From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from us-smtp-delivery-124.mimecast.com (us-smtp-delivery-124.mimecast.com [170.10.133.124]) by sourceware.org (Postfix) with ESMTPS id BA6163857000 for ; Tue, 9 Aug 2022 18:52:50 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.1 sourceware.org BA6163857000 Received: from mail-io1-f70.google.com (mail-io1-f70.google.com [209.85.166.70]) by relay.mimecast.com with ESMTP with STARTTLS (version=TLSv1.2, cipher=TLS_ECDHE_RSA_WITH_AES_256_GCM_SHA384) id us-mta-638-ofE0xxx_MdyfDP-1PmcsnA-1; Tue, 09 Aug 2022 14:52:48 -0400 X-MC-Unique: ofE0xxx_MdyfDP-1PmcsnA-1 Received: by mail-io1-f70.google.com with SMTP id u5-20020a6b4905000000b00681e48dbd92so6856387iob.21 for ; Tue, 09 Aug 2022 11:52:48 -0700 (PDT) X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20210112; h=content-transfer-encoding:in-reply-to:from:references:cc:to :content-language:subject:user-agent:mime-version:date:message-id :x-gm-message-state:from:to:cc; bh=R9Uo5gnvYEzlZseaCdPOKT7nchMTVrHBf4FkOjQDRQI=; b=xRlLy4w/bxkRUGg2heY3LSi3yG20DjVwGXgO2yDV2fOyAa7viiqyrWD3XGN1UBLuV0 +s9ZulCoOAvsr8ZMXQgAsy/UoYQ0mQgEppEJhVObzHuwdUK7oQAoADFk+yGfR40lO7Bm 6ierpqW2VSzbN8Z1hu4gYHMYemJqBtjjHKjVrpu8W1rNDXweMzlzfDuXyUouNBdIt7T+ bhdJMJhznMDfwXGxeKFN+qOvE2I8vBxnU9cBsoBqhg4iylshAkwwJpyhyBNBSgeNbbod 2EXdNzNHGVxBdZWhXWr6rgoPNxVJWws5ndXPdTLWKCq1uB8hJ5kQ1NlSRln8JXmWd4nG kbXQ== X-Gm-Message-State: ACgBeo30vQyDs3CH74HtOEn+KVDd30Ne6li/aM2b56Y9nAYzqDdHvEX6 uU8e75TjRkLKiwR2RbjxySGQm9A/rtpocLh+ZoE2CqoGeaPriTr1LO4/L+VxvCaVJ9co5YyWFTW FfHBFZb/y7fMzsf65bA== X-Received: by 2002:a05:6602:27c6:b0:657:7e7a:11f3 with SMTP id l6-20020a05660227c600b006577e7a11f3mr10176170ios.40.1660071167582; Tue, 09 Aug 2022 11:52:47 -0700 (PDT) X-Google-Smtp-Source: AA6agR4fetQm6mufy36BocPZsW153tcR2g8nRY0pOFf+rDHvb9whWYxMD+OmxPjkqvBaPJUgMgYIeQ== X-Received: by 2002:a05:6602:27c6:b0:657:7e7a:11f3 with SMTP id l6-20020a05660227c600b006577e7a11f3mr10176162ios.40.1660071167164; Tue, 09 Aug 2022 11:52:47 -0700 (PDT) Received: from [192.168.0.135] ([192.24.49.145]) by smtp.gmail.com with ESMTPSA id p63-20020a022942000000b00342b327d709sm6257206jap.71.2022.08.09.11.52.45 (version=TLS1_3 cipher=TLS_AES_128_GCM_SHA256 bits=128/128); Tue, 09 Aug 2022 11:52:46 -0700 (PDT) Message-ID: <55ed3acb-ac0c-d7b7-0e9b-58591cf992a2@redhat.com> Date: Tue, 9 Aug 2022 14:52:46 -0400 MIME-Version: 1.0 User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101 Thunderbird/91.11.0 Subject: Re: [PATCH] tree-optimization/106514 - revisit m_import compute in backward threading To: Richard Biener , gcc-patches@gcc.gnu.org References: <37255.122080909012202281@us-mta-359.us.mimecast.lan> From: Andrew MacLeod In-Reply-To: <37255.122080909012202281@us-mta-359.us.mimecast.lan> X-Mimecast-Spam-Score: 0 X-Mimecast-Originator: redhat.com Content-Language: en-US Content-Type: text/plain; charset=UTF-8; format=flowed Content-Transfer-Encoding: 8bit X-Spam-Status: No, score=-12.7 required=5.0 tests=BAYES_00, DKIMWL_WL_HIGH, DKIM_SIGNED, DKIM_VALID, DKIM_VALID_AU, DKIM_VALID_EF, GIT_PATCH_0, NICE_REPLY_A, RCVD_IN_DNSWL_NONE, SPF_HELO_NONE, SPF_NONE, TXREP, T_SCC_BODY_TEXT_LINE autolearn=ham autolearn_force=no version=3.4.6 X-Spam-Checker-Version: SpamAssassin 3.4.6 (2021-04-09) on server2.sourceware.org X-BeenThere: gcc-patches@gcc.gnu.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Gcc-patches mailing list List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Tue, 09 Aug 2022 18:52:52 -0000 On 8/9/22 09:01, Richard Biener wrote: > This revisits how we compute imports later used for the ranger path > query during backwards threading. The compute_imports function > of the path solver ends up pulling the SSA def chain of regular > stmts without limit and since it starts with just the gori imports > of the path exit it misses some interesting names to translate > during path discovery. In fact with a still empty path this > compute_imports function looks like not the correct tool. I don't really know how this works in practice.  