From patchwork Fri Jul 26 10:40:33 2024 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Manolis Tsamis X-Patchwork-Id: 1965257 Return-Path: X-Original-To: incoming@patchwork.ozlabs.org Delivered-To: patchwork-incoming@legolas.ozlabs.org Authentication-Results: legolas.ozlabs.org; dkim=pass (2048-bit key; unprotected) header.d=vrull.eu header.i=@vrull.eu header.a=rsa-sha256 header.s=google header.b=I45SX8go; dkim-atps=neutral Authentication-Results: legolas.ozlabs.org; spf=pass (sender SPF authorized) smtp.mailfrom=gcc.gnu.org (client-ip=2620:52:3:1:0:246e:9693:128c; helo=server2.sourceware.org; envelope-from=gcc-patches-bounces~incoming=patchwork.ozlabs.org@gcc.gnu.org; receiver=patchwork.ozlabs.org) Received: from server2.sourceware.org (server2.sourceware.org [IPv6:2620:52:3:1:0:246e:9693:128c]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange X25519 server-signature ECDSA (secp384r1) server-digest SHA384) (No client certificate requested) by legolas.ozlabs.org (Postfix) with ESMTPS id 4WVknR2RJgz1yXx for ; Fri, 26 Jul 2024 20:41:15 +1000 (AEST) Received: from server2.sourceware.org (localhost [IPv6:::1]) by sourceware.org (Postfix) with ESMTP id 6BB48383E52A for ; Fri, 26 Jul 2024 10:41:13 +0000 (GMT) X-Original-To: gcc-patches@gcc.gnu.org Delivered-To: gcc-patches@gcc.gnu.org Received: from mail-lj1-x232.google.com (mail-lj1-x232.google.com [IPv6:2a00:1450:4864:20::232]) by sourceware.org (Postfix) with ESMTPS id 34613385DC1B for ; Fri, 26 Jul 2024 10:40:46 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.2 sourceware.org 34613385DC1B Authentication-Results: sourceware.org; dmarc=none (p=none dis=none) header.from=vrull.eu Authentication-Results: sourceware.org; spf=pass smtp.mailfrom=vrull.eu ARC-Filter: OpenARC Filter v1.0.0 sourceware.org 34613385DC1B Authentication-Results: server2.sourceware.org; arc=none smtp.remote-ip=2a00:1450:4864:20::232 ARC-Seal: i=1; a=rsa-sha256; d=sourceware.org; s=key; t=1721990449; cv=none; b=RUm62HD0jhe/zRBh/fZ6pO7+e8NkLrpIn85JaVEdARLadjHtapVIOl5JjmpDe5L/TCq5yUQgvfKZFZlUpt7zVEeqsoXdnao45yVvv7aw8G5bDjd5pfrFyxLlMpsErsbAN6s4YZCi0rRuu8uwG/+kv+6XYW8E7d9atySa1InsHa0= ARC-Message-Signature: i=1; a=rsa-sha256; d=sourceware.org; s=key; t=1721990449; c=relaxed/simple; bh=Za9lCteKh7lbG0wXa6eUJOUlJ+mIi0xVG3zkDvZ4st8=; h=DKIM-Signature:From:To:Subject:Date:Message-Id:MIME-Version; b=KbSUOIRFdlr1XSJw6FSTU/EVeZcKri2HE5OEeK3bBb+X/bqBu+598jcluW0tMN3WzfcC3GJdXDeimvfT7zxyVpw9TmpXDEq0f74L7+3xGP3GYNXQLIihfhGPWoqwdSV2ZR+9U5Al4eaBDQgWbwLkKzzJOC3ZhB4f+JhssYpi7K4= ARC-Authentication-Results: i=1; server2.sourceware.org Received: by mail-lj1-x232.google.com with SMTP id 38308e7fff4ca-2ef2ed592f6so11925641fa.0 for ; Fri, 26 Jul 2024 03:40:46 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=vrull.eu; s=google; t=1721990444; x=1722595244; darn=gcc.gnu.org; h=content-transfer-encoding:mime-version:references:in-reply-to :message-id:date:subject:cc:to:from:from:to:cc:subject:date :message-id:reply-to; bh=w4SxUy7BJDgo3DllO64PTDX3GrlGCkFNuq8McAX3wlc=; b=I45SX8goqpTmieUiJ8DDzdjpX0ZCntgAB0Cbwgh03KKgkyD8R1TvIz/cCBLCVBElCi evACgWijWUvxATH9Mh/ku+WRPx+8cGkZAgrWSEfMmAclObkz2p4LLJQ+X54N+ON+N9rn YRJhwA9+09WRHeMqu7UyGv6FxMM53XqrLtGPSqXNQCbG7UbJ9aVM3g9iwFqBZX5qRIEt fZqNq4dmiNSBQp4Ij0qEYVQH3jQ/fYY20b8fEG+jUt8VWGUTIiZwXbUtXyF4Sof5mIlZ 2HZGsEpEgsyNbmyxJ4o7/NkvYlkrSVCnSOjG4YX/0UCE88gulMd99DNJg5ofIyv064/W hI9Q== X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20230601; t=1721990444; x=1722595244; h=content-transfer-encoding:mime-version:references:in-reply-to :message-id:date:subject:cc:to:from:x-gm-message-state:from:to:cc :subject:date:message-id:reply-to; bh=w4SxUy7BJDgo3DllO64PTDX3GrlGCkFNuq8McAX3wlc=; b=NBdb9gR/T7WMnavmbLxhcJJNPPBMHtMzVsXcI7M5xnqKja6YtDtU1U87DoXfIsy33+ cyX5MlXYjrHgSTQDwNdBd9JTK7rluMGsI74WWWBHiHqT9CsGmikV70jEEJR5kpn+rOaJ kJWpZ3xBIwop4aZP1vxTbQSVkqIOOliJk2fX44LxFlqK67UffWRUjP21U1dSDW1GJLg4 W3zhvGQjwov5Nfdr+twd8oAV2DLmPHwwd45K6NQCvbNE02yqujlvasPIbCUy3fildtAM 4Q8ou6hLc42aE+m8cxN1kGMUNhGZR9sB3VknMeACfD8X/DVAVqh5vyQg0mvHSmOh/vAk a5FQ== X-Gm-Message-State: AOJu0YxLr2X5tMX9DoV3X+buoCVwWYkqxl1Uotx8xWab6rzDMVhcET30 JPj0I+HyuqxaIDRNt8Xq9yJGuS4GwOVy4292qY1KjyJVG7ekurDKOUGQx9fiPF4j9kLlBrE2yxq eUW4= X-Google-Smtp-Source: AGHT+IFkO12OqSl4zBdcE3wWOFCZJA8TxWvmMN7YkRrJln6j37GwQKgAymTTc6p+z04Ao8Fr5AobaQ== X-Received: by 2002:a2e:900b:0:b0:2ef:2b45:b71d with SMTP id 38308e7fff4ca-2f03dba5d54mr32912151fa.24.1721990444032; Fri, 26 Jul 2024 03:40:44 -0700 (PDT) Received: from helsinki-03.engr ([2a01:4f9:6b:2a47::2]) by smtp.gmail.com with ESMTPSA id 38308e7fff4ca-2f03cf0e38fsm4249731fa.4.2024.07.26.03.40.43 (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Fri, 26 Jul 2024 03:40:43 -0700 (PDT) From: Manolis Tsamis To: gcc-patches@gcc.gnu.org Cc: Robin Dapp , Philipp Tomsich , Jiangning Liu , Richard Sandiford , Richard Biener , Jeff Law , Manolis Tsamis Subject: [RESEND PATCH v5 1/3] ifcvt: handle sequences that clobber flags in noce_convert_multiple_sets Date: Fri, 26 Jul 2024 12:40:33 +0200 Message-Id: <20240726104035.4100769-2-manolis.tsamis@vrull.eu> X-Mailer: git-send-email 2.34.1 In-Reply-To: <20240726104035.4100769-1-manolis.tsamis@vrull.eu> References: <20240726104035.4100769-1-manolis.tsamis@vrull.eu> MIME-Version: 1.0 X-Spam-Status: No, score=-11.9 required=5.0 tests=BAYES_00, DKIM_SIGNED, DKIM_VALID, DKIM_VALID_AU, DKIM_VALID_EF, GIT_PATCH_0, RCVD_IN_DNSWL_NONE, SPF_HELO_NONE, SPF_PASS, TXREP 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: 2.1.30 Precedence: list List-Id: Gcc-patches mailing list List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: gcc-patches-bounces~incoming=patchwork.ozlabs.org@gcc.gnu.org This is an extension of what was done in PR106590. Currently if a sequence generated in noce_convert_multiple_sets clobbers the condition rtx (cc_cmp or rev_cc_cmp) then only seq1 is used afterwards (sequences that emit the comparison itself). Since this applies only from the next iteration it assumes that the sequences generated (in particular seq2) doesn't