Type-driven code generation for OCaml >=4.02
OCaml Makefile



deriving is a library simplifying type-driven code generation on OCaml >=4.02.

deriving includes a set of useful plugins: show, eq, ord, enum, iter, map, fold, make, yojson, protobuf.

Sponsored by Evil Martians.


deriving can be installed via OPAM:

opam install ppx_deriving

Buildsystem integration

To use deriving, only one modification is needed: you need to require via ocamlfind the package corresponding to the deriving plugin. This will both engage the syntax extension and link in the runtime components of the deriving plugin, if any.

For example, if you are using ocamlbuild, add the following to _tags to use the default deriving plugins:

<src/*>: package(ppx_deriving.std)

If you are using another buildsystem, just make sure it passes -package ppx_deriving.whatever to ocamlfind.


From a user's perspective, deriving is triggered by a [@@deriving plugin] annotation attached to a type declaration in structure or signature:

type point2d = float * float
[@@deriving show]

It's possible to invoke several plugins by separating their names with commas:

type point3d = float * float * float
[@@deriving show, eq]

It's possible to pass options to a plugin by appending a record to plugin's name:

type t = string
[@@deriving yojson { strict = true }]

It's possible to make deriving ignore a missing plugin rather than raising an error by passing an optional = true option, for example, to enable conditional compilation:

type addr = string * int
[@@deriving yojson { optional = true }]

It's also possible for many plugins to derive a function directly from a type, without declaring it first.

open OUnit2
let test_list_sort ctxt =
  let sort = List.sort [%derive.ord: int * int] in
  assert_equal ~printer:[%derive.show: (int * int) list]
               [(1,1);(2,0);(3,5)] (sort [(2,0);(3,5);(1,1)])

The [%derive.x:] syntax can be shortened to [%x:], given that the deriver x exists and the payload is a type. If these conditions are not satisfied, the extension node will be left uninterpreted to minimize potential conflicts with other rewriters.

Working with existing types

At first, it may look like deriving requires complete control of the type declaration. However, a lesser-known OCaml feature allows to derive functions for any existing type. Using Pervasives.fpclass as an example, show can be derived as follows:

# module M = struct
  type myfpclass = fpclass = FP_normal | FP_subnormal | FP_zero | FP_infinite | FP_nan
  [@@deriving show]
module M :
    type myfpclass =
      fpclass =
      | FP_subnormal
      | FP_zero
      | FP_infinite
      | FP_nan
    val pp_myfpclass : Format.formatter -> fpclass -> unit
    val show_myfpclass : fpclass -> bytes
# M.show_myfpclass FP_normal;;
- : bytes = "FP_normal"

The module is used to demonstrate that show_myfpclass really accepts Pervasives.fpclass, and not just M.myfpclass.

The need to repeat the type definition may look tedious, but consider this: if the definition was automatically imported from the declaration point, how would you attach attributes to refine the behavior of the deriving plugin?

Nevertheless, for the case where no attributes need to be attached, it is possible to use ppx_import to automatically pull in the type definition.

Plugin conventions

It is expected that all deriving plugins will follow the same conventions, thus simplifying usage.

  • By default, the functions generated by a plugin for a type foo are called fn_foo or foo_fn. However, if the type is called type t, the function will be named foo. The defaults can be overridden by an affix = true|false plugin option.

  • There may be additional attributes attached to the AST. In case of a plugin named eq and attributes named compare and skip, the plugin must recognize all of compare, skip, eq.compare, eq.skip, deriving.eq.compare and deriving.eq.skip annotations. However, if it detects that at least one namespaced (e.g. eq.compare or deriving.eq.compare) attribute is present, it must not look at any attributes located within a different namespace. As a result, different ppx rewriters can avoid interference even if the attribute names they use overlap.

  • A typical plugin should handle tuples, records, normal and polymorphic variants; builtin types: int, int32, int64, nativeint, float, bool, char, string, bytes, ref, list, array, option, lazy_t and their Mod.t aliases; Result.result available since 4.03 or in the result opam package; abstract types; and _. For builtin types, it should have customizable, sensible default behavior. This default behavior should not be used if a type has a [@nobuiltin] attribute attached to it, and the type should be treated as abstract. For abstract types, it should expect to find the functions it would derive itself for that type.

