[v8,1/5] asm-generic: Improve csum_fold

Message ID 20231027-optimize_checksum-v8-1-feb7101d128d@rivosinc.com
State New
Headers
Series riscv: Add fine-tuned checksum functions |

Commit Message

Charlie Jenkins Oct. 27, 2023, 10:43 p.m. UTC
  This csum_fold implementation introduced into arch/arc by Vineet Gupta
is better than the default implementation on at least arc, x86, and
riscv. Using GCC trunk and compiling non-inlined version, this
implementation has 41.6667%, 25% fewer instructions on riscv64, x86-64
respectively with -O3 optimization. Most implmentations override this
default in asm, but this should be more performant than all of those
other implementations except for arm which has barrel shifting and
sparc32 which has a carry flag.

Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>
Reviewed-by: David Laight <david.laight@aculab.com>
---
 include/asm-generic/checksum.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
  

Comments

Al Viro Oct. 27, 2023, 11:10 p.m. UTC | #1
On Fri, Oct 27, 2023 at 03:43:51PM -0700, Charlie Jenkins wrote:
>  /*
>   * computes the checksum of a memory block at buff, length len,
>   * and adds in "sum" (32-bit)
> @@ -31,9 +33,7 @@ extern __sum16 ip_fast_csum(const void *iph, unsigned int ihl);
>  static inline __sum16 csum_fold(__wsum csum)
>  {
>  	u32 sum = (__force u32)csum;
> -	sum = (sum & 0xffff) + (sum >> 16);
> -	sum = (sum & 0xffff) + (sum >> 16);
> -	return (__force __sum16)~sum;
> +	return (__force __sum16)((~sum - ror32(sum, 16)) >> 16);
>  }

Will (~(sum + ror32(sum, 16))>>16 produce worse code than that?
Because at least with recent gcc this will generate the exact thing
you get from arm inline asm...
  
Charlie Jenkins Oct. 28, 2023, 12:04 a.m. UTC | #2
On Sat, Oct 28, 2023 at 12:10:36AM +0100, Al Viro wrote:
> On Fri, Oct 27, 2023 at 03:43:51PM -0700, Charlie Jenkins wrote:
> >  /*
> >   * computes the checksum of a memory block at buff, length len,
> >   * and adds in "sum" (32-bit)
> > @@ -31,9 +33,7 @@ extern __sum16 ip_fast_csum(const void *iph, unsigned int ihl);
> >  static inline __sum16 csum_fold(__wsum csum)
> >  {
> >  	u32 sum = (__force u32)csum;
> > -	sum = (sum & 0xffff) + (sum >> 16);
> > -	sum = (sum & 0xffff) + (sum >> 16);
> > -	return (__force __sum16)~sum;
> > +	return (__force __sum16)((~sum - ror32(sum, 16)) >> 16);
> >  }
> 
> Will (~(sum + ror32(sum, 16))>>16 produce worse code than that?
> Because at least with recent gcc this will generate the exact thing
> you get from arm inline asm...

Yes that will produce worse code because an out-of-order processor will be able to
leverage that ~sum and ror32(sum, 16) can be computed independently of
each other. There are more strict data dependencies in (~(sum +
ror32(sum, 16))>>16.

- Charlie
  

Patch

diff --git a/include/asm-generic/checksum.h b/include/asm-generic/checksum.h
index 43e18db89c14..ad928cce268b 100644
--- a/include/asm-generic/checksum.h
+++ b/include/asm-generic/checksum.h
@@ -2,6 +2,8 @@ 
 #ifndef __ASM_GENERIC_CHECKSUM_H
 #define __ASM_GENERIC_CHECKSUM_H
 
+#include <linux/bitops.h>
+
 /*
  * computes the checksum of a memory block at buff, length len,
  * and adds in "sum" (32-bit)
@@ -31,9 +33,7 @@  extern __sum16 ip_fast_csum(const void *iph, unsigned int ihl);
 static inline __sum16 csum_fold(__wsum csum)
 {
 	u32 sum = (__force u32)csum;
-	sum = (sum & 0xffff) + (sum >> 16);
-	sum = (sum & 0xffff) + (sum >> 16);
-	return (__force __sum16)~sum;
+	return (__force __sum16)((~sum - ror32(sum, 16)) >> 16);
 }
 #endif