Re: [PATCH for-4.20] automation/gitlab: disable coverage from clang randconfig

[Roger Pau Monné <roger.pau@xxxxxxxxxx>]

On Tue, Jan 14, 2025 at 05:48:20PM +0000, Andrew Cooper wrote:
> On 14/01/2025 5:43 pm, Roger Pau Monne wrote:
> > If randconfig enables coverage support the build times out due to GNU LD
> > taking too long.  For the time being prevent coverage from being enabled in
> > clang randconfig job.
> >
> > Signed-off-by: Roger Pau Monné <roger.pau@xxxxxxxxxx>
> 
> Acked-by: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>

Thanks.

> > ---
> > Cc: Oleksii Kurochko <oleksii.kurochko@xxxxxxxxx>
> > ---
> > I will fix the orphaned section stuff separately, as I'm considering just
> > removing LLVM coverage support because the llvm coverage format is not
> > stable, and the code to dump it has already become stale.  However I need
> > to think about it, and in the short term disabling coverage support from
> > randconfig is more straightforward.
> 
> Oh, so it's broken too?  Unless the fix is trivial, we should have a
> Kconfig level disable on it.  No point letting people turn on something
> that's broken.

It's not build time broken, but the format of the buffer that we
return in llvm.c dump() function has gotten out of sync with the
native format in modern clang versions, and hence coverage tools will
refuse to parse it.

I think newer versions of llvm/clang have now included an internal
function to dump the buffer, so that we no longer have to attempt to
generate it from Xen, I need to see whether that works.  Otherwise the
only remaining option is to convert the llvm data into gcov format
(which is stable), IIRC this is what Linux did.

Thanks, Roger.