[LLVMdev] Sanitizers libs in Compiler-RT

[Chandler Carruth]

On Thu, Jan 30, 2014 at 1:54 PM, Renato Golin <renato.golin at linaro.org>wrote:

> On 30 January 2014 21:50, Reid Kleckner <rnk at google.com> wrote:
>>
>> I don't see any compelling reason to split the sanitizers out today.
>>
>
> Clear, next.
>

Just as a side note, I had some thoughts about the organization of this
stuff a while back that I wanted to replay here.

Fundamentally, I feel like we're in particular getting a few things
conflated. Maybe separating them out helps:

- There is the repository "compiler-rt" that holds all of the runtime
libraries that (if desired) need to be shipped along side the compiler.
They're separated so that they can be omitted when they aren't needed, and
potentially to isolate an unusual desired property: we would really like to
(eventually) build them with the just-built compiler rather than the host
compiler (where possible, clearly this requires the target compiler to be
executable on the host).

- There is the core runtime library. Historically this was called
'compiler-rt' informally, but perhaps better called 'libclang_rt', which
provides the core necessary runtime library facilities to compile C or C++
applications. It's analogous to libgcc but without some of the unwinding
code (as I understand it, there may be details I'm wrong about here or
glossing over, but it's not relevant to the organization of things).

- There are several optional runtime libraries to support specific compiler
/ toolchain features. These include the sanitizers and profiling libraries.

I think all of the libraries here make sense in the same repository because
of the shared concerns of building runtime libraries. For example, it would
be useful to compile them with the just-built-clang when target is
executable on the host, and it would be useful to automatically cross
compile versions of them for as many targets as are possible on the given
host and supported by the host compiler (if used) and the just-built-clang.

However, the organization of the tree is ... very hard to understand.
originally, there was only the one 'libclang_rt' library that had its
generic C99 implementation in 'lib', and architecture-specific assembly
routines when desirable in architecture subdirectories of lib. When we
added new libraries, we put them in subdirectories of lib, making the whole
thing kind of a mess. My suggestion to fix this was to create a "core"
subdirectory of lib to contain the code used for 'libclang_rt'. "core" is a
terrible name, but i've no better. suggestions welcome there. Then we would
have a more sensible organization of the 'lib' tree.

I also think it might be useful to have a single large test tree (much like
with llvm or clang) that has subdirectories for the various tests rather
than test directories under lib/asan/ and friends, but maybe the sanitizer
folks have objections to that. consistency seems a compelling reason here,
but there might be other compelling concerns.


Now, the build system has always been problematic because the build system
for this tree is *hard* and no one who has worked on it has really had the
time to do an extremely thorough job and finish all aspects. It's a huge
project. Right now, the makefile system does a good job of using the
just-built-clang and can do a limited amount of cross-building runtimes for
other targets. But the makefile system of compiler-rt is also terribly,
terribly complex, doesn't follow the conventions of LLVM's makefiles, and
is generally painful to maintain and update, so folks have been reluctant
to flesh out its support for new libraries and other new things.

The CMake system is much cleaner in some respects (a bit less opaque to the
folks trying to maintain it), but CMake makes it much harder to use the
just-built-clang, especially for the C++ runtime code. The consequence is
that we've never finished either the cross-building or
just-built-clang-hosting features that are desirable. Any work toward these
would be really awesome to see, but is a huge pile of work. Finally, when I
was originally doing the CMake build for this I didn't understand what
really needed to be done to build and use the core 'libclang_rt' library,
and so I don't think I got it right. Some folks have sent patches to
improve it, but I suspect it still really needs more work to be a solid
system to use instead of libgcc. So I'm really excited about your emails.
=] Having a more self-contained stack would be a significant improvement.


I think the obvious incremental steps are to disable building any parts of
compiler-rt that don't build cleanly. There should never be a requirement
for you to port an optional runtime library unless you need it. =] I think
having good ports is important, but that should never block progress
getting other thinsg ported and working well.

If it helps to reorganize things, I'm happy to even help there as much as I
can. I agree that the organization isn't great, but I *really* didn't want
to fight the makefile build system to do the reorganization myself, so its
something that has lingered too long.


Sorry for the long ramble, but hopefully this gives you some of tho context.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20140131/1505c51a/attachment.html>