On Fri, Jun 10, 2011 at 9:19 AM, Ira Rosen <ira.ro...@linaro.org> wrote: > Hi, > > The test in PR 49318 fails because the vectorizer recognizes address > computation sequence as a widening-multiplication pattern, while such > sequence is not relevant to vectorization. The problem is that the > vectorizer doesn't check if a statement is going to be vectorized > before replacing it with a pattern. Moreover, the vectorizer first > detects the patterns and only after that looks for relevant > statements. Changing the order is not a good option, since statements > relevance is defined also by their belonging to a pattern. > > This patch solves the problem by removing pattern statements that were > created for statements that are not supposed to be vectorized. > > Bootstrapped with vectorization enabled on powerpc64-suse-linux and > tested on powerpc64-suse-linux and x86_64-suse-linux. > Committed.
Ick, yeah. I remember running into this ordering issue when doing the multi-vector size reorgs... In the end I think we should not generate the pattern stmt during pattern matching but only mark the relevant statements with a pattern kind. Say, for each pattern we have a "main" statement that has related stmts belonging to the pattern that define uses of the "main" statement - mark those to refer to that "main" statement. For that "main" statement simply record an enum value, like, widening_mult. Then only at vectorized statement generation time actually generate the vectorized form of the pattern statement. Thus, the non-vectorized pattern statements would never be generated. So, separate analysis and transform more properly. Of course I didn't get around to implement the above ... (we'd have a new vectorizable_pattern worker). Richard. > Ira > > ChangeLog: > > PR tree-optimization/49318 > * tree-vect-loop.c (vect_determine_vectorization_factor): > Remove irrelevant pattern statements. > > testsuite/ChangeLog: > > PR tree-optimization/49318 > * gcc.dg/vect/pr49318.c: New test. >
