(1) This is generally true, although transformations such as loop interchanging and collapsing can involve outer loops too. Conversely, loop materialization enables vectorization of straightline code.
(2) For double-precision data, loops with just two iterations can still be vectorized, although the compiler may, by default, reject such short running loops for efficiency reasons, like misaligned or too much setup/cleanup overhead. Use the vectorization diagnostics (-vec_report2) to inspect the reason why loop are not vectorized, and experiment with pragmas and other compiler hints to see if efficient vectorization is possible. Perhaps you can post an example?
Hope this helps.
Aart Bik
http://www.aartbik.com/
