[lldb][gnustep][PDB] Parse ObjC base classes and recognize NSObject type

[sgraenitz]

Summary

While C++ base classes are registered with CreateBaseClassSpecifier(), we have to use SetObjCSuperClass() for ObjC. The isa member of NSObject is a zero-length UDT. Special handling for ObjC id type will be added in one of the next patches.

[DavidSpickett]

I don't know anything about Objective C internals, perhaps you can provide some reference that talks about this?

lldb/test/Shell/Expr/objc-gnustep-print-pdb.m

44

Could you use label in the source so you don't have to update this line number each each time?

I forget if labels are included in the debug info. In the python tests we have this more heavyweight thing that searches for a particular comment to find the line.

[sgraenitz]

In D146297#4202172, @DavidSpickett wrote:

I don't know anything about Objective C internals, perhaps you can provide some reference that talks about this?

Honestly, I am mostly looking what the DWARF parser does and try to mimic it in PDB. I haven't done much research for a formal reference. My impression is that the implementation is the reference here. (Please correct me if I am wrong!)

Yeah, let's wait and see if I get feedback from ObjC and PDB maintainers.

lldb/source/Plugins/SymbolFile/PDB/PDBASTParser.cpp

349

Maybe this is ok without a dedicated language type check? IIUC the worst thing that could happen, in case we really have zero-length types with such names in C++ code, is a small performance drop.

BTW the reference for this seems to be: https://github.com/llvm/llvm-project/blob/release/16.x/lldb/source/Plugins/SymbolFile/DWARF/DWARFASTParserClang.cpp#L709-L710 😄

lldb/test/Shell/Expr/objc-gnustep-print-pdb.m

44

It's a minor effort on my side to update the line number and it's solid. We cannot break on ok, since the function name doesn't get parsed with a correct basename yet. Main would work, but it doesn't seem more readable:

-o "b main" -o "n" -o "n"

In the python tests we have this more heavyweight thing that searches for a particular comment to find the line.

If possible I'd like to stay away from the Python tests. I don't think we need their advanced features. lit is a lot easier for me (and more widespread in LLVM).

68

With this change we get base class and the member offsets right (all are zero-initialized).

[compnerd]

lldb/source/Plugins/SymbolFile/PDB/PDBASTParser.cpp

399

Does it make sense to drop a note here that the 0-length UDT is used as a marker for the NSObject root? Although, shouldn't NSObject's isa be NSObject? Or does GNUStep not do that?

lldb/source/Plugins/SymbolFile/PDB/PDBASTParser.cpp

GetDeclFromContextByName(const clang::ASTContext &ast,
 
   return *result.begin();
 }
 
 static bool IsAnonymousNamespaceName(llvm::StringRef name) {
   return name == "`anonymous namespace'" || name == "`anonymous-namespace'";
 }
 
+static bool IsSpecialNameObjC(llvm::StringRef name) {
+  return name == "objc_object" || name == "objc_class";
+}
+
 static clang::CallingConv TranslateCallingConvention(PDB_CallingConv pdb_cc) {
   switch (pdb_cc) {
   case llvm::codeview::CallingConvention::NearC:
     return clang::CC_C;
   case llvm::codeview::CallingConvention::NearStdCall:
     return clang::CC_X86StdCall;
   case llvm::codeview::CallingConvention::NearFast:
     return clang::CC_X86FastCall;
case PDB_SymType::UDT: {
     assert(udt);
 
     // Note that, unnamed UDT being typedef-ed is generated as a UDT symbol
     // other than a Typedef symbol in PDB. For example,
     //    typedef union { short Row; short Col; } Union;
     // is generated as a named UDT in PDB:
     //    union Union { short Row; short Col; }
     // Such symbols will be handled here.
 
