On Fri, 2018-02-02 at 16:50 +0100, Greg KH wrote:
> On Fri, Feb 02, 2018 at 04:37:55PM +0200, Jani Nikula wrote:
> > On Fri, 02 Feb 2018, Greg KH <gre...@linuxfoundation.org> wrote:
> > > On Fri, Feb 02, 2018 at 12:44:38PM +0200, Jani Nikula wrote:
> > >>
> > >> +Knut, Fengguang
> > >>
> > >> On Fri, 02 Feb 2018, Greg KH <gre...@linuxfoundation.org> wrote:
> > >> > - If clang now builds the kernel "cleanly", yes, I want to take
> > >> > warning fixes in the stable tree. And even better yet, if you
> > >> > keep working to ensure the tree is "clean", that would be
> > >> > wonderful.
> > >>
> > >> So we can run sparse using 'make C=1' and friends, or other static
> > >> analysis tools using 'make CHECK=foo C=1', as long as the passed command
> > >> line params work. There was work by Knut to extend this make checker
> > >> stuff . Since mixing different HOSTCC's in a single workdir seems
> > >> like a bad idea, I wonder how hard it would be to make clang work like
> > >> this:
> > >>
> > >> $ make CHECK=clang C=1
> > >>
> > >> Or using Knut's wrapper. Feels like that could increase the use of clang
> > >> for static analysis of patches.
> > >
> > > Why not just build with clang itself:
> > > make CC=clang
> > Same as HOSTCC, mixing different CC's in a single build dir seems like a
> > bad idea. Sure, everyone can setup a separate build dir for clang, but
> > IMHO having 'make CHECK=clang C=1' work has least resistance. YMMV.
> "O=some_output_dir" is your friend. If you aren't doing that already
> for your test builds, you don't know what you are missing :)
I use O= a lot myself - so good not to have all the output files "pollute" the
tree, and to be able to switch branches and compile without having to recompile
by having multiple O= set up.
I think what my runchecks wrapper script brings in addition is the ability to
to a number
of checks which may or may not pass, even return error codes, from the same
and configure what errors to fix now and what to postpone/ignore (and thus not
As an example, I just tried clang (on v4.15-rc6) with:
and it fails to compile for me in arch/x86/xen/mmu_pv.o.
If I'd want to just make sure that some patches did not introduce new errors
I would waste some time with unrelated errors, and there will be noise in the
consuming personal "cycles".
I haven't really looked at the details of much of what clang outputs of errors
yet, but I
can imagine that specific errors reported by clang might be useful to correct
even in old
kernels, where some files inevitably will fail to compile like this.
This would be easy to handle with runchecks using a few exceptions for those
problems/files not yet fixed, allowing a run to easily detect (while compiling
with gcc as
the main compiler) that no new clang errors were introduced of any other kind
> greg k-h