Skip to content

benjamin-hodgson/Sawmill

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Sawmill

Simple tools for working with immutable trees, based on Uniform Boilerplate and List Processing and developed at Stack Overflow.

Installing

Sawmill is available on Nuget. API docs are hosted on my website.

Tutorial

Sawmill contains functions which make it easy to work with immutable tree-shaped data such as abstract syntax trees. It factors out the boilerplate associated with recursively traversing a tree, allowing you to write queries and transformations which get straight to the point.

Sawmill is designed to be extremely simple and lightweight (it's built as a set of extension methods for a single simple interface); it works well with modern C# features like lambdas and pattern matching (the days of the clunky old visitor pattern are over!); and it doesn't get in the way when you need to go it alone and write traversals without Sawmill's help.

I've written a step-by-step tutorial on the library's core idea on my blog.

Getting started

For example, suppose you're working with a simple language of arithmetic expressions featuring literal numbers, variables, addition, and unary subtraction. Each syntactic construct corresponds to a subclass of an Expr base type, so an expression like (2 + x) + (-4) would be represented as new Add(new Add(new Lit(2), new Var(x)), new Neg(new Lit(4))).

For your tree type to work with Sawmill, it must implement the IRewritable<T> interface. An object is rewritable if it knows how to access its collection of immediate children; accordingly, IRewritable<T> contains GetChildren and SetChildren methods. Implementations of IRewritable should ensure that the rewritable type conforms to the following two-point specification:

  • You get out what you put in - x.SetChildren(children).GetChildren() == children
  • Setting twice is the same as setting once - x.SetChildren(children1).SetChildren(children2) == x.SetChildren(children2)

See below for a full example of how to implement the Expr type outlined above. You can also use the supplied AutoRewriter or RewriterBuilder classes to assist in implementing IRewritable.

Querying a tree

Here's a function to extract a list of the variables mentioned in a given Expr (for example, a compiler writer might want to find the variables captured by a lambda expression):

IEnumerable<string> GetVariables(Expr expr)
{
    switch (expr)
    {
        case Lit l:
            return Enumerable.Empty<string>();
        case Var v:
            return new[] { v.Name };
        case Add a:
            return GetVariables(a.Left).Concat(GetVariables(a.Right));
        case Neg n:
            return GetVariables(n.Operand);
    }
    throw new ArgumentOutOfRangeException(nameof(expr));
}

The only interesting line of code here is the Var case. The rest is just boilerplate to recursively call GetVariables on the nodes' children and combine the results. With Sawmill, GetVariables is one line of simple, direct code:

IEnumerable<string> GetVariables(Expr expr)
    => expr.SelfAndDescendants().OfType<Var>().Select(v => v.Name);

SelfAndDescendants returns an enumerable containing the current node and all of the nodes in the rest of the tree. The example above uses the standard OfType and Select LINQ methods to find the names of all the variables mentioned in expr. It also has two cousins, DescendantsAndSelf and SelfAndDescendantsBreadthFirst, which differ in the order in which they yield nodes.

By the way, you can totally "go it alone" and write complex or performance-critical traversals without Sawmill's help. You can just use explicit recursion, as in the first example. Sawmill is intended to be useful, not opinionated.

Transforming a tree

Sawmill makes tree transformations easier too. Here's an example of a simple optimisation pass, which removes double-negatives:

Expr RemoveDoubleNegation(Expr expr)
{
    switch (expr)
    {
        case Neg n1 when n.Operand is Neg n2:
            return RemoveDoubleNegation(n2);
        case Neg n:
            return new Neg(RemoveDoubleNegation(n.Operand));
        case Lit l:
            return l;
        case Var v:
            return v;
        case Add a:
            return new Add(RemoveDoubleNegation(a.Left), RemoveDoubleNegation(a.Right));
    }
    throw new ArgumentOutOfRangeException(nameof(expr));
}

Once again, Sawmill tackles the boilerplate - recursively taking each node apart and putting them back together - so you can focus on the important part of your operation.

Expr RemoveDoubleNegation(Expr expr)
    => expr.Rewrite(node =>
        node is Neg n1 && n1.Operand is Neg n2
            ? n2.Operand
            : node
    );

Rewrite takes a transformation function and rebuilds a tree by applying the function to every node in the tree. For example, given a representation of the expression (2 + x) + (-4) and a transformer function, expr.Rewrite(transformer) is equivalent to:

transformer(new Add(
    transformer(new Add(
        transformer(new Lit(2)),
        transformer(new Var("x"))
    )),
    transformer(new Neg(
        transformer(new Lit(4))
    ))
))

So the transformation function gets applied to every node in the tree exactly once. Rewrite is a mapping operation, like LINQ's Select.

Sawmill takes care to avoid rebuilding parts of the tree which the transformation function leaves unchanged, so Rewrite will typically be more efficient than a naïve handwritten implementation.

Sawmill also contains tools for some more niche operations:

Putting an expression into a normal form

Normalising an expression typically involves repeatedly applying a set of rewrite rules until they can't be applied any more.

