Preparing and Distributing modules¶

There are three or four different ways to use CFFI in a project. In order of complexity:

The “in-line”, “ABI mode”:

import cffi

ffi = cffi.FFI()
ffi.cdef("C-like declarations")
lib = ffi.dlopen("libpath")

# use ffi and lib here

The “out-of-line”, but still “ABI mode”, useful to organize the code and reduce the import time:

# in a separate file "package/foo_build.py"
import cffi

ffi = cffi.FFI()
ffi.set_source("package._foo", None)
ffi.cdef("C-like declarations")

if __name__ == "__main__":
    ffi.compile()

Running python foo_build.py produces a file _foo.py, which can then be imported in the main program:

from package._foo import ffi
lib = ffi.dlopen("libpath")

# use ffi and lib here

The “out-of-line”, “API mode” gives you the most flexibility to access a C library at the level of C, instead of at the binary level:
```
# in a separate file "package/foo_build.py"
import cffi

ffi = cffi.FFI()
ffi.set_source("package._foo", "real C code")   # <=
ffi.cdef("C-like declarations with '...'")

if __name__ == "__main__":
    ffi.compile()
```
Running python foo_build.py produces a file _foo.c and invokes the C compiler to turn it into a file _foo.so (or _foo.pyd or _foo.dylib). It is a C extension module which can be imported in the main program:
```
from package._foo import ffi, lib
# no ffi.dlopen()

# use ffi and lib here
```

Finally, you can (but don’t have to) use CFFI’s Distutils or Setuptools integration when writing a setup.py. For Distutils (only in out-of-line API mode):

# setup.py (requires CFFI to be installed first)
from distutils.core import setup

import foo_build   # possibly with sys.path tricks to find it

setup(
    ...,
    ext_modules=[foo_build.ffi.distutils_extension()],
)

For Setuptools (out-of-line, but works in ABI or API mode; recommended):

# setup.py (with automatic dependency tracking)
from setuptools import setup

setup(
    ...,
    setup_requires=["cffi>=1.0.0"],
    cffi_modules=["package/foo_build.py:ffi"],
    install_requires=["cffi>=1.0.0"],
)

Note that CFFI actually contains two different FFI classes. The page Using the ffi/lib objects describes the common functionality. It is what you get in the from package._foo import ffi lines above. On the other hand, the extended FFI class is the one you get from import cffi; ffi = cffi.FFI(). It has the same functionality (for in-line use), but also the extra methods described below (to prepare the FFI).

The reason for this split of functionality is that a regular program using CFFI out-of-line does not need to import the cffi pure Python package at all. (Internally it still needs _cffi_backend, a C extension module that comes with CFFI; this is why CFFI is also listed in install_requires=.. above. In the future this might be split into a different PyPI package that only installs _cffi_backend.)

Note that a few small differences do exist: notably, from _foo import ffi returns an object of a type written in C, which does not let you add random attributes to it (nor does it have all the underscore-prefixed internal attributes of the Python version). Similarly, the lib objects returned by the C version are read-only, apart from writes to global variables. Also, lib.__dict__ no longer works (it now tries to look up a hypothetical symbol __dict__ from the C library); use instead dir(lib).

ffi.cdef(): declaring types and functions¶

ffi.cdef(source): parses the given C source. It registers all the functions, types, constants and global variables in the C source. The types can be used immediately in ffi.new() and other functions. Before you can access the functions and global variables, you need to give ffi another piece of information: where they actually come from (which you do with either ffi.dlopen() or ffi.set_source()).

The C source is parsed internally (using pycparser). This code cannot contain #include. It should typically be a self-contained piece of declarations extracted from a man page. The only things it can assume to exist are the standard types:

char, short, int, long, long long (both signed and unsigned)
float, double, long double
intN_t, uintN_t (for N=8,16,32,64), intptr_t, uintptr_t, ptrdiff_t, size_t, ssize_t
wchar_t (if supported by the backend)
_Bool and bool (equivalent). If not directly supported by the C compiler, this is declared with the size of unsigned char.
FILE. You can declare C functions taking a FILE * argument and call them with a Python file object. If needed, you can also do c_f = ffi.cast("FILE *", fileobj) and then pass around c_f.
all common Windows types are defined if you run on Windows (DWORD, LPARAM, etc.). Changed in version 0.9: the types TBYTE TCHAR LPCTSTR PCTSTR LPTSTR PTSTR PTBYTE PTCHAR are no longer automatically defined; see ffi.set_unicode().
New in version 0.9.3: the other standard integer types from stdint.h, like intmax_t, as long as they map to integers of 1, 2, 4 or 8 bytes. Larger integers are not supported.

