enso/docs/style-guide/haskell.md
2020-07-21 13:59:40 +01:00

77 KiB
Raw Permalink Blame History

layout title category tags order
style-guide Haskell Style Guide style-guide
style-guide
4

Haskell Style Guide

Like many style guides, this Haskell style guide exists for two primary reasons. The first is to provide guidelines that result in a consistent code style across all of the Enso codebases, while the second is to guide people towards a style that is expressive while still easy to read and understand.

In general, it aims to create a set of 'zero-thought' rules in order to ease the programmer burden; there is usually only one way to lay out code correctly.

Code Formatting

This section explains the rules for visually laying out your code. They provide a robust set of guidelines for creating a consistent visual to the code.

Whitespace

The rules for whitespace in the Enso codebases are relatively simple:

  • 4 spaces are used for indentation, with no tabs.
  • There should not be any trailing whitespace.
  • There should be no spurious whitespace within the lines, unless it is used for alignment as discussed below.

Line Wrapping

In order to provide visual consistency across our codebases, and also to contribute to making our code easier to scan, we enforce that all code should be wrapped to 80 characters width at a maximum.

The nature of Haskell, however, means that it can sometimes be unclear where to break lines. We use the following guidelines:

  • Wrap all lines to a maximum length of 80 characters.

  • Break the lines on operators where possible, rather than wrapping function arguments.

    -- This
    foo <- veryLongFunction1 veryLongArgument1
        $ veryLongFunction2 veryLongArgument2 veryLongArgument3
    
    -- Not this
    foo <- veryLongFunction1 veryLongArgument1 $ veryLongFunction2
        veryLongArgument2 veryLongArgument3
    
  • When you have a choice of operators on which you could break, choose the one with the highest precedence. We find that this makes code significantly more readable.

    -- This
    potentialPkgRoot <- liftIO $ Directory.canonicalizePath
        =<< (canPath </>) <$> pkgRootFromExe @a
    
    -- Not this
    potentialPkgRoot <- liftIO $ Directory.canonicalizePath =<< (canPath </>)
        <$> pkgRootFromExe @a
    
  • Wrap operators to the start of the line, rather than leaving them trailing on a line.

    -- This
    foo <- veryLongFunction1 veryLongArgument1
        $ veryLongFunction2 veryLongArgument2 veryLongArgument3
    
    -- Not this
    foo <- veryLongFunction1 veryLongArgument1 $
        veryLongFunction2 veryLongArgument2 veryLongArgument3
    
  • Function signatures should wrap on the => and ->, and in the context of a doc comment should have each argument on a separate line.

  • Lists (and all list-like constructs e.g. constraint tuples, import lists) should be wrapped with a leading comma, aligned with the opening bracket, and a space between the opening bracket and the first item. This is also used in record declarations.

    -- This
    myFunctionWithAVeryLongName :: forall a m . ( SomeConstraintOnA a
                                                , SomeMonadConstraint m )
        => a -> SomeOtherType -> m a
    
    -- Not this
    myFunctionWithAVeryLongName :: forall a m . (SomeConstraintOnA a,
                                                SomeMonadConstraint m)
        => a -> SomeOtherType -> m a
    
  • If all else fails, wrap the lines using your best effort (usually what you find to be most readable). This may result in discussion during code review, but will provide a learning experience to augment this guide with more examples.

Please do not shorten sensible names in order to make things fit into a single line. We would much prefer that the code wraps to two lines and that naming remains intelligible than names become so shortened as to be useless.

Alignment

When there are multiple lines that are visually similar, we try to align the similar portions of the lines vertically.

people   <- getAllPeople <$> worlds
names    <- getName      <$> people
surnames <- getSurnames  <$> names

This should only be done when the lines don't need to be wrapped. If you have lines long enough that this visual justification would cause them to wrap, you should prefer to not wrap the lines and forego the visual alignment.

Furthermore, if you have to wrap a visually similar line such that it now spans multiple lines, it no longer counts as visually similar, and hence subsequent lines should not be aligned with it.

Naming

Enso has some fairly simple general naming conventions, though the sections below may provide more rules for use in specific cases.

  • Types are written using UpperCamelCase.
  • Variables and function names are written using camelCase.
  • If a name contains an initialism or acronym, all parts of that initialism should be of the same case: httpRequest or makeHTTPRequest.
  • Short variable names such as a and b should only be used in contexts where there is no other appropriate name (e.g. flip (a, b) = (b, a)). They should never be used to refer to temporary data in a where or let expression.
  • Any function that performs an unsafe operation that is not documented in its type (e.g. head : [a] -> a, which fails if the list is empty), must be named using the word 'unsafe' (e.g. unsafeHead). For more information on unsafe function usage, see the section on safety. The one exception to this rule is for functions which fail intentionally on a broken implementation (e.g. "should not happen"-style fatal crashes).
  • Naming should use American English spelling.

