This is the mail archive of the
newlib@sourceware.org
mailing list for the newlib project.
Re: Optimized arm string routines
On Tue, 2009-01-20 at 10:21 +0000, Richard Earnshaw wrote:
> I've recently written the following hand-optimized versions of the
> string functions strlen, strcpy and strcmp for ARM (including support
> for thumb1 and thumb2). These are attached for inclusion in newlib.
>
> They should perform significantly better than the C coded versions since
> they invariably use fewer call-saved registers (a major overhead on
> short strings). They also handle more cases than the C versions before
> falling back to simple byte loops.
>
> The thumb1 variants are all optimized for space. The inner loops become
> very complex with the restricted instruction set and it's unlikely to
> give any significant performance benefit.
>
> R.
>
> <date> Richard Earnshaw <rearnsha@arm.com>
>
> * libc/machine/arm/arm_asm.h: New file.
> * libc/machine/arm/strlen.c: New file.
> * libc/machine/arm/strcpy.c: New file.
> * libc/machine/arm/strcmp.c: New file.
> * libc/machine/arm/Makefile.am: Add new string routines.
>
>
Whops, the strcpy routine had a bug in it, which I managed to introduce
late yesterday during final tweaking :-(
Please use this version rather than the one in the original patch
R.
/*
* Copyright (c) 2008 ARM Ltd
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* 3. The name of the company may not be used to endorse or promote
* products derived from this software without specific prior written
* permission.
*
* THIS SOFTWARE IS PROVIDED BY ARM LTD ``AS IS'' AND ANY EXPRESS OR IMPLIED
* WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
* IN NO EVENT SHALL ARM LTD BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
* SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
* TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
* PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
* NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
* SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include "arm_asm.h"
#include <_ansi.h>
#include <string.h>
#ifdef __thumb2__
#define magic1(REG) "#0x01010101"
#define magic2(REG) "#0x80808080"
#else
#define magic1(REG) #REG
#define magic2(REG) #REG ", lsl #7"
#endif
char* __attribute__((naked))
strcpy (char* dst, const char* src)
{
asm (
#if !(defined(__OPTIMIZE_SIZE__) || defined (PREFER_SIZE_OVER_SPEED) || \
(defined (__thumb__) && !defined (__thumb2__)))
"optpld r1\n\t"
"eor r2, r0, r1\n\t"
"mov ip, r0\n\t"
"tst r2, #3\n\t"
"bne 4f\n\t"
"tst r1, #3\n\t"
"bne 3f\n"
"5:\n\t"
#ifndef __thumb2__
"str r5, [sp, #-4]!\n\t"
"mov r5, #0x01\n\t"
"orr r5, r5, r5, lsl #8\n\t"
"orr r5, r5, r5, lsl #16\n\t"
#endif
"str r4, [sp, #-4]!\n\t"
"tst r1, #4\n\t"
"ldr r3, [r1], #4\n\t"
"beq 2f\n\t"
"sub r2, r3, "magic1(r5)"\n\t"
"eors r2, r2, r3\n\t"
"tst r2, "magic2(r5)"\n\t"
"itt eq\n\t"
"streq r3, [ip], #4\n\t"
"ldreq r3, [r1], #4\n"
"bne 1f\n\t"
/* Inner loop. We now know that r1 is 64-bit aligned, so we
can safely fetch up to two words. This allows us to avoid
load stalls. */
".p2align 2\n"
"2:\n\t"
"optpld r1, #8\n\t"
"ldr r4, [r1], #4\n\t"
"sub r2, r3, "magic1(r5)"\n\t"
"eors r2, r2, r3\n\t"
"tst r2, "magic2(r5)"\n\t"
"sub r2, r4, "magic1(r5)"\n\t"
"bne 1f\n\t"
"str r3, [ip], #4\n\t"
"eors r2, r2, r4\n\t"
"tst r2, "magic2(r5)"\n\t"
"itt eq\n\t"
"ldreq r3, [r1], #4\n\t"
"streq r4, [ip], #4\n\t"
"beq 2b\n\t"
"mov r3, r4\n"
"1:\n\t"
#ifdef __ARMEB__
"rors r3, r3 #24\n\t"
#endif
"strb r3, [ip], #1\n\t"
"tst r3, #0xff\n\t"
#ifdef __ARMEL__
"ror r3, r3, #8\n\t"
#endif
"bne 1b\n\t"
"ldr r4, [sp], #4\n\t"
#ifndef __thumb2__
"ldr r5, [sp], #4\n\t"
#endif
"RETURN\n"
/* Strings have the same offset from word alignment, but it's
not zero. */
"3:\n\t"
"tst r1, #1\n\t"
"beq 1f\n\t"
"ldrb r2, [r1], #1\n\t"
"strb r2, [ip], #1\n\t"
"cmp r2, #0\n\t"
"it eq\n"
"RETURN eq\n"
"1:\n\t"
"tst r1, #2\n\t"
"beq 5b\n\t"
"ldrh r2, [r1], #2\n\t"
#ifdef __ARMEB__
"tst r2, #0xff00\n\t"
"iteet ne\n\t"
"strneh r2, [ip], #2\n\t"
"lsreq r2, r2, #8\n\t"
"streqb r2, [ip]\n\t"
"tstne r2, #0xff\n\t"
#else
"tst r2, #0xff\n\t"
"itet ne\n\t"
"strneh r2, [ip], #2\n\t"
"streqb r2, [ip]\n\t"
"tstne r2, #0xff00\n\t"
#endif
"bne 5b\n\t"
"RETURN\n"
/* src and dst do not have a common word-alignement. Fall back to
byte copying. */
"4:\n\t"
"ldrb r2, [r1], #1\n\t"
"strb r2, [ip], #1\n\t"
"cmp r2, #0\n\t"
"bne 4b\n\t"
"RETURN"
#elif !defined (__thumb__) || defined (__thumb2__)
"mov r3, r0\n\t"
"1:\n\t"
"ldrb r2, [r1], #1\n\t"
"strb r2, [r3], #1\n\t"
"cmp r2, #0\n\t"
"bne 1b\n\t"
"RETURN"
#else
"mov r3, r0\n\t"
"1:\n\t"
"ldrb r2, [r1]\n\t"
"add r1, r1, #1\n\t"
"strb r2, [r3]\n\t"
"add r3, r3, #1\n\t"
"cmp r2, #0\n\t"
"bne 1b\n\t"
"RETURN"
#endif
);
}