Re: Help: Unwinding the C++ stack...throw, longjmp & threads

[Jason Merrill <jason@cygnus.com>]

>>>>> George T Talbot <george@moberg.com> writes:

 > I'm working on a solution for deferred pthread_cancel() under Linux. 
 > What I want is to unwind the stack when the thread is cancelled so that
 > C++ destructors for objects on the stack get called.

 > If I recompile the C library with the -fexceptions flag, I can throw an
 > exception in a cancellation handler, and this does what I want.  My
 > platform is Linux x86.

 > However, after talking to Ulrich Drepper at LWCE, he says that
 > recompiling the C library with -fexceptions slows it down by around 5%. 
 > So, obviously, he won't accept a patch to the main-line source to do
 > this, with very good reason.

That doesn't make sense.  Compiling C code with -fexceptions should not
affect the code at all; it just creates unwind tables in another section.

 > I would like to write a function which unwinds the stack similarly to
 > what happens at a throw in C++, but which wouldn't require recompiling
 > existing libraries written in C.

 > Is this possible?

Yes; gdb does it by looking at the code for the prologue and figuring out
from there how registers have been saved.  See i386_frame_find_saved_regs
in gdb/i386-tdep.c.  Adding generic unwind support would be most welcome.

 > Can anyone point me to the source for how throw unwinds the stack?

See throw_helper in gcc/libgcc2.c.

 > Alternatively, could someone put me in touch with the folks that work on
 > C++ exceptions in gcc on the x86 processor?

That would be Andrew Macleod <amacleod@cygnus.com> and myself.

 > Also, does setjmp()/longjmp() assure that C++ destructors for
 > stack-based objects are called?

You mean, explicit calls to setjmp and longjmp in user code?  No.

Jason