Imports

Organising imports properly means that it's easy to find the provenance of a given function even in the absence of IDE-style tooling. We organise our imports in four sections, each of which may be omitted if empty.

  1. Re-Exports: These are the modules that are to be re-exported from the current module. We import these qualified under a name X (for export), and then re-export these in the module header (see below for an example).
  2. Preludes: As we recommend the use of -XNoImplicitPrelude, we then explicitly import the prelude in use. This is almost always going to be Prologue as described in the section on libraries below.
  3. Qualified Imports: A list of all modules imported qualified. The as portion of the import expressions should be vertically aligned.
  4. Unqualified Imports: These must always have an explicit import list. There are no circumstances under which we allow a truly unqualified import. The import lists should be vertically aligned.

Imports within each section should be listed in alphabetical order, and should be vertically aligned.

When we have a module that exports a type the same as its name, we import the module qualified as its name, but we also import the primary type from the module unqualified. This can be seen with Map in the examples below.

This example is for a module that re-exports some names:

module Enso.MyModule (module Enso.MyModule, module X) where

import Enso.MyModule.Class as X (foo, bar)

import Prologue

import qualified Control.Monad.State as State
import qualified Data.Map            as Map

import Data.Map  (Map)
import Rectangle (Rectangle)
import Vector    (Vector (Vector), test)

However, in the context where your module doesn't re-export anything, you can use the simplified form:

module Enso.MyModule where

import Prologue

import qualified Control.Monad.State as State
import qualified Data.Map            as Map

import Data.Map  (Map)
import Rectangle (Rectangle)
import Vector    (Vector (Vector), test)

Exports

There is nothing more frustrating than having a need to use a function in a module that hasn't been exported. To that end, we do not allow for restricted export lists in our modules.

Instead, if you want to indicate that something is for internal use, you need to define it in an internal module. For a module named Enso.MyModule, we can define internal functions and data-types in Enso.MyModule.Internal. This means that these functions can be imported by clients of the API if they need to, but that we provide no guarantees about API stability when using those functions.

Section Headers

In order to visually break up the code for easier 'visual grepping', we organise it using section headers. These allow us to easily find the section that we are looking for, even in a large file.

For each type defined in a file, it can be broken into sections as follows:

--------------------
-- === MyType === --
--------------------

-- === Definition === --
{- The definition of the type goes here -}


-- === API === --
{- The API of the type goes here -}


-- === Instances === --
{- Any instances for the type go here -}

The section header must be preceded by three blank lines, while the subsection headers (except the first) should be preceded by two blank lines. Any of these subsections may be omitted if they don't exist, and a file may contain multiple of these sections as relevant.

Auto-Formatting

While we have attempted to use haskell auto-formatters to enforce many of the above stylistic choices in this document, none have been found to be flexible enough for our needs. However, as tools evolve or new ones emerge, we are open to revisiting this decision; if you know of a tool that would let us automate the above stylistic rules, then please speak up.

Commenting

Comments in code are a tricky area to get right as we have found that comments often expire quickly, and in absence of a way to validate them, remain incorrect for long periods of time. In order to best deal with this problem, we make the keeping of comments up-to-date into an integral part of our programming practice while also limiting the types and kinds of comments we allow.

Comments across the Enso codebases fall into three main types:

  • Documentation Comments: API documentation for all language constructs that can have it (functions, typeclasses, and so on).
  • Source Notes: Detailed explorations of design reasoning that avoid cluttering the code itself.
  • Tasks: Things that need doing or fixing in the codebase.

When we write comments, we try to follow one general guideline. A comment should explain what and why, without mentioning how. The how should be self-explanatory from reading the code, and if you find that it is not, that is a sign that the code in question needs refactoring.

Code should be written in such a way that it guides you over what it does, and comments should not be used as a crutch for badly-designed code.

Documentation Comments

One of the primary forms of comment that we allow across the Enso codebases is the doc comment. Every language construct that can have an associated doc comment should do so. These are intended to be consumed by users of the API, and use the standard Haddock syntax. Doc comments should:

  • Provide a short one-line explanation of the object being documented.
  • Provide a longer description of the object, including examples where relevant.
  • Explain the arguments to a function where relevant.

They should not reference internal implementation details, or be used to explain choices made in the function's implementation. See Source Notes below for how to indicate that kind of information.

