vgvassilev wrote: > > One thing I don't understand about this PR is why we need Compiler.h in > > Clang -- wouldn't the LLVM definitions in their Compiler.h work for Clang > > as well? (This would probably be worth explaining in the patch summary.) > > The symbol lookup in PE/COFF is two level and symmetric, bound to the module. > As a result, each module needs to explicitly specify the ABI interface > (contract). However, given that the interface is annotated differently > (`__declspec(dllexport)` for the producer, `__declspec(dllimport)` for the > consumer), we need to flip the definition. When doing so, we need to know > which module is being built. We cannot just use the conjunction of the module > identifier to select between the two variants as there may be > interdependencies with dynamic linking (e.g. `#if defined(A_ABI) || > defined(B_ABI)` doesn't work if B depends on A and dynamically links to A). > > I do agree that we could put this in the commit message, though, this is > likely something that deserves a note in the developer documentation because > I fear that other developers are going to have this question repeatedly once > this work is completed. The biggest point of contention with this work was > the fact that it involves an ongoing maintenance cost as the ABI boundary is > now being more rigidly defined which is/was not needed to support it on Unix > due to the (global) single level binding.
@fsfod, can you capture this comment in the commit message and the documentation as suggested? https://github.com/llvm/llvm-project/pull/108276 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
