[llvm-dev] Inclusion of the ORC runtime in compiler-rt.

Mon Apr 19 15:58:52 PDT 2021

Hi Petr,

The primary difference, at least from my perspective, is that the ABI
> between the compiler and compiler-rt is unstable and these runtimes are
> version locked to the compiler...

This is the case for the ORC runtime as well, at least for now.

This can help guide the appropriate location for ORC runtime, but as you
> can also see, there are no hard rules and either compiler-rt or a
> standalone top-level project can be made to work.
>

If we decide to go with a new top-level project, I wouldn't try to reuse
> compiler-rt CMake build. That build has a number of issues which we're
> trying to fix but it might take a while before the situation improves.
> Instead, I'd start with the simplest possible CMake setup and then see if
> you could reuse common parts of libc and libcxx builds, and if so we can
> extract those into a common location as needed.

A couple of the properties that I was hoping to rely on are support for
platform specific assembly and multi-slice archives as output. Compiler-rt
seems to support those nicely, but I can't see equivalent support in any of
the other runtimes. Do you know whether the compiler-rt build issues
permeate the cmake code that supports those features? If so, are there any
cleaner implementations of those features in other runtimes?

-- Lang.

On Mon, Apr 19, 2021 at 11:52 AM Petr Hosek <phosek at google.com> wrote:

