hoppy-docs-0.3.2: C++ FFI generator - Documentation

Safe HaskellNone




The Hoppy User's Guide



Hoppy is a foreign function interface (FFI) generator for interfacing Haskell with C++. It lets developers specify C++ interfaces in pure Haskell, and generates code to expose that functionality to Haskell. Hoppy is made up of a few different packages that provide interface definition data structures and code generators, some runtime support for Haskell bindings, and interface definitions for the C++ standard library.

Bindings using Hoppy have three Cabal packages:

  • A Haskell generator program (in /myproject-generator) that knows the interface definition and generates code for the next two parts.
  • A C++ library (in /myproject-cpp) that gets compiled into a shared object containing the C++ half of the bindings.
  • A Haskell library (in /myproject) that links against the C++ library and exposes the bindings.

The path names are suggested subdirectories of a project, and are used in this document, but are not required. Only the latter two items need to be packaged and distributed to users of the binding (plus Hoppy itself which is a dependency of the generated bindings).

Getting started

This section provides a gentle introduction to working with Hoppy.

Project setup

To set up a new Hoppy project, it's recommended to start with the project in the example/ directory as a base. It is a minimal project that defines a C++ function to reverse a std::string, exposes that to Haskell via a library, and provides a demo program that uses the library. The example/install.sh script simply compiles and installs the generator, C++, and Haskell packages in turn.

The generator package specifies the C++ interface to be exposed, using the functions and data types described in the rest of this section.

The C++ package is a mostly empty, primarily containing a Setup.hs file that invokes Hoppy build hooks, and the C++ code we're binding to. When building this package, Hoppy generates some C++ code and then relies on a Makefile we provide for linking it together with any code we provided (see example/example-cpp/cpp/Makefile). If you are relying on a system library, you can link to it in the Makefile.

The Haskell package is even more empty than the C++ one. It contains a similar Setup.hs to invoke Hoppy. Nothing else is included in the package's library, although you are free to add your own Haskell modules. The executable ties everything together by calling the C++ code. It reverses the characters of each input line it sees.

To publish this project, one would upload all three packages to Hackage. (But make sure to rename it first!)

A first binding

A complete C++ API is specified using Haskell data structures in Foreign.Hoppy.Generator.Spec. At the top level is the Interface type. An interface contains Modules which correspond to portions of functionality of their interface; that is, collections of classes, functions, files, etc. Functionality can be grouped arbitrarily into modules and doesn't have to follow the structure of existing C++ files. Modules contain Exports which refer to concrete bound entities (Functions, Classes, etc.).

For starters, we will look at a single class. Let's write a binding for std::string. An initial version could start as follows.

import Foreign.Hoppy.Generator.Spec

c_string :: Class
c_string =
  addReqIncludes [includeStd "string"] $
  makeClass (ident1 "std" "string") (Just $ toExtName "StdString")
  [ mkCtor "new" []
  , mkConstMethod "at" [intT] charT
  , mkConstMethod "string" [] sizeT

There is quite a bit to look at here, so let's work through it.

First, everything that can be exported has two names besides the name used for the Haskell binding (c_string above). Identifiers are used to specify the qualified C++ names of exports, including namespaces and template arguments. For this example, our identifier is std::string, which we specify with the ident1 call above. The number indicates the number of leading namespace components. ident can be used for top-level entities.

Exported entities also each have an external name that uniquely identifies it within an interface. This name can be different from the name of the C++ entity the export is referring to. An external name is munged by the code generators and must be a valid identifier in all languages a set of bindings will use, so it is restricted to characters in the range [a-zA-Z0-9_], and must start with an alphabetic character. Character case in external names will be preserved as much as possible in generated code, although case conversions are sometimes necessary (e.g. Haskell requiring identifiers to begin with upper or lower case characters). In the example, the toExtName call specifies an explicit external name for the class. Nothing may be provided to automatically derive an external name from the given identifier. The derived name is based on the last component of the identifier, which in this case is just string. Converting this to a Haskell type name gives String, which collides with the built-in string type, so we give an explicit external name instead.

The third argument to makeClass is a list of superclasses. std::string does not derive from any classes so we leave this empty in the example. When specifying interfaces in Hoppy, only publicly accessible components need to (and in fact, can) be referenced by Hoppy: public base classes, public methods and variables, but never protected or private entities.

The final argument to makeClass is a list of entities within the class. Here we can specify constructors, methods, and variables that the class exposes, via the ClassEntity type. There are a few sets of methods for building class entities:

  • Basic forms: makeCtor, makeMethod, and makeClassVariable are the core functions for building class entities. These are fully general, and take parameters both for the C++ and external names, as well as staticness, constness, etc. There is also makeFnMethod for defining a method that is actually backed by a C++ function, not a method of the class; this can be used when manual wrapping of a method is required, to wrap a method with a function but make it look like a method.
  • Convenience forms: mkCtor, mkMethod, mkConstMethod, mkStaticMethod, mkProp, mkBoolIsProp, mkBoolHasProp, mkStaticProp, and mkClassVariable only take the C++ name, and derive the external name from it, as well as assuming other parameters (staticness, etc.). These are what you typically use, unless you need overloading, in which case use the overloading forms below.
  • Overloading forms: mkMethod', mkConstMethod', and mkStaticMethod' are convenience functions for overloaded methods. Overloading is handled by defining multiple exports all pointing to a single C++ entity. These in turn become separate Haskell functions. Because external names must be unique though, a different external name must be provided for each overloaded form; this is the second argument to these functions.
  • Unwrapped forms: Underscore forms for all of the above are provided as well (e.g. mkMethod_ and mkMethod'_) that return the actual object they create (Method, Ctor, ClassVariable) instead of wrapping the result in a ClassEntity as is usually desired.

