[llvm] [CodeGen] Mark mem intrinsic loads and stores as dereferenceable (PR #80184)

[Derek Schuff via llvm-commits]

dschuff wrote:

@jayfoad / @arsenm The code motion interpretation is the only one that makes sense to me. The description in the [header](https://github.com/llvm/llvm-project/blob/720007e4b782c2446c5f10dd1e7733f4b6b323f1/llvm/include/llvm/CodeGen/MachineMemOperand.h#L144) says that it doesn't trap, which would allow code motion of side-effecting instructions across it.

@arsenm there is one new change since I first uploaded, the one to llvm/test/CodeGen/AMDGPU/memcpy-scalar-load.ll. It looks to me like it might still be correct, but I'd appreciate if you could take a look.

@yonghong-song see my comment above about the BPF test. Does any of that ring any bells to you?

https://github.com/llvm/llvm-project/pull/80184