Open
Description
I have a piece of code that computes different floating point results depending on whether the computation happens at compilation time or at runtime (see https://godbolt.org/z/q65q4f7c3 ).
When the values are known at compile time the compiler contracts the operation and computes the result. When they are not known, the expression is computed at runtime without contraction.
#include <iostream>
#include <cmath>
#include <iomanip>
int main() {
double x = 30.508474576271183309;
double y = 6.1016949152542370172;
double r = (x-std::trunc(x/y)*y);
std::cout << std::setprecision(17);
std::cout << "compile time (w/ fma): " << r << std::endl;
volatile double *xx = &x;
volatile double *yy = &y;
r = (*xx-std::trunc(*xx/(*yy))*(*yy));
std::cout << "runtime (w/o fma): " << r << std::endl;
r = std::fma(std::trunc(*xx/(*yy)),-(*yy),*xx);
std::cout << "runtime (w/ explicit fma): " << r << std::endl;
return 0;
}
The output is (with -O3 -ffp-contract=on
):
compile time (w/ fma): -1.7763568394002505e-15
runtime (w/o fma): 0
runtime (w/ explicit fma): -1.7763568394002505e-15