Source Notes

Source Notes is a mechanism for moving detailed design information about a piece of code out of the code itself. In doing so, it retains the key information about the design while not impeding the flow of the code.

Source notes are detailed comments that, like all comments, explain both the what and the why of the code being described. In very rare cases, it may include some how, but only to refer to why a particular method was chosen to achieve the goals in question.

A source note comment is broken into two parts:

  1. Referrer: This is a small comment left at the point where the explanation is relevant. It takes the following form: -- Note [Note Name], where Note Name is a unique identifier across the codebase. These names should be descriptive, and make sure you search for it before using it, in case it is already in use.
  2. Source Note: This is the comment itself, which is a large block comment placed after the first function in which it is referred to in the module. It uses the haskell block-comment syntax {- ... -}, and the first line names the note using the same referrer as above: {- Note [Note Name]. The name(s) in the note are underlined using a string of the ~ (tilde) character.

A source note may contain sections within it where necessary. These are titled using the following syntax: == Note [Note Name (Section Name)], and can be referred to from a referrer much as the main source note can be.

Sometimes it is necessary to reference a source note in another module, but this should never be done in-line. Instead, a piece of code should reference a source note in the same module that references the other note while providing additional context.

An example, taken from the GHC codebase, can be seen below.

prepareRhs :: SimplEnv -> OutExpr -> SimplM (SimplEnv, OutExpr)
-- Adds new floats to the env iff that allows us to return a good RHS
prepareRhs env (Cast rhs co)    -- Note [Float Coercions]
  | (ty1, _ty2) <- coercionKind co      -- Do *not* do this if rhs is unlifted
  , not (isUnLiftedType ty1)            -- seUnsae Note [Float Coercions (Unlifted)]
  = do  { (env', rhs') <- makeTrivial env rhs
        ; return (env', Cast rhs' co) }

        ...more equations for prepareRhs....

