replace-megaparsec

• Usage Examples
• In the Shell
• Alternatives
• Benchmarks
• Hypothetically Asked Questions

replace-megaparsec is for finding text patterns, and also replacing or splitting on 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.

replace-megaparsec can be used in the same sort of “string splitting” situations in which one would use Python re.split or Perl split.

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. Parse, don’t validate.

• Regular expressions are only able to pattern-match regular grammars. Megaparsec parsers are able pattern-match context-free grammars, and even context-sensitive grammars, 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

These usage examples can be run in a live Jupyter notebook hosted by mybinder.org. Click the badge to launch:

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

The examples depend on these imports.

import Data.Void
import Replace.Megaparsec
import Text.Megaparsec
import Text.Megaparsec.Char
import Text.Megaparsec.Char.Lexer

Split strings with splitCap

Find all pattern matches, capture the matched text and the parsed result

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
splitCap (match hexparser) "0xA 000 0xFFFF"

[Right ("0xA",10), Left " 000 ", Right ("0xFFFF",65535)]

Find all pattern matches, capture only the locations of the matched patterns

Find all of the sections of the stream which are letters. Capture a list of the offsets of the beginning of every pattern match.

import Data.Either
let letterOffset = getOffset <* some letterChar :: Parsec Void String Int
rights $ splitCap letterOffset " a  bc "

[1,4]

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.

import Data.Functor (void)
import Data.Bifunctor (second)
let parens :: Parsec Void String ()
    parens = do
        char '('
        manyTill
            (void (noneOf "()") <|> void parens)
            (char ')')
        pure ()

second fst <$> splitCap (match parens) "(()) (()())"

[Right "(())",Left " ",Right "(()())"]

Edit strings 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 occurances with newline.

let crnl = chunk "\r\n" :: Parsec Void String String
streamEdit crnl (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.

let somelet = some letterChar :: Parsec Void String String
streamEdit somelet (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 with streamEditT

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

import System.Environment (getEnv)
let bracevar = char '{' *> manyTill anySingle (char '}') :: ParsecT Void String IO String
streamEditT bracevar getEnv "- {HOME} -"

"- /home/jbrock -"

Context-sensitive pattern match and edit the matches with streamEditT

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
        let i' = i+1
        put i'
        if i'==3 then pure [x] else empty

flip evalState 0 $ streamEditT capThird (pure . fmap toUpper) "a a a a a"

"a a A a a"

Pattern match, edit the matches, and count the edits with streamEditT

Find and capitalize no more than three letters in a string, and return the edited string along with the number of letters capitalized. To enable the editor function to remember how many letters it has capitalized, we’ll run streamEditT in the State monad from the mtl package. Use this technique to get the same functionality as Python re.subn.

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

let editThree :: Char -> MTL.State Int String
    editThree x = do
        i <- get
        if i<3
            then do
                put $ i+1
                pure [toUpper x]
            else pure [x]

flip runState 0 $ streamEditT letterChar editThree "a a a a a"

("A A A a a",3)

Non-greedy pattern repetition

This is not a feature of this library, but it’s a useful technique to know.

How do we do non-greedy repetition of a pattern p, like we would in Regex by writing p*??

By using the manyTill_ combinator. To repeat pattern p non-greedily, write manyTill_ p q where q is the entire rest of the parser.

For example, this parse fails because many repeats the pattern letterChar greedily.

flip parseMaybe "aab" $ do
  a <- many letterChar
  b <- single 'b'
  pure (a,b)

Nothing

To repeat pattern letterChar non-greedily, use manyTill_.

flip parseMaybe "aab" $ do
  (a,b) <- manyTill_ letterChar $ do
    single 'b'
  pure (a,b)

Just ("aa",'b')

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 lts-16
  --package megaparsec
  --package replace-megaparsec
-}
-- https://docs.haskellstack.org/en/stable/GUIDE/#script-interpreter

import Data.Void
import Text.Megaparsec
import Text.Megaparsec.Char
import Text.Megaparsec.Char.Lexer
import Replace.Megaparsec

main = interact $ streamEdit (decimal :: Parsec Void String Int) (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 3.7.4 re.sub repl function	89.23ms	23.98ms
Perl 5 s///ge	180.65ms	5.02ms
Replace.Megaparsec.streamEdit String	441.94ms	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	> 10min	14.29ms
Control.Lens.Regex.ByteString.match	> 10min	4.27ms
Control.Lens.Regex.Text.match	> 10min	14.74ms

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 for pattern-matching 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 monadic parsers are great for finding small signal patterns in a stream of otherwise noisy text.

   Our 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.