[Mingw-w64-public] setjmp/longjmp blues

[Roland S. <rol...@on...>]

Hi...

Presently I am migrating a big codebase from GCC 4.4.3 to 4.8.2 using 
the current trunk mingw-w64 crt and headers. I am using a self built 
toolchain with SEH for 64bit windows and sjlj on 32bit windows.

Everything appears to be fine when compiling for 32bit. But I am/was 
facing some trouble when running ObjectiveC code on 64 bit windows when 
using ObjectiveC (GNUStep) exceptions.

We are still using traditional ObjC exceptions.

NS_DURING
  <code which can throw an excpetion>
NS_HANDLER
  <exception handling code>
NS_ENDHANDLER

All these 3 NS_ constructs are macros wrapping around setjmp(). With 
resolved macros it looks like this:

{NSHandler NSLocalHandler;
_NSAddHandler(&NSLocalHandler);
if (!setjmp(NSLocalHandler.jumpState))
{
    <code which can throw an excpetion>
_NSRemoveHandler(&NSLocalHandler);
}
else
{
  NSException *localException = NSLocalHandler.exception
  <exception handling code>
}
}

The exception is "thrown" always from within an instance of the 
NSException class, which code is located in the gnustep-base dll. There 
the longjmp(jumpState,1) is performed.

With our version of gcc 4.4.3 (which Kai knows very well ;-)) everything 
is fine. Not so with gcc 4.8.2 in "release" mode (without symbols). Here 
the resulting binary crashes (most of the times) when an exception is 
raised (exactly when the longjmp() is performed).

After nearly loosing all of my remaining hairs in the last 3 days I 
found a workaround. As far as I can see in setjmp.h setjmp()/longjmp() 
wraps around _setjmp(jmp_buf,__builtin_frame_adress(0)) and longjmp() 
from msvcrt.dll.

When I use gcc's __builtin_setjmp() / __builtin_longjmp() instead of the 
msvcrt supplied versions it works again. Has anyone seen something mad 
like that before?

Just 2 more informations:

1.
When I do a normal setjmp()/longjmp() combo manually inside of an ObjC 
method everything is fine too. Works as expected. I assume when I cross 
dll borders (setjmp() in one dll and longjmp() in another) something is 
going mad. I would like to understand what is going on here.

2.
When I dump out the jmp_buf supplied to setjmp() and look at it it looks 
perfectly ok. Frameadress/Rip are correct. Even all other registers look 
ok. jmp_buf is also not corrupted in between of setjmp() and longjmp().

And only win64 bit is affected not 32bit. Presently I am compiling with 
optimization completely turned off and nearly no other gcc options 
enabled beside of -fno-omit-frame-pointer.

Thanks for your help,

Roland