In tasks such as parsing there is often a need to construct a result representation of some kind, e.g. a parse tree. This system is concerned with flexible construction and processing of different result representations while avoiding coupling between producers and consumers of such results.
Staying with the parsing example, the result of a successful parse is some sort of (abstract) syntax tree (AST). Most parsing code in Common Lisp seems to do this in one of two ways: nested list structures or a tree of (class or structure) instances. Both approaches have advantages and disadvantages
- On the one hand, list-based parse results are well suited for
debugging since they pretty print nicely and unit tests since they
are
equal
comparable. - On the other hand list-based results are not suitable for CLOS-dispatch while instances are.
- Both kinds of results are well suited for AST processing using pattern matching (e.g. with optima).
In practice, much parsing code seems to be written for one particular consumer of the produced AST. This fact usually seems to inform the choice of result representation.
This system employs the “builder” design pattern to enable a flexible result representation with little effort for consumers and producers. A “builder protocol” is concerned with the construction of results while a “un-builder protocol” is concerned with destructuring and traversing the constructed representations.
Since this is a probably a common case, we will use the construction of a simplistic AST from the output of an equally simplistic parser as an example.
The example code in the following sections can be loaded into the
cl-user
package and assumes that the alexandria
system is
loaded.
The nodes of the AST we want to construct are either literals or operator applications with two operands and are both expressions:
(defclass expression () ())
(defclass literal (expression)
((%value :initarg :value :reader literal-value)))
(defclass operator (expression)
((%operands :accessor operator-operands :initform '())))
Note that the value
slot of the literal
is initialized using
the :value
initarg while the operands
slot of the operator
class is initialized to the empty lists but allows for later
mutation via (setf operator-operands)
. The rationale is that
literal
instances can be constructed in one make-instance
call
while operator
instance may be constructed before their operand
nodes, thus requiring mutation to attach these operand nodes once
they have been constructed.
A simple implementation of the builder protocol for these nodes looks like this:
(defclass ast-builder () ())
(defmethod architecture.builder-protocol:make-node
((builder ast-builder)
(kind (eql :literal))
&key value)
(make-instance 'literal :value value))
(defmethod architecture.builder-protocol:make-node
((builder ast-builder)
(kind (eql :operator))
&key)
(make-instance 'operator))
(defmethod architecture.builder-protocol:relate
((builder ast-builder)
(relation (eql :operand))
(left operator)
(right expression)
&key)
(alexandria:appendf (operator-operands left) (list right))
left)
We can already use this builder without a parser:
(let* ((builder (make-instance 'ast-builder))
(operands (list (architecture.builder-protocol:make+finish-node
builder :literal :value 5)
(architecture.builder-protocol:make+finish-node
builder :literal :value 6)))
(operator (architecture.builder-protocol:make-node builder :operator)))
(architecture.builder-protocol:finish-node
builder :operator
(reduce (lambda (l r)
(architecture.builder-protocol:relate
builder :operand l r))
operands :initial-value operator)))
#<OPERATOR {100E5961}>
The following is a more compact (but equivalent behind the scenes) spelling of the above code:
(architecture.builder-protocol:with-builder ((make-instance 'ast-builder))
(architecture.builder-protocol:node* (:operator)
(* :operand (list (architecture.builder-protocol:node* (:literal :value 5))
(architecture.builder-protocol:node* (:literal :value 6))))))
#<OPERATOR {1019F0E013}>
We will use a parser for a very simple expressions in polish notation:
EXPRESSION ::= OPERATOR | LITERAL LITERAL ::= '0' | '1' | '2' | '3' | '4' | '5' | '6' | '7' | '8' | '9' OPERATOR ::= '+' EXPRESSION EXPRESSION
The parser is straightforward: it has a local function for each
element of the grammar and uses the builder protocol like in the
previous example. Since we now parse an actual source text, source
locations of constructed result nodes can be recorded using the
:bounds
initarg. Note that the ast-builder
we defined in the
previous section receives the :bounds
initarg in make-node
calls, but does not store it anywhere. If storing source locations
in AST nodes was desired, a %source
slot could be added to the
expression
class and the value of the :bounds
keyword argument
could be passed to make-instance
as the :source
initarg.