clobber the condition rtx itself before using it in the if_then_else, which is only true in specific cases (currently only register/subregister moves are allowed). This patch changes this so it also tests if seq2 clobbers cc_cmp/rev_cc_cmp in the current iteration. It also checks whether the resulting sequence clobbers the condition attached to the jump. This makes it possible to include arithmetic operations in noce_convert_multiple_sets. It also makes the code that checks whether the condition is used outside of the if_then_else emitted more robust. gcc/ChangeLog: * ifcvt.cc (check_for_cc_cmp_clobbers): Use modified_in_p instead. (noce_convert_multiple_sets_1): Don't use seq2 if it clobbers cc_cmp. Punt if seq clobbers cond. Refactor the code that sets read_comparison. Signed-off-by: Manolis Tsamis --- (no changes since v1) gcc/ifcvt.cc | 127 ++++++++++++++++++++++++++++++++------------------- 1 file changed, 79 insertions(+), 48 deletions(-) diff --git a/gcc/ifcvt.cc b/gcc/ifcvt.cc index 58ed42673e5..58c34aaf1ee 100644 --- a/gcc/ifcvt.cc +++ b/gcc/ifcvt.cc @@ -3592,20 +3592,6 @@ noce_convert_multiple_sets (struct noce_if_info *if_info) return true; } -/* Helper function for noce_convert_multiple_sets_1. If store to - DEST can affect P[0] or P[1], clear P[0]. Called via note_stores. */ - -static void -check_for_cc_cmp_clobbers (rtx dest, const_rtx, void *p0) -{ - rtx *p = (rtx *) p0; - if (p[0] == NULL_RTX) - return; - if (reg_overlap_mentioned_p (dest, p[0]) - || (p[1] && reg_overlap_mentioned_p (dest, p[1]))) - p[0] = NULL_RTX; -} - /* This goes through all relevant insns of IF_INFO->then_bb and tries to create conditional moves. In case a simple move sufficis the insn should be listed in NEED_NO_CMOV. The rewired-src cases should be @@ -3731,36 +3717,71 @@ noce_convert_multiple_sets_1 (struct noce_if_info *if_info, creating an additional compare for each. If successful, costing is easier and this sequence is usually preferred. */ if (cc_cmp) - seq2 = try_emit_cmove_seq (if_info, temp, cond, - new_val, old_val, need_cmov, - &cost2, &temp_dest2, cc_cmp, rev_cc_cmp); + { + seq2 = try_emit_cmove_seq (if_info, temp, cond, + new_val, old_val, need_cmov, + &cost2, &temp_dest2, cc_cmp, rev_cc_cmp); + + /* The if_then_else in SEQ2 may be affected when cc_cmp/rev_cc_cmp is + clobbered. We can't safely use the sequence in this case. */ + for (rtx_insn *iter = seq2; iter; iter = NEXT_INSN (iter)) + if (modified_in_p (cc_cmp, iter) + || (rev_cc_cmp && modified_in_p (rev_cc_cmp, iter))) + { + seq2 = NULL; + break; + } + } /* The backend might have created a sequence that uses the - condition. Check this. */ + condition as a value. Check this. */ + + /* We cannot handle anything more complex than a reg or constant. */ + if (!REG_P (XEXP (cond, 0)) && !CONSTANT_P (XEXP (cond, 0))) + read_comparison = true; + + if (!REG_P (XEXP (cond, 1)) && !CONSTANT_P (XEXP (cond, 1))) + read_comparison = true; + rtx_insn *walk = seq2; - while (walk) + int if_then_else_count = 