[lldb-dev] Some API test failures are really opaque/could be improved

Thu Nov 4 08:14:45 PDT 2021

Is someone currently working on fixing this? FWIW, I think David's
change seems to go in the right direction (when I originally looked at
this I also ended up on the wrong rpath but I thought it was some
other code that set the wrong value. Didn't realize we have two places
where this happens). I think David's diff is better than we currently
have so maybe we should just turn this into a review?

Am Di., 26. Okt. 2021 um 06:43 Uhr schrieb David Blaikie <dblaikie at gmail.com>:
>
> On Mon, Oct 25, 2021 at 1:28 PM Louis Dionne <ldionne at apple.com> wrote:
>>
>> I believe the issue is probably not related so much to LLVM_ENABLE_PROJECTS vs LLVM_ENABLE_RUNTIMES, but rather to the fact that LLVM_ENABLE_RUNTIMES uses per-target runtime directories now (hasn't always been the case), which basically means that libc++ ends up in `<prefix>/lib/<target-triple>/libc++.so` instead of `<prefix>/lib/libc++.so`.
>
>
> Ish, yes. It's a bug in LLVM_ENABLE_RUNTIMES that isn't present in LLVM_ENABLE_PROJECTS, so for now if I want to run the lldb pretty printer tests for libc++ on Linux it seems the only way I can is by using the deprecated functionality of libc++ in LLVM_ENABLE_PROJECTS.
>
> Consider this a bug report (looks like a bug in the lldb CMake configuration, not in libc++'s build itself, but something to figure out if Linux lldb devs are going to use libc++ +.ENABLE_RUNTIMES path) on that deprecation?
>
>>
>> I think you either want to specify the per-target library dir when running against libc++, or you want to disable that and use `LLVM_ENABLE_PER_TARGET_RUNTIME_DIR=OFF` when configuring the runtimes. In all cases, you want to be using `LLVM_ENABLE_RUNTIMES` and not `LLVM_ENABLE_PROJECTS`, since the latter is deprecated now.
>
>
> I didn't enable LLVM_ENABLE_PER_TARGET_RUNTIME_DIR myself/in the root cmake config. It looks like it's hardcoded(?) into the ENABLE_RUNTIMES sub-build? https://github.com/llvm/llvm-project/blob/e5fb79b31424267704e9d2d9674089fd7316453e/llvm/runtimes/CMakeLists.txt#L76 I'm not sure there's any way to override that from the root? And in any case I'd have thought the defaults would need to/be intended to work correctly on supported platforms?
>
> So something in lldb's dir handling (maybe some general infrastructure in LLVM could use some improvement to provide an LLVM_RUNTIME_LIBS_DIR, or similar? that could then be used from other places - rather than libc++, for instance, creating that directory for itself based on LLVM_LIBS_DIR and LLVM_ENABLE_PER_TARGET_RUNTIME_DIR, etc) needs some fixes to support the current defaults/hardcoded modes on Linux?
>
>>
>> Cheers,
>> Louis
>>
>> On Oct 25, 2021, at 13:57, David Blaikie <dblaikie at gmail.com> wrote:
>>
>> +Louis Dionne - perhaps the libcxx and lldb folks would be interesting in finding a suitable way to address this issue, since currently either option (using libcxx in ENABLE_PROJECTS or using it in ENABLE_RUNTIMES) is incomplete - if I use ENABLE_RUNTIMES I get the libcxx testing run against the just-built clang and generally this is the "supported" configuration, but then some lldb tests fail because they can't find libcxx.so.1 (on Linux) - and using ENABLE_PROJECTS means not using the just-built clang for libcxx tests (so missing the libcxx breakages caused by my array name change) but do use the just-built libcxx in lldb tests and find failures there...
