chriswailes / RLTK

The Ruby Language Toolkit
http://chriswailes.github.io/RLTK/
University of Illinois/NCSA Open Source License
453 stars 35 forks source link

Welcome to the Ruby Language Toolkit

RLTK is a collection of classes and methods designed to help programmers work with languages in an easy to use and straightforward manner. This toolkit provides the following features:

In addition, RLTK includes several ready-made lexers and parsers and a Turing-complete language called Kazoo for use in your code and as examples for how to use the toolkit.

Why Use RLTK

Here are some reasons to use RLTK to build your lexers, parsers, and abstract syntax trees, as well as generating LLVM IR and native object files:

Lexers

To create your own lexer using RLTK you simply need to subclass the {RLTK::Lexer} class and define the rules that will be used for matching text and generating tokens. Here we see a simple lexer for a calculator:

class Calculator < RLTK::Lexer
  rule(/\+/) { :PLS }
  rule(/-/)  { :SUB }
  rule(/\*/) { :MUL }
  rule(/\//) { :DIV }

  rule(/\(/) { :LPAREN }
  rule(/\)/) { :RPAREN }

  rule(/[0-9]+/) { |t| [:NUM, t.to_i] }

  rule(/\s/)
end

The {RLTK::Lexer.rule} method's first argument is the regular expression used for matching text. The block passed to the function is the action that executes when a substring is matched by the rule. These blocks must return the type of the token (which must be in ALL CAPS; see the Parsers section), and optionally a value. In the latter case you must return an array containing the type and value, which you can see an example of in the Calculator lexer shown above. The values returned by the proc object are used to build a {RLTK::Token} object that includes the type and value information, as well as information about the line number the token was found on, the offset from the beginning of the line to the start of the token, and the length of the token's text. If the type value returned by the proc is nil the input is discarded and no token is produced.

The {RLTK::Lexer} class provides both {RLTK::Lexer.lex} and {RLTK::Lexer.lex_file}. The {RLTK::Lexer.lex} method takes a string as its argument and returns an array of tokens, with an end of stream token automatically added to the result. The {RLTK::Lexer.lex_file} method takes the name of a file as input, and lexes the contents of the specified file.

The Lexing Environment

The proc objects passed to the {RLTK::Lexer.rule} methods are evaluated inside an instance of the {RLTK::Lexer::Environment} class. This gives you access to methods for manipulating the lexer's state and flags (see bellow). You can also subclass the environment inside your lexer to provide additional functionality to your rule blocks. When doing so you need to ensure that you name your new class Environment like in the following example:

class MyLexer < RLTK::Lexer
  ...

  class Environment < Environment
    def helper_function
      ...
    end

  ...
  end
end

Using States

The lexing environment may be used to keep track of state inside your lexer. When rules are defined they are defined inside a given state, which is specified by the second parameter to {RLTK::Lexer.rule}. The default state is cleverly named :default. When the lexer is scanning the input string for matching rules, it only considers the rules for the given state.

The methods used to manipulate state are:

States may be used to easily support nested comments.

class StateLexer < RLTK::Lexer
  rule(/a/) { :A }
  rule(/\s/)

  rule(/\(\*/) { push_state(:comment) }

  rule(/\(\*/, :comment) { push_state(:comment) }
  rule(/\*\)/, :comment) { pop_state            }
  rule(/./,    :comment)
end

By default the lexer will start in the :default state. To change this, you may use the {RLTK::Lexer.start} method.

Using Flags

The lexing environment also maintains a set of flags. This set is manipulated using the following methods:

When rules are defined they may use a third parameter to specify a list of flags that must be set before the rule is considered when matching substrings. An example of this usage follows:

class FlagLexer < RLTK::Lexer
  rule(/a/) { set_flag(:a); :A }

  rule(/\s/)

  rule(/b/, :default, [:a])     { set_flag(:b); :B }
  rule(/c/, :default, [:a, :b]) { :C               }
end

Instantiating Lexers

In addition to using the {RLTK::Lexer.lex} class method you may also instantiate lexer objects. The only difference then is that the lexing environment used between subsequent calls to {RLTK::Lexer#lex} is the same object, and therefor allows you to keep persistent state.

