[LLVMdev] RFC/Heads Up: Deprecating External Build Support

[Daniel Dunbar]

Hi John,

I knew I would hear from you. :)

On Tue, Oct 11, 2011 at 2:38 PM, John Criswell <criswell at illinois.edu> wrote:
> On 10/11/11 3:24 PM, Daniel Dunbar wrote:
>>
>> Hi all,
>>
>> I would like to deprecate the LLVM project's "official" support of
>> setting up the Makefiles / autoconf configurations in such a way that
>> external projects are encouraged to leverage them in their own build.
>>
>> I am mostly referring to the things documented in docs/Projects.html
>> and projects/sample.
>>
>> My justification:
>>
>> (1) I believe few people use these capabilities.
>
> I use these capabilities extensively in numerous research and open-source
> projects.  Those projects include:

As it happens, I knew of two people who used this feature -- myself
(KLEE) and you. :)

> 1) Automatic Pool Allocation (publicly available in the poolalloc module)
> 2) SAFECode (publicly available in the safecode module)
> 3) Giri (publicly available but currently of no real consequence)
> 4) A diagnosis tool developed by one of our graduate students
> 5) An integer overflow detector tool
>
> The first two I care very much about; we are tracking mainline and want to
> keep them working so that we can have something for people to use when LLVM
> 3.0 hits the streets.
>
> Additionally, SAFECode copies Clang into its own source tree (for a variety
> of reasons which I'll skip over for the purposes of this discussion).  The
> project build system actually makes this very easy; very few Clang Makefiles
> needed modification to enable Clang to compile within SAFECode.  If my
> projects have to create their own build system, then I'll have to duplicate
> LLVM's build machinery in the SAFECode project to build Clang.
>
> The other projects I listed are less important; I merely list them so that
> you have an idea of how much work you'll make for me if you remove the
> Projects functionality.
> :)
>
>>
>> (1a) The primary consumer that people have used in the past was
>> llvm-test-suite. That was broken in its own way, and the test suite
>> can now be used independently of an LLVM installation. This makes it
>> much more useful as a production test suite for compilers (vs. a test
>> suite for a research project).
>
>
>>
>> (2) They are fragile and really not a very good way to structure an
>> external project. It is much better for external projects to just
>> structure themselves as any other open source project would and use
>> tools like llvm-config to gather build information.
>
> IMHO, the LLVM Projects setup is a very good way for people to write their
> own LLVM passes without having to patch the LLVM source tree.  It alleviates
> the need to create a whole new build system, it allows one to build LLVM and
> one's project(s) all in a single go, and one's project source code is
> organized just as it would be within LLVM's source base (more or less).

I wouldn't agree it is a "very good" way, but I agree it is a way.

The problem is that the "solution" we are providing for this problem
is too intimately tied to the build system of LLVM itself.

If the community thinks this is an important problem for LLVM to
address, then I think the ideal solution is that we provide a set of
autoconf/makefile/whatever foo that uses llvm-config and so on, but
doesn't try and do things like include the LLVM Makefiles or the LLVM
autoconf files. This would also provide a migration solution for
existing projects using the current functionality.

> I suspect that removing the Project functionality will also increase the
> demand for a switch to git as people will prefer to write their LLVM passes
> as patches to the LLVM source base instead of as a separate project.
>
>>
>> (3) The official support is not particularly well known, and is easy
>> to break. Considering Makefiles and autoconf files as part of a
>> project API is painful, at best, and makes refactoring of project
>> structure difficult.
>
>
> Building the Projects functionality was a pain (I'm the one that originally
> added the support for it, so I know).  If it's getting in the way of
> important changes to the build system, then perhaps it should be removed
> despite the fact that it'll create more work (headaches) for me.
>
> That said, I ask for two things:
>
> 1) Please don't remove functionality like this right before an LLVM release.
>  Both SAFECode and Poolalloc use this feature, and I don't want to have to
> scramble at the last minute to make them work again with the upcoming LLVM
> 3.0.  If it has to go, please wait until development for LLVM 3.1 begins
> (waiting longer would be even better).

I have no intention of removing or breaking it now, I just want to
deprecate it (as in, declare somewhere that we don't officially
support it, and stop publicizing it).

> 2) Please only remove it if it's currently getting in the way of some
> important bug fix, refactoring, or enhancement.  If you're removing it just
> because LLVM Core doesn't need it, then you're not fixing anything in LLVM,
> and you're breaking things that use LLVM.

Right, I completely agree with this, and am completely familiar with
the pain of maintaining an external-to-LLVM client. I'll do my best to
maintain things when/if I ever start actually making changes.

 - Daniel

>
> -- John T.
>
>>
>> Object now, or hold your peace.
>>
>>  - Daniel
>> _______________________________________________
>> LLVM Developers mailing list
>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
>