From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (qmail 78954 invoked by alias); 8 Jul 2015 17:37:51 -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 77672 invoked by uid 89); 8 Jul 2015 17:37:46 -0000 Authentication-Results: sourceware.org; auth=none X-Virus-Found: No X-HELO: mx1.redhat.com Received: from mx1.redhat.com (HELO mx1.redhat.com) (209.132.183.28) by sourceware.org (qpsmtpd/0.93/v0.84-503-g423c35a) with (AES256-GCM-SHA384 encrypted) ESMTPS; Wed, 08 Jul 2015 17:37:46 +0000 Received: from int-mx09.intmail.prod.int.phx2.redhat.com (int-mx09.intmail.prod.int.phx2.redhat.com [10.5.11.22]) by mx1.redhat.com (Postfix) with ESMTPS id 4CFB23824E3; Wed, 8 Jul 2015 17:37:45 +0000 (UTC) Received: from redhat.com (ovpn-204-65.brq.redhat.com [10.40.204.65]) by int-mx09.intmail.prod.int.phx2.redhat.com (8.14.4/8.14.4) with ESMTP id t68HbfA2004424 (version=TLSv1/SSLv3 cipher=AES256-SHA bits=256 verify=NO); Wed, 8 Jul 2015 13:37:44 -0400 Date: Wed, 08 Jul 2015 17:37:00 -0000 From: Marek Polacek To: Richard Biener Cc: GCC Patches , Jakub Jelinek Subject: Re: RFC: Add ADDR_EXPR lowering (PR tree-optimization/66718) Message-ID: <20150708173740.GC16027@redhat.com> References: <20150703132147.GD10139@redhat.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: User-Agent: Mutt/1.5.23 (2014-03-12) X-SW-Source: 2015-07/txt/msg00644.txt.bz2 On Fri, Jul 03, 2015 at 03:41:29PM +0200, Richard Biener wrote: > On Fri, 3 Jul 2015, Marek Polacek wrote: > > > This patch implements a new pass, called laddress, which deals with > > lowering ADDR_EXPR assignments. Such lowering ought to help the > > vectorizer, but it also could expose more CSE opportunities, maybe > > help reassoc, etc. It's only active when optimize != 0. > > > > So e.g. > > _1 = (sizetype) i_9; > > _7 = _1 * 4; > > _4 = &b + _7; > > instead of > > _4 = &b[i_9]; > > > > This triggered 14105 times during the regtest and 6392 times during > > the bootstrap. > > > > The fallout (at least on x86_64) is surprisingly small, i.e. none, just > > gcc.dg/vect/pr59984.c test (using -fopenmp-simd) ICEs, but that is due > > to a bug in the vectorizer. Jakub has a patch and knows the details. > > As the test shows, we're now able to vectorize ADDR_EXPR of non-invariants > > (that was the motivation of this pass). > > > > This doesn't introduce any kind of verification nor PROP_laddress. > > Don't know if we want that, but hopefully it can be done as a follow-up > > if we do. > > Yes. At the moment nothing requires lowered address form so this is > merely an optimization (and not a bug for some later pass to > re-introduce un-lowered non-invariant addresses). I can imagine > that for example IVOPTs could be simplified if we didn't have this > kind of addresses in the IL. > > > Do we want to move some optimizations into this new pass, e.g. > > from fwprop? > > I think we might want to re-try forwprop_into_addr_expr before lowering > the address. Well, but that's maybe just over-cautionous. > > > Thoughts? > > Please move the pass before crited, crited and pre are supposed to > go together. Done. > Otherwise looks ok to me. I renamed the file to gimple-laddress.c then and adjusted the timevar. Another change is that for x86_64 we don't need -mavx at all, so I dropped that. The test is now restricted to x86_64/i?86; on aarch64/ppc64 we aren't able to vectorize all the functions. Bootstrapped/regtested on x86_64-linux + ppc64-linux, ok for trunk? 