Note that simple reductions, like level 1 BLAS, are memory-bound, and optimizing is generally only worthwhile for short arrays. (Contrast level 3, i.e. matrix-matrix operations.) Also it's generally arithmetically incorrect to vectorize them. That said, the BLAS test suites I've tried do pass when vectorized with gcc -funsafe-math-optimizations or the incorrect-by-default Intel compiler.

The general technique of restructuring arrays of data structures is a standard optimization. However, if you just write the computation described straight down in Fortran, (at least recent) gfortran -Ofast unrolls and vectorizes it. Hooray for Fortran types.