[LLVMdev] The nsw story

[David A. Greene]

Dan Gohman <gohman at apple.com> writes:

> And it would still mean that non-volatile store instructions suddenly
> can have side effects beyond just writing to their addressed memory.
> That would be surprising.

No more surprising than any hardware architecture that implements this
kind of poison tracking.  You see the fault or undefined behavior at the
store.  Yes, a compiler IR is not an ISA but we do get to define the IR
semantics.  It's a contract where we get to specify all of the rules.
:)

If you're more worried about things like the above rather than the high
static analysis cost these various suggestions are trying to address,
that's simply a particular tradeoff you'd rather make.  There's no rule
that says any particular semantic is "surprising."

We make the tradeoffs we want and implement the semantics based on that.

> I think Dave also suggested making select instructions be
> observation points. That would mean that select instructions would
> have side effects. Again, that would be surprising.

How so?  It certainly is an observable point.  A select is just a branch
by another name.  If the program behavior was undefined in the presence
of a branch, why shoudln't it be undefined in the presence of a select?

                          -Dave