Re: [RFC PATCH 0/6] Deduplicating RISCV cmpxchg.h macros

[Conor Dooley]

On Wed, Mar 22, 2023 at 12:42:56AM -0300, Leonardo Bras Soares Passos wrote:
> On Tue, Mar 21, 2023 at 4:48 PM Conor Dooley <conor@xxxxxxxxxx> wrote:
> > On Tue, Mar 21, 2023 at 02:01:59PM -0300, Leonardo Bras Soares Passos wrote:
> > > On Tue, Mar 21, 2023 at 4:49 AM Conor Dooley <conor.dooley@xxxxxxxxxxxxx> wrote:
> > > > On Tue, Mar 21, 2023 at 03:30:35AM -0300, Leonardo Brás wrote:

> > > > > For patch 5/6 it actually adds many more lines, but tracking (some of) the
> > > > > errors gave me no idea why.
> > > >
> > > > Probably just sparse being unhappy with some way the macros were
> > > > changed - but some of it ("Should it be static?" bits) look very much
> > > > like the patch causing things to be rebuilt only for the "after the
> > > > patch" build, but somehow not for the "before" build.
> > >
> > > Humm, not sure I could understand that last part:
> > > What I get is that the patch 5/6 is causing things to be rebuilt, and
> > > it was not like that on 1-4/6.
> > > Is that what you said?
> > > If so, and you are doing it as an incremental build, changing the
> > > macros in 5/6 should be triggering rebuilds, but it does not make
> > > sense to not be rebuilt in 1-4/6 as they change the same macros.
> >
> > Right, it is an incremental build.
> > First it builds the tree with a patch applied, then it checks out HEAD~1
> > and builds that tree. Then it goes back to the tree with the patch
> > applied and builds it again. The output from builds 2 & 3 are compared
> > to see if any errors snuck in.
> > In theory, that should ensure that, as builds 2 & 3 have had the same
> > diff to the previous one just in opposite directions, the same files get
> > rebuilt - but I am a little worried that ccache gets involved sometimes
> > and leads to the before/after builds not being exactly the same.
> 
> Makes sense to me.

Oh, the other thing that I didn't notice until now is that, if, as
in your case, HEAD~1 for the patch currently being tested does not build
but the HEAD build does, the HEAD build will have more sparse coverage
than the HEAD~1 one. Cos of that, sparse is likely to find more issues in
the after build, and you end up with +1000 error count like patch 5 in
the v1.
If you look at your most recent version, that doesn't have build issues,
patch 5 gets the all-clear:
https://patchwork.kernel.org/project/linux-riscv/list/?series=732217

Not sure why I didn't notice that before, I completely forgot that 4/6
had build issues and it should've been immediately obvious to me why 5/6
had a bunch of extra sparse warnings...

Attachment: signature.asc
Description: PGP signature