I've noticed regressions on target armeb-none-linux-gnueabihf --with-mode arm --with-cpu cortex-a9 --with-fpu neon-fp16: gcc.dg/vect/vect-alias-check.c -flto -ffat-lto-objects scan-tree-dump-times vect "vectorized 1 loops" 1 gcc.dg/vect/vect-alias-check.c scan-tree-dump-times vect "vectorized 1 loops" 1 gcc.dg/vect/vect-live-5.c -flto -ffat-lto-objects scan-tree-dump-times vect "vectorized 1 loops" 1 gcc.dg/vect/vect-live-5.c scan-tree-dump-times vect "vectorized 1 loops" 1 In my logs I can see: PASS: gcc.dg/vect/vect-live-5.c execution test gcc.dg/vect/vect-live-5.c: pattern found 0 times FAIL: gcc.dg/vect/vect-live-5.c scan-tree-dump-times vect "vectorized 1 loops" 1 PASS: gcc.dg/vect/vect-alias-check.c (test for excess errors) gcc.dg/vect/vect-alias-check.c: pattern found 0 times FAIL: gcc.dg/vect/vect-alias-check.c scan-tree-dump-times vect "vectorized 1 loops" 1 This appeared between r11-1908 and r11-1952.
Bisect identified commit g30fdaead5b7880c4e9f140618e26ad1c545642d5
I will have a look. How do I need to configure to reproduce with a cc1 cross?
I mentioned the configuration flags above: --target armeb-none-linux-gnueabihf --with-mode arm --with-cpu cortex-a9 --with-fpu neon-fp16
*** Bug 96129 has been marked as a duplicate of this bug. ***
By checking the case vect-alias-check.c on the sparc machine, although vectorizer does the versioning for the alignment requirement and is able to set the misaligned flag off for misaligned DRs, it bypasses the DRs *b_20(D) and MEM[(int *)b_20(D) + 4B], as they satisfy the condition check !vect_relevant_for_alignment_p. if (aligned_access_p (dr_info) || !vect_relevant_for_alignment_p (dr_info)) continue; More specific, it's due to these two DR's steps are unchanged in the loop (b[0] and b[1]). /* Scatter-gather and invariant accesses continue to address individual scalars, so vector-level alignment is irrelevant. */ if (STMT_VINFO_GATHER_SCATTER_P (stmt_info) || integer_zerop (DR_STEP (dr_info->dr))) return false; A simple fix seems to be: diff --git a/gcc/tree-vect-data-refs.c b/gcc/tree-vect-data-refs.c index 2b4421b5fb4..d5f52929f89 100644 --- a/gcc/tree-vect-data-refs.c +++ b/gcc/tree-vect-data-refs.c @@ -1186,7 +1186,7 @@ vect_update_misalignment_for_peel (dr_vec_info *dr_info, /* Return true if alignment is relevant for DR_INFO. */ -static bool +bool vect_relevant_for_alignment_p (dr_vec_info *dr_info) { stmt_vec_info stmt_info = dr_info->stmt; diff --git a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c index cec5c601268..8a04f55fdb1 100644 --- a/gcc/tree-vect-stmts.c +++ b/gcc/tree-vect-stmts.c @@ -2369,7 +2369,8 @@ get_load_store_type (vec_info *vinfo, stmt_vec_info stmt_info, return false; } - if (*alignment_support_scheme == dr_unaligned_unsupported) + if (*alignment_support_scheme == dr_unaligned_unsupported + && vect_relevant_for_alignment_p (STMT_VINFO_DR_INFO (stmt_info))) { if (dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h index 7c6de8397b3..067222ffe39 100644 --- a/gcc/tree-vectorizer.h +++ b/gcc/tree-vectorizer.h @@ -1935,6 +1935,7 @@ extern tree vect_get_new_ssa_name (tree, enum vect_var_kind, extern tree vect_create_addr_base_for_vector_ref (vec_info *, stmt_vec_info, gimple_seq *, tree, tree = NULL_TREE); +bool vect_relevant_for_alignment_p (dr_vec_info *dr_info); /* In tree-vect-loop.c. */ extern widest_int vect_iv_limit_for_partial_vectors (loop_vec_info loop_vinfo);
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>: https://gcc.gnu.org/g:446703ccc2ce09e9710964f781a21030bf3c2d54 commit r11-6717-g446703ccc2ce09e9710964f781a21030bf3c2d54 Author: Richard Biener <rguenther@suse.de> Date: Fri Jan 15 14:20:43 2021 +0100 tree-optimization/96376 - do not check alignment for invariant loads The testcases show that we fail to disregard alignment for invariant loads. The patch handles them like we handle gather and scatter. 2021-01-15 Richard Biener <rguenther@suse.de> PR tree-optimization/96376 * tree-vect-stmts.c (get_load_store_type): Disregard alignment for VMAT_INVARIANT.
Should be fixed now.