Remove useless test and initialisation in name to hash computation

Remove useless test and initialisation in name to hash computation

[Damien Guibouret]

Hello,

I think there is a minor improvment that could be performed on name to hash 
computation in hash.c. Before first byte modulo 4 the computed value is 
reinitialised, but it is already correctly initialised before loop and after 
having processed last byte modulo 4. So the test and initialisation seems useless.

For the kernel, this lead to following change (sorry I do not have a git version 
of it, so it is a simple diff):
---------------------------------------------------------------
--- fs/ext4/hash.c.orig 2017-07-24 20:41:53.000000000 +0200
+++ fs/ext4/hash.c      2017-07-24 20:42:23.000000000 +0200
@@ -79,8 +79,6 @@ static void str2hashbuf_signed(const cha
         if (len > num*4)
                 len = num * 4;
         for (i = 0; i < len; i++) {
-               if ((i % 4) == 0)
-                       val = pad;
                 val = ((int) scp[i]) + (val << 8);
                 if ((i % 4) == 3) {
                         *buf++ = val;
@@ -107,8 +105,6 @@ static void str2hashbuf_unsigned(const c
         if (len > num*4)
                 len = num * 4;
         for (i = 0; i < len; i++) {
-               if ((i % 4) == 0)
-                       val = pad;
                 val = ((int) ucp[i]) + (val << 8);
                 if ((i % 4) == 3) {
                         *buf++ = val;
---------------------------------------------------------------

For e2fsprogs, the 2 functions are combined in one, so there is only one change:
---------------------------------------------------------------
diff --git a/lib/ext2fs/dirhash.c b/lib/ext2fs/dirhash.c
index c4ac94e..4ba3f35 100644
--- a/lib/ext2fs/dirhash.c
+++ b/lib/ext2fs/dirhash.c
@@ -154,8 +154,6 @@ static void str2hashbuf(const char *msg, int len, __u32 
*buf, int num,
      if (len > num*4)
          len = num * 4;
      for (i=0; i < len; i++) {
-        if ((i % 4) == 0)
-            val = pad;
          if (unsigned_flag)
              c = (int) ucp[i];
          else
---------------------------------------------------------------

Regards,

Damien

Re: Remove useless test and initialisation in name to hash computation

[Lukas Czerner]

On Tue, Jul 25, 2017 at 08:40:03PM +0200, Damien Guibouret wrote:
> Hello,
> 
> I think there is a minor improvment that could be performed on name to hash
> computation in hash.c. Before first byte modulo 4 the computed value is
> reinitialised, but it is already correctly initialised before loop and after
> having processed last byte modulo 4. So the test and initialisation seems
> useless.
> 
> For the kernel, this lead to following change (sorry I do not have a git
> version of it, so it is a simple diff):
> ---------------------------------------------------------------
> --- fs/ext4/hash.c.orig 2017-07-24 20:41:53.000000000 +0200
> +++ fs/ext4/hash.c      2017-07-24 20:42:23.000000000 +0200
> @@ -79,8 +79,6 @@ static void str2hashbuf_signed(const cha
>         if (len > num*4)
>                 len = num * 4;
>         for (i = 0; i < len; i++) {
> -               if ((i % 4) == 0)
> -                       val = pad;
>                 val = ((int) scp[i]) + (val << 8);
>                 if ((i % 4) == 3) {
>                         *buf++ = val;
> @@ -107,8 +105,6 @@ static void str2hashbuf_unsigned(const c
>         if (len > num*4)
>                 len = num * 4;
>         for (i = 0; i < len; i++) {
> -               if ((i % 4) == 0)
> -                       val = pad;
>                 val = ((int) ucp[i]) + (val << 8);
>                 if ((i % 4) == 3) {
>                         *buf++ = val;
> ---------------------------------------------------------------
> 
> For e2fsprogs, the 2 functions are combined in one, so there is only one change:
> ---------------------------------------------------------------
> diff --git a/lib/ext2fs/dirhash.c b/lib/ext2fs/dirhash.c
> index c4ac94e..4ba3f35 100644
> --- a/lib/ext2fs/dirhash.c
> +++ b/lib/ext2fs/dirhash.c
> @@ -154,8 +154,6 @@ static void str2hashbuf(const char *msg, int len, __u32
> *buf, int num,
>      if (len > num*4)
>          len = num * 4;
>      for (i=0; i < len; i++) {
> -        if ((i % 4) == 0)
> -            val = pad;
>          if (unsigned_flag)
>              c = (int) ucp[i];
>          else
> ---------------------------------------------------------------

Hi Damien,

the change looks ok to me, have you done any testing to quantify the
improvement, or validate the change ?

