Perhaps it would be wisest to take a multipronged approach:
- push a version into testing with:
- the known aliasing issue workaround (char* instead of void*)
- and, -O2 for just the problematic plugin, and -O3 for everything
else, in particular for the library proper
- keep the unstable version at -O3 entirely, and rev to current
upstream CVS HEAD including char* fix
- try to isolate the problem with a short snippet of code to report
as a GCC issue
- via contact w/ the GCC people, figure out what the "right" fix is
here. My guess is that there is some magic incantation like
__attribute__(sacrifice_red_heifers,eye_of_newt,caldron_bubble)
that avoids the need to use weirdo casts or lie about types.
--Barak.
--
To UNSUBSCRIBE, email to [email protected]
with a subject of "unsubscribe". Trouble? Contact [email protected]
