From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (qmail 1392 invoked by alias); 25 Jun 2019 12:59:07 -0000 Mailing-List: contact gcc-patches-help@gcc.gnu.org; run by ezmlm Precedence: bulk List-Id: List-Archive: List-Post: List-Help: Sender: gcc-patches-owner@gcc.gnu.org Received: (qmail 1384 invoked by uid 89); 25 Jun 2019 12:59:07 -0000 Authentication-Results: sourceware.org; auth=none X-Spam-SWARE-Status: No, score=-10.2 required=5.0 tests=AWL,BAYES_00,GIT_PATCH_2,GIT_PATCH_3,KAM_ASCII_DIVIDERS,KAM_NUMSUBJECT,SPF_PASS autolearn=ham version=3.3.1 spammy=reusing X-HELO: mx1.suse.de Received: from mx2.suse.de (HELO mx1.suse.de) (195.135.220.15) by sourceware.org (qpsmtpd/0.93/v0.84-503-g423c35a) with ESMTP; Tue, 25 Jun 2019 12:59:06 +0000 Received: from relay2.suse.de (unknown [195.135.220.254]) by mx1.suse.de (Postfix) with ESMTP id A3EEBAFF9; Tue, 25 Jun 2019 12:59:03 +0000 (UTC) Date: Tue, 25 Jun 2019 12:59:00 -0000 From: Richard Biener To: gcc-patches@gcc.gnu.org cc: Jan Hubicka Subject: [PATCH] Try fix PR90911 Message-ID: User-Agent: Alpine 2.20 (LSU 67 2015-01-07) MIME-Version: 1.0 Content-Type: text/plain; charset=US-ASCII X-SW-Source: 2019-06/txt/msg01550.txt.bz2 PR90911 reports a slowdown of 456.hmmer with the recent introduction of vectorizer versioning of outer loops, more specifically the case of re-using if-conversion created versions. The patch below fixes things up to adjust the edge probability and scale the loop bodies in two steps, delaying scalar_loop scaling until all peeling is done. This restores profile-mismatches to the same state as it was on the GCC 9 branch and seems to fix the observed slowdown of 456.hmmer. Boostrap & regtest running on x86_64-unknown-linux-gnu. Honza, does this look OK? Thanks, Richard. 2019-06-25 Richard Biener * tree-vectorizer.h (_loop_vec_info::scalar_loop_scaling): New field. (LOOP_VINFO_SCALAR_LOOP_SCALING): new. * tree-vect-loop.c (_loop_vec_info::_loop_vec_info): Initialize scalar_loop_scaling. (vect_transform_loop): Scale scalar loop profile if needed. * tree-vect-loop-manip.c (vect_loop_versioning): When re-using the loop copy from if-conversion adjust edge probabilities and scale the vectorized loop body profile, queue the scalar profile for updating after peeling. Index: gcc/tree-vectorizer.h =================================================================== --- gcc/tree-vectorizer.h (revision 272636) +++ gcc/tree-vectorizer.h (working copy) @@ -548,6 +548,9 @@ typedef struct _loop_vec_info : public v /* Mark loops having masked stores. */ bool has_mask_store; + /* Queued scaling factor for the scalar loop. */ + profile_probability scalar_loop_scaling; + /* If if-conversion versioned this loop before conversion, this is the loop version without if-conversion. */ struct loop *scalar_loop; @@ -603,6 +606,7 @@ typedef struct _loop_vec_info : public v #define LOOP_VINFO_PEELING_FOR_NITER(L) (L)->peeling_for_niter #define LOOP_VINFO_NO_DATA_DEPENDENCIES(L) (L)->no_data_dependencies #define LOOP_VINFO_SCALAR_LOOP(L) (L)->scalar_loop +#define LOOP_VINFO_SCALAR_LOOP_SCALING(L) (L)->scalar_loop_scaling #define LOOP_VINFO_HAS_MASK_STORE(L) (L)->has_mask_store #define LOOP_VINFO_SCALAR_ITERATION_COST(L) (L)->scalar_cost_vec #define LOOP_VINFO_SINGLE_SCALAR_ITERATION_COST(L) (L)->single_scalar_iteration_cost Index: gcc/tree-vect-loop.c =================================================================== --- gcc/tree-vect-loop.c (revision 272636) +++ gcc/tree-vect-loop.c (working copy) @@ -835,6 +835,7 @@ _loop_vec_info::_loop_vec_info (struct l operands_swapped (false), no_data_dependencies (false), has_mask_store (false), + scalar_loop_scaling (profile_probability::uninitialized ()), scalar_loop (NULL), orig_loop_info (NULL) { @@ -8562,6 +8563,10 @@ vect_transform_loop (loop_vec_info loop_ epilogue = vect_do_peeling (loop_vinfo, niters, nitersm1, &niters_vector, &step_vector, &niters_vector_mult_vf, th, check_profitability, niters_no_overflow); + if (LOOP_VINFO_SCALAR_LOOP (loop_vinfo) + && LOOP_VINFO_SCALAR_LOOP_SCALING (loop_vinfo).initialized_p ()) + scale_loop_frequencies (LOOP_VINFO_SCALAR_LOOP (loop_vinfo), + LOOP_VINFO_SCALAR_LOOP_SCALING (loop_vinfo)); if (niters_vector == NULL_TREE) { Index: gcc/tree-vect-loop-manip.c =================================================================== --- gcc/tree-vect-loop-manip.c (revision 272636) +++ gcc/tree-vect-loop-manip.c (working copy) @@ -3114,8 +3114,17 @@ vect_loop_versioning (loop_vec_info loop GSI_SAME_STMT); } - /* ??? if-conversion uses profile_probability::always () but - prob below is profile_probability::likely (). */ + /* if-conversion uses profile_probability::always () for both paths, + reset the paths probabilities appropriately. */ + edge te, fe; + extract_true_false_edges_from_block (condition_bb, &te, &fe); + te->probability = prob; + fe->probability = prob.invert (); + /* We can scale loops counts immediately but have to postpone + scaling the scalar loop because we re-use it during peeling. */ + scale_loop_frequencies (loop_to_version, prob); + LOOP_VINFO_SCALAR_LOOP_SCALING (loop_vinfo) = prob.invert (); + nloop = scalar_loop; if (dump_enabled_p ()) dump_printf_loc (MSG_NOTE, vect_location,