flatparse: High-performance parsing from strict bytestrings

[ library, mit, parsing ] [ Propose Tags ]
Versions [RSS],,,,,,,,,,,,,,,,,,,,,,,,,
Dependencies base (>=4.7 && <5), bytestring, containers, integer-gmp, template-haskell, utf8-string (>=1.0.2 && <1.1) [details]
License MIT
Copyright 2021 András Kovács
Author András Kovács
Maintainer puttamalac@gmail.com
Category Parsing
Home page https://github.com/AndrasKovacs/flatparse#readme
Bug tracker https://github.com/AndrasKovacs/flatparse/issues
Source repo head: git clone https://github.com/AndrasKovacs/flatparse
Uploaded by AndrasKovacs at 2024-03-10T21:11:01Z
Distributions LTSHaskell:, NixOS:, Stackage:
Reverse Dependencies 9 direct, 25 indirect [details]
Downloads 3635 total (98 in the last 30 days)
Rating (no votes yet) [estimated by Bayesian average]
Your Rating
  • λ
  • λ
  • λ
Status Docs available [build log]
Last success reported on 2024-03-10 [all 1 reports]

Readme for flatparse-

[back to package description]


Hackage CI

flatparse is a high-performance parsing library, supporting parsing for programming languages, human-readable data and machine-readable data. The "flat" in the name refers to the ByteString parsing input, which has pinned contiguous data, and also to the library internals, which avoids indirections and heap allocations whenever possible. flatparse is generally lower-level than parsec-style libraries, but it is possible to build higher-level features (such as source spans, hints, indentation parsing) on top of it, without making any compromises in performance.


It is advised to build with -fllvm option when using this package, since that can result in significant speedups (20-40% from what I've seen). Additionally, you can enable -fllvm for flatparse specifically by enabling the llvm package flag. However, this has minor impact, since almost all parser code will be typically inlined into modules outside flatparse, and compiled there.

Features and non-features

  • Excellent performance. On microbenchmarks, flatparse is around 10 times faster than attoparsec or megaparsec. On larger examples with heavier use of source positions and spans and/or indentation parsing, the performance difference grows to 20-30 times. Compile times and executable sizes are also significantly better with flatparse than with megaparsec or attoparsec. flatparse internals make liberal use of unboxed tuples and GHC primops. As a result, pure validators (parsers returning ()) in flatparse are not difficult to implement with zero heap allocation.
  • No incremental parsing, and only strict ByteString is supported as input. However, it can be still useful to convert from Text, String or other types to ByteString, and then use flatparse for parsing, since flatparse performance usually more than makes up for the conversion costs.
  • Only little-endian 64 bit systems are currently supported as the host machine. This may change in the future. Getting good performance requires architecture-specific optimizations; I've only considered the most common setting at this point. However, flatparse does include primitive integer parsers with specific endianness.
  • Support for fast source location handling, indentation parsing and informative error messages. flatparse provides a low-level interface to these. Batteries are not included, but it should be possible for users to build custom solutions, which are more sophisticated, but still as fast as possible. In my experience, the included batteries in other libraries often come with major unavoidable overheads, and often we still have to extend existing machinery in order to scale to production features.
  • The backtracking model of flatparse is different to parsec libraries, and is more close to the nom library in Rust. The idea is that parser failure is distinguished from parsing error. The former is used for control flow, and we can backtrack from it. The latter is used for unrecoverable errors, and by default it's propagated to the top. flatparse does not track whether parsers have consumed inputs. In my experience, what we really care about is the failure/error distinction, and in parsec or megaparsec the consumed/non-consumed separation is often muddled and discarded in larger parser implementations. By default, basic flatparse parsers can fail but can not throw errors, with the exception of the specifically error-throwing operations. Hence, flatparse users have to be mindful about grammar, and explicitly insert errors where it is known that the input can't be valid.

flatparse comes in two flavors: FlatParse.Basic and FlatParse.Stateful. Both support a custom error type. Also, both come in three modes, where we can respectively run IO actions, ST actions, or no side effects. The modes are selected by a state token type parameter on the parser types.

  • FlatParse.Basic only supports the above features. If you don't need indentation parsing, this is sufficient.
  • FlatParse.Stateful additionally supports a built-in Int worth of internal state and an additional custom reader environment. This can support a wide range of indentation parsing features. There is a slight overhead in performance and code size compared to Basic. However, in small parsers and microbenchmarks the difference between Basic and Stateful is often reduced to near zero by GHC and/or LLVM optimization.


Informative tutorials are work in progress. See src/FlatParse/Examples for a lexer/parser example with acceptably good error messages.


Pull requests are welcome. I'm fairly quick to add PR authors as collaborators.

Some benchmarks

Execution times below. See source code in bench. Compiled with GHC 9.4.4 -O2 -fllvm. Executed on Intel 1165G7 CPU at 28W power draw. Uses nightly-2023-02-06 Stackage snapshot for the involved packages.

benchmark runtime
sexp/fpbasic 1.93 ms
sexp/fpstateful 2.00 ms
sexp/attoparsec 21.82 ms
sexp/megaparsec 59.60 ms
sexp/parsec 79.81 ms
long keyword/fpbasic 0.1 ms
long keyword/fpstateful 0.1 ms
long keyword/attoparsec 2.43 ms
long keyword/megaparsec 5.2 ms
long keyword/parsec 10.02 ms
numeral csv/fpbasic 0.72 ms
numeral csv/fpstateful 0.56 ms
numeral csv/attoparsec 10.52 ms
numeral csv/megaparsec 19.77 ms
numeral csv/parsec 26.46 ms

Object file sizes for each module containing the s-exp, long keyword and numeral csv benchmarks.

library object file size (bytes)
fpbasic 20656
fpstateful 26664
attoparsec 69384
megaparsec 226232
parsec 117696