[PATCH 1/1] core-kernel: use multiply instead of shifts in hash_64

[PATCH 1/1] core-kernel: use multiply instead of shifts in hash_64

[Andrew Hunter]

hash_64(val) = val * (a 64-bit constant).  It is "optimized" by
replacing the multiply by a bunch of shifts and adds.  On modern
machines, this is not an optimization; remove it.

Running this hash function in a independent benchmark, it's about three times
as fast (1ns vs 3ns) with a multiply as with a shift on Westmere. It's also
considerably smaller (and since we inline this function often, that matters.)

Signed-off-by: Andrew Hunter <ahh@google.com>
Reviewed-by: Paul Turner <pjt@google.com>
---
 include/linux/hash.h |    6 ++++--
 1 files changed, 4 insertions(+), 2 deletions(-)

diff --git a/include/linux/hash.h b/include/linux/hash.h
index b80506b..daabc3d 100644
--- a/include/linux/hash.h
+++ b/include/linux/hash.h
@@ -34,7 +34,9 @@
 static inline u64 hash_64(u64 val, unsigned int bits)
 {
 	u64 hash = val;
-
+#if BITS_PER_LONG == 64
+	hash *= GOLDEN_RATIO_PRIME_64;
+#else
 	/*  Sigh, gcc can't optimise this alone like it does for 32 bits. */
 	u64 n = hash;
 	n <<= 18;
@@ -49,7 +51,7 @@ static inline u64 hash_64(u64 val, unsigned int bits)
 	hash += n;
 	n <<= 2;
 	hash += n;
-
+#endif
 	/* High bits are more random, so use them. */
 	return hash >> (64 - bits);
 }
-- 
1.7.7.3

[PATCH 1/1] core-kernel: use multiply instead of shifts in hash_64

[Andrew Hunter]

hash_64(val) = val * (a 64-bit constant).  It is "optimized" by
replacing the multiply by a bunch of shifts and adds.  On modern
machines, this is not an optimization; remove it.

Running this hash function in a independent benchmark, it's about three times
as fast (1ns vs 3ns) with a multiply as with a shift on Westmere. It's also
considerably smaller (and since we inline this function often, that matters.)

Signed-off-by: Andrew Hunter <ahh@google.com>
Reviewed-by: Paul Turner <pjt@google.com>
---
 include/linux/hash.h |    6 ++++--
 1 files changed, 4 insertions(+), 2 deletions(-)

diff --git a/include/linux/hash.h b/include/linux/hash.h
index b80506b..daabc3d 100644
--- a/include/linux/hash.h
+++ b/include/linux/hash.h
@@ -34,7 +34,9 @@
 static inline u64 hash_64(u64 val, unsigned int bits)
 {
 	u64 hash = val;
-
+#if BITS_PER_LONG == 64
+	hash *= GOLDEN_RATIO_PRIME_64;
+#else
 	/*  Sigh, gcc can't optimise this alone like it does for 32 bits. */
 	u64 n = hash;
 	n <<= 18;
@@ -49,7 +51,7 @@ static inline u64 hash_64(u64 val, unsigned int bits)
 	hash += n;
 	n <<= 2;
 	hash += n;
-
+#endif
 	/* High bits are more random, so use them. */
 	return hash >> (64 - bits);
 }
-- 
1.7.7.3

Re: [PATCH 1/1] core-kernel: use multiply instead of shifts in hash_64

[Michael Tokarev]

On 03.07.2012 00:25, Andrew Hunter wrote:
> diff --git a/include/linux/hash.h b/include/linux/hash.h
> index b80506b..daabc3d 100644
> --- a/include/linux/hash.h
> +++ b/include/linux/hash.h
> @@ -34,7 +34,9 @@
>  static inline u64 hash_64(u64 val, unsigned int bits)
>  {
>  	u64 hash = val;
> -
> +#if BITS_PER_LONG == 64
> +	hash *= GOLDEN_RATIO_PRIME_64;
> +#else
>  	/*  Sigh, gcc can't optimise this alone like it does for 32 bits. */

Hmm.  Does this comment make sense here now?

Thanks,

/mjt

Re: [PATCH 1/1] core-kernel: use multiply instead of shifts in hash_64

[Andrew Hunter]

On Tue, Jul 10, 2012 at 6:35 AM, Michael Tokarev <mjt@tls.msk.ru> wrote:
> On 03.07.2012 00:25, Andrew Hunter wrote:
>> diff --git a/include/linux/hash.h b/include/linux/hash.h
>> index b80506b..daabc3d 100644
>> --- a/include/linux/hash.h
>> +++ b/include/linux/hash.h
>> @@ -34,7 +34,9 @@
>>  static inline u64 hash_64(u64 val, unsigned int bits)
>>  {
>>       u64 hash = val;
>> -
>> +#if BITS_PER_LONG == 64
>> +     hash *= GOLDEN_RATIO_PRIME_64;
>> +#else
>>       /*  Sigh, gcc can't optimise this alone like it does for 32 bits. */
>
> Hmm.  Does this comment make sense here now?
>

I haven't checked what output gcc provides for 32-bit kernels with
this or a literal multiply.  It's not even clear what optimization is
_asked_ for here (possibly the reduction of strength that we probably
don't even want.)