[LLVMdev] nsw is still logically inconsistent

[David A. Greene]

Dan Gohman <gohman at apple.com> writes:

>> I'm not following.  If the promotion to i64 produces a different value,
>> then the nsw smeantic was violated, leading to undefined behavior, as
>> you note.  That that point all bets are off.  Divide by zero certainly
>> is a perfectly valid expression of undefined behavior.  If we had a
>> delayed check we would have to put it somewhere before the udiv.  We
>> would probably need some kind of fixup path _a_la_ IA64's check
>> instruction.
>
> The original code was well-behaved for all inputs.
> The final code has undefined behavior for some inputs.
>
> Fixup paths don't really seem to fit here. We're never going
> to want fixup paths in the final output, for example. And
> there is no way to fix up undefined behavior.

My point is that hoisting the div above its control dependence is
illegal unless you have some way of mitigating the poison path.  Or we
define new "poison" IR operations as Rafael suggested.  I'm sort of
intrigued by that idea.  It's similar to a check operation but we could
define whatever semantics we want (produce a safe value if poison,
etc.).

Again, it's undefined behavior so we get to define it.

                             -Dave