The key is that if you do this, different optimization levels produce different numerical results, and also some problems in the code become unfixable; you can'

The key is that if you do this, different optimization levels produce different numerical results, and also some problems in the code become unfixable; you can't group computations according to their expected value ranges because the compiler will ungroup them again, incorrectly assuming that FP addition and multiplication are associative. Certainly for some applications it's fine, but a test of whether it's fine would be, for example, that it works just as well with single-precision as double-precision and even 16-bit floats would be fine, for example weights in an NN.