Fork me on GitHub

FastParse 2.0.4


Fast to write, Fast running Parsers in Scala

import fastparse._, NoWhitespace._
def number[_: P]: P[Int] = P( CharIn("0-9").rep(1).!.map(_.toInt) )
def parens[_: P]: P[Int] = P( "(" ~/ addSub ~ ")" )
def factor[_: P]: P[Int] = P( number | parens )

def divMul[_: P]: P[Int] = P( factor ~ (CharIn("*/").! ~/ factor).rep ).map(eval)
def addSub[_: P]: P[Int] = P( divMul ~ (CharIn("+\\-").! ~/ divMul).rep ).map(eval)
def expr[_: P]: P[Int]   = P( addSub ~ End )
val Parsed.Success(2, _) = parse("1+1", expr(_))
val Parsed.Success(15, _) = parse("(1+1*2)+3*4", expr(_))
val Parsed.Success(21, _) = parse("((1+1*2)+(3*4*5))/3", expr(_))
val Parsed.Failure(expected, failIndex, extra) = parse("1+1*", expr(_))
assert(
  failIndex == 4,
  extra.trace().longAggregateMsg == 
  """Expected expr:1:1 / addSub:1:1 / divMul:1:3 / factor:1:5 / ([0-9] | "("):1:5, found """""
)

FastParse is a Scala library for parsing strings and bytes into structured data. This lets you easily write a parser for any arbitrary textual data formats (e.g. program source code, JSON, ...) and have the parsers run at an acceptable speed, with great error debuggability and error reporting. Features include:

FastParse is a project by Li Haoyi. If you use FastParse and enjoyed it, please chip in to support our development at https://www.patreon.com/lihaoyi.

The following sections will introduce you to FastParse and how to use it.

Getting Started


To begin using FastParse, add the following to your build config

"com.lihaoyi" %% "fastparse" % "2.0.4" // SBT
ivy"com.lihaoyi::fastparse:2.0.4" // Mill

To use with Scala.js, you'll need

"com.lihaoyi" %%% "fastparse" % "2.0.4" // SBT
ivy"com.lihaoyi::fastparse::2.0.4" // Mill

Writing Parsers


Basic

The simplest parser matches a single string:

import fastparse._, NoWhitespace._
def parseA[_: P] = P("a")

val Parsed.Success(value, successIndex) = parse("a", parseA(_))
assert(value == (), successIndex == 1)

val f @ Parsed.Failure(label, index, extra) = parse("b", parseA(_))
assert(
  label == "",
  index == 0,
  f.msg == """Position 1:1, found "b""""
)

Such a parser returns a Parsed.Success if the input matches the string, and otherwise returns a Parsed.Failure.

As you can see, by default the Parsed.Success contains a (): Unit, unless you use Capture or Map described below.

You can also wrap the strings in an IgnoreCase("...") if you want the matching to be case-insensitive.

The P(...) lets you write mutually recursive parsers, without running into circular initialization problems, and gives the parser a nice, human-readable name (from the val it is being assigned to) which appears in parse error messages. In general, every time you assign a parser to a val, you should wrap it in P(...).

Failures

import fastparse._, NoWhitespace._
def parseA[_: P] = P( ("a" | "b").? ~ "c" )
val f @ Parsed.Failure(failureString, index, extra) = parse("d", parseA(_))

assert(
  failureString == "",
  index == 0,
  f.msg == """Position 1:1, found "d""""
)

// .trace() collects additional metadata to use for error reporting
val trace = f.trace()

// `.msg` records the last parser that failed, which is "c", and
// `.longMsg` also shows the parsing stack at point of failure
assert(
  trace.label == "\"c\"",
  trace.index == 0,
  trace.msg == """Expected "c":1:1, found "d"""",
  trace.longMsg == """Expected parseA:1:1 / "c":1:1, found "d""""
)

// aggregateMsg and longAggregateMsg record all parsers
// failing at the position, "a" | "b" | "c",
assert(
  trace.aggregateMsg == """Expected ("a" | "b" | "c"):1:1, found "d"""",
  trace.longAggregateMsg == """Expected parseA:1:1 / ("a" | "b" | "c"):1:1, found "d""""
)

When parsing, dealing with failures is an important part of making things work. Fastparse provides three levels of error reporting you can ask for when a parse fails:

  1. The default Parsed.Failure only tells you the position of the error message, without any clues what went wrong at that position.
  2. You can pass in the flag verboseFailures = true to fastparse.parse to get a better description of what caused the parse failure. This slows the parse down by a significant amount (takes ~50% longer than a normal parse), You can then use the .msg or .longMsg to see what's going on.
  3. You can call failure.trace() on an existing failure to perform a second parse over the initial input. This also gives you access to failure.trace().msg and failure.trace().longMsg, but additionally provides .aggregateMsg and .longAggregateMsg for even more details about what went wrong.

In general, you may either want to run your parses with verboseFailures enabled, or run in the default fast mode and fall back to .trace() when something goes wrong. Often the number of parse failures is small compared to the number of successes, and this approach ensures the common success case is as fast as possible.

Sequence

def ab[_: P] = P( "a" ~ "b" )

val Parsed.Success(_, 2) = parse("ab", ab(_))

val Parsed.Failure(_, 1, _) = parse("aa", ab(_))

You can combine two parsers with the ~ operator. This creates a new parser that only succeeds if both left and right parsers succeed one after another.

Repeat

def ab[_: P] = P( "a".rep ~ "b" )
val Parsed.Success(_, 8) = parse("aaaaaaab", ab(_))
val Parsed.Success(_, 4) = parse("aaaba", ab(_))

def abc[_: P] = P( "a".rep(sep="b") ~ "c")
val Parsed.Success(_, 8) = parse("abababac", abc(_))
val Parsed.Failure(_, 3, _) = parse("abaabac", abc(_))

def ab4[_: P] = P( "a".rep(min=2, max=4, sep="b") )
val Parsed.Success(_, 7) = parse("ababababababa", ab4(_))

def ab2exactly[_: P] = P( "ab".rep(exactly=2) )
val Parsed.Success(_, 4) = parse("abab", ab2exactly(_))

def ab4c[_: P] = P ( "a".rep(min=2, max=4, sep="b") ~ "c" )
val Parsed.Failure(_, 1, _) = parse("ac", ab4c(_))
val Parsed.Success(_, 4) = parse("abac", ab4c(_))
val Parsed.Success(_, 8) = parse("abababac", ab4c(_))
val Parsed.Failure(_, 7, _) = parse("ababababac", ab4c(_))

The .rep method creates a new parser that attempts to parse the given parser zero or more times. If you want to parse something a given number of times, you can use .rep(min = 2, max = 4) o r the shorter .rep(1) for one or more times, in addition there is exactly parameter that if it's defined behaves like min and max equals to it. You can optionally provide an argument which acts as a separator between the usages of the original parser, such as a comma in .rep(sep = ",").

Optional

def option[_: P] = P( "c".? ~ "a".rep(sep="b").! ~ End)

val Parsed.Success("aba", 3) = parse("aba", option(_))
val Parsed.Success("aba", 3) = parse("aba", option(_))

Similar to .rep is .?, which creates a new parser that attempts to parse the given parser zero or 1 times.

Either

def either[_: P] = P( "a".rep ~ ("b" | "c" | "d") ~ End)

val Parsed.Success(_, 6) = parse("aaaaab", either(_))
val f @ Parsed.Failure(_, 5, _) = parse("aaaaae", either(_))
val trace = f.trace().longAggregateMsg
assert(
  f.toString == """Parsed.Failure(Position 1:6, found "e")""",
  trace == """Expected either:1:1 / ("a" | "b" | "c" | "d"):1:6, found "e""""
)

The | operator tries the parser on the left, and if that fails, tries the one on the right, failing only if both parsers fail.

End, Start

def noEnd[_: P] = P( "a".rep ~ "b")
def withEnd[_: P] = P( "a".rep ~ "b" ~ End)

val Parsed.Success(_, 4) = parse("aaaba", noEnd(_))
val Parsed.Failure(_, 4, _) = parse("aaaba", withEnd(_))

The End parser only succeeds if at the end of the input string. By default, a Parser does not need to consume the whole input, and can succeed early consuming a portion of it (exactly how much input was consumed is stored in the Success#index attribute). By using End, we can make the parse fail if it doesn't consume everything

There is also a similar Start parser, which only succeeds at the start of the input

def ab[_: P] = P( (("a" | Start) ~ "b").rep ~ End).!

val Parsed.Success("abab", 4) = parse("abab", ab(_))
val Parsed.Success("babab", 5) = parse("babab", ab(_))

val Parsed.Failure(_, 2, _) = parse("abb", ab(_))

Pass, Fail

val Parsed.Success((), 0) = parse("asdad", Pass(_))
val Parsed.Failure(_, 0, _) = parse("asdad", Fail(_))

These two parsers always succeed, or always fail, respectively. Neither consumes any input.

