From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: by sourceware.org (Postfix, from userid 1971) id E22323858D37; Wed, 19 Jan 2022 14:15:05 +0000 (GMT) DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org E22323858D37 MIME-Version: 1.0 Content-Transfer-Encoding: 7bit Content-Type: text/plain; charset="utf-8" From: Andre Simoes Dias Vieira To: gcc-cvs@gcc.gnu.org Subject: [gcc r12-6740] vect: Fix epilogue mode skipping X-Act-Checkin: gcc X-Git-Author: Andre Vieira X-Git-Refname: refs/heads/master X-Git-Oldrev: ffc7f200adbdf47f14b3594d9b21855c19cf797a X-Git-Newrev: f4ca0a53be18dfc7162fd5dcc1e73c4203805e14 Message-Id: <20220119141505.E22323858D37@sourceware.org> Date: Wed, 19 Jan 2022 14:15:05 +0000 (GMT) X-BeenThere: gcc-cvs@gcc.gnu.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Gcc-cvs mailing list List-Unsubscribe: , List-Archive: List-Help: List-Subscribe: , X-List-Received-Date: Wed, 19 Jan 2022 14:15:06 -0000 https://gcc.gnu.org/g:f4ca0a53be18dfc7162fd5dcc1e73c4203805e14 commit r12-6740-gf4ca0a53be18dfc7162fd5dcc1e73c4203805e14 Author: Andre Vieira Date: Wed Jan 19 14:11:32 2022 +0000 vect: Fix epilogue mode skipping gcc/ChangeLog: PR tree-optimization/103997 * tree-vect-loop.cc (vect_analyze_loop): Fix mode skipping for epilogue vectorization. Diff: --- gcc/tree-vect-loop.cc | 24 +++++++++++++++++++----- 1 file changed, 19 insertions(+), 5 deletions(-) diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc index 0b2785a5ed6..4860bfd3344 100644 --- a/gcc/tree-vect-loop.cc +++ b/gcc/tree-vect-loop.cc @@ -3004,6 +3004,12 @@ vect_analyze_loop (class loop *loop, vec_info_shared *shared) unsigned int mode_i = 0; unsigned HOST_WIDE_INT simdlen = loop->simdlen; + /* Keep track of the VF for each mode. Initialize all to 0 which indicates + a mode has not been analyzed. */ + auto_vec cached_vf_per_mode; + for (unsigned i = 0; i < vector_modes.length (); ++i) + cached_vf_per_mode.safe_push (0); + /* First determine the main loop vectorization mode, either the first one that works, starting with auto-detecting the vector mode and then following the targets order of preference, or the one with the @@ -3011,6 +3017,10 @@ vect_analyze_loop (class loop *loop, vec_info_shared *shared) while (1) { bool fatal; + unsigned int last_mode_i = mode_i; + /* Set cached VF to -1 prior to analysis, which indicates a mode has + failed. */ + cached_vf_per_mode[last_mode_i] = -1; opt_loop_vec_info loop_vinfo = vect_analyze_loop_1 (loop, shared, &loop_form_info, NULL, vector_modes, mode_i, @@ -3020,6 +3030,12 @@ vect_analyze_loop (class loop *loop, vec_info_shared *shared) if (loop_vinfo) { + /* Analyzis has been successful so update the VF value. The + VF should always be a multiple of unroll_factor and we want to + capture the original VF here. */ + cached_vf_per_mode[last_mode_i] + = exact_div (LOOP_VINFO_VECT_FACTOR (loop_vinfo), + loop_vinfo->suggested_unroll_factor); /* Once we hit the desired simdlen for the first time, discard any previous attempts. */ if (simdlen @@ -3100,12 +3116,10 @@ vect_analyze_loop (class loop *loop, vec_info_shared *shared) { /* If the target does not support partial vectors we can shorten the number of modes to analyze for the epilogue as we know we can't pick a - mode that has at least as many NUNITS as the main loop's vectorization - factor, since that would imply the epilogue's vectorization factor - would be at least as high as the main loop's and we would be - vectorizing for more scalar iterations than there would be left. */ + mode that would lead to a VF at least as big as the + FIRST_VINFO_VF. */ if (!supports_partial_vectors - && maybe_ge (GET_MODE_NUNITS (vector_modes[mode_i]), first_vinfo_vf)) + && maybe_ge (cached_vf_per_mode[mode_i], first_vinfo_vf)) { mode_i++; if (mode_i == vector_modes.length ())