When InstCombine replaces a memcpy with loads+stores it does not copy over the llvm.mem.parallel_loop_access from the memcpy instruction.
BTW, currently our memcpy's don't get annotated with llvm.mem.parallel_loop_access by clang (just opened a PR for this one: 28980). So this scenario is not yet reproducible automatically (need to add the llvm.mem.parallel_loop_access manually to the memcpy in the .ll).
Can add here that this is related to PR28981, for the broader context.