Index

def finder[_: P] = P( "hay".rep ~ Index ~ "needle" ~ "hay".rep )

val Parsed.Success(9, _) = parse("hayhayhayneedlehay", finder(_))

Always succeeds, and provides the current index of the parse into the input string. e.g. useful for providing source locations for AST nodes. Consumes no input.

Capture

def capture1[_: P] = P( "a".rep.! ~ "b" ~ End)

val Parsed.Success("aaa", 4) = parse("aaab", capture1(_))

def capture2[_: P] = P( "a".rep.! ~ "b".! ~ End)

val Parsed.Success(("aaa", "b"), 4) = parse("aaab", capture2(_))

def capture3[_: P] = P( "a".rep.! ~ "b".! ~ "c".! ~ End)

val Parsed.Success(("aaa", "b", "c"), 5) = parse("aaabc", capture3(_))

def captureRep[_: P] = P( "a".!.rep ~ "b" ~ End)

val Parsed.Success(Seq("a", "a", "a"), 4) = parse("aaab", captureRep(_))

def captureOpt[_: P] = P( "a".rep ~ "b".!.? ~ End)

val Parsed.Success(Some("b"), 4) = parse("aaab", captureOpt(_))

So far, all the parsers go over the input text but do not return any useful value: the Success#value attribute was always (). In order to make them do so, you use the .! operation to capture the section of the input string the parser parsed.

Note the types of each parser:

In general, if you have a parser of type TupleN, capturing one more section turns it into a TupleN+1. Furthermore, if you capture within a .rep or .? optional parser, it becomes a P[Seq[T]] or P[Option[T]] respectively

AnyChar

def ab[_: P] = P( "'" ~ AnyChar.! ~ "'" )

val Parsed.Success("-", 3) = parse("'-'", ab(_))

val Parsed.Failure(stack, 2, _) = parse("'-='", ab(_))

This parser parses any single character successfully. It almost always succeeds, except if there simply aren't any characters left to parse.

There is also a plural AnyChars(count: Int) version that parses exactly count characters in a row, regardless of what they are.

Positive Lookahead

def keyword[_: P] = P( ("hello" ~ &(" ")).!.rep )

val Parsed.Success(Seq("hello"), _) = parse("hello ", keyword(_))
val Parsed.Success(Seq(), __) = parse("helloX", keyword(_))

The &(...) operator wraps a parser, only succeeds if it succeeds, but consumes no input. Useful for doing checks like "these characters must be followed by a whitespace, but don't consume the whitespace"

Negative Lookahead

def keyword[_: P] = P( "hello" ~ !" " ~ AnyChar ~ "world" ).!

val Parsed.Success("hello-world", _) = parse("hello-world", keyword(_))
val Parsed.Success("hello_world", _) = parse("hello_world", keyword(_))

val Parsed.Failure(_, 5, _) = parse("hello world", keyword(_))

The !... operator wraps a parser and only succeeds if it fails, also consuming no input. Useful to combine with other parsers like AnyChar to restrict the things that they can parse.

Map

def binary[_: P] = P( ("0" | "1" ).rep.! )
def binaryNum[_: P] = P( binary.map(Integer.parseInt(_, 2)) )

val Parsed.Success("1100", _) = parse("1100", binary(_))
val Parsed.Success(12, _) = parse("1100", binaryNum(_))

Up till now, we've only dealt with

.map lets you convert an arbitrary P[T] into a P[V] by providing a T => V function. This is useful for converting the strings and tuples/seqs/options of strings into more useful data-structures.

FlatMap

def leftTag[_: P] = P( "<" ~ (!">" ~ AnyChar).rep(1).! ~ ">")
def rightTag[_: P](s: String) = P( "</" ~ s.! ~ ">" )
def xml[_: P] = P( leftTag.flatMap(rightTag) )

val Parsed.Success("a", _) = parse("<a></a>", xml(_))
val Parsed.Success("abcde", _) = parse("<abcde></abcde>", xml(_))

val failure = parse("<abcde></edcba>", xml(_)).asInstanceOf[Parsed.Failure]
assert(
  failure.trace().longAggregateMsg == """Expected xml:1:1 / rightTag:1:8 / "abcde":1:10, found "edcba>""""
)

.flatMap allows you to dynamically choose a parser to continue with, given the result of the current parser. The example above uses it to parse balanced XML tags. .flatMap can be used to parse indentation-based grammars, and is used to do so in Scalatex.

Note that the function given to .flatMap is evaluated every time this parser is tried. You should be conscious of the cost of re-creating the resultant parser every time, since FastParse parsers are somewhat expensive to initialize despite being fast per-run. If possible, store the parsers somewhere before-hand or memo-ize/cache them to avoid initializing them wastefully.

As is normal in Scala, you can replace a .flatMap call with a for-comprehension, as below:

def leftTag[_: P] = P( "<" ~ (!">" ~ AnyChar).rep(1).! ~ ">" )
def rightTag[_: P](s: String) = P( "</" ~ s.! ~ ">" )
def xml[_: P] = P(
  for{
    s <- leftTag
    right <- rightTag(s)
  } yield right
)

val Parsed.Success("a", _) = parse("<a></a>", xml(_))
val Parsed.Success("abcde", _) = parse("<abcde></abcde>", xml(_))

val failure = parse("<abcde></edcba>", xml(_)).asInstanceOf[Parsed.Failure]
assert(
  failure.trace().longAggregateMsg == """Expected xml:1:1 / rightTag:1:8 / "abcde":1:10, found "edcba>""""
)

Which is equivalent and behaves exactly the same.

Note that .flatMap consumes whitespace between the first and second parsers; in cases where you do not want to do this, use .flatMapX

Filter

def digits[_: P] = P(CharPred(c => '0' <= c && c <= '9').rep(1).!).map(_.toInt)
def even[_: P] = P( digits.filter(_ % 2 == 0) )
val Parsed.Success(12, _) = parse("12", even(_))
val failure = parse("123", even(_)).asInstanceOf[Parsed.Failure]

.filter allows you to supply a predicate T => Boolean which is applied to the successful result of the current parser. If the predicate is true the filtered parser succeeds otherwise it fails. The example above uses .filter on digits to parse only even numbers successfully while odd numbers will fail. If the current parser fails then that failure is simply passed along.

Opaque

Sometimes it's useful to hide parser's implementation details and provide a higher-level error message. .opaque achieves exactly that.

def digit[_: P] = CharIn("0-9")
def letter[_: P] = CharIn("A-Z")
def twice[T, _: P](p: => P[T]) = p ~ p
def errorMessage[T](p: P[_] => P[T], str: String) =
  parse(str, p).asInstanceOf[Parsed.Failure].trace().longAggregateMsg

// Portuguese number plate format since 2006
def numberPlate[_: P] = P(twice(digit) ~ "-" ~ twice(letter) ~ "-" ~ twice(digit))

val err1 = errorMessage(numberPlate(_), "11-A1-22")
assert(err1 == """Expected numberPlate:1:1 / [A-Z]:1:5, found "1-22"""")

// Suppress implementation details from the error message
def opaqueNumberPlate[_: P] = numberPlate.opaque("<number-plate>")

val err2 = errorMessage(opaqueNumberPlate(_), "11-A1-22")
assert(err2 == """Expected <number-plate>:1:1, found "11-A1-22"""")

Parsers marked as .opaque only succeed or fail as a single entity and leaves no traces of underlying parsers on the stack.

Log

val logged = collection.mutable.Buffer.empty[String]
implicit val logger = Logger(logged.append(_))

def DeepFailure[_: P] = P( "C" ).log
def Foo[_: P] = P( (DeepFailure | "A") ~ "B".!).log

parse("AB", Foo(_))

val allLogged = logged.mkString("\n")

val expected =
  """+Foo:1:1, cut
    |  +DeepFailure:1:1
    |  -DeepFailure:1:1:Failure(DeepFailure:1:1 / "C":1:1 ..."AB")
    |-Foo:1:1:Success(1:3, cut)
    |
""".stripMargin.trim
assert(allLogged == expected)

Debugging Parsers is often done with the .log() method, which logs output whenever the parser is tried, and when it succeeds or fails, together with the location and other data when these things happen (traces on failures, results on successes, the presence of Cuts, ...). You can define custom loggers as we've done here, or you can just leave it to by default print to stdout.

