<table border="1" cellspacing="0" cellpadding="8">
    <tr>
        <th>Issue</th>
        <td>
            <a href=https://github.com/llvm/llvm-project/issues/193340>193340</a>
        </td>
    </tr>

    <tr>
        <th>Summary</th>
        <td>
            [MLGO] Build fails with modern TensorFlow due to unlinked constant/metadata object files
        </td>
    </tr>

    <tr>
      <th>Labels</th>
      <td>
            cmake,
            mlgo
      </td>
    </tr>

    <tr>
      <th>Assignees</th>
      <td>
      </td>
    </tr>

    <tr>
      <th>Reporter</th>
      <td>
          nickdesaulniers
      </td>
    </tr>
</table>

<pre>
    When building LLVM with MLGO enabled (-DTENSORFLOW_AOT_PATH=...) using a recent version of TensorFlow, the build fails during the final link step with undefined symbol errors related to constant buffers.

Newer versions of TensorFlow's AOT compiler (saved_model_cli / tfcompile) generate multiple object files to separate code, metadata, and constants (e.g., InlinerSizeModel.o, InlinerSizeModel_metadata.o, and InlinerSizeModel_constants.o). 

Currently, llvm/cmake/modules/TensorFlowCompile.cmake only tracks the primary object file (${obj_file}). Because the _constants.o and _metadata.o files are not added to GENERATED_OBJS, they are not included in the final archive (libLLVMAnalysis.a), leading to linkage failures.

## Error Messages:

```
   ld.lld: error: undefined symbol: __tfcompile_llvm_InlinerSizeModel_Constant_0_constant_buffer_contents 
   >>> referenced by MLInlineAdvisor.cpp 
   >>> MLInlineAdvisor.cpp.o:(llvm::InlinerSizeModel::EmbeddedConstantBuffers()) in archive lib/libLLVMAnalysis.a
```

## Steps to Reproduce:
1. Install a recent TensorFlow pip package. In my case, 2.21.0 is broken. (`python3 -c "import tensorflow as tf; print(tf.__version__)"`)
2. Configure LLVM with MLGO enabled (e.g., -DTENSORFLOW_AOT_PATH=$TF_PIP -DLLVM_INLINER_MODEL_PATH=...).
3. Build Clang.

## Proposed Solution:
Update TensorFlowCompile.cmake to dynamically detect if the installed TensorFlow version emits these extra object files (either by checking the TF version via Python or by probing the output of a dummy compilation during configuration). If present, these extra objects must be added to the OUTPUT of the add_custom_command and appended to the GENERATED_OBJS list.
</pre>
<img width="1" height="1" alt="" src="http://email.email.llvm.org/o/eJx8Vt1u4zwOfRrlhqjhyEnTXOTCzc9sF21TtJmdS0O26EQTWTIkObPZp19Qzl-nMx-QALFMUofk4WGE92prEGds_MjGi4Hows66mVHVXqIXnTYKnR-UVh5nP3ZooOyUlsps4fn5Py_wS4UdvDx_WwMaUWqUwPjD3WKzfP1Yv6-e1z-KfL0p3vLNv1i2SJKE8Sl0ntwFOKzQBDig88oasDVs0HjrVtr-YnwOYYf9bVALpT3IzpEjHdfKCA1amT34gG0PozMSa2VQgj82pdWAzlnnwaEWASUEC5U1PggToOzqGp1PWJqzNH_FX-jOQPzvSCYe8vUGKtu0SqOjDL04oCwaK1EXlVbA-ApCfbKgHLdo0ImA0HQ6qFYj2PInVgFqpdETFI-tiBaVlUjpNhiEFEHQb2HkBaunCzHZJvTiyWhl0H2o_-EL3Z7YP50W51j9a4r2xeQSnmymCfSVmHfOoQn6SH5aHxrGV1Uj9sj4qrGy0-gZX12rM-9TTqINWKOPEJyo9j62qXWqEe54mzslw_iITR5t-bOgEzZZRACPWInOY3S8RRfh32R0KqFwCMYGEFL2vf22fF2-55vlolg__vvjRKDjxU6ZSndkqswNhYSrduoQUWlVEqVzI_TRK58IxqexCigi34ONhBNbjHzsHJ7pw3jGeAZLohu8oPdii55l-entfXr6pDkAaJloLVmW9_SkH78Tl86K4sKngvpQfGng_FSiIr1Uq-hpTc8BI3PilSxb9h9wWKNDU6GE8ggvz33QXB6Uty6p2vaLxx9sEku58YdIjyxnWf47tv502ZRIzTkDfexnLhJgSkOizKX-WpWMr7624HP5rrX-CNjGMXrH1lnZVdjXe5jAE92m9VVgrmyFVrXQioqaSIbQHKESPo4fT_gwSUF5KJ3do0kiU-_T9hh21mRwVwHjXDWtdQFCDFlTSOEh1Cx7JLKbwPhDqJOiOGlJUcRUOcHnU5bmPIG5NbXadg7_QUHP4_43JWV8tFkVb09vcLegKMXT6_PT6_K9eFkvls-f9ZZImiXwGJV0roXZfqbtm7Ot9Sjhw-ouKGv6Sn5vJanT30Y9WJBHIxpVCa2PIDHQhKs6zpbqW4DytvZnncdGhSgPHgH_G5z4LI2UvQo7dMTQaofV_iz6m9UlxkEJeIuNARsNW2fLs53tQtsFUnEBsmuoxxG6oOTOW6Q6dSEeRv15qqF16JGaOP8DPg9N5wOUeJUcum39ffP2fUO30ZOQsqg6H2xTVLZpSLrit23R3Dh91irQyodkIGeZnGZTMcDZcDK5f0ink-l4sJtVQ3H_MOYyHU2zOnvI-HBcTsYPI16PUj7EyUDNeMrv0xEfcj68z8ZJPU05H6bT-5HMJhkv2SjFRiid0Mgm1m0HyvsOZ8Nplo3SgRYlah__AXB-UnvO-Jxx3uitpYfxYuBm5H1XdlvPRilB9td4QQUd_0MQk9l4caJbv7gjxWlZOnNLCNlFGnUkHnu8LjxaNSe1_0SNQef0bBdCG7WVrxhfbVXYdWVSWdpTp3UVUbbOkiPjq5gpLa1TsocZ_38AAAD__z0d_hs">