semantic/docs/coding-style.md

Haskell is a syntactically-flexible language, which gives the programmer a tremendous amount of leeway regarding the appearance of their code. This is a set of best practices that we use in semantic and its related projects.

This file draws from the style guides written by Johan Tibbel and Kowainik.

General guidelines

Make your code look like the code around it. Consistency is the name of the game.

Use stylish-haskell for general formatting. We have our own style defined in .stylish-haskell.yaml, and it automates many uninteresting style debates: it will format LANGUAGE pragmas, alphabetize and align imports, etc. Atom, Emacs, vim, and most other editors can run it automatically. You should also set up your editor to remove trailing whitespace.

Our CI systems ensure that all patches pass hlint's muster. We have our own set of rules in .hlint.yaml.

We strongly recommend adding Haddock documentation to any function/data type, unless its purpose is immediately apparent from its name. Comments should describe the "why", type signatures should describe the "what", and the code should describe the "how".

Formatting

2 spaces everywhere. Tabs are forbidden. Haskell indentation can be unpredictable, so generally stick with what your editor suggests. There is no hard line-length limit, though if you go beyond 110 or 120 you should generally split it up, especially for type signatures.

Use applicative notation when constructing simple data types.

thing :: Parser Foo

-- Broke:
thing = do
  a <- bar
  b <- baz
  pure (Foo a b)

-- Woke:
thing = Foo <$> bar <*> baz

Overreliance on applicative notation can create code that is difficult to read. Don't use applicative notation in combination with operator sections. If in doubt, write it with do notation and see if it's more immediately comprehensible.

Avoid the Applicative instance for functions. That means you, Rob.

Use leading commas for records, exports, and lists.

Leading commas make it easy to add and remove fields without introducing syntax errors, and properly aligned records are easy to read:

data Pos = Pos
  { posLine   :: Int
  , posColumn :: Int
  }

Align typographical symbols.

-> in case statements and signatures, = in functions, and :: in records should be aligned. Your editor can help with this. In certain situations, aligning symbols may decrease readability, e.g. complicated case statements. Use your best judgment.

Naming

Locally bound variables (such as the arguments to functions, or helpers defined in a where clause) can have short names, such as x or go. Globally bound functions and variables should have descriptive names.

You'll often find yourself implementing functions that conflict with Prelude definitions. If this is the case, avoid adding a prefix to these functions, and instead import them qualified.

-- Broke
foo = heapLookup thing
-- Woke
foo = Heap.lookup thing

Unlike many Haskell projects, we rely in places on variable shadowing (especially in open-recursive functions). Avoid variable shadowing if possible, as it can lead to unintuitive error messages; you are free to disable shadowing on a per-file basis with {-# OPTIONS_GHC -Wshadow #-}

Functions

Don't go buckwild with infix operators.

Sensible use of infix operators can provide serious readability benefits, but often the best tool is just a named function. If you're defining new operators, make sure that you have a solid justification for doing so.

Avoid list comprehensions.

In almost all cases, map, filter, fold, and the [] monad are more flexible and readable.

Don't go buckwild with point-free definitions.

Point-free style can help or hinder readability, depending on the context. If a function is expressed naturally with the . operator, then do so, but if you have to gyrate the definition to write it point-free, then you should probably just write out the variable names. If you are reviewing someone else's PR and find a point-free definition hard to read, ask them to simplify/clarify it.

Prefer . and $ to parentheses.

Parentheses can make a function harder to edit, since parentheses have to be balanced. The composition and application operators (. and $) can reduce clunkiness.

-- Broke
f (g (h x))
-- Woke
f $ g $ h x
-- Bespoke
f . g . h $ x

Do not use partial functions.

hlint will catch several classes of partial functions (head, fromJust, etc.). Do not use error if at all possible, and never use undefined.

Data Types

Prefer newtypes to types.

newtype values are zero-cost to construct and eliminate, and provide more informative error messages than type synonyms. Only use type for convenience aliases to existing types.

Don't use String.

String is almost always the wrong choice. If your type represents human-readable strings, use Text; if you have a blob of bytes, use ByteString. -XOverloadedStrings is enabled globally to make this easy.

Use -XDerivingStrategies when using -XGeneralizedNewtypeDeriving or -XDeriveAnyClass.

Subtle bugs can creep in if you fail to specify the correct strategy, so prefer specifying an explicit strategy even if GHC doesn't require it. If all the classes you're deriving are the stock classes (Eq, Ord, Show, etc.), there's no need to specify a strategy.

Only use record selectors on single-constructor types.

The following code generates two partial functions, which is bad:

data Bad = Evil { getInt :: Int }
         | Bad  { getFloat :: Float }

If you need fields that properly take failure into account, consider using lens and generating Lenses and Traversals, which avoid calls to error.

An exception to this case is when record selectors are present to provide clarity regarding field names. In this case, prefix the field names with _ (so that GHC will avoid warning you), and export only the constructors, like so:

module Thing (Foo (Bar, Baz)) where

data Foo = Bar { _thing1 :: String
               , _thing2 :: String
               }
         | Baz { _thingRed :: Float
               , _thingBlue :: Float
               }

lens

The lens library is large and intimidating, and poorly-written lens code can be impossible to maintain. Here are some suggestions to keep complexity at bay:

• Prefer hand-written lenses to those generated by Template Haskell, except if the hand-written lens is significantly less reasonable than a TH splice.
• Only use these infix operators: ^. (view), ^? (preview), .~ (set), %~ (over), and ^.. (toListOf). Prefer prefix functions in other cases.
• Prefer specific imports: if all you need is ^., import just Control.Lens.Getter rather than Control.Lens.
• Only use lenses when you have to view and manipulate deeply nested data types. If you can get away with a plain old record, do so.

Miscellany

• Prefer eliminators like maybe and either to explicit pattern-matching.
• Prefer guards and the bool eliminator to if-then-else statements.
• Prefer where to let, except in the case of nested wheres.
• Don't use {-# ANN … #-} to disable hlint warnings, as it can slow down compilation. If you need to disable lints in a file, do so in .hlint.yaml.