{- Note [Float Coercions]
~~~~~~~~~~~~~~~~~~~~~~~~~
When we find the binding
        x = e `cast` co
we'd like to transform it to
        x' = e
        x = x `cast` co         -- A trivial binding
There's a chance that e will be a constructor application or function, or
something like that, so moving the coercion to the usage site may well cancel
the coercions and lead to further optimisation.
        ...more stuff about coercion floating...

== Note [Float Coercions (Unlifted)]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
        ...explanations of floating for unlifted types...
-}

A source note like this is useful whenever you have design decisions to explain, but can also be used for:

  • Formulae and Algorithms: If your code makes use of a mathematical formula, or algorithm, it should note where the design element came from, preferably with a link.
  • Safety: Sometimes it is necessary to use an unsafe API in a context where it is trivially made safe. You should always use a source note to explain why its usage is safe in this context.

TODO Comments

We follow a simple convention for TODO comments in our codebases:

  • The line starts with TODO or FIXME.
  • It is then followed by the author's initials [ARA], or for multiple people [ARA, WD], in square brackets.
  • It is then followed by an explanation of what needs to be done.

For example:

-- TODO [ARA] This is a bit of a kludge. Instead of X it should to Y, accounting
-- for the fact that Z.

Other Comment Usage

There are, of course, a few other situations where commenting is very useful:

  • Commenting Out: You may comment out code while developing it, but if you commit any commented out code, it should be accompanied by an explanation of why said code can't just be deleted.
  • Bugs: You can use comments to indicate bugs in our code, as well as third-party bugs. In both cases, the comment should link to the issue tracker where the bug has been reported.

Program Design

Any good style guide goes beyond purely stylistic rules, and also talks about design styles to use in code.

Libraries

The Enso project has many internal libraries that are useful, but we have found that maintaining these on Hackage while they are under such active development is counterproductive.

Instead, libraries live in the lib/ folder of the primary project with which they are associated (Enso, Enso Studio, or Dataframes). These libraries may be freely used by others of our projects by depending on a git commit of the project that they live in. All of these are safe to use.

Prologue

Prologue is our replacement for Haskell's Prelude. For the most part it is compatible with the prelude, though it is designed with a safe API as the first port of call.

As a rule of thumb, if the prelude exports a partial function, that function has been made total in Prologue. This usually takes the form of returning Maybe, rather than throwing an error (e.g. head :: [a] -> Maybe a). In the case where a function has been redefined like this, the original version is available using an unsafe name (e.g. unsafeHead in the case above).

Prologue also exports additional useful functionality from across the Haskell ecosystem, such as utilities for working with Lenses and for writing type-level computation.

It is highly recommended that you scan the code of Prologue.

Safety

It is incredibly important that we can trust the code that we use, and hence we tend to disallow the definition of unsafe functions in our public API. When defining an unsafe function, you must account for the following:

  • It must be named unsafeX, as discussed in the section on naming.
  • Unsafe functions should only be used in the minimal scope in which it can be shown correct, not in larger pieces of code.
  • Unsafe function definition must be accompanied by a source note explaining why it is not defined safely (e.g. performance).
  • Unsafe function usage must be accompanied by a source note explaining why this usage of it is safe.

Furthermore, we do not allow for code containing pattern matches that can fail.

Control.Monad.Exception

We have our own exception framework based on ExceptT that encodes exception usage at the type level. This ensures that all synchronous exceptions must be dealt with.

It is defined in lib/exception/ and contains utilities for declaring that a function throws an exception, as well as throwing and catching exceptions.

The primary part of this API is the Throws constraint, which can be passed both a single exception type or a list of exceptions. It is a monadic exception framework.

myFunction :: Throws '[MyErrorOne, MyErrorTwo] m => ArgType -> m ReturnType

We encourage our programmers to define their own exception types, and when doing so they should use the following guidelines:

  • We name them using 'Error' rather than 'Exception', so MyError, rather than MyException.
  • We always provide an instance of Exception for our exception type.
  • We avoid encoding error information as strings, instead passing a strongly typed representation of the problem around. This often means that we end up re-wrapping an error thrown inside our function.

Modules

Unlike much of the Haskell ecosystem, we tend to design modules to be imported qualified rather than unqualified. This means that we have a few rules to keep in mind:

  • When designing a module that exports a type, the module should be named after that type. If it exports multiple types, there should be a primary type, or the other types should be factored out into their own modules.
  • We import modules as their name. If you have a module Enso.Space.MyType, we import it qualified as MyType.
  • Functions should be named with the assumption of being used qualified. This means that we rarely refer to the module name in the function name (e.g. State.run rather than State.runState).

Data Declarations

When declaring data types in the Enso codebases, please make sure to keep the following rules of thumb in mind:

  • For single-constructor types:

    • Write the definition across multiple lines.
    • Always name your fields.
