replace-megaparsec

Find, replace, and edit text patterns with Megaparsec parsers

https://github.com/jamesdbrock/replace-megaparsec

Version on this page:1.2.1.0
LTS Haskell 22.14:1.5.0.1
Stackage Nightly 2024-03-28:1.5.0.1
Latest on Hackage:1.5.0.1

See all snapshots replace-megaparsec appears in

BSD-2-Clause licensed and maintained by James Brock
This version can be pinned in stack with:replace-megaparsec-1.2.1.0@sha256:ce54a2b349f96513341ef26012a8ecbb7523d91e2c872fcc681704fb8038d69b,2389

Module documentation for 1.2.1.0

replace-megaparsec

Hackage Stackage Nightly Stackage LTS

replace-megaparsec is for finding text patterns, and also editing and replacing the found patterns. This activity is traditionally done with regular expressions, but replace-megaparsec uses megaparsec parsers instead for the pattern matching.

replace-megaparsec can be used in the same sort of “pattern capture” or “find all” situations in which one would use Python re.findall or Perl m//, or Unix grep.

replace-megaparsec can be used in the same sort of “stream editing” or “search-and-replace” situations in which one would use Python re.sub, or Perl s///, or Unix sed, or awk.

See replace-attoparsec for the attoparsec version.

Why would we want to do pattern matching and substitution with parsers instead of regular expressions?

  • Haskell parsers have a nicer syntax than regular expressions, which are notoriously difficult to read.

  • Regular expressions can do “group capture” on sections of the matched pattern, but they can only return stringy lists of the capture groups. Parsers can construct typed data structures based on the capture groups, guaranteeing no disagreement between the pattern rules and the rules that we’re using to build data structures based on the pattern matches.

    For example, consider scanning a string for numbers. A lot of different things can look like a number, and can have leading plus or minus signs, or be in scientific notation, or have commas, or whatever. If we try to parse all of the numbers out of a string using regular expressions, then we have to make sure that the regular expression and the string-to-number conversion function agree about exactly what is and what isn’t a numeric string. We can get into an awkward situation in which the regular expression says it has found a numeric string but the string-to-number conversion function fails. A typed parser will perform both the pattern match and the conversion, so it will never be in that situation.

  • Regular expressions are only able to pattern-match regular grammers. Megaparsec parsers are able pattern-match context-free grammers, and even context-sensitive grammers, if needed. See below for an example of lifting a Parser into a State monad for context-sensitive pattern-matching.

  • The replacement expression for a traditional regular expression-based substitution command is usually just a string template in which the Nth “capture group” can be inserted with the syntax \N. With this library, instead of a template, we get an editor function which can perform any computation, including IO.

Usage Examples

Try the examples in ghci by running cabal v2-repl in the replace-megaparsec/ root directory.

The examples depend on these imports.

import Replace.Megaparsec
import Text.Megaparsec
import Text.Megaparsec.Char
import Text.Megaparsec.Char.Lexer

Parsing with sepCap family of parser combinators

The following examples show how to match a pattern to a string of text and separate it into sections which match the pattern, and sections which don’t match.

Pattern match, capture only the parsed result with sepCap

Separate the input string into sections which can be parsed as a hexadecimal number with a prefix "0x", and sections which can’t. Parse the numbers.

let hexparser = chunk "0x" >> hexadecimal :: Parsec Void String Integer
parseTest (sepCap hexparser) "0xA 000 0xFFFF"
[Right 10,Left " 000 ",Right 65535]

Pattern match, capture only the matched text with findAll

Just get the strings sections which match the hexadecimal parser, throw away the parsed number.

let hexparser = chunk "0x" >> hexadecimal :: Parsec Void String Integer
parseTest (findAll hexparser) "0xA 000 0xFFFF"
[Right "0xA",Left " 000 ",Right "0xFFFF"]

Pattern match, capture the matched text and the parsed result with findAllCap

Capture the parsed hexadecimal number, as well as the string section which parses as a hexadecimal number.

