My first guess is that it should be (because addition on FP gives the nearest representable value), but I'd also guess that any small differences here are irrel

My first guess is that it should be (because addition on FP gives the nearest representable value), but I'd also guess that any small differences here are irrelevant. I wonder how many of those optimizations (e.g. maybe this exact one) would be done by compilers with -ffast-math.