>From my limited testing the improvement seems to be around 11% for me
which would be nice to have. Also my test on 466k strings looks ok
as well.

Are you willing to send out properly formated patch for kernel and
e2fsprogs ?

Thanks!
-Lukas

> 
> Regards,
> 
> Damien

Re: Remove useless test and initialisation in name to hash computation

[Damien Guibouret]

Lukas Czerner wrote:
> On Tue, Jul 25, 2017 at 08:40:03PM +0200, Damien Guibouret wrote:
> 
>>Hello,
>>
>>I think there is a minor improvment that could be performed on name to hash
>>computation in hash.c. Before first byte modulo 4 the computed value is
>>reinitialised, but it is already correctly initialised before loop and after
>>having processed last byte modulo 4. So the test and initialisation seems
>>useless.
>>
>>For the kernel, this lead to following change (sorry I do not have a git
>>version of it, so it is a simple diff):
>>---------------------------------------------------------------
>>--- fs/ext4/hash.c.orig 2017-07-24 20:41:53.000000000 +0200
>>+++ fs/ext4/hash.c      2017-07-24 20:42:23.000000000 +0200
>>@@ -79,8 +79,6 @@ static void str2hashbuf_signed(const cha
>>        if (len > num*4)
>>                len = num * 4;
>>        for (i = 0; i < len; i++) {
>>-               if ((i % 4) == 0)
>>-                       val = pad;
>>                val = ((int) scp[i]) + (val << 8);
>>                if ((i % 4) == 3) {
>>                        *buf++ = val;
>>@@ -107,8 +105,6 @@ static void str2hashbuf_unsigned(const c
>>        if (len > num*4)
>>                len = num * 4;
>>        for (i = 0; i < len; i++) {
>>-               if ((i % 4) == 0)
>>-                       val = pad;
>>                val = ((int) ucp[i]) + (val << 8);
>>                if ((i % 4) == 3) {
>>                        *buf++ = val;
>>---------------------------------------------------------------
>>
>>For e2fsprogs, the 2 functions are combined in one, so there is only one change:
>>---------------------------------------------------------------
>>diff --git a/lib/ext2fs/dirhash.c b/lib/ext2fs/dirhash.c
>>index c4ac94e..4ba3f35 100644
>>--- a/lib/ext2fs/dirhash.c
>>+++ b/lib/ext2fs/dirhash.c
>>@@ -154,8 +154,6 @@ static void str2hashbuf(const char *msg, int len, __u32
>>*buf, int num,
>>     if (len > num*4)
>>         len = num * 4;
>>     for (i=0; i < len; i++) {
>>-        if ((i % 4) == 0)
>>-            val = pad;
>>         if (unsigned_flag)
>>             c = (int) ucp[i];
>>         else
>>---------------------------------------------------------------
> 
> 
> Hi Damien,
> 
> the change looks ok to me, have you done any testing to quantify the
> improvement, or validate the change ?
> 
>>From my limited testing the improvement seems to be around 11% for me
> which would be nice to have. Also my test on 466k strings looks ok
> as well.
> 
> Are you willing to send out properly formated patch for kernel and
> e2fsprogs ?
> 
> Thanks!
> -Lukas
> 
> 
>>Regards,
>>
>>Damien
> 
> 
Hello,

I just called old and new versions of functions with two different inputs (one 
multiple of 4 and one not) and check results were equals, nothing more. So not 
very formal, just a quick check but that allows covering all the code of these 
functions. For timing measurement I did nothing.
I will see to prepare a more official patch this week-end.

Regards,

Damien