Re: [PATCH v2 0/8] riscv: Support compiling the kernel with more extensions

[Charlie Jenkins <charlie@rivosinc.com>]

On Fri, May 10, 2024 at 09:25:37AM +0100, Conor Dooley wrote:
> On Thu, May 09, 2024 at 03:55:12PM -0700, Charlie Jenkins wrote:
> > On Thu, May 09, 2024 at 11:08:34PM +0100, Conor Dooley wrote:
> 
> > > Maybe if you read what I wrote you'd see what I was getting at, or maybe
> > > not as I might not have been sufficiently clear. I'm not saying that this
> > > particular optimisation is not worth having, but rather than I wanted to
> > 
> > I seem to frequently give you the impression that I don't read what you
> > say before responding.
> 
> Does it happen frequently? I don't really recall it annoying me before.
> 
> > What we each view as "important" in the kernel is
> > different so we come from different places when approaching a problem. I
> > respond in the way that I do not because I am not listening to you, but
> > simply because I have a different opinion and I am not explaining that
> > properly.
> 
> If you're trying to describe a different opinion, responding to the bit
> I was talking about, as you do below in your latest response is ideal.
> Responding inline but not actually addressing the points I was making
> did make me think you were [un]intentionally ignoring what I was trying
> to say.
> 
> > > see why this particular optimisation was worth maintaining 3 code paths
> > 
> > I interpreted the "3 code paths" as with Zbb + 64 bit, with Zbb + 32
> > bit, and without Zbb. I directly responded to that by saying that we
> > could eliminate all of the code paths that are not Zbb + 64 bit could be
> > eliminated. I should have given performance numbers for these alternate
> > cases too, and somebody should have asked. I agree that performance
> > needs justification, and I understand that I did not provide ample
> > justification in this patch. All other architectures optimized these
> > code paths so I figured that was sufficient justification for riscv to
> > do the same, but I understand that it is not.
> 
> And hey, if you look at the commit in question, who acked it? I'm just
> saying that I think we should have a higher standard going forwards.
> 
> > > for but the commit message does not detail any of the benefits, and
> > > looking at the cover I discovered that it was not tested in hardware nor
> > > seemingly with a real test case.
> > 
> > It's hard when riscv currently is very focused on microcontrollers.
> 
> Zbb is actually implemented in hardware, so testing it in the real world
> is not a barrier. Palmer probably has a JH7110 board that someone gave
> to him is not using...
> 
> > These changes are in order to help future hardware that is more focused
> > about performance.
> 
> I'm not replying to most of your response rn, got other stuff to do, but
> what I was trying to say is that I think the point at which optimisations
> for future hardware/extensions should be merged is the point at which
> those extensions are no longer future.
> 
> > That's why I contribute this upstream with the hope
> > that other people, like me, care about performance.
> 
> Heh, that could be read to mean that I do not care about performance,
> which wouldn't be true.
> 
> Cheers,
> Conor.


This is all a good perspective for me to keep in mind when I look into
performance improvements. I will try to get my hands on some hardware
that I can use to test so I can have some real numbers!

- Charlie