The declarations can also contain “...” at various places; these are placeholders that will be completed by the compiler. More information about it below in Letting the C compiler fill the gaps.

Note that all standard type names listed above are handled as defaults only (apart from the ones that are keywords in the C language). If your cdef contains an explicit typedef that redefines one of the types above, then the default described above is ignored. (This is a bit hard to implement cleanly, so in some corner cases it might fail, notably with the error Multiple type specifiers with a type tag. Please report it as a bug if it does.)

Multiple calls to ffi.cdef() are possible. Beware that it can be slow to call ffi.cdef() a lot of times, a consideration that is important mainly in in-line mode.

New in version 0.8.2: The ffi.cdef() call takes an optional argument packed: if True, then all structs declared within this cdef are “packed”. If you need both packed and non-packed structs, use several cdefs in sequence.) This has a meaning similar to __attribute__((packed)) in GCC. It specifies that all structure fields should have an alignment of one byte. (Note that the packed attribute has no effect on bit fields so far, which mean that they may be packed differently than on GCC. Also, this has no effect on structs declared with "...;"—next section.)

ffi.set_unicode(enabled_flag): Windows: if enabled_flag is True, enable the UNICODE and _UNICODE defines in C, and declare the types TBYTE TCHAR LPCTSTR PCTSTR LPTSTR PTSTR PTBYTE PTCHAR to be (pointers to) wchar_t. If enabled_flag is False, declare these types to be (pointers to) plain 8-bit characters. (These types are not predeclared at all if you don’t call set_unicode().) New in version 0.9.

The reason behind this method is that a lot of standard functions have two versions, like MessageBoxA() and MessageBoxW(). The official interface is MessageBox() with arguments like LPTCSTR. Depending on whether UNICODE is defined or not, the standard header renames the generic function name to one of the two specialized versions, and declares the correct (unicode or not) types.

Usually, the right thing to do is to call this method with True. Be aware (particularly on Python 2) that, afterwards, you need to pass unicode strings as arguments instead of not byte strings. (Before cffi version 0.9, TCHAR and friends where hard-coded as unicode, but UNICODE was, inconsistently, not defined by default.)

ffi.dlopen(): loading libraries in ABI mode¶

ffi.dlopen(libpath, [flags]): this function opens a shared library and returns a module-like library object. Use this when you are fine with the limitations of ABI-level access to the system. In case of doubt, read again ABI versus API in the overview.

You can use the library object to call the functions previously declared by ffi.cdef(), to read constants, and to read or write global variables. Note that you can use a single cdef() to declare functions from multiple libraries, as long as you load each of them with dlopen() and access the functions from the correct one.

The libpath is the file name of the shared library, which can contain a full path or not (in which case it is searched in standard locations, as described in man dlopen), with extensions or not. Alternatively, if libpath is None, it returns the standard C library (which can be used to access the functions of glibc, on Linux).

Let me state it again: this gives ABI-level access to the library, so you need to have all types declared manually exactly as they were while the library was made. No checking is done. Mismatches can cause random crashes.

Note that only functions and global variables live in library objects; the types exist in the ffi instance independently of library objects. This is due to the C model: the types you declare in C are not tied to a particular library, as long as you #include their headers; but you cannot call functions from a library without linking it in your program, as dlopen() does dynamically in C.

For the optional flags argument, see man dlopen (ignored on Windows). It defaults to ffi.RTLD_NOW.

This function returns a “library” object that gets closed when it goes out of scope. Make sure you keep the library object around as long as needed. (Alternatively, the out-of-line FFIs have a method ffi.dlclose(lib).)

Note: the old version of ffi.dlopen() from the in-line ABI mode tries to use ctypes.util.find_library() if it cannot directly find the library. The newer out-of-line ffi.dlopen() no longer does it automatically; it simply passes the argument it receives to the underlying dlopen() or LoadLibrary() function. If needed, it is up to you to use ctypes.util.find_library() or any other way to look for the library’s filename. This also means that ffi.dlopen(None) no longer work on Windows; try instead ffi.dlopen(ctypes.util.find_library('c')).

ffi.set_source(): preparing out-of-line modules¶

ffi.set_source(module_name, c_header_source, [**keywords...]): prepare the ffi for producing out-of-line an external module called module_name. New in version 1.0.

ffi.set_source() by itself does not write any file, but merely records its arguments for later. It can therefore be called before or after ffi.cdef().

