1
1
mirror of https://github.com/github/semantic.git synced 2024-11-30 06:07:23 +03:00
semantic/docs/assignment.md
2020-04-20 13:59:52 -04:00

12 KiB
Raw Blame History

Note that this document describes a process that is now deprecated. For more information, see documentation on adding a new language.

What is Assignment?

"Assignment" refers to the part of our system that parses parse trees. The step preceding assignment uses tree-sitter to parse source code and output rose trees labeled with symbols in the language's grammar and source locations (represented as byte range and span). Assignment is a second layer of parsing required to get these ASTs in a shape appropriate for our Haskell project and to support the types of analyses we'd like to perform further along. Assignment represents a partial map from AST nodes onto another structure, typically terms.

We do this for a few reasons:

  1. Generalization: this approach lets us reason about all languages in a standardized way. We can build necessary machinery to define the type of each specific language independently, while retaining the ability to share computations on the overlapping constructs (such as if statements).
  2. Type system: Assignment gives us a way to represent the semantics of a language such that we know its behavior at runtime. Dynamic languages such as Ruby or JavaScript have looser rules which produce unpredictable results or may perform implicit type conversion at runtime. Having a finite list of items in syntax trees gives us a way to place a hard limitation on expected behavior.

Relevant files

Relevant ghci commands

  • Assignment AST: quieterm <$> parseFile javaParser "/example.java"
  • Tree-sitter AST: fmap nodeSymbol <$> parseFile javaASTParser "/example.java"

How we construct an assignment

Assignments for supported language are defined in a file that is named something like this: Language.InsertSomeLanguage.Assignment. Working in this file is akin to writing another grammar, one that parses the tree-sitter parse tree instead of source code.

Assignments typically take on this pattern:

someLanguageConstruct :: Assignment
someLanguageConstruct = makeTerm <$> symbol NodeNameOfSymbolToMatch <*> children (SyntaxDataType <$> field1 <*> field2)

The building blocks that compose this DSL come from: Assigning.Assignment, explained below.

The underlying machinery of Assigning.Assignment

Assignments are matched based on symbol, sequence and hierarchy. This gives us specificity (ex., in @x = y@, both @x@ and @y@ might have the same symbol, @Identifier@, but we can remove the possibility of doubt by assigning the left hand side to a variable declaration and the right hand side to a variable reference).

Assignments can be any of the following primitive rules

Assigning.Assignment gives us the primitive rules:

symbol:

  • Produces the current node's location by matching against a specific symbol in the source language's grammar.
  • Succeeds iff a) a current node exists, and b) its symbol is equal to the argument symbol. This rule fails if no node is found, for example at the end of any branch once we've matched all nodes.
  • Matching a symbol rule does not advance past the current node, meaning you can match against a symbol and also match against the node's children. This also means some care must be taken, as repeating a symbol with many or some will never advance past the current node and could infinitely loop.

location:

  • Produces the current node's location (byte Range and Span)
  • Always succeeds. If there is no current node (i.e. if matching has advanced past the root node or past the last child node when operating within a children rule), the location is instead the end of the most recently matched node, specified as a zero-width Range and Span. location rules don't advance past the current node, meaning that you can both match a node's location and other properties. Dissimilar to source which requires we match a node—the location rule is only used for bookkeeping and thus does not place that hard expectation.

source:

  • Produces a node's source as a ByteString.
  • source rules succeed whenever a current node exists (i.e. matching has not advanced past the root node or past the last child node when operating within a children rule).
  • source is intended to match leaf nodes such as comments. source rules advance past the current node.

children:

  • children rules apply their argument (an assignment) to the children of the current node, succeeding iff a) a current node exists, b) the argument assignment matches the children, and c) there are no (regular) nodes left over (the distinction between nodes is given below in the Matching tokens section), producing the result of matching the argument assignment against the children.

Alternative instance:

  • Allows you to define the way you expect to get a ByteString, failing if it doesn't meet the standards you've defined. Insert example
  • Allows you to accept 0 or more alternatives—0 meaning accept nothing and failing on the choice between them. empty assignments always fail. This can be used (in combination with the 'Monad' instance) to, for example, fail if a source assignment produces an ill-formatted ByteString. However, see below re: committed choice.

Applicative instance:

  • pure (or via the 'Monad' instance, return) assignments always succeed, producing the passed value. They do not advance past the current node. In combination with the Alternative instance, pure can provide default values when optional syntax is not present in the AST.