    • Always generate lenses.
    data Rectangle = MkRectangle
        { _width  :: Double
        , _height :: Double
        } deriving (Eq, Ord, Show)
    makeLenses ''Rectangle
    
  • For multiple-constructor data-types:

    • Write the definition across multiple lines.
    • Never name your fields.
    • Generate prisms only when necessary.
    data Shape
        = ShapeCircle Circle
        | ShapeRect   Rectangle
        deriving (Eq, Ord, Show)
    
  • Always prefer named fields over unnamed ones. You should only use unnamed fields if one or more of the following hold:

    • Your data type is one where you are are sure that separate field access will never be needed.
    • You are defining a multiple-constructor data type.
  • Sort deriving clauses in alphabetical order, and derive the following for your type if logically correct:

    • General Types: Eq, Generic, NFData, Ord, Show.
    • Parametric 1-Types: Applicative, Alternative, Functor.
    • Monads: Monad, MonadFix.
    • Monad Transformers: MonadTrans.

Lenses

The Enso codebases make significant use of Lenses, and so we have some rules for their use:

  • Always use the makeLenses wrapper exported from Prologue.
  • Always generate lenses for single-constructor data types.
  • Never generate lenses for multi-constructor data types (though you may sometimes want to generate prisms).
  • Fields in data types should be named with a single underscore.
  • If you have multiple types where the fields need the same name, the Prologue lens wrappers will disambiguate the names for you as follows as long as you use a double underscore in the data declaration (e.g. __x).
data Vector = Vector
    { __x :: Double
    , __y :: Double
    , __z :: Double
    } deriving (Show)
makeLenses ''Vector

data Point = Point
    { __x :: Double
    , __y :: Double
    } deriving (Show)
makeLenses ''Point

This will generate lenses with names like vector_x, vector_y, and point_x, point_y.

Testing and Benchmarking

New code should always be accompanied by tests. These can be unit, integration, or some combination of the two, and they should always aim to test the new code in a rigorous fashion.

  • We tend to use HSpec, but also make use of QuickCheck for property-based testing.
  • Tests should be declared in the project configuration so they can be trivially run, and should use the mechanisms HSpec provides for automatic test discovery.
  • A test file should be named after the module it tests. If the module is named Enso.MyModule, then the test file should be named Enso.MyModuleSpec.

Any performance-critical code should also be accompanied by a set of benchmarks. These are intended to allow us to catch performance regressions as the code evolves, but also ensure that we have some idea of the code's performance in general.

  • We use Criterion for our benchmarks.
  • We measure time, but also memory usage and CPU time where possible.
  • Where relevant, benchmarks may set thresholds which, when surpassed, cause the benchmark to fail. These thresholds should be set for a release build, and not for a development build.

Do not benchmark a development build as the data you get will often be entirely useless.

Warnings, and Lints

In general, we aim for a codebase that is free of warnings and lints, and we do this using the following ideas:

Warnings

New code should introduce no new warnings onto main. You may build with warnings on your own branch, but the code that is submitted as part of a PR should not introduce new warnings. You should also endeavour to fix any warnings that you come across during development.

Sometimes it is impossible to fix a warning (e.g. TemplateHaskell generated code often warns about unused pattern matches). In such cases, you are allowed to suppress the warning at the module level using an OPTIONS_GHC pragma, but this must be accompanied by a source note explaining why the warning cannot be fixed otherwise.

Lints

We also recommend using HLint on your code as a stylistic guide, as we find that its suggestions in general lead to more readable code. If you don't know how to set up automatic linting for your editor, somebody will be able to help.

An example of an anti-pattern that HLint will catch is the repeated-$. Instead of foo $ bar $ baz $ bam quux, you should write foo . bar. baz $ bam quux to use function composition.

Language Extensions

Much like any sophisticated Haskell codebase, Enso makes heavy use of the GHC language extensions. We have a broad swath of extensions that are enabled by default across our projects, and a further set which are allowed whenever necessary. We also have a set of extensions that are allowed with care, which must be used sparingly.

When enabling a non-default extension, we never do it at the project or package level. Instead, they are enabled on a file-by-file basis using a LANGUAGE pragma. You may also negate default extensions, if necessary, using this same technique.

It should be noted that not all of the extensions listed below are available across all compiler versions. If you are unsure whether an extension is available to you, we recommend checking the GHC Users Guide entry for that extension (linked from the extension's table below).

Default Extensions

The following language extensions are considered to be so safe, or to have such high utility, that they are considered to be Enso's set of default extensions. You can find said set of extensions for Enso itself defined in a common configuration file.

AllowAmbiguousTypes

Name AllowAmbiguousTypes
Flag -XAllowAmbiguousTypes

This extension is particularly useful in the context of -XTypeApplications where the use of type applications can disambiguate the call to an ambiguous function.

We often use the design pattern where a function has a free type variable not used by any of its arguments, which is then applied via type applications. This would not be possible without -XAllowAmbiguousTypes.

ApplicativeDo

Name ApplicativeDo
Flag -XApplicativeDo