Generally, if a parser is doing something wrong, the workflow is:

It's a non-trivial process, but it is generally not hard to figure out what's happening this way.

Utilities

Fastparse provides tools for many common use cases, so you do not need to implement them yourself. These utilities are optimized and likely faster than whatever you would come up with on the spot, so you should use them whenever possible

CharPred

def cp[_: P] = P( CharPred(_.isUpper).rep.! ~ "." ~ End )

val Parsed.Success("ABC", _) = parse("ABC.", cp(_))
val Parsed.Failure(_, 2, _) = parse("ABc.", cp(_))

CharPred takes a Char => Boolean predicate and creates a parser that parses any single character that satisfies that predicate. e.g. you can use any of the helpful methods on scala.Char to check if a Char isUpper, isDigit, isControl, etc. etc.

CharIn

def ci[_: P] = P( CharIn("abc", "xyz").rep.! ~ End )

val Parsed.Success("aaabbccxyz", _) = parse("aaabbccxyz", ci(_))
val Parsed.Failure(_, 7, _) = parse("aaabbccdxyz.", ci(_))


def digits[_: P] = P( CharIn("0-9").rep.! )

val Parsed.Success("12345", _) = parse("12345abcde", digits(_))
val Parsed.Success("123", _) = parse("123abcde45", digits(_))

Similar to CharPred, except you pass in literal strings containing regex-style character ranges. This could be things like "abcd" to accept the first four alphabets, or "a-zA-Z" to accept all alphabets both lower and uppercase. Since the "-" character is used to express character ranges, it needs to be escaped if you want to match it directly e.g. "+\\-" to match either "+" or "-". The backslash character also needs to be similarly escaped e.g. "\\\\" to match a single backslash

CharsWhile

def cw[_: P] = P( CharsWhile(_ != ' ').! )

val Parsed.Success("12345", _) = parse("12345", cw(_))
val Parsed.Success("123", _) = parse("123 45", cw(_))

A repeated version of CharPred: this parser continually chomps away at characters as long as they continue passes the given predicate.

This is a very fast parser, ideal for quickly consuming large numbers of characters.

CharsWhileIn

def cw[_: P] = P( CharsWhileIn("123456789").! )

val Parsed.Success("12345", _) = parse("12345", cw(_))
val Parsed.Success("123", _) = parse("123 45", cw(_))

A combination of CharsWhile and CharIn, this parser continues consuming characters as long as they are within the set you characters you passed to it. Like CharIn, this parser takes in literal strings representing regex-style character ranges that it accepts

StringIn

def si[_: P] = P( StringIn("cow", "cattle").!.rep(1) )

val Parsed.Success(Seq("cow", "cattle"), _) = parse("cowcattle", si(_))
val Parsed.Success(Seq("cow"), _) = parse("cowmoo", si(_))
val Parsed.Failure(_, _, _) = parse("co", si(_))

Quickly parses one of any number of strings that you give it. Behind the scenes, it converts the list of strings into a Trie so it can attempt to parse all of them in a single pass.

As a result, this is much faster to execute than if you had combined the individual strings with "cow" | "cattle" | ....

There is also a StringInIgnoreCase parser you can use if you want to match things case-insensitively.

Cuts

A "cut" (a ~/ b or a./ in Fastparse) is a marker in a recursive-descent parser that states "you cannot backtrack past this point". In FastParse, this serves two purposes:

No Cuts

def alpha[_: P] = P( CharIn("a-z") )
def nocut[_: P] = P( "val " ~ alpha.rep(1).! | "def " ~ alpha.rep(1).!)

val Parsed.Success("abcd", _) = parse("val abcd", nocut(_))

val failure = parse("val 1234", nocut(_)).asInstanceOf[Parsed.Failure]
val trace = failure.trace().longAggregateMsg
assert(
  failure.index == 0,
  trace == """Expected nocut:1:1 / ("val " ~ alpha.rep(1).! | "def "):1:1, found "val 1234""""
)

Above we have a naive scala definition parser: it either parses a val or def, a space, and its (lower-case only) name. On a success this works as expected, and extracts the name. However, on a failure, something odd happens: the deepest parser on-failure is shown to be the entire Either, rather than just the alpha that came after "val ". Why is that?

By default, the parse has an opportunity to backtrack whenever it enters a

e.g. in the case of p1 | p2, if it tries to parse p1 and fails, it then tries to parse p2. If that fails, all that FastParse knows is that one of them should have succeeded. Specifically, FastParse does not know that after successfully parsing "val ", that only the left branch of the Either is viable! Thus it has no choice but to offer both alternatives in the error message.

Cuts

def alpha[_: P] = P( CharIn("a-z") )
def nocut[_: P] = P( "val " ~/ alpha.rep(1).! | "def " ~/ alpha.rep(1).!)

val Parsed.Success("abcd", _) = parse("val abcd", nocut(_))

val failure = parse("val 1234", nocut(_)).asInstanceOf[Parsed.Failure]
val trace = failure.trace().longAggregateMsg
assert(
  failure.index == 4,
  trace == """Expected nocut:1:1 / alpha:1:5 / [a-z]:1:5, found "1234""""
)

Cuts are added using the ~/ operator, which is similar to the Sequence operator ~. Once the parse has crossed a cut, it can no longer backtrack past the point at which the cut occured. Hence, in this case you can see that it no longer backtracks to index 0, out of the enclosing Either parser and offering that in the error trace. Instead, it shows a much more precise error: at index 4, expecting one of the small set of alphanumeric characters.

In general, if you know that a parser is "committed" to one branch after parsing to a certain point, adding a cut will greatly improve the error message by ensuring that the parser itself knows that. Good places to add cuts include places like after keywords in a programming language parser, where a keyword can be followed by only one thing and anything else is an error.

Rep Cuts

def alpha[_: P] = P( CharIn("a-z") )
def stmt[_: P] = P( "val " ~ alpha.rep(1).! ~ ";" ~ " ".rep )
def stmts[_: P] = P( stmt.rep(1) ~ End )

val Parsed.Success(Seq("abcd"), _) = parse("val abcd;", stmts(_))
val Parsed.Success(Seq("abcd", "efg"), _) = parse("val abcd; val efg;", stmts(_))

val failure = parse("val abcd; val ", stmts(_)).asInstanceOf[Parsed.Failure]
val trace = failure.trace().longAggregateMsg
assert(
  failure.index == 10,
  trace == """Expected stmts:1:1 / (" " | stmt.rep(1) | end-of-input):1:11, found "val """"
)

A similar problem occurs inside Repeat or Optional parsers, where the parser will give up and backtrack out if it fails, even if it really should succeed. Again, adding cuts would result in a more precise error message:

def alpha[_: P] = P( CharIn("a-z") )
def stmt[_: P] = P( "val " ~/ alpha.rep(1).! ~ ";" ~ " ".rep )
def stmts[_: P] = P( stmt.rep(1) ~ End )

val Parsed.Success(Seq("abcd"), _) = parse("val abcd;", stmts(_))
val Parsed.Success(Seq("abcd", "efg"), _) = parse("val abcd; val efg;", stmts(_))

val failure = parse("val abcd; val ", stmts(_)).asInstanceOf[Parsed.Failure]
val trace = failure.trace().longAggregateMsg
assert(
  failure.index == 14,
  trace == """Expected stmts:1:1 / stmt:1:11 / alpha:1:15 / [a-z]:1:15, found """""
)

Another case where you may want to pay attention is when you are using delimiters with your .rep calls:

def digits[_: P] = P( CharIn("0-9").rep(1) )
def tuple[_: P] = P( "(" ~ digits.!.rep(sep=",") ~ ")" )

val Parsed.Success(Seq("1", "23"), _) = parse("(1,23)", tuple(_))

val failure = parse("(1,)", tuple(_)).asInstanceOf[Parsed.Failure]
val trace = failure.trace().longAggregateMsg
assert(
  failure.index == 2,
  trace == """Expected tuple:1:1 / ([0-9] | ")"):1:3, found ",)""""
)

in many (but not all!) cases, if a delimiter is parsed, you want to commit to parsing one more iteration of the Repeat. However, by default, it backtracks out of the Repeat entirely and starts trying to parse the next item in sequence (in this case the ")" giving the behavior shown above.