Generated C++ bindings for exported entities usually need #includes in order to access those entities. This is done with Include and Reqs types. When defining bindings, all exportable types have an instance of the HasReqs typeclass, and addReqIncludes can be used to add includes. includeStd produces #include <...> statements, and includeLocal produces #include "...".

This use of addReqIncludes also indicates a common pattern for writing class bindings. After constructing a Class with makeClass, there are a number of functions that modify the class definition in various ways. These functions' types always end in ... -> Class -> Class, so that they can be chained easily. Among others, these functions include:

The main point to using all of these functions is to chain them on the result of makeClass, but to bind the Haskell binding to final resulting value. For instance, if we have the following class:

c_NotFoundException :: Class
c_NotFoundException =
  addReqIncludes [includeStd "exceptions.hpp"] $
  classMakeException $
  makeClass (ident "NotFoundException") Nothing []
  [ mkCtor "newCopy" [objT c_NotFoundException]

Then, addReqIncludes and classMakeException modify the class object, and the constructor definition makes use of the resulting object. This works as intended. In some cases the order of modifiers is important -- for example, marking a class as an exception class requires that there be a copy constructor defined beforehand -- but usually order of modification does not matter.

Another point of note is the c_ prefix used on these two classes. A suggested naming convention for entities is:

Hoppy follows this convention, but you are not required to in your own bindings. It enables the use of proper casing on the actual entity name, and avoids collision with existing Haskell names.

Given all this, we can improve our std::string binding. The at() method provides a non-const overload that returns a reference to a requested character. Let's have two versions of at(), as well as expose the fact that std::string is assignable, comparable, copyable, and equatable, with a second version:

c_string :: Class
c_string =
  addReqIncludes [includeStd "string"] $
  classAddFeatures [Assignable, Comparable, Copyable, Equatable] $
  makeClass (ident1 "std" "string") (Just $ toExtName "StdString")
  [ mkCtor "new" []
  , mkConstMethod' "at" "at" [intT] $ refT charT
  , mkConstMethod' "at" "get" [intT] charT
  , mkConstMethod "string" [] sizeT


Let's take a break from std::string for a moment and talk about how we represent data types in Hoppy.

All C++ types are represented with the Type data type, values of which are in the Foreign.Hoppy.Generator.Types module. This includes primitive numeric types, object types, function types, pointers and references, void, the const qualifier, etc.

A Hoppy Type value has a corresponding C++ type, and also what we refer to as a C type, and possibly a Haskell type. The C++ type is, of course, whatever C++ type the Type structure represents. The Haskell type is the Haskell data type that the C++ value gets converted to and from when using the bindings from Haskell code. The C type is the C data type that gets passed over the gateway between C++ and Haskell, because these types are the common denominator between C++ and Haskell. When passing values back and forth between C++ and Haskell, generally, primitive types are converted to equivalent types on both ends, and pointer types in C++ are represented by corresponding pointer types in Haskell. Other types have special rules. Conversion code is generated on both sides of the gateway to perform the necessary conversions.

For numbers, the Haskell FFI provides numeric types in Foreign.C for interfacing with C directly. Hoppy maps C++ numbers to these types, with the exception of bool, int, float, and double, which map to their native Haskell equivalents instead, for convenience (Bool, Int, Float, Double).

The mapping between C++ and Haskell types is as follows. Types that we haven't covered are listed here for completeness. Some of these aren't really types, but are modifiers that are only useful in certain situations.

  • voidT: C++ uses void. The C type is unspecified. Haskell uses ().
  • boolT, intT, floatT, doubleT: C++ and C use bool, int, float, double. Haskell uses Bool, Int, Float, Double.
  • charT, ucharT, and other primitive numeric types: Identical C++ and C types; Haskell uses built-in foreign data types. See Foreign.Hoppy.Generator.Types for more info.
  • enumT: C++ uses the enum type. C uses int. Haskell uses a generated data type for the enum.
  • bitspaceT: C++ and C use a specified type, usually int. Haskell uses a generated data type for the bitspace.
  • ptrT: C++ and C use the raw pointer type. The Haskell type depends on the pointed-to type. If it's an object pointer, then Haskell uses the generated handle data type for the class. If it's a function pointer, then Haskell uses a FunPtr of the Haskell function in IO with C types. Otherwise, Haskell uses a Ptr to the C type of the pointed type.
  • refT: C++ uses the reference type. Everything else is the same as ptrT.
  • fnT: Function types aren't useful raw, and need to be wrapped in ptrT to be useable in parameter and return types.
  • callbackT: C++ uses the callback's generated functor class. C uses an internal implementation pointer. Haskell uses a function in IO of the Haskell types of the parameters and return types.
  • objT: For objects passed by value (not wrapped in ptrT or refT), C++ uses the object type (or a const reference), C uses a pointer to the object, and Haskell uses a type configured for the class.
  • objToHeapT: Copies objects to the heap.
  • toGcT: Assigns objects to the garbage collector.
  • constT: Wraps types in const.

Many of these types (enumerations, object types, functions, callbacks) are discussed in later sections.