> We have other runtimes that are not a part of compiler-rt but provide ABI
> used by the compiler (for example libcxxabi). The primary difference, at
> least from my perspective, is that the ABI between the compiler and
> compiler-rt is unstable and these runtimes are version locked to the
> compiler (that's why they're installed in
> lib/clang/<version>/lib/<target>/libclang_rt.<name>.a) whereas the runtimes
> that live outside of compiler-rt typically provide standard ABI and you can
> have multiple interchangeable implementations (for example libcxxrt or
> libstdcxx).
>
> There are exceptions to this rule as always. For example, builtins library
> is supposed to be ABI compatible with libgcc. libFuzzer on the other hand
> doesn't provide any ABI used by the compiler, it's just a regular library
> and as far as I can tell, the only reason it lives in compiler-rt is
> convenience (and it's also the first library I'd like to see extracted out
> of compiler-rt).
>
> This can help guide the appropriate location for ORC runtime, but as you
> can also see, there are no hard rules and either compiler-rt or a
> standalone top-level project can be made to work.
>
> If we decide to go with a new top-level project, I wouldn't try to reuse
> compiler-rt CMake build. That build has a number of issues which we're
> trying to fix but it might take a while before the situation improves.
> Instead, I'd start with the simplest possible CMake setup and then see if
> you could reuse common parts of libc and libcxx builds, and if so we can
> extract those into a common location as needed.
>
> On Mon, Apr 19, 2021 at 10:32 AM Lang Hames via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
>
>> Arguments in favor are that the build requirements are very similar, and
>>> there are some conceptual parallels (the ORC runtime is providing
>>> implementations for entry points generated by the compiler and linker,
>>> among other things). On the other hand the implementations are all JIT
>>> specific, which is definitely different from everything else in compiler-rt.
>>
>>
>> After a night to think about it I think I'd rephrase this: *The ORC
>> runtime is a JIT-specific compiler runtime library*. From a JIT
>> developer's point of view it is an excellent fit for compiler-rt, and from
>> a static compiler developer's point of view it's a non-entity, so just dead
>> weight.
>>
>> If compiler-rt were broken up then it would make sense to have the ORC
>> runtime be a separate subproject and client of the common complier-rt build
>> infrastructure. That would be the ideal solution. Until then I think
>> compiler-rt seems like the best home for it -- making the ORC runtime its
>> own subproject now would duplicate compiler-rt's build system only for us
>> to have to reconcile it later (or worse, maintain the duplication
>> indefinitely).
>>
>> Petr, Eric -- Do you have a sense of how difficult it would be to lift
>> compiler-rt's common build infrastructure out? Being new to compiler-rt I'm
>> wary of tackling that, but if you think it'd be easy I'm happy to work with
>> you to try to get it done.
>>
>> -- Lang.
>>
>>
>> On Sun, Apr 18, 2021 at 9:50 PM Lang Hames <lhames at gmail.com> wrote:
>>
>>> Hi Chris,
>>>
>>> My understanding is that compiler-rt is an umbrella project that builds
>>> a number of libraries (builtins, asan, tsan, fuzzer, etc.), and my thought
>>> was that the ORC runtime could fit in as an addition to that set. Arguments
>>> in favor are that the build requirements are very similar, and there are
>>> some conceptual parallels (the ORC runtime is providing implementations for
>>> entry points generated by the compiler and linker, among other things). On
>>> the other hand the implementations are all JIT specific, which is
>>> definitely different from everything else in compiler-rt.
>>>
>>> If not compiler-rt, is there anywhere else that you think this project
>>> would fit? Would it make sense to introduce it as a new top-level project?
>>>
>>> -- Lang.
>>>
>>> On Sun, Apr 18, 2021 at 9:08 PM Chris Lattner <clattner at nondot.org>
>>> wrote:
>>>
>>>> Hey Lang,
>>>>
>>>> Is your goal here to make this part fo compiler_rt the generated
>>>> library, or part of the subproject?  This seems conceptually very different
>>>> than compiler_rt (which was supposed to be entry points implicitly
>>>> generated by the compiler).  Should this be its “own thing”?
>>>>
>>>> -Chris
>>>>
>>>> On Apr 17, 2021, at 1:51 PM, Lang Hames via llvm-dev <
>>>> llvm-dev at lists.llvm.org> wrote:
>>>>
>>>> Hi All,
>>>>
>>>> I've broken the compiler-rt cmake changes and new directories out of
>>>> the ORC runtime prototype and posted them for review in
>>>> https://reviews.llvm.org/D100711.
>>>>
>>>> Most of this was adapted from xray's cmake files and project layout.