With a cut, the error is improved:

def digits[_: P] = P( CharIn("0-9").rep(1) )
def tuple[_: P] = P( "(" ~ digits.!.rep(sep=","./) ~ ")" )

val Parsed.Success(Seq("1", "23"), _) = parse("(1,23)", tuple(_))

val failure = parse("(1,)", tuple(_)).asInstanceOf[Parsed.Failure]
val index = failure.index
val trace = failure.trace().longAggregateMsg
assert(
  index == 3,
  trace == """Expected tuple:1:1 / digits:1:4 / [0-9]:1:4, found ")""""
)

The ./ operator can be used without a following parser as a shortcut for ~/ Pass. Compare the previous example with the following one:

def digits[_: P] = P( CharIn("0-9").rep(1) )
def tuple[_: P] = P( "(" ~ digits.!.rep(sep=","./) ~ ")" )

val Parsed.Success(Seq("1", "23"), _) = parse("(1,23)", tuple(_))

val failure = parse("(1,)", tuple(_)).asInstanceOf[Parsed.Failure]
val trace = failure.trace().longAggregateMsg
assert(
  failure.index == 3,
  trace == """Expected tuple:1:1 / digits:1:4 / [0-9]:1:4, found ")""""
)

Isolating Cuts

Because cuts prevent backtracking throughout the entire parser, they make it difficult to compose arbitrary parsers:

def digit[_: P] = P( CharIn("0-9") )
def time1[_: P] = P( ("1".? ~ digit) ~ ":" ~/ digit ~ digit ~ ("am" | "pm") )
def time2[_: P] = P( (("1" | "2").? ~ digit) ~ ":" ~/ digit ~ digit )
val Parsed.Success((), _) = parse("12:30pm", time1(_))
val Parsed.Success((), _) = parse("17:45", time2(_))
def time[_: P] = P( time1 | time2 ).log
val Parsed.Success((), _) = parse("12:30pm", time(_))
val failure = parse("17:45", time(_)).asInstanceOf[Parsed.Failure]
assert(failure.index == 5)  // Expects am or pm

In the above case, time1 and time2 are arbitrary parsers containing Cuts. By default, that means that once you've crossed a cut, you can no longer backtrack. However, there are cases where you want to use an existing parser (for example time1) in a situation where you want to allow it to backtrack, but you don't want to rewrite it identically but without cuts. In this case it's trivial, but if time1 was larger you would need to rewrite all of it as well as all of its transitive sub-parsers to make sure that not a single one had a cut inside!

To explicitly isolate a cut to one branch of a parser, place that branch within NoCut. Cuts within that branch will prevent backtracking inside that branch, but if that branch fails alternate branches will be tried as normal.

def digit[_: P] = P( CharIn("0-9") )
def time1[_: P] = P( ("1".? ~ digit) ~ ":" ~/ digit ~ digit ~ ("am" | "pm") )
def time2[_: P] = P( (("1" | "2").? ~ digit) ~ ":" ~/ digit ~ digit )
val Parsed.Success((), _) = parse("12:30pm", time1(_))
val Parsed.Success((), _) = parse("17:45", time2(_))
def time[_: P] = P( NoCut(time1) | time2 )
val Parsed.Success((), _) = parse("12:30pm", time(_))
val Parsed.Success((), _) = parse("17:45", time(_))

Streaming Parsing


In addition to the parsing strings, you can also parse "streaming" data from Iterators. To do so, simply pass in an Iterator[String] instead of a String to the fastparse.parse method.

import NoWhitespace._
def p[_: P] = P( "ab" ~/ "cd".rep().! ~ "ef" | "z" )

val Parsed.Success(res, i) = parse(Iterator("ab", "cd", "cd", "cd", "ef"), p(_))

assert(res == "cdcdcd")

Note that fastparse does not parse Iterator[Char] or Iterator[Byte]s for performance reasons: most input sources make data available in chunks, such as network packets or lines from file on the disk. By parsing chunks, FastParse better matches any underlying data source, and itself has better performance parsing larger chunks.

Streaming parsing still needs to buffer input in-memory: in particular, parsers like Optional, Repeat or Either means a parser may backtrack, and thus FastParse needs to buffer any input from where such a parsers starts parsing. Other parsers like Capture do not backtrack, but need to buffer data in order to return the data that gets captured. Using Cuts to prevent backtracking, apart from making Debugging Parsers easier, also allows FastParse to flush parts of the buffer that it no longer needs to backtrack into.

In general every cut in your parser possibly reduces the memory used to buffer input for iterator parsing

Streaming Parsing Buffer Size

This first benchmark measures the maximum size of buffered input when using streaming parsing, for some sample parsers we have in the test suite, for input-chunks of size 1 and 1024:

Parser Maximum buffer
for 1-sized chunk
Maximum buffer
for 1024-sized chunk
Size of input Used memory
ScalaParse155525231478941.4%
PythonParse20062867685583.6%

As you can see, for these "typical" parsers, some input needs to be buffered to allow backtracking, but it turns out to be only a few percent of the total file size.

These parsers make heavy use of backtracking operators like Either or Repeat, but also make heavy use of Cuts. This lets FastParse drop buffered input when it knows it can no longer backtrack.

Another thing to note is the chunk size: a smaller chunk size reduces the granularity of the chunks that get buffered, reducing the buffer size. However, this comes at a performance cost, as you can see below...

Streaming Parsing Performance

This next benchmark measures the effect of streaming parsing on runtime performance, using two different chunk-sizes, compared to the performance of non-streaming parsing:

Parser Score on the plain parsing Score on the iterator parsing
for 1-sized chunk
Score on the iterator parsing
for 1024-sized chunk
ScalaParse433343
PythonParse1150600890

Here, we can see that streaming parsing has a non-trivial effect on performance: ScalaParse seems unaffected by a chunks of size 1024, and takes a 25% performance hit for chunks of size 1, but PythonParse takes a significant hit (25%, 47%). While smaller chunk sizes results in smaller buffers, it also comes with a performance cost. Exactly how big you want your input chunks to be is up to you to decide: FastParse will accept an iterator of chunks as large or as small as you want.

In general, Streaming Parsing it always going to be a performance hit over parsing a single String you pre-loaded in memory. The point of streaming parsing is to handle cases where you can't/don't-want-to load everything in memory. In that case, if the choice is between slightly-slower parsing or an OutOfMemory error, streaming parsing is a good option to have.

Streaming Parsing Limitations

Apart from the performance/memory tradeoff mentioned above, streaming parsing has some limitations that it is worth being aware of:

Example Parsers


Above, we've already covered all the individual bits and pieces that make writing a parser possible. But how does that fit together? Let's take a look at some examples.

Math

import fastparse._, NoWhitespace._
def number[_: P]: P[Int] = P( CharIn("0-9").rep(1).!.map(_.toInt) )
def parens[_: P]: P[Int] = P( "(" ~/ addSub ~ ")" )
def factor[_: P]: P[Int] = P( number | parens )

def divMul[_: P]: P[Int] = P( factor ~ (CharIn("*/").! ~/ factor).rep ).map(eval)
def addSub[_: P]: P[Int] = P( divMul ~ (CharIn("+\\-").! ~/ divMul).rep ).map(eval)
def expr[_: P]: P[Int]   = P( addSub ~ End )

This is a small arithmetic expression parser, the same one shown at the top of this page. It parses only whole integers, parentheses, +-*/, and no whitespace.

