I'm posting these patches to feel out the likelihood of this approach to
C++ exception compatibility ever getting adopted into Guile. They still
need some work in terms of function documentation, and one test failure.
----------------------------------------------------------------------
Currently, the use of libguile functions from C++ is fraught with
danger. C++ destructors are not triggered by guile exceptions, and the
dynwind mechanism is ill-suited for application to C++ data structures
(in particular, STL structures).
In my experience, writing a C++ function that takes data from guile code
in a truly safe manner roughly doubles the required amount of code,
because of the generally-required pattern:
// ...
// Some code to check that a SCM really is an int.
int dummy = scm_to_int(my_scm);
// Below this point, no guile exceptions may be thrown.
// Now, we can create a std::vector, or what-have-you, and put our
// int into it.
std::vector<int> my_vector;
my_vector.push_back(scm_to_int(my_scm));
// ...
This sequence of patches addresses this shortcoming by using hidden C++
exceptions to perform the stack unwinding when a guile exception is
thrown, so that the pre-checking demonstrated above is not needed.
# How is this implemented?
This sequence of patches does four things:
- Enables autotools C++ support.
- Adds -fexceptions to the compiler command line to enable C++
exceptions to traverse C stack frames.
- Adds extern "C" guards to the internal libguile headers.
- Adds a single C++ source file to libguile that contains set/longjmp
workalikes for prompt and exception handling. Prompt and
exception-related uses of set/longjmp are replaced with these
functions. (This is the third patch, and the real meat of the work).
# What are the downsides of this prototype implementation?
The use of -fexceptions and a C++ source file limit the choice of
compiler. GCC, Clang, and ICC should be fine. I have no idea about
some of the more esoteric compilers.
On most platforms, -fexceptions entails some space overhead. In my
testing, libguile gets bumped from 4.8 MiB to about 5 MiB.
The last two test cases in test-out-of-memory ("the kicker" and "ouch")
fail. I'm still investigating the root cause. All other tests pass.
