From abc5b28db164dc2d807750cb2baae91e288c84a9 Mon Sep 17 00:00:00 2001 From: =?utf8?q?Rafa=C5=82=20Mi=C5=82ecki?= Date: Sun, 3 Jul 2022 13:22:00 +0200 Subject: [PATCH] bcm53xx: use -falign-functions=32 for kernel compilation MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit Northstar SoCs have pretty small CPU caches and their performance is heavily affected by cache hits & misses. It means that all kind of random code changes can affect performance as they often reorganize (change alignment & possibly reorder) kernel symbols. It was discussed in ARM / net mailinglists: 1. ARM router NAT performance affected by random/unrelated commits [1] [2] 2. Optimizing kernel compilation / alignments for network performance [3] [4] It seems that -falign-functions can be used as a partial workaround. It doesn't solve all cases (e.g. documented watchdog one [5]) but it surely helps with many of them. A complete long term solution may be PGO (profile-guided optimization) but it isn't available at this point. [1] https://lkml.org/lkml/2019/5/21/349 [2] https://www.spinics.net/lists/linux-block/msg40624.html [3] https://lore.kernel.org/linux-arm-kernel/066fc320-dc04-11a4-476e-b0d11f3b17e6@gmail.com/T/ [4] https://www.spinics.net/lists/netdev/msg816103.html [5] http://lists.openwrt.org/pipermail/openwrt-devel/2022-July/038989.html Signed-off-by: Rafał Miłecki --- config/Config-devel.in | 1 + 1 file changed, 1 insertion(+) diff --git a/config/Config-devel.in b/config/Config-devel.in index a068f694e5..87f9157106 100644 --- a/config/Config-devel.in +++ b/config/Config-devel.in @@ -76,6 +76,7 @@ menuconfig DEVEL config KERNEL_CFLAGS string "Kernel extra CFLAGS" if DEVEL + default "-falign-functions=32" if TARGET_bcm53xx default "" config EXTERNAL_KERNEL_TREE -- 2.30.2