Dropping to the raw x86 PREFETCHT0/T1/T2/NTA instructions via inline assembly is what JIT compilers like LuaJIT and hand-tuned codecs (x264/x265) do when they need finer control over cache-level placement than __builtin_prefetch's abstraction allows. This is genuinely low-level work — Agner Fog's optimization manuals devote entire sections to picking the right temporal hint, since prefetching into the wrong cache level can waste memory bandwidth instead of saving it.