on 4162113 help sign in
378 lines 13 kB history raw
1******************************
2New Foreign Function Interface
3******************************
5.. sectionauthor:: Edwin Brady
7Ever since Idris has had multiple backends compiling to different
8target languages on potentially different platforms, we have had the
9problem that the foreign function interface (FFI) was written under
10the assumption of compiling to C. As a result, it has been hard to
11write generic code for multiple targets, or even to be sure that if
12code compiles that it will run on the expected target.
14As of 0.9.17, Idris will have a new foreign function interface (FFI)
15which is aware of multiple targets. Users who are working with the
16default code generator can happily continue writing programs as before
17with no changes, but if you are writing bindings for an external
18library, writing a back end, or working with a non-C back end, there
19are some things you will need to be aware of, which this page
20describes.
22The ``IO'`` monad, and ``main``
23===============================
25The ``IO`` monad exists as before, but is now specific to the C
26backend (or, more precisely, any backend whose foreign function calls
27are compatible with C.) Additionally, there is now an ``IO'`` monad,
28which is parameterised over a FFI descriptor:
30.. code-block:: idris
32 data IO' : (lang : FFI) -> Type -> Type
34The Prelude defines two FFI descriptors which are imported
35automatically, for C and JavaScript/Node, and defines ``IO`` to use
36the C FFI and ``JS_IO`` to use the JavaScript FFI:
38.. code-block:: idris
40 FFI_C : FFI
41 FFI_JS : FFI
43 IO : Type -> Type
44 IO a = IO' FFI_C a
46 JS_IO : Type -> Type
47 JS_IO a = IO' FFI_JS a
49As before, the entry point to an Idris program is ``main``, but the
50type of ``main`` can now be any implementation of ``IO'``, e.g. the
51following are both valid:
53.. code-block:: idris
55 main : IO ()
56 main : JS_IO ()
58The FFI descriptor includes details about which types can be
59marshalled between the foreign language and Idris, and the "target" of
60a foreign function call (typically just a String representation of the
61function's name, but potentially something more complicated such as an
62external library file or even a URL).
64FFI descriptors
65===============
67An FFI descriptor is a record containing a predicate which holds when
68a type can be marshalled, and the type of the target of a foreign
69call:
71.. code-block:: idris
73 record FFI where
74 constructor MkFFI
75 ffi_types : Type -> Type
76 ffi_fn : Type
78For C, this is:
80.. code-block:: idris
82 ||| Supported C integer types
83 public export
84 data C_IntTypes : Type -> Type where
85 C_IntChar : C_IntTypes Char
86 C_IntNative : C_IntTypes Int
87 C_IntBits8 : C_IntTypes Bits8
88 C_IntBits16 : C_IntTypes Bits16
89 C_IntBits32 : C_IntTypes Bits32
90 C_IntBits64 : C_IntTypes Bits64
92 ||| Supported C function types
93 public export
94 data C_FnTypes : Type -> Type where
95 C_Fn : C_Types s -> C_FnTypes t -> C_FnTypes (s -> t)
96 C_FnIO : C_Types t -> C_FnTypes (IO' FFI_C t)
97 C_FnBase : C_Types t -> C_FnTypes t
99 ||| Supported C foreign types
100 public export
101 data C_Types : Type -> Type where
102 C_Str : C_Types String
103 C_Float : C_Types Double
104 C_Ptr : C_Types Ptr
105 C_MPtr : C_Types ManagedPtr
106 C_Unit : C_Types ()
107 C_Any : C_Types (Raw a)
108 C_FnT : C_FnTypes t -> C_Types (CFnPtr t)
109 C_IntT : C_IntTypes i -> C_Types i
111 ||| A descriptor for the C FFI. See the constructors of `C_Types`
112 ||| and `C_IntTypes` for the concrete types that are available.
113 %error_reverse
114 public export
115 FFI_C : FFI
116 FFI_C = MkFFI C_Types String String
118Foreign calls
119=============
121To call a foreign function, the ``foreign`` function is used. For
122example:
124.. code-block:: idris
126 do_fopen : String -> String -> IO Ptr
127 do_fopen f m
128 = foreign FFI_C "fileOpen" (String -> String -> IO Ptr) f m
130The ``foreign`` function takes an FFI description, a function name (the
131type is given by the ``ffi_fn`` field of ``FFI_C`` here), and a function
132type, which gives the expected types of the remaining arguments. Here,
133we're calling an external function ``fileOpen`` which takes, in the C, a
134``char*`` file name, a ``char*`` mode, and returns a file pointer. It is
135the job of the C back end to convert Idris ``String`` to C ``char*``
136and vice versa.
