I didn't look at the code, but I often use something similar in assembly when hyper-optimizing loops. The fastest memcpy-like loop structure I've found on x64 involves using a negative offset from the end of the region, which is then incremented until it hits zero.

The advantage of this construct is that you can use the flags set by the increment to test for loop termination, rather than needing an additional comparison. The problem in C is that the compiler tends to "unoptimize" it for you after seemingly innocuous changes.

Anyway, the loop ends up looking something like this:

  char *end = start + size;
  size_t neg = -size;
  do { 
    register = *(end + neg);
    process(register);
    neg += advance;
  } while (neg);
    

For certain loops this construct shaves off 1 cycle per iteration, at the cost of ~1 cycle of initial setup cost. I'm wagering that he was doing something similar, but just kept the name "size" after the negation, rather than switching to "neg" as I did here.