Haskell grammar for tree-sitter.

• Haskell 2010 Language Report – Syntax References
• GHC Language Extensions

These extensions are supported ✅, unsupported ❌ or not applicable because they don't involve parsing ➖️:

• AllowAmbiguousTypes ➖️
• ApplicativeDo ➖️
• Arrows ❌
• BangPatterns ✅
• BinaryLiterals ✅
• BlockArguments ✅
• CApiFFI ✅
• ConstrainedClassMethods ✅
• ConstraintKinds ✅
• CPP ✅
• CUSKs ✅
• DataKinds ✅
• DatatypeContexts ✅
• DefaultSignatures ✅
• DeriveAnyClass ➖️
• DeriveDataTypeable ➖️
• DeriveFoldable ➖️
• DeriveFunctor ➖️
• DeriveGeneric ➖️
• DeriveLift ➖️
• DeriveTraversable ➖️
• DerivingStrategies ✅
• DerivingVia ✅
• DisambiguateRecordFields ➖️
• DuplicateRecordFields ➖️
• EmptyCase ✅
• EmptyDataDecls ✅
• EmptyDataDeriving ✅
• ExistentialQuantification ✅
• ExplicitForAll ✅
• ExplicitNamespaces ✅
• ExtendedDefaultRules ➖️
• FlexibleContexts ✅
• FlexibleInstances ✅
• ForeignFunctionInterface ✅
• FunctionalDependencies ✅
• GADTs ✅
• GADTSyntax ✅
• GeneralisedNewtypeDeriving ➖️
• GHCForeignImportPrim ✅
• Haskell2010 ➖️
• Haskell98 ➖️
• HexFloatLiterals ✅
• ImplicitParams ✅
• ImplicitPrelude ➖️
• ImportQualifiedPost ✅
• ImpredicativeTypes ➖️
• IncoherentInstances ➖️
• InstanceSigs ✅
• InterruptibleFFI ✅
• KindSignatures ✅
• LambdaCase ✅
• LexicalNegation ❌
• LiberalTypeSynonyms ✅
• LinearTypes ✅
• ListTuplePuns ✅
• MagicHash ✅
• Modifiers ❌
• MonadComprehensions ➖️
• MonadFailDesugaring ➖️
• MonoLocalBinds ➖️
• MonomorphismRestriction ➖️
• MultiParamTypeClasses ✅
• MultiWayIf ✅
• NamedFieldPuns ✅
• NamedWildCards ✅
• NegativeLiterals ➖️
• NondecreasingIndentation ✅
• NPlusKPatterns ➖️
• NullaryTypeClasses ✅
• NumDecimals ➖️
• NumericUnderscores ✅
• OverlappingInstances ➖️
• OverloadedLabels ✅
• OverloadedLists ➖️
• OverloadedRecordDot ✅
• OverloadedRecordUpdate ✅
• OverloadedStrings ➖️
• PackageImports ✅
• ParallelListComp ✅
• PartialTypeSignatures ✅
• PatternGuards ✅
• PatternSynonyms ✅
• PolyKinds ➖️
• PostfixOperators ➖️
• QualifiedDo ✅
• QuantifiedConstraints ✅
• QuasiQuotes ✅
• Rank2Types ✅
• RankNTypes ✅
• RebindableSyntax ➖️
• RecordWildCards ➖️
• RecursiveDo ✅
• RequiredTypeArguments ✅
• RoleAnnotations ✅
• Safe ➖️
• ScopedTypeVariables ✅
• StandaloneDeriving ✅
• StandaloneKindSignatures ✅
• StarIsType ✅
• StaticPointers ❌
• Strict ➖️
• StrictData ✅
• TemplateHaskell ✅
• TemplateHaskellQuotes ✅
• TraditionalRecordSyntax ➖️
• TransformListComp ✅
• Trustworthy ➖️
• TupleSections ✅
• TypeAbstractions ✅
• TypeApplications ✅
• TypeData ✅
• TypeFamilies ✅
• TypeFamilyDependencies ✅
• TypeInType ✅
• TypeOperators ✅
• TypeSynonymInstances ➖️
• UnboxedSums ✅
• UnboxedTuples ✅
• UndecidableInstances ➖️
• UndecidableSuperClasses ➖️
• UnicodeSyntax ✅
• UnliftedFFITypes ➖️
• UnliftedNewtypes ✅
• Unsafe ➖️
• ViewPatterns ✅

Preprocessor #elif and #else directives cannot be handled correctly, since the parser state would have to be manually reset to what it was at the #if. As a workaround, the code blocks in the alternative branches are parsed as part of the directives.

The grammar contains several supertypes, which group multiple other node types under a single name.

Supertype names do not occur as extra nodes in parse trees, but they can be used in queries in special ways:

• As an alias, matching any of their subtypes
• As prefix for one of their subtypes, matching its symbol only when it occurs as a production of the supertype

For example, the query (expression) matches the nodes infix, record, projection, constructor, and the second and third variable in this tree for cats <> Cat {mood = moods.sleepy}:

(infix
  (variable)
  (operator)
  (record
    (constructor)
    (field_update
      (field_name (variable))
      (projection (variable) (field_name (variable)))))))))

The two occurrences of variable in field_name (mood and sleepy) are not expressions, but record field names part of a composite record expression.

Matching variable nodes specifically that are expressions is possible with the second special form. A query for (expression/variable) will match only the other two, cats and moods.

