Description
In the attached reproducer case, I have 2 small libraries, called libv1 and libv2. Each of them contain a single header called header.h containing a single define LIB_VALUE, which expands to either 1 or 2 depending on the library.
There is another library, includer, which is made of a single header, includer.h. All it does is include header.h.
These use module maps defined in lib.modulemap and includer.modulemap respectively. 2 Overlay files are used to use either the libv1 or libv2 library. Finally main.cc does nothing but output the content of the LIB_VALUE macro it get from includer.h.
In the build script, we build main.cc with modules enabled, first with libv1, then libv2, using the same module cache. This results in 2 binaries that both output 1.
We then build using a new module cache, this time first building main.cc with libv2, then libv1. This results in 2 binaries that both output 2.
The problem seems to come from the fact that Clang does not perform any path check when loading a module from the cache. This is particularly problematic for Swift on Windows, which relies on module maps to parse the Windows SDK headers. These are versioned in different sub-directories but use the same module map file, resulting in compilation failures when parts of the SDK are already present in the module cache but for another Windows SDK version.