[llvm] Revert "LangRef: Clarify llvm.minnum and llvm.maxnum about sNaN and signed zero (#112852)" (PR #168838)

[Phoebe Wang via llvm-commits]

phoebewang wrote:

> > The point is since all targets share the same defination, a change on that needs consensus made through all targets. The hardware details (especially there are differences among different targets) are not the only consideration for the defination. And I explained them in details in the following 2 bullets.
> 
> This has been discussed at length in the previous threads, and this has never worked correctly across all targets. The operation you want is minimumnum

We need to clarify what is correct firstly. No matter how IEEE defines it, LLVM allows SNaN to be treated as QNAN in non constrained FP mode. Hence the current behavior is correct to me. To strictly match with IEEE, we can only leverage constrained intrinsics.

https://github.com/llvm/llvm-project/pull/168838