This boils down to using LLVMs `prefetch` intrinsic [1]. [1]: https://llvm.org/docs/LangRef.html#llvm-prefetch-intrinsic