The following delays sinking of loads within the same innermost loop when it was unconditional before. That's a not uncommon issue preventing vectorization when masked loads are not available.
Bootstrapped and tested on x86_64-unknown-linux-gnu. I have a followup patch improving sinking that without this would cause more of the problematic sinking - now that we have a second sink pass after loop opts this looks like a reasonable approach? OK? Thanks, Richard. PR tree-optimization/92335 * tree-ssa-sink.cc (select_best_block): Before loop optimizations avoid sinking unconditional loads/stores in innermost loops to conditional executed places. * gcc.dg/tree-ssa/ssa-sink-10.c: Disable vectorizing. * gcc.dg/tree-ssa/predcom-9.c: Clone from ssa-sink-10.c, expect predictive commoning to happen instead of sinking. * gcc.dg/vect/pr65947-3.c: Adjust. --- gcc/testsuite/gcc.dg/tree-ssa/predcom-9.c | 20 ++++++++++++++++++++ gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-10.c | 2 +- gcc/testsuite/gcc.dg/vect/pr65947-3.c | 6 +----- gcc/tree-ssa-sink.cc | 12 ++++++++++++ 4 files changed, 34 insertions(+), 6 deletions(-) create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/predcom-9.c diff --git a/gcc/testsuite/gcc.dg/tree-ssa/predcom-9.c b/gcc/testsuite/gcc.dg/tree-ssa/predcom-9.c new file mode 100644 index 00000000000..b0fb0e2d4c5 --- /dev/null +++ b/gcc/testsuite/gcc.dg/tree-ssa/predcom-9.c @@ -0,0 +1,20 @@ +/* { dg-do compile } */ +/* { dg-options "-O2 -fdump-tree-sink-details -fdump-tree-pcom-details" } */ + +int x[1024], y[1024], z[1024], w[1024]; +void foo (void) +{ + int i; + for (i = 1; i < 1024; ++i) + { + int a = x[i]; + int b = y[i]; + int c = x[i-1]; + int d = y[i-1]; + if (w[i]) + z[i] = (a + b) + (c + d); + } +} + +/* { dg-final { scan-tree-dump-not "Sinking # VUSE" "sink1" } } */ +/* { dg-final { scan-tree-dump "Executing predictive commoning without unrolling" "pcom" } } */ diff --git a/gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-10.c b/gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-10.c index 535cb3208f5..a35014be038 100644 --- a/gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-10.c +++ b/gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-10.c @@ -1,5 +1,5 @@ /* { dg-do compile } */ -/* { dg-options "-O2 -fdump-tree-sink-details -fno-tree-pre" } */ +/* { dg-options "-O2 -fdump-tree-sink-details -fno-tree-vectorize -fno-tree-pre" } */ int x[1024], y[1024], z[1024], w[1024]; void foo (void) diff --git a/gcc/testsuite/gcc.dg/vect/pr65947-3.c b/gcc/testsuite/gcc.dg/vect/pr65947-3.c index f1bfad65c22..6b4077e1a62 100644 --- a/gcc/testsuite/gcc.dg/vect/pr65947-3.c +++ b/gcc/testsuite/gcc.dg/vect/pr65947-3.c @@ -51,10 +51,6 @@ main (void) return 0; } -/* Since the fix for PR97307 which sinks the load of a[i], preventing - if-conversion to happen, targets that cannot do masked loads only - vectorize the inline copy. */ -/* { dg-final { scan-tree-dump-times "LOOP VECTORIZED" 2 "vect" { target vect_masked_load } } } */ -/* { dg-final { scan-tree-dump-times "LOOP VECTORIZED" 1 "vect" { target { ! vect_masked_load } } } } */ +/* { dg-final { scan-tree-dump-times "LOOP VECTORIZED" 2 "vect" } } */ /* { dg-final { scan-tree-dump-times "optimizing condition reduction with FOLD_EXTRACT_LAST" 2 "vect" { target vect_fold_extract_last } } } */ /* { dg-final { scan-tree-dump-not "condition expression based on integer induction." "vect" } } */ diff --git a/gcc/tree-ssa-sink.cc b/gcc/tree-ssa-sink.cc index cf0a32a954b..dcbe05b3b03 100644 --- a/gcc/tree-ssa-sink.cc +++ b/gcc/tree-ssa-sink.cc @@ -220,6 +220,18 @@ select_best_block (basic_block early_bb, if (bb_loop_depth (best_bb) < bb_loop_depth (early_bb)) return best_bb; + /* Avoid turning an unconditional load/store into a conditional one when we + still might want to perform vectorization. */ + if (best_bb->loop_father == early_bb->loop_father + && loop_outer (best_bb->loop_father) + && !best_bb->loop_father->inner + && gimple_vuse (stmt) + && flag_tree_loop_vectorize + && !(cfun->curr_properties & PROP_loop_opts_done) + && dominated_by_p (CDI_DOMINATORS, best_bb->loop_father->latch, early_bb) + && !dominated_by_p (CDI_DOMINATORS, best_bb->loop_father->latch, best_bb)) + return early_bb; + /* Get the sinking threshold. If the statement to be moved has memory operands, then increase the threshold by 7% as those are even more profitable to avoid, clamping at 100%. */ -- 2.35.3