>>>> I'm not a CMake expert, so I expect there's room for improvement here, but
>>>> otherwise I'm hoping it's a pretty canonical "new compiler-rt library".
>>>>
>>>> Kind Regards,
>>>> Lang.
>>>>
>>>> On Wed, Apr 14, 2021 at 10:34 AM Lang Hames <lhames at gmail.com> wrote:
>>>>
>>>>> Hi All,
>>>>>
>>>>> Petr -- since the ORC runtime's dependencies are similar to
>>>>> libFuzzer's, is there any reason not to land the ORC runtime in compiler-rt
>>>>> now and then break it out again later? If the compiler-rt refactor is
>>>>> likely to happen soon then it's worth waiting, otherwise I think landing it
>>>>> in compiler-rt sooner rather than later is the best option, so that any
>>>>> kinks in the integration can be worked out before any future compiler-rt
>>>>> refactor.
>>>>>
>>>>> -- Lang.
>>>>>
>>>>>
>>>>>
>>>>> On Mon, Apr 12, 2021 at 4:38 PM Lang Hames <lhames at gmail.com> wrote:
>>>>>
>>>>>> From this it sounds like "convenient reusing of the build system"
>>>>>>> rather than "should be included in compiler-rt as a library"? If that's the
>>>>>>> case maybe making it clear or lifting the common build system support out
>>>>>>> might be maintainable without the "this is a runtime library for the
>>>>>>> system" sort of thing?
>>>>>>
>>>>>>
>>>>>> Yeah. It sounds like in an ideal world we'd lift out the common build
>>>>>> system support, then have a new set of sub-projects that re-use that
>>>>>> generic build system.
>>>>>>
>>>>>> Do you have any sense of how difficult it would be to lift out that
>>>>>> common build system code? If that's relatively easy then maybe the right
>>>>>> approach is to do that first, then land the ORC runtime. Otherwise the ORC
>>>>>> runtime could go in to compiler-rt for now, then be split out with the rest
>>>>>> of compiler-rt when it's broken up -- it doesn't require any meaningful
>>>>>> changes to existing compiler-rt code, so it should be very easy to break
>>>>>> back out again later.
>>>>>>
>>>>>> -- Lang.
>>>>>>
>>>>>>
>>>>>> On Mon, Apr 12, 2021 at 3:51 PM Eric Christopher <echristo at gmail.com>
>>>>>> wrote:
>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> On Mon, Apr 12, 2021 at 6:36 PM Lang Hames <lhames at gmail.com> wrote:
>>>>>>>
>>>>>>>> Hi Petr,
>>>>>>>>
>>>>>>>> I'd like to better understand the structure of the ORC runtime and
>>>>>>>>> its dependencies (both build and runtime). Does it use the C or C++
>>>>>>>>> standard library? Does it depend on other parts of LLVM? Do you plan on
>>>>>>>>> reusing some of the existing compiler-rt libraries like sanitizer_common?
>>>>>>>>
>>>>>>>>
>>>>>>>> The ORC runtime currently uses the C++ standard library.
>>>>>>>> Since the ORC runtime needs to communicate with the LLVM ORC
>>>>>>>> library it also currently uses some header-only includes from LLVM. It does
>>>>>>>> not depend on any LLVM libraries. We could duplicate this code, but I'd
>>>>>>>> prefer to share it if possible.
>>>>>>>> I have not used sanitizer_common, but some parts of it look like
>>>>>>>> they may be useful.
>>>>>>>>
>>>>>>>> I gravitated towards implementing the ORC runtime in compiler-rt
>>>>>>>> because I need to be able to write parts of it in platform-specific
>>>>>>>> assembly (which compiler-rt supports), and because the runtime should be
>>>>>>>> build for all targets, not just the host (which seems to be the standard
>>>>>>>> way that compiler-rt is configured).
>>>>>>>>
>>>>>>>>
>>>>>>> From this it sounds like "convenient reusing of the build system"
>>>>>>> rather than "should be included in compiler-rt as a library"? If that's the
>>>>>>> case maybe making it clear or lifting the common build system support out
>>>>>>> might be maintainable without the "this is a runtime library for the
>>>>>>> system" sort of thing?
>>>>>>>
>>>>>>> -eric
>>>>>>>
>>>>>>>
>>>>>>>> To give a bit more background on why I'm interested, compiler-rt
>>>>>>>>> has grown fairly organically this has been making the maintenance more and
>>>>>>>>> more difficult, at least from the build perspective. There are some
>>>>>>>>> runtimes that only use C, some that use C++, some that use C++ standard
>>>>>>>>> library. When building compiler-rt together with other runtimes like libc
>>>>>>>>> or libc++, it's difficult to pick up the right order which is why we have
>>>>>>>>> several entry points into compiler-rt's build system to build different
>>>>>>>>> subsets and that's been a maintenance headache.
>>>>>>>>
>>>>>>>>
>>>>>>>>> I've been thinking about this quite a bit recently and I
>>>>>>>>> repeatedly came to the conclusion that compiler-rt would ideally be broken
>>>>>>>>> up into several subprojects, but that should probably be discussed as a
>>>>>>>>> separate topic. However, understanding the build and runtimes dependencies