Ways to combine assignments

  1. The Functor instance maps values from the AST (Loc, ByteString, etc.) onto another structure.

  2. The Applicative instance assigns sequences of (sibling) AST nodes in order, as well as providing pure assignments.

  3. The Alternative instance chooses between a set of assignments, as well as providing empty assignments (see above). See below re: committed choice for best practices for efficiency & error reporting when it comes to assigning multiple alternatives. Most high-level assignments (e.g. “declaration” or “statement” assignments) consist of choices among two or more Applicative chains of assignments, mirroring the structure of the parsers choices. The Alternative instance also enables repetitions via the 'many' (≥ 0 repetitions) and 'some' (≥ 1 repetition) methods. Finally, the 'optional' function uses the 'Alternative' instance to assign a value in 'Maybe', succeeding with 'Nothing' when unmatched.

  4. The Monad instance allows assignments to depend on the results of earlier assignments. In general, most assignments should not be written using the Monad instance; however, some specific situations require it, e.g. assigning 'x += y' to be equivalent to 'x = x + y'. need example of this--I thought the monad instance allows us to chain things together such that we don't have to use do notation

Assignment best practices

Because of their flexibility, the same assignment can often be written in multiple different ways. The following best practices should ensure efficient assignment with clear error messages for ill-formed AST.

Committed choice

Assignments can represent alternatives as either committed or uncommitted choices, both written with <|>. “Committed” in this context means that a failure in one of the alternatives will not result in backtracking followed by an attempt of one of the other alternatives; thus, committed choice is more efficient. (By the same token, it enables much better error messages since backtracking erases most of the relevant context.) Committed choices are constructed via the following rules:

  1. empty is dropped from choices.
  2. symbol rules construct a committed choice (with only a single alternative).
  3. fmap (and by extension <$> and <$) of a committed choice is a committed choice.
  4. <*> (and by extension *> and <*) with a committed choice on the left is a committed choice.
  5. >>= (and by extension >>, =<<, and <<) of a committed choice is a committed choice. It may be helpful to think of this and the above rule for <*> as “sequences starting with committed choices remain committed choices.”
  6. <|> of two committed choices is a committed choice.

Finally, if a given choice is not a committed choice, it is an uncommitted choice.

Given the above, it is a good idea to always start an assignment for a given piece of syntax with either a symbol rule or an fmap over a symbol rule. When assigning multiple pieces of syntax, place any known uncommitted choices at the (rightmost) end of the chain; <|> is left-associative, so this guarantees that youre adding at most one uncommitted choice on top of the ones already present.

Matching tokens

AST symbols are classified by their 'symbolType' as either 'Regular', 'Anonymous', or 'Auxiliary'. 'Regular' is for the symbols of explicitly named productions in the grammar; 'Anonymous' is for unnamed productions of content such as tokens. 'Auxiliary' never appears in ASTs. Most of the time, assignments are only concerned with the named productions, and thus will be using 'Regular' symbols. Therefore, when matching a committed choice of all-'Regular' symbols, nodes with 'Anonymous' symbols will be skipped. However, in some cases grammars dont provide a named symbol for e.g. every kind of infix operator, and thus the only way to differentiate between them is by means of a symbol rule for an 'Anonymous' token. In these cases, and before every other kind of assignment, the 'Anonymous' nodes will not be skipped so that matching can succeed.

This means that in addition to the rule of thumb for committed choices (see above), it's a good idea to try to match 'Regular' symbols up front, and only match 'Anonymous' ones in the middle of a chain. That will ensure that you dont have to make redundant effort to explicitly skip 'Anonymous' nodes ahead of multiple alternatives, and can instead rely on them being automatically skipped except when explicitly required.

Data.Syntax

This is the file from where we get combinators (makeTerm, contextualize, emptyTerm, handleError, infixContext, makeTerm, makeTerm', makeTerm'', makeTerm1, parseError, postContextualize).

The difference between makeTerm, makeTerm' and makeTerm1

The makeTerm family of functions all construct terms whose Range and Span are at least as large as the union of all of their child terms Ranges and Spans. In the case of makeTerm and makeTerm', it also includes the Range and Span matched by their first argument, typically produced by a symbol rule. By way of contrast, makeTerm1 is named by analogy with foldr1, and only takes a single argument, which is expected to be a non-empty piece of syntax, i.e. a piece of syntax which has one or more child terms; the resulting terms Range and Span are then precisely the union of the child terms.

Data.Syntax.SomeSubsetOfLanguageFeatures

These are files where data types for syntaxes are defined for use in your assignment.

Parsing.Parser

Contains data constructors required to produce ASTs.

  • ASTParser - parser producing AST using tree-sitter language grammar.
  • AssignmentParser - parser producing à la carte term given an AST-producing parser and an Assignment onto Terms in some syntax type.
  • javaParser - parses according to what's defined in Language.Java.Assignment.