RE: [PATCH] lib/x86: Optimise csum_partial of buffers that are not multiples of 8 bytes.

From: David Laight <redacted>

Sent: 13 December 2021 18:01

Add in the trailing bytes first so that there is no need to worry
about the sum exceeding 64 bits.

This is an alternate version that (mostly) compiles to reasonable code.
I've also booted a kernel with it - networking still works!

https://godbolt.org/z/K6vY31Gqs

I changed the while (len >= 64) loop into an
if (len >= 64) do (...) while(len >= 64) one.
But gcc makes a pigs breakfast of compiling it - it optimises
it so that it is while (ptr < lim) but adds a lot of code.
So I've done that by hand.
Then it still makes a meal of it because it refuses to take
'buff' from the final loop iteration.
An assignment to the limit helps.

Then there is the calculation of (8 - (len & 7)) * 8.
gcc prior to 9.2 just negate (len & 7) then use leal 56(,%rs1,8),%rcx.
But later ones and fail to notice.
Even given (64 + 8 * -(len & 7)) clang fails to use leal.

I'm not even sure the code clang generates is right:
(%rsi is (len & 7))
        movq    -8(%rsi,%rax), %rdx
        leal    (,%rsi,8), %ecx
        andb    $56, %cl
        negb    %cl
        shrq    %cl, %rdx

The 'negb' is the wrong size of the 'andb'.
It might be ok if it is assuming the cpu ignores the high 2 bits of %cl.
But that is a horrid assumption to be making.

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)