Rules for alignment and non-NULLness of references

[RalfJung]

Hmm, but if the @llvm.memcpy call states an alignment (greater than one), LLVM is presumably allowed to optimize under the assumption that the alignment is correct, so even though nothing is read it might still be UB.

True. It would be helpful to know when and where exactly LLVM starts making alignment assumptions (in the IR generated by rustc) -- then we could define Rust's alignment assumptions accordingly. Currently miri does not check empty accesses, but it does check on every & -- which can't be entirely right, because you are supposed to be able to do &unaligned.field as *const _.