Wrapping up the string binding

Let's package up the std::string binding to get a buildable example. We define a module to export the class, and an interface to collect our modules.

import Foreign.Hoppy.Generator.Main
import Foreign.Hoppy.Generator.Spec

main :: IO ()
main = defaultMain interfaceResult

interfaceResult :: Either String Interface
interfaceResult = do
  iface <- interface "example" [mod_string]
  interfaceAddHaskellModuleBase ["Example"] iface

mod_string :: Module
mod_string =
  moduleModify' (makeModule "mystring" "mystring.hpp" "mystring.cpp") $
  moduleAddExports [ExportClass c_string]

Each Module produces separate C++ files to be compiled (a header and a source file), and a separate Haskell module. The Haskell module will be named Example.Mystring, which is the concatenation of the Haskell module base path defined on the interface, and the case-corrected name of the module. A custom module path can replace the default with moduleAddHaskellName.


Let's say we have a C++ function to compute the hypotenuse of a triangle:

double hypotenuse(double x, double y);

To declare a binding for this, we can write:

f_hypotenuse :: Function
f_hypotenuse =
  makeFn "hypotenuse" Nothing Pure [doubleT, doubleT] doubleT

Like the first two arguments to makeClass, the first two arguments to makeFn are its C++ name, and an optional external name that will be derived from the C++ name if absent.

The third argument indicates whether or not the function is pure. Nonpure generates a function in IO, and Pure generates a pure function (using unsafePerformIO internally). Most functions should be marked as nonpure, either because they have side effects or because you want to control the order of their execution with respect to other functions with side effects, but in this case (assuming the implementation of hypontenuse() just performs the calculation) we can mark it as pure.

Finally, there are the parameter type list and return type.

In Haskell, this will generate the following function, no surprises here:

hypotenuse :: Double -> Double -> Double


There is a lot more to using classes than there is with functions, so we'll spend the next several sections discussing how objects work.

Generated bindings

Now that we've seen what is generated for a function, let's see what is generated for our std::string binding. Here is the definition again:

c_string :: Class
c_string =
  addReqIncludes [includeStd "string"] $
  classAddFeatures [Assignable, Comparable, Copyable, Equatable] $
  makeClass (ident1 "std" "string") (Just $ toExtName "StdString")
  [ mkCtor "new" []
  , mkConstMethod' "at" "at" [intT] $ refT charT
  , mkConstMethod' "at" "get" [intT] charT
  , mkConstMethod "string" [] sizeT

The first thing that is generated are data types called handles that represent pointers to instances. For each class, there are two handle types, a const and a nonconst version. These are distinct types so that we can enforce const-safety. Functions for casting constness are created as well.

data StdString
data StdStringConst

castStdStringToConst :: StdString -> StdStringConst
castStdStringToNonconst :: StdStringConst -> StdString

instance Eq StdString
instance Eq StdStringConst
instance Ord StdString
instance Ord StdStringConst
instance Show StdString
instance Show StdStringConst

These instances operate on the underlying pointer. There is also a Foreign.Hoppy.Runtime module in the hoppy-runtime package that holds various types and code needed at runtime, and generated bindings make extensive use of this package. There are a number of class-related typeclasses there that also contain these types.

instance CppPtr StdString
instance CppPtr StdStringConst
instance Deletable StdString
instance Deletable StdStringConst
instance Copyable StdString StdString
instance Copyable StdStringConst StdString

CppPtr is a typeclass for all handle types. Unless a class is marked as having a private destructor (classSetDtorPrivate), you will be able to delete it with delete; deletability also enables garbage collection, discussed later. Finally, if a class has a copy constructor defined for it (either manually with mkCtor or via the Copyable class feature), then it gets instances of Copyable.

Next, a set of typeclasses are generated to hold values representing the class, strings in this case.

class StdStringValue a

class CppPtr this => StdStringConstPtr this where
  toStdStringConst :: this -> StdStringConst

class StdStringConstPtr this => StdStringPtr this where
  toStdString :: this -> StdString

instance {OVERLAPPABLE} StdStringConstPtr a => StdStringValue a  -- (sic)

instance StdStringConstPtr StdStringConst

instance StdStringConstPtr StdString
instance StdStringPtr StdString

instance StdStringValue a => Assignable StdString a

The three typeclasses contain string-like types. StdStringValue is the most general, and contains types that can be represented as a std::string. For now this includes the handle types; we'll see how to make good use of this typeclass in a bit.

The other two typeclasses contain std::string handles. StdStringConstPtr contains all handles that via C++ implicit casting could be converted to a StdStringConst, and likewise for StdStringPtr and StdString. This gives the instances above. The typeclasses' methods can be used to upcast, and optionally add const.

In a hierarchy with derived classes, this would be more complicated. For an arbitrary class, the const typeclass has as superclasses the const typeclasses for all of the C++ class's superclasses (or just CppPtr if this list is empty). The non-const typeclass has as superclasses the non-const typeclasses for all of the C++ class's superclasses, plus the current const typeclass. Instances will be generated for all handles as appropriate. This formalises const handles as parallel to nonconst handles in the hierarchy, but also above.

The overlappable instance just says that all std::string handles, const or nonconst, are std::string values. Because we defined an assignment operator (operator=), we get an Assignable instance, so that if we have a StdStringValue value, we can assign it to a StdString handle by calling that operator.