In ABI mode, you call ffi.set_source(module_name, None). The argument is the name (or dotted name inside a package) of the Python module to generate. In this mode, no C compiler is called.

In API mode, the c_header_source argument is a string that will be pasted into the .c file generated. This piece of C code typically contains some #include, but may also contain more, like definitions for custom “wrapper” C functions. The goal is that the .c file can be generated like this:

// C file "module_name.c"
#include <Python.h>

...c_header_source...

...magic code...

where the “magic code” is automatically generated from the cdef(). For example, if the cdef() contains int foo(int x); then the magic code will contain logic to call the function foo() with an integer argument, itself wrapped inside some CPython or PyPy-specific code.

The keywords arguments to set_source() control how the C compiler will be called. They are passed directly to distutils or setuptools and include at least sources, include_dirs, define_macros, undef_macros, libraries, library_dirs, extra_objects, extra_compile_args and extra_link_args. You typically need at least libraries=['foo'] in order to link with libfoo.so or libfoo.so.X.Y, or foo.dll on Windows. The sources is a list of extra .c files compiled and linked together (the file module_name.c shown above is always generated and automatically added as the first argument to sources). See the distutils documentations for more information about the other arguments.

An extra keyword argument processed internally is source_extension, defaulting to ".c". The file generated will be actually called module_name + source_extension. Example for C++ (but note that there are still a few known issues of C-versus-C++ compatibility):

ffi.set_source("mymodule", '''
extern "C" {
    int somefunc(int somearg) { return real_cpp_func(somearg); }
}
''', source_extension='.cpp')

Letting the C compiler fill the gaps¶

If you are using a C compiler (“API mode”), then:

functions taking or returning integer or float-point arguments can be misdeclared: if e.g. a function is declared by cdef() as taking a int, but actually takes a long, then the C compiler handles the difference.
other arguments are checked: you get a compilation warning or error if you pass a int * argument to a function expecting a long *.
similarly, most other things declared in the cdef() are checked, to the best we implemented so far; mistakes give compilation warnings or errors.

Moreover, you can use “...” (literally, dot-dot-dot) in the cdef() at various places, in order to ask the C compiler to fill in the details. These places are:

structure declarations: any struct { } that ends with “...;” as the last “field” is partial: it may be missing fields and/or have them declared out of order. This declaration will be corrected by the compiler. (But note that you can only access fields that you declared, not others.) Any struct declaration which doesn’t use “...” is assumed to be exact, but this is checked: you get an error if it is not correct.
New in version 1.1: integer types: the syntax “typedef int... foo_t;” declares the type foo_t as an integer type whose exact size and signedness is not specified. The compiler will figure it out. (Note that this requires set_source(); it does not work with verify().) The int... can be replaced with long... or unsigned long long... or any other primitive integer type, with no effect. The type will always map to one of (u)int(8,16,32,64)_t in Python, but in the generated C code, only foo_t is used.
unknown types: the syntax “typedef ... foo_t;” declares the type foo_t as opaque. Useful mainly for when the API takes and returns foo_t * without you needing to look inside the foo_t. Also works with “typedef ... *foo_p;” which declares the pointer type foo_p without giving a name to the opaque type itself. Note that such an opaque struct has no known size, which prevents some operations from working (mostly like in C). You cannot use this syntax to declare a specific type, like an integer type! It declares opaque struct-like types only. In some cases you need to say that foo_t is not opaque, but just a struct where you don’t know any field; then you would use “typedef struct { ...; } foo_t;”.
array lengths: when used as structure fields or in global variables, arrays can have an unspecified length, as in “int n[...];”. The length is completed by the C compiler. This is slightly different from “int n[];”, because the latter means that the length is not known even to the C compiler, and thus no attempt is made to complete it. New in version 1.1: support for multidimensional arrays: “int n[...][...];”.
enums: if you don’t know the exact order (or values) of the declared constants, then use this syntax: “enum foo { A, B, C, ... };” (with a trailing “...”). The C compiler will be used to figure out the exact values of the constants. An alternative syntax is “enum foo { A=..., B, C };” or even “enum foo { A=..., B=..., C=... };”. Like with structs, an enum without “...” is assumed to be exact, and this is checked.
integer constants and macros: you can write in the cdef the line “#define FOO ...”, with any macro name FOO but with ... as a value. Provided the macro is defined to be an integer value, this value will be available via an attribute of the library object. The same effect can be achieved by writing a declaration static const int FOO;. The latter is more general because it supports other types than integer types (note: the C syntax is then to write the const together with the variable name, as in static char *const FOO;).

