Question

I tried to find out the speed difference between plain loops, loop loops and builtin rep loops. I wrote three programs to compare the behavior:

Program 1

_start: xor %ecx,%ecx
0:      not %ecx
        dec %ecx
        jnz 0b
        mov $1,%eax
        xor %ebx,%ebx
        int $0x80       # syscall 1: exit

Program 2

_start: xor %ecx,%ecx
        not %ecx
        loop .
        mov $1,%eax
        xor %ebx,%ebx
        int $0x80

Program 3

_start: xor %ecx,%ecx
        not %ecx
        rep nop # Do nothing but decrement ecx
        mov $1,%eax
        xor %ebx,%ebx
        int $0x80

It turned out the third program doesn’t work as expected, and some recherche tells me, that rep nop aka pause does something completely unrelated.

What are the rep, repz and repnz prefixes doing, when the instruction following them is not a string instruction?

Answer 1

It depends. rep ret is sometimes used to avoid bad performance of jumping directly to a ret on certain AMD processors. The rep (F3) and repne (F2) prefixes are also used as Mandatory Prefix for many SSE instructions (for example they change packed-single variants to scalar-singe or scalar-double variants). pause (spin lock hint) is an alias of rep nop. Some other new instructions use a “fake rep prefix” as well (popcnt, crc32, vmxon, etc). The “fake” or Mandatory Prefix comes before the optional REX prefix, so it can’t be said to be part of the opcode, it really is a prefix.

Other operations generate an #UD if prefixed with a rep.