The real and imag-pointers should be marked with "restrict"-keyword to enable aliasing optimizations. This goes for all other pointers also. And as others have already said, leaving optimzation at -O1 and not targeting your architecture (-march) can also have big impact.

    real[j + halfsize] = real[j] - tpre;
    imag[j + halfsize] = imag[j] - tpim;        // 1. Reading imag[j]
    real[j] += tpre;                            // 2. Writing real[j], which could be same data as imag[j]
    imag[j] += tpim;                            // 3. Must re-read imag[j] even though we read it in #1, because real[j] might alias it