[LLVMdev] The nsw story

[Dan Gohman]

On Dec 5, 2011, at 11:55 AM, Paul Robinson wrote:
> 
> On Thu, Dec 1, 2011 at 9:37 AM, David A. Greene <greened at obbligato.org> wrote:
> Dan Gohman <gohman at apple.com> writes:
> 
> > Prohibiting poison values from propogating through memory would mean
> > that the reg2mem pass would no longer be a semantics-preserving pass.
> 
> Or it means you couldn't demote those values.
> 
> If reg2mem is constructing spill slots, or address-not-taken locals, I could
> see classifying those as okay-to-poison. You would not want statics or globals
> or pointer-addressed memory to be poisonable, as those should be treated
> as being observable locations.
> 
> So, it's not really "all of memory" that could be poisoned, only these
> compiler-generated things.

This would mean that it's not valid to do store-to-load propagation
in the optimizer if the pointee might be a global variable. That
would mean the loss of real optimizations.

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.

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.

Dan