Always Import FieldDecls from Records

When we switched this path over to use lazy member loading, a
subsequent commit introduced a check that members were canonical. In
certain extremely twisty scenarios involving multiple modular frameworks
pulling in non-modular headers, this check can fail which results in us
silently dropping these members on the floor.

It's ultimately correct to only pull in the canonical members, but real-world
examples of "system" frameworks evading modularity diagnostics exist so this
is strictly a regression.

Drop the canonicity check for FieldDecls.

rdar://86740970

Author: CodaFi

lib/ClangImporter/ImportDecl.cpp

@@ -9769,13 +9769,21 @@ void ClangImporter::Implementation::loadAllMembersOfRecordDecl(
   llvm::SmallVector<Decl *, 16> members;
   for (const clang::Decl *m : clangRecord->decls()) {
     auto nd = dyn_cast<clang::NamedDecl>(m);
+    if (!nd)
+      continue;
+
     // Currently, we don't import unnamed bitfields.
     if (isa<clang::FieldDecl>(m) &&
         cast<clang::FieldDecl>(m)->isUnnamedBitfield())
       continue;
 
-    if (nd && nd == nd->getCanonicalDecl() &&
-        nd->getDeclContext() == clangRecord &&
+    // Make sure we always pull in record fields. Everything else had better
+    // be canonical. Note that this check mostly catches nested C++ types since
+    // we import nested C struct types by C's usual convention of chucking them
+    // into the global namespace.
+    const bool isCanonicalInContext =
+        (isa<clang::FieldDecl>(nd) || nd == nd->getCanonicalDecl());
+    if (isCanonicalInContext && nd->getDeclContext() == clangRecord &&
         isVisibleClangEntry(nd))
       insertMembersAndAlternates(nd, members);
   }