Currently, it is not supported to find automatically which of the various integer or float types you need at which place. If a type is named, and an integer type, then use typedef int... the_type_name;. In the case of function arguments or return type, when it is a simple integer/float type, it may be misdeclared (if you misdeclare a function void f(long) as void f(int), it still works, but you have to call it with arguments that fit an int). But it doesn’t work any longer for more complex types (e.g. you cannot misdeclare a int * argument as long *) or in other locations (e.g. a global array int a[5]; must not be misdeclared long a[5];). CFFI considers all types listed above as primitive (so long long a[5]; and int64_t a[5] are different declarations).

ffi.compile() etc.: compiling out-of-line modules¶

You can use one of the following functions to actually generate the .py or .c file prepared with ffi.set_source() and ffi.cdef().

Note that these function won’t overwrite a .py/.c file with exactly the same content, to preserve the mtime. In some cases where you need the mtime to be updated anyway, delete the file before calling the functions.

ffi.compile(tmpdir=’.’): explicitly generate the .py or .c file, and (if .c) compile it. The output file is (or are) put in the directory given by tmpdir. In the examples given here, we use if __name__ == "__main__": ffi.compile() in the build scripts—if they are directly executed, this makes them rebuild the .py/.c file in the current directory. (Note: if a package is specified in the call to set_source(), then a corresponding subdirectory of the tmpdir is used.)

ffi.emit_python_code(filename): generate the given .py file (same as ffi.compile() for ABI mode, with an explicitly-named file to write). If you choose, you can include this .py file pre-packaged in your own distributions: it is identical for any Python version (2 or 3).

ffi.emit_c_code(filename): generate the given .c file (for API mode) without compiling it. Can be used if you have some other method to compile it, e.g. if you want to integrate with some larger build system that will compile this file for you. You can also distribute the .c file: unless the build script you used depends on the OS or platform, the .c file itself is generic (it would be exactly the same if produced on a different OS, with a different version of CPython, or with PyPy; it is done with generating the appropriate #ifdef).

ffi.distutils_extension(tmpdir=’build’, verbose=True): for distutils-based setup.py files. Calling this creates the .c file if needed in the given tmpdir, and returns a distutils.core.Extension instance.

For Setuptools, you use instead the line cffi_modules=["path/to/foo_build.py:ffi"] in setup.py. This line asks Setuptools to import and use a helper provided by CFFI, which in turn executes the file path/to/foo_build.py (as with execfile()) and looks up its global variable called ffi. You can also say cffi_modules=["path/to/foo_build.py:maker"], where maker names a global function; it is called with no argument and is supposed to return a FFI object.

ffi.include(): combining multiple CFFI interfaces¶

ffi.include(other_ffi): includes the typedefs, structs, unions, enums and constants defined in another FFI instance. This is meant for large projects where one CFFI-based interface depends on some types declared in a different CFFI-based interface.

Note that you should only use one ffi object per library; the intended usage of ffi.include() is if you want to interface with several inter-dependent libraries. For only one library, make one ffi object. (You can write several cdef() calls over the same ffi from several Python files, if one file would be too large.)

For out-of-line modules, the ffi.include(other_ffi) line should occur in the build script, and the other_ffi argument should be another FFI that comes from another build script. When the two build scripts are turned into generated files, say _ffi.so and _other_ffi.so, then importing _ffi.so will internally cause _other_ffi.so to be imported. At that point, the real declarations from _other_ffi.so are combined with the real declarations from _ffi.so.