>>>>>>>>> of the ORC runtime could help us decide whether it should be a part of
>>>>>>>>> compiler-rt or be a separate subproject.
>>>>>>>>
>>>>>>>>
>>>>>>>> That makes sense to me. I think of the ORC runtime as a
>>>>>>>> compiler-rt-style runtime with a libc++ dependency. In that sense I think
>>>>>>>> it's similar to libFuzzer, and whatever solution we come up with for
>>>>>>>> libFuzzer would probably also be applicable to the ORC runtime too.
>>>>>>>>
>>>>>>>> -- Lang.
>>>>>>>>
>>>>>>>> On Mon, Apr 12, 2021 at 2:36 PM Petr Hosek <phosek at google.com>
>>>>>>>> wrote:
>>>>>>>>
>>>>>>>>> I'd like to better understand the structure of the ORC runtime and
>>>>>>>>> its dependencies (both build and runtime). Does it use the C or C++
>>>>>>>>> standard library? Does it depend on other parts of LLVM? Do you plan on
>>>>>>>>> reusing some of the existing compiler-rt libraries like sanitizer_common?
>>>>>>>>>
>>>>>>>>> To give a bit more background on why I'm interested, compiler-rt
>>>>>>>>> has grown fairly organically this has been making the maintenance more and
>>>>>>>>> more difficult, at least from the build perspective. There are some
>>>>>>>>> runtimes that only use C, some that use C++, some that use C++ standard
>>>>>>>>> library. When building compiler-rt together with other runtimes like libc
>>>>>>>>> or libc++, it's difficult to pick up the right order which is why we have
>>>>>>>>> several entry points into compiler-rt's build system to build different
>>>>>>>>> subsets and that's been a maintenance headache.
>>>>>>>>>
>>>>>>>>> I've been thinking about this quite a bit recently and I
>>>>>>>>> repeatedly came to the conclusion that compiler-rt would ideally be broken
>>>>>>>>> up into several subprojects, but that should probably be discussed as a
>>>>>>>>> separate topic. However, understanding the build and runtimes dependencies
>>>>>>>>> of the ORC runtime could help us decide whether it should be a part of
>>>>>>>>> compiler-rt or be a separate subproject.
>>>>>>>>>
>>>>>>>>> On Mon, Apr 12, 2021 at 12:26 PM Lang Hames <lhames at gmail.com>
>>>>>>>>> wrote:
>>>>>>>>>
>>>>>>>>>> Hi All,
>>>>>>>>>>
>>>>>>>>>> I'd like to add the ORC runtime library (preview available at
>>>>>>>>>> https://github.com/lhames/llvm-project/tree/orc-runtime-preview)
>>>>>>>>>> to compiler-rt.
>>>>>>>>>>
>>>>>>>>>> Background:
>>>>>>>>>>
>>>>>>>>>> ORC, like MCJIT, can link JIT'd code either into the current
>>>>>>>>>> process ("in-process" mode) or into a remote executor process
>>>>>>>>>> ("cross-process" mode). Some JIT features require support code in the
>>>>>>>>>> executor process, but the existing ORC libraries are only linked into the
>>>>>>>>>> JIT process. This has made cross-process mode support for those features
>>>>>>>>>> (which include static initializers, thread local variables, exception
>>>>>>>>>> handling, and others) awkward or impractical to implement. The ORC runtime
>>>>>>>>>> library aims to provide the necessary support code in a form that is
>>>>>>>>>> loadable by the JIT itself, which should allow these features to work
>>>>>>>>>> uniformly in both modes.
>>>>>>>>>>
>>>>>>>>>> My prototype branch of the ORC runtime (available at
>>>>>>>>>> https://github.com/lhames/llvm-project/tree/orc-runtime-preview)
>>>>>>>>>> has advanced to the point where it can provide uniform support for static
>>>>>>>>>> initializers, destructors, exceptions, thread locals, and language
>>>>>>>>>> registration for Objective C and Swift code. This support is all
>>>>>>>>>> MachO/Darwin only so far, but should be easily adaptable for ELF/Linux/BSD
>>>>>>>>>> support.
>>>>>>>>>>
>>>>>>>>>> Proposal:
>>>>>>>>>>
>>>>>>>>>> The proof of concept implementation has been very successful, so
>>>>>>>>>> I would like to move future development to the LLVM main branch so that
>>>>>>>>>> others can benefit from this.
>>>>>>>>>>
>>>>>>>>>> Before I start posting patches, though:
>>>>>>>>>>
>>>>>>>>>> Does anyone see any problems with including this in compiler-rt?
>>>>>>>>>>
>>>>>>>>>> Does anyone think that there is a more reasonable home for the
>>>>>>>>>> ORC runtime within the llvm-project? I considered LLVM itself, or a new
>>>>>>>>>> top-level project, but neither seemed as natural a fit as compiler-rt.
>>>>>>>>>>
>>>>>>>>>> Finally, if everyone is happy for it to be included in principle,
>>>>>>>>>> are there any volunteers to review ORC runtime patches?
>>>>>>>>>>
>>>>>>>>>> Regards,
>>>>>>>>>> Lang.
>>>>>>>>>>
>>>>>>>>> _______________________________________________
>>>> LLVM Developers mailing list
>>>> llvm-dev at lists.llvm.org
>>>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>>>
>>>>
>>>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20210419/c6fda82b/attachment-0001.html>