let hexparser = chunk "0x" >> hexadecimal :: Parsec Void String Integer
parseTest (findAllCap hexparser) "0xA 000 0xFFFF"
[Right ("0xA",10),Left " 000 ",Right ("0xFFFF",65535)]

Pattern match, capture only the locations of the matched patterns

Find all of the sections of the stream which match a string of spaces. Print a list of the offsets of the beginning of every pattern match.

import Data.Either
let spaceoffset = getOffset <* space1 :: Parsec Void String Int
parseTest (return . rights =<< sepCap spaceoffset) " a  b  "
[0,2,5]

Pattern match balanced parentheses

Find groups of balanced nested parentheses. This is an example of a “context-free” grammar, a pattern that can’t be expressed by a regular expression. We can express the pattern with a recursive parser.

let parens :: Parsec Void String ()
    parens = do
        char '('
        manyTill
            (void (noneOf "()") <|> void parens)
            (char ')')
        return ()

parseTest (findAll parens) "(()) (()())"
[Right "(())",Left " ",Right "(()())"]

Edit text strings by running parsers with streamEdit

The following examples show how to search for a pattern in a string of text and then edit the string of text to substitute in some replacement text for the matched patterns.

Pattern match and replace with a constant

Replace all carriage-return-newline instances with newline.

streamEdit (chunk "\r\n") (const "\n") "1\r\n2\r\n"
"1\n2\n"

Pattern match and edit the matches

Replace alphabetic characters with the next character in the alphabet.

streamEdit (some letterChar) (fmap succ) "HAL 9000"
"IBM 9000"

Pattern match and maybe edit the matches, or maybe leave them alone

Find all of the string sections s which can be parsed as a hexadecimal number r, and if r≤16, then replace s with a decimal number. Uses the match combinator.

let hexparser = chunk "0x" >> hexadecimal :: Parsec Void String Integer
streamEdit (match hexparser) (\(s,r) -> if r<=16 then show r else s) "0xA 000 0xFFFF"
"10 000 0xFFFF"

Pattern match and edit the matches with IO

Find an environment variable in curly braces and replace it with its value from the environment.

import System.Environment
streamEditT (char '{' *> manyTill anySingle (char '}')) getEnv "- {HOME} -"
"- /home/jbrock -"

Context-sensitive pattern match and edit the matches

Capitalize the third letter in a string. The capthird parser searches for individual letters, and it needs to remember how many times it has run so that it can match successfully only on the third time that it finds a letter. To enable the parser to remember how many times it has run, we’ll compose the parser with a State monad from the mtl package. (Run in ghci with cabal v2-repl -b mtl). Because it has stateful memory, this parser is an example of a “context-sensitive” grammar.

import qualified Control.Monad.State.Strict as MTL
import Control.Monad.State.Strict (get, put, evalState)
import Data.Char (toUpper)

let capthird :: ParsecT Void String (MTL.State Int) String
    capthird = do
        x <- letterChar
        i <- get
        put (i+1)
        if i==3 then return [x] else empty

flip evalState 1 $ streamEditT capthird (return . fmap toUpper) "a a a a a"
"a a A a a"

In the Shell

If we’re going to have a viable sed replacement then we want to be able to use it easily from the command line. This Stack script interpreter script will find decimal numbers in a stream and replace them with their double.

#!/usr/bin/env stack
{- stack
  script
  --resolver nightly-2019-09-13
  --package megaparsec
  --package replace-megaparsec
-}
-- https://docs.haskellstack.org/en/stable/GUIDE/#script-interpreter

import Text.Megaparsec
import Text.Megaparsec.Char
import Text.Megaparsec.Char.Lexer
import Replace.Megaparsec

main = interact $ streamEdit decimal (show . (*2))

If you have The Haskell Tool Stack installed then you can just copy-paste this into a file named doubler.hs and run it. (On the first run Stack may need to download the dependencies.)

