Thomas Witzel wrote:
The option -fprefetch-loop-arrays generates prefetch commands for
non-vectorized code,
but not for vectorized one. Is that the intended functionality ? Is
there a way to get prefetching also
for the vectorized routines ?
Thanks, Thomas
Example: (compiled with g++-4.5.0)
g++ -O3 -fcx-fortran-rules -fprefetch-loop-arrays -mtune=core2
-march=core2 -mssse3 -S -c ../test_loop.cpp
The code for a complex multiplication loop done this way:
void f(std::complex<float> *a, std::complex<float> *b, std::complex<float> *r)
{
for(std::size_t s=0; s<N; s++)
r[s] = a[s]*b[s];
}
Is generated two-fold, one vectorized (.L3) and one not (L5):
It's certainly hard to guess the effect of pre-fetching only in the
remainder loop (early 32-bit pentium4 style?). As you have set
-mtune=core2, it seems reasonable the compiler would not optimize for
Athlon-32, which may have been the most recent common CPU without
effective hardware prefetch for vectorized loops.
I don't really expect gcc to attempt further optimization specific to
-mssse3, now that it's about 2 years out of production.
