Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S932085Ab1EUACQ (ORCPT ); Fri, 20 May 2011 20:02:16 -0400 Received: from mga01.intel.com ([192.55.52.88]:58587 "EHLO mga01.intel.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754835Ab1EUACI (ORCPT ); Fri, 20 May 2011 20:02:08 -0400 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="4.65,244,1304319600"; d="scan'208";a="5514053" From: Andi Kleen To: linux-kernel@vger.kernel.org Cc: akpm@linux-foundation.org, airlied@linux.ie, Andi Kleen Subject: [PATCH 01/12] Force always inline for gcc 4.5 when optimizing for size Date: Fri, 20 May 2011 17:01:11 -0700 Message-Id: <1305936082-21304-1-git-send-email-andi@firstfloor.org> X-Mailer: git-send-email 1.7.4.4 Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 3110 Lines: 79 From: Andi Kleen I found that gcc 4.5 didn't inline a lot of inlines with CONFIG_OPTIMIZE_INLINING and CONFIG_CC_OPTIMIZE_FOR_SIZE. It was quite common to have very small inlines to be out of line, or worse inline statics in include files to be out of line with a copy for every file using it too. This is handily visible in a function graph trace for might_fault: 10) | might_fault() { 10) | _cond_resched() { 10) | should_resched() { 10) | need_resched() { 10) 0.063 us | test_ti_thread_flag(); 10) 0.643 us | } 10) 1.238 us | } 10) 1.845 us | } 10) 2.438 us | } Note all of these functions are very small and should be definitely inlined in each other. In many cases even copy_from_user ends up out of line now which is really bad! If I switch to -O2 it is also not quite as bad, but since a lot of people use -Os I was trying to fix it up. So this patch forces inlining with gcc 4.4 with -Os. Unfortunately it costs some code size with just this patch. text data bss dec hex filename 11507035 1940276 1191936 14639247 df608f vmlinux-O2 10189858 1908124 1187840 13285822 cab9be vmlinux-Os-force 9808525 1940204 1187840 12936569 c56579 vmlinux-Os-orig But after some starring on bloat-o-meter it turned out only some subsystems (in my kernel) had a real problem. The biggest offender was DRM. I fixed those up manually by removing inlines. With these changes (and disabling DRM debugging, which is on by default) I get a kernel with force inline that is a few KB smaller. With DRM debugging enabled it's about 50k larger (nearly all of it in DRM, mostly radeon). I hope the default for this can be changed. I haven't tested earlier gcc 4.x versions, but they may need the same treatment. Signed-off-by: Andi Kleen --- include/linux/compiler-gcc.h | 5 ++++- 1 files changed, 4 insertions(+), 1 deletions(-) diff --git a/include/linux/compiler-gcc.h b/include/linux/compiler-gcc.h index cb4c1eb..0f2b513 100644 --- a/include/linux/compiler-gcc.h +++ b/include/linux/compiler-gcc.h @@ -40,9 +40,12 @@ /* * Force always-inline if the user requests it so via the .config, * or if gcc is too old: + * When optimizing for size on gcc 4.5 always force inlining too. */ #if !defined(CONFIG_ARCH_SUPPORTS_OPTIMIZED_INLINING) || \ - !defined(CONFIG_OPTIMIZE_INLINING) || (__GNUC__ < 4) + !defined(CONFIG_OPTIMIZE_INLINING) || (__GNUC__ < 4) || \ + (defined(CONFIG_CC_OPTIMIZE_FOR_SIZE) && \ + (__GNUC__ == 4 && __GNUC_MINOR__ == 5)) # define inline inline __attribute__((always_inline)) # define __inline__ __inline__ __attribute__((always_inline)) # define __inline __inline __attribute__((always_inline)) -- 1.7.4.4 -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/