2015-07-08 Marek Polacek PR tree-optimization/66718 * Makefile.in (OBJS): Add gimple-laddress.o. * passes.def: Schedule pass_laddress. * timevar.def (DEFTIMEVAR): Add TV_GIMPLE_LADDRESS. * tree-pass.h (make_pass_laddress): Declare. * gimple-laddress.c: New file. * gcc.dg/vect/vect-126.c: New test. diff --git gcc/Makefile.in gcc/Makefile.in index 89eda96..1817025 100644 --- gcc/Makefile.in +++ gcc/Makefile.in @@ -1255,6 +1255,7 @@ OBJS = \ gimple-expr.o \ gimple-iterator.o \ gimple-fold.o \ + gimple-laddress.o \ gimple-low.o \ gimple-match.o \ generic-match.o \ diff --git gcc/gimple-laddress.c gcc/gimple-laddress.c index e69de29..c8036b9 100644 --- gcc/gimple-laddress.c +++ gcc/gimple-laddress.c @@ -0,0 +1,137 @@ +/* Lower and optimize address expressions. + Copyright (C) 2015 Free Software Foundation, Inc. + Contributed by Marek Polacek + +This file is part of GCC. + +GCC is free software; you can redistribute it and/or modify it under +the terms of the GNU General Public License as published by the Free +Software Foundation; either version 3, or (at your option) any later +version. + +GCC is distributed in the hope that it will be useful, but WITHOUT ANY +WARRANTY; without even the implied warranty of MERCHANTABILITY or +FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received a copy of the GNU General Public License +along with GCC; see the file COPYING3. If not see +. */ + +#include "config.h" +#include "system.h" +#include "coretypes.h" +#include "alias.h" +#include "predict.h" +#include "tm.h" +#include "function.h" +#include "dominance.h" +#include "cfg.h" +#include "basic-block.h" +#include "tree-ssa-alias.h" +#include "symtab.h" +#include "tree.h" +#include "stringpool.h" +#include "tree-ssanames.h" +#include "fold-const.h" +#include "gimple-expr.h" +#include "gimple.h" +#include "gimplify.h" +#include "gimple-iterator.h" +#include "gimplify-me.h" +#include "tree-pass.h" + + +namespace { + +const pass_data pass_data_laddress = +{ + GIMPLE_PASS, /* type */ + "laddress", /* name */ + OPTGROUP_NONE, /* optinfo_flags */ + TV_GIMPLE_LADDRESS, /* tv_id */ + ( PROP_cfg | PROP_ssa ), /* properties_required */ + 0, /* properties_provided */ + 0, /* properties_destroyed */ + 0, /* todo_flags_start */ + 0, /* todo_flags_finish */ +}; + +class pass_laddress : public gimple_opt_pass +{ +public: + pass_laddress (gcc::context *ctxt) + : gimple_opt_pass (pass_data_laddress, ctxt) + {} + + /* opt_pass methods: */ + opt_pass * clone () { return new pass_laddress (m_ctxt); } + virtual bool gate (function *) { return optimize != 0; } + virtual unsigned int execute (function *); + +}; // class pass_laddress + +unsigned int +pass_laddress::execute (function *fun) +{ + basic_block bb; + + FOR_EACH_BB_FN (bb, fun) + { + for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi);) + { + gimple stmt = gsi_stmt (gsi); + if (!is_gimple_assign (stmt) + || gimple_assign_rhs_code (stmt) != ADDR_EXPR + || is_gimple_invariant_address (gimple_assign_rhs1 (stmt))) + { + gsi_next (&gsi); + continue; + } + + /* Lower ADDR_EXPR assignments: + _4 = &b[i_9]; + into + _1 = (sizetype) i_9; + _7 = _1 * 4; + _4 = &b + _7; + This ought to aid the vectorizer and expose CSE opportunities. + */ + + tree expr = gimple_assign_rhs1 (stmt); + HOST_WIDE_INT bitsize, bitpos; + tree base, offset; + machine_mode mode; + int volatilep = 0, unsignedp = 0; + base = get_inner_reference (TREE_OPERAND (expr, 