First and Longest Match

A RLTK::Lexer may be told to select either the first substring that is found to match a rule or the longest substring to match any rule. The default behavior is to match the longest substring possible, but you can change this by using the {RLTK::Lexer.match_first} method inside your class definition as follows:

class MyLexer < RLTK::Lexer
  match_first

  ...
end

Match Data

Because it isn't RLTK's job to tell you how to write lexers and parsers, the MatchData object from a pattern match is available inside the Lexer::Environment object via the match accessor.

Context-Free Grammars

The {RLTK::CFG} class provides an abstraction for context-free grammars. For the purpose of this class terminal symbols appear in ALL CAPS, and non-terminal symbols appear in all lowercase. Once a grammar is defined the {RLTK::CFG#first_set} and {RLTK::CFG#follow_set} methods can be used to find first and follow sets.

Defining Grammars

A grammar is defined by first instantiating the {RLTK::CFG} class. The {RLTK::CFG#production} and {RLTK::CFG#clause} methods may then be used to define the productions of the grammar. The production method can take a Symbol denoting the left-hand side of the production and a string describing the right-hand side of the production, or the left-hand side symbol and a block. In the first usage a single production is created. In the second usage the block may contain repeated calls to the clause method, each call producing a new production with the same left-hand side but different right-hand sides. {RLTK::CFG#clause} may not be called outside of {RLTK::CFG#production}. Bellow we see a grammar definition that uses both methods:

grammar = RLTK::CFG.new

grammar.production(:s) do
  clause('A G D')
  clause('A a C')
  clause('B a D')
  clause('B G C')
end

grammar.production(:a, 'b')
grammar.production(:b, 'G')

Extended Backus–Naur Form

The RLTK::CFG class understands grammars written in the extended Backus–Naur form. This allows you to use the *, +, and ? operators in your grammar definitions. When each of these operators are encountered additional productions are generated. For example, if the right-hand side of a production contained NUM* a production of the form num_star -> | NUM num_star is added to the grammar. As such, your grammar should not contain productions with similar left-hand sides (e.g. foo_star, bar_question, or baz_plus).

As these additional productions are added internally to the grammar a callback functionality is provided to let you know when such an event occurs. The callback proc object can either be specified when the CFG object is created, or by using the {RLTK::CFG#callback} method. The callback will receive three arguments: the production generated, the operator that triggered the generation, and a symbol (:first or :second) specifying which clause of the production this callback is for.

Helper Functions

Once a grammar has been defined you can use the following functions to obtain information about it:

Parsers

To create a parser using RLTK simply subclass RLTK::Parser, define the productions of the grammar you wish to parse, and call finalize. During finalization RLTK will build an LALR(1) parsing table, which may contain conflicts that can't be resolved with LALR(1) lookahead sets or precedence/associativity information. Traditionally, when parser generators such as YACC encounter conflicts during parsing table generation they will resolve shift/reduce conflicts in favor of shifts and reduce/reduce conflicts in favor of the production that was defined first. This means that the generated parsers can't handle ambiguous grammars.

RLTK parsers, on the other hand, can handle all context-free grammars by forking the parse stack when shift/reduce or reduce/reduce conflicts are encountered. This method is called the GLR parsing algorithm and allows the parser to explore multiple different possible derivations, discarding the ones that don't produce valid parse trees. GLR parsing is more expensive, in both time and space requirements, but these penalties are only payed when a parser for an ambiguous grammar is given an input with multiple parse trees, and as such most parsing should proceed using the faster LALR(1) base algorithm.

Defining a Grammar

Let us look at the simple prefix calculator included with RLTK:

class PrefixCalc < RLTK::Parser
  production(:e) do
    clause('NUM') {|n| n}

    clause('PLS e e') { |_, e0, e1| e0 + e1 }
    clause('SUB e e') { |_, e0, e1| e0 - e1 }
    clause('MUL e e') { |_, e0, e1| e0 * e1 }
    clause('DIV e e') { |_, e0, e1| e0 / e1 }
  end

  finalize
end

The parser uses the same method for defining productions as the {RLTK::CFG} class. In fact, the parser forwards the {RLTK::Parser.production} and {RLTK::Parser.clause} method invocations to an internal {RLTK::CFG} object after removing the parser specific information. To see a detailed description of grammar definitions please read the Context-Free Grammars section bellow.

