arm-trusted-firmware

Alexei Fedorov e3f2b1a932 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>	4 years ago
..
memset.S	plat/arm: Introduce and use libc_asm.mk makefile	4 years ago

Alexei Fedorov e3f2b1a932 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>

4 years ago

memset.S

plat/arm: Introduce and use libc_asm.mk makefile

4 years ago