Things to note:

def eval(tree: (Int, Seq[(String, Int)])) = {
  val (base, ops) = tree
  ops.foldLeft(base){ case (left, (op, right)) => op match{
    case "+" => left + right case "-" => left - right
    case "*" => left * right case "/" => left / right
  }}
}

This is a small example, but it works. We check it to verify that every parse results in the expected integer:

val Parsed.Success(2, _) = parse("1+1", expr(_))
val Parsed.Success(15, _) = parse("(1+1*2)+3*4", expr(_))
val Parsed.Success(21, _) = parse("((1+1*2)+(3*4*5))/3", expr(_))
val Parsed.Failure(expected, failIndex, extra) = parse("1+1*", expr(_))
assert(
  failIndex == 4,
  extra.trace().longAggregateMsg == 
  """Expected expr:1:1 / addSub:1:1 / divMul:1:3 / factor:1:5 / ([0-9] | "("):1:5, found """""
)

Try it out yourself! Remember that it does not handle whitespace:

Whitespace Handling

import SingleLineWhitespace._
def eval(tree: (Int, Seq[(String, Int)])): Int = {
  val (base, ops) = tree
  ops.foldLeft(base){ case (left, (op, right)) => op match{
    case "+" => left + right case "-" => left - right
    case "*" => left * right case "/" => left / right
  }}
}
def number[_: P]: P[Int] = P( CharIn("0-9").rep(1).!.map(_.toInt) )
def parens[_: P]: P[Int] = P( "(" ~/ addSub ~ ")" )
def factor[_: P]: P[Int] = P( number | parens )

def divMul[_: P]: P[Int] = P( factor ~ (CharIn("*/").! ~/ factor).rep ).map(eval)
def addSub[_: P]: P[Int] = P( divMul ~ (CharIn("+\\-").! ~/ divMul).rep ).map(eval)
def expr[_: P]: P[Int]   = P( " ".rep ~ addSub ~ " ".rep ~ End )

To handle whitespace and other non-significant characters with FastParse, we can replace the normal import NoWhitespace._ with a whitespace consumer that picks up the whitespace we want, e.g. SingleLineWhitespace for skipping spaces, tabs, MultiLineWhitespace for also skipping newlines, and ScriptWhitespace/JavaWhitespace/ScalaWhitespace for also skipping various sorts of comments (#-delimited, // and /* */ delimited, and allowing nested /* */s respectively)

The whitespace consumer affects the ~ and .rep operators to consume all non-trailing whitespace and ignoring it.

Here it is in action:

def check(str: String, num: Int) = {
  val Parsed.Success(value, _) = parse(str, expr(_))
  assert(value == num)
}

* - check("1+1", 2)
* - check("1+   1*   2", 3)
* - check("(1+   1  *  2)+(   3*4*5)", 63)
* - check("15/3", 5)
* - check("63  /3", 21)
* - check("(1+    1*2)+(3      *4*5)/20", 6)
* - check("((1+      1*2)+(3*4*5))/3", 21)

You can also define your own custom whitespace consumer, if none of bundled ones fit your needs:

implicit val whitespace = { implicit ctx: ParsingRun[_] =>
  CharsWhileIn(" \t", 0)
}
def eval(tree: (Int, Seq[(String, Int)])): Int = {
  val (base, ops) = tree
  ops.foldLeft(base){ case (left, (op, right)) => op match{
    case "+" => left + right case "-" => left - right
    case "*" => left * right case "/" => left / right
  }}
}
def number[_: P]: P[Int] = P( CharIn("0-9").rep(1).!.map(_.toInt) )
def parens[_: P]: P[Int] = P( "(" ~/ addSub ~ ")" )
def factor[_: P]: P[Int] = P( number | parens )

def divMul[_: P]: P[Int] = P( factor ~ (CharIn("*/").! ~/ factor).rep ).map(eval)
def addSub[_: P]: P[Int] = P( divMul ~ (CharIn("+\\-").! ~/ divMul).rep ).map(eval)
def expr[_: P]: P[Int]   = P( " ".rep ~ addSub ~ " ".rep ~ End )

Or try it yourself:

Indentation Grammars

def eval(tree: (String, Seq[Int])) = tree match{
  case ("+", nums) => nums.reduceLeft(_+_)
  case ("-", nums) => nums.reduceLeft(_-_)
  case ("*", nums) => nums.reduceLeft(_*_)
  case ("/", nums) => nums.reduceLeft(_/_)
}

/**
  * Parser for an indentation-based math syntax. Parens are no longer
  * necessary, and the whole parser is parametrized with the current
  * depth of indentation
  */
class Parser(indent: Int){
  def number[_: P]: P[Int] = P( CharIn("0-9").rep(1).!.map(_.toInt) )

  def deeper[_: P]: P[Int] = P( " ".rep(indent + 1).!.map(_.length) )
  def blockBody[_: P]: P[Seq[Int]] = "\n" ~ deeper.flatMapX(i =>
    new Parser(indent = i).factor.rep(1, sep = ("\n" + " " * i)./)
  )
  def block[_: P]: P[Int] = P( CharIn("+\\-*/").! ~/ blockBody).map(eval)

  def factor[_: P]: P[Int] = P( number | block )

  def expr[_: P]: P[Int]   = P( block ~ End )
}
def expr[_: P] = new Parser(indent = 0).expr

Here is a grammar that is used to parse a simple indentation-based math grammar. To understand the grammar it is trying to parse, it is worth looking at the test data:

def check(str: String, num: Int) = {
  val Parsed.Success(value, _) = parse(str, expr(_))
  assert(value == num)
}

check(
  """+
    |  1
    |  1
  """.stripMargin.trim,
  2
)
check(
  """+
    |  1
    |  *
    |    1
    |    2
  """.stripMargin.trim,
  3
)

As you can see, it is basically a prefix math evaluator, where you use indentation to pass the numbers or expressions to each operator to operate on.

As for the parser, the novel things are:

Note how there is no pre-processing, and no lexining phase where the lexer has to guess where in the token stream to inject synthetic indent and dedent tokens, Everything happens in a single pass.

Try it out!

Json

sealed trait Val extends Any {
  def value: Any
  def apply(i: Int): Val = this.asInstanceOf[Arr].value(i)
  def apply(s: java.lang.String): Val =
    this.asInstanceOf[Obj].value.find(_._1 == s).get._2
}
case class Str(value: java.lang.String) extends AnyVal with Val
case class Obj(value: (java.lang.String, Val)*) extends AnyVal with Val
case class Arr(value: Val*) extends AnyVal with Val
case class Num(value: Double) extends AnyVal with Val
case object False extends Val{
  def value = false
}
case object True extends Val{
  def value = true
}
case object Null extends Val{
  def value = null
}
import fastparse._, NoWhitespace._
def stringChars(c: Char) = c != '\"' && c != '\\'

def space[_: P]         = P( CharsWhileIn(" \r\n", 0) )
def digits[_: P]        = P( CharsWhileIn("0-9") )
def exponent[_: P]      = P( CharIn("eE") ~ CharIn("+\\-").? ~ digits )
def fractional[_: P]    = P( "." ~ digits )
def integral[_: P]      = P( "0" | CharIn("1-9")  ~ digits.? )

def number[_: P] = P(  CharIn("+\\-").? ~ integral ~ fractional.? ~ exponent.? ).!.map(
  x => Js.Num(x.toDouble)
)

def `null`[_: P]        = P( "null" ).map(_ => Js.Null)
def `false`[_: P]       = P( "false" ).map(_ => Js.False)
def `true`[_: P]        = P( "true" ).map(_ => Js.True)

def hexDigit[_: P]      = P( CharIn("0-9a-fA-F") )
def unicodeEscape[_: P] = P( "u" ~ hexDigit ~ hexDigit ~ hexDigit ~ hexDigit )
def escape[_: P]        = P( "\\" ~ (CharIn("\"/\\\\bfnrt") | unicodeEscape) )

def strChars[_: P] = P( CharsWhile(stringChars) )
def string[_: P] =
  P( space ~ "\"" ~/ (strChars | escape).rep.! ~ "\"").map(Js.Str)

def array[_: P] =
  P( "[" ~/ jsonExpr.rep(sep=","./) ~ space ~ "]").map(Js.Arr(_:_*))

def pair[_: P] = P( string.map(_.value) ~/ ":" ~/ jsonExpr )

def obj[_: P] =
  P( "{" ~/ pair.rep(sep=","./) ~ space ~ "}").map(Js.Obj(_:_*))

def jsonExpr[_: P]: P[Js.Val] = P(
  space ~ (obj | array | string | `true` | `false` | `null` | number) ~ space
)

This is a somewhat larger example than the math parser shown above. In it, we parse a JSON expression from a string, including all the proper handling for whitespace and error-handling built in.

