So here's what happens with empty. If you run it from the shell, first the shell will fork, then try to exec empty. But the exec fails, since empty doesn't begin with a magic value. Therefore the exec call returns an error. Now the shell picks up this error, and then tries to run the program again, this time as an argument to an invocation of the shell (i.e., it does an exec of /bin/bash, passing it "empty" as a parameter). This is why empty ends up taking longer to run.

This is the normal pattern, just in case you forget to put "#!/bin/bash" at the top of the script, so that the script can be run anyway. This is also a source of confusion for some sysadmins, when a script works from the command line but not from something like a cron script.

Before the system can execute any one of these programs, it has fork the bash process, call exec to load the new binary in the process memory, load the standard library and map it to the adress space, open stdin / stdout, run the program, close stdin / stdout, wait for the process termination ...

Comparing to all this, a move instruction in userland won't really make a difference.

Edit: List on system calls required to execute trivial.c (on my Linux):

execve, brk, access, mmap, access, open, open, open, open, open, stat, open, stat, open, stat, open, stat, open, fstat, mmap, close, access, open, read, fstat, mmap, mprotect, mmap, mmap, close, mmap, mmap, mmap, arch_prctl, mprotect, mprotect, munmap, exit_group

I get the same results:

    $ touch empty

    $ chmod +x empty 

    $ time ./empty 

        real    0m0.005s
        user    0m0.001s
        sys     0m0.001s

    $ echo "int main(){return 0;}" > trivial.c

    $ gcc trivial.c -o trivial

    $ time ./trivial

        real    0m0.002s
        user    0m0.001s
        sys     0m0.002s