>>
>> On Wed, Oct 20, 2021 at 1:57 PM David Blaikie <dblaikie at gmail.com> wrote:
>>>
>>> On Tue, Oct 19, 2021 at 4:55 PM David Blaikie <dblaikie at gmail.com> wrote:
>>>>
>>>> On Tue, Oct 19, 2021 at 9:08 AM Raphael Isemann <teemperor at gmail.com> wrote:
>>>>>
>>>>> Actually the RPATH theory is wrong, but the LLVM_ENABLE_PROJECT
>>>>> workaround *should* still work.
>>>>
>>>>
>>>> I'll give that a go (it's running at the moment) though I guess this is inconsistent with the direction libcxx is moving in for building, re: https://groups.google.com/g/llvm-dev/c/tpuLxk_ipLw
>>>
>>>
>>> Yep, it does work with LLVM_ENABLE_PROJECT rather than LLVM_ENABLE_RUNTIME.
>>>
>>> Specifically the test binary is linked with an rpath to the just-built lib directory that ensures the just-built libc++.so is found:
>>>
>>> /usr/local/google/home/blaikie/dev/llvm/build/release/bin/clang main.o -g -O0 -fno-builtin -m64  -I/usr/local/google/home/blaikie/dev/llvm/src/lldb/packages/Python/lldbsuite/test/make/../../../../../include -I/usr/local/google/home/blaikie/dev/llvm/src/lldb/test/API/functionalities/data-formatter/data-formatter-stl/libcxx/list -I/usr/local/google/home/blaikie/dev/llvm/src/lldb/packages/Python/lldbsuite/test/make -include /usr/local/google/home/blaikie/dev/llvm/src/lldb/packages/Python/lldbsuite/test/make/test_common.h -fno-limit-debug-info  -gsplit-dwarf    -stdlib=libc++ -Wl,-rpath,/usr/local/google/home/blaikie/dev/llvm/build/release/./lib --driver-mode=g++ -o "a.out"
>>>
>>> Oh, actually it passes the same rpath when using LLVM_ENABLE_RUNTIME, but the libc++.so.1 is in a different place: ./lib/x86_64-unknown-linux-gnu/libc++.so.1
>>>
>>> Looks like this rpath setting happens here: (changing this to a junk argument causes the test to fail to build as expected)
>>> https://github.com/llvm/llvm-project/blob/618583565687f5a494066fc902a977f6057fc93e/lldb/packages/Python/lldbsuite/test/make/Makefile.rules#L400
>>>
>>> And it gets the LLVM_LIBS_DIR from here: https://github.com/llvm/llvm-project/blob/207998c242c8c8a270ff22a5136da87338546725/lldb/test/API/lit.cfg.py#L163
>>>
>>> So maybe we need to pass down the default target triple too, since that seems to be how libc++ is deciding where to put the library? ( https://github.com/llvm/llvm-project/blob/207998c242c8c8a270ff22a5136da87338546725/libcxx/CMakeLists.txt#L424 ) at least on non-apple :/ (or maybe there's some way to make the connection between the two less brittle - for libc++'s build to export some variable that lldb can use, or for LLVM to provide something for both to use?)
>>>
>>> Yeah, applying this change does work for me, but wouldn't work on Apple for instance (where libcxx doesn't add the default target triple to the path):
>>> $ git diff
>>> diff --git lldb/test/API/lit.site.cfg.py.in lldb/test/API/lit.site.cfg.py.in
>>> index 987078a53edb..e327429b7ff9 100644
>>> --- lldb/test/API/lit.site.cfg.py.in
>>> +++ lldb/test/API/lit.site.cfg.py.in
>>> @@ -3,7 +3,7 @@
>>>  config.llvm_src_root = "@LLVM_SOURCE_DIR@"
>>>  config.llvm_obj_root = "@LLVM_BINARY_DIR@"
>>>  config.llvm_tools_dir = "@LLVM_TOOLS_DIR@"
>>> -config.llvm_libs_dir = "@LLVM_LIBS_DIR@"
>>> +config.llvm_libs_dir = "@LLVM_LIBS_DIR@/@LLVM_DEFAULT_TARGET_TRIPLE@"
>>>  config.llvm_shlib_dir = "@SHLIBDIR@"
>>>  config.llvm_build_mode = "@LLVM_BUILD_MODE@"
>>>  config.lit_tools_dir = "@LLVM_LIT_TOOLS_DIR@"
>>>
>>> Thoughts?
>>
>>