It is important to note that the proc objects associated with productions should evaluate to the value you wish the left-hand side of the production to take.

The default starting symbol of the grammar is the left-hand side of the first production defined (in this case, e). This can be changed using the {RLTK::Parser.start} function when defining your parser.

Make sure you call finalize at the end of your parser definition, and only call it once.

Shortcuts

RLTK provides several shortcuts for common grammar constructs. Right now these shortcuts include the {RLTK::Parser.list} and {RLTK::Parser.nonempty_list} methods. A list may contain 0, 1, or more elements, with an optional token or tokens separating each element. A non-empty list contains at least 1 element. An empty list with only a single list element and an empty separator is equivalent to the Kleene star. Similarly, a list with only a single list element and an empty separator is equivalent to the Kleene plus.

This example shows how these shortcuts may be used to define a list of integers separated by a :COMMA token:

class ListParser < RLTK::Parser
  nonempty_list(:int_list, :INT, :COMMA)

  finalize
end

If you wanted to define a list of floats or integers you could define your parser like this:

class ListParser < RLTK::Parser
  nonempty_list(:mixed_list, [:INT, :FLOAT], :COMMA)

  finalize
end

If you don't want to require a separator you can do this:

class ListParser < RLTK::Parser
  nonempty_list(:mixed_nonsep_list, [:INT, :FLOAT])

  finalize
end

You can also use separators that are made up of multiple tokens:

class ListParser < RLTK::Parser
  nonempty_list(:mixed_nonsep_list, [:INT, :FLOAT], 'COMMA NEWLINE?')

  finalize
end

A list may also contain multiple tokens between the separator:

class ListParser < RLTK::Parser
  nonempty_list(:foo_bar_list, 'FOO BAR', :COMMA)

  finalize
end

Lastly, you can mix all of these features together:

class ListParser < RLTK::Parser
  nonempty_list(:foo_list, ['FOO BAR', 'FOO BAZ+'], :COMMA)

  finalize
end

The productions generated by these shortcuts will always evaluate to an array. In the first two examples above the productions will produce a 1-D array containing the values of the INT or FLOAT tokens. In the last two examples the productions foo_bar_list and foo_list will produce 2-D arrays where the top level array is composed of tuples corresponding to the values of FOO, and BAR or one or more BAZs.

Precedence and Associativity

To help you remove ambiguity from your grammars RLTK lets you assign precedence and associativity information to terminal symbols. Productions then get assigned precedence and associativity based on either the last terminal symbol on the right-hand side of the production, or an optional parameter to the {RLTK::Parser.production} or {RLTK::Parser.clause} methods. When an {RLTK::Parser} encounters a shift/reduce error it will attempt to resolve it using the following rules:

  1. If there is precedence and associativity information present for all reduce actions involved and for the input token we attempt to resolve the conflict using the following rule. If not, no resolution is possible and the parser generator moves on. This conflict will later be reported to the programmer.

  2. The precedence of the actions involved in the conflict are compared (a shift action's precedence is based on the input token), and the action with the highest precedence is selected. If two actions have the same precedence the associativity of the input symbol is used: left associativity means we select the reduce action, right associativity means we select the shift action, and non-associativity means that we have encountered an error.

To assign precedence to terminal symbols you can use the {RLTK::Parser.left}, {RLTK::Parser.right}, and {RLTK::Parser.nonassoc} methods inside your parser class definition. Later declarations of associativity have higher levels of precedence than earlier declarations of the same associativity.

Let's look at the infix calculator example now:

class InfixCalc < RLTK::Parser

  left  :PLS, :SUB
  right :MUL, :DIV

  production(:e) do
    clause('NUM') { |n| n }

    clause('LPAREN e RPAREN') { |_, e, _| e }

    clause('e PLS e') { |e0, _, e1| e0 + e1 }
    clause('e SUB e') { |e0, _, e1| e0 - e1 }
    clause('e MUL e') { |e0, _, e1| e0 * e1 }
    clause('e DIV e') { |e0, _, e1| e0 / e1 }
  end