This extension allows desugaring of do-notation based on applicative operations (<$>, <*>, and join) as far as is possible. This will preserve the original semantics as long as the type has an appropriate applicative instance.

Applicative operations are often easier to optimise than monadic ones, so if you can write a computation using applicatives please do. This is the same reason that we prefer pure to return.

BangPatterns

Name BangPatterns
Flag -XBangPatterns

This extension allows for strict pattern matching, where the type being matched against is evaluated to WHNF before the match takes place. This is very useful in performance critical code where you want more control over strictness and laziness.

BinaryLiterals

Name BinaryLiterals
Flag -XBinaryLiterals

This extensions allow for binary literals to be written using the 0b prefix. This can be very useful when writing bit-masks, and other low-level code.

ConstraintKinds

Name ConstraintKinds
Flag -XConstraintKinds

This allows any types which have kind Constraint to be used in contexts (in functions, type-classes, etc). This works for class constraints, implicit parameters, and type quality constraints. It also enables type constraint synonyms.

All of these are very useful.

DataKinds

Name DataKinds
Flag -XDataKinds

This extension enables the promotion of data types to be kinds. All data types are promoted to kinds and the value constructors are promoted to type constructors.

This is incredibly useful, and used heavily in the type-level programming that makes the Enso codebase so expressive and yet so safe.

DefaultSignatures

Name DefaultSignatures
Flag -XDefaultSignatures

When you declare a default in a typeclass, it conventionally has to have exactly the same type signature as the typeclass method. This extension lifts this restriction to allow you to specify a more-specific signature for the default implementation of a typeclass method.

DeriveDataTypeable

Name DeriveDataTypeable
Flag -XDeriveDataTypeable

This extension enables deriving of the special kind-polymorphic Typeable typeclass. Instances of this class cannot be written by hand, and they associate type representations with types. This is often useful for low-level programming.

DeriveFoldable

Name DeriveFoldable
Flag -XDeriveFoldable

This enables deriving of the Foldable typeclass, which represents structures that can be folded over. This allows automated deriving for any data type with kind Type -> Type.

DeriveFunctor

Name DeriveFunctor
Flag -XDeriveFunctor

This enables automated deriving of the Functor typeclass for any data type with kind Type -> Type.

DeriveGeneric

Name DeriveGeneric
Flag -XDeriveGeneric

Enables automated deriving of the Generic typeclass. Generic is a typeclass that represents the structure of data types in a generic fashion, allowing for generic programming.

DeriveTraversable

Name DeriveTraversable
Flag -XDeriveTraversable

Enables automated deriving of the Traversable typeclass that represents types that can be traversed. It is a valid derivation for any data type with kind Type -> Type.

DerivingStrategies

Name DerivingStrategies
Flag -XDerivingStrategies

Under certain circumstances it can be ambiguous as to which method to use to derive an instance of a class for a data type. This extension allows users to manually supply the strategy by which an instance is derived.

If it is not specified, it uses the defaulting rules as described at the above link.

DerivingVia

Name DerivingVia
Flag -XDerivingVia

This allows deriving a class instance for a type by specifying another type of equal runtime representation (such that there exists a Coercible instance between the two). It is indicated by use of the via deriving strategy, and requires the specification of another type (the via-type) to coerce through.

DuplicateRecordFields

Name DuplicateRecordFields
Flag -XDuplicateRecordFields

This extension allows definitions of records with identically named fields. This is very useful in the context of Prologue's makeLenses wrapper as discussed above in the section on lenses.

EmptyDataDecls

Name EmptyDataDecls
Flag -XEmptyDataDecls

Allows the definition of data types with no value constructors. This is very useful in conjunction with -XDataKinds to allow the creation of more safety properties in types through the use of rich kinds.

FlexibleContexts

Name FlexibleContexts
Flag -XFlexibleContexts

This enables the use of complex constraints in class declarations. This means that anything with kind Constraint is usable in a class declaration's context.

FlexibleInstances

Name FlexibleInstances
Flag -XFlexibleInstances

This allows the definition of typeclasses with arbitrarily-nested types in the instance head. This, like many of these extensions, is enabled by default to support rich type-level programming.

Functional Dependencies

Name FunctionalDependencies
Flag -XFunctionalDependencies

Despite this extension being on the 'defaults' list, this is only for the very rare 1% of cases where Functional Dependencies allow you to express a construct that Type Families do not.

You should never need to use a Functional Dependency, and if you think you do it is likely that your code can be expressed in a far more clean manner by using Type Families.

GeneralizedNewtypeDeriving

Name GeneralizedNewtypeDeriving
Flag -XGeneralizedNewtypeDeriving

This enables the generalised deriving mechanism for newtype definitions. This means that a newtype can inherit some instances from its representation. This has been somewhat superseded by -XDerivingVia

InstanceSigs

Name InstanceSigs
Flag -XInstanceSigs

This extension allows you to write type signatures in the instance definitions for type classes. This signature must be identical to, or more polymorphic than, the signature provided in the class definition.

LambdaCase

Name LambdaCase
Flag -XLambdaCase

Enables \case as an alternative to case <...> of. This often results in much cleaner code.

LiberalTypeSynonyms

Name LiberalTypeSynonyms
Flag -XLiberalTypeSynonyms

This extension moves the type synonym validity check to after the synonym is expanded, rather than before. This makes said synonyms more useful in the context of type-level programming constructs.

