[llvm] LangRef: Clarify behaviors of nsz in fast math flag (PR #137567)

[Matt Arsenault via llvm-commits]

https://github.com/arsenm commented:

Calling this out as a special case for min/max isn't the right way to express this. Given that this is the one standardized area with fuzzy signed zero handling, I think we should try to fit the definition such that this works (i.e. it's not permitted to randomly introduce a 0 sign bit).

I'm not aware of any transforms taking advantage of producing a -0 where there wasn't one to begin with. That is, we are allowed to produce either sign zero if one of the operation inputs is a -0, or the operation would naturally produce a -0 based on its inputs

https://github.com/llvm/llvm-project/pull/137567