  finalize
end

The standard order of mathematical operations tells us that the correct way to group the operations in the expression 2 + 3 * 4 is 2 + (3 * 4). However, our grammar tells us that (2 + 3) * 5 is also a valid way to parse the expression, leading to a shift/reduce error in the parser. To get rid of the shift/reduce error we need some way to tell the parser how to distinguish between these two parse trees. This is where associativity comes in. If the parser has already read NUM PLS NUM and the current symbol is a MUL symbol we want to tell the parser to shift the new MUL symbol onto the stack and continue on. We do this by making the MUL symbol right associative. When the parser generator encounters a shift/reduce error it looks at the token currently being read. If it has no associativity information, the error can't be resolved; if the token is left associative, it will remove the shift action from the parser (leaving only the reduce action); if the token is right associative, it will remove the reduce action from the parser (leaving only the shift action).

Now, let us consider the expression 3 - 2 - 1. Here, the correct way to parse the expression is (3 - 2) - 1. To ensure that this case is selected over 3 - (2 - 1) we can make the SUB token left associative. This will cause the symbols NUM SUB NUM to be reduced before the second SUB symbol is shifted onto the parse stack.

Not that, to resolve a shift/reduce or reduce/reduce conflict, precedence and associativity information must be present for all actions involved in the conflict. As such, it isn't enough to simply make the MUL and DIV tokens right associative; we must also make the PLS and SUB tokens left associative.

Token Selectors

In many cases productions contain tokens who's value is unimportant. In such situations passing nil to the production's action is not useful. To prevent this happening you may use token selectors. By placing a period (.) in front of a token you can indicate to the parser that the following token is important and you wish for its value to be passed to the action. In the following example selectors are used to only pass the sub-expressions' values to the action:

class InfixCalc < RLTK::Parser

  left  :PLS, :SUB
  right :MUL, :DIV

  production(:e) do
    clause('NUM') { |n| n }

    clause('LPAREN .e RPAREN') { |e| e }

    clause('.e PLS .e') { |e0, e1| e0 + e1 }
    clause('.e SUB .e') { |e0, e1| e0 - e1 }
    clause('.e MUL .e') { |e0, e1| e0 * e1 }
    clause('.e DIV .e') { |e0, e1| e0 / e1 }
  end

  finalize
end

Argument Passing for Actions

By default the proc objects associated with productions are passed one argument for each symbol on the right-hand side of the production. This can lead to long, unwieldy argument lists. To change this behaviour you can use the {RLTK::Parser.default_arg_type} method, which accepts the :splat (default) and :array arguments. Any production actions that are defined after a call to {RLTK::Parser.default_arg_type} will use the argument passing method currently set as the default. You can switch between the different argument passing methods by calling {RLTK::Parser.default_arg_type} repeatedly.

Individual productions may specify the argument type used by their action via the arg_type parameter. If the {RLTK::Parser.production} method is passed an argument type and a block, any clauses defined inside the block will use the argument type specified by the arg_type parameter.

The Parsing Environment

The parsing environment is the context in which the proc objects associated with productions are evaluated, and can be used to provide helper functions and to keep state while parsing. To define a custom environment simply subclass {RLTK::Parser::Environment} inside your parser definition as follows:

class MyParser < RLTK::Parser
  ...

  class Environment < Environment
    def helper_function
    ...
    end

  ...
  end

  finalize
end

(The definition of the Environment class may occur anywhere inside the MyParser class definition.)

Instantiating Parsers

In addition to using the {RLTK::Parser.parse} class method you may also instantiate parser objects. The only difference then is that the parsing environment used between subsequent calls to {RLTK::Parser#parse} is the same object, and therefor allows you to keep persistent state.

Finalization Options

The {RLTK::Parser.finalize} method has several options that you should be aware of:

Parsing Options

The {RLTK::Parser.parse} and {RLTK::Parser#parse} methods also have several options that you should be aware of:

Parse Trees

The above section briefly mentions the parse_tree option. So that this neat feature doesn't get lost in the rest of the documentation here is the tree generated by the Kazoo parser from Chapter 7 of the tutorial when it parses the line def fib(a) if a < 2 then 1 else fib(a-1) + fib(a-2);:

Kazoo parse tree.

Parsing Exceptions

Calls to {RLTK::Parser.parse} may raise one of four exceptions:

Error Productions

Warning: this is the least tested feature of RLTK. If you encounter any problems while using it, please let me know so I can fix any bugs as soon as possible.

