llvm-undname: Add support for demangling typeinfo names

[thakis]

Summary

typeinfo names aren't symbols but string constant contents
stored in compiler-generated typeinfo objects, but llvm-cxxfilt
can demangle these for Itanium names.

In the MSVC ABI, these are just a '.' followed by a mangled
type -- this means they don't start with '?' like all MS-mangled
symbols do.

[hans]

llvm/include/llvm/Demangle/MicrosoftDemangle.h

161 ↗

(On Diff #221076)

Off topic: Do you know why the demangler uses its own StringView class instead of StringRef like the rest of llvm?

(Looks like it comes from libcxxabi via this commit: r327859)

[mstorsjo]

llvm/include/llvm/Demangle/MicrosoftDemangle.h

161 ↗

(On Diff #221076)

The demangler can't use the normal llvm classes, as it is supposed to be able to run on top of plain C/C++ standard libraries (because it's shared with the demangling code in libcxxabi).

[hans]

llvm/include/llvm/Demangle/MicrosoftDemangle.h

161 ↗

(On Diff #221076)

Ah, thanks for clarifying!

llvm/trunk/include/llvm/Demangle/MicrosoftDemangle.h

public:
 
   void dumpBackReferences();
 
 private:
   SymbolNode *demangleEncodedSymbol(StringView &MangledName,
                                     QualifiedNameNode *QN);
   SymbolNode *demangleDeclarator(StringView &MangledName);
   SymbolNode *demangleMD5Name(StringView &MangledName);
+  SymbolNode *demangleTypeinfoName(StringView &MangledName);
 
   VariableSymbolNode *demangleVariableEncoding(StringView &MangledName,
                                                StorageClass SC);
   FunctionSymbolNode *demangleFunctionEncoding(StringView &MangledName);
 
   Qualifiers demanglePointerExtQualifiers(StringView &MangledName);
 
   // Parser functions. This is a recursive-descent parser.

llvm/trunk/lib/Demangle/MicrosoftDemangle.cpp

SymbolNode *Demangler::demangleMD5Name(StringView &MangledName) {
 
   StringView MD5(Start, MangledName.begin());
   SymbolNode *S = Arena.alloc<SymbolNode>(NodeKind::Md5Symbol);
   S->Name = synthesizeQualifiedName(Arena, MD5);
 
   return S;
 }
 
+SymbolNode *Demangler::demangleTypeinfoName(StringView &MangledName) {
+  assert(MangledName.startsWith('.'));
+  MangledName.consumeFront('.');
+
+  TypeNode *T = demangleType(MangledName, QualifierMangleMode::Result);
+  if (Error || !MangledName.empty()) {
+    Error = true;
+    return nullptr;
+  }
+  return synthesizeVariable(Arena, T, "`RTTI Type Descriptor Name'");
+}
+
 // Parser entry point.
 SymbolNode *Demangler::parse(StringView &MangledName) {
+  // Typeinfo names are strings stored in RTTI data. They're not symbol names.
+  // It's still useful to demangle them. They're the only demangled entity
+  // that doesn't start with a "?" but a ".".
+  if (MangledName.startsWith('.'))
+    return demangleTypeinfoName(MangledName);
+
   if (MangledName.startsWith("??@"))
     return demangleMD5Name(MangledName);
 
   // MSVC-style mangled symbols must start with '?'.
   if (!MangledName.startsWith('?')) {
     Error = true;
     return nullptr;
   }

llvm/trunk/test/Demangle/invalid-manglings.test

 ; CHECK-EMPTY:
 ; CHECK-NEXT: ??_A
 ; CHECK-NEXT: error: Invalid mangled name
 
 ??_P
 ; CHECK-EMPTY:
 ; CHECK-NEXT: ??_P
 ; CHECK-NEXT: error: Invalid mangled name
+
+.?AUBase@@@8
+; CHECK-EMPTY:
+; CHECK-NEXT: .?AUBase@@@8
+; CHECK-NEXT: error: Invalid mangled name

llvm/trunk/test/Demangle/ms-operators.test

 ; CHECK: void * __cdecl operator new[](unsigned __int64, class klass &)
 
 ??_V@YAXPEAXAEAVklass@@@Z
 ; CHECK: void __cdecl operator delete[](void *, class klass &)
 
 ??_R0?AUBase@@@8
 ; CHECK: struct Base `RTTI Type Descriptor'
 
+.?AUBase@@
+; CHECK: struct Base `RTTI Type Descriptor Name'
+
 ??_R1A@?0A@EA@Base@@8
 ; CHECK: Base::`RTTI Base Class Descriptor at (0, -1, 0, 64)'
 
 ??_R2Base@@8
 ; CHECK: Base::`RTTI Base Class Array'
 
 ??_R3Base@@8
 ; CHECK: Base::`RTTI Class Hierarchy Descriptor'