arm-trusted-firmware/lib/libc/aarch64/memset.S

65 lines
1.6 KiB
ArmAsm
Raw Normal View History

plat/arm: Introduce and use libc_asm.mk makefile Trace analysis of FVP_Base_AEMv8A 0.0/6063 model running in Aarch32 mode with the build options listed below: TRUSTED_BOARD_BOOT=1 GENERATE_COT=1 ARM_ROTPK_LOCATION=devel_ecdsa KEY_ALG=ecdsa ROT_KEY=plat/arm/board/common/rotpk/arm_rotprivk_ecdsa.pem shows that when auth_signature() gets called 71.99% of CPU execution time is spent in memset() function written in C using single byte write operations, see lib\libc\memset.c. This patch introduces new libc_asm.mk makefile which replaces C memset() implementation with assembler version giving the following results: - for Aarch32 in auth_signature() call memset() CPU time reduced to 20.56%. The number of CPU instructions (Inst) executed during TF-A boot stage before start of BL33 in RELEASE builds for different versions is presented in the tables below, where: - C TF-A: existing TF-A C code; - C musl: "lightweight code" C "implementation of the standard library for Linux-based systems" https://git.musl-libc.org/cgit/musl/tree/src/string/memset.c - Asm Opt: assemler version from "Arm Optimized Routines" project https://github.com/ARM-software/optimized-routines/blob/ master/string/arm/memset.S - Asm Linux: assembler version from Linux kernel https://github.com/torvalds/linux/blob/master/arch/arm/lib/memset.S - Asm TF-A: assembler version from this patch Aarch32: +-----------+------+------+--------------+----------+ | Variant | Set | Size | Inst | Ratio | +-----------+------+------+--------------+----------+ | C TF-A | T32 | 16 | 2122110003 | 1.000000 | | C musl | T32 | 156 | 1643917668 | 0.774662 | | Asm Opt | T32 | 84 | 1604810003 | 0.756233 | | Asm Linux | A32 | 168 | 1566255018 | 0.738065 | | Asm TF-A | A32 | 160 | 1525865101 | 0.719032 | +-----------+------+------+--------------+----------+ AArch64: +-----------+------+------------+----------+ | Variant | Size | Inst | Ratio | +-----------+------+------------+----------+ | C TF-A | 28 | 2732497518 | 1.000000 | | C musl | 212 | 1802999999 | 0.659836 | | Asm TF-A | 140 | 1680260003 | 0.614917 | +-----------+------+------------+----------+ This patch modifies 'plat\arm\common\arm_common.mk' by overriding libc.mk makefile with libc_asm.mk and does not effect other platforms. Change-Id: Ie89dd0b74ba1079420733a0d76b7366ad0157c2e Signed-off-by: Alexei Fedorov <Alexei.Fedorov@arm.com>
2020-09-01 15:38:32 +01:00
/*
* Copyright (c) 2020, Arm Limited. All rights reserved.
*
* SPDX-License-Identifier: BSD-3-Clause
*/
#include <asm_macros.S>
.global memset
/* -----------------------------------------------------------------------
* void *memset(void *dst, int val, size_t count)
*
* Copy the value of 'val' (converted to an unsigned char) into
* each of the first 'count' characters of the object pointed to by 'dst'.
*
* Returns the value of 'dst'.
* -----------------------------------------------------------------------
*/
func memset
cbz x2, exit /* exit if 'count' = 0 */
mov x3, x0 /* keep x0 */
tst x0, #7
b.eq aligned /* 8-bytes aligned */
/* Unaligned 'dst' */
unaligned:
strb w1, [x3], #1
subs x2, x2, #1
b.eq exit /* exit if 0 */
tst x3, #7
b.ne unaligned /* continue while unaligned */
/* 8-bytes aligned */
aligned:cbz x1, x1_zero
bfi w1, w1, #8, #8 /* propagate 'val' */
bfi w1, w1, #16, #16
bfi x1, x1, #32, #32
x1_zero:ands x4, x2, #~0x3f
b.eq less_64
write_64:
.rept 4
stp x1, x1, [x3], #16 /* write 64 bytes in a loop */
.endr
subs x4, x4, #64
b.ne write_64
less_64:tbz w2, #5, less_32 /* < 32 bytes */
stp x1, x1, [x3], #16 /* write 32 bytes */
stp x1, x1, [x3], #16
less_32:tbz w2, #4, less_16 /* < 16 bytes */
stp x1, x1, [x3], #16 /* write 16 bytes */
less_16:tbz w2, #3, less_8 /* < 8 bytes */
str x1, [x3], #8 /* write 8 bytes */
less_8: tbz w2, #2, less_4 /* < 4 bytes */
str w1, [x3], #4 /* write 4 bytes */
less_4: tbz w2, #1, less_2 /* < 2 bytes */
strh w1, [x3], #2 /* write 2 bytes */
less_2: tbz w2, #0, exit
strb w1, [x3] /* write 1 byte */
exit: ret
endfunc memset