138The argument types and return type given here must be present in the
139``fn_types`` predicate of the ``FFI_C`` description for the foreign
140call to be valid.
142**Note** The arguments to ``foreign`` *must* be known at compile time,
143because the foreign calls are generated statically. The ``%inline``
144directive on a function can be used to give hints to help this, for
145example a shorthand for calling external JavaScript functions:
147.. code-block:: idris
149 %inline
150 jscall : (fname : String) -> (ty : Type) ->
151 {auto fty : FTy FFI_JS [] ty} -> ty
152 jscall fname ty = foreign FFI_JS fname ty
154C callbacks
155-----------
156It is possible to pass an Idris function to a C function taking a function
157pointer by using ``CFnPtr`` in the function type. The Idris function is passed
158to ``MkCFnPtr`` in the arguments. The example below shows declaring the C standard
159library function ``qsort`` which takes a pointer to a comparison function.
161.. code-block:: idris
163 myComparer : Ptr -> Ptr -> Int
164 myComparer = ...
166 qsort : Ptr -> Int -> Int -> IO ()
167 qsort data elems elsize = foreign FFI_C "qsort"
168 (Ptr -> Int -> Int -> CFnPtr (Ptr -> Ptr -> Int) -> IO ())
169 data elems elsize (MkCFnPtr myComparer)
171There are a few limitations to callbacks in the C FFI. The foreign function can't
172take the function to make a callback of as an argument. This will give a
173compilation error:
175.. code-block:: idris
177 -- This does not work
178 example : (Int -> ()) -> IO ()
179 example f = foreign FFI_C "callbacker" (CFnPtr (Int -> ()) -> IO ()) f
181The other big limitation is that it doesn't support IO functions. Use
182``unsafePerformIO`` to wrap them (i.e. to make an IO function usable as a callback, change the return type
183from IOr to r, and change the = do to = unsafePerformIO $ do).
185There are two special function names:
186``%wrapper`` returns the function pointer that wraps an Idris function. This
187is useful if the function pointer isn't taken by a C function directly but
188should be inserted into a data structure. A foreign declaration using
189``%wrapper`` must return ``IO Ptr``.
191.. code-block:: idris
193 -- this returns the C function pointer to a qsort comparer
194 example_wrapper : IO Ptr
195 example_wrapper = foreign FFI_C "%wrapper" (CFnPtr (Ptr -> Ptr -> Int) -> IO Ptr)
196 (MkCFnPtr myComparer)
198``%dynamic`` calls a C function pointer with some arguments. This is useful if
199a C function returns or data structure contains a C function pointer, for example
200structs of function pointers are common in object-oriented C such as in COM or the
201Linux kernel. The function type contains an extra ``Ptr`` at the start for the
202function pointer. ``%dynamic`` can be seen as a pseudo-function that calls the
203function in the first argument, passing the remaining arguments to it.
205.. code-block:: idris
207 -- we have a pointer to a function with the signature int f(int), call it
208 example_dynamic : Ptr -> Int -> IO Int
209 example_dynamic fn x = foreign FFI_C "%dynamic" (Ptr -> Int -> IO Int) fn x
211If the foreign name is prefixed by a ``&``, it is treated as a pointer to the
212global variable with the following name. The type must be just ``IO Ptr``.
214.. code-block:: idris
216 -- access the global variable errno
217 errno : IO Ptr
218 errno = foreign FFI_C "&errno" (IO Ptr)
220If the foreign name is prefixed by a ``#``, the name is pasted in literally. This is
221useful to access constants that are preprocessor definitions (like ``INT_MAX``).
223.. code-block:: idris
225 %include C "limits.h"
227 -- access the preprocessor definition INT_MAX
228 intMax : IO Int
229 intMax = foreign FFI_C "#INT_MAX" (IO Int)
231 main : IO ()
232 main = print !intMax
234For more complicated interactions with C (such as reading and setting fields
235of a C struct), there is a module CFFI available in the contrib package.
237FFI implementation
238------------------
240In order to write bindings to external libraries, the details of how
241``foreign`` works are unnecessary --- you simply need to know that
242``foreign`` takes an FFI descriptor, the function name, and its
243type. It is instructive to look a little deeper, however:
245The type of ``foreign`` is as follows:
247.. code-block:: idris
249 foreign : (ffi : FFI)
250 -> (fname : ffi_fn f)
251 -> (ty : Type)
252 -> {auto fty : FTy ffi [] ty}
253 -> ty
255The important argument here is the implicit ``fty``, which contains a
256proof (``FTy``) that the given type is valid according to the FFI
257description ``ffi``:
259.. code-block:: idris
261 data FTy : FFI -> List Type -> Type -> Type where
262 FRet : ffi_types f t -> FTy f xs (IO' f t)
263 FFun : ffi_types f s -> FTy f (s :: xs) t -> FTy f xs (s -> t)
265Notice that this uses the ``ffi_types`` field of the FFI descriptor
266--- these arguments to ``FRet`` and ``FFun`` give explicit proofs that
267the type is valid in this FFI. For example, the above ``do_fopen``
268builds the following implicit proof as the ``fty`` argument to
269``foreign``:
271.. code-block:: idris
273 FFun C_Str (FFun C_Str (FRet C_Ptr))
275Compiling foreign calls
276=======================
278(This section assumes some knowledge of the Idris internals.)
