[Dwarf-discuss] Using CFI to end a backtrace

[Todd Allen]

> Several times recently I have wanted a "clean" way to indicate to GDB
> that a backtrace was over - give up, go away, nothing more to see here.
> This is useful in startup code, kernel exception handlers, et cetera.
> There's no immediately obvious way to use .debug_frame for this;
> however, it seems to me that marking the return address column as
> DW_CFA_undefined should do what I want.  If we can't find the return
> address, then there's no point in unwinding further.
> 
> Does this interpretation make sense?  If so, should it be documented,
> so that others don't need to reinvent it?
> 

It makes sense to me.  But it's architecture-specific.  Technically, there's
no rule that an architecture has to have a return address or PC column,
although I can't imagine a real one without something like that.  I'm not
sure how it would be documented, except maybe just as a suggestion.

BTW, we've done this by indicating a memory location that has a NULL in it.
That saves us from the compiler having to know every routine beyond which one
can't walkback.  It then becomes a choice belonging to the linker or runtime
system.  Of course, it only works if NULL isn't likely to be a valid PC.

-- 
Todd Allen
Concurrent Computer Corporation