Commands, et al.
Ffidl Version 0.5
Ffidl allows you to call C functions using pure Tcl wrappers. You specify a function name, a library, a list of argument types, and a return type, and Ffidl takes care of the nasty details of converting a Tcl command call into a C function call for you. So, if you have a shared library and a specification of the entries in the library, you can wrap the library into a Tcl extension with Ffidl and pure Tcl.
Ffidl 0.5 adds support for Windows and synchronous callbacks from C functions into Tcl on the x86 architecture, the option of using the GPL'ed ffcall package, introduces a ::ffidl:: namespace, renames ffidl-proc to ::ffidl::callout, and adds some options to ::ffidl::info.
Ffidl uses libffi-2.00-beta , with substantial patches, to dynamically construct calls to C functions from Tcl, and on the x86 architecture, to dynamically construct calls from C functions into Tcl. Libffi has officially been ported to:
Ffidl can also be configured (--enable-ffcall) to use ffcall-1.6 a GPL'ed foreign function package which implements both callouts and callbacks on:
Ffidl uses dlopen() and dlsym(), or their Windows approximations LoadLibrary() and GetProcAddress, to load dynamic libraries and discover the locations of functions.
Ffidl was developed under linux-2.0.36, linux-2.2.5, and Windows 95 using Tcl8.2.1. It should be able to run on any system with a stubs enabled Tcl, libffi or ffcall support, and a libdl implementation.
Ffidl 0.5 is a second alpha release. There have been about 50 downloads of Ffidl 0.1, the first alpha, in the past weeks, but only one person has reported on his experiences. There are configuration details which you will need to attend to by hand in the current release. The initial development turned up a few bugs in libffi under linux-x86, so users on other architectures should be alert for similar problems. There are several open design issues still to be resolved, so there may be changes in the interfaces in future releases. There are several common Tcl/Tk platforms which need ports of libffi or implementations of libdl.
I'm releasing at this early point because it's really too much fun to keep to myself, I need assistance verifying that the implementation works on all the architectures supported by libffi, and the open design issues could use some discussion.
Ffidl defines five Tcl commands in the Ffidl package: ::ffidl::callout, ::ffidl::callback, ::ffidl::symbol, ::ffidl::typedef, and ::ffidl::info; exports one function from the Ffidl shared library: ffidl_pointer_pun; and defines two helper procs in the Ffidlrt package in demos/ffidlrt.tcl: ::ffidl::find-lib, and ::ffidl::find-type, which are currently just stubs of their true form.
These interfaces should be considered subject to revision.
The Ffidl builtin types include the scalar C types in both their unsized forms and as explicitly bit sized types, and a variety of pointer treatments. Note that some types are only valid in certain contexts: arguments (arg), return (ret), or struct elements (elt).
In addition to the builtin types, the ::ffidl::typedef command may be used to define new types. Aliases for existing types may be used where ever the existing type may be used. Structured aggregates may be used as arguments, returns, or elements of other structures.
Building on linux has been painless, but building on Windows has yet to be replicated by anyone but myself. The binaries directory contains some prebuilt libs.
Installation on linux consists of:
tar xzvf ffidl-0.5.tar.gz cd ffidl-0.5 configure && makeInstallation under windows consists of:
unzip ffidl05.zip cd ffidl-0.5 CC="gcc -mno-cygwin" ./configure make -f Makefile.gnuThe Windows build is very, very picky. It assumes that you have Cygwin-B20 installed with Mumit Khan's gcc-2.95 update, and with the libtclstub82.a from an install of Jan Nijtman's tcl82*plus.exe .
The Windows Makefile.gnu contains instructions for compiling ffcall, but it doesn't link as a dll under cygwin.
Three custom configure options are implemented for selecting between libffi and ffcall, --enable-libffi and --enable-ffcall, and for excluding callbacks, --disable-callbacks.
(cd libffi-2.00-beta && ./ffitest)will run libffi's test suite.
(cd ffcall-1.6 && make check)will run ffcall's self test.
At this point you should edit demos/ffidlrt.tcl and look at the table of libraries in ::ffidlrt::libs and the table of types in ::ffidlrt::types. Either or both of these tables will probably need attention if you go further. You may need, for instance, to find, build, or install libraries for gmp and gdbm, or to adjust the pathnames for libc, libm, tcl, and tk. (The ever useful Mumit Khan has a gmp port for Windows, gmp-dll-2.0.2-mingw32.zip, available for download, which will allow Windozens to run the multiprecision demos.)
make -f Makefile.gnu testwill build a test shared lib and run a series of scripts in demos/ which run to completion without error messages on my system. It is not quite a systematic test suite, but it does exercise a good deal of Ffidl's capabilities, and it has turned up some problems with libffi.
The configure for Ffidl started from the TEA sample configuration, but it needs more work. It doesn't test that libdl is actually available. It builds the ffidl_test library and runs the test scripts in demos/, but it wasn't clear how Makefile.in was supposed to support this or cleaning up test binaries. Most glaringly, I couldn't get the Windows configure to work at all, so I gave up and wrote a makefile.gnu by hand.
The included copy of libffi-2.00-beta has been rewritten from a copy retrieved via cvs on October 28, 1999. The distributed libffi-2.00-beta erroneously makes "long int" a 64 bit integer on x86, and the inline assembly for implementing closures miscompiled on two out of two gcc's. If you have installed a copy of libffi on your machine, beware of finding the unpatched headers in /usr/local/include before the patched ones.
