ISO 9899 7.13.2.1 (3) says that "[...] the values of objects of automatic
storage duration that are local to the function containing the invocation of the
[...] setjmp macro that do not have volatile-qualified type and have been
changed between the setjmp invocation and longjmp call are indeterminate."

We should probably be marking all auto variables in the scope of a setjmp
volatile, to be safe. So far, this has not bitten us, but this seems to prove
only that we have been lucky.

This should definately be fixed, but we probably want to make it conditional
somehow.  If we don't, every variable in a function which uses 'invoke' will
have to be marked volatile.  That's not cool.

Also, note that this only effects platforms which do not use the "snapshot of
register state" way of implementing setjmp/longjmp.  I do not believe this
effects sparc or X86.

Fixed like so:
http://mail.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20031110/009568.html

Any targets which do not require volatile markers on variables which setjmp
should extend this patch to avoid slowing down code that uses exception
handling.  I don't know whether the Sparc fits in this class or not, so I
conservatively didn't include it.

-Chris