Aldys off this week, so he can comment when he returns. The original premise was along the line of recognizing that only changes to a GORI import name to a block can affect the branch at the end of the block.  ie, if the path doesn't change any import to block A, then the branch at the end of block A will not change either.    Likewise, if it does change an import, then we look at whether the branch can be threaded.    Beyond that basic premise, I dont know what all it does. I presume the unbounded def chain is for local defs within a block that in turn feeds the import to another block.   Im not sure why we need to do much with those..  again, its only the import to the defchain that can affect the outcome t the end of the chain.. and if it changes, then you need to recalculate the entire chain.. but that would be part of the normal path walk.  I suspect ther eis also some pruning that can be done there, as GORi reflects "can affect the range" not "will affect the range". Perhaps whats going on is that all those local elements are being added up front to the list of interesting names?  That would certainly blow up the bitmaps and loops and such. Im sure Aldy will pitch in when he returns from vacation. > > The following instead implements what it does during the path discovery > and since we add the exit block we seed the initial imports and > interesting names from just the exit conditional. When we then > process interesting names (aka imports we did not yet see the definition > of) we prune local defs but add their uses in a similar way as > compute_imports would have done. > > The patch also properly unwinds m_imports during the path discovery > backtracking and from a debugging session I have verified the two > sets evolve as expected now while previously behaving slightly erratic. > > Fortunately the m_imports set now also is shrunken significantly for > the PR69592 testcase (aka PR106514) so that there's overall speedup > when increasing --param max-jump-thread-duplication-stmts as > 15 -> 30 -> 60 -> 120 from 1s -> 2s -> 13s -> 27s to with the patch > 1s -> 2s -> 4s -> 8s. > > This runs into a latent issue in X which doesn't seem to expect > any PHI nodes with a constant argument on an edge inside the path. > But we now have those as interesting, for example for the ICEing > g++.dg/torture/pr100925.C which just has sth like > > if (i) > x = 1; > else > x = 5; > if (x == 1) > ... > > where we now have the path from if (i) to if (x) and the PHI for x > in the set of imports to consider for resolving x == 1 which IMHO > looks exactly like what we want. The path_range_query::ssa_range_in_phi > papers over the issue and drops the range to varying instead of > crashing. I didn't want to mess with this any further in this patch > (but I couldn't resist replacing the loop over PHI args with > PHI_ARG_DEF_FROM_EDGE, so mind the re-indenting). > > Bootstrapped and tested on x86_64-unknown-linux-gnu w/o the > path_range_query::ssa_range_in_phi fix, now re-running with. > > OK? > > Thanks, > Richard. > > PR tree-optimization/106514 > * tree-ssa-threadbackward.cc (back_threader::find_paths_to_names): > Compute and unwind both m_imports and interesting on the fly during > path discovery. > (back_threader::find_paths): Compute the original m_imports > from just the SSA uses of the exit conditional. Drop > handling single_succ_to_potentially_threadable_block. > * gimple-range-path.cc (path_range_query::ssa_range_in_phi): Handle > constant PHI arguments without crashing. Use PHI_ARG_DEF_FROM_EDGE. > --- > gcc/gimple-range-path.cc | 52 ++++++++--------- > gcc/tree-ssa-threadbackward.cc | 104 ++++++++++++++++++++++++++------- > 2 files changed, 106 insertions(+), 50 deletions(-) > > diff --git a/gcc/gimple-range-path.cc b/gcc/gimple-range-path.cc > index 43e7526b6fc..b4376011ea8 100644 > --- a/gcc/gimple-range-path.cc > +++ b/gcc/gimple-range-path.cc > @@ -276,8 +276,6 @@ void > path_range_query::ssa_range_in_phi (vrange &r, gphi *phi) > { > tree name = gimple_phi_result (phi); > - basic_block bb = gimple_bb (phi); > - unsigned nargs = gimple_phi_num_args (phi); > > if (at_entry ()) > { > @@ -287,6 +285,7 @@ path_range_query::ssa_range_in_phi (vrange &r, gphi *phi) > // Try to fold the phi exclusively with global or cached values. > // This will get things like PHI <5(99), 6(88)>. We do this by > // calling range_of_expr with no context. > + unsigned nargs = gimple_phi_num_args (phi); > Value_Range arg_range (TREE_TYPE (name)); > r.set_undefined (); > for (size_t i = 0; i < nargs; ++i) > @@ -303,36 +302,31 @@ path_range_query::ssa_range_in_phi (vrange &r, gphi *phi) > return; > } > > + basic_block bb = gimple_bb (phi); > basic_block prev = prev_bb (); > edge e_in = find_edge (prev, bb); > - > - for (size_t i = 0; i < nargs; ++i) > - if (e_in == gimple_phi_arg_edge (phi, i)) > - { > - tree arg = gimple_phi_arg_def (phi, i); > - // Avoid using the cache for ARGs defined in this block, as > - // that could create an ordering problem. > - if (ssa_defined_in_bb (arg, bb) || !get_cache (r, arg)) > - { > - if (m_resolve) > - { > - Value_Range tmp (TREE_TYPE (name)); > - // Using both the range on entry to the path, and the > - // range on this edge yields significantly better > - // results. > - if (defined_outside_path (arg)) > - range_on_path_entry (r, arg); > - else > - r.set_varying (TREE_TYPE (name)); > - m_ranger->range_on_edge (tmp, e_in, arg); > - r.intersect (tmp); > - return; > - } > + tree arg = PHI_ARG_DEF_FROM_EDGE (phi, e_in); > + // Avoid using the cache for ARGs defined in this block, as > + // that could create an ordering problem. > + if (ssa_defined_in_bb (arg, bb) || !get_cache (r, arg)) > + { > + if (m_resolve) > + { > + Value_Range tmp (TREE_TYPE (name)); > + // Using both the range on entry to the path, and the > + // range on this edge yields significantly better > + // results. > + if (TREE_CODE (arg) == SSA_NAME > + && defined_outside_path (arg)) > + range_on_path_entry (r, arg); > + else > r.set_varying (TREE_TYPE (name)); > - } > - return; > - } > - gcc_unreachable (); > + m_ranger->range_on_edge (tmp, e_in, arg); > + r.intersect (tmp); > + return; > + } > + r.set_varying (TREE_TYPE (name)); > + } > } > > // If NAME is defined in BB, set R to the range of NAME, and return > diff --git a/gcc/tree-ssa-threadbackward.cc b/gcc/tree-ssa-threadbackward.cc > index 4cd4d21c712..a544e97b2af 100644 > --- a/gcc/tree-ssa-threadbackward.cc > +++ b/gcc/tree-ssa-threadbackward.cc > @@ -482,32 +482,76 @@ back_threader::find_paths_to_names (basic_block bb, bitmap interesting, > { > // For further greedy searching we want to remove interesting > // names defined in BB but add ones on the PHI edges for the > - // respective edges. We do this by starting with all names > + // respective edges and adding imports from those stmts. > + // We do this by starting with all names > // not defined in BB as interesting, collecting a list of > // interesting PHIs in BB on the fly. Then we iterate over > // predecessor edges, adding interesting PHI edge defs to > // the set of interesting names to consider when processing it. > auto_bitmap new_interesting; > + auto_vec new_imports; > auto_vec interesting_phis; > bitmap_iterator bi; > unsigned i; > + auto_vec worklist; > EXECUTE_IF_SET_IN_BITMAP (interesting, 0, i, bi) > { > tree name = ssa_name (i); > gimple *def_stmt = SSA_NAME_DEF_STMT (name); > + /* Imports remain interesting. */ > if (gimple_bb (def_stmt) != bb) > - bitmap_set_bit (new_interesting, i); > - else if (gphi *phi = dyn_cast (def_stmt)) > { > - tree res = gimple_phi_result (phi); > - if (!SSA_NAME_OCCURS_IN_ABNORMAL_PHI (res)) > - interesting_phis.safe_push (phi); > + bitmap_set_bit (new_interesting, i); > + continue; > + } > + worklist.quick_push (name); > + while (!worklist.is_empty ()) > + { > + tree name = worklist.pop (); > + gimple *def_stmt = SSA_NAME_DEF_STMT (name); > + /* Newly discovered imports are interesting. */ > + if (gimple_bb (def_stmt) != bb) > + { > + bitmap_set_bit (new_interesting, SSA_NAME_VERSION (name)); > + continue; > + } > + /* Local PHIs participate in renaming below. */ > + if (gphi *phi = dyn_cast (def_stmt)) > + { > + tree res = gimple_phi_result (phi); > + if (!SSA_NAME_OCCURS_IN_ABNORMAL_PHI (res)) > + interesting_phis.safe_push (phi); > + } > + /* For other local defs process their uses, amending > + imports on the way. */ > + else if (is_gimple_assign (def_stmt)) > + { > + for (unsigned j = 1; j < gimple_num_ops (def_stmt); ++j) > + { > + tree rhs = gimple_op (def_stmt, j); > + if (TREE_CODE (rhs) == SSA_NAME > + && bitmap_set_bit (m_imports, > + SSA_NAME_VERSION (rhs))) > + { > + /* ??? We probably want to track a 'visited' > + flag separately and add to imports only > + when the def is handled by ranger. The > + current way adds us defs that are neither > + PHIs nor "interesting" assigns, like for > + example loads. */ > + new_imports.safe_push (SSA_NAME_VERSION (rhs)); > + worklist.safe_push (rhs); > + } > + } > + } > } > } > + worklist.release (); > if (!bitmap_empty_p (new_interesting) > || !interesting_phis.is_empty ()) > { > - auto_vec unwind (interesting_phis.length ()); > + auto_vec unwind (interesting_phis.length ()); > + auto_vec imports_unwind (interesting_phis.length ()); > edge_iterator iter; > edge e; > FOR_EACH_EDGE (e, iter, bb->preds) > @@ -525,22 +569,31 @@ back_threader::find_paths_to_names (basic_block bb, bitmap interesting, > { > tree def = PHI_ARG_DEF_FROM_EDGE (phi, e); > if (TREE_CODE (def) == SSA_NAME) > - if (bitmap_set_bit (new_interesting, > - SSA_NAME_VERSION (def))) > - { > - bitmap_set_bit (m_imports, SSA_NAME_VERSION (def)); > - unwind.quick_push (def); > - } > + { > + int ver = SSA_NAME_VERSION (def); > + if (bitmap_set_bit (new_interesting, ver)) > + { > + if (bitmap_set_bit (m_imports, ver)) > + imports_unwind.quick_push (ver); > + unwind.quick_push (ver); > + } > + } > } > find_paths_to_names (e->src, new_interesting, overall_paths); > - // Restore new_interesting. We leave m_imports alone since > - // we do not prune defs in BB from it and separately keeping > - // track of which bits to unwind isn't worth the trouble. > - for (tree def : unwind) > - bitmap_clear_bit (new_interesting, SSA_NAME_VERSION (def)); > + // Restore new_interesting. > + for (int def : unwind) > + bitmap_clear_bit (new_interesting, def); > unwind.truncate (0); > + // Restore and m_imports. > + for (int def : imports_unwind) > + bitmap_clear_bit (m_imports, def); > + imports_unwind.truncate (0); > } > } > + /* m_imports tracks all interesting names on the path, so when > + backtracking we have to restore it. */ > + for (int j : new_imports) > + bitmap_clear_bit (m_imports, j); > } > else if (dump_file && (dump_flags & TDF_DETAILS)) > fprintf (dump_file, " FAIL: Search space limit %d reached.\n", > @@ -566,15 +619,24 @@ back_threader::find_paths (basic_block bb, tree name) > && gimple_code (stmt) != GIMPLE_SWITCH)) > return; > > - if (EDGE_COUNT (bb->succs) > 1 > - || single_succ_to_potentially_threadable_block (bb)) > + if (EDGE_COUNT (bb->succs) > 1) > { > m_last_stmt = stmt; > m_visited_bbs.empty (); > m_path.truncate (0); > m_name = name; > - m_solver->compute_imports (m_imports, bb); > > + // We compute imports of the path during discovery starting > + // just with names used in the conditional. > + bitmap_clear (m_imports); > + ssa_op_iter iter; > + FOR_EACH_SSA_TREE_OPERAND (name, stmt, iter, SSA_OP_USE) > + bitmap_set_bit (m_imports, SSA_NAME_VERSION (name)); > + > + // Interesting is the set of imports we still not have see > + // the definition of. So while imports only grow, the > + // set of interesting defs dwindles and once empty we can > + // stop searching. > auto_bitmap interesting; > bitmap_copy (interesting, m_imports); > find_paths_to_names (bb, interesting, 1);