[clang] [llvm] [Transforms][Utils][PromoteMem2Reg] Propagate nnan flag on par with the nsz flag (PR #114271)

[Paul Osmialowski via cfe-commits]

pawosm-arm wrote:

> > Yet we see a value in the change proposed here too hence a good reason for merging it.
> 
> I've clearly misunderstood the context of the rebase. What is the value in merging this change if the original issue has already been resolved? I guess this fix is simpler and thus might help compile time but then that wasn't the original intent.

There are two views at this situation:

One, I can imagine many ways the `nnan` flag could not be retrieved if it isn't ensured (the way my patch is doing it), Dave's patch solves the problem more elaborate way we know it works in specific situations, yet how can we be sure there aren't other places in which `nnan` should be easily visible if a transformation should succeed?

The other is that there's a general movement away from having the function attributes (in favor of using only the instruction flags) and thus by adding another use we make this transition harder.

My PR will catch all cases and has very low complexity. Seems I need @nikic opinion on the subject, whether we should land it or abandon it.


https://github.com/llvm/llvm-project/pull/114271