We can't allow vectorization of widneing reduction patterns like DOT_PROD
and WIDEN_SUM when their results are used in the loop. This is because
these idioms compute N widened results and then reduce them into N/2
results by summing-up pairs of results. So if we need to use (say store)
the original N results we cannot use this idiom. (In innermost-loop
vectorization this cannot happen because a reduction cannot be used in the
loop. However, in outer-loop vectorization, a reduction in the inner-loop
can be used in the outer-loop).

Bootstrapped with vectorization enabled and tested on the vectorizer
testcases on powerpc64-linux. Committed to autovect branch.

dorit

        * tree-vect-analyze.c (vect_mark_relevant): Don't mark widening
        reduction patterns as relevant if they are used in the outer-loop
        (we want to vectorize the original sequence instead).
        (vect_mark_stmts_to_be_vectorized): Add an assert.
        * tree-vect-transform.c (vect_finalize_reduction): Set a stmt_info.

        * gcc.dg/vect/vect-outer-4e.c: New test.
        * gcc.dg/vect/vect-outer-4f.c: New test.
        * gcc.dg/vect/vect-outer-4g.c: New test.
        * gcc.dg/vect/no-section-anchors-vect-outer-4h.c: New test.
        * gcc.dg/vect/vect-outer-4i.c: New test.
        * gcc.dg/vect/vect-outer-4j.c: New test.
        * gcc.dg/vect/vect-outer-4k.c: New test.
        * gcc.dg/vect/vect-outer-4l.c: New test.
        * gcc.dg/vect/vect-outer-4m.c: New test.

(See attached file: autovectfix.txt)