For example, to put an arithmetic expression into negation normal form, so that all of the minus signs appear only next to variables or literal numbers, you distribute - over + (so -(3+2) becomes (-3)+(-2)). Since performing this distribution might produce more places where the result of an addition is negated (consider -((1+2)+3) -> (-(1+2))+(-3)), you need to do so repeatedly until you can't do it any more.

RewriteIter packages up this pattern. It applies a transformation function to every node in the tree from bottom to top, repeating this until the function is a no-op for each node in the tree. (In other words, x.RewriteIter(f).DescendantsAndSelf().All(n => f(n) == n) == true.)

Expr ToNegationNormalForm(Expr expr)
    => expr.RewriteIter(node => 
        node is Neg n && n.Operand is Add a
            ? new Add(new Neg(a.Left), new Neg(a.Right))
            : n
    );

It'd be pretty tedious to write this operation by hand as a recursive function!

Reducing a tree to a value

LINQ has the Aggregate method, which passes an accumulator value along an enumerable, using an aggregation function to combine elements. But while an element of an enumerable has only one predecessor, a node of a tree may have many children. So Sawmill's Fold method passes multiple accumulator values up a tree, using an aggregation function to flatten them into a single value.

Here's an example of compiling our expression tree into code for a hypothetical stack machine.

string Compile(Expr expr)
    => expr.Fold<Expr, string>(
        (n, children) => n switch
        {
            Lit l => "PUSH " + l.Value + ";",
            Var v => "LOAD " + v.Name + ";",
            Add a => children[0] + children[1] + "ADD;",
            Neg n => children[0] + "NEGATE;",
            _ => throw new ArgumentOutOfRangeException(nameof(n))
        }
    );

Replacing individual nodes in a tree

There are several "InContext" extension methods:

  • ChildrenInContext
  • SelfAndDescendantsInContext
  • DescendantsAndSelfInContext
  • SelfAndDescendantsInContextBreadthFirst

These all have a return type of IEnumerable<(T item, Func<T, T> replace)>: a list of tuples containing a node and a function to build a new tree with a different node in its place. This might be useful in mutation testing, where you want to see all the ways you can change a Lit node in a tree. You can think of the function as representing the node's context in the tree; calling the function with a new node "plugs the hole" that was created by removing the node from the tree.

Inspecting and replacing a node and its neighbours

The Cursor() method generalises the InContext methods by returning a Cursor<T> - a mutable builder object representing a focus on a particular node in a tree. You can efficiently replace the currently focused node by setting the cursor's Focus property. The Up, Down, Left and Right methods allow you to efficiently move the cursor's focus to the current node's parent, the current node's first child, and the current node's next and previous siblings, respectively. This is useful if you need to make a complex sequence of edits to a particular area in a tree, such as if a user is editing part of a text file. Moving the cursor all the way back to the Top rebuilds the whole tree with the new nodes in place of the old ones.

Implementing IRewritable<T>

To use Sawmill with your own expression types, you implement the IRewritable interface. You explain how to read and write the nodes' immediate children, and Sawmill does the boring work of recursively traversing the children's children.

abstract class Expr : IRewritable<Expr>
{
    public abstract int CountChildren();
    public abstract void GetChildren(Span<Expr> span);
    public abstract Expr SetChildren(ReadOnlySpan<Expr> newChildren);
}
// literal numbers are leaf nodes; they have no children
class Lit : Expr
{
    public int Value { get; }
    
    public Lit(int value)
    {
        Value = value;
    }

    public override int CountChildren() => 0;
    public override void GetChildren(Span<Expr> span)
    {
    }
    public override Expr SetChildren(ReadOnlySpan<Expr> newChildren)
        => this;
}
// variables also have no children
class Var : Expr
{
    public string Name { get; }
    
    public Var(string name)
    {
        Name = name;
    }

    public override int CountChildren() => 0;
    public override void GetChildren(Span<Expr> span)
    {
    }
    public override Expr SetChildren(ReadOnlySpan<Expr> newChildren)
        => this;
}
class Neg : Expr
{
    public Expr Operand { get; }
    
    public Neg(Expr operand)
    {
        Operand = operand;
    }

    public override int CountChildren() => 1;
    public override void GetChildren(Span<Expr> span)
    {
        span[0] = Operand;
    }
    public override Expr SetChildren(ReadOnlySpan<Expr> newChildren)
        => new Neg(newChildren[0]);
}
class Add : Expr
{
    public Expr Left { get; }
    public Expr Right { get; }
    
    public Add(Expr left, Expr right)
    {
        Left = left;
        Right = right;
    }

    public override int CountChildren() => 2;
    public override void GetChildren(Span<Expr> span)
    {
        span[0] = Left;
        span[1] = Right;
    }
    public override Expr SetChildren(ReadOnlySpan<Expr> newChildren)
        => new Add(newChildren[0], newChildren[1]);
}

If you can't implement IRewritable

There's also an IRewriter<T> interface, which is useful if you can't change the tree type to implement IRewritable. Sawmill comes bundled with IRewriter implementations (and extension methods) for some tree-shaped objects in the BCL, namely Expression, XElement, and XmlNode. In the box you'll also find RewriterBuilder, which is a domain-specific language for buiding IRewriter implementations, and an experimental reflection-based AutoRewriter.