-    // Some UDT with trival ctor has zero length. Just ignore.
-    if (udt->getLength() == 0)
-      return nullptr;
-
     // Ignore unnamed-tag UDTs.
     std::string name =
         std::string(MSVCUndecoratedNameParser::DropScope(udt->getName()));
     if (name.empty())
       return nullptr;
 
+    // Some UDT with trival ctor has zero length. Just ignore.
+    if (udt->getLength() == 0 && !IsSpecialNameObjC(name))
+      return nullptr;
+
     auto decl_context = GetDeclContextContainingSymbol(type);
 
     // PDB has no attribute to encode the language per symbol. We assume
     // all ObjCInterfaceDecls to derive from specific base classes.
     LanguageType lang = eLanguageTypeC_plus_plus;
     auto *symbol_file = static_cast<SymbolFilePDB *>(m_ast.GetSymbolFile());
     if (symbol_file->IsaNSObjectOrNSProxy(*udt))
       lang = eLanguageTypeObjC;
if (!base_comp_type.GetCompleteType()) {
           ":: Class '{0}' has a base class '{1}' "
           "which does not have a complete definition.",
           record_type.GetTypeName().GetCString(),
           base_comp_type.GetTypeName().GetCString());
       if (TypeSystemClang::StartTagDeclarationDefinition(base_comp_type))
         TypeSystemClang::CompleteTagDeclarationDefinition(base_comp_type);
     }
 
+    if (TypeSystemClang::IsObjCObjectOrInterfaceType(base_comp_type)) {
+      m_ast.SetObjCSuperClass(record_type, base_comp_type);
+      assert(bases_enum.getNext() == nullptr && "Single inheritance only");
+      return;
+    }
+
     auto access = TranslateMemberAccess(base->getAccess());
 
     auto is_virtual = base->isVirtualBaseClass();
 
     std::unique_ptr<clang::CXXBaseSpecifier> base_spec =
         m_ast.CreateBaseClassSpecifier(base_comp_type.GetOpaqueQualType(),
                                        access, is_virtual,
                                        record_kind == clang::TTK_Class);

lldb/test/Shell/Expr/objc-gnustep-print-pdb.m

 - (int)ok;
 @end
 @implementation TestObj
 - (int)ok {
   return self ? 0 : 1;
 }
 @end
 
-// RUN: %lldb -b -o "b objc-gnustep-print-pdb.m:67" -o "run" -o "p ptr" -o "p *ptr" -- %t | FileCheck %s
+// RUN: %lldb -b -o "b objc-gnustep-print-pdb.m:72" -o "run" -o "p ptr" -o "p *ptr" -- %t | FileCheck %s
 //
-// CHECK: (lldb) b objc-gnustep-print-pdb.m:67
-// CHECK: Breakpoint {{.*}} at objc-gnustep-print-pdb.m:67
+// CHECK: (lldb) b objc-gnustep-print-pdb.m:72
+// CHECK: Breakpoint {{.*}} at objc-gnustep-print-pdb.m:72
 //
 // CHECK: (lldb) run
 // CHECK: Process {{[0-9]+}} stopped
-// CHECK: frame #0: {{.*}}`main  at objc-gnustep-print-pdb.m:67
+// CHECK: frame #0: {{.*}}`main  at objc-gnustep-print-pdb.m:72
 //
 // CHECK: (lldb) p ptr
 // CHECK: (TestObj *) $0 = 0x{{[0-9]+}}
 //
 // CHECK: (lldb) p *ptr
 // CHECK: (TestObj) $1 = {
-// CHECK:   _int
-// CHECK:   _float
-// CHECK:   _char
-// CHECK:   _ptr_void
-// CHECK:   _ptr_nsobject
-// CHECK:   _id_objc
+// CHECK:   NSObject = {
+// CHECK:     isa = 0x{{[0-9]+}}
+// CHECK:     refcount
+// CHECK:   }
+// CHECK:   _int = 0
+// CHECK:   _float = 0
+// CHECK:   _char = '\0'
+// CHECK:   _ptr_void = 0x{{0+}}
+// CHECK:   _ptr_nsobject = nil
+// CHECK:   _id_objc = nil
 // CHECK: }
 
 int main() {
   TestObj *ptr = [TestObj new];
   return [ptr ok];
 }