Yeah, machine code written by a human is generally slower than that produced by a compiler. Particularly when you've got an architecture with oodles of registers, like the PowerPC's 32 or even the ARM's and AMD64's 16. Then there's stuff like pipelining, load latency hiding and so forth -- ain't no way that a human can write and refactor assembly code and get the kind of speed that a compiler can produce for code every time it's compiled.
That said, low-level optimizations typically matter for shit these days anymore, with cache misses taking up 70% of a process' execution time. Unless you've got a loop where you spend 90% of the program's time, just give the compiler "-O2 -fomit-frame-pointer -march=k8" and be done with it.