280When writing a back end, we now need to know how to compile
281``foreign``. We'll skip the details here of how a ``foreign`` call
282reaches the intermediate representation (the IR), though you can look
283in ``IO.idr`` in the ``prelude`` package to see a bit more detail ---
284a ``foreign`` call is implemented by the primitive function
285``mkForeignPrim``. The important part of the IR as defined in
286``Lang.hs`` is the following constructor:
288.. code-block:: idris
290 data LExp = ...
291 | LForeign FDesc -- Function descriptor
292 FDesc -- Return type descriptor
293 [(FDesc, LExp)]
295So, a ``foreign`` call appears in the IR as the ``LForeign``
296constructor, which takes a function descriptor (of a type given by the
297``ffi_fn`` field in the FFI descriptor), a return type descriptor
298(given by an application of ``FTy``), and a list of arguments with
299type descriptors (also given by an application of ``FTy``).
301An ``FDesc`` describes an application of a name to some arguments, and
302is really just a simplified subset of an ``LExp``:
304.. code-block:: idris
306 data FDesc = FCon Name
307 | FStr String
308 | FUnknown
309 | FApp Name [FDesc]
311There are corresponding structures in the lower level IRs, such as the
312defunctionalised, simplified and bytecode forms.
314Our ``do_fopen`` example above arrives in the ``LExp`` form as:
316.. code-block:: idris
318 LForeign (FStr "fileOpen") (FCon (sUN "C_Ptr"))
319 [(FCon (sUN "C_Str"), f), (FCon (sUN "C_Str"), m)]
321(Assuming that ``f`` and ``m`` stand for the ``LExp`` representations
322of the arguments.) This information should be enough for any back end
323to marshal the arguments and return value appropriately.
325.. note::
327 When processing ``FDesc``, be aware that there may be implicit
328 arguments, which have not been erased. For example, ``C_IntT`` has
329 an implicit argument ``i``, so will appear in an ``FDesc`` as
330 something of the form ``FApp (sUN "C_IntT") [i, t]`` where ``i`` is
331 the implicit argument (which can be ignored) and ``t`` is the
332 descriptor of the integer type. See ``CodegenC.hs``, specifically
333 the function ``toFType``, to see how this works in practice.
335JavaScript FFI descriptor
336=========================
338The JavaScript FFI descriptor is a little more complex, because the
339JavaScript FFI supports marshalling functions. It is defined as
340follows:
342.. code-block:: idris
344 mutual
345 data JsFn t = MkJsFn t
347 data JS_IntTypes : Type -> Type where
348 JS_IntChar : JS_IntTypes Char
349 JS_IntNative : JS_IntTypes Int
351 data JS_FnTypes : Type -> Type where
352 JS_Fn : JS_Types s -> JS_FnTypes t -> JS_FnTypes (s -> t)
353 JS_FnIO : JS_Types t -> JS_FnTypes (IO' l t)
354 JS_FnBase : JS_Types t -> JS_FnTypes t
356 data JS_Types : Type -> Type where
357 JS_Str : JS_Types String
358 JS_Float : JS_Types Double
359 JS_Ptr : JS_Types Ptr
360 JS_Unit : JS_Types ()
361 JS_FnT : JS_FnTypes a -> JS_Types (JsFn a)
362 JS_IntT : JS_IntTypes i -> JS_Types i
364The reason for wrapping function types in a ``JsFn`` is to help the
365proof search when building ``FTy``. We hope to improve proof search
366eventually, but for the moment it works much more reliably if the
367indices are disjoint! An example of using this appears in `IdrisScript
368<https://github.com/idris-hackers/IdrisScript>`__ when setting
369timeouts:
371.. code-block:: idris
373 setTimeout : (() -> JS_IO ()) -> (millis : Int) -> JS_IO Timeout
374 setTimeout f millis = do
375 timeout <- jscall "setTimeout(%0, %1)"
376 (JsFn (() -> JS_IO ()) -> Int -> JS_IO Ptr)
377 (MkJsFn f) millis
378 pure $ MkTimeout timeout