When an RLTK parser encounters a token for which there are no more valid actions (and it is on the last parse stack / possible parse-tree path) it will enter error handling mode. In this mode the parser pops states and input off of the parse stack (the parser is a pushdown automaton after all) until it finds a state that has a shift action for the ERROR terminal. A dummy ERROR terminal is then placed onto the parse stack and the shift action is taken. This error token will have the position information of the token that caused the parser to enter error handling mode. Additional tokens may have been discarded after this token.

If the input (including the ERROR token) can be reduced immediately the associated error handling proc is evaluated and we continue parsing. If no shift or reduce action is available the parser will being shifting tokens off of the input stack until a token appears with a valid action in the current state, in which case parsing resumes as normal.

The value of an ERROR non-terminal will be an array containing all of the tokens that were discarded while the parser was searching for a valid action.

The example below, based on one of the unit tests, shows a very basic usage of error productions:

class ErrorCalc < RLTK::Parser
  left  :ERROR
  right :PLS, :SUB, :MUL, :DIV, :NUM

  production(:e) do
    clause('NUM') {|n| n}

    clause('e PLS e') { |e0, _, e1| e0 + e1 }
    clause('e SUB e') { |e0, _, e1| e0 - e1 }
    clause('e MUL e') { |e0, _, e1| e0 * e1 }
    clause('e DIV e') { |e0, _, e1| e0 / e1 }

    clause('e PLS ERROR e') { |e0, _, err, e1| error("#{err.len} tokens skipped."); e0 + e1 }
  end

  finalize
end

A Note on Token Naming

In the world of RLTK both terminal and non-terminal symbols may contain only alphanumeric characters and underscores. The differences between terminal and non-terminal symbols is that terminals are ALL_UPPER_CASE and non-terminals are all_lower_case.

ASTNode

The {RLTK::ASTNode} base class is meant to be a good starting point for implementing your own abstract syntax tree nodes. By subclassing {RLTK::ASTNode} you automagically get features such as tree comparison, notes, value accessors with type checking, child node accessors and each and map methods (with type checking), and the ability to retrieve the root of a tree from any member node.

To create your own AST node classes you subclass the {RLTK::ASTNode} class and then use the {RLTK::ASTNode.child} and {RLTK::ASTNode.value} methods. By declaring the children and values of a node the class will define the appropriate accessors with type checking, know how to pack and unpack a node's children, and know how to handle constructor arguments.

Here we can see the definition of several AST node classes that might be used to implement binary operations for a language:

class Expression < RLTK::ASTNode; end

class Number < Expression
  value :value, Fixnum
end

class BinOp < Expression
  value :op, String

  child :left,  Expression
  child :right, Expression
end

The assignment functions that are generated for the children and values perform type checking to make sure that the AST is well-formed. The type of a child must be a subclass of the {RLTK::ASTNode} class, whereas the type of a value can be any Ruby class. While child and value objects are stored as instance variables it is unsafe to assign to these variables directly, and it is strongly recommended to always use the accessor functions.

When instantiating a subclass of {RLTK::ASTNode} the arguments to the constructor should be the node's values (in order of definition) followed by the node's children (in order of definition). If a constructor is given fewer arguments then the number of values and children the remaining arguments are assumed to be nil. Example:

class Foo < RLTK::ASTNode
  value :a, Fixnum
  child :b, Bar
  value :c, String
  child :d, Bar
end

class Bar < RLTK::ASTNode
  value :a, String
end

Foo.new(1, 'baz', Bar.new)

Lastly, the type of a child or value can be defined as an array of objects of a specific type as follows:

class Foo < RLTK::ASTNode
  value :strings, [String]
end

Tree Iteration and Mapping

RLTK Abstract Syntax Trees may be traversed in three different ways:

The order you wish to traverse the tree can be specified by passing the appropriate symbol to {RLTK::ASTNode#each}: :pre, :post, or :level.

You can also map one tree to another tree using the {RLTK::ASTNode#map} and {RLTK::ASTNode#map!} methods. In the former case a new tree is created and returned; in the latter case the current tree is transformed and the result of calling the provided block on the root node is returned. These methods will always visit nodes in post-order, so that all children of a node are visited before the node itself.