MonadComprehensions

Name GeneralizedNewtypeDeriving
Flag -XGeneralizedNewtypeDeriving

Enables a generalisation of the list comprehension notation that works across any type that is an instance of Monad.

MultiParamTypeClasses

Name MultiParamTypeClasses
Flag -XMultiParamTypeClasses

Enables the ability to write type classes with multiple parameters. This is very useful for type-level programming, and to express relationships between types in typeclasses.

MultiWayIf

Name MultiWayIf
Flag -XMultiWayIf

This extension allows GHC to accept conditional expressions with multiple branches, using the guard-style notation familiar from function definitions.

NamedWildCards

NegativeLiterals

NoImplicitPrelude

Name NoImplicitPrelude
Flag -XNoImplicitPrelude

Disables the implicit import of the prelude into every module. This enables us to use Prologue, our own custom prelude (discussed in the section on prologue).

NumDecimals

Name NumDecimals
Flag -XNumDecimals

Enables writing integer literals using exponential syntax.

OverloadedLabels

Name OverloadedLabels
Flag -XOverloadedLabels

Enables support for Overloaded Labels, a type of identifier whose type depends both on its literal text and its kind. This is similar to -XOverloadedStrings.

OverloadedStrings

Name OverloadedStrings
Flag -XOverloadedStrings

Enables overloading of the native String type. This means that string literals are given their type based on contextual information as, and a string literal can be used to represent any type that is an instance of IsString.

PatternSynonyms

Name PatternSynonyms
Flag -XPatternSynonyms

Pattern synonyms enable giving names to parametrized pattern schemes. They can also be thought of as abstract constructors that dont have a bearing on data representation. They can be unidirectional or bidirectional, and are incredibly useful for defining clean APIs to not-so-clean data.

QuasiQuotes

Name QuasiQuotes
Flag -XQuasiQuotes

Quasi-quotation allows patterns and expressions to be written using programmer-defined concrete syntax. This extension enables the use of quotations in Haskell source files.

RankNTypes

Name RankNTypes
Flag -XRankNTypes

Enables the ability to express arbitrary-rank polymorphic types (those with a forall which is not on the far left of the type). These are incredibly useful for defining clean and safe APIs.

RecursiveDo

Name RecursiveDo
Flag -XRecursiveDo

This extension enables recursive binding in do-notation for any monad which is an instance of MonadFix. Bindings introduced in this context are recursively defined, much as for an ordinary let-expression.

ScopedTypeVariables

Name ScopedTypeVariables
Flag -XScopedTypeVariables

This enables lexical scoping of type variables introduced using an explicit forall in the type signature of a function. With this extension enabled, the scope of this variables is extended to the function body.

StandaloneDeriving

Name StandaloneDeriving
Flag -XStandaloneDeriving

Allows the creation of deriving declarations that are not directly associated with the class that is being derived. This is useful in the context where you need to create orphan instances, or to derive some non-default classes.

Strict

Name Strict
Flag -XStrict

We have found that making our code strict-by-default allows us to reason much more easily about its performance. When we want lazy evaluation, we use a combination of the negation flags and lazy pattern matching to achieve our goals.

When disabling strict for a module using -XNoStrict, you also need to add -XNoStrictData.

StrictData

Name StrictData
Flag -XStrictData

Much like the above, this helps with reasoning about performance, but needs to be explicitly disabled in contexts where the strictness is undesirable.

TemplateHaskell

Name TemplateHaskell
Flag -XTemplateHaskell

Enables the usage of Template Haskell, including the syntax for splices and quotes. TH is a meta-language that allows for generating Haskell code from arbitrary input.

TupleSections

Name TupleSections
Flag -XTupleSections

Much like we can do operator sections to partially apply operators, this extension enables partial application of tuple constructors.

TypeApplications

Name TypeApplications
Flag -XTypeApplications

This extension allows you to use visible type application in expressions. This allows for easily providing types that are ambiguous (or otherwise) to GHC in a way that doesn't require writing complete type signatures. We make heavy use of type applications in our type-level programming and API.

These should always be used as an alternative to Proxy, as they are just as useful for passing type information around without provision of data, and lead to nice and clean APIs.

TypeFamilies

Name TypeFamilies
Flag -XTypeFamilies

Type Families can be thought of as type-level functions, or functions on types. They are slightly more verbose than functional dependencies, but provide much better reusability, clearer contexts, and are far easier to compose. They should always be used in preference to functional dependencies.

When using Type Families, please keep the following things in mind:

  • Prefer open type families to closed type families.

  • Use closed type families if you want a fall-back when checking types.

    type family SumOf where
        SumOf Vector a      = Vector
        SumOf a      Vector = Vector
        SumOf a      a      = a
    
  • Do not use closed type families unless you are absolutely sure that your type family should not be able to be extended in the future.

TypeFamilyDependencies

Name TypeFamilyDependencies
Flag -XTypeFamilyDependencies

This extension allows type families to be annotated with injectivity information using syntax similar to that used for functional dependencies. This information is used by GHC during type-checking to resolve the types of expressions that would otherwise be ambiguous.

TypeOperators

Name TypeOperators
Flag -XTypeOperators

