https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94548
--- Comment #2 from fabrice salvaire <fabrice.salvaire at orange dot fr> --- Yes I missed this important point on 8-bit architecture ... This line doesn't also work for some reasons const unsigned long int f0 = (8*(10ULL)^(6ULL)) / (1000*256ULL); but this one works const unsigned long int f0 = (8*1000000ULL) / (1000*256ULL); It means the compiler computation of constants is a bit error prone on such architectures versus x64. There is no way in C to say compute the right value in infinite-bit then store the result in 16-bit. I am surprised by the fact the compiler doesn't warn for overflow ? Since I found the warning option "-Wno-overflow — Do not warn about compile-time overflow in constant expressions" I should get a warning ???
