Besides the issues with bytecode, interpreters, and all that other good stuff, I've found the the most serious slow down in any of my programs has been my (bad) design decisions. I find that bottleneck, research the problem, and usually come up with something that does it orders of magnitude faster.

I guess it would be nice to pass around bytecode, but not being able to hasn't really stopped me from doing anything. Although a few other people have had other experiences, I've found most of my performance suckers happen at runtime, so the compilation time wasn't that big of a problem.

The problem I dread, however, is that the same bytecode acts differently with different versions of a virtual machine (or the other way around). I experienced that situation in the early days of Java, althougth that was mostly because they were changing the names of a lot of things. I feel pretty good about passing the source around and letting the latest parser-interpreter combination figure it out.