The following patch deals with the testsuite fallout of the patch forcing LOOP_VECTORIZED () versioning in GIMPLE if-conversion. We no longer see the if-converted body when doing BB vectorization.
While the real fix would be to teach BB vectorization about conditions (and thus if-conversion) itself a quick workaround to restore previous behavior is to simply run BB vectorization for if-converted loop bodies from inside the loop vectorizer itself. That fixes the testcase and restores behavior for all if-converted (but not also unrolled) loops. Bootstrap & regtest in progress on x86_64-unknown-linux-gnu. Richard. 2016-11-21 Richard Biener <rguent...@suse.de> PR tree-optimization/78396 * tree-vectorizer.c (vectorize_loops): If an innermost loop didn't vectorize try vectorizing an if-converted body using BB vectorization. * gcc.dg/vect/bb-slp-cond-1.c: Adjust. diff --git a/gcc/testsuite/gcc.dg/vect/bb-slp-cond-1.c b/gcc/testsuite/gcc.dg/vect/bb-slp-cond-1.c index 35811bd..ddad853 100644 --- a/gcc/testsuite/gcc.dg/vect/bb-slp-cond-1.c +++ b/gcc/testsuite/gcc.dg/vect/bb-slp-cond-1.c @@ -1,4 +1,5 @@ /* { dg-require-effective-target vect_condition } */ +/* { dg-additional-options "-fdump-tree-vect-details" } */ #include "tree-vect.h" @@ -41,5 +42,10 @@ int main () return 0; } -/* { dg-final { scan-tree-dump-times "basic block vectorized" 1 "slp1" { target vect_element_align } } } */ +/* Basic blocks of if-converted loops are vectorized from within the loop + vectorizer pass. In this case it is really a deficiency in loop + vectorization data dependence analysis that causes us to require + basic block vectorization in the first place. */ + +/* { dg-final { scan-tree-dump-times "basic block vectorized" 1 "vect" { target vect_element_align } } } */ diff --git a/gcc/tree-vectorizer.c b/gcc/tree-vectorizer.c index 35d7a3e..b390664 100644 --- a/gcc/tree-vectorizer.c +++ b/gcc/tree-vectorizer.c @@ -540,6 +540,7 @@ vectorize_loops (void) || loop->force_vectorize) { loop_vec_info loop_vinfo, orig_loop_vinfo = NULL; + gimple *loop_vectorized_call = vect_loop_vectorized_call (loop); vectorize_epilogue: vect_location = find_loop_location (loop); if (LOCATION_LOCUS (vect_location) != UNKNOWN_LOCATION @@ -558,6 +559,33 @@ vectorize_epilogue: if (loop_constraint_set_p (loop, LOOP_C_FINITE)) vect_free_loop_info_assumptions (loop); + /* If we applied if-conversion then try to vectorize the + BB of innermost loops. + ??? Ideally BB vectorization would learn to vectorize + control flow by applying if-conversion on-the-fly, the + following retains the if-converted loop body even when + only non-if-converted parts took part in BB vectorization. */ + if (flag_tree_slp_vectorize != 0 + && loop_vectorized_call + && ! loop->inner) + { + basic_block bb = loop->header; + for (gimple_stmt_iterator gsi = gsi_start_bb (bb); + !gsi_end_p (gsi); gsi_next (&gsi)) + { + gimple *stmt = gsi_stmt (gsi); + gimple_set_uid (stmt, -1); + gimple_set_visited (stmt, false); + } + if (vect_slp_bb (bb)) + { + dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location, + "basic block vectorized\n"); + fold_loop_vectorized_call (loop_vectorized_call, + boolean_true_node); + ret |= TODO_cleanup_cfg; + } + } continue; } @@ -575,7 +603,6 @@ vectorize_epilogue: break; } - gimple *loop_vectorized_call = vect_loop_vectorized_call (loop); if (loop_vectorized_call) set_uid_loop_bbs (loop_vinfo, loop_vectorized_call); if (LOCATION_LOCUS (vect_location) != UNKNOWN_LOCATION