Re: [dpdk-dev] [PATCH v2 2/2] build: support building ABI versioned files twice

[Bruce Richardson <bruce.richardson@intel.com>]

On Tue, Oct 01, 2019 at 05:53:05PM +0100, Bruce Richardson wrote:
> On Tue, Oct 01, 2019 at 03:23:47PM +0200, Andrzej Ostruszka wrote:
> > Thanks Bruce for the patch.  I like the idea of splitting versioning out
> > of rte_compat.h, but I have some comments.
> > 
> > On 9/27/19 10:59 PM, Bruce Richardson wrote:
> > [...]
> > > --- a/config/common_base
> > > +++ b/config/common_base
> > > @@ -111,6 +111,7 @@ CONFIG_RTE_MAX_VFIO_CONTAINERS=64
> > >  CONFIG_RTE_MALLOC_DEBUG=n
> > >  CONFIG_RTE_EAL_NUMA_AWARE_HUGEPAGES=n
> > >  CONFIG_RTE_USE_LIBBSD=n
> > > +CONFIG_RTE_USE_FUNCTION_VERSIONING=y
> > 
> > I'm not fond of this config option - it is not really an option to be
> > changed by the user.  I would prefer to just add flag to CFLAGS in
> > mk/target/generic/rte.vars.mk.
> > 
> 
> Ok, that sounds reasonable enough.

Done in V3.

> 
> > >  #
> > >  # Recognize/ignore the AVX/AVX512 CPU flags for performance/power testing.
> > > diff --git a/config/rte_config.h b/config/rte_config.h
> > > index 0bbbe274f..b63a2fdea 100644
> > > --- a/config/rte_config.h
> > > +++ b/config/rte_config.h
> > > @@ -31,9 +31,6 @@
> > >  
> > >  /****** library defines ********/
> > >  
> > > -/* compat defines */
> > > -#define RTE_BUILD_SHARED_LIB
> > > -
> > 
> > So now everything builds "as static lib" (but with "-fPIC") apart from
> > those libraries that use symbol versioning.  I'm OK with that however
> > I'd like to note that code might be using RTE_BUILD_SHARED_LIB and do
> > different things e.g. app/test-bbdev/test_bbdev_perf.c.  I know that was
> > already the case - just wanted to say that aloud to make sure we are all
> > aware of this :).
> 
> Thanks for pointing this out, I wasn't aware of it! Doing a git grep this
> seems to be the only place in a C file (other than rte_config.h and
> rte_compat.h) where the SHARED_LIB flag is being checked. I'll need to
> follow up on that to see what the logic is there, because it seems strange
> to require such a check.
> 

This #ifdef can be removed, see patchset: 
http://patches.dpdk.org/project/dpdk/list/?series=6699

This then leaves function versioning as the only use-case where we need
different code paths for static vs shared builds.

> > 
> > > diff --git a/doc/guides/contributing/coding_style.rst b/doc/guides/contributing/coding_style.rst
> > > index 449b33494..e95a1a2be 100644
> > > --- a/doc/guides/contributing/coding_style.rst
> > > +++ b/doc/guides/contributing/coding_style.rst
> > > @@ -948,6 +948,13 @@ reason
> > >  	built. For missing dependencies this should be of the form
> > >  	``'missing dependency, "libname"'``.
> > >  
> > > +use_function_versioning
> > > +	**Default Value = false**.
> > > +	Specifies if the library in question has ABI versioned functions. If it
> > > +	has, this value should be set to ensure that the C files are compiled
> > > +	twice with suitable parameters for each of shared or static library
> > > +	builds.
> > > +
> > 
> > Maybe a different name for this option?  In general an "ideal
> > theoretical" solution would be for build system to figure out on its own
> > that separate build is necessary automatically - but that might incur
> > some performance penalty (additional grep'ing of sources or so). 
> 
> I was thinking about that, and how we can do it automatically. The trouble
> is that for correctness we would need to recheck every file after it had
> changed, and since the result of the check means that we have different
> build steps it would basically mean doing a full reconfiguration for every
> file change. That's not really practical, hence this proposed solution.
> 

I've not made any changes here for the V3. However, if we want to reduce
the number of changes required, we could always switch to having the
rte_function_versioning.h header file included on the basis of the flag in
the meson.build file. Having the C flag compile vary based on the meson one
is normal, having the inverse is problematic because of what I explained
above - you'd basically need to reconfigure to check after each file
change.

Personally, I don't think changing things to auto-include the header is
worth it, hence the fact of no-change in v3.

Regards,
/Bruce