Re: [PATCH] volatile global register variable

[David Edelsohn <dje at watson dot ibm dot com>]

>>>>> Mark Mitchell writes:

Mark> I'm not sure how to solve the broader problems that David and Ian have
Mark> raised.  David points out that being sure never to optimize volatile
Mark> register variables may be hard; Ian points out that we already have to
Mark> do that for volatile memory, so maybe it's not too hard.  I don't know
Mark> how hard it might be, but it does seem like abstractly the right thing;
Mark> the rules for a volatile register variable ought to be the same as the
Mark> rules for any other volatile object.

	My concern is not about volatile, but about global register
variables having overloaded and ambiguous semantics.  I specifically think
we should avoid global register variables implying volatile.  Some
constituency will not be happy -- either wanting more aggressive
optimization or being unhappy about an optimization transformation.  If a
programmer really wants a volatile global register variable, he or she
should use the volatile modifier.

	We can ensure that volatile is used consistently throughout the
compiler.  I want to avoid GCC testing "volatile or global register"
because someone is going to forget to add the latter and introduce a
latent bug.  That likely will lead to long arguments about what the
semantics should be.  I want to stop that thread before it starts.

David