A PPC native Roadshow will likely do better, because the TCP/IP stack still spends a lot of time copying data from one memory buffer to another one. The copying code which Roadshow uses is the fastest 68k implementation I know of, and it doesn't seem to do too poorly in emulated form. Still, a PPC native copying routine would likely do a lot better.
PPC chips are not actually very good at copying speed. So if you got any advantage from a PPC native implementation it would likely be due to other reasons.
Another way of saying it is: JITs work really really well at small loops that execute over and over again millions of times. They have a lot of trouble executing "random" code that does not run in a loop, or only executes a few times in each loop.
Because of contractual obligations, however, I am unable to provide a PPC native variant of Roadshow for MorphOS...
Google Translate version: Hyperion will sue my a$$ off!
