Messages in this thread | | | From | David Laight <> | Subject | RE: x86/csum: Remove unnecessary odd handling | Date | Sun, 7 Jan 2024 12:11:18 +0000 |
| |
From: Eric Dumazet > Sent: 06 January 2024 10:26 ... > On a related note, at least with clang, I found that csum_ipv6_magic() > is needlessly using temporary on-stack storage, > showing a stall on Cascade Lake unless I am patching add32_with_carry() : > > diff --git a/arch/x86/include/asm/checksum_64.h > b/arch/x86/include/asm/checksum_64.h > index 407beebadaf45a748f91a36b78bd1d023449b132..c3d6f47626c70d81f0c2ba401d85050b09a39922 > 100644 > --- a/arch/x86/include/asm/checksum_64.h > +++ b/arch/x86/include/asm/checksum_64.h > @@ -171,7 +171,7 @@ static inline unsigned add32_with_carry(unsigned > a, unsigned b) > asm("addl %2,%0\n\t" > "adcl $0,%0" > : "=r" (a) > - : "0" (a), "rm" (b)); > + : "0" (a), "r" (b)); > return a; > }
Try replacing: return csum_fold( (__force __wsum)add32_with_carry(sum64 & 0xffffffff, sum64>>32)); with: return csum_fold((__force __wsum)((sum64 + ror64(sum64, 32)) >> 32));
Should be less instructions as well. (shift, add, shift v shift, mov, and, add, add) Although both might be 3 clocks.
The best C version of csum_fold (from IIRC arc) is also likely to be better than the x86 asm one - certainly no worse.
David
- Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
| |