On regtest there is only 1 bit of room in the targets. So how about:

const uint32 DIFF_WEIGHT_PRECISION = 1000000;
last_target /= DIFF_WEIGHT_PRECISION; // outside the loop

uint32_t adj_time_i = time_i * ((target_i / DIFF_WEIGHT_PRECISION) / last_target).GetLow64();

Looks good except we need a temporary rather than updating last_target. Will do.

Sorry

uint32_t adj_time_i = ((time_i * (target_i / DIFF_WEIGHT_PRECISION)) / last_target).GetLow64();

I notice that it divides by last_target inside the loop, then multiplies the result by last_target outside the loop.

Seems like the calculations could be done more efficiently removing the division inside the loop (and the subsequent multiplication outside)

Great catch. This will also remove the need for fixed precision, we'll just need to accumulate in a 256-bit var.