the input value and affects program control-flow in the target language; the E Bool expression corresponds to the condition in the generated if-statement.

• The second half of the pair constructs a value of the sum type p which can be dependent on the input value. This provides the functionality of toSensorData from earlier.

The following would be a possible definition of a Partition instance for our example use case:

{-# LANGUAGE BinaryLiterals        #-}
{-# LANGUAGE MultiParamTypeClasses #-}
{-# LANGUAGE NumericUnderscores    #-}

instance Partition Word16 SensorData where
    partition =
        [ \ v -> ( testBitE 15 v
                 , Sensor (castE (v &. tempMask))
                          (castE (v &. humidityMask >>. 8))
                 )
        , \ v -> ( notE (testBitE 15 v)
                 , Error (v &. errorMask)
                 )
        ]
      where
        -- Language extensions BinaryLiterals and NumericUnderscores enables
        -- nicer literals for binary values.
        tempMask, humidityMask, errorMask :: E Word16
        tempMask     = 0b0000_0000_1111_1111
        humidityMask = 0b0111_1111_0000_0000
        errorMask    = 0b0111_1111_1111_1111

In each of the corresponding E Bool fields in our pairs, we implement the check for the control bit in our input data, and apply constructors Sensor and Error in the second half of the pair accordingly. testBitE checks whether a given bit is set, and castE and >>. are straightforward C translations of type-casting and bitwise shift-right operations.

When constructing our needsWatering program, match will apply all the functions given by partition to a placeholder value ESym, thus creating values of SensorData using both our constructors:

-- Simplified snippet from the definition of 'match`
let branches = map ($ ESym) $ partition @a @p

Then, when generating the C code, the compiler can figure out how to replace the ESym values with that of the scrutinee.

Below is the corresponding C output (slightly edited for clarity) for the pattern match and match usage in needsWatering:

uint16_t _scrut1;

bool v0(uint16_t arg) {
    _scrut1 = arg;
    bool res;

    if (((1 << 15) & _scrut1) != 0) {
        res = ((int8_t) (_scrut1 & 255)) > 30 || ((uint8_t) ((_scrut1 & 32512) >> 8)) < 25;
    } else if (!(((1 << 15) & _scrut1) != 0)) {
        res = false;
    } else {
        fprintf(stderr, "No match on: `_scrut1`\n");
        exit(1);
    }
    return res;
}

We can see that the right-hand sides of the lambda-case function from needsWatering has been applied in some form on the scrutinee, in combination with whatever transformation was applied by partition.

On a quick note, one might wonder why the variable _scrut is necessary, seeing as it immediately gets assigned the function argument arg. Why not use arg directly? The reason is only an implementation convenience to avoid scoping errors that can otherwise occur when nestling calls to match. The current compiler creates single-parameter function defintions to handle each call to match. In the generated C code, this means that a variable from an outer match is invisible inside the function body generated from an inner match, even though this differs from the Haskell scoping:

f :: E Int -> Estate (E Bool)
f x = matchM x $ \case
    A2 b n -> b
    A1 n   -> match n $ \case
        A1 _   -> b  -- <- b refers to a field in the outer match.
        A2 t _ -> t

Lessen duplicate computation

Our needsWatering example looks fairly OK at this point, but there exists a potential issue related to our technique of pre-applying both partition and the lambda function on which match is applied. The issue manifests when fields of constructors are used multiple times in the body of the match call.

Consider a different (contrived) example, and the generated C code (again, edited for clarity):

contrived :: E Int -> Estate (E Bool)
contrived v = match v $ \case
    T1 n -> n + n >. n
    T2   -> valE False

data T = T1 (E Int) | T2

instance Partition Int T where
    partition =
        [ \ v -> (v >=. 0, T1 (v * 10 - 2 * 10 - 2))
        , \ v -> (v <.  0, T2)
        ]

int _scrut1;

bool v0(int arg) {
    _scrut1 = arg;
    bool res;

    if (_scrut1 >= 0) {
        // Duplicate computations!
        res = (((((_scrut1 * 10) - (2 * 10)) - 2)
               + (((_scrut1 * 10) - (2 * 10)) - 2))
               > (((_scrut1 * 10) - (2 * 10)) - 2));
    } else if (_scrut1 < 0) {
        res = false;
    } else {
        fprintf(stderr, "No match on: `_scrut1`\n");
        exit(1);
    }

    return res;
}

In the body of the match call, we only performed two operations, + and >.. Yet, because we reference the expression n multiple times, its entire definition (as defined by partition) shows up multiple times as well! This is a problem in several aspects. For one, it clutters the code and highlights the lack of a relation between the fields of our constructors and corresponding variables in the generated code. More importantly, however, it leads to extra, unnecessary computation in the generated code.