Things to note:

def apply(i: Int): Val = this.asInstanceOf[Arr].value(i)
def apply(s: java.lang.String): Val =
  this.asInstanceOf[Obj].value.find(_._1 == s).get._2

We can verify that this parser builds the JSON tree that we expect:

val Parsed.Success(value, _) =
  parse("""{"omg": "123", "wtf": 12.4123}""", jsonExpr(_))

assert(value == Js.Obj("omg" -> Js.Str("123"), "wtf" -> Js.Num(12.4123)))

And that it provides good error messages in the case of mal-formed JSON, even for moderately-sized fragemnts

{
    "firstName": "John",
    "lastName": "Smith",
    "age": 25,
    "address": {
        "streetAddress": "21 2nd Street",
        "city": "New York",
        "state": "NY",
        "postalCode": 10021
    },
    "phoneNumbers":
        {
            "type": "home",
            "number": "212 555-1234"
        },
        {
            "type": "fax",
            "number": "646 555-4567"
        }
    ]
}
jsonExpr:1:1 / obj:2:9 / pair:16:19 / string:16:19 / "\"":17:17 ..."{\n        "

Here, we're missing a square bracket after the "phoneNumbers" key, and so the parser expects to find a single JSON expression. It finds a JSON object, and then fails reporting that it expected to find the next key (a string), but instead found "{\n" at that index.

Try it out!

ScalaParse

ScalaParse is a parser for the entire Scala programming language, written using FastParse. This is notable for a few reasons:

ScalaParse does not currently generate an AST. As you can see, the parse result above is listed as undefined. However, that does not make it useless! Even without generating an AST, ScalaParse can be used to:

Using ScalaParse

To begin using ScalaParse, add
"com.lihaoyi" %% "scalaparse" % "2.0.4"

To your SBT configuration. To use with Scala.js, you'll need

"com.lihaoyi" %%% "scalaparse" % "2.0.4"

PythonParse

There is now an example Python parser available under a subproject in the repo. This is a good example of a real-world parser: parsing knotty syntax (including indentation-delimited blocks!), building an AST, and with heavy unit tests.

PythonParse is currently compatible enough to parse all the python sources in Zulip, Ansible, Changes, Django, and Flask. It isn't published yet on maven central, but feel free to look at it if you want an idea of how to write a complex, real parser.

"com.lihaoyi" %%% "pythonparse" % "2.0.4"

CssParse

CssParse is a parser that parses CSS files into an abstract syntax tree (AST). The implementation is too large to show in-line, but can be found here:

This AST can then be used for a variety of reasons: you could analyze the CSS to try and find bugs, transform the CSS in some custom way (e.g. prefixing class-names with the name of the file) or just re-formatting the CSS when printing it back out.

CssParse compiles to Javascript via Scala.js, and we have a demo here that demonstrates the use of CssParse as a CSS pretty-printer. Enter some CSS in the box below, no matter how it's formatted or minified, and CssParse will add the necessary spaces and tabs to make the file readable and good-looking.

As mentioned above, CssParse builds and AST that stores information about tags and rules in the given CSS, this AST isn't complete, because of complexity of initial CSS format, but it provides all the essential information about basic elements of file (blocks, selectors, rules). The correctness of CssParse is tested by parsing and then printing several huge files including CSS from Bootstrap and Primer.

This is available on Maven Central as

"com.lihaoyi" %%% "cssparse" % "2.0.4"

API Highlights


Parsing Results

The result of a parser comes in two flavors of Parsed; the first is a success (Parsed.Success) and the second is a failure (Parsed.Failure). Parsed.Success provides the parsed value - the value you are probably most interested in - and the index in the input string till where the parse was performed. Parsed.Failure allows you to retrieve the last parser that failed and the index where it failed. Additionally, failure provides an Parsed.Failure.extra field that provides precise details about the failure, in particular, and most importantly a complete stack trace of the involved parsers, which is accessible via extra.traced.

sealed trait AndOr
case object And extends AndOr
case object Or extends AndOr
def and[_: P] = P(IgnoreCase("And")).map(_ => And)
def or[_: P] = P(IgnoreCase("Or")).map(_ => Or)
def andOr[_: P] = P(and | or)

def check(input: String, expectedOutput: String) = {
  val folded = parse(input, andOr(_)).fold(
    (_, _, _) => s"Cannot parse $input as an AndOr",
    (v, _) => s"Parsed: $v"
  )
  assert(folded == expectedOutput)

It is also possible to pattern match over Parsed, however, you may experience spurious warnings related to SI-4440. In order to prevent these warnings import fastparse.core.Result in versions 0.2.x and import fastparse.core.Parsed in higher versions than 0.2.x.

An overview of Parsed:

package fastparse

import fastparse.internal.Util

/**
  * The outcome of a [[ParsingRun]] run, either a success (with value and index) or
  * failure (with associated debugging metadata to help figure out what went
  * wrong).
  *
  * Doesn't contain any information not already present in [[ParsingRun]], but
  * packages it up nicely in an immutable case class that's easy for external
  * code to make use of.
  */
sealed abstract class Parsed[+T](val isSuccess: Boolean){
  def fold[V](onFailure: (String, Int, Parsed.Extra) => V, onSuccess: (T, Int) => V): V
  def get: Parsed.Success[T]
}

object Parsed{
  def fromParsingRun[T](p: ParsingRun[T]): Parsed[T] = {
    if (p.isSuccess) Parsed.Success(p.successValue.asInstanceOf[T], p.index)
    else Parsed.Failure(
      Option(p.lastFailureMsg).fold("")(_()),
      p.index,
      Parsed.Extra(p.input, p.startIndex, p.index, p.originalParser, p.failureStack)
    )
  }

  /**
    * The outcome of a successful parse
    *
    * @param value The value returned by the parse
    * @param index The index at which the parse completed at
    */
  final case class Success[+T](value: T, index: Int) extends Parsed[T](true){
    def get = this
    def fold[V](onFailure: (String, Int, Extra) => V, onSuccess: (T, Int) => V) = onSuccess(value, index)
    override def toString() = s"Parsed.Success($value, $index)"
  }

  /**
    * The outcome of a failed parse
    *
    * @param label A hint as to why the parse failed. Defaults to "",
    *                     unless you set `verboseFailures = true` or call
    *                     `.trace()` on an existing failure
    * @param index The index at which the parse failed
    * @param extra Metadata about the parse; useful for re-running the parse
    *              to trace out a more detailed error report
    */
  final case class Failure(label: String,
                           index: Int,
                           extra: Extra) extends Parsed[Nothing](false){
    def get = throw new Exception("Parse Error, " + msg)
    def fold[V](onFailure: (String, Int, Extra) => V, onSuccess: (Nothing, Int) => V) = onFailure(label, index, extra)
    override def toString() = s"Parsed.Failure($msg)"

    /**
      * Displays the failure message excluding the parse stack
      */
    def msg = {
      label match{
        case "" =>
          "Position " + extra.input.prettyIndex(index) +
          ", found " + Failure.formatTrailing(extra.input, index)
        case s => Failure.formatMsg(extra.input, List(s -> index), index)
      }
    }

    /**
      * Displays the failure message including the parse stack, if possible
      */
    def longMsg = {
      if (extra.stack.nonEmpty) {
        Failure.formatMsg(extra.input, extra.stack ++ List(label -> index), index)
      } else throw new Exception(
        "`.longMsg` requires the parser to be run with `verboseFailures = true`, " +
        "or to be called via `.trace().longMsg` or `.trace().longAggregateMsg`"
      )
    }

    /**
      * Re-runs the failed parse with `verboseFailures` turned on and failure
      * aggregation enabled. This allows Fastparse to provide much more
      * detailed error messages, at a cost of taking ~2x as long than the
      * original parse
      */
    def trace() = extra.trace
  }

  object Failure{
    def formatMsg(input: ParserInput, stack: List[(String, Int)], index: Int) = {
      "Expected " + Failure.formatStack(input, stack) +
      ", found " + Failure.formatTrailing(input, index)
    }
    def formatStack(input: ParserInput, stack: List[(String, Int)]) = {
      stack.map{case (s, i) => s"$s:${input.prettyIndex(i)}"}.mkString(" / ")
    }
    def formatTrailing(input: ParserInput, index: Int) = {
      Util.literalize(input.slice(index, index + 10))
    }
  }

  case class Extra(input: ParserInput,
                   startIndex: Int,
                   index: Int,
                   originalParser: ParsingRun[_] => ParsingRun[_],
                   stack: List[(String, Int)]) {
    @deprecated("Use .trace instead")
    def traced = trace