Code Generation

RLTK supports the generation of native code and LLVM IR, as well as JIT compilation and execution, through the {RLTK::CG} module. This module is built on top of bindings to LLVM and provides much, though not all, of the functionality of the LLVM libraries.

Acknowledgments and Discussion

Before we get started with the details, I would like to thank Jeremy Voorhis. The bindings present in RLTK are really a fork of the great work that he did on ruby-llvm.

Why did I fork ruby-llvm, and why might you want to use the RLTK bindings over ruby-llvm? There are a couple of reasons:

Before you dive into generating code, here are some resources you might want to look over to build up some background knowledge on how LLVM works:

LLVM

Since RLTK's code generation functionality is built on top of LLVM the first step in generating code is to inform LLVM of the target architecture. This is accomplished via the {RLTK::CG::LLVM.init} method, which is used like this: RLTK::CG::LLVM.init(:PPC). The {RLTK::CG::Bindings::ARCHS} constant provides a list of supported architectures. This call must appear before any other calls to the RLTK::CG module.

If you would like to see what version of LLVM is targeted by your version of RLTK you can either call the {RLTK::CG::LLVM.version} method or looking at the {RLTK::LLVM_TARGET_VERSION} constant.

Modules

Modules are one of the core building blocks of the code generation module. Functions, constants, and global variables all exist inside a particular module and, if you use the JIT compiler, a module provides the context for your executing code. New modules can be created using the {RLTK::CG::Module#initialize RLTK::CG::Module.new} method. While this method is overloaded you, as a library user, will always pass it a string as its first argument. This allows you to name your modules for easier debugging later.

Once you have created you can serialize the code inside of it into bitcode via the {RLTK::CG::Module#write_bitcode} method. This allows you to save partially generated code and then use it later. To load a module from bitcode you use the {RLTK::CG::Module.read_bitcode} method.

Types

Types are an important part of generating code using LLVM. Functions, operations, and other constructs use types to make sure that the generated code is sane. All types in RLTK are subclasses of the {RLTK::CG::Type} class, and have class names that end in "Type". Types can be grouped into to categories: fundamental and composite.

Fundamental types are those like {RLTK::CG::Int32Type} and {RLTK::CG::FloatType} that don't take any arguments when they are created. Indeed, these types are represented using a Singleton class, and so the new method is disabled. Instead you can use the instance method to get an instantiated type, or simply pass in the class itself whenever you need to reference the type. In this last case, the method you pass the class to will instantiate the type for you.

Composite types are constructed from other types. These include the {RLTK::CG::ArrayType}, {RLTK::CG::FunctionType}, and other classes. These types you must instantiate directly before they can be used, and you may not simply pass the type class as the type argument to functions inside the RLTK::CG module.

For convenience, the native integer type of the host platform is made available via {RLTK::CG::NativeIntType}.

Values

The {RLTK::CG::Value} class is the common ancestor of many classes inside the RLTK::CG module. The main way in which you, the library user, will interact with them is when creating constant values. Here is a list of some of value classes you might use:

Again, for convenience, the native integer class of the host platform is made available via {RLTK::CG::NativeInt}.

Functions

Functions in LLVM are much like C functions; they have a return type, argument types, and a body. Functions may be created in several ways, though they all require a module in which to place the function.

The first way to create functions is via a module's function collection:

mod.functions.add('my function', RLTK::CG::NativeIntType, [RLTK::CG::NativeIntType, RLTK::CG::NativeIntType])

Here we have defined a function named 'my function' in the mod module. It takes two native integers as arguments and returns a native integer. It is also possible to define the type of a function ahead of time and pass it to this method:

type = RLTK::CG::FunctionType.new(RLTK::CG::NativeIntType, [RLTK::CG::NativeIntType, RLTK::CG::NativeIntType])
mod.functions.add('my function', type)

Functions may also be created directly via the {RLTK::CG::Function#initialize RLTK::CG::Function.new} method, though a reference to a module is still necessary:

mod = Module.new('my module')
fun = Function.new(mod, 'my function', RLTK::CG::NativeIntType, [RLTK::CG::NativeIntType, RLTK::CG::NativeIntType])

or

mod  = Module.new('my module')
type = RLTK::CG::FunctionType.new(RLTK::CG::NativeIntType, [RLTK::CG::NativeIntType, RLTK::CG::NativeIntType])
fun  = Function.new(mod, 'my function', type)