We can also attempt downcasting using dynamic_cast. If std::string had superclasses, their handles would have instances of the following typeclasses:

class StdStringSuper a where
  downToStdString :: a -> StdString

class StdStringSuperConst a where
  downToStdStringConst :: a -> StdStringConst

Like dynamic_cast, these methods will return nullptr if the given object is not of an appropriate type.

Passing and returning objects

So how do we work with functions that expect objects? Suppose we have the following function:

void reverse(std::string&);

Then if we define:

f_reverse :: Function
f_reverse = makeFn "reverse" Nothing Nonpure [refT $ objT c_string] voidT

We'll get the following binding:

reverse :: StdStringPtr this => this -> IO ()

Unlike primitive types, parameters for object types use typeclass constraints to accept a range of types, so that subclasses' handles can also be passed.

There are five different object types that can be used in parameter and return types:

  1. (refT $ constT $ objT c_string) is const std::string&.
  2. (refT $ objT c_string) is std::string&.
  3. (ptrT $ constT $ objT c_string) is const std::string*.
  4. (ptrT $ objT c_string) is std::string*.
  5. (objT c_string) is std::string.

When used as a parameter type, cases 1 and 3 generate a Haskell parameter type of StdStringConstPtr a => a, and cases 2 and 4 generate a Haskell parameter type of StdStringPtr a => a. There is no distinction between references and pointers in Haskell, but if the C++ function expects a mutable object, then so does the Haskell binding. Case 5 generates a Haskell parameter type of StdStringValue a => a.

You should use the type that the actual function expects, with one exception. When the type of a parameter of a C++ function is const C& (case 1), it is recommended to just declare it as C (case 5), since this is shorter and is fully equivalent when passing a handle (but is also open to the object having a conversion added in the future). This exception does not apply to function return types.

When returning values from a function, cases 1 and 3 return a StdStringConst, cases 2 and 4 return a StdString, and what case 5 returns depends on the class's defined conversion (discussed later).

Garbage collection

By default, object lifetimes are managed manually. They are created with constructors and eventually destroyed with delete (or explicitly scoped via withScopedPtr). Alternatively, ownership of objects can be passed to the Haskell garbage collector, to be deleted when no references are left from Haskell memory to the object.

This is tracked internally by handles. Handles can either be unmanaged (as they are initially) or managed (by the collector). For simplicity, this is not reflected in a handle's type. A managed handle can be created from an unmanaged handle by calling toGc. This assigns the object to be tracked by the collector, and delete will be called on the object once no more handles are left pointing to it. toGc returns a new handle, and existing unmanaged handles for the object should no longer be used, since they will be dangling pointers once the object is destroyed. There are some points of caution around using this function that are worth knowing about; see the function documentation for more info.

If you want to pass an object to the collector immediately upon creation, chain its constructor call with (toGc =<<). This is not done by default because we don't support revoking the collector's watch over an object, and there are times when you want to work with manually managed objects.

toGcT may be used when defining a function to make an object being passed into Haskell be managed by the garbage collector explicitly. But rather than using toGcT with value objects, it's better to use classSetConversionToGc. There is also a lesser-used objToHeapT for copying a temporary onto the heap for Haskell code to manage without giving it to the garbage collector (and a corresponding classSetConversionToHeap).


Object pointer and reference types (refT . objT, ptrT . constT . objT, etc.) use handles in Haskell to refer to objects living in C++ memory. By-value object types (not pointers or references, just the by-value object types, objT directly) are treated differently. When an object is taken or returned by value, this typically indicates a lightweight or short-lived, easily copied object, and Hoppy provides a few different behaviours to choose from to handle these. For instance, you can choose to use objects returned by value via garbage-collected handles, or you can define automatic conversions to and from a Haskell type.

This is all controlled by the ClassConversion object that lives on each Class. Within, ClassHaskellConversion has three fields that control by-value behaviour:

The Haskell type here is the same as the Haskell type for objT mentioned earlier, in the section on types. With the other two fields, we introduce the concept of class convertibility. A class can be convertible in zero, one, or both directions to and from C++, depending on which of the latter two fields above are specified. Both of the conversion logic fields, if present, require that a Haskell type be specified as well.

All of these fields include Generator in their types. This is a Haskell code generation monad that supports line-based output and also manages module imports and exports. Refer to ClassHaskellConversion and Foreign.Hoppy.Generator.Language.Haskell for more detail on writing custom conversions.

For ease of use with std::string, we want strings in C++ to convert to strings in Haskell and vice versa, and we'll let the garbage collector handle the Haskell strings. We can do this by specifying the Haskell type as String, and writing conversions:

c_string :: Class
c_string =
  addReqIncludes [includeStd "string"] $
  classAddFeatures [Assignable, Comparable, Copyable, Equatable] $
    { classHaskellConversionType = Just $ do
        addImports $ hsWholeModuleImport "Prelude"
        return $ HsTyCon $ UnQual $ HsIdent "String"
    , classHaskellConversionToCppFn = Just $ do
        addImports $ mconcat [hsWholeModuleImport "Prelude", hsWholeModuleImport "Foreign.C"]
        sayLn "flip withCString stdString_newFromCString"
    , classHaskellConversionFromCppFn = Just $ do
        addImports $ mconcat [hsImport1 "Control.Monad" "(<=<)", hsWholeModuleImport "Foreign.C"]
        sayLn "peekCString <=< stdString_c_str"
    } $
  makeClass (ident1 "std" "string") (Just $ toExtName "StdString")
  [ mkCtor "new" []
  , mkCtor "newFromCString" [ptrT $ constT charT]
  , mkConstMethod' "at" "at" [intT] $ refT charT
  , mkConstMethod' "at" "get" [intT] charT
  , mkConstMethod "c_str" [] $ ptrT $ constT charT
  , mkConstMethod "string" [] sizeT