What we would like is for each field of our constructors to be assigned to a variable in the generated code, so that they can be reused. For this to happen, the compiler needs to know when an expression corresponds to a field of a constructor, for example by having expressions tagged with some kind of dedicated "tag" constructor. Implementing this is a bit tricky though; the match function itself does not know enough about the type of the value it operates on to meaningfully inspect each constructor field, though it might be possible through generic programming techniques such as GHC generics or similar.

Instead, the simplest way would be to apply a tag constructor EField in the defintion of partition, where the value is constructed:

EField :: ArgId -> E a -> E a
-- Simplified type of ArgId to demonstrate the idea; the actual type is
-- slightly different, but conceptually the same.
type ArgId = String

instance Partition Int T where
    partition =
        [ \ v -> (v >=. 0, T1 (EField "T1_1" (v * 10 - 2 * 10 - 2)))
        , \ v -> (v <.  0, T2)  -- ^ -- Tag constructor
        ]

EField wraps an expression with a tag, indicating to the compiler that it should assign the entire sub-expression to a variable, and use that variable in the future if the expression is referenced again. This works, but it has the downside of introducing additional boilerplate to the user, which is both unergonomic and potentially error-prone, especially so for larger types and for constructor with more fields. The EField tag only serves its purpose internally anyway, so it should be ideally be hidden away from the user.

Here is where we can use smart constructors to hide the tagging of field expressions. A smart constructor is a normal function that wraps the construction of a value, possibly performing additional work that can be hidden away from the user.

With a smart constructor, we could instead implement partition like this:

instance Partition Int T where
    partition =
        [ \ v -> (v >=. 0, _T1 (v * 10 - 2 * 10 - 2))
        , \ v -> (v <.  0, T2)
        ]

-- Smart constructor for constructor T1.
_T1 :: E Int -> T
_T1 a = let t1 = EField "T1_1"
        in T1 (t1 a)

Now the implementation of partition is minimally impacted, as we switch from writing T1 to _T1. Of course, we now instead have an entire separate smart constructor function that needs to be written instead, so we have only really moved the problem elsewhere.

Thankfully, these smart constructors follow a pattern, and the smart constructor name, type, number of parameters, and number tags to create, can all be derived from the type definition of value that it constructs. While the tagging would be difficult to do in match due to operating on a polymorphc type, we can generate our smart constructors using Template Haskell.

With this, we can add a single line to our data type declaration to generate smart constructors for it:

{-# LANGUAGE TemplateHaskell #-}

data T = T1 (E Int) | T2
$(mkConstructors ''T)

instance Partition Int T where
    partition =
        [ \ v -> (v >=. 0, _T1 (v * 10 - 2 * 10 - 2))
        , \ v -> (v <.  0, _T2)  -- Not necessary (0 fields) but looks uniform :)
        ]

As a small detail, the generated code from the TH function is slightly different from our hand-written one:

_T1 :: E Int -> Estate T
_T1 v0 = do
    tag1 <- newFieldTag
    pure $ T1 (tag1 a)

newFieldTag generates a unique tag name, which will be unique in the entire program. This results in the smart constructor having the return type Estate a instead of simply a, which also changes the definition of partition accordingly:

partition :: [E a -> (E Bool, Estate p)]
--                            ^^^^^^

Using the smart constructors, our generated code now looks like this instead:

int _scrut1;
int _scrut1_field1;

bool v0(int arg) {
    _scrut1 = arg;
    bool res;
    _scrut1_field1 = ((_scrut1 * 10) - (2 * 10)) - 2;

    if (_scrut1 >= 0) {
        res = (_scrut1_field1 * _scrut1_field1) > _scrut1_field1;
    } else if (_scrut1 < 0) {
        res = false;
    } else {
        fprintf(stderr, "No match on: `_scrut1`\n");
        exit(1);
    }
    return res;
}

The duplicate computations from earlier are gone and the relation to the original E program is a bit clearer as well!