It would be nice to get rid of the 'op' local variable. I think it would help my work on bug 442379. In particular, GCC seems to have a hard time optimizing this variable in the call-threading version because read-write dependences cross op dispatch. Also, eventually I will want to stop updating regs.pc as well, so this is a good first step. The perf results look OK in SunSpider as well: 1% speedup overall. Most benchmarks are a few percent faster, a few are a few percent slower. The worst is math-cordic, which is 4.6% slower. I don't know why that happens, but is may be that GCC doesn't CSE *regs.pc when it is used more than once in a case. We could consider applying that optimization manually.

This was fixed in bug 910782.