• voxel@sopuli.xyz
        link
        fedilink
        English
        arrow-up
        1
        arrow-down
        1
        ·
        edit-2
        5 months ago

        both apple m cpus and snapdragon have hardware magic to make translation viable when paired with usermode tools like Rozetta or Prism
        otherwise it’s painfully slow, as demonstrated by both box64 and fex.

          • voxel@sopuli.xyz
            link
            fedilink
            English
            arrow-up
            1
            ·
            edit-2
            5 months ago

            well apple cpus have a hardware flag which slightly changes side effects of some existing arm instructions AND completely changes the memory model to one that imitates Intel x86 for the current thread. (TSO)

            with this flag enabled, the code can be almost fully recompiled ahead-of-time (from x86_64 to arm64) with minimal overhead (no need to remap memory addresses etc), with no reason to resort to e.g. jit recompilation on the fly (which is exactly what box64 does all the time)

            i assume snapdragon and microsoft are doing a similar thing (it definitely involves some hw magic too, maybe not exactly the same as in apple) since it’s seemingly the only way to achieve parity with apple in terms of performance