Lastly, whenever you use one of these methods to create a function you may give it a block to be executed inside the context of the function object. This allows for easier building of functions:

mod.functions.add('my function', RLTK::CG::NativeIntType, [RLTK::CG::NativeIntType, RLTK::CG::NativeIntType]) do
  bb = blocks.append('entry)'
  ...
end

Basic Blocks

Once a function has been added to a module you will need to add {RLTK::CG::BasicBlock BasicBlocks} to the function. This can be done easily:

bb = fun.blocks.append('entry')

We now have a basic block that we can use to add instructions to our function and get it to actually do something. You can also instantiate basic blocks directly:

bb = RLTK::CG::BasicBlock.new(fun, 'entry')

The Builder

Now that you have a basic block you need to add instructions to it. This is accomplished using a {RLTK::CG::Builder builder}, either directly or indirectly.

To add instructions using a builder directly (this is most similar to how it is done using C/C++) you create the builder, position it where you want to add instructions, and then build them:

fun = mod.functions.add('add', RLTK::CG::NativeIntType, [RLTK::CG::NativeIntType, RLTK::CG::NativeIntType])
bb  = fun.blocks.append('entry')

builder = RLTK::CG::Builder.new

builder.position_at_end(bb)

# Generate an add instruction.
inst0 = builder.add(fun.params[0], fun.params[1])

# Generate a return instruction.
builder.ret(inst0)

You can get rid of some of those references to the builder by using the {RLTK::CG::Builder#build} method:

fun = mod.functions.add('add', RLTK::CG::NativeIntType, [RLTK::CG::NativeIntType, RLTK::CG::NativeIntType])
bb  = fun.blocks.append('entry')

builder = RLTK::CG::Builder.new

builder.build(bb) do
  ret add(fun.params[0], fun.params[1])
end

To get rid of more code:

fun = mod.functions.add('add', RLTK::CG::NativeIntType, [RLTK::CG::NativeIntType, RLTK::CG::NativeIntType])
bb  = fun.blocks.append('entry')

RLTK::CG::Builder.new(bb) do
  ret add(fun.params[0], fun.params[1])
end

or

fun = mod.functions.add('add', RLTK::CG::NativeIntType, [RLTK::CG::NativeIntType, RLTK::CG::NativeIntType])
fun.blocks.append('entry') do
  ret add(fun.params[0], fun.params[1])
end

or even

mod.functions.add('add', RLTK::CG::NativeIntType, [RLTK::CG::NativeIntType, RLTK::CG::NativeIntType]) do
  blocks.append('entry') do |fun|
    ret add(fun.params[0], fun.params[1])
  end
end

In the last two examples a new builder object is created for the block. It is possible to specify the builder to be used:

builder = RLTK::CG::Builder.new

mod.functions.add('add', RLTK:CG::NativeIntType, [RLTK::CG::NativeIntType, RLTK::CG::NativeIntType]) do
  blocks.append('entry', builder) do |fun|
    ret add(fun.params[0], fun.params[1])
  end
end

For an example of where this is useful, see the Kazoo tutorial.

The Contractor

An alternative to using the {RLTK::CG::Builder} class is to use the {RLTK::CG::Contractor} class, which is a subclass of the Builder and includes the Filigree::Visitor module. (Get it? It's a visiting builder!) By subclassing the Contractor you can define blocks of code for handling various types of AST nodes and leave the selection of the correct code up to the {RLTK::CG::Contractor#visit} method. In addition, the :at and :rcb options to the visit method make it much easier to manage the positioning of the Builder.

