Fix for PR2966: va_start/va_end use the *address* of the

va_list, not the value.  Patch fixes an assert on 
x86-64 and a wrong-code bug on x86.  I'm not completely sure, 
but I think it's a regression from r58306.

Does anyone run regression tests regularly on x86-64? The crash should 
have been picked up there.

llvm-svn: 58366