Unable to run inlining-demo

[srivathsang]

I have been trying to run the inlining-demo, but keep running into the same error while building LLVM for Fuchsia. The error happens specifically when I try to generate runtimes with the given python script:

$ python3 scripts/clang/generate_runtimes.py --clang-prefix=$LLVM_INSTALLDIR --sdk-dir=$IDK_DIR --build-id-dir=$LLVM_INSTALLDIR/lib/.build-id > $LLVM_INSTALLDIR/lib/runtime.json

ld.lld: error: unable to find library -lc++
ld.lld: error: unable to find library -lunwind
clang++: error: ld.lld command failed with exit code 1 (use -v to see invocation)
ld.lld: error: unable to find library -lc++
ld.lld: error: unable to find library -lunwind
clang++: error: ld.lld command failed with exit code 1 (use -v to see invocation)
ld.lld: error: unable to find library -lc++
ld.lld: error: cannot open /home/sgovind7/llvm-install/lib/clang/20/lib/x86_64-unknown-fuchsia/libclang_rt.asan.so: No such file or directory
ld.lld: error: cannot open /home/sgovind7/llvm-install/lib/clang/20/lib/x86_64-unknown-fuchsia/libclang_rt.asan-preinit.a: No such file or directory
ld.lld: error: cannot open /home/sgovind7/llvm-install/lib/clang/20/lib/x86_64-unknown-fuchsia/libclang_rt.asan_static.a: No such file or directory
ld.lld: error: unable to find library -lunwind
clang++: error: ld.lld command failed with exit code 1 (use -v to see invocation)
ld.lld: error: unable to find library -lc++
ld.lld: error: cannot open /home/sgovind7/llvm-install/lib/clang/20/lib/x86_64-unknown-fuchsia/libclang_rt.asan.so: No such file or directory
ld.lld: error: cannot open /home/sgovind7/llvm-install/lib/clang/20/lib/x86_64-unknown-fuchsia/libclang_rt.asan-preinit.a: No such file or directory
ld.lld: error: cannot open /home/sgovind7/llvm-install/lib/clang/20/lib/x86_64-unknown-fuchsia/libclang_rt.asan_static.a: No such file or directory
ld.lld: error: unable to find library -lunwind
clang++: error: ld.lld command failed with exit code 1 (use -v to see invocation)
ld.lld: error: unable to find library -lc++
ld.lld: error: cannot open /home/sgovind7/llvm-install/lib/clang/20/lib/x86_64-unknown-fuchsia/libclang_rt.ubsan_standalone.so: No such file or directory
ld.lld: error: unable to find library -lunwind
clang++: error: ld.lld command failed with exit code 1 (use -v to see invocation)
ld.lld: error: unable to find library -lc++
ld.lld: error: cannot open /home/sgovind7/llvm-install/lib/clang/20/lib/x86_64-unknown-fuchsia/libclang_rt.ubsan_standalone.so: No such file or directory
ld.lld: error: unable to find library -lunwind
clang++: error: ld.lld command failed with exit code 1 (use -v to see invocation)

Fuchsia hash: 42f5675c4
LLVM hash: 8280651ad

P.S: I have disabled generating runtimes for other targets (ARM and RISC), but having them also produces the same error

[mtrofin]

See issue #373 - the build fixes on Fuchsia haven't landed yet, will update there.

[srivathsang]

I tried using the -DLLVM_INLINER_MODEL_PATH=download flag while building llvm from source and got the following error:

CMake Warning:
  Manually-specified variables were not used by the project:

    LLVM_INLINER_MODEL_PATH

The cmake command that I use is:

 cmake -DLLVM_ENABLE_PROJECTS=clang -DLLVM_INLINER_MODEL_PATH=download  -DCMAKE_BUILD_TYPE=Release -G "Unix Makefiles" ../llvm

[boomanaiden154]

The download functionality was removed llvm/llvm-project@24c6c35. I believe setting LLVM_INLINER_MODEL_PATH to the URL removed in the commit should work.

You also need to make sure that you enable building with MLGO release mode. For that you need to set TENSORFLOW_AOT_PATH in the CMake invocation, like done in the demos.

[srivathsang]

I still get the same error stating that the specified variable was not used. I tried ccmake and couldn't find the flags LLVM_INLINER_MODEL_PATH nor LLVM_RAEVICT_MODEL_PATH. Have these been removed in any subsequent commits?

I have the following paths set before running cmake:

$ echo $TENSORFLOW_AOT_PATH
~/.local/share/virtualenvs/ml-compiler-opt-ms369JFz/lib/python3.8/site-packages/tensorflow
$ echo $TFLITE_PATH
~/tflite

[mtrofin]

Have you tried passing to cmake -DTENSORFLOW_AOT_PATH=~/.local/share/virtualenvs/ml-compiler-opt-ms369JFz/lib/python3.8/site-packages/tensorflow (in addition to the rest of the cmake flags).

[srivathsang]

Apologies for the delayed response.
Thanks for pointing that out, I was able to build it without that previous warning message.
How can I check if the inlining model is being used by the newly built clang on my custom test program?

[boomanaiden154]

If you built LLVM with assertions enabled, you should be able to pass -debug-only=inline-ml to LLVM and get some output. There are also some optimization remarks. There's also the enable-scc-inline-advisor-printing pass that should enable printing the advisor decisions.

If your binary size changes when enabling the ML inliner compared to when not enabling it though, that's usually a good sign the flags are having some effect.

Hi! During the inlining demo where I'd use clang both for training and corpus collection following these instructions I was encountering the same CMake warning. Not sure what fixed it, probably how I passed the DTENSORFLOW_AOT_PATH. Was hoping this would be the reason why the binaries generated with and without the policy have the same size for all but one file, where the baseline actually performed marginally better. Is this expected with this setup?

[mtrofin]

Followed up in #303 (comment)