0; + while (walk && !read_comparison) { - rtx set = single_set (walk); + rtx exprs_to_check[2]; + unsigned int exprs_count = 0; - if (!set || !SET_SRC (set)) + rtx set = single_set (walk); + if (set && XEXP (set, 1) + && GET_CODE (XEXP (set, 1)) == IF_THEN_ELSE) { - walk = NEXT_INSN (walk); - continue; + /* We assume that this is the cmove created by the backend that + naturally uses the condition. */ + exprs_to_check[exprs_count++] = XEXP (XEXP (set, 1), 1); + exprs_to_check[exprs_count++] = XEXP (XEXP (set, 1), 2); + if_then_else_count++; } + else if (NONDEBUG_INSN_P (walk)) + exprs_to_check[exprs_count++] = PATTERN (walk); - rtx src = SET_SRC (set); + /* Bail if we get more than one if_then_else because the assumption + above may be incorrect. */ + if (if_then_else_count > 1) + { + read_comparison = true; + break; + } - if (XEXP (set, 1) && GET_CODE (XEXP (set, 1)) == IF_THEN_ELSE) - ; /* We assume that this is the cmove created by the backend that - naturally uses the condition. Therefore we ignore it. */ - else + for (unsigned int i = 0; i < exprs_count; i++) { - if (reg_mentioned_p (XEXP (cond, 0), src) - || reg_mentioned_p (XEXP (cond, 1), src)) - { - read_comparison = true; - break; - } + subrtx_iterator::array_type array; + FOR_EACH_SUBRTX (iter, array, exprs_to_check[i], NONCONST) + if (*iter != NULL_RTX + && (reg_overlap_mentioned_p (XEXP (cond, 0), *iter) + || reg_overlap_mentioned_p (XEXP (cond, 1), *iter))) + { + read_comparison = true; + break; + } } walk = NEXT_INSN (walk); @@ -3788,22 +3809,32 @@ noce_convert_multiple_sets_1 (struct noce_if_info *if_info, return false; } - if (cc_cmp) - { - /* Check if SEQ can clobber registers mentioned in - cc_cmp and/or rev_cc_cmp. If yes, we need to use - only seq1 from that point on. */ - rtx cc_cmp_pair[2] = { cc_cmp, rev_cc_cmp }; - for (walk = seq; walk; walk = NEXT_INSN (walk)) + /* Although we use temporaries if there is register overlap of COND and + TARGET, it is possible that SEQ modifies COND anyway. For example, + COND may use the flags register and if INSN clobbers flags then + we may be unable to emit a valid sequence (e.g. in x86 that would + require saving and restoring the flags register). */ + if (!second_try) + for (rtx_insn *iter = seq; iter; iter = NEXT_INSN (iter)) + if (modified_in_p (cond, iter)) { - note_stores (walk, check_for_cc_cmp_clobbers, cc_cmp_pair); - if (cc_cmp_pair[0] == NULL_RTX) - { - cc_cmp = NULL_RTX; - rev_cc_cmp = NULL_RTX; - break; - } + end_sequence (); + return false; } + + if (cc_cmp && seq == seq1) + { + /* Check if SEQ can clobber registers mentioned in cc_cmp/rev_cc_cmp. + If yes, we need to use only SEQ1 from that point on. + Only check when we use SEQ1 since we have already tested SEQ2. */ + for (rtx_insn *iter = seq; iter; iter = NEXT_INSN (iter)) + if (modified_in_p (cc_cmp, iter) + || (rev_cc_cmp && modified_in_p (rev_cc_cmp, iter))) + { + cc_cmp = NULL_RTX; + rev_cc_cmp = NULL_RTX; + break; + } } /* End the sub sequence and emit to the main sequence. */