From owner-freebsd-arm@FreeBSD.ORG Fri Dec 28 17:13:44 2012 Return-Path: Delivered-To: freebsd-arm@freebsd.org Received: from mx1.freebsd.org (mx1.freebsd.org [69.147.83.52]) by hub.freebsd.org (Postfix) with ESMTP id 8B45F418; Fri, 28 Dec 2012 17:13:44 +0000 (UTC) (envelope-from freebsd@damnhippie.dyndns.org) Received: from duck.symmetricom.us (duck.symmetricom.us [206.168.13.214]) by mx1.freebsd.org (Postfix) with ESMTP id D2CEA8FC16; Fri, 28 Dec 2012 17:13:43 +0000 (UTC) Received: from damnhippie.dyndns.org (daffy.symmetricom.us [206.168.13.218]) by duck.symmetricom.us (8.14.5/8.14.5) with ESMTP id qBSHDafK044048; Fri, 28 Dec 2012 10:13:36 -0700 (MST) (envelope-from freebsd@damnhippie.dyndns.org) Received: from [172.22.42.240] (revolution.hippie.lan [172.22.42.240]) by damnhippie.dyndns.org (8.14.3/8.14.3) with ESMTP id qBSHDXh3076321; Fri, 28 Dec 2012 10:13:33 -0700 (MST) (envelope-from freebsd@damnhippie.dyndns.org) Subject: Re: FreeBSD/armv6z/clang on Raspberry Pi 512MB (with U-Boot + ubldr) From: Ian Lepore To: Adrian Chadd In-Reply-To: References: <2BA73CBF02B04DD19D08CDFC556B8750@ad.peach.ne.jp> <1356624694.1144.67.camel@revolution.hippie.lan> <234E1E18AC1C4A3985D6C570F78698E6@ad.peach.ne.jp> <9ED42265200A41B1BD637682720E0E9D@ad.peach.ne.jp> Content-Type: text/plain; charset="us-ascii" Date: Fri, 28 Dec 2012 10:13:33 -0700 Message-ID: <1356714813.1144.92.camel@revolution.hippie.lan> Mime-Version: 1.0 X-Mailer: Evolution 2.32.1 FreeBSD GNOME Team Port Content-Transfer-Encoding: 7bit Cc: David Chisnall , freebsd-arm@freebsd.org X-BeenThere: freebsd-arm@freebsd.org X-Mailman-Version: 2.1.14 Precedence: list List-Id: Porting FreeBSD to the StrongARM Processor List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Fri, 28 Dec 2012 17:13:44 -0000 It shouldn't be padding, the pointer is aligned correctly. The problem is that it's optimizing a pair of word stores into a single doubleword store when the alignment of the pair doesn't allow that optimization. Some arm hardware allows unaligned access, other hardware doesn't, so maybe we should be setting some clang optimization flag to let it know what's legal for the target platform? -- Ian On Fri, 2012-12-28 at 08:47 -0800, Adrian Chadd wrote: > .. the compiler should know the alignment for each of those types and > pad the structure appropriately, right? > > david - what's the "right" behaviour here? Surely clang should be > inserting 4 bytes of padding before that pointer? > > > > Adrian > > > On 28 December 2012 08:41, Daisuke Aoyama wrote: > >>>> This is complete version of systimer patch. > >>>> It should fix stopping interrupt and related things such as USB LAN is > >>>> unstable, > >>>> SSH is closed suddenly. > >>>> (I've not yet finished all test at this time, but at least portsnap > >>>> fetch is > >>>> success. > >>>> Now extracting the ports without any problems.) > > > > > > Finished without any problems. > > > > > > I'm checking the alignment of clang now. It seems no difference of location > > of data structure. > > However, access method is different. > > > > I learned clang will combine two loads into one op. > > This is a reason why the alignment seems difference between clang and gcc. > > Also, a reason why clang binary is smaller than gcc code. > > > > According to ARM manual, strd alignment is: > > "The address must be a multiple of eight for doubleword transfers." > > > > uboot/lib/api_public.h: > > ---------------------------------------------------------------------- > > struct sys_info { > > unsigned long clk_bus; > > unsigned long clk_cpu; > > unsigned long bar; > > struct mem_region *mr; /* << mr offset is 12!! (not DW > > aligned) */ > > int mr_no; /* number of memory regions */ > > }; > > > > uboot/lib/glue.c: > > ---------------------------------------------------------------------- > > struct sys_info * > > ub_get_sys_info(void) > > { > > int err = 0; > > > > memset(&si, 0, sizeof(struct sys_info)); > > si.mr = mr; > > si.mr_no = UB_MAX_MR; > > memset(&mr, 0, sizeof(mr)); > > > > if (!syscall(API_GET_SYS_INFO, &err, (u_int32_t)&si)) > > return (NULL); > > > > return ((err) ? NULL : &si); > > } > > > > ---------------------------------------------------------------------- > > clang -O2 output (7 steps): > > bl memset ; ATPCS uses r0-r3 as parameters > > ldr r0, .LCPI6_1 ; mr > > mov r1, #5 ; UB_MAX_MR > > mov r2, #60 ; sizeof(mr) > > strd r0, r1, [r5, #12] ; r5 aligned but strd requires DW(8byte) > > alignment (faulted here) > > mov r1, #0 > > bl memset ; memset(&mr, 0, sizeof(mr)); > > > > clang final binary size(2.8% smaller): > > -rwxr-xr-x 1 root wheel 235984 Dec 28 23:49 ubldr > > ---------------------------------------------------------------------- > > gcc 4.2 -O2 output (9 steps): > > bl memset ; ATPCS uses r0-r3 as parameters > > ldr ip, .L162+4 ; mr > > mov r3, #5 ; UB_MAX_MR > > mov r1, r4 ; r4 is zero > > mov r0, ip ; << what?? > > mov r2, #60 ; sizeof(mr) > > str r3, [r6, #16] ; r6 aligned same as clang > > str ip, [r6, #12] ; r6 aligned same as clang > > bl memset ; memset(&mr, 0, sizeof(mr)); > > > > gcc final binary size: > > -rwxr-xr-x 1 root wheel 242810 Dec 28 18:22 ubldr > > ---------------------------------------------------------------------- > > I don't know gcc 4.3 or newer, but probably output is more smart. > > It seems that there is no reason to use ip in this case. > > > > Does any one know how to prevent above clang output? > > (or how to solve this issue for all codes.) > > > > Thanks > > > > -- > > Daisuke Aoyama > > > > > > _______________________________________________ > > freebsd-arm@freebsd.org mailing list > > http://lists.freebsd.org/mailman/listinfo/freebsd-arm > > To unsubscribe, send any mail to "freebsd-arm-unsubscribe@freebsd.org" > _______________________________________________ > freebsd-arm@freebsd.org mailing list > http://lists.freebsd.org/mailman/listinfo/freebsd-arm > To unsubscribe, send any mail to "freebsd-arm-unsubscribe@freebsd.org"