[llvm] [SystemZ] Don't lower float/double ATOMIC_[LOAD|STORE] to [LOAD|STORE] (PR #75879)

[James Y Knight via llvm-commits]

jyknight wrote:

I don't really know what arsenm or @preames has/had in mind, but I think there is value in all the backends working in a similar way here, rather than SystemZ doing something different.

The reason I'm skeptical about merging the two operations is that there is a significant difference between LOAD and ATOMIC_LOAD beyond just the memory-ordering: a LOAD (even in some cases when it's volatile!) can be split (both by SelectionDAG expansions, and potentially within the hardware instruction that was selected). An ATOMIC_LOAD cannot be. Given the difference, I'm not sure merging the two into one opcode actually makes things simpler.

https://github.com/llvm/llvm-project/pull/75879