lzma: Fix decompression speed regression

Uncompressing a 1.7Mbytes FIT image on U-boot 2023.04 takes
approx 7s on a powerpc 8xx.
The same on U-boot 2023.07-rc6 takes approx 28s unless watchdog
is disabled.

During that decompression, LzmaDec_DecodeReal() calls schedule
1.6 million times, that is every 4µs in average.

In the past it used to be a call to WATCHDOG_RESET() which was
just calling hw_watchdog_reset().

But the combination of commit 29caf9305b ("cyclic: Use schedule()
instead of WATCHDOG_RESET()") and commit 26e8ebcd7c ("watchdog:
mpc8xxx: Make it generic") results in an heavier processing.

However, there is absolutely no point in calling schedule() that
often.

By moving and keeping only one call to schedule() in the main
loop the number of calls is reduced to 1.2 million which is still
too much. So add logic to only call schedule every 1024 times.
That leads to a call to schedule approx every 6ms which is still
far enough to entertain the watchdog which has a 1s timeout on
powerpc 8xx.

powerpc 8xx being one of the slowest targets we have today in
U-boot, and most other watchdogs having a timeout of one minutes
instead of one second like the 8xx, this fix should not have
negative impact on other targets.

Fixes: 29caf9305b ("cyclic: Use schedule() instead of WATCHDOG_RESET()")
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Reviewed-by: Simon Glass <sjg@chromium.org>
This commit is contained in:
Christophe Leroy 2023-07-05 10:34:26 +02:00 committed by Tom Rini
parent 820801eacc
commit ad47974707

View file

@ -152,8 +152,7 @@ static int MY_FAST_CALL LzmaDec_DecodeReal(CLzmaDec *p, SizeT limit, const Byte
const Byte *buf = p->buf;
UInt32 range = p->range;
UInt32 code = p->code;
schedule();
unsigned int loop = 0;
do
{
@ -162,6 +161,9 @@ static int MY_FAST_CALL LzmaDec_DecodeReal(CLzmaDec *p, SizeT limit, const Byte
unsigned ttt;
unsigned posState = processedPos & pbMask;
if (!(loop++ & 1023))
schedule();
prob = probs + IsMatch + (state << kNumPosBitsMax) + posState;
IF_BIT_0(prob)
{
@ -177,8 +179,6 @@ static int MY_FAST_CALL LzmaDec_DecodeReal(CLzmaDec *p, SizeT limit, const Byte
state -= (state < 4) ? state : 3;
symbol = 1;
schedule();
do { GET_BIT(prob + symbol, symbol) } while (symbol < 0x100);
}
else
@ -188,8 +188,6 @@ static int MY_FAST_CALL LzmaDec_DecodeReal(CLzmaDec *p, SizeT limit, const Byte
state -= (state < 10) ? 3 : 6;
symbol = 1;
schedule();
do
{
unsigned bit;
@ -321,8 +319,6 @@ static int MY_FAST_CALL LzmaDec_DecodeReal(CLzmaDec *p, SizeT limit, const Byte
UInt32 mask = 1;
unsigned i = 1;
schedule();
do
{
GET_BIT2(prob + i, i, ; , distance |= mask);
@ -335,8 +331,6 @@ static int MY_FAST_CALL LzmaDec_DecodeReal(CLzmaDec *p, SizeT limit, const Byte
{
numDirectBits -= kNumAlignBits;
schedule();
do
{
NORMALIZE
@ -409,8 +403,6 @@ static int MY_FAST_CALL LzmaDec_DecodeReal(CLzmaDec *p, SizeT limit, const Byte
const Byte *lim = dest + curLen;
dicPos += curLen;
schedule();
do
*(dest) = (Byte)*(dest + src);
while (++dest != lim);
@ -418,8 +410,6 @@ static int MY_FAST_CALL LzmaDec_DecodeReal(CLzmaDec *p, SizeT limit, const Byte
else
{
schedule();
do
{
dic[dicPos++] = dic[pos];