package duff

  1. Overview
  2. Docs
Rabin's fingerprint and diff algorithm in OCaml


Dune Dependency






This library provides a pure implementation of Rabin's fingerprint and diff algorithm in OCaml.

It follows libXdiff C library. It is used by ocaml-git project.

Published: 10 Apr 2022


Duff – libXdiff implementation in OCaml

Duff is a little library to implement libXdiff in OCaml. This library is a part of the ocaml-git project. This code is a translation of diff-delta.c available on the git project in OCaml. So, it respects some git's constraints unlike libXdiff.


This library let the user to calculate an index from a source (a hash-table) which can be computed with a blob. Then, from index (which represents your source) and a blob, we generate a list of Copy and Insert elements.

  • Copy (off, len) means to take a slice of len bytes from your source at off (absolute offset) and copy it.

  • Insert (off, len) means to store a slice of len bytes from your blob at off (absolute offset) and copy it.

From this information, we can have a tiny representation of your blob which can be reconstruct with your source. The goal is to store Copy opcode with off and len, and Insert opcode which contains a slice of your blob.

Finally, to produce a PACK file in git or ocaml-git, we use this algorithm and this representation to optimize storage of your blobs (cf. git gc).


You can see an example of duff in bin directory. It's an executable to represent a thin representation of your file. Then, you can reconstruct it with patch sub-command.

This is an example to use duff:

$ ./duff.exe diff source target > target.xduff
$ ./duff.exe patch source < target.xduff >
$ diff target
$ echo $?

The internal format used is close to what git does internally (without zlib layer). However, it does not correspond to an official format. The binary is not optimized to be used in a production environment but feedback and improvement on it are welcome.


Because this project is used by ocaml-git, we have some limitations:

  • We compute at most 0xFFFFFFFE bytes from source

  • An insert block can not be bigger than 0x10000 bytes

For example, libXdiff computes a bigger source than this implementation. Then, limitation about insert block depends on the PACK (git) file format. So, don't ask me to compute bigger source or merge and produce bigger insert block - these constraints is outside the scope of this library.

From this limitation, Copy opcode have an offset between 0x0 and 0xFFFFFFE and off + len is lower than 0xFFFFFFFE.


We provide a fuzzer to randomly test this library. Currently (4/9/2018), afl-fuzz did not find any bugs and it computed 67.7k cycles (117 paths).

Dependencies (3)

  1. fmt >= "0.8.7"
  2. dune >= "2.0.0"
  3. ocaml >= "4.07.0"

Dev Dependencies (4)

  1. crowbar with-test
  2. hxd with-test & >= "0.3.1"
  3. bigstringaf with-test
  4. alcotest with-test

Used by (1)

  1. carton




Innovation. Community. Security.