The demos directory contains several small and medium size examples of Ffidl bindings to shared libraries, and some code for making comparisons to other ways of doing the same thing.
ffidlrt.tcl will need attention unless you're running on my machine. There are two functions, ::ffidl::find-lib and ::ffidl::find-type, which abstract library names and system typedefs out of the rest of the code. However, the abstraction is currently limited to the correct results for my Linux box. You'll need to rewrite the mapping for your own machine.
ffidlrt.tcl also contains some examples of binding into the Tcl core itself.
tkphoto.tcl allows extraction and insertion of photo image pixels as binary data. See test-tkphoto.tcl for an example.
The gdbm.tcl extension should be plug compatible with Tclgdbm0.6, a C coded Tcl extension for manipulating gdbm files. Since gdbm passes and returns structures, it also tests the Ffidl struct code.
The gmp*.tcl extensions make a nice example. The main Gmp package wraps all the exported mpz_*, mpq_*, and mpf_* entries from the Gnu multiple precision library. The subsidiary Gmp[zqf] packages use the Gmp package to define arbitrary precision integers, rationals, and floats which are represented as strings. This isn't the most efficient way to do arbitrary precision arithmetic, but it is convenient, it does avoid needing to know what type mp_limb_t and mp_size_t actually are, and it does show how to use the underlying library if you want to build something more efficient.
Performance appears to be excellent, but I can't take any credit because libffi is doing most of the work. The demos/time-libm.tcl script compares ::ffidl::callout wrapped libm functions to the Tcl expr versions of the same functions. On UNIX, the "make test" target will build libmathswig0.5.so to provide a SWIG wrapped libm for comparison purposes. If you're running on Linux-x86 or Windows you can install Robin Becker's ::dll for another data point. time-libm.tcl will time them on the same functions.
The Ffidl bindings to libm run a little faster than ::dll bindings and a little slower than SWIG wrappers, all of them coming in slower than expr itself. The bottom line is that all three extensions run better than 1.5 times the speed of expr over the same functions.
A few issues have been closed since the initial release.
The Windows port is done.
Callbacks are implemented for the x86 architecture.
But there are many open issues.
Port to macintosh and other architectures.
Importing libdl compatability code from Tcl. It would be real nice if the Tcl core exported some neatly packaged libdl abstraction, since it has the wherewithal implemented for load already.
Finding the right library is a pain. dlopen("libm.so") finds libm on my machine, but dlopen("libc.so") returns an error string decorated with binary characters while dlopen("libc.so.6") works. If you work with shared libraries you build yourself, it's not an issue, but for all the standard stuff there is no standard. In demos/ffidlrt.tcl the ::ffidl::find-lib function provides an abstraction for at least removing these issues one layer away from your Ffidl bindings to the library, but the implementation of the abstraction hasn't gone farther than listing where I find my standard libraries.
Discovering what type a type is is also a pain. Include headers are typically so heavily conditionalized, that one needs to search and search to find which typedef is actually implemented. In demos/ffidlrt.tcl the ::ffidl::find-type function abstracts these issues out of the Ffidl bindings, but again the implementation of the abstraction will need some work.
A backend for SWIG which generates Ffidl bindings might be useful.
There are some more pointer types which ought to be defined: a variant of pointer-var which requires an unshared value; a pointer to a native character string -- but couldn't that be pushed back to the Tcl layer?
Writing Tcl extensions with Ffidl is very much like writing C code in Tcl. I'm not sure what the actual required skill set is. But if you're not sure what you're doing, you might be in over your head. In any case, try not to take the core dumps personally.
Loading snippets of code into a Tcl interpreter with Ffidl loaded could be very hazardous, as in downloading "Try ME!" scripts from the web. There is no Ffidl_SafeInit(), we'd probably need signed scripts to even begin to consider such a thing.
I've looked at SWIG and at dll and seen that they very carefully duplicate any shared Tcl_Obj before attempting a conversion to Int or Double. I've also looked at the source for Tcl's expr command, and it converts objects to Double or Int and only duplicates shared objects when it finds a valid Int or a Double with an existing string representation. Ffidl only duplicates shared objects when processing pointer-var, though I'm open to explanations why it should do otherwise. It seems that if you pass a parameter to a typed function that you shouldn't be upset if the parameter is converted to that type.
Hmm, this is a really pared to the bone. It would be nice for newbies and experimenters and the careless if Ffidl implemented a debugging mode which verified that constraints were observed: 1) that Tcl_Obj string reps were not modified, 2) that Tcl_Obj bytearray reps were not modified outside their allocated sizes, and so on. This could be done by switching in an alternate implementation of tcl_ffidl_call() which made copies and verified the constraints after the call.
Some naming consistency in the demos. I seem to be reinventing my Ffidl extension style each time I start a new example.
Some style consistency in the tests. The tests just run, some generate descriptions, some report what they've done, some say nothing, some give summaries.
Robin Becker's ::dll package, which does much the same thing as Ffidl, provided the immediate inspiration for this work and pointed to the solution of some of the design issues for me. And Robin hisself has been very helpful.
Anthony Green's libffi package provided most of the initial implementation of Ffidl.
Bruno Haible's ffcall-1.6 package for the clisp system provided an alternate implementation and a truly amazing example of cpp macrology.
Ffidl Version 0.5, Copyright © 1999 by Roger E Critchlow Jr, Santa Fe, NM, USA
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the ``Software''), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL ROGER E CRITCHLOW JR BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
Note that libffi is Copyright © 1996-1999 Cygnus Solutions and bears a similar copyright notice, license, and non-warranty.