Ewgenij Gawrilow wrote:
David Daney wrote:
Ewgenij Gawrilow wrote:
Hello,
is somebody out there who has consciously used the asynchronous unwind
tables? The GCC documentation does not reveal much more about this
feature than the mere statement of its existence.
What I eventually need is the possibility to catch certain signals and
unwind a part of the stack down to a certain frame in a mixed C / C++
program.
A naive approach of throwing an exception from the signal handler does
not work, neither does a siglongjump from the handler unwind the stack
frames between it and the frame where sigsetjump was executed (although
some C++ ABI drafts suggested this behaviour).
The main problem is probably the kernel-generated frame of the signal
handler, lying between the frame where unwinding starts and the catching
frame. If I set up the signal handler to use the alternative stack,
this frame won't be an obstacle any more; but I didn't find any clue how
to tell the unwinding machinery (_Unwind_RaiseException and the
relatives) as it is implemented in libgcc, that it has to scan over a
different stack.
libgcj (GCC's java runtime library) does exactly what you describe for
synchronous signals (SIGSEGV, SIGFPE). Throwing exceptions through such
signal frames does work for both i386 and x86_64 (as well as many other
architectures) on linux. Unwinding from arbitrary asynchronous signals
(SIGUSR1 etc.) is not supported.
The code that does it can be found in libjava/prims.cc
Do you (or somebody else) know by chance what's the major difference between the signal handler
frames generated by SIGFPE and, say, SIGINT? Is it documented anywhere?
The problem isn't the signal handler frames. The unwinder can unwind
through them all.
The problem is that you must have accurate unwind tables in 100% of your
code (including libc, ld.so, and all other possible libraries that you
might be calling) if you want to be able to unwind from asynchronous
signals.
With SIGFPE and SIGSEGV, the compiler knows exactly where they might be
generated and only needs accurate unwind tables at those locations.
David Daney