The usage of ffi.include() is the cdef-level equivalent of a #include in C, where a part of the program might include types and functions defined in another part for its own usage. You can see on the ffi object (and associated lib objects on the including side) the types and constants declared on the included side. In API mode, you can also see the functions and global variables directly. In ABI mode, these must be accessed via the original other_lib object returned by the dlopen() method on other_ffi.

ffi.cdef() limitations¶

All of the ANSI C declarations should be supported in cdef(), and some of C99. (This excludes any #include or #ifdef.) Known missing features that are GCC or MSVC extensions:

Any __attribute__ or #pragma pack(n)
Additional types: complex numbers, special-size floating and fixed point types, vector types, and so on. You might be able to access an array of complex numbers by declaring it as an array of struct my_complex { double real, imag; }, but in general you should declare them as struct { ...; } and cannot access them directly. This means that you cannot call any function which has an argument or return value of this type (this would need added support in libffi). You need to write wrapper functions in C, e.g. void foo_wrapper(struct my_complex c) { foo(c.real + c.imag*1j); }, and call foo_wrapper rather than foo directly.
Thread-local variables (access them via getter/setter functions)
Function pointers with non-default calling conventions (e.g. on Windows, “stdcall”).

Note that since version 0.8, declarations like int field[]; in structures are interpreted as variable-length structures. Declarations like int field[...]; on the other hand are arrays whose length is going to be completed by the compiler. You can use int field[]; for array fields that are not, in fact, variable-length; it works too, but in this case, as CFFI believes it cannot ask the C compiler for the length of the array, you get reduced safety checks: for example, you risk overwriting the following fields by passing too many array items in the constructor.

Debugging dlopen’ed C libraries¶

A few C libraries are actually hard to use correctly in a dlopen() setting. This is because most C libraries are intented for, and tested with, a situation where they are linked with another program, using either static linking or dynamic linking — but from a program written in C, at start-up, using the linker’s capabilities instead of dlopen().

This can occasionally create issues. You would have the same issues in another setting than CFFI, like with ctypes or even plain C code that calls dlopen(). This section contains a few generally useful environment variables (on Linux) that can help when debugging these issues.

export LD_TRACE_LOADED_OBJECTS=all

provides a lot of information, sometimes too much depending on the setting. Output verbose debugging information about the dynamic linker. If set to all prints all debugging information it has, if set to help prints a help message about which categories can be specified in this environment variable

export LD_VERBOSE=1

(glibc since 2.1) If set to a nonempty string, output symbol versioning information about the program if querying information about the program (i.e., either LD_TRACE_LOADED_OBJECTS has been set, or --list or --verify options have been given to the dynamic linker).

export LD_WARN=1

(ELF only)(glibc since 2.1.3) If set to a nonempty string, warn about unresolved symbols.

ffi.verify(): in-line API-mode¶

ffi.verify() is supported for backward compatibility, but is deprecated. ffi.verify(c_header_source, tmpdir=.., ext_package=.., modulename=.., flags=.., **kwargs) makes and compiles a C file from the ffi.cdef(), like ffi.set_source() in API mode, and then immediately loads and returns the dynamic library object. Some non-trivial logic is used to decide if the dynamic library must be recompiled or not; see below for ways to control it.

The c_header_source and the extra keyword arguments have the same meaning as in ffi.set_source().

One remaining use case for ffi.verify() would be the following hack to find explicitly the size of any type, in bytes, and have it available in Python immediately (e.g. because it is needed in order to write the rest of the build script):

ffi = cffi.FFI()
ffi.cdef("const int mysize;")
lib = ffi.verify("const int mysize = sizeof(THE_TYPE);")
print lib.mysize

Extra arguments to ffi.verify():

tmpdir controls where the C files are created and compiled. Unless the CFFI_TMPDIR environment variable is set, the default is directory_containing_the_py_file/__pycache__ using the directory name of the .py file that contains the actual call to ffi.verify(). (This is a bit of a hack but is generally consistent with the location of the .pyc files for your library. The name __pycache__ itself comes from Python 3.)
ext_package controls in which package the compiled extension module should be looked from. This is only useful after distributing ffi.verify()-based modules.
The tag argument gives an extra string inserted in the middle of the extension module’s name: _cffi_<tag>_<hash>. Useful to give a bit more context, e.g. when debugging.
The modulename argument can be used to force a specific module name, overriding the name _cffi_<tag>_<hash>. Use with care, e.g. if you are passing variable information to verify() but still want the module name to be always the same (e.g. absolute paths to local files). In this case, no hash is computed and if the module name already exists it will be reused without further check. Be sure to have other means of clearing the tmpdir whenever you change your sources.
source_extension has the same meaning as in ffi.set_source().
The optional flags argument has been added in version 0.9; see man dlopen (ignored on Windows). It defaults to ffi.RTLD_NOW. (With ffi.set_source(), you would use sys.setdlopenflags().)
The optional relative_to argument is useful if you need to list local files passed to the C compiler:
```
ext = ffi.verify(..., sources=['foo.c'], relative_to=__file__)
```
The line above is roughly the same as:
```
ext = ffi.verify(..., sources=['/path/to/this/file/foo.c'])
```
except that the default name of the produced library is built from the CRC checkum of the argument sources, as well as most other arguments you give to ffi.verify() – but not relative_to. So if you used the second line, it would stop finding the already-compiled library after your project is installed, because the '/path/to/this/file' suddenly changed. The first line does not have this problem.

