From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (qmail 22597 invoked by alias); 8 Jun 2010 15:57:28 -0000 Received: (qmail 22587 invoked by uid 22791); 8 Jun 2010 15:57:26 -0000 X-SWARE-Spam-Status: No, hits=-1.2 required=5.0 tests=AWL,BAYES_05,T_RP_MATCHES_RCVD X-Spam-Check-By: sourceware.org Received: from mail.codesourcery.com (HELO mail.codesourcery.com) (38.113.113.100) by sourceware.org (qpsmtpd/0.43rc1) with ESMTP; Tue, 08 Jun 2010 15:57:21 +0000 Received: (qmail 20282 invoked from network); 8 Jun 2010 15:57:19 -0000 Received: from unknown (HELO ?84.152.236.246?) (bernds@127.0.0.2) by mail.codesourcery.com with ESMTPA; 8 Jun 2010 15:57:19 -0000 Message-ID: <4C0E6863.6010702@codesourcery.com> Date: Tue, 08 Jun 2010 15:57:00 -0000 From: Bernd Schmidt User-Agent: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.9.1.9) Gecko/20100425 Thunderbird/3.0.4 MIME-Version: 1.0 To: Richard Earnshaw CC: ramana.radhakrishnan@arm.com, GCC Patches Subject: Re: ARM ldm/stm peepholes References: <4BCD9301.2060605@codesourcery.com> <1272010406.6783.72.camel@e200593-lin.cambridge.arm.com> <1272027779.1977.22.camel@e200601-lin.cambridge.arm.com> In-Reply-To: <1272027779.1977.22.camel@e200601-lin.cambridge.arm.com> Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: 7bit 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 X-SW-Source: 2010-06/txt/msg00781.txt.bz2 With "official" performance numbers apparently unavailable, Ramana asked me to run some benchmarks, with a focus on when to enable generation of ldm insns on Cortex-A9 (the assumption being that on other cores, ldm should always be a win, and so should stm everywhere). I've tested three variants of the patch against SPEC2k on an A9 board - first the full patch, then a variant which restrict generation of ldm insns to sequences of 3 or more, and one variant which only allows ldm for 4 insn sequences. Both the full patch and the 3/4-only variant are quite close to an unpatched compiler. The 3/4-only variant is 3.5% better on 164.gzip vs. the full patch, which translates into a 0.3% improvement in overall score. The 4-only variant is significantly worse, with large drops in 164.gzip and 253.perlbmk, for an overall 1% lower score. Given these results, what would you like me to do with the patch? Leave it as-is? Modify it to disallow 2-insn ldms on Cortex-A9? Something else? Bernd