Re: [PATCH net-next 3/3] eth: bnxt: handle invalid Tx completions more gracefully

[Paolo Abeni <pabeni@redhat.com>]

On Wed, 2023-07-12 at 09:34 -0700, Jakub Kicinski wrote:
> On Wed, 12 Jul 2023 08:50:06 +0200 Paolo Abeni wrote:
> > Surely not a big deal. But some users (possibly most of them!) have
> > older compiler.
> 
> I checked GCC 10 and GCC 9, and the code is the same :(
> Any idea on how old do we need to go?

I guess that would be more then enough!

> > Including an assignment in the test code, I get this
> > additional difference:
> > 
> > -   c:	80 4b 09 01          	orb    $0x1,0x9(%rbx)
> > +   c:	c6 43 09 01          	movb   $0x1,0x9(%rbx)
> > 
> > with the bitfield using the 'or' operation. Not a big deal, but the
> > other option is slightly better ;)
> 
> Is there really any difference whether one changes a byte or ors
> in a bit? Either way it's a partial update of a word.

Really not a big deal, but 'or' fetches memory and then store it, while
move [immediate] is a single store. In case of a cache miss, 'or'
should stall, while 'mov' should not. In general with 'mov' there
should be less pressure on the cache and/or bus.

Cheers,

Paolo