$ chmod u+x doubler.hs
$ echo "1 6 21 107" | ./doubler.hs
2 12 42 214

Alternatives

Some libraries that one might consider instead of this one.

http://hackage.haskell.org/package/regex-applicative

http://hackage.haskell.org/package/pcre-heavy

http://hackage.haskell.org/package/lens-regex-pcre

http://hackage.haskell.org/package/regex

http://hackage.haskell.org/package/pipes-parse

http://hackage.haskell.org/package/stringsearch

http://hackage.haskell.org/package/substring-parser

http://hackage.haskell.org/package/pcre-utils

http://hackage.haskell.org/package/template

Benchmarks

These benchmarks are intended to measure the wall-clock speed of everything except the actual pattern-matching. Speed of the pattern-matching is the responsibility of the megaparsec and attoparsec libraries.

The benchmark task is to find all of the one-character patterns x in a text stream and replace them by a function which returns the constant string oo. So, like the regex s/x/oo/g.

We have two benchmark input cases, which we call dense and sparse.

The dense case is one megabyte of alternating spaces and xs like

x x x x x x x x x x x x x x x x x x x x x x x x x x x x

The sparse case is one megabyte of spaces with a single x in the middle like

                         x

Each benchmark program reads the input from stdin, replaces x with oo, and writes the result to stdout. The time elapsed is measured by perf stat, and the best observed time is recorded.

See replace-benchmark for details.

Program dense sparse
Python re.sub¹ 89.23ms 23.98ms
Perl s///ge² 180.65ms 5.60ms
Replace.Megaparsec.streamEdit String 454.95ms 375.04ms
Replace.Megaparsec.streamEdit ByteString 529.99ms 73.76ms
Replace.Megaparsec.streamEdit Text 547.47ms 139.21ms
Replace.Attoparsec.ByteString.streamEdit 394.12ms 41.13ms
Replace.Attoparsec.Text.streamEdit 515.26ms 46.10ms
Text.Regex.Applicative.replace String 1083.98ms 646.40ms
Text.Regex.PCRE.Heavy.gsub Text ⊥³ 14.76ms

¹ Python 3.7.4

² This is perl 5, version 28, subversion 2 (v5.28.2) built for x86_64-linux-thread-multi

³ Does not finish.

Hypothetically Asked Questions

  1. Could we write this library for parsec?

    No, because the match combinator doesn’t exist for parsec. (I can’t find it anywhere. Can it be written?)

  2. Is this a good idea?

    You may have heard it suggested that monadic parsers are better when the input stream is mostly signal, and regular expressions are better when the input stream is mostly noise.

    The premise of this library is: that sentiment is outdated; monadic parsers are great for finding small patterns in a stream of otherwise uninteresting text; and the reluctance to forego the speedup opportunities afforded by restricting ourselves to regular grammars is an old superstition about opportunities which remain mostly unexploited anyway. The performance compromise of allowing stack memory allocation (a.k.a pushdown automata, a.k.a context-free grammar) was once considered controversial for general-purpose programming languages. I think we can now resolve that controversy the same way for pattern matching languages.

Changes

Revision history for replace-megaparsec

1.2.0.0 – 2019-10-31

Benchmark improvements

Specializations of the sepCap function, guided by replace-benchmark.

New benchmarks

Program dense sparse
Replace.Megaparsec.streamEdit String 454.95ms 375.04ms
Replace.Megaparsec.streamEdit ByteString 529.99ms 73.76ms
Replace.Megaparsec.streamEdit Text 547.47ms 139.21ms

Old benchmarks

Program dense sparse
Replace.Megaparsec.streamEdit String 454.95ms 375.04ms
Replace.Megaparsec.streamEdit ByteString 611.98ms 433.26ms
Replace.Megaparsec.streamEdit Text 592.66ms 353.32ms

1.1.5.0 – 2019-10-08

1.1.0.0 – 2019-09-01

  • Add benchmark suite.
  • In streamEditT, replace fold with mconcat. The benchmarks now show linear scaling instead of quadratic.

1.0.1.0 – 2019-08-28

  • Add test suite.
  • sepCap will treats sep as failing if it succeeds but consumes no input.

1.0.0.0 – 2019-08-24

  • First version.