But the original benchmarks weren't multithreaded, which should make that a nonissue? (Is it possible to generate code assuming a weaker memory model, and then

But the original benchmarks weren't multithreaded, which should make that a nonissue? (Is it possible to generate code assuming a weaker memory model, and then invalidate that upon creation of a thread? Might get dicey with shared memory, though...)