One of the current limitations on LLVM on Windows is that you cannot use LLVM_BUILD_LLVM_DYLIB: https://github.com/llvm/llvm-project/blob/main/llvm/tools/llvm-shlib/CMakeLists.txt#L14-L16 I am interested in trying to see if we can lift this limitation. There are others in the community that also seem to be interested in seeing LLVM being possible to use as a DLL on Windows and the topic does come up on the mailing lists every so often.
When you build a distribution of a LLVM based toolchain currently, the result on Windows is ~2GiB for a trimmed down toolset. This is largely due to the static linking used for all the tools. I would like to be able to use the shared LLVM build for building a toolset on Windows.
Unlike Unix platforms, the default on Windows is that all symbols are treated as
dso_local (that is
-fvisibility-default=hidden). Symbols which are meant to participate in dynamic linking are to be attributed as
__declspec(dllexport) in the module and
__declspec(dllimport) external to the module. This is similar to Unix platforms where
__attribute__((__visibility__(...))) controls the same type of behaviour with
For the case of distributions, it would remain valuable to minimize the number of shared objects to reduce the files that require to be shipped but also to minimize the number of cross-module calls which are not entirely free (i.e. PLT+GOT or IAT costs). At the same time, the number of possible labels which can be exposed from a single module on Windows is limited to 64K. Experience from MSys2 indicates that LLVM with all the backends is likely to exceed this count (with a subset of targets, the number already is close to 60K). This means that it may be that we would need two libraries on Windows.
With the LLVM community being diverse, people often build on different platforms with different configurations, and I am concerned that adding more differences in how we build libraries complicates how maintainable LLVM is. I would suggest that we actually change the behavior of the Unix builds to match that of Windows by building with
-fvisibility-default=hidden. Although this is a change, it is not without value. By explicitly marking the interfaces which are vended by a library and making everything else internal, it does enable some potential optimization options for the compiler and linker (to be clear, I am not suggesting that this will have a guaranteed benefit, just that it can potentially enable additional opportunities for optimizations and size reductions). This should incidentally help static linking.
In order to achieve this, we would need to have a module specific annotation to indicate what symbols are meant to be used outside of the module when built in a shared configuration. The same annotation would apply to all targets and is expected to be applied uniformly. This of course has a cost associated with it: the public interfaces would need to be decorated appropriately. However, by having the same behaviour on all the platforms, developers would not be impacted by the platform differences in their day-to-day development. The only time that developers would need to be aware of this is when they are working on the module boundary, that is, changes which do not change the API surface of LLVM would not need to consider the annotations.
Concretely, what I believe is required to enable building with LLVM_BUILD_LLVM_DYLIB on Windows is:
- introduce module specific decoration (e.g. LLVM_SUPPORT_ABI, …) to mark public interfaces of shared library modules
- decorate all the public interfaces of the shared library modules with the new decoration
- switching the builds to use
I believe that these can be done mostly independently and staged in the order specified. Until the last phase, it would have no actual impact on the builds. However, by staging it, we could allow others to experiment with the option while it is under development, and allows for an easier path for switching the builds over.
Although this would enable LLVM_BUILD_LLVM_DYLIB on Windows, give us better uniformity between Windows and non-Windows platforms, potentially enable additional optimization benefits, improve binary sizes for a distribution of the toolchain (though less on Linux where distributors are already using the build configuration ignoring the official suggestions in the LLVM guides), and help with runtime costs of the toolchain (by making the core of the tools a shared library, the backing pages can now be shared across multiple instances), it is not entirely without downsides. The primary downsides that I see are:
- it becomes less enticing to support both LLVM_BUILD_LLVM_DYLIB and BUILD_SHARED_LIBS: while technically possible, interfaces will need to be decorated for both forms of the build
- LLVM_DYLIB_COMPONENTS becomes less tractable: in theory it is possible to apply enough CPP magic to determine where a symbol is homed, but allowing a symbol to be homed in a shared or static library is significantly more complex
- BUILD_SHARED_LIBS becomes more expensive to maintain: the decoration is per-module, which requires that we would need to decorate the symbols of each module with module specific annotations as well
One argument that people make for BUILD_SHARED_LIBS is that it reduces the overall time build-test cycle. With the combination of lld, DWARF Fission, and LLVM_BUILD_LLVM_DYLIB, I believe that most of the benefits still can be had. The cost of linking all the tools is amortized across the link of a single library, which while not as small as the a singular library, is offset by the following:
- The LLVM_BUILD_LLVM_DYLIB would not require the re-linking of all the libraries for each tool.
- DWARF Fission would avoid the need to relink all of the DWARF information.
- lld is faster than the gold and bfd linkers
Header changes would still ripple through the system as before, requiring rebuilding the transitive closure. Source file changes do not have the same impact of course.
For those would like a more concrete example of what a change like this may shape up into: https://reviews.llvm.org/D109192 contains
LLVMSupportExports.h which has the expected structure for declaring the decoration macros with the rest of the change primarily being focused on applying the decoration. Please ignore the CMake changes as they are there to ensure that the CI validates this without changing the configuration and not intended to be part of the final version of the change.