  • If a type is parametric, the generated functions accept an argument for every type variable before all other arguments.

Plugin: show

show derives a function that inspects a value; that is, pretty-prints it with OCaml syntax. However, show offers more insight into the structure of values than the Obj-based pretty printers (e.g. Printexc), and more flexibility than the toplevel printer.

# type t = [ `A | `B of int ] [@@deriving show];;
type t = [ `A | `B of i ]
val pp : Format.formatter -> [< `A | `B of i ] -> unit = <fun>
val show : [< `A | `B of i ] -> bytes = <fun>
# show (`B 1);;
- : bytes = "`B (1)"

For an abstract type ty, show expects to find a pp_ty function in the corresponding module.

show allows to specify custom formatters for types to override default behavior. A formatter for type t has a type Format.formatter -> t -> unit:

# type file = {
  name : string;
  perm : int     [@printer fun fmt -> fprintf fmt "0o%03o"];
} [@@deriving show];;
# show_file { name = "dir"; perm = 0o755 };;
- : bytes = "{ name = \"dir\"; perm = 0o755 }"

It is also possible to use [@polyprinter]. The difference is that for a type int list, [@printer] should have a signature formatter -> int list -> unit, and for [@polyprinter] it's ('a -> formatter -> unit) -> formatter -> 'a list -> unit.

[@opaque] is a shorthand for [@printer fun fmt _ -> Format.pp_print_string fmt "<opaque>"].

The function fprintf is locally defined in the printer.

By default all constructors are printed with prefix which is dot-separated filename and module path. For example

# module X = struct type t = C [@@deriving show] end;;
# X.(show C);;
- : Ppx_deriving_runtime.string = "X.C"

This code will create printers which return the string X.C, X is a module path and C is a constructor name. File's name is omitted in the toplevel. To skip all module paths the one needs to derive show with option with_path (which defaults to true)

# module X = struct type t = C [@@deriving show { with_path = false }] end;;
# X.(show C);;
- : Ppx_deriving_runtime.string = "C"

Plugins: eq and ord

eq derives a function comparing values by semantic equality; structural or physical depending on context. ord derives a function defining a total order for values, returning a negative value if lower, 0 if equal or a positive value is greater. They're similar to Pervasives.(=) and Pervasives.compare, but are faster, allow to customize the comparison rules, and never raise at runtime. eq and ord are short-circuiting.

# type t = [ `A | `B of int ] [@@deriving eq, ord];;
type t = [ `A | `B of int ]
val equal : [> `A | `B of int ] -> [> `A | `B of int ] -> bool = <fun>
val compare : [ `A | `B of int ] -> [ `A | `B of int ] -> int = <fun>
# equal `A `A;;
- : bool = true
# equal `A (`B 1);;
- : bool = false
# compare `A `A;;
- : int = 0
# compare (`B 1) (`B 2);;
- : int = -1

For variants, ord uses the definition order. For builtin types, properly monomorphized (=) is used for eq, or corresponding Mod.compare function (e.g. String.compare for string) for ord. For an abstract type ty, eq and ord expect to find an equal_ty or compare_ty function in the corresponding module.

eq and ord allow to specify custom comparison functions for types to override default behavior. A comparator for type t has a type t -> t -> bool for eq or t -> t -> int for ord. If an ord comparator returns a value outside -1..1 range, the behavior is unspecified.

# type file = {
  name : string [@equal fun a b -> String.(lowercase a = lowercase b)];
  perm : int    [@compare fun a b -> compare b a]
} [@@deriving eq, ord];;
type file = { name : bytes; perm : int; }
val equal_file : file -> file -> bool = <fun>
val compare_file : file -> file -> int = <fun>
# equal_file { name = "foo"; perm = 0o644 } { name = "Foo"; perm = 0o644 };;
- : bool = true
# compare_file { name = "a"; perm = 0o755 } { name = "a"; perm = 0o644 };;
- : int = -1

Plugin: enum

enum is a plugin that treats variants with argument-less constructors as enumerations with an integer value assigned to every constructor. enum derives functions to convert the variants to and from integers, and minimal and maximal integer value.