First we add newFromCString and c_str which we need to write the conversions. Whenever writing Haskell generator code, you need to import whatever types you want to use, in each of these individual actions, because they are used in different places during generation. The Haskell code that Hoppy writes only uses qualified imports, under aliases that begin with Hoppy (with the exception of some standard infix operators which it imports unqualified). It does this to avoid conflicts between generated names and built-in ones, so you could use an external name of String if you wanted to. You are welcome to import modules wholesale, as the example here does.

stdString_newFromCString and stdString_c_str used in the conversion code are the generated methods. Conversion code is produced in the same Haskell module as the rest of the Haskell code for a class, so the class's methods are always present. The to-C++ conversion takes a Haskell String, converts it to a temporary C string, and creates and returns a std::string from that. Conversely, the from-C++ conversion takes a const std::string, grabs its C string, and creates a Haskell string from it.

So how does this let us use strings more easily? If we look back to what was generated for our string class earlier, we had a StdStringValue typeclass that contained general std::string-like values. When we define a conversion to C++, then the Haskell String also becomes an instance of this typeclass. So whenever a C++ function takes an argument of std::string, const std::string&, or const std::string*, we can now pass a String and have it work automatically. When a C++ function returns a std::string, it will now automatically be converted to a String. For example, given a function:

std::string reverse(const std::string&);

And a method binding:

makeFn "reverse" Nothing Nonpure [objT c_string] $ objT c_string

We get the following Haskell function and instance:

reverse :: StdStringValue a => a -> IO String

instance {OVERLAPPING} StdStringValue String  -- (sic)

Note that when returning an objT from a function, there is no choice whether the Haskell type or a handle is returned; the conversion is always performed.

If we just wanted to use the StdString handle as the Haskell type for the class, but have objects returned to Haskell be garbage collected, then in our class definition we could use classSetConversionToGc instead of classSetHaskellConversion. This would change reverse to have the following signature, and no StdStringValue instance would be generated for String:

reverse :: StdStringValue a => a -> IO StdString

API versioning

Hoppy provides API versioning support in the Foreign.Hoppy.Generator.Version module. This is mainly done with the collect, just, and test functions, which are a simple wrapper around collecting a list of optional values:

type Filtered = Maybe

collect :: [Filtered a] -> [a]
none :: Filtered a
just :: a -> Filtered a
test :: Bool -> a -> Filtered a

These can be used anywhere a list is provided to Hoppy to filter based on some criteria. For example, the std::pair binding in hoppy-std defines a swap method conditionally based on the version of the C++ standard being used:

c_pair :: Class
c_pair =
  ... $
  makeClass ... $
  [ ...
  , test (activeCppVersion >= Cpp2011) $ mkMethod "swap" [refT $ objT c_pair] voidT

It is up to you to decide how to pass to your feature flags into your generator (whether by environment variables, Cabal flags, etc.). If you use environment variables, you will need to use unsafePerformIO to access them, since binding definitions don't have access to IO. For an example, see the implementation of activeCppVersion.


This section describes the behaviour of the code generators, and documents what they output for each type of export.

The code generators live at Foreign.Hoppy.Generator.Language.<language>. The top-level module for a language is internal to Hoppy and contains the bulk of the generator. General submodules expose functionality that can control generator behaviour.


The C++ code generator generates C++ bindings that other languages' bindings will link against. This generator lives in Foreign.Hoppy.Generator.Language.Cpp, with internal parts in Foreign.Hoppy.Generator.Language.Cpp.Internal.

Module structure

Generated modules consist of a source and a header file. The source file contains all of the bindings for foreign languages to make use of. The header file contains things that may be depended on from other generated modules. Currently this consists only of generated callback classes.

Cycles between generated C++ modules are not supported. This can currently only happen because of #include cycles involving callbacks, since callbacks are the only Exports that can be referenced by other generated C++ code. Also, C++ callbacks that handle exceptions depend on the interface's exception support module (see interfaceExceptionSupportModule).

Object passing

ptrT :: Type -> Type
refT :: Type -> Type
objT :: Class -> Type
constT :: Type -> Type

We consider all of the following cases as passing an object, both into and out of C++, and independently, as an argument and as a return value:

  1. objT _
  2. refT (constT (objT _))
  3. refT (objT _)
  4. ptrT (constT (objT _))
  5. ptrT (objT _)

The first is equivalent to constT (objT _). When passing an argument from a foreign language to C++, the first two are equivalent, and it's recommended to use the first, shorter form (T and const T& are functionally equivalent in C++, and are the same as far as what values foreign bindings will accept).

When passing any of the above types as an argument in either direction, an object is passed between C++ and a foreign language via a pointer. Cases 1, 2, and 4 are passed as const pointers. For a foreign language passing a objT _ to C++, this means converting a foreign value to a temporary C++ object. Passing a objT _ argument into or out of C++, the caller always owns the object.

