On Tue, Jun 15, 2021 at 3:18 PM David Laight [off-list ref] wrote:

From: Bin Meng

quoted

Sent: 15 June 2021 14:09

On Tue, Jun 15, 2021 at 4:57 PM David Laight [off-list ref] wrote:

quoted

...

quoted

quoted

I'm surprised that the C loop:

quoted

+             for (; count >= bytes_long; count -= bytes_long)
+                     *d.ulong++ = *s.ulong++;

ends up being faster than the ASM 'read lots' - 'write lots' loop.

I believe that's because the assembly version has some unaligned
access cases, which end up being trap-n-emulated in the OpenSBI
firmware, and that is a big overhead.

Ah, that would make sense since the asm user copy code
was broken for misaligned copies.
I suspect memcpy() was broken the same way.

I'm surprised IP_NET_ALIGN isn't set to 2 to try to
avoid all these misaligned copies in the network stack.
Although avoiding 8n+4 aligned data is rather harder.

That's up to the network driver, indeed I have a patch already for the
BeagleV one:

https://lore.kernel.org/netdev/20210615012107.577ead86@linux.microsoft.com/T/ (local)

Misaligned copies are just best avoided - really even on x86.
The 'real fun' is when the access crosses TLB boundaries.

-- 
per aspera ad upstream