# type insn = Const | Push | Pop | Add [@@deriving enum];;
type insn = Const | Push | Pop | Add
val insn_to_enum : insn -> int = <fun>
val insn_of_enum : int -> insn option = <fun>
val min_insn : int = 0
val max_insn : int = 3
# insn_to_enum Pop;;
- : int = 2
# insn_of_enum 3;;
- : insn option = Some Add

By default, the integer value associated is 0 for lexically first constructor, and increases by one for every next one. It is possible to set the value explicitly with [@value 42]; it will keep increasing from the specified value.

Plugins: iter, map and fold

iter, map and fold are three closely related plugins that generate code for traversing polymorphic data structures in lexical order and applying a user-specified action to all values corresponding to type variables.

# type 'a btree = Node of 'a btree * 'a * 'a btree | Leaf [@@deriving iter, map, fold];;
type 'a btree = Node of 'a btree * 'a * 'a btree | Leaf
val iter_btree : ('a -> unit) -> 'a btree -> unit = <fun>
val map_btree : ('a -> 'b) -> 'a btree -> 'b btree = <fun>
val fold_btree : ('a -> 'b -> 'a) -> 'a -> 'b btree -> 'a = <fun>
# let tree = (Node (Node (Leaf, 0, Leaf), 1, Node (Leaf, 2, Leaf)));;
val tree : int btree = Node (Node (Leaf, 0, Leaf), 1, Node (Leaf, 2, Leaf))
# iter_btree (Printf.printf "%d\n") tree;;
- : unit = ()
# map_btree ((+) 1) tree;;
- : int btree = Node (Node (Leaf, 1, Leaf), 2, Node (Leaf, 3, Leaf))
# fold_btree (+) 0 tree;;
- : int = 3

Plugin: make

make is a plugin that generates record constructors. Given a record, a function is generated that accepts all fields as labelled arguments and (); alternatively, if one field is specified as [@main], it is accepted last. The fields which have a default value (fields of types 'a option, 'a list, and fields with [@default] annotation) are mapped to optional arguments; the rest are mandatory. A field of form xs: ('a * 'a list) [@split] corresponds to two arguments: mandatory argument x and optional argument xs with types 'a and 'a list correspondingly.

type record = {
  opt  : int option;
  lst  : int list;
  def  : int [@default 42];
  args : (int * int list) [@split];
  norm : int;
} [@@deriving make];;
val make_record :
  ?opt:int ->
  ?lst:int list ->
  ?def:int ->
  arg:int ->
  ?args:int list ->
  norm:int ->
  unit ->

Building ppx drivers

By default, deriving dynlinks every plugin, whether invoked as a part of a batch compilation or from the toplevel. If this is unsuitable for you for some reason, it is possible to precompile a ppx rewriter executable that includes several deriving plugins:

$ ocamlfind opt -predicates ppx_driver -package ppx_deriving_foo -package ppx_deriving_bar \
                -package ppx_deriving.main -linkpkg -linkall -o ppx_driver

Currently, the resulting ppx driver still depends on Dynlink as well as retains the ability to load more plugins.

Developing plugins

This section only explains the tooling and best practices. Anyone aiming to implement their own deriving plugin is encouraged to explore the existing ones, e.g. eq or show.

Tooling and environment

A deriving plugin is packaged as a Findlib library; this library should include a peculiar META file. As an example, let's take a look at a description of a yojson plugin:

version = "1.0"
description = "[@@deriving yojson]"
exists_if = "ppx_deriving_yojson.cma"
# The following part affects batch compilation and toplevel.
# The plugin package may require any runtime component it needs.
requires(-ppx_driver) = "ppx_deriving yojson"
ppxopt(-ppx_driver) = "ppx_deriving,./ppx_deriving_yojson.cma"
# The following part affects ppx driver compilation.
requires(ppx_driver) = "ppx_deriving.api"
archive(ppx_driver, byte) = "ppx_deriving_yojson.cma"
archive(ppx_driver, native) = "ppx_deriving_yojson.cmxa"

The module(s) provided by the package in the ppxopt variable must register the derivers using Ppx_deriving.register "foo" during loading. Any number of derivers may be registered; careful registration would allow a yojson deriver to support all three of [@@deriving yojson], [@@deriving of_yojson] and [@@deriving to_yojson], as well as [%derive.of_yojson:] and [%derive.to_yojson:].

