http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47001
Ira Rosen <irar at il dot ibm.com> changed: What |Removed |Added ---------------------------------------------------------------------------- Status|UNCONFIRMED |ASSIGNED Last reconfirmed| |2010.12.20 09:00:13 CC| |irar at il dot ibm.com AssignedTo|unassigned at gcc dot |irar at gcc dot gnu.org |gnu.org | Ever Confirmed|0 |1 --- Comment #1 from Ira Rosen <irar at il dot ibm.com> 2010-12-20 09:00:13 UTC --- The vectorizer tries to do SLP for reduction here, but there is only one scalar load for both reductions, which is not supported, but also is not checked. I am testing a patch that adds such check. Index: tree-vect-slp.c =================================================================== --- tree-vect-slp.c (revision 167365) +++ tree-vect-slp.c (working copy) @@ -1002,7 +1002,35 @@ vect_supported_load_permutation_p (slp_i if (!bad_permutation) { - /* This permutaion is valid for reduction. Since the order of the + /* Check that all the loads are different. */ + load_index = sbitmap_alloc (group_size); + sbitmap_zero (load_index); + for (k = 0; k < group_size; k++) + { + first_group_load_index = VEC_index (int, load_permutation, k); + if (TEST_BIT (load_index, first_group_load_index)) + { + bad_permutation = true; + break; + } + + SET_BIT (load_index, first_group_load_index); + } + + if (!bad_permutation) + for (k = 0; k < group_size; k++) + if (!TEST_BIT (load_index, k)) + { + bad_permutation = true; + break; + } + + sbitmap_free (load_index); + } + + if (!bad_permutation) + { + /* This permutation is valid for reduction. Since the order of the statements in the nodes is not important unless they are memory accesses, we can rearrange the statements in all the nodes according to the order of the loads. */ Ira