Question

Do pointer offsets take any extra time? Which is preferable?

mov rcx, length
dec rcx
mov rax, qword ptr [rsi+rcx*8]

or

mov rcx, length
mov rax, qword ptr [rsi+rcx*8-8]

Answer 1

Using an offest should be preferred, on a modern architecture this can be faster, but never slower.

• It is one instruction less. Instruction decoding is a limited resource.
• The length of the code sequence is one byte shorter, becuase the offset is in the char range. This needs less cache memory.
• The dependency chain is one instruction shorter. In your first sequence the address calculationj hast to wait for the result of the decrement.
• Address calculation is done in separate units. There is no speed penalty when using complex addressing modes.

On the other hand there is a chance the the decrement will be reordered to be executed earlier, so that is won’t make a difference speed wise.
See more background information at http://www.agner.org/optimize/#manuals .