Re: [PATCH 5/5] First version of REPL Ctrl-C trampoline for Woe32

[Bruno Haible]

José E. Marchesi wrote:
> I am CCing Bruno.  He may have some ideas about this.

Yes I have a couple of ideas about this.

1) Move the 3 methods Throw, RaiseInThread, DontOptimize out of the class
   ThreadAbort. We don't like C++. Use plain C syntax whereever possible.
   (You don't need a C++ class in order to use 'throw' and 'catch'.
   You can use them in plain static functions in a .cpp file.)

2) Once this is done, the only use of class ThreadAbort is in the functions
   Throw and pk_tramp_cmd_exec. What happens here is unwinding the stack.
   Unwinding a stack that contains C++ function invocations can only be
   done with 'throw' and 'catch'. But poke is a C program; it does not
   contain C++. (This is on purpose.) For C programs, 'longjmp' and 'setjmp'
   are the equivalent. So, rewrite the Throw and pk_tramp_cmd_exec to use
   a jmp_buf instead. (This does not need to be the same jmp_buf as in
   pk-repl.c. Better do the simplifications one by one.)

Does the Ctrl-C handling still work? I hope so. Then:

3) The RaiseInThread function is a dirty hack, but it is the way to do it.
   GNU libsigsegv does a similar hack. The hairy part is to get the stack
   alignment right.

Why does this actually work? At the point where pk_tramp_cmd_exec returns,
callee-saved registers need to have certain values. The RaiseInThread
functions dealt with the program counter and stack pointer; the longjmp
restored the remaining register values.

4) Now try to eliminate one of the two jmp_bufs. We have one already
   in pk_repl.c. Use that with setjmp/longjmp on platforms where sigsetjmp/
   siglongjmp does not exist.

5) If and only if the approaches above don't work, I would suggest to use
   the 'ansi-c++-opt' module instead of AC_PROG_CXX. Then poke can be
   built on Unix systems without a C++ compiler.

Bruno