Re: [PATCH v2] RISC-V: fix Andes errata build issues

[Charlie Jenkins <charlie@rivosinc.com>]

On Wed, May 15, 2024 at 06:30:36PM +0100, Conor Dooley wrote:
> On Wed, May 15, 2024 at 10:18:43AM -0700, Charlie Jenkins wrote:
> > On Wed, May 15, 2024 at 05:56:30PM +0100, Conor Dooley wrote:
> > > On Wed, May 15, 2024 at 09:49:24AM -0700, Charlie Jenkins wrote:
> > > > On Wed, May 15, 2024 at 05:09:34PM +0100, Conor Dooley wrote:
> > > > > From: Conor Dooley <conor.dooley@microchip.com>
> > > > > 
> > > > > Commit e47c37c24024 ("riscv: Introduce vendor variants of extension
> > > > > helpers") added includes for the new vendor_extensions.h header in
> > > > > the T-Head and SiFive errata handling code but didn't do so for Andes,
> > > > > resulting in allmodconfig build issues when commit 589e2fc85850
> > > > > ("riscv: Convert xandespmu to use the vendor extension framework")
> > > > > added a user of a macro defined there.
> > > > > 
> > > > > Fixes: 589e2fc85850 ("riscv: Convert xandespmu to use the vendor extension framework")
> > > > > Signed-off-by: Conor Dooley <conor.dooley@microchip.com>
> > > 
> > > > 
> > > > I was going to fix this in my next version but was waiting for the
> > > > reviews on the thead stuff. I wasn't anticipating these patches to be
> > > > able to jump the queue :)
> > > 
> > > Yah, the reason for that is I asked him to take the non-vector parts of
> > > the series as 6.10 material so that we'd have less stuff movin' around
> > > in cpufeatures.c so that Clement's Zc* + validation changes wouldn't run
> > > into a bunch of conflicts etc. Same reason that I pushed for getting
> > > Andy's vector subset stuff merged today, but that mighta been before you
> > > hopped in.
> > > 
> > > Cheers,
> > > Conor.
> > 
> > Yes I was a couple minutes late to the meeting, whoops.
> 
> 
> It's prob at like 0600 for you, so w/e.
> 
> > The subset of
> > patches that was pulled into for-next is odd to me because there is some
> > of the thead enablement code as part of the vendor extension enablement
> > so that there was a user for it. Since the subset on Palmer's for-next
> > does not have the rest of the thead code there is only a
> > half-implementation of the thead code, it allows the kernel to probe for
> > xtheadvector but it doesn't probe anywhere.
> 
> I dunno, I think that reporting that the extension is there constitutes a
> user, it's not gonna be dead code. There's plenty of extensions for
> which all we do is detect them and nothing more.
> 
> > In my opinion, a better solution would be for me to get rid of the thead
> > code entirely from those patches. So that there is still a user, I can
> > replace the thead code with the andes versions.
> 
> The Andes stuff is in the subset he applied though, so...
> > 
> > Since Palmer already pulled in those changes maybe it's too late. There
> > is not a critical problem here, but it seems like it's bad practice to
> > introduce code without a user.
> 
> ...there is actually a "real" user in xandespmu. I did miss that

I meant there is no user of the xtheadvector addition.

> "riscv: Extend cpufeature.c to detect vendor extensions" actually
> contained the xtheadvector detection though, rather than just the
> infrastructure. I think it is probably harmless to have it, but
> shouldn't be too hard to quickly drop the thead bits either I suppose
> if you're worried about it.

And the adding vlenb to the DT patches is unrelated to the subset of the
series that was pulled into Palmer's for-next so spinning that off into
a different series would be more logical. This is kind of a pointless
rabbit hole I am getting into, but when we start splitting up series
the code contained in the patches start to diverge from the cover
letters that end up in the merge commits.

- Charlie

> 
> Cheers,
> Conor.



_______________________________________________
linux-riscv mailing list
linux-riscv@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-riscv