[llvm] [AMDGPU] Fix isCanonicalized for IEEE-mode-dependent sNaN quieting (PR #208733)

[Wooseok Lee via llvm-commits]

wooseoklee wrote:

> The premise of the the description is wrong. These are canonicalizing operations, and aren't any different from fadd. The bigger problems are 1. the quieting only happens for ieee=0 (or for gfx12+)

I tried to address by gating isCanonicalized on MODE.IEEE=1 || GFX12+, confirmed by hardware testing across gfx90a, gfx950, gfx1100, and gfx1201.

> the way this optimization is dubious given what the IR rules are. Technically, we need to disallow canonicalize dropping behaviors by definition of the DAG operations, or perform these folds in a post-selection pass.

This is not currently implement with this PR as it is a larger architectural change, leaving it as a future work. 

https://github.com/llvm/llvm-project/pull/208733