(defun parse (stream builder)
(labels ((expression ()
(let ((c (peek-char nil stream)))
(cond ((char= c #\+)
(operator))
((digit-char-p c)
(literal)))))
(literal ()
(let ((start (stream-file-position stream))
(c (read-char stream)))
(architecture.builder-protocol:make-node
builder :literal
:value (parse-integer (string c))
:bounds (cons start (1+ start)))))
(operator ()
(let ((start (stream-file-position stream))
(c (read-char stream))
(operands (list (expression) (expression)))
(end (stream-file-position stream)))
(declare (ignore c))
(architecture.builder-protocol:finish-node
builder :operator
(reduce (lambda (l r)
(architecture.builder-protocol:relate
builder :operator-operand l r))
operands
:initial-value (architecture.builder-protocol:make-node
builder :operator
:bounds (cons start end)))))))
(expression)))
As before, the various builder method calls can be written
compactly using the node
macro:
(defun parse2 (stream builder)
(labels ((expression ()
(let ((c (peek-char nil stream)))
(cond ((char= c #\+)
(operator))
((digit-char-p c)
(literal)))))
(literal ()
(let ((start (stream-file-position stream))
(c (read-char stream)))
(architecture.builder-protocol:node
(builder :literal :value (parse-integer (string c))
:bounds (cons start (1+ start))))))
(operator ()
(let ((start (stream-file-position stream))
(c (read-char stream))
(operands (list (expression) (expression)))
(end (stream-file-position stream)))
(declare (ignore c))
(architecture.builder-protocol:node
(builder :operator :bounds (cons start end))
(* :operand operands)))))
(expression)))
The with-builder
macro allows writing the node
macro calls
without supplying the builder
argument:
(architecture.builder-protocol:with-builder (BUILDER)
(architecture.builder-protocol:node* (:KIND :INITARG …)
(* :RELATION …)))
When developing or testing result producers like parsers, it can be
convenient to produce a list-based result since it pretty-prints
nicely without any extra effort and can be equal
-compared in unit
tests without depending on a more heavyweight representation such
as instances of AST node classes.
For these cases, the architecture.builder-protocol
system
provides a builtin list
builder:
(parse (make-string-input-stream "++123") 'list)
(:OPERATOR
(:OPERATOR-OPERAND
(((:OPERATOR
(:OPERATOR-OPERAND
(((:LITERAL NIL :VALUE 1 :BOUNDS (2 . 3)))
((:LITERAL NIL :VALUE 2 :BOUNDS (3 . 4)))))
:BOUNDS (1 . 4)))
((:LITERAL NIL :VALUE 3 :BOUNDS (4 . 5)))))
:BOUNDS (0 . 5))
This may be slightly off-topic, but a nice hack for printing
arbitrary results produced by the list
builder can be done
using the =utilities.print-tree= system:
(defun my-print-tree (tree &optional (stream *standard-output*))
(utilities.print-tree:print-tree
stream tree
(utilities.print-tree:make-node-printer
(lambda (stream depth node)
(declare (ignore depth))
(destructuring-bind (kind relations &rest slots) node
(declare (ignore relations))
(format stream "~A~@[ @~A~]"
kind (getf slots :bounds))
(alexandria:remove-from-plist slots :bounds)))
(lambda (stream depth node)
(declare (ignore depth))
(destructuring-bind (kind relations &rest slots) node
(declare (ignore kind relations))
(format stream "~{~A: ~A~^~@:_~}"
(alexandria:remove-from-plist slots :bounds))))
(lambda (node)
(loop :for (relation nodes) :on (second node) :by #'cddr
:appending (mapcar #'car nodes))))))
Putting these pieces together, we can achieve the following:
(my-print-tree (parse (make-string-input-stream "++123") 'list))
OPERATOR @(0 . 5) ├─OPERATOR @(1 . 4) │ ├─LITERAL @(2 . 3) │ │ VALUE: 1 │ └─LITERAL @(3 . 4) │ VALUE: 2 └─LITERAL @(4 . 5) VALUE: 3
The system architecture.builder-protocol.print-tree
implements a
more complete version (not restricted to the list
builder, among
other things) of this idea:
(defun print-tree (tree)
(fresh-line)
(architecture.builder-protocol.print-tree:print-tree
'list tree *standard-output*))
(print-tree (parse (make-string-input-stream "++123") 'list))
OPERATOR @(0 . 5) ├─OPERATOR-OPERAND: OPERATOR @(1 . 4) │ ├─OPERATOR-OPERAND: LITERAL @(2 . 3) │ │ VALUE: 1 │ └─OPERATOR-OPERAND: LITERAL @(3 . 4) │ VALUE: 2 └─OPERATOR-OPERAND: LITERAL @(4 . 5) VALUE: 3
The generic function walk-nodes
can be used to traverse trees of
nodes built using the build protocol. It uses the “un-build”
protocol and can thus handle arbitrary tree representations.
