Addressing two unrelated arrays with [base+index*scale] without UB

I'm trying to figure out how to get GCC or Clang to use [base+index*scale] addressing to index two unrelated arrays, like this, to save an add instruction:

;; rdi is srcArr, rsi is srcArr - dstArr, rax is len of arraysLoop:  mov     ecx, dword ptr [rdi + 4*rsi]  mov     dword ptr [rdi], ecx  add     rdi, 4  cmp     rdi, rax  jb      .Loop

The C++ code below achieves this, but because dst and src are unrelated, subtracting them is UB, even though the difference is never dereferenced. (I think this will nonetheless work on all x86/64 hardware though.)

#include <cstddef>#include <cstring>void offset(float* __restrict__ dst, const float* __restrict__ src, size_t n) {    const float *enddst = dst + n;    const std::ptrdiff_t offset = src - dst;    while (dst < enddst) {        std::memcpy(dst, /* src */ dst + offset, sizeof(float));        dst++;    }}

(Ignore the silly loop body, it's just for illustration as something using the addresses.)

Is there a way to do this without UB?

godbolt with non-UB and UB version