Like Reid, I also think that the default should be on. On our system,
we need to be able to handle asynchronous exceptions that might occur in
routine prologue/epilogue as well as supporting debuggers, etc that all
depend on the unwind information for correct behavior.
In my opinion, we don't even do enough to handle
-fasychronous-unwind-tables correctly. The clang driver simply turns
that into -funwind-tables.
Our OpenVMS work will be working on getting sufficient .cfi directives
in place to support our needs. We discussed this a little bit a few
months ago with respect to tweaking the compact form but have decided to
rely on the DWARF CFI style EH data for prologue/epilogue regions.
John