This means the code takes advantage of CPU pipelines and branch predictions, which means the source/destination pointer might be de-referenced and loaded into a register before the number of bytes is being inspected, or the assembly is executed with a 0 value, which still makes the CPU do a load or store on the src or dest pointer. This is being done to gain performance, and doing things in another order or explicitly validating the input decreases performance.
If then the src/dest pointer is NULL, the above scenario plays out badly, but in order to not sacrifice performance, the burden is placed on the caller to correctly use memcpy() instead of memcpy() detecting invalid use.
You might not agree with this as a design decision but I'm not arguing that this is the right trade-off, just answering:
> I don't think your optimization actually exists. In what possible world can we generate better code by assuming that for memcpy(x,y,n), when n==0, x!=nullptr||y!=nullptr?