Re: Default NO_CLEAN=yes in 15+

On Tue, 23 Jul 2024, John Baldwin wrote:

> On 7/23/24 16:08, Shawn Webb wrote:
>> On Tue, Jul 23, 2024 at 03:58:13PM -0400, John Baldwin wrote:
>>> The buildworld and buildkernel targets include a "clean" step before 
>>> building
>>> objects dating back before my time to 'make world' (I haven't looked to 
>>> see
>>> how far back it goes).  To permit incremental builds, this step can be 
>>> skipped
>>> via NO_CLEAN=yes.  This step is a bit unusual in build systems however. 
>>> Most
>>> build systems have separate commands for building vs cleaning (e.g. 'make 
>>> all'
>>> vs 'make clean') and over time FreeBSD's build system has gained dedicated
>>> clean targets as well (cleanworld and cleankernel).  For myself, I always
>>> use NO_CLEAN=yes when building worlds and kernels.  If I need a clean 
>>> build
>>> I use the dedicated clean targets (e.g. cleanworld) first.  In particular,
>>> cleanworld/cleankernel are far more efficient since they use a single
>>> recursive 'rm' whereas the "clean" step involves a full tree walk with
>>> nested make invocations of the 'cleandir' target.
>>> 
>>> A few years ago, Ed Maste added a MK_CLEAN option to src.opts.mk to as a
>>> WITH/WITHOUT knob for the "clean" step similar to NO_CLEAN=yes.  To 
>>> preserve
>>> existing behavior this knob currently defaults to on, but I know Ed's goal
>>> was to eventually flip the default so that NO_CLEAN builds would be the
>>> default.  I would like us to do that starting in 15.
>> 
>> It would make sense to me to default MK_CLEAN=no in release branches.
>> Perhaps stable branches, too. While I don't hold a strong opinion on
>> the matter, I would prefer MK_CLEAN=yes to remain the default on the
>> main branch.
>> 
>> I can't give tangible examples, but I remember running into weird
>> issues occasionally when using `make buildworld WITHOUT_CLEAN=yes` in
>> main. I probably should do a better job at documenting those
>> (infrequent) issues when they arise.
>
> To be clear, the suggestion is that when you hit an issue, just run
> 'make cleanworld' rather than relying on the omission of NO_CLEAN=yes
> to do this for you (and much slower at that).  Have you used any other
> build systems where 'make' does an implicit 'make clean' before it
> builds?  I have never encountered another where this is true.  'clean' is
> always a separate target from 'all' in my experience.

Can we (if not in 15 but then in 16) simply:

(1) document clean builds as make cleanblah .. && make blah
(2) document that if we do not run make cleanblah we'll get an
     incremental build
(3) if an incremental build fails (CI will detect as well) -- do as Ed
     said
(4) make sure re and so are aware of the extra make cleanblah needed?
(5) remove NO_CLEAN and MK_CLEAN entirely (there's no point in flipping
     the default if you can choose to run a make step or not)?

I think that is basically your suggestion?

+101 for that from here.

/bz

-- 
Bjoern A. Zeeb                                                     r15:7