pymetaterp 1.1

A python parser that builds python ASTs in 502 lines of python without using modules

pymetaterp

This is a python AST builder that uses no Python modules. A longer stackless version is available for easier porting. single_file.py is a stand-alone 502 lines script.

Its (also) just another parsing expression grammar (PEG) based parser with one major difference. The parsed grammar is interpreted instead of compiled. This makes it easy to modify the language (by editing its grammar) as well as the language that grammar is written in (and the language of that grammar).

[This is a pre-release of sorts. There are probably some errors and missing information.]

Download and run

git clone https://github.com/asrp/pymetaterp
cd pymetaterp
python single_file.py

or

python single_file.py filename.py

This will print out the AST of the given file (or single_file.py's own AST). Sample beginning of the output:

file_input
  regular_assign
    testlist
      NAME
        str 'NAME'
      NAME
        str 'FLAGS'

To run files from the library

python test/boot_test.py 
python test/python_parse_test.py 

Files

single_file.py is mainly for demonstration. This module is otherwise separated into files. There are many files but they are mostly separate. The import dependencies is

util.py
  boot.py
  boot_stackless.py
    python.py

Other files have no imports. To get something useful, you'll have to import multiple files. See test/python_parse_test.py and test/boot_test.py for some examples.

Repl

An obvious thing missing is the grammar read-eval-print loop (repl) so the interpreter can be fed one rule at a time, parsing subsequence input using the rules seen so far.

Source reading order

I'd suggest reading boot.py and bootstrap in boot_grammar.py first. The two form the core and together with boot_tree.py, they can regenerate boot_tree.

Then boot_stackless is the same as boot.py but doesn't use the Python call stack/recursion for parsing.

python.py adds functionality to the boot.py interpreter. diff in boot_grammar.py adds the syntax for those.

Finally, python_grammar.py contains the python grammar to be finally parsed.

Python language parsed

The module builds the AST for Python 2.x programs. It is able to parse all of Python 2.x (in fact, it contains a slightly modified version of the Python 2.x grammar) but is less lenient with whitespaces. For example, parsing

from my_module import (var1, var2,
                       var3, var4)

gives an error.

Since this is a pre-release, there are likely bugs with parts of the language I don't use so often. It can build the AST for all files included here.

Gramamr language differences

The beginning of boot_grammar.py self-describes the grammar. Its a PEG so all "or" (|) returns the first match and "and" and "quantified" (*, +, ?) are greedy.

name = (letter | '_') (letter | digit | '_')*
expr = apply | exactly | token | parenthesis | output

exactly! = "'" {(escaped_char | ~'\'' anything)*} "'"
token! = "\"" {(escaped_char | ~'"' anything)*} "\""
escaped_char! = '\\' {'n'|'r'|'t'|'b'|'f'|'"'|'\''|'\\'}
apply! = ('\t'|' ')* {name}
parenthesis = "(" {or} ")"
output! = "{" {or} "}"

not = "~" {expr=negation} | expr
quantified = not (('*' | '+' | '?')=quantifier)?
bound = quantified ('=' {name=inline})?
and = bound*
or = and ("|" {and})*

rule = spaces {name=rule_name '!'?=flags and=args ("=" {or})}
grammar = {rule*} spaces

The main difference from other PEG.

• output rule: a {b c} d will match the concatenation of a b c d but only return what matched b c.
• quantifier collapse: letter letter* returns a list rather than a pair with the second element being a list matching letter*.
• nested and collapse: a (b (c d)) e has the same output as a b c d e (see inline below if some pairs need to be explicitly grouped).
• node collapsing: nodes in the output with only one child are replaced by their parent, unless the ! ("don't collapse") flag is set for that node.
• inline: shamelessly taken Ohm but with a slightly different interpretation. expression=name creates a node named name wrapping the output of expression.
• rule replacement: having a second rule_name = expression line replaces the first definition of rule_name (instead of appending into an or).
• two basic tokens: there are two basic token types: 'a' (single quote) and "a" (double quote). The double quoted token allows whitespace before matching.

Regenerating boot_tree.py

Create some tree match_tree using Interpreter.match and call save on the result.

match_tree.save("tree.py")

Left recursion

While "PEG/packrat parsers can support left-recursion", the tree output isn't the one we want. The python functions reformat_binary and reformat_atom fixes a parsed tree's ouput.

Source oddities

Two hard-coded rules

        if root[NAME] == "anything":
            return pop(self.input)
        elif root[NAME] == "void":
            return

Hard-coded semantics for tokens

        if name == "token":
            while pop(self.input) in self.whitespace:
                if self.input[0][self.input[1]] == '\\':
                    pop(self.input)
            self.input[1] -= 1
        if name == "token" and root[0].isalpha():
            top = pop(self.input)
            if top.isalnum() or top == '_':
                raise MatchError("Prefix matched but didn't end.")
            self.input[1] -= 1

Optimization

Some effort were made to make these files short (especially single_file.py) but not too much. There are still some asserts around and commented print statements that can be useful for debugging. The final goal is, of course, to reduce the program's complexity and verbosity, not its line count.

Missing features

Features/bloat from a longer version of this program not (yet?) moved over:

• Debugging tree of nodes visited and their input and output
• Function arguments (its in the grammar but not the interpreter)
• Nested list inputs (its also in the grammar but not the interpreter)
• name, args, flags, body as parameters instead of positional children
• Memoization
• Matched input start and end positions
• Exact python expression matching for predicate, action and rule value. balanced is used as a simpler heuristic for now.

Removing features

To get a smaller file with just the basics.

patch -R pymetaterp/python.py < patches/python_pos.patch
patch -R pymetaterp/python.py < patches/python_memoizer.patch
patch -R pymetaterp/boot_stackless.py < patches/boot_pos.patch
patch -R pymetaterp/boot_stackless.py < patches/boot_memoizer.patch

Readings

• Ometa - Warth's thesis reads very well.
• PEG and packrat parser
• Packrat Parsers Can Support Left Recursion

Other similar projects

• parsimonious
• Pymeta
• pyparsing