0), &bitsize, + &bitpos, &offset, &mode, &unsignedp, + &volatilep, false); + gcc_assert (base != NULL_TREE && (bitpos % BITS_PER_UNIT) == 0); + if (offset != NULL_TREE) + { + if (bitpos != 0) + offset = size_binop (PLUS_EXPR, offset, + size_int (bitpos / BITS_PER_UNIT)); + offset = force_gimple_operand_gsi (&gsi, offset, true, NULL, + true, GSI_SAME_STMT); + base = build_fold_addr_expr (base); + base = force_gimple_operand_gsi (&gsi, base, true, NULL, + true, GSI_SAME_STMT); + gimple g = gimple_build_assign (gimple_assign_lhs (stmt), + POINTER_PLUS_EXPR, base, offset); + gsi_replace (&gsi, g, false); + } + gsi_next (&gsi); + } + } + + return 0; +} + +} // anon namespace + +gimple_opt_pass * +make_pass_laddress (gcc::context *ctxt) +{ + return new pass_laddress (ctxt); +} diff --git gcc/passes.def gcc/passes.def index 0d8356b..5cd07ae 100644 --- gcc/passes.def +++ gcc/passes.def @@ -213,6 +213,7 @@ along with GCC; see the file COPYING3. If not see form if possible. */ NEXT_PASS (pass_cse_sincos); NEXT_PASS (pass_optimize_bswap); + NEXT_PASS (pass_laddress); NEXT_PASS (pass_split_crit_edges); NEXT_PASS (pass_pre); NEXT_PASS (pass_sink_code); diff --git gcc/testsuite/gcc.dg/vect/vect-126.c gcc/testsuite/gcc.dg/vect/vect-126.c index e69de29..e269ad3 100644 --- gcc/testsuite/gcc.dg/vect/vect-126.c +++ gcc/testsuite/gcc.dg/vect/vect-126.c @@ -0,0 +1,63 @@ +/* PR tree-optimization/66718 */ +/* { dg-do compile } */ + +int *a[1024], b[1024]; +struct S { int u, v, w, x; }; +struct S c[1024]; +int d[1024][10]; + +void +f0 (void) +{ + for (int i = 0; i < 1024; i++) + a[i] = &b[0]; +} + +void +f1 (void) +{ + for (int i = 0; i < 1024; i++) + { + int *p = &b[0]; + a[i] = p + i; + } +} + +void +f2 (int *p) +{ + for (int i = 0; i < 1024; i++) + a[i] = &p[i]; +} + +void +f3 (void) +{ + for (int i = 0; i < 1024; i++) + a[i] = &b[i]; +} + +void +f4 (void) +{ + int *p = &c[0].v; + for (int i = 0; i < 1024; i++) + a[i] = &p[4 * i]; +} + +void +f5 (void) +{ + for (int i = 0; i < 1024; i++) + a[i] = &c[i].v; +} + +void +f6 (void) +{ + for (int i = 0; i < 1024; i++) + for (unsigned int j = 0; j < 10; j++) + a[i] = &d[i][j]; +} + +/* { dg-final { scan-tree-dump-times "vectorized 1 loops in function" 7 "vect" { target { i?86-*-* x86_64-*-* } } } } */ diff --git gcc/timevar.def gcc/timevar.def index efac4b7..ff22909 100644 --- gcc/timevar.def +++ gcc/timevar.def @@ -275,6 +275,7 @@ DEFTIMEVAR (TV_GIMPLE_SLSR , "straight-line strength reduction") DEFTIMEVAR (TV_VTABLE_VERIFICATION , "vtable verification") DEFTIMEVAR (TV_TREE_UBSAN , "tree ubsan") DEFTIMEVAR (TV_INITIALIZE_RTL , "initialize rtl") +DEFTIMEVAR (TV_GIMPLE_LADDRESS , "address lowering") /* Everything else in rest_of_compilation not included above. */ DEFTIMEVAR (TV_EARLY_LOCAL , "early local passes") diff --git gcc/tree-pass.h gcc/tree-pass.h index 2808dad..c47b22e 100644 --- gcc/tree-pass.h +++ gcc/tree-pass.h @@ -393,6 +393,7 @@ extern gimple_opt_pass *make_pass_cd_dce (gcc::context *ctxt); extern gimple_opt_pass *make_pass_call_cdce (gcc::context *ctxt); extern gimple_opt_pass *make_pass_merge_phi (gcc::context *ctxt); extern gimple_opt_pass *make_pass_split_crit_edges (gcc::context *ctxt); +extern gimple_opt_pass *make_pass_laddress (gcc::context *ctxt); extern gimple_opt_pass *make_pass_pre (gcc::context *ctxt); extern unsigned int tail_merge_optimize (unsigned int); extern gimple_opt_pass *make_pass_profile (gcc::context *ctxt); Marek