Re: [Qemu-devel] [PATCH v4] buildsys: Move crypto cflags/libs to per object variables

[Fam Zheng <famz@redhat.com>]

On Fri, 09/08 11:36, Daniel P. Berrange wrote:
> On Fri, Sep 08, 2017 at 06:27:01PM +0800, Fam Zheng wrote:
> > On Fri, 09/08 11:05, Daniel P. Berrange wrote:
> > > On Wed, Sep 06, 2017 at 08:49:00PM +0800, Fam Zheng wrote:
> > > > This patch groups the crypto objects into a few .mo objects based on
> > > > functional submodules, and moves inclusion conditions to *-objs
> > > > variables, then moves the global cflags/libs to the *-cflags and *-libs
> > > > variables.
> > > > 
> > > > For init.o and cipher.o, which may or may not need the library flags
> > > > depending on config, adding flags and libs unconditionally doesn't hurt,
> > > > because if the library is not available, the variables are empty.  This
> > > > makes less code.
> > > > 
> > > > Signed-off-by: Fam Zheng <famz@redhat.com>
> > > > 
> > > > ---
> > > > 
> > > > v4: Merge into one patch which is supposedly easier to manage and
> > > > review, and use .mo appraoch to avoid $(foreach) and $(eval) magics.
> > > 
> > > I don't think using  .mo is suitable here. You've used it as a generic
> > > mechanism for grouping .o files, but that is not what it does. There
> > > are special semantics around .mo rules that affect how the final
> > > binaries are linked.
> > 
> > Using .mo is okay here, but after a hindsight I think grouping by library
> > (nettle.mo, gcrypt.mo, etc.) is better than grouping by functionality, for
> > modularization in the future. But that also means assigning the cflags/libs
> > variable cannot be simplified like this.
> > 
> > > 
> > > eg looking back at the description of .mo files 
> > > 
> > > [quote]
> > > commit c261d774fb9093d00e0938a19f502fb220f62718
> > > Author: Fam Zheng <famz@redhat.com>
> > > Date:   Mon Sep 1 18:35:10 2014 +0800
> > > 
> > > [...snip...]
> > > 
> > >     3) When linking an executable, those .mo files in its "-y" variables are
> > >        filtered out, and replaced by one or more -Wl,-u,$symbol flags. This
> > >        is done in the added macro "process-archive-undefs".
> > >     
> > >        These "-Wl,-u,$symbol" flags will force ld to pull in the function
> > >        definition from the archives when linking.
> > >     
> > >        Note that the .mo objects, that are actually meant to be linked in
> > >        the executables, are already expanded in unnest-vars, before the
> > >        linking command. So we are safe to simply filter out .mo for the
> > >        purpose of pulling undefined symbols.
> > >     
> > >        process-archive-undefs works as this: For each ".mo", find all the
> > >        undefined symbols in it, filter ones that are defined in the
> > >        archives. For each of these symbols, generate a "-Wl,-u,$symbol" in
> > >        the link command, and put them before archive names in the command
> > >        line.
> > > [/quote]
> > > 
> > > Based on this, I don't think I can ack this patch, because it can
> > > have unexpected consequences.
> > 
> > This described the process-archive-undefs semantics of .mo, but not the essence
> > of it.  Basically .mo is just partial linking with the additional services of
> > -cflags, -libs and the above -Wl,-u thing. I cannot think of any unexpected
> > consequences with this change. We've had sdl.mo in ui/Makefile.objs for long,
> > just for the same purpose of this patch, with no problem.
> 
> While I'm in favour of moving the linker/compiler flags out of the global
> vars, I'm not convinced this impl is a step forward.
> 
> We already have a mechanism for grouping object files - the 'NNNN-obj-y'
> variables we use throughout our Makefiles.
> 
> This patch is adding a second level of grouping purely to work around the
> fact that we can't set linker/compiler flags on the NNN-obj-y variables
> we use. I think this second level of grouping makes the makefiles more
> complex than they ought to be.

Not quite, it is actually a required step to modularization, which I'm inclined
to get my hands on next. That is also why .mo was introduced.

> 
> IOW, I'd rather see the rules fixed so that we can set variables against
> the existing grouping we have. eg
> 
>    crypto-obj-y-cflags := ...
>    crypto-obj-y-libs := ...
> 
> so we avoid having to introduce second level groups every time we want
> to set these cflags/libs.

This is certainly true, but taking the modularization work into account, .mo
based -cflags and -libs are more natural and consistent. IMO we already have the
latter, so other mechanisms are not really necessary. Remember how complex the
general unnest-vars code is?  I believe adding support to crypto-obj-y-cflags is
more complex than (ab)using .mo objects, even if just for flags/libs
localization.

If you don't like introducing {nettle,gcrypt,gnutls}.mo for now, we can probably
defer it to the time when crypto subsystem is modularized.

Either way let's drop this patch for now.

Fam