When returning an object, again, pointers are always what is passed across the language boundary in either direction. Returning a objT _ transfers ownership: a C++ function returning a objT _ will copy the object to the heap, and return a pointer to the object which the caller owns; a callback returning a objT _ will internally create a C++ object from a foreign value, and hand that object off to the C++ side (which will return it and free the temporary).

Object lifetimes can be managed by a foreign language's garbage collector. toGcT is a special type that is only allowed in certain forms, and only when passing a value from C++ to a foreign language (i.e. returning from a C++ function, or C++ invoking a foreign callback), to put the object under the collector's management. Only object types are allowed:

  1. toGcT (objT cls)
  2. toGcT (refT (constT (objT cls)))
  3. toGcT (refT (objT cls))
  4. toGcT (ptrT (constT (objT cls)))
  5. toGcT (ptrT (objT cls))

Cases 2-5 are straightforward: the existing object is given to the collector. Case 1 without the toGcT would cause the object to be converted, but instead here the (temporary) object gets copied to the heap, and a managed pointer to the heap object is returned. Case 1 is useful when you want to pass a handle that has a non-trivial C++ representation (so you don't define a conversion for it), but it's still a temporary that you don't want users to have to delete manually.

Objects are always managed manually unless given to a garbage collector. In particular, constructors always return unmanaged pointers. When a managed pointer is passed into C++, that it is managed is lost in the FFI conversion, and if this pointer is then passed back into the foreign language, it will arrive in an unmanaged state (although the object is still managed, and it should not be assigned to the collector a second time).


data Callback = Callback ExtName [Type] Type ...  -- Parameter and return types.

callbackT :: Callback -> Type

We want to call some foreign code from C++. There are two choices for doing so, described below. Declaring a callback provides support for both types of invocation.

Function pointer: Function pointers are expressed with a ptrT (fnT ...) type. Foreign runtimes' FFIs can provide a means for creating raw function pointers directly (Haskell's does with FunPtr). Hoppy provides an optional layer that performs the necessary type conversions, but only the foreign half of the conversions, so only C types can be used within function pointer types (this is a limitation of speaking over a C FFI; an error is signaled when trying to use a type that requires C<->C++ conversion). The other downside of using function pointers is that C++ provides no lifetime tracking, and because in general foreign code can't know how long some C++ code is going to hold a function pointer, it's necessary to manage the lifetime of the pointer manually.

C++ functor: This is the preferred method for calling into foreign code. This type is expressed with callbackT. It wraps the function pointer support above in C++ functors that add automatic lifetime tracking.

Internally, we create a class G that takes a foreign function pointer and implements operator(), performing the necessary conversions around invoking the pointer. In the event that the function pointer is dynamically allocated (as in Haskell), then this class also ties the lifetime of the function pointer to the lifetime of the class. But this would cause problems for passing this object around by value, so instead we make G non-copyable and non-assignable, allocate our G instance on the heap, and create a second class F that holds a shared_ptr<G> and whose operator() calls through to G.

This way, the existance of the F and G objects are invisible to the foreign language, and (for now) passing these callbacks back to the foreign language is not supported.

When a binding is declared to take a callback type, the generated foreign side of the binding will take a foreign function (the callback) with foreign-side types, and use a function (Haskell: callbackName) generated for the callback type to wrap the callback in a foreign function that does argument decoding and return value encoding: this wrapped function will have C-side types. The binding will then create a G object (above) for this wrapped function (Haskell: using callbackName'), and pass a G pointer into the C side of the binding. The binding will decode this C pointer by wrapping it in a temporary F object, and passing that to the C++ function. The C++ code is free to copy this F object as much as it likes. If it doesn't store a copy somewhere before returning, then the when the temporary F object is destructed, the G object will get deleted.


The Haskell code generator lives in Foreign.Hoppy.Generator.Language.Haskell, with internal parts in Foreign.Hoppy.Generator.Language.Haskell.Internal.

Central to generated Haskell bindings is the idea of type sidedness and the HsTypeSide enum. When a value is passed to or from C++, it needs to be converted so that the receiving language knows what to do with it. The C++ side of bindings just exchanges C types across the language boundary and does not do conversions, so it is up to the Haskell side to do so. Internally, the Haskell generator refers to types exchanged with C++ as C-side types, and types the bindings exchange with user Haskell code as Haskell-side types. These are both Haskell types! The terminology is overlapped a bit but generally, type or C++ type refers to a Type, and in the context of the Haskell generator, C-side or Haskell-side apply to a HsType, calculated from a Type and a HsTypeSide using cppTypeToHsTypeAndUse. For many primitive C++ types, the C-side and Haskell-side types are the same.

Module structure

The result of generating a Hoppy module is a single Haskell module that contains bindings for everything exported from the Hoppy module. The Haskell module name is the concatenation of the interface's interfaceHaskellModuleBase and the module's moduleHaskellName.

The contents of the module depends on the what Exports the module has.

Variable exports

A Variable is exposed in Haskell as a getter function and a setter function. For a variable with external name foo with Haskell-side type Bar, the following functions are created:

foo_get :: IO Bar
foo_set :: Bar -> IO ()

Enum exports

A CppEnum is exposed in Haskell as an enumerable data type. For an enum defined as follows:

alignment :: CppEnum
alignment =
  makeEnum (ident "Alignment") Nothing
  [ (0, ["left", "align"])
  , (1, ["center", "align"])
  , (2, ["right", "align"])

the following data type will be generated:

data Alignment =
  | Alignment_CenterAlign
  | Alignment_RightAlign

with instances for Bounded, Enum, Eq, Ord, and Show.

