From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1754991Ab3KLSEG (ORCPT ); Tue, 12 Nov 2013 13:04:06 -0500 Received: from unicorn.mansr.com ([81.2.72.234]:38715 "EHLO unicorn.mansr.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751972Ab3KLSEB convert rfc822-to-8bit (ORCPT ); Tue, 12 Nov 2013 13:04:01 -0500 From: =?iso-8859-1?Q?M=E5ns_Rullg=E5rd?= To: Nicolas Pitre Cc: =?iso-8859-1?Q?M=E5ns_Rullg=E5rd?= , Ben Dooks , Russell King - ARM Linux , Stephen Boyd , linux-kernel@vger.kernel.org, Christopher Covington , Jean-Christophe PLAGNIOL-VILLARD , linux-arm-kernel@lists.infradead.org Subject: Re: [PATCH v2] ARM: Use udiv/sdiv for __aeabi_{u}idiv library functions References: <1383951632-6090-1-git-send-email-sboyd@codeaurora.org> <528193A0.7050505@codeaurora.org> <20131112140436.GK16735@n2100.arm.linux.org.uk> <52823877.60806@codethink.co.uk> Date: Tue, 12 Nov 2013 18:03:56 +0000 In-Reply-To: (Nicolas Pitre's message of "Tue, 12 Nov 2013 09:55:09 -0500 (EST)") Message-ID: User-Agent: Gnus/5.13 (Gnus v5.13) Emacs/24.3 (gnu/linux) MIME-Version: 1.0 Content-Type: text/plain; charset=iso-8859-1 Content-Transfer-Encoding: 8BIT Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Nicolas Pitre writes: > On Tue, 12 Nov 2013, Måns Rullgård wrote: > >> Nicolas Pitre writes: >> >> > On Tue, 12 Nov 2013, Ben Dooks wrote: >> > >> >> Given these are single instructoins for ARM, is it possible we could >> >> make a table of all the callers and fix them up when we initialise >> >> as we do for the SMP/UP case and for page-offset? >> > >> > Not really. Calls to those functions are generated by the compiler >> > implicitly when a divisor operand is used and therefore we cannot >> > annotate those calls. We'd have to use special accessors everywhere to >> > replace the standard division operand (like we do for 64 by 32 bit >> > divisions) but I doubt that people would accept that. >> >> It might be possible to extract this information from relocation tables. > > True, but only for individual .o files. Once the linker puts them > together the information is lost, and trying to infer what the linker > has done is insane. > > Filtering the compiler output to annotate idiv calls before it is > assembled would probably be a better solution. OK, here's an extremely ugly hootenanny of a patch. It seems to work on an A7 Cubieboard2. I would never suggest actually doing this, but maybe it can be useful for comparing performance against the more palatable solutions. diff --git a/arch/arm/Makefile b/arch/arm/Makefile index 7397db6..cf1cd30 100644 --- a/arch/arm/Makefile +++ b/arch/arm/Makefile @@ -113,7 +113,7 @@ endif endif # Need -Uarm for gcc < 3.x -KBUILD_CFLAGS +=$(CFLAGS_ABI) $(CFLAGS_THUMB2) $(arch-y) $(tune-y) $(call cc-option,-mshort-load-bytes,$(call cc-option,-malignment-traps,)) -msoft-float -Uarm +KBUILD_CFLAGS +=$(CFLAGS_ABI) $(CFLAGS_THUMB2) $(arch-y) $(tune-y) $(call cc-option,-mshort-load-bytes,$(call cc-option,-malignment-traps,)) -msoft-float -Uarm -include asm/divhack.h KBUILD_AFLAGS +=$(CFLAGS_ABI) $(AFLAGS_THUMB2) $(arch-y) $(tune-y) -include asm/unified.h -msoft-float CHECKFLAGS += -D__arm__ diff --git a/arch/arm/include/asm/divhack.h b/arch/arm/include/asm/divhack.h new file mode 100644 index 0000000..c750b78 --- /dev/null +++ b/arch/arm/include/asm/divhack.h @@ -0,0 +1,23 @@ +__asm__ (".macro dobl tgt \n" + " .ifc \\tgt, __aeabi_idiv \n" + " .L.sdiv.\\@: \n" + " .pushsection .sdiv_tab.init, \"a\", %progbits \n" + " .word .L.sdiv.\\@ \n" + " .popsection \n" + " .endif \n" + " .ifc \\tgt, __aeabi_uidiv \n" + " .L.udiv.\\@: \n" + " .pushsection .udiv_tab.init, \"a\", %progbits \n" + " .word .L.udiv.\\@ \n" + " .popsection \n" + " .endif \n" + " bl \\tgt \n" + ".endm \n" + ".macro defbl \n" + " .macro bl tgt \n" + " .purgem bl \n" + " dobl \\tgt \n" + " defbl \n" + " .endm \n" + ".endm \n" + "defbl \n"); diff --git a/arch/arm/kernel/setup.c b/arch/arm/kernel/setup.c index 067815c1..b3a3fe1 100644 --- a/arch/arm/kernel/setup.c +++ b/arch/arm/kernel/setup.c @@ -375,6 +375,18 @@ static void __init cpuid_init_hwcaps(void) case 1: elf_hwcap |= HWCAP_IDIVT; } + + if (!IS_ENABLED(CONFIG_THUMB2_KERNEL) && (elf_hwcap & HWCAP_IDIVA)) { + extern u32 __sdiv_tab_start, __sdiv_tab_end; + extern u32 __udiv_tab_start, __udiv_tab_end; + u32 *div; + + for (div = &__sdiv_tab_start; div < &__sdiv_tab_end; div++) + *(u32 *)*div = 0xe710f110; + + for (div = &__udiv_tab_start; div < &__udiv_tab_end; div++) + *(u32 *)*div = 0xe730f110; + } } static void __init feat_v6_fixup(void) diff --git a/arch/arm/kernel/vmlinux.lds.S b/arch/arm/kernel/vmlinux.lds.S index 43a31fb..3d5c103 100644 --- a/arch/arm/kernel/vmlinux.lds.S +++ b/arch/arm/kernel/vmlinux.lds.S @@ -176,6 +176,8 @@ SECTIONS CON_INITCALL SECURITY_INITCALL INIT_RAM_FS + __sdiv_tab_start = .; *(.sdiv_tab.init); __sdiv_tab_end = .; + __udiv_tab_start = .; *(.udiv_tab.init); __udiv_tab_end = .; } #ifndef CONFIG_XIP_KERNEL .exit.data : { -- Måns Rullgård mans@mansr.com