[llvm] [AMDGPU][SDAG] Try folding "lshr i64 + mad" to "mad_[iu]64_[iu]32" (PR #119218)

[Jay Foad via llvm-commits]

jayfoad wrote:

> This is essentially a revert of #76285.

I wouldn't go that far. It fixes a code quality regression caused by that patch in the specific case of the divisor $C0$ being $2^{32}$.

https://github.com/llvm/llvm-project/pull/119218