prepare BUILDER Prepare BUILDER for result construction, return a builder. The default method just returns BUILDER.
finish BUILDER VALUES Finalize and return VALUES produced by BUILDER as multiple values. VALUES is a list of objects that should be returned as multiple values and constitute the overall result of an object tree construction with BUILDER. The first element of VALUES which becomes the first return value is the constructed tree itself (which often coincides with the root node). Additional values are optional and their presence and meaning depend on BUILDER. The default method just returns VALUES.
wrap BUILDER THUNK Call THUNK with an appropriate dynamic environment for BUILDER. A method on this generic function could, for example, bind special variables around the construction of a result object tree. The existing default methods do not specialize the BUILDER parameter and specialize the THUNK parameter to `cl:function' and `cl:symbol'. These default methods just call THUNK.
make-node BUILDER KIND &REST INITARGS &KEY &ALLOW-OTHER-KEYS Use BUILDER to make a result tree node of kind KIND and return it. As a convention, when supplied, the value of the :bounds keyword argument is of the form (START . END) and can be used to indicate the input range for which the tree is constructed.
finish-node BUILDER KIND NODE Use BUILDER to perform finalization for NODE. Return the modified NODE or an appropriate newly created object.
relate BUILDER RELATION LEFT RIGHT &REST ARGS &KEY &ALLOW-OTHER-KEYS Establish RELATION between nodes LEFT and RIGHT and return the resulting modified LEFT node (or an appropriate newly created object). ARGS can be used to supply additional information about the relation that is available from neither LEFT nor RIGHT. In a typical case, RELATION could be :child, LEFT being the parent node and RIGHT being the child node.
add-relations BUILDER NODE RELATIONS Use BUILDER to add relations according to RELATIONS to NODE. RELATIONS is a list of relation specifications of the form (CARDINALITY RELATION-NAME RIGHT &rest ARGS) which are translated into `relate' calls in which NODE is the "left" argument to `relate'. CARDINALITY has to be of type `relation-cardinality' and is interpreted as follows: ? RIGHT is a single node or `nil'. If RIGHT is `nil', `relate' is not called. 1 RIGHT is a single node. * RIGHT is a (possibly empty) sequence of nodes. ARGS must be of the form :KEY₁ (VALUE₁₁ VALUE₁₂ …) :KEY₂ (VALUE₂₁ VALUE₂₂ …) … . The `relate' call for the k-th element of RIGHT will receive the keyword arguments :KEY₁ VALUEₖ₁ :KEY₂ VALUEₖ₂ …. If the value list for a given key would be a repetition of a particular value VALUE, the circular list #1=(VALUE . #1#) may be used as a replacement for that value list. (:map . KEY) RIGHT is a (possible empty) sequence of nodes that should be "zipped" with a sequence of keys (see below) to form a set of key-values pair and thus a map. The sequence of keys is the value of the property whose indicator is KEY in the ARGS plist. The two sequences must be of the same length. Elements at corresponding positions will be paired in a "zipping" operation as described above. RELATION-NAME does not have to be unique across the elements of RELATIONS. This allows multiple "right" nodes to be related to NODE via a given RELATION-NAME with CARDINALITY * in multiple RELATIONS entries, potentially with different ARGS. The modified NODE or a new node is returned.
make+finish-node BUILDER KIND &REST INITARGS &KEY &ALLOW-OTHER-KEYS Convenience function for constructing and immediately finishing a node.
make+finish-node+relations BUILDER KIND INITARGS RELATIONS Use BUILDER to create a KIND, INITARGS node, relate it via RELATIONS. RELATIONS is processed as described for `add-relations'. `finish-node' is called on the created node. The created node is returned.
node-kind BUILDER NODE Return the kind of NODE w.r.t. BUILDER. The return value is EQ to the KIND argument used to create NODE with BUILDER.
node-initargs BUILDER NODE Return a plist of initargs for NODE w.r.t. BUILDER. The returned list is EQUAL to the list of keyword arguments pass to the MAKE-NODE call that, using BUILDER, constructed NODE.