Here we can see how easy it is to define a block that builds the instructions for binary operations:

on Binary do |node|
  left  = visit node.left
  right = visit node.right

  case node
    when Add then fadd(left, right, 'addtmp')
    when Sub then fsub(left, right, 'subtmp')
    when Mul then fmul(left, right, 'multmp')
    when Div then fdiv(left, right, 'divtmp')
    when LT  then ui2fp(fcmp(:ult, left, right, 'cmptmp'), RLTK::CG::DoubleType, 'booltmp')
  end
end

AST nodes whose translation requires the generation of control flow will require the creation of new BasicBlocks and the repositioning of the builder. This can be easily managed:

on If do |node|
  cond_val = visit node.cond
  fcmp :one, cond_val, ZERO, 'ifcond'

  start_bb = current_block
  fun      = start_bb.parent

  then_bb               = fun.blocks.append('then')
  then_val, new_then_bb = visit node.then, at: then_bb, rcb: true

  else_bb               = fun.blocks.append('else')
  else_val, new_else_bb = visit node.else, at: else_bb, rcb: true

  merge_bb = fun.blocks.append('merge', self)
  phi_inst = build(merge_bb) { phi RLTK::CG::DoubleType, {new_then_bb => then_val, new_else_bb => else_val}, 'iftmp' }

  build(start_bb) { cond cond_val, then_bb, else_bb }

  build(new_then_bb) { br merge_bb }
  build(new_else_bb) { br merge_bb }

  returning(phi_inst) { target merge_bb }
end

More extensive examples of how to use the Contractor class can be found in the Kazoo tutorial chapters.

Execution Engines

Once you have generated your code you may want to run it. RLTK provides bindings to both the LLVM interpreter and JIT compiler to help you do just that. Creating a JIT compiler is pretty simple.

mod = RLTK::CG::Module.new('my module')
jit = RLTK::CG::JITCompiler(mod)

mod.functions.add('add', RLTK:CG::NativeIntType, [RLTK::CG::NativeIntType, RLTK::CG::NativeIntType]) do
  blocks.append('entry', nil, nil, self) do |fun|
    ret add(fun.params[0], fun.params[1])
  end
end

Now you can run your 'add' function like this:

jit.run(fun, 1, 2)

The result will be a {RLTK::CG::GenericValue} object, and you will want to use its {RLTK::CG::GenericValue#to_i #to_i} and {RLTK::CG::GenericValue#to_f #to_f} methods to get the Ruby value result.

Tutorial

What follows is an in-depth example of how to use the Ruby Language Toolkit. This tutorial will show you how to use RLTK to build a lexer, parser, AST nodes, and compiler to create a toy language called Kazoo. The tutorial is based on the LLVM Kaleidoscope tutorial, but has been modified to:

The Kazoo toy language is a procedural language that allows you to define functions, use conditionals, and perform basic mathematical operations. Over the course of the tutorial we’ll extend Kazoo to support the if/then/else construct, for loops, JIT compilation, and a simple command line interface to the JIT.

Because we want to keep things simple the only datatype in Kazoo is a 64-bit floating point type (a C double or a Ruby float). As such, all values are implicitly double precision and the language doesn’t require type declarations. This gives the language a very nice and simple syntax. For example, the following example computes Fibonacci numbers:

def fib(x)
  if x < 3 then
    1
  else
    fib(x-1) + fib(x-2)

The tutorial is organized as follows:

Before starting this tutorial you should know about regular expressions, the basic ideas behind lexing and parsing, and be able to read context-free grammar (CFG) definitions. By the end of this tutorial we will have written 372 lines of source code and have a JIT compiler for a Turing complete language.

Provided Lexers and Parsers

The following lexer and parser classes are included as part of RLTK:

Contributing

If you are interested in contributing to RLTK there are many aspects of the library that you can work on. A detailed TODO list can be found here. If you are looking for smaller units of work feel free to:

Lastly, I love hearing back from users. If you find any part of the documentation unclear or incomplete let me know. It is also helpful to me to know how people are using the library, so if you are using RLTK in your project send me an email. This lets me know what features are being used and where I should focus my development efforts.

News

Aaaaand we're back. Development of RLTK has been on hold for a while as I worked on other projects. If you want to see what I've been up to, you can check out Clang's new -Wconsumed flag and the Filigree gem.

The next version of RLTK is going to be updated to require Ruby 2.0 as well as LLVM 3.4. Previous versions of RLTK required my LLVM-ECB libarary to expose extra LLVM features through the C bindings; this is no longer necessary as this functionality has been moved into LLVM proper. If anyone has any requests for new or improved features for RLTK version 3.0, let me know.