Bitspace exports

Bitspaces, unlike enums, materialize in Haskell using a single data constructor and bindings for values, rather than multiple data constructors. A bitspace declaration such as

formatFlags :: Bitspace
formatFlags =
  makeBitspace (toExtName "Format") intT
  [ (1, ["format", "letter"])
  , (2, ["format", "jpeg"])
  , (4, ["format", "c"])

will generate the following:

newtype Format

instance Bits Format
instance Bounded Format
instance Eq Format
instance Ord Format
instance Show Format

fromFormat :: Format -> CInt

class IsFormat a where
  toFormat :: a -> Format

instance IsFormat CInt

format_FormatLetter :: Format
format_FormatJpeg :: Format
format_FormatC :: Format

Function exports

For a Function export, a single Haskell function will be generated named after the external name of the export. The function will take the Haskell-side types of its arguments, and return the Haskell-side type of its return type. If the function is Nonpure then it will return a value in IO, otherwise it will return a pure value using unsafePerformIO.

For most Types, the corresponding Haskell parameter type will be a concrete type. This differs for objects (and references and pointers to them), where typeclass constraints are used to implement C++ parameter type contravariance. See the section on Haskell object passing for more details.

Callback exports

Declared callbacks provide support for callback types (callbackT) as well as function pointers (ptrT (fnT ...)) in Haskell.

Callback types manifest directly as Haskell function types in IO. Function pointers manifest as FunPtrs around Haskell function types in IO.

No runtime support is exposed to the user for working with internal Haskell callback types (some machinery is generated however). For function pointer types, a function callbackName_newFunPtr is exposed from the callback's module that makes it easy to wrap anonymous functions in FunPtrs that perform the Haskell side of conversions, with code like the following:

-- Generator bindings

cb_intCallback = makeCallback "IntCallback" [intT] intT

f_funPtrTest = makeFn "funPtrTest" Nothing Nonpure [ptrT $ fnT [intT] intT] intT

f_callbackTest = makeFn "callbackTest" Nothing Nonpure [callbackT cb_intCallback] intT
-- Test program

import Foreign.C (CInt)
import Foreign.Hoppy.Runtime (withScopedFunPtr)

-- Generated things:
intCallback_newFunPtr :: (Int -> IO Int) -> IO (FunPtr (CInt -> IO CInt))
funPtrTest :: FunPtr (CInt -> IO CInt) -> Int
callbackTest :: (Int -> IO Int) -> Int

-- Driver code:
callFunPtrTest = withScopedFunPtr (intCallback_newFunPtr $ return . (* 2)) funPtrTest
callCallbackTest = callbackTest $ return . (* 2)

Class exports

Classes expose quite a few things to the user. Take a simple class definition such as:

compressor :: Class

zipper :: Class
zipper =
  makeClass (ident "Zipper") Nothing [compressor]
  [ mkCtor "new" [] ]
  [ mkStaticMethod "canZip" [] boolT
  , mkConstMethod "hasZipped" [] voidT
  , mkMethod "zip" [] voidT

Let's focus on zipper. Two data types will be generated that represent const and non-const pointers to Zipper objects:

data Zipper
data ZipperConst

Internally, these types hold Ptrs, and they can be converted to Ptrs with toPtr (though this conversion is lossy for pointers managed by the garbage collector, see the section on object passing).

Several typeclass instances are generated for both types:

  • Eq, Ord, and Show compare and render based on the underlying pointer address.
  • CppPtr and Deletable instances provide object management.
  • A single Decodable (Ptr Zipper) Zipper instance is generated for converting raw Ptrs into object handles. This is the opposite operation of toPtr.
  • If the class -- Zipper in this case -- has an operator= method that takes either a objT zipper or a refT (constT (objT zipper)), then an instance ZipperValue a => Assignable Zipper a is generated to allow assigning of general zipper-like values to Zipper objects; see below for an explanation of ZipperValue. This instance is for the non-const Zipper only.

There will also be some typeclasses generated, for types that represent Zipper objects:

class ZipperValue a where
  withZipperPtr :: a -> (ZipperConst -> IO b) -> IO b

instance CompressorPtrConst a => ZipperValue a

class CompressorPtrConst a => ZipperPtrConst a where
  toZipperConst :: a -> ZipperConst

class (ZipperPtrConst a, CompressorPtr a) => ZipperPtr a where
  toZipper :: a -> Zipper

instance ZipperPtrConst ZipperConst
instance ZipperPtr Zipper
... instances required by superclasses ...

Ignoring the first typeclass and instance for a moment, the two Ptr typeclasses represent const and non-const pointers respectively, and allow upcasting pointer types. The const typeclass has as superclasses the const typeclasses for all of the C++ class's superclasses (or just CppPtr if this list is empty). The non-const typeclass has as superclasses the non-const typeclasses for all of the C++ class's superclasses, plus the current const typeclass. Instances will be generated for all of the appropriate typeclasses for Zipper and ZipperConst, all the way up to CppPtr.

The ZipperValue class represents general Zipper values, of which pointers are one type (hence the first instance above). Values of these types can be converted to a temporary const pointer. If Zipper were to have a native Haskell type (see classHaskellConversion), then an additional instance would be generated for that type. This second instance in this case is overlapping, and the above instance is overlappable. These typeclasses allow for mixing pointer, reference, and object types when calling C++ functions.

