Use specific instructions for moving bytes around in a word. This speeds
things up, and as a side-effect, slightly lowers code size.
ARIA_P3 and ARIA_P1 are now 1 single-cycle instruction each (those
instructions are available in all architecture versions starting from v6-M).
Note: ARIA_P3 was already translated to a single instruction by Clang 3.8 and
armclang 6.5, but not arm-gcc 5.4 nor armcc 5.06.
ARIA_P2 is already efficiently translated to the minimal number of
instruction (1 in ARM mode, 2 in thumb mode) by all tested compilers
Manually compiled and inspected generated code with the following compilers:
arm-gcc 5.4, clang 3.8, armcc 5.06 (with and without --gnu), armclang 6.5.
Size reduction (arm-none-eabi-gcc -march=armv6-m -mthumb -Os): 5288 -> 5044 B
Effect on executing time of self-tests on a few boards:
FRDM-K64F (Cortex-M4): 444 -> 385 us (-13%)
LPC1768 (Cortex-M3): 488 -> 432 us (-11%)
FRDM-KL64Z (Cortex-M0): 1429 -> 1134 us (-20%)
Measured using a config.h with no cipher mode and the following program with
aria.c and aria.h copy-pasted to the online compiler:
#include "mbed.h"
#include "aria.h"
int main() {
Timer t;
t.start();
int ret = mbedtls_aria_self_test(0);
t.stop();
printf("ret = %d; time = %d us\n", ret, t.read_us());
}
(A similar commit for Arm follows.)
Use specific instructions for moving bytes around in a word. This speeds
things up, and as a side-effect, slightly lowers code size.
ARIA_P3 (aka reverse byte order) is now 1 instruction on x86, which speeds up
key schedule. (Clang 3.8 finds this but GCC 5.4 doesn't.)
I couldn't find an Intel equivalent of ARM's ret16 (aka ARIA_P1), so I made it
two instructions, which is still much better than the code generated with
the previous mask-shift-or definition, and speeds up en/decryption. (Neither
Clang 3.8 nor GCC 5.4 find this.)
Before:
O aria.o ins
s 7976 43,865
2 10520 37,631
3 13040 28,146
After:
O aria.o ins
s 7768 33,497
2 9816 28,268
3 11432 20,829
For measurement method, see previous commit:
"aria: turn macro into static inline function"
This decreases the size with -Os by nearly 1k while
not hurting performance too much with -O2 and -O3
Before:
O aria.o ins
s 8784 41,408
2 11112 37,001
3 13096 27,438
After:
O aria.o ins
s 7976 43,865
2 10520 37,631
3 13040 28,146
(See previous commit for measurement details.)
Besides documenting types better and so on, this give the compiler more room
to optimise either for size or performance.
Here are some before/after measurements of:
- size of aria.o in bytes (less is better)
- instruction count for the selftest function (less is better)
with various -O flags.
Before:
O aria.o ins
s 10896 37,256
2 11176 37,199
3 12248 27,752
After:
O aria.o ins
s 8784 41,408
2 11112 37,001
3 13096 27,438
The new version allows the compiler to reach smaller size with -Os while
maintaining (actually slightly improving) performance with -O2 and -O3.
Measurements were done on x86_64 (but since this is mainly about inlining
code, this should transpose well to other platforms) using the following
helper program and script, after disabling CBC, CFB and CTR in config.h, in
order to focus on the core functions.
==> st.c <==
#include "mbedtls/aria.h"
int main( void ) {
return mbedtls_aria_self_test( 0 );
}
==> p.sh <==
#!/bin/sh
set -eu
ccount () {
(
valgrind --tool=callgrind --dump-line=no --callgrind-out-file=/dev/null --collect-atstart=no --toggle-collect=main $1
) 2>&1 | sed -n -e 's/.*refs: *\([0-9,]*\)/\1/p'
}
printf "O\taria.o\tins\n"
for O in s 2 3; do
GCC="gcc -Wall -Wextra -Werror -Iinclude"
$GCC -O$O -c library/aria.c
$GCC -O1 st.c aria.o -o st
./st
SIZE=$( du -b aria.o | cut -f1 )
INS=$( ccount ./st )
printf "$O\t$SIZE\t$INS\n"
done
We're not absolutely consistent in the rest of the library, but we tend to use
C99-style comments less often.
Change to use C89-style comments everywhere except for end-of-line comments
