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

[Jonas Paulsson via llvm-commits]

JonPsson1 wrote:

> I don't know if this patch also works, but why not translate atomic_load_8/etc patterns directly in tablegen to instructions, and skip the generic LOAD entirely, like e.g. X86 does:
> 

On SystemZ, the memory instructions are only atomic if naturally aligned, so the non-atomic ISD opcode would still be needed for unaligned accesses.

I however now see a problem with this patch as the reg/mem folding does not seem to be working as I had expected, which need to look into.

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