Hi all,
I've made some progress updating the callback interface for each backend to work with callback pointers in a manner that should be safe when saving/reloading images.
The branch is up at:
http://slacknet.com/~jamesjb/cffi-callback-fix
which can be obtained via 'darcs get', or pulled into an existing repository, or browsed over the web.
The only significant visible change is the addition of an exported %CALLBACK function to the CFFI-SYS layer, which returns a pointer for a callback defined by %DEFCALLBACK. CFFI:CALLBACK is implemented in terms of CFFI-SYS:%CALLBACK instead of using the symbol's property list.
This should also fix some problems on (at least) OpenMCL where the native callback was defined with a gensym'ed name, which means redefinitions of CFFI callbacks would not actually redefine the native callback.
In CMU CL, almost every test fails from a saved image, with or without these changes---I'm not sure what's up with that yet.
Allegro, CLISP, OpenMCL, and LispWorks now pass all expected tests from a saved image. (In CLISP, you need to reload foreign libraries by hand after loading the saved image. We might want to add another initialization hook to reload libraries loaded by CFFI...)
SBCL passes all tests from a saved image with the PURIFY workaround, and should work out of the box when the underlying callback bug is fixed.
I can't get the ECL backend to build, and I don't have access to Corman, so they are untested.
Jörg, in the CLISP implementation, I'm a little concerned about using FFI::EXEC-ON-STACK to convert the functions to callbacks, but I couldn't find another way to get the effect of FFI:WITH-C-VAR at runtime, given a parsed type.
Comments?
James