node-relations BUILDER NODE Return a list of relations of NODE w.r.t. BUILDER. Each relation is of one of the forms RELATION-NAME (RELATION-NAME . CARDINALITY) where RELATION-NAME names the relation and CARDINALITY is of type `relation-cardinality'. When the first form is used, i.e. CARDINALITY is not present, it is assumed to be `*'. CARDINALITY values are interpreted as follows: ? The relation designated by RELATION-NAME with NODE as the "left" node has zero or one "right" nodes. 1 The relation designated by RELATION-NAME with NODE as the "left" node has exactly one "right" node. * The relation designated by RELATION-NAME with NODE as the "left" node has zero or more "right" nodes. (:map . KEY) The relation designated by RELATION-NAME with NODE as the "left" node has zero or more "right" nodes with the additional constraint that the relation parameters for each such node must contain a unique value for the key KEY. . This cardinality information is reflected by the return values of (node-relation BUILDER RELATION-NAME NODE).
node-relation BUILDER RELATION NODE Return two values: 1) a single node or a sequence of nodes related to NODE via RELATION w.r.t. BUILDER 2) `nil' or a same-length sequence of arguments of the relations. RELATION must be of one of the forms RELATION-NAME (RELATION-NAME . CARDINALITY) where RELATION-NAME names the relation and CARDINALITY is of type `relation-cardinality'. The second form is accepted for convenience so that, for example, relation descriptions returned by `node-relations' can be used as arguments to this function. CARDINALITY is not processed by this function except that a `type-error' may be signaled if CARDINALITY is not of type `relation-cardinality'. If the cardinality of RELATION is 1 or `?', the first return value is a single node. Otherwise the first return value is a sequence of nodes. Again, note that the cardinality of RELATION here refers to the actual cardinality as known by BUILDER, not information encoded in RELATION by the caller supplying RELATION as (RELATION-NAME . CARDINALITY). Each element in the sequence of relation arguments is EQUAL to the list of arguments passed to the RELATE call that, using BUILDER, established the relation between NODE and the related node.
walk-nodes BUILDER FUNCTION ROOT Call FUNCTION on nodes of the tree ROOT constructed by BUILDER. Return whatever FUNCTION returns when called for ROOT. The lambda-list of FUNCTION must be compatible to (recurse relation relation-args node kind relations &rest initargs) where RELATION and RELATION-ARGS are the relation and its arguments connecting NODE to the previously visited node, NODE is the node currently being visited, KIND is the kind returned by `node-kind' for BUILDER and NODE. RELATIONS are the relations returned by `node-relations' for BUILDER and NODE. INITARGS are the initargs returned by `node-initargs' for BUILDER and NODE. RECURSE is a function with the lambda-list (&key relations function) that can be called, optionally with a list of relations, to traverse the nodes related to NODE by that relation. If a list of relations is not supplied via the :relations keyword parameter, all relations are traversed. The :function keyword parameter allows performing the traversal with a different function instead of FUNCTION. Calls of this function return a list of elements each of which is the result for the corresponding element of RELATIONS. The result for a relation is either the return value of FUNCTION if the cardinality of the relation is 1 or ? or a list of such return values if the cardinality is * or :map. If FUNCTION is an instance of `peeking', call the "peeking" function stored in FUNCTION before the ordinary walk function (also stored in FUNCTION) is called. The lambda-list of the "peeking" function must be compatible to (builder relation relation-args node) (i.e. it does not receive kind, initargs or relations). This function can control whether NODE should be processed normally, replaced with something else, processed with a different builder or ignored: Its return values are interpreted as follows: NIL Forego processing of NODE, in particular do not call `node-kind', `node-relations', `node-initargs' or the walk function for NODE. T [* * * BUILDER] Continue processing as if there was no "peeking" function. If non-NIL, BUILDER specifies a builder that should be used instead of the current builder to process the NODE and its ancestors. INSTEAD KIND INITARGS RELATIONS [BUILDER] Continue processing as if NODE had been replaced by INSTEAD and builder had returned KIND, INITARGS and RELATIONS. In particular do not call `node-kind', `node-relations', `node-initargs' for NODE. If non-NIL, BUILDER specifies a builder that should be used instead of the current builder to process INSTEAD and its ancestors. Depending on FUNCTION, potentially return a list-of-lists of the same shape as the traversed tree containing return values of FUNCTION.