    /**
      * Re-runs the failed parse with aggregation turned on. This is the
      * slowest of Fastparse's error reporting mode, taking ~2x as long
      * as the original parse, but provides the greatest detail in the error
      * message
      */
    def trace() = {
      input.checkTraceable()
      TracedFailure.fromParsingRun(
        parseInputRaw[Any](
          input,
          originalParser,
          startIndex = startIndex,
          traceIndex = index,
          enableLogging = false,
          verboseFailures = true
        )
      )
    }
  }

  object TracedFailure{

    def fromParsingRun[T](p: ParsingRun[T]) = {
      assert(!p.isSuccess)
      TracedFailure(
        p.failureAggregate.reverse.map(_()).distinct,
        Parsed.fromParsingRun(p).asInstanceOf[Failure]
      )
    }
  }

  /**
    * A decorated [[Failure]] with extra metadata; provides a much more
    * detailed, through verbose, of the possible inputs that may have been
    * expected at the index at which the parse failed.
    *
    * @param failureAggregate contains not just the parsers which were present
    *                         when the parse finally failed, but also any other
    *                         parsers which may have earlier been tried at the
    *                         same index.
    * @param failure The raw failure object
    */
  case class TracedFailure(failureAggregate: Seq[String],
                           failure: Failure){
    def label = failure.label
    def index = failure.index
    def input = failure.extra.input
    def stack = failure.extra.stack
    def combinedAggregateString = failureAggregate match{
      case Seq(x) => x
      case items => items.mkString("(", " | ", ")")
    }

    @deprecated("Use .msg instead")
    def trace = longAggregateMsg
    /**
      * Displays the failure message excluding the parse stack
      */
    def msg = failure.msg
    /**
      * Displays the failure message including the parse stack, if possible
      */
    def longMsg = failure.longMsg
    /**
      * Displays the aggregate failure message, excluding the parse stack
      */
    def aggregateMsg = Failure.formatMsg(input, List(combinedAggregateString -> index), index)

    /**
      * Displays the aggregate failure message, including the parse stack
      */
    def longAggregateMsg = Failure.formatMsg(input, stack ++ Seq(combinedAggregateString -> index), index)
  }
}

Note how Failure only contains the parser which failed and a single index where the parse failed. Further debugging information is available via the Failure.Extra class. Especially the TracedFailure that is lazily-computed via Extra.traced, provides valuable information: It performs a whole new parse on the input data with additional instrumentation, and provides additional insight into why the parse failed.

Computing the Extra.traced data is not done by default for performance reasons: the additional run takes about 3x longer than the initial run due to the instrumentation, for a total of 4x slowdown. If you want the information for debugging, though, it will be there.

Performance


FastParse will never be able to compete with hand-written recursive descent parsers for speed, but for most use cases it is plenty fast enough. Here's a comparison of FastParse with alternatives, using Parboiled2's JSON parsing benchmark, which parses a ~21,500 line JSON file:

BenchmarkScore
fastparse159.5
circe332.4
argonaut149.1
uJson266.6
json4s100.9
play-json226.6
scala-parser-combinators0.9

These numbers are the number of iterations/second of parsing a sample test.json file, averaged over 200 runs. As you can see, the FastParse based parser comes within a factor of 4 of the fastest hand written parser (Jackson), is just as fast as the Parboiled2 based parser (slightly faster/slower depending if full tracing is enabled), and is almost 100x faster than the scala-parser-combinators library.

In exchange for the perf hit compared to hand-rolled solutions, you get the short, super-simple parser definition, and excellent error free error reporting. While for super-high-performance use cases you may still want a hand-rolled parser, for many ad-hoc situations a FastParse parser would do just fine.

A similar speed ratio can be seen in parsing a sample Scala file comparing FastParse andScalac's inbuilt hand-written Scala-language parser:

BenchmarkScore
fastparse203
scalac754

Or comparing Fastparse's Python parser with the Jython Python parser:

BenchmarkScore
fastparse406
jython472

In all these cases, you can see that the iterations-per-second performance of Fastparse parsers is comparable to various production quality parsers. While the Fastparse parser may be a few times slower, it is nonetheless competitive, and at the same time is usually less than 1/10th as much code to write, understand and debug.

Improving Performance

There are many ways to improve performance of your FastParse parsers. If you study the example parsers included in the repo, those already have many of these techniques applied, and if you follow the same style you'll probably do ok. Nevertheless, here are some concrete tips:

Profiling

Since FastParse Parsers are just methods, you can use standard profiling techniques to show where in the parser time is being spent. For example, here is the JProfiler profile of the ScalaParse Scala syntax parser:

Using standard tools, you can easily dig into what parts of your parser are slow and why

Debugging Parsers


The vast majority of your time working with FastParse, your parsers will be incorrect. This is almost by definition, because once your parser is correct, you'll be done and can go do something else with your life!

Even if your parsers are correct, often you'll find yourself parsing broken input:

No matter what goes wrong, someone will need to figure it out and fix it. Thus FastParse puts a lot of effort into making working with broken parsers and input as easy as possible.

Let's take an example Parser:

object Foo{

  def plus[_: P] = P( "+" )
  def num[_: P] = P( CharIn("0-9").rep(1) ).!.map(_.toInt)
  def side[_: P] = P( "(" ~ expr ~ ")" | num )
  def expr[_: P]: P[Int] = P( side ~ plus ~ side ).map{case (l, r) => l + r}
}

This is a simple parser that parses some basic arithmetic expressions: 1+2, (1+2)+(3+4), etc. It's simpler than Math parser shown at the top of the page, as it does not handle multiple operators in a row e.g. 1+2+3+4, nor does it handle operator precedence. Nevertheless it will be enough to show how error handling and debugging works.

If we run the parser on a bad input, though, we get this:

parse("(1+(2+3x))+4", Foo.expr(_)),
"""Parsed.Failure(Position 1:1, found "(1+(2+3x))")"""

As you can see, the error message is pretty generic: "i had a syntax error at line 1 column 1 (index 0)". Why does it tell us that?

Using Cuts

The answer is that as far as FastParse knows, you could have wanted either the "(" ~ expr ~ ")" or the num at that position, and it doesn't know which one. Thus even though it starts off parsing a paren, when that branch eventually fails (it tries to parse a ")" at index 7, but finds a "x") it backtracks out of the "(" ~ expr ~ ")" parser and then tries to parse num. When that fails, it doesn't know which side was "meant" to succeed, and so it gives up and just tells you both sides failed to parse.

Although FastParse doesn't know which branch was meant to succeed, we know that once we've parsed a "(", it can no longer parse a number! Thus there's no point in backtracking and trying that side of the |. We can tell FastParse this fact by adding Cuts ~/ after "("

object Foo{

  def plus[_: P] = P( "+" )
  def num[_: P] = P( CharIn("0-9").rep(1) ).!.map(_.toInt)
  def side[_: P] = P( "(" ~/ expr ~ ")" | num )
  def expr[_: P]: P[Int] = P( side ~ plus ~ side ).map{case (l, r) => l + r}
}

Now, once FastParse sees a "(", it can no longer backtrack! Thus it knows that whatever error occurs later, it must be because it failed to parse a ")" and not because num failed. Then the error message becomes much more precise and useful:

parse("(1+(2+3x))+4", Foo.expr(_)),
"""Parsed.Failure(Position 1:8, found "x))+4")"""

Using Log

We can add Log calls to make FastParse tell us a lot more about what a parser is doing. For example, if we want to know whenever a side or expr is being attempted, we can add .log() to those to parsers to find out:

object Foo{

  def plus[_: P] = P( "+" )
  def num[_: P] = P( CharIn("0-9").rep(1) ).!.map(_.toInt)
  def side[_: P] = P( "(" ~/ expr ~ ")" | num ).log
  def expr[_: P]: P[Int] = P( side ~ plus ~ side ).map{case (l, r) => l + r}.log
}

Note that .log by default prints out the name of the enclosing parser method. This is what you want if you are logging the body of an entire method like DeepFailure or Foo, but if you want to log the individual bits and pieces like "A" or "B" you'll need to pass in the name of the thing you are logging manually (as we have done here)

Then when you run it on an invalid input:

parse("(1+(2+3x))+4", Foo.expr(_))

You get a dump of everything the logged parsers are trying to do

"""+Foo:1:1, cut
  |  +DeepFailure:1:1
  |  -DeepFailure:1:1:Failure(DeepFailure:1:1 / "C":1:1 ..."AB")
  |-Foo:1:1:Success(1:3, cut)
  |

+ is when a parser is started, - is when it finishes with either a success or a failure. In the case of failure, it tells you what the stack was when it failed.