Type operators is a simple extension that allows for the definition of types with operators as their names. Much like you can define term-level operators, this lets you define type-level operators. This is a big boon for the expressiveness of type-level APIs.

UnicodeSyntax

Name UnicodeSyntax
Flag -XUnicodeSyntax

Enables unicode syntax for certain parts of the Haskell language.

ViewPatterns

Name ViewPatterns
Flag -XViewPatterns

View patterns provide a mechanism for pattern matching against abstract types by letting the programmer execute arbitrary logic as part of a pattern match. This is very useful for the creation of clean APIs.

Allowed Extensions

These extensions can be used in your code without reservation, but are not enabled by default because they may interact negatively with other parts of the codebase.

BlockArguments

Name BlockArguments
Flag -XBlockArguments

Block arguments allow expressions such as do, \, if, case, and let, to be used as both arguments to operators and to functions. This can often make code more readable than it otherwise would be.

GADTs

Name GADTs
Flag -XGADTs

This enables Generalised Algebraic Data Types, which expand upon normal data definitions to allow both contexts for constructors and richer return types. When this extension is enabled, there is a new style of data declaration available for declaring GADTs.

HexFloatLiterals

Name HexFloatLiterals
Flag -XHexFloatLiterals

Enables the ability to specify floating point literals using hexadecimal to ensure that no rounding or truncation takes place.

MagicHash

Name MagicHash
Flag -XMagicHash

Allows the use of # as a postfix modifier to identifiers. This allows the programmer to refer to the names of many of GHC's internal unboxed types for use in surface Haskell.

NumericUnderscores

Name NumericUnderscores
Flag -XNumericUnderscores

This extension allows breaking up of long numeric literals using underscores (e.g 1_000_000_000), which can often aid readability.

PolyKinds

Name PolyKinds
Flag -XPolyKinds

This extension enables users to access the full power of GHC's kind system, and allows for programming with kinds as well as types and values. It expands the scope of kind inference to bring additional power, but is sometimes unable to infer types at the value level as a result.

You should only enable -XPolyKinds in contexts where you need the power that it brings.

Quantified Constraints

Name QuantifiedConstraints
Flag -XQuantifiedConstraints

Quantified constraints bring additional expressiveness to the constraint language used in contexts in GHC Haskell. In essence, it allows for contexts to contain nested contexts, and hence for users to express more complex constraints than they would otherwise be able to.

RoleAnnotations

Name RoleAnnotations
Flag -XRoleAnnotations

Role annotations allow programmers to constrain the type inference process by specifying the roles of the class and type parameters that they declare.

UnboxedSums

Name UnboxedSums
Flag -XUnboxedSums

Enables the syntax for writing anonymous, unboxed sum types. These can be very useful for writing performance critical code, as they can be used as a standard anonymous sum type, including in pattern matching and at the type level.

UnboxedTuples

Name UnboxedTuples
Flag -XUnboxedTuples

This extension enables the syntax and use of unboxed tuples. This can be thought of as a dual to the above -XunboxedSums as it allows for the declaration and manipulation of unboxed product types.

Allowed With Care

If you make use of any of these extensions in your code, you should accompany their usage by a source note that explains why they are used.

CApiFFI

Name CApiFFI
Flag -XCApiFFI

Enables the capi calling convention for foreign function declarations. This should only be used when you need to call a foreign function using the C-level API, rather than across the platform's ABI. This enables the programmer to make use of preprocessor macros and the like, as the call is resolved as if it was against the C language.

CPP

Name CPP
Flag -XCPP

Enables the C preprocessor. We strongly discourage use of the preprocessor, but it is sometimes unavoidable when you need to do purely string-based preprocessing of a Haskell source file. It should only be used if you have no other solution to your problem.

PostfixOperators

Name PostfixOperators
Flag -XPostfixOperators

Enables the definition of left-sections for postfix operators. Please take care if you enable this that it does not lead to unclear code. You should not export a postfix operator from a module, as we do not condone enabling this throughout the entire codebase.

StaticPointers

Name StaticPointers
Flag -XStaticPointers

Allows static resolution of a limited subset of expressions to a value at compile time. This allows for some precomputation of values during compilation for later lookup at runtime. While this can be useful for some low-level code, much care must be taken when it is used.

UndecidableInstances

Name UndecidableInstances
Flag -XUndecidableInstances

This extension permits the definition of typeclass instances that could potentially lead to non-termination of the type-checker. This is sometimes necessary to define the instance you want, but care must be taken to ensure that you only enable this extension when you are sure that your instances are terminating.

UndecidableSuperclasses

Name UndecidableSuperclasses
Flag -XUndecidableSuperclasses

Permits the definition of superclass constraints which can potentially lead to the non-termination of the type-checker. Much like the above, this is sometimes necessary but should only be enabled when you are sure that you will not cause the typechecker to loop.

Disallowed Extensions

If a language extension hasn't been listed in the above sections, then it is considered to be disallowed throughout the Enso codebases. If you have a good reason to want to use one of these disallowed extensions, please talk to Ara or Wojciech to discuss its usage.

If an extension not listed above is implied by one of the extensions listed above (e.g. -XRankNTypes implies -XExplicitForall), then the implied extension is also considered at least as safe as the category the implying extension is in.