From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (qmail 26631 invoked by alias); 18 Nov 2004 01:32:30 -0000 Mailing-List: contact gcc-bugs-help@gcc.gnu.org; run by ezmlm Precedence: bulk List-Archive: List-Post: List-Help: Sender: gcc-bugs-owner@gcc.gnu.org Received: (qmail 26607 invoked by uid 48); 18 Nov 2004 01:32:27 -0000 Date: Thu, 18 Nov 2004 01:32:00 -0000 Message-ID: <20041118013227.26606.qmail@sourceware.org> From: "jgrimm2 at us dot ibm dot com" To: gcc-bugs@gcc.gnu.org In-Reply-To: <20041012174801.17950.steinmtz@us.ibm.com> References: <20041012174801.17950.steinmtz@us.ibm.com> Reply-To: gcc-bugzilla@gcc.gnu.org Subject: [Bug rtl-optimization/17950] Over Aggressive Use of Data Cache Touch Instructions During FDO X-Bugzilla-Reason: CC X-SW-Source: 2004-11/txt/msg02109.txt.bz2 List-Id: ------- Additional Comments From jgrimm2 at us dot ibm dot com 2004-11-18 01:32 ------- Hmmm... looks like a couple things (maybe more) come in to play here. First, value-profile code doesn't seem to merge prefetches (among other things). That is, it should be pretty easy (crude?) to track the last prefetch and not prefetch addresses that look close. Doing so seems to help this test case quite a bit (96 prefetches down to 24), but probably not enough..... as it seems that the loop unroller comes through afterwards and does its damage too. Might need to teach the unroller about prefetch a bit too?? -- What |Removed |Added ---------------------------------------------------------------------------- CC| |jgrimm2 at us dot ibm dot | |com http://gcc.gnu.org/bugzilla/show_bug.cgi?id=17950