The port of the triton-shared pass to our raise-block-pointer pass revealed some limitations: - Unstructured memory access masks cannot not be directly forwarded to block pointer access. A more advanced handling is required. - Modulo are not handled in a safe way that meets production standards (see https://github.com/intel/intel-xpu-backend-for-triton/pull/1570#discussion_r1666781372 ) These two limitations must be handled safely. Compiler hints could help handling these two limitations.