diff -uprN a/gcc/testsuite/gcc.dg/vect/pr96698.c b/gcc/testsuite/gcc.dg/vect/pr96698.c --- a/gcc/testsuite/gcc.dg/vect/pr96698.c 1970-01-01 08:00:00.000000000 +0800 +++ b/gcc/testsuite/gcc.dg/vect/pr96698.c 2020-08-27 17:53:24.396000000 +0800 @@ -0,0 +1,19 @@ +/* { dg-do compile } */ + +void test(int a, int* i) +{ + for (; a < 5; ++a) + { + int b = 0; + int c = 0; + for (; b != -11; b--) + for (int d = 0; d ==0; d++) + { + *i += c & a; + c = b; + } + } +} + +/* We should be able to vectorize the inner cycle. */ +/* { dg-final { scan-tree-dump "OUTER LOOP VECTORIZED" "vect" { target vect_int } } } */ diff -uprN a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c --- a/gcc/tree-vect-loop.c 2020-08-27 09:25:58.000000000 +0800 +++ b/gcc/tree-vect-loop.c 2020-08-27 18:41:41.016000000 +0800 @@ -4325,7 +4325,8 @@ info_for_reduction (stmt_vec_info stmt_i { stmt_info = vect_orig_stmt (stmt_info); gcc_assert (STMT_VINFO_REDUC_DEF (stmt_info)); - if (!is_a (stmt_info->stmt)) + if (!is_a (stmt_info->stmt) + || !VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_info))) stmt_info = STMT_VINFO_REDUC_DEF (stmt_info); gphi *phi = as_a (stmt_info->stmt); if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_double_reduction_def) @@ -8622,6 +8623,43 @@ vect_transform_loop (loop_vec_info loop_ } } + /* Fill in backedge defs of reductions. */ + for (unsigned i = 0; i < loop_vinfo->reduc_latch_defs.length (); ++i) + { + stmt_vec_info stmt_info = loop_vinfo->reduc_latch_defs[i]; + stmt_vec_info orig_stmt_info = vect_orig_stmt (stmt_info); + stmt_vec_info phi_info + = STMT_VINFO_VEC_STMT (STMT_VINFO_REDUC_DEF (orig_stmt_info)); + stmt_vec_info vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); + gphi *phi + = dyn_cast (STMT_VINFO_REDUC_DEF (orig_stmt_info)->stmt); + edge e = loop_latch_edge (gimple_bb (phi_info->stmt)->loop_father); + do + { + add_phi_arg (as_a (phi_info->stmt), + gimple_get_lhs (vec_stmt->stmt), e, + gimple_phi_arg_location (phi, e->dest_idx)); + phi_info = STMT_VINFO_RELATED_STMT (phi_info); + vec_stmt = STMT_VINFO_RELATED_STMT (vec_stmt); + } + while (phi_info); + gcc_assert (!vec_stmt); + } + for (unsigned i = 0; i < loop_vinfo->reduc_latch_slp_defs.length (); ++i) + { + slp_tree slp_node = loop_vinfo->reduc_latch_slp_defs[i].first; + slp_tree phi_node = loop_vinfo->reduc_latch_slp_defs[i].second; + gphi *phi = as_a (SLP_TREE_SCALAR_STMTS (phi_node)[0]->stmt); + e = loop_latch_edge (gimple_bb (phi)->loop_father); + gcc_assert (SLP_TREE_VEC_STMTS (phi_node).length () + == SLP_TREE_VEC_STMTS (slp_node).length ()); + for (unsigned j = 0; j < SLP_TREE_VEC_STMTS (phi_node).length (); ++j) + add_phi_arg (as_a (SLP_TREE_VEC_STMTS (phi_node)[j]->stmt), + gimple_get_lhs + (SLP_TREE_VEC_STMTS (slp_node)[j]->stmt), + e, gimple_phi_arg_location (phi, e->dest_idx)); + } + /* Stub out scalar statements that must not survive vectorization. Doing this here helps with grouped statements, or statements that are involved in patterns. */ diff -uprN a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h --- a/gcc/tree-vectorizer.h 2020-08-27 09:25:57.000000000 +0800 +++ b/gcc/tree-vectorizer.h 2020-08-27 17:53:24.400000000 +0800 @@ -575,6 +575,11 @@ typedef struct _loop_vec_info : public v stmt in the chain. */ auto_vec reduction_chains; + /* The vectorized stmts defining the latch values of the reduction + they are involved with. */ + auto_vec reduc_latch_defs; + auto_vec > reduc_latch_slp_defs; + /* Cost vector for a single scalar iteration. */ auto_vec scalar_cost_vec; diff -uprN a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c --- a/gcc/tree-vect-stmts.c 2020-08-27 09:25:58.000000000 +0800 +++ b/gcc/tree-vect-stmts.c 2020-08-27 17:53:24.400000000 +0800 @@ -10213,8 +10213,8 @@ vect_transform_stmt (stmt_vec_info stmt_ if (STMT_VINFO_TYPE (stmt_info) == store_vec_info_type) return is_store; - /* If this stmt defines a value used on a backedge, update the - vectorized PHIs. */ + /* If this stmt defines a value used on a backedge, record it so + we can update the vectorized PHIs later. */ stmt_vec_info orig_stmt_info = vect_orig_stmt (stmt_info); stmt_vec_info reduc_info; if (STMT_VINFO_REDUC_DEF (orig_stmt_info) @@ -10234,32 +10234,13 @@ vect_transform_stmt (stmt_vec_info stmt_ && (PHI_ARG_DEF_FROM_EDGE (phi, e) == gimple_get_lhs (orig_stmt_info->stmt))) { - stmt_vec_info phi_info - = STMT_VINFO_VEC_STMT (STMT_VINFO_REDUC_DEF (orig_stmt_info)); - stmt_vec_info vec_stmt = STMT_VINFO_VEC_STMT (stmt_info); - do - { - add_phi_arg (as_a (phi_info->stmt), - gimple_get_lhs (vec_stmt->stmt), e, - gimple_phi_arg_location (phi, e->dest_idx)); - phi_info = STMT_VINFO_RELATED_STMT (phi_info); - vec_stmt = STMT_VINFO_RELATED_STMT (vec_stmt); - } - while (phi_info); - gcc_assert (!vec_stmt); + as_a (vinfo)->reduc_latch_defs.safe_push (stmt_info); } else if (slp_node && slp_node != slp_node_instance->reduc_phis) { - slp_tree phi_node = slp_node_instance->reduc_phis; - gphi *phi = as_a (SLP_TREE_SCALAR_STMTS (phi_node)[0]->stmt); - e = loop_latch_edge (gimple_bb (phi)->loop_father); - gcc_assert (SLP_TREE_VEC_STMTS (phi_node).length () - == SLP_TREE_VEC_STMTS (slp_node).length ()); - for (unsigned i = 0; i < SLP_TREE_VEC_STMTS (phi_node).length (); ++i) - add_phi_arg (as_a (SLP_TREE_VEC_STMTS (phi_node)[i]->stmt), - gimple_get_lhs (SLP_TREE_VEC_STMTS (slp_node)[i]->stmt), - e, gimple_phi_arg_location (phi, e->dest_idx)); + as_a (vinfo)->reduc_latch_slp_defs.safe_push + (std::make_pair (slp_node, slp_node_instance->reduc_phis)); } }