Note that during development, every time you change the C sources that you pass to cdef() or verify(), then the latter will create a new module file name, based on two CRC32 hashes computed from these strings. This creates more and more files in the __pycache__ directory. It is recommended that you clean it up from time to time. A nice way to do that is to add, in your test suite, a call to cffi.verifier.cleanup_tmpdir(). Alternatively, you can manually remove the whole __pycache__ directory.

An alternative cache directory can be given as the tmpdir argument to verify(), via the environment variable CFFI_TMPDIR, or by calling cffi.verifier.set_tmpdir(path) prior to calling verify.

Upgrading from CFFI 0.9 to CFFI 1.0¶

CFFI 1.0 is backward-compatible, but it is still a good idea to consider moving to the out-of-line approach new in 1.0. Here are the steps.

ABI mode if your CFFI project uses ffi.dlopen():

import cffi

ffi = cffi.FFI()
ffi.cdef("stuff")
lib = ffi.dlopen("libpath")

and if the “stuff” part is big enough that import time is a concern, then rewrite it as described in the out-of-line but still ABI mode above. Optionally, see also the setuptools integration paragraph.

API mode if your CFFI project uses ffi.verify():

import cffi

ffi = cffi.FFI()
ffi.cdef("stuff")
lib = ffi.verify("real C code")

then you should really rewrite it as described in the out-of-line, API mode above. It avoids a number of issues that have caused ffi.verify() to grow a number of extra arguments over time. Then see the distutils or setuptools paragraph. Also, remember to remove the ext_package=".." from your setup.py, which was sometimes needed with verify() but is just creating confusion with set_source().

The following example should work both with old (pre-1.0) and new versions of CFFI—supporting both is important to run on PyPy, because CFFI 1.0 does not work in PyPy < 2.6:

# in a separate file "package/foo_build.py"
import cffi

ffi = cffi.FFI()
C_HEADER_SRC = '''
    #include "somelib.h"
'''
C_KEYWORDS = dict(libraries=['somelib'])

if hasattr(ffi, 'set_source'):
    ffi.set_source("package._foo", C_HEADER_SRC, **C_KEYWORDS)

ffi.cdef('''
    int foo(int);
''')

if __name__ == "__main__":
    ffi.compile()

And in the main program:

try:
    from package._foo import ffi, lib
except ImportError:
    from package.foo_build import ffi, C_HEADER_SRC, C_KEYWORDS
    lib = ffi.verify(C_HEADER_SRC, **C_KEYWORDS)

(FWIW, this latest trick can be used more generally to allow the import to “work” even if the _foo module was not generated.)

Writing a setup.py script that works both with CFFI 0.9 and 1.0 requires explicitly checking the version of CFFI that we can have—it is hard-coded as a built-in module in PyPy:

if '_cffi_backend' in sys.builtin_module_names:   # PyPy
    import _cffi_backend
    requires_cffi = "cffi==" + _cffi_backend.__version__
else:
    requires_cffi = "cffi>=1.0.0"

Then we use the requires_cffi variable to give different arguments to setup() as needed, e.g.:

if requires_cffi.startswith("cffi==0."):
    # backward compatibility: we have "cffi==0.*"
    from package.foo_build import ffi
    extra_args = dict(
        ext_modules=[ffi.verifier.get_extension()],
        ext_packages="...",    # if needed
    )
else:
    extra_args = dict(
        setup_requires=[requires_cffi],
        cffi_modules=['package/foo_build.py:ffi'],
    )
setup(
    name=...,
    ...,
    install_requires=[requires_cffi],
    **extra_args
)