clang-offload-bundler mixes symbols from different objects when unbundling an archive

[LU-JOHN]

Change to clang/lib/Driver/OffloadBundler.cpp in commit #81700 causes clang-offload-bundler to join symbols from different objects when unbundling an archive:

Source file s.cpp:

void bazball() {
}

Source file t.cpp:

void foobar() {
}

Commands:

clang++ -c s.cpp -o s.o;
clang-offload-bundler -type=o -targets=host-x86_64-unknown-linux-gnu -input=s.o -output=s_bundled.o;
clang++ -c t.cpp -o t.o;
clang-offload-bundler -type=o -targets=host-x86_64-unknown-linux-gnu -input=t.o -output=t_bundled.o;
llvm-ar cr bundled.a s_bundled.o t_bundled.o
clang-offload-bundler -unbundle --targets=host-x86_64-unknown-linux-gnu -type=aoo -input=bundled.a -output=t_list.txt; nm `cat t_list.txt`

The output of the nm command without the change is:

bundled-742762.o:
0000000000000000 T _Z7bazballv
bundled-61506b.o:
0000000000000000 T _Z6foobarv

The output with the change is:

/tmp/bundled-e7ca35.o:s_bundled.o:
0000000000000000 T _Z7bazballvt_bundled.o:
0000000000000000 T _Z6foobarv

/tmp/bundled-5ea078.o:s_bundled.o:
0000000000000000 T _Z7bazballvt_bundled.o:
0000000000000000 T _Z6foobarv

[jsji]

FYI. @yxsamliu

[LU-JOHN]

Sorry, I need to analyze this issue more carefully. It depends on some non-community changes.

[LU-JOHN]

I have a concern about the code in commit #81700 at:

https://github.com/yxsamliu/llvm-project/blob/1e9c29c4794ed6d60cf54c902cd0f854afd1aace/clang/lib/Driver/OffloadBundler.cpp#L718

This code is used to get a processed bundle in an ObjectFileHandler by using objcopy to process BundlerConfig.InputFileNames.front(). My concern is how does this code guarantee that BundlerConfig.InputFileNames.front() is connected to this instance of ObjectFileHandler? This caused a problem for me when unbundling an archive with multiple object files. I hit this problem when using the command sequence in my earlier comment. I have not been able to convert the command sequence into one that works with community tools, but I believe it should be possible when processing a multi-object archive.

[yxsamliu]

Good catch. The code assumes unbundling a non-archive file which has only one input. That is why it breaks for archive.

[yxsamliu]

Are you using some downstream llvm branch? Clang-offload-bundler in the trunk does not accept -type=aoo and also does not allow extract bundle entry for --targets=host-x86_64-unknown-linux-gnu

clang-offload-bundler -unbundle --targets=host-x86_64-unknown-linux-gnu -type=a -input=bundled.a -output=t_list.txt
clang-offload-bundler: error: no compatible code object found for the target 'host-x86_64-unknown-linux-gnu-' in heterogeneous archive library: bundled.a

[LU-JOHN]

Yes, it is an INTEL customization. I was not able to figure out how to extract a multi-object archive with the official clang-offload-bundler. Is it possible?