The CPU JIT bug was a good catch. I just found out that this code produced wrong results on CPU backend (thanks @Myhroun for detecting weird numbers in a larger code base):

#include <arrayfire.h>
int main() {
    double data_[] = {1., 2., 3., 4.};
    af::array a{2, 2, data_};
    af::array b{2, 2, data_};
    a + a;  // remove this line and everything is fine
    af::print("difference", af::flat(a) - af::flat(b));  // nonzero difference!
}

And while browsing through the arrayfire code, I saw this commit that fixed the issue for me. Thank you and I am looking forward to a new release (or backport :-) ).