[llvm] [AMDGPU] Add a target option to disable aggressive FMA fusion (PR #161450)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> This is because FMA fusion may cause numerical errors when we want bitwise accuracy. 

This is a property of the incoming IR and cannot be addressed by changing backend costing heuristics. What IR are you using? You have to change that 

https://github.com/llvm/llvm-project/pull/161450