@llvm/issue-subscribers-clang-modules

I don't understand the issue for the reproducer. Why do we add -fmodule-file=nonexistent=ThisDoesntShowOnTheCommandLine.pcm?

This means that all such compilations trigger -Wread-modules-implicitly when an interface depends on some other module.

If the interface depends on other modules, we should add the corresponding -fmodule=<module-name>=<path-to-BMI>. This should be what we discussed. Do I miss anything?

@ChuanqiXu9 Sorry my example was misleading.

If the interface depends on other modules, we should add the corresponding -fmodule==. This should be what we discussed. Do I miss anything?

This is the intended behavior, but the pcm -> object compilation seems to ignore -fmodule-file. The following triggers -Wread-modules-implicitly:

echo "export module A:partition;" > a-partition.cppm

clang++ a-partition.cppm -v -std=c++20 \
    --precompile -o a-partition.pcm

printf '
export module A;

import :partition;
' > a.cppm

clang++ a.cppm -v -std=c++20 \
    -fmodule-file=A:partition=a-partition.pcm \
    --precompile -o a.pcm

# -fmodule-file is ignored.
clang++ a.pcm -v -std=c++20 -c \
    -fmodule-file=A:partition=a-partition.pcm \
    -o a.interface.o

Do you mean when you run the following line:

# -fmodule-file is ignored.
clang++ a.pcm -v -std=c++20 -c \
    -fmodule-file=A:partition=a-partition.pcm \
    -o a.interface.o

you can still see the warning even if the option -fmodule-file is specified?

Then this should be a true bug.

you can still see the warning even if the option -fmodule-file is specified? Then this should be a true bug.

Yes.

I couldn't figure out what caused this for a long time, but it seems that only during the pcm -> o step the -fmodule-file flag is dropped. At least that's my assumption. When inspecting the output of the invocations with -v, there is no -fmodule-file printed for the last compilation step.