MOVing between two memory addresses
Your suspicion is correct, you can't move from memory to memory.
Any general-purpose register will do. Remember to PUSH the register if you are not sure what's inside it and to restore it back once done.
That's correct, x86 machine code can't encode an instruction with two explicit memory operands (arbitrary addresses specified in [])
- Why isn't movl from memory to memory allowed?
- What x86 instructions take two (or more) memory operands?
Whats the recommended register
Any register you don't need to save/restore.
In all the mainstream 32-bit and 64-bit calling conventions, EAX, ECX, and EDX are call-clobbered, so AL, CL, and DL are good choices. For a byte or word copy, you typically want a movzx load into a 32-bit register, then an 8-bit or 16-bit store. This avoids a false dependency on the old value of the register. Only use a narrow 16 or 8-bit mov load if you actively want to merge into the low bits of another value. x86's movzx is the analogue of instructions like ARM ldrb.
movzx ecx, byte [rdi] ; load CL, zero-extending into RCX
mov [rdi+10], cl
In 64-bit mode, SIL, DIL, r8b, r9b and so on are also fine choices, but require a REX prefix in the machine code for the store so there's a minor code-size reason to avoid them.
Generally avoid writing AH, BH, CH, or DH for performance reasons, unless you've read and understood the following links and any false dependencies or partial-register merging stalls aren't going to be a problem or happen at all in your code.
- Why doesn't GCC use partial registers?
- How exactly do partial registers on Haswell/Skylake perform? Writing AL seems to have a false dependency on RAX, and AH is inconsistent
(or should I use the stack instead)?
First of all, you can't push a single byte at all, so there's no way you could do a byte load / byte store from the stack. For a word, dword, or qword (depending on CPU mode), you could push [src] / pop [dst], but that's a lot slower than copying via a register. It introduces an extra store/reload store-forwarding latency before the data can be read from the final destination, and takes more uops.
Unless somewhere on the stack is the desired destination and you can't optimize that local variable into a register, in which case push [src] is just fine to copy it there and allocate stack space for it.
See https://agner.org/optimize/ and other x86 performance links in the x86 tag wiki