On 26/04/2018 11:26, Jack Andrews wrote: > This reduces it to a one-platform problem. gcc -O different to gcc > Maybe I'm being idealistic, but why should optimization change results? Because, for example on x86 platforms, 'gcc -O0' will use the 80-bit x87 stack regs, while 'gcc -O2' will use the 64-bit SSE regs. It is better to give up the notion that floating point computation are exact, and accept the fact that small errors do change the results on different implementations (and as pointed out, even on the same implementation with different options). It is also worth pointing out that sometimes these small errors accumulate into huge errors. Floating point is tricky. cf. https://en.wikipedia.org/wiki/Loss_of_significance Regards.