It is possible to test the plugin without installing it by instructing deriving to load it directly; the compiler should be invoked as ocamlfind c -package ppx_deriving -ppxopt ppx_deriving,src/ppx_deriving_foo.cma .... The file extension is replaced with .cmxs automatically for native builds. This can be integrated with buildsystem, e.g. for ocamlbuild:

let () = dispatch (
  | After_rules ->
    (* Assuming files tagged with deriving_foo are already tagged with
       package(ppx_deriving) or anything that uses it, e.g. package(ppx_deriving.std). *)
    flag ["ocaml"; "compile"; "deriving_foo"] &
      S[A"-ppxopt"; A"ppx_deriving,src/ppx_deriving_foo.cma"]
  | _ -> ()

Alternatively, you can quickly check the code generated by a ppx rewriter packaged with ocamlfind by running the toplevel as ocaml -dsource or utop -dsource, which will unparse the rewritten syntax tree into OCaml code and print it before executing.

Goals of the API

deriving is a thin wrapper over the ppx rewriter system. Indeed, it includes very little logic; the goal of the project is 1) to provide common reusable abstractions required by most, if not all, deriving plugins, and 2) encourage the deriving plugins to cooperate and to have as consistent user interface as possible.

As such, deriving:

  • Completely defines the syntax of [@@deriving] annotation and unifies the plugin discovery mechanism;
  • Provides an unified, strict option parsing API to plugins;
  • Provides helpers for parsing annotations to ensure that the plugins interoperate with each other and the rest of the ecosystem.

Using the API

Complete API documentation is available online.


A very important aspect of a syntax extension is hygiene. Consider a case where a deriving plugin makes assumptions about the interface provided by the List module: it will normally work as expected, but not in case where someone shadows the List identifier! This happens quite often in the OCaml ecosystem, e.g. the Jane Street [Core] library encourages developers to use open Core.Std.

Additionally, if your deriving plugin inserts user-provided expressions into the generated code, a name you are using internally may accidentally collide with a user-defined name.

With deriving, both of these problems are solved in three easy steps:

  • Create a quoter:

    let quoter = Ppx_deriving.create_quoter () in
  • Pass the user-provided expressions, if any, through the quoter, such as by using a helper function:

    let attr_custom_fn attrs =
      Ppx_deriving.(attrs |> attr ~deriver "custom_fn" |> Arg.(get_attr ~deriver expr)
                          |> quote ~quoter)
  • Wrap the generated code:

    let expr_of_typ typ =
      let quoter = ...
      and expr = ... in
      Ppx_deriving.sanitize ~quoter expr

    If the plugin does not accept user-provided expressions, sanitize expr could be used instead.


The following is a list of tips for developers trying to use the ppx interface:

  • Module paths overwhelm you? Open all of the following modules, they don't conflict with each other: Longident, Location, Asttypes, Parsetree, Ast_helper, Ast_convenience.
  • Need to insert some ASTs? See ppx_metaquot; it is contained in the ppx_tools.metaquot package.
  • Need to display an error? Use Ppx_deriving.raise_errorf ~loc "Cannot derive Foo: (error description)" (doc); keep it clear which deriving plugin raised the error!
  • Need to derive a function name from a type name? Use Ppx_deriving.mangle_type_decl and Ppx_deriving.mangle_lid.
  • Need to fetch an attribute from a node? Use Ppx_deriving.attr ~prefix "foo" nod.nod_attributes (doc); this takes care of interoperability.
  • Put all functions derived from a set of type declarations into a single let rec block; this reflects the always-recursive nature of type definitions.
  • Need to handle polymorphism? Use Ppx_deriving.poly_fun_of_type_decl for derived functions, Ppx_deriving.poly_arrow_of_type_decl for signatures, and Ppx_deriving.poly_apply_of_type_decl for "forwarding" the arguments corresponding to type variables to another generated function.
  • Need to display a full path to a type, e.g. for an error message? Use Ppx_deriving.path_of_type_decl.
  • Need to apply a sequence or a binary operator to variant, tuple or record elements? Use Ppx_deriving.fold_exprs.
  • Don't forget to display an error message if your plugin doesn't parse any options.


deriving is distributed under the terms of MIT license.