[RFC] [PATCH] [timer] Optimise apply_slack() for size and speed.

[RFC] [PATCH] [timer] Optimise apply_slack() for size and speed.

[Chinmay V S]

To apply proper slack, the original algorithm used to prepare a mask and
then apply the mask to obtain the appropriately rounded-off absolute
time the timer expires.

This patch modifies the masking logic to a bit-shift logic, therby
reducing the complexity and number of operations. Thus obtaining a minor
speed-up.

Signed-off-by: Chinmay V S <chinmay.v.s@pathpartnertech.com>
---
 kernel/timer.c |    8 ++------
 1 files changed, 2 insertions(+), 6 deletions(-)

diff --git a/kernel/timer.c b/kernel/timer.c
index a297ffc..0379658 100644
--- a/kernel/timer.c
+++ b/kernel/timer.c
@@ -785,9 +785,7 @@ EXPORT_SYMBOL(mod_timer_pending);
  * Algorithm:
  *   1) calculate the maximum (absolute) time
  *   2) calculate the highest bit where the expires and new max are different
- *   3) use this bit to make a mask
- *   4) use the bitmask to round down the maximum time, so that all last
- *      bits are zeros
+ *   3) round down the maximum time, so that all the lower bits are zeros
  */
 static inline
 unsigned long apply_slack(struct timer_list *timer, unsigned long expires)
@@ -811,9 +809,7 @@ unsigned long apply_slack(struct timer_list *timer, unsigned long expires)
 
 	bit = find_last_bit(&mask, BITS_PER_LONG);
 
-	mask = (1 << bit) - 1;
-
-	expires_limit = expires_limit & ~(mask);
+	expires_limit = (expires_limit >> bit) << bit;
 
 	return expires_limit;
 }
-- 
1.7.5.4

Re: [RFC] [PATCH] [timer] Optimise apply_slack() for size and speed.

[Sebastian Andrzej Siewior]

* Chinmay V S | 2012-01-31 15:07:32 [+0530]:

>This patch modifies the masking logic to a bit-shift logic, therby
>reducing the complexity and number of operations. Thus obtaining a minor
>speed-up.
>diff --git a/kernel/timer.c b/kernel/timer.c
>index a297ffc..0379658 100644
>--- a/kernel/timer.c
>+++ b/kernel/timer.c
>@@ -785,9 +785,7 @@ EXPORT_SYMBOL(mod_timer_pending);
>@@ -811,9 +809,7 @@ unsigned long apply_slack(struct timer_list *timer, unsigned long expires)
> 
> 	bit = find_last_bit(&mask, BITS_PER_LONG);
> 
>-	mask = (1 << bit) - 1;
>-
>-	expires_limit = expires_limit & ~(mask);
>+	expires_limit = (expires_limit >> bit) << bit;

The question is whether a shift left is more efficient compared to an
and operation. Besides that you save atleast one operation because you
don't need to load -1 into a register for creating a mask. So it looks
like less code.

Sebastian

Re: [RFC] [PATCH] [timer] Optimise apply_slack() for size and speed.

[Arjan van de Ven]

On 2/2/2012 10:31 PM, Chinmay V S wrote:
> Here is the dis-assembly (build for x86) :
> 
> Original code snippet :
> mov    %eax,%ecx
> mov    $0x1,%edx
> shl    %cl,%edx
> sub    $0x1,%edx
> mov    %edx,-0x10(%ebp)
> not    %edx
> and    %esi,%edx
> mov    %edx,-0x14(%ebp)
> 
> Patched code snippet :
> mov    %eax,%ecx
> shrl   %cl,-0x14(%ebp)
> shll   %cl,-0x14(%ebp)
> 
> As is evident, we are able to reduce 5 instructions by using a bit-shift
> logic (compared to a masking logic).
> 

this code is so not performance critical to this level, that we should
optimize for readability, not for the output of a compiler.