The general strategy for adding .logs is:

  1. Is my parser misbehaving? This is usually obvious from seeing parse failures when there shouldn't be
  2. Are any sub-parsers which I believe should be succeeding/failing/getting-called, aren't? Add logging to the sub-parsers. You can do this at the definition-site of the sub-parsers as shown above, or to the use-site e.g. side.log("SIDE 1") ~ plus ~ side.log("SIDE 2") if the parser is used in multiple places and you only want to log this usage.
  3. Look at the logging, see some parser behaving strangely. Go to 1.

In general, you do not want to add too many .log() calls otherwise the output becomes noisy and verbose and hard to understand. Adding them incrementally helps you narrow down the problem while still keeping the output readable.

Tracing

By default, on failure, FastParse only provides the index and the last parser which failed at that index. This is information FastParse already has and is thus cheap to provide, and often is enough to show what went wrong, and where.

Often you want something more, though, and for that FastParse provides tracing, as described in the documentation of Parsing Results. By computing the .traced value on a Failure, FastParse will perform a whole second parse on the original input, starting from the same index, but with additional tracing code to keep track of what's happening. This typically costs ~2x as much as the original parse, so isn't done by default, but it's trivial to ask for it.

For example, this is tracing being done on an invalid input we know will fail:

val Parsed.Failure(stack, idx, extra) = parse("(1+(2+3x))+4", Foo.expr(_))

We know that this input will fail, because our grammar (defined earlier) does not contain an "x" in it! It only handles numbers and "+" and parentheses. expected gives us the parser that it needed in order to continue (with a readable .toString) and idx the index that it failed at:

> (idx, expected) // Last index and last parser at which it failed
(")", 7)

But there is also the extra result which contains additional information you can use, e.g. the extra.traced result.

> extra.traced.trace // The named parsers in the stack when it failed
Expected expr:1:1 / side:1:1 / expr:1:2 / side:1:4 / ([0-9] | ")"):1:8, found "x))+4"

> extra.traced.stack // Same as .trace, but as a List[Frame] rather than String
List(
    (([0-9] | ")"), 7),
    (side, 3),
    (expr, 1),
    (side, 0),
    (expr, 0)
)

As you can see, tracing gives us a much more detailed view: every parser in the stack when the parse failed, what indices they were tried at. Apart from getting it as a readable string via .trace, you can also get it as structured data via .stack in case you want to manipulate it programmatically.

Instrumenting Parsers

FastParse provides an instrument argument, which you can use to inject callbacks that let you run code before or after every named parse (defined inside a P) is called. This can be used for a wide variety of things. For example, you could use it to count how many times each parser in the example Math parser gets called when you parse a simple arithmetic expression:

val callCount = mutable.Map.empty[String, Int]


val instrument = new Instrument {
  def beforeParse(parser: String, index: Int): Unit = {
    callCount(parser) = callCount.getOrElse(parser, 0) + 1
  }
  def afterParse(parser: String, index: Int, success: Boolean): Unit = ()
}

parse("((1+1*2)+(3*4*5))/3", expr(_), instrument = instrument)

val expectedCallCount = Map(
  "expr" -> 1,
  "addSub" -> 4,
  "divMul" -> 6,
  "factor" -> 10,
  "number" -> 10,
  "parens" -> 3
)
assert(callCount == expectedCallCount)

This is useful for ad-hoc investigation of what your parser is doing: you can easily count how many times various parsers are being run, whether they're succeeding and failing and at what indices.

The above example runs code before each named P(...) parser is run. You can also run code after the parser is run:

val resultCount = mutable.Map.empty[(String, Boolean), Int]
val instrument = new Instrument {
  def beforeParse(parser: String, index: Int): Unit = ()
  def afterParse(parser: String, index: Int, success: Boolean): Unit = {
    val resultKey = (parser, success)
    resultCount(resultKey) = resultCount.getOrElse(resultKey, 0) + 1
  }
}

In this case, we are using it to count not just how many times each named P(...) parser is run, but also how many times it succeeds and fails. We can parse a "good" input and see that most of the attempts by named parser succeed:

// Good Parse
parse("((1+1*2)+(3*4*5))/3", expr(_), instrument = instrument)

val expectedResultCount = Map(
  ("expr", true) -> 1,
  ("addSub", true) -> 4,
  ("divMul", true) -> 6,
  ("factor", true) -> 10,
  ("number", true) -> 7,
  ("number", false) -> 3,
  ("parens", true) -> 3
)
assert(resultCount == expectedResultCount)

And we can parse invalid input, and see that it results in many of the named parsers failing repeatedly in the course of the parse:

// Bad Parse
resultCount.clear()
parse("((1+1*2)+(3*4*))/3", expr(_), instrument = instrument)

val expectedResultCount2 = Map(
  ("expr", false) -> 1,

  ("addSub", true) -> 1,
  ("addSub", false) -> 3,

  ("divMul", true) -> 3,
  ("divMul", false) -> 3,

  ("factor", true) -> 6,
  ("factor", false) -> 3,

  ("number", true) -> 5,
  ("number", false) -> 4,

  ("parens", true) -> 1,
  ("parens", false) -> 3
)
assert(resultCount == expectedResultCount2)

Which is what you would expect, since the parse overall failed! This measurement of pass/fail rates for each parser is a useful tool for understanding how the overall parser behaves at runtime: which individual parsers are being used over and over and are worth optimizing, and which are called only a handful of times and not bothering with. If an individual parser is being called over and over, you could optimize its internals to make it faster, or you could try to refactor the rest of the overall parser so this individual parser gets called fewer times.

Instrumenting parsers typically isn't something you want to use on a "production" parser. Nevertheless, instrument it is still a useful tool to have available in your toolbox if you are unsure of what a parser is doing, and can be used to give you insight into the runtime behavior of your parsers which is hard to get otherwise.

Use Cases

What's the purpose of all this detailed error reporting? The goal is three-fold:

In general, FastParse's error reporting is detailed and structured. As a developer, most of your time spent interacting with your parser is when it is incorrect and throwing errors at you. As a user, most of your time spent interacting with the parser is when your input is incorrect and it is throwing errors at you. This is almost self-evident, since once your parser is correct or your input is correct you're done and go do other things

Thus, FastParse makes an effort to make the error reporting both detailed and structured. This means as a developer you can quickly diagnose problems, and (if you wish to) put in effort to use the structured errors to help your users diagnose problems. That makes life better for everybody.

Comparisons


FastParse differs from all other parser-combinator libraries in the Scala universe, in quite substantial ways:

Internals


FastParse 2.0.4 is implemented as a set of methods that perform a recursive-descent parse on the given input, with all book-keeping information maintained in the fastparse.ParsingRun[T] objects (abbreviated fastparse.P[T]). ParsingRuns are mutable, heavily stateful objects that keep track of everything related to the parse: the current index, whether backtracking is allowed, any value we may want to return, etc.. By defining your parsers as:

def myParser[_: P]: P[T] = P( ... )

We ensure that a ParsingRun object is required to start the parse, the same instance is made available implicitly to any other parsers myParser may call and finally returned with the relevant value of type T stored in it.

Inlining

FastParse heavily relies on inlining to achieve performance; many of the FastParse operators such as a ~ b, !a, a.rep etc. are implemented using macros which inline their implementation code at the call-site.

This inlining should be mostly transparent to you, as the operators would parse the same input and return the same values as if they were not inlined. The only thing you may notice is these operators do not appear in stack traces or profiles, as their code is inlined as part of the enclosing method (e.g. def myParser).

Opacity

Fastparse parsers are opaque: as plain methods def myParser[_: P]: P[T], the only thing you can do is invoke them. You cannot programmatically inspect the body of a parser, see what other parsers it depends on, or perform any transformations on it. This in turn allows us to perform additional optimizations to improve performance of the parsing run.

Synchronous

As plain methods, Fastparse parsers are synchrous: calling a parser method does not return until the parse is complete. Even if parsing streaming input, FastParse will block on the input iterator until it either provides the text to parse or is exhausted.

Stack-Limited

FastParse's parsing methods use the normal JVM method-call-stack to perform their recursive descent. This means that excessively deep or nested parses can cause a stack-overflow, which can be mitigated by the normal JVM flags to increase the stack size in memory.

Change Log


2.0.4

1.0.0

0.4.4

0.4.3

0.4.2

0.4.1

0.4.0

0.3.7

0.3.6

0.3.5

0.3.4

0.3.3

0.3.2

0.3.1

0.2.1

0.2.0

0.1.7

0.1.6