[ Thanks to Dave
for this link. ]
“[Q:] Will the Code Morphing software adjust its x86
instructions to software optimized for MMX, SSE, and 3Dnow? In
other words will the code morphing software be able to allow the
Crusoe to emulate generic x86 only or will it be flexible as to
what the best emulation technique would be for a given
application?”
“[A:] We support MMX in hardware, but for now, I think it would
be criminally wrong for our CPUID to advertise that it supports
3DNow or SSE. That suggests a certain floating point performance
that would need specifically designed hardware to handle those
instructions.
“It is pretty flexible. For instance, it might make sense for us
to use MMX instructions instead of regular integer instructions if
we notice a lot of integer operation parallelism and little use of
the flags. We have the option of making that optimization. It would
be a lot harder for a traditional CPU to do this efficiently,
because it doesn’t have the bird’s-eye-view we do of large pieces
of code.”>
“For instance, when we notice a jmp instruction that always goes
to the same place, or even if it just mostly goes to the same
place, we often just delete the jmp and attach the destination code
at the end of that translation. The net effect is not only a
0-cycle branch, but an opportunity to get even better
scheduling.”