The grammar's supertypes consist of the following sets:

• expression

  Rules that are valid in any expression position, excluding type applications, explicit types and expression signatures.

• pattern

  Rules that are valid in any pattern position, excluding type binders, explicit types and pattern signatures.

• type

  Types that are either atomic (have no ambiguous associativity, like bracketed constructs, variables and type constructors), applied types or infix types.

• quantified_type

  Types prefixed with a forall, context or function parameter.

• constraint

  Almost the same rules as type, but mirrored for use in contexts.

• constraints

  Analog of quantified_type, for constraints with forall or context.

• type_param

  Atomic nodes in type and class heads, like the three nodes following A in data A @k a (b :: k).

• declaration

  All top-level declarations, like functions and data types.

• decl

  Shorthand for declarations that are also valid in local bindings (let and where) and in class and instance bodies, except for fixity declarations. Consists of signature, function and bind.

• class_decl and instance_decl

  All declarations that are valid in classes and instances, which includes associated type and data families.

• statement

  Different forms of do-notation statements.

• qualifier

  Different forms of list comprehension qualifiers.

• guard

  Different forms of guards in function equations and case alternatives.

The main driver for generating and testing the parser for this grammar is the tree-sitter CLI. Other components of the project require additional tools, described below.

Some are made available through npm – for example, npx tree-sitter runs the CLI. If you don't have tree-sitter available otherwise, prefix all the commands in the following sections with npx.

The CLI writes the shared library containing the parser to the directory denoted by $TREE_SITTER_LIBDIR. If that variable is unset, it defaults to $HOME/.cache/tree-sitter/lib.

In order to avoid clobbering this global directory with development versions, you can set the env var to a local path:

export TREE_SITTER_LIBDIR=$PWD/.lib

The javascript file grammar.js contains the entry point into the grammar's production rules. Please consult the tree-sitter documentation for a comprehensive introduction to the syntax and semantics.

Parsing starts with the first item in the rules field:

{
  rules: {
    haskell: $ => seq(
      optional($.header),
      optional($._body),
    ),
  }
}

The first step in the development workflow converts the javascript rule definitions to C code in src/parser.c:

$ tree-sitter generate

Two byproducts of this process are written to src/grammar.json and src/node-types.json.

The C code is automatically compiled by most of the test tools mentioned below, but you can instruct tree-sitter to do it in one go:

$ tree-sitter generate --build

If you've set $TREE_SITTER_LIBDIR as mentioned above, the shared object will be written to $PWD/.lib/haskell.so.

Aside from the generated src/parser.c, tree-sitter will also compile and link src/scanner.c into this object. This file contains the external scanner, which is a custom extension of the built-in lexer whose purpose is to handle language constructs that cannot be expressed (efficiently) in the javascript grammar, like Haskell layouts.

The parser can be compiled to WebAssembly as well, which requires emscripten:

$ tree-sitter build --wasm

The resulting binary is written to $PWD/tree-sitter-haskell.wasm.

The most fundamental test infrastructure for tree-sitter grammars consists of a set of code snippets with associated reference ASTs stored in ./test/corpus/*.txt.

$ tree-sitter test

Individual tests can be run by specifying (a substring of) their description with -f:

$ tree-sitter test -f 'module: exports empty'

The project contains several other types of tests:

• test/parse/run.bash [update] [test names ...] parses the files in test/parse/*.hs and compares the output with test/parse/*.target. If update is specified as the first argument, it will update the .target file for the first failing test.

• test/query/run.bash [update] [test names ...] parses the files in test/query/*.hs, applies the queries in test/query/*.query and compares the output with test/query/*.target, similar to test/parse.

• test/rust/parse-test.rs contains a few tests that use tree-sitter's Rust API to extract the test ranges for terminals in a slightly more convenient way. This requires cargo to be installed, and can be executed with cargo test (which also runs the tests in bindings/rust).

• test/parse-libs [wasm] clones a set of Haskell libraries to test/libs and parses the entire codebase. When invoked as test/parse-libs wasm, it will use the WebAssembly parser. This requires bc to be installed.

• test/parse-lib name [wasm] parses only the library name in that directory (without cloning the repository).

The shared library built by tree-sitter test includes debug symbols, so if the scanner segfaults you can just run coredumpctl debug to inspect the backtrace and memory:

newline_lookahead () at src/scanner.c:2583
2583                ((Newline *) 0)->indent = 5;
(gdb) bt
#0  newline_lookahead () at src/scanner.c:2583
#1  0x00007ffff7a0740e in newline_start () at src/scanner.c:2604
#2  scan () at src/scanner.c:2646
#3  eval () at src/scanner.c:2684
#4  tree_sitter_haskell_external_scanner_scan (payload=<optimized out>, lexer=<optimized out>,
    valid_symbols=<optimized out>) at src/scanner.c:2724
#5  0x0000555555772488 in ts_parser.lex ()

For more control, launch gdb tree-sitter and start the process with run test -f 'some test', and set a breakpoint with break tree_sitter_haskell_external_scanner_scan.

To disable optimizations, run tree-sitter test --debug-build.

The test and parse commands offer two modes for obtaining detailed information about the parsing process.

With tree-sitter test --debug, every lexer step and shift/reduce action is printed to stderr.

With tree-sitter test --debug-graph, the CLI will generate an HTML file showing a graph representation of every step. This requires graphviz to be installed.