For downcasting, separate const and non-const typeclasses are generated with instances for all direct and indirect superclasses of Zipper:

-- Enables downcasting from any non-const superclass of Zipper.
class ZipperSuper a where
  downToZipper :: a -> Zipper

-- Enables downcasting from any const superclass of Zipper.
class ZipperSuperConst a where
  downToZipperConst :: a -> ZipperConst

instance ZipperSuper Compressor
... instances for other non-const superclasses ...
instance ZipperSuperConst CompressorConst
... instances for other const superclasses ...

The downcast functions are wrappers around dynamic_cast, and will return a null pointer if the argument is not a supertype of the target type.

Finally, Haskell functions are generated for all of the class's constructors and methods. These work much the same as function exports, but non-static methods take a this object as the first argument. Const methods take a ZipperValue on the assumption that it's safe to create a temporary C++ object from a Haskell value if necessary to call a const method. Non-const methods take a ZipperPtr, since it's potentially a mistake to perform side-effects on a temporary object that is thrown away immediately.

zipper_new :: IO Zipper
zipper_canZip :: IO Bool
zipper_hasZipped :: ZipperValue this => this -> IO Bool
zipper_zip :: ZipperPtr this => this -> IO Bool

Module dependencies

While generated C++ modules get their objects from #includes of underlying headers and only depend on each other in the case of callbacks, Haskell modules depend on each other any time something in one references something in another (somewhat mirroring the dependency graph of the binding definitions), so cycles are much more common (for example, when a C++ interface uses a forward class declaration to break an #include cycle). Fortunately, GHC supports dependency cycles, so Hoppy automatically detects and breaks cycles with the use of .hs-boot files. The boot files contain everything that could be used from another generated module, for example class casting functions needed to coerce pointers to the right type for a foreign call, or enum data declarations. The result of this cycle breaking is deterministic: for each non-trivial strongly connected component in the module dependency graph, .hs-boot files are generated for all modules, and all .hs files' dependencies within the SCC import .hs-boot files.

Object passing

All of the comments about argument passing for the C++ generator apply here. The following types are used for passing arguments from Haskell to C++:

 C++ type   | Pass over FFI | HsCSide  | HsHsSide
 Foo        | Foo const*    | FooConst | FooValue a => a
 Foo const& | Foo const*    | FooConst | FooValue a => a
 Foo&       | Foo*          | Foo      | FooPtr a => a
 Foo const* | Foo const*    | FooConst | FooValue a => a
 Foo*       | Foo*          | Foo      | FooPtr a => a

FooPtr contains pointers to nonconst Foo (and all subclasses). FooValue contains pointers to const and nonconst Foo (and all subclasses), as well as the convertible Haskell type, if there is one. The rationale is that FooValue is used where the callee will not modify the argument, so both a const pointer to an existing object, and a fresh const pointer to a temporary on the case of passing a Foo, are fine. Because functions taking Foo& and Foo* may modify their argument, we disallow passing a temporary converted from a Haskell value implicitly; withCppObj can be used for this.

For values returned from C++, and for arguments and return values in callbacks, the HsCSide column above is the exposed type; polymorphism as in the HsHsSide column is not provided.

Object pointer types in Haskell hide whether they are managed (garbage collected) or unmanaged pointers in their runtime representation. The APIs that bindings expose to Haskell users should generally not require them to be concerned about object lifetimes, and also having separate data types for managed pointers would balloon the size of bindings. Unmanaged objects can be converted to managed objects with toGc; after calling this function, the value it returns should always be used in place of any existing pointers.


C++ exceptions can caught and thrown in Haskell. C++ entities that deal with exceptions need to be marked as such, for Hoppy to generate the support code for them. To work with exceptions at all, you need to pick one of your Hoppy modules to contain some runtime support code, using interfaceSetExceptionSupportModule. C++ functions that throw need to be marked with the specific exceptions that they throw, using handleExceptions. Callbacks that want to be able to throw need to be marked with callbackSetThrows, after which they are allowed to throw any exception classes defined in the interface. Exception handling in both directions can also be set up at the module and interface levels using handleExceptions, interfaceSetCallbacksThrow, and moduleSetCallbacksThrow.

Classes can be marked as being exception classes with classMakeException. Exception classes need to be copyable, so make sure to define a copy constructor (use Copyable).

C++ exceptions in Haskell are handled with throwCpp and catchCpp. While they use Haskell exceptions under the hood, do not use throw and catch to work with them; this may leak C++ objects.

Catching a wildcard (i.e. catch (...)) is supported, but no information is available about the caught value.

Implementation-wise, an in-flight C++ exception in Haskell always owns the object (which is on the heap). An exception coming from C++ into Haskell (it's a heap temporary) will be given to the garbage collector. Hence, for ease of use, caught exceptions should always be garbage-collected. Also, when throwing from Haskell, throwing will always take ownership of the object. If throwCpp gets a non-GCed object, then it will be given to the garbage collector; and then the exception will be thrown as a Haskell exception. If the exception propagates out to a callback and back into C++, then a temporary non-GCed copy will be passed over the gateway, and rethrown as a value object on the C++ side.

In the above strategy, when throwing an exception from Haskell that propagates to C++, it is wasteful to make the thrown object GCed, just to have to create a non-GCed copy. So when we throw from Haskell, we don't actually assign to the garbage collector immediately (if it's not already); instead, we delay the toGc call until catchCpp.