From ad47974707da5377bab9011a887ec8e92f2b7ddb Mon Sep 17 00:00:00 2001
From: Christophe Leroy <christophe.leroy@csgroup.eu>
Date: Wed, 5 Jul 2023 10:34:26 +0200
Subject: [PATCH] lzma: Fix decompression speed regression
MIME-Version: 1.0
Content-Type: text/plain; charset=utf8
Content-Transfer-Encoding: 8bit

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 29caf9305b6 ("cyclic: Use schedule()
instead of WATCHDOG_RESET()") and commit 26e8ebcd7cb ("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: 29caf9305b6 ("cyclic: Use schedule() instead of WATCHDOG_RESET()")
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Reviewed-by: Simon Glass <sjg@chromium.org>
---
 lib/lzma/LzmaDec.c | 18 ++++--------------
 1 file changed, 4 insertions(+), 14 deletions(-)

diff --git a/lib/lzma/LzmaDec.c b/lib/lzma/LzmaDec.c
index 341149f766..a90b35c6a9 100644
--- a/lib/lzma/LzmaDec.c
+++ b/lib/lzma/LzmaDec.c
@@ -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];
-- 
2.39.5