Re:Not that fast (Score 1)

You can think of the x86 front-end as a dynamic hardware compiler to the VLIW core. The optimizations end up handling a smaller instruction window than you would have with a software compiler, do to the smaller window you have to operate, but remember the x86 code was already compiled with parallelism in mind as well, so it's quite likely that the parallel instructions lie very close to one another.

I've read that some studies have shown that this sort of dynamic recompilation of code can lead to a reasonable improvement in performance on average (10% to 15% in a study x86 to x86 recompilation) even with well optimized code.