Search code examples
c++visual-studio-2013x86shared-ptr

x86/C++ - Pointer To Pointer: Const being violated by compiler?


I was working on a shared pointer (called Handle) implementation for my student project's game engine, and we ran into a bug that we couldn't explain. For some reason, at a certain point in our factory, there was an invalid internal pointer being passed to a factory through a handle, which was causing a crash during our archetype loading. We debugged through the process for hours, and broke any complex statements down to their most simple versions for easier debugging. I finally isolated the problem down to the Handle class' copy constructor. However, it still seemed like there was some intermediate step where the internal pointer was being freed. I read every article I could find on what could cause this problem, and didn't find anything. Finally, I decided to look at the disassembly and see if I could figure out what was happening. Here's the copy constructor without the disassembly.

template <class TYPE>
Handle<TYPE>::Handle(const Handle<TYPE>& rhs, bool weak) : m_weak(weak), m_ref_count(rhs.m_ref_count), m_ptr(rhs.m_ptr)
{
  if(!m_weak)
    ++(*m_ref_count);
}

This is the copy constructor with the disassembly.

template <class TYPE>
Handle<TYPE>::Handle(const Handle<TYPE>& rhs, bool weak) : m_weak(weak), m_ref_count(rhs.m_ref_count), m_ptr(rhs.m_ptr)
{
013FFF50 push         ebp
013FFF51 mov          ebp,esp
013FFF53 sub          esp,0CCh
013FFF59 push         ebx
013FFF5A push         esi
013FFF5B  push        edi  
013FFF5C  push        ecx  
013FFF5D  lea         edi,[ebp-0CCh]  
013FFF63  mov         ecx,33h  
013FFF68  mov         eax,0CCCCCCCCh  
013FFF6D  rep stos    dword ptr es:[edi]  
013FFF6F  pop         ecx  
013FFF70  mov         dword ptr [this],ecx  
013FFF73  mov         eax,dword ptr [this]  
013FFF76  mov         cl,byte ptr [weak]  
013FFF79  mov         byte ptr [eax],cl  
013FFF7B  mov         eax,dword ptr [this]  
013FFF7E  mov         ecx,dword ptr [rhs]  
013FFF81  mov         edx,dword ptr [ecx+4]  
013FFF84  mov         dword ptr [eax+4],edx  
013FFF87  mov         eax,dword ptr [this]  
013FFF8A  mov         ecx,dword ptr [rhs]  
013FFF8D  mov         edx,dword ptr [ecx+8]  
013FFF90  mov         dword ptr [eax+8],edx  
  if(!m_weak)
013FFF93  mov         eax,dword ptr [this]  
013FFF96  movzx       ecx,byte ptr [eax]  
013FFF99  test        ecx,ecx  
013FFF9B  jne         Handle<Component>::Handle<Component>+60h (013FFFB0h)  
    ++(*m_ref_count);
013FFF9D  mov         eax,dword ptr [this]  
013FFFA0  mov         ecx,dword ptr [eax+4]  
013FFFA3  mov         edx,dword ptr [ecx]  
013FFFA5  add         edx,1  
013FFFA8  mov         eax,dword ptr [this]  
013FFFAB  mov         ecx,dword ptr [eax+4]  
013FFFAE  mov         dword ptr [ecx],edx  
}

The problem (as far as I can tell) is here:

013FFF6D  rep stos    dword ptr es:[edi]

As everyone knows, this is used to clear memory quickly. However, this instruction is clearing the pointer that the rhs Handle's internal pointer-to-pointer points to, and not only that, but it's violating const, as the passed in Handle is a const reference. I'm not sure if it's a VS issue, some kind of flag that's set in our build settings that's causing it to happen, but the copy constructor works fine until we try to make archetypes with it.

Any help would be greatly appreciated!

EDIT 1: I have further researched the issue after some of the comments brought to light that I might be returning a temporary variable somewhere in my code. I wasn't able to find anything that would suggest I was doing so, but I did sift through some more disassembly to see if I could find any hints. Here is our CreateGenericArchetype function, which is where the problem is cropping up.

Handle<Object> ObjectFactory::CreateGenericArchetype(const std::wstring &ArchetypeName)
{
  /* Create The Object */
  auto archetype = mComponentFactoryMap.find(std::wstring(L"Object"));
  Handle<Component> created_component = archetype->second->CreateArchetypeComponent();
  Handle<Object> retVal = static_handle_cast<Object>(created_component);
  retVal->mArchetypeName = ArchetypeName;
  mArchetypeMap.insert(std::pair<std::wstring, Handle<Object>>(ArchetypeName, retVal));

  return Handle<Object>(retVal, true);
}

The line that's causing all of our problems (at the moment) is this:

Handle<Component> created_component = archetype->second->CreateArchetypeComponent();

Naturally, I extended my debugging to CreateArchetypeComponent as well, so here is that too:

Handle<Component> ComponentTypeFactory<Object>::CreateArchetypeComponent() const
{
  Object* created_object = new Object;
  Component* casted_object = static_cast<Component*>(created_object);
  Handle<Component> newComponent(casted_object);
  newComponent->mType = mType;
  newComponent->mID = GetNewID();
  return newComponent;
}

Object inherits from Component, so the static_cast downcast of the Object to Component is valid, and the constructor successfully constructs the Handle from the Component*. The issue crops up when we try to return the Handle we constructed inside the function. Here is the disassembly of that interaction:

return newComponent;
005602AD  push        0  
005602AF  lea         eax,[newComponent]  
005602B2  push        eax  
005602B3  mov         ecx,dword ptr [ebp+8]  
005602B6  call        Handle<Component>::Handle<Component> (04EA050h)  
005602BB  mov         ecx,dword ptr [ebp-110h]  
005602C1  or          ecx,1  
005602C4  mov         dword ptr [ebp-110h],ecx  
005602CA  mov         byte ptr [ebp-4],0  
005602CE  lea         ecx,[newComponent]  
005602D1  call        Handle<Component>::~Handle<Component> (04F2E7Bh)  
005602D6  mov         eax,dword ptr [ebp+8]  
}
00EB02D9  push        edx  
00EB02DA  mov         ecx,ebp  
00EB02DC  push        eax  
00EB02DD  lea         edx,ds:[0EB0318h]  
00EB02E3  call        @_RTC_CheckStackVars@8 (0E4BA9Eh)  
00EB02E8  pop         eax  
00EB02E9  pop         edx  
00EB02EA  mov         ecx,dword ptr [ebp-0Ch]  
00EB02ED  mov         dword ptr fs:[0],ecx  
00EB02F4  pop         ecx  
00EB02F5  pop         edi  
00EB02F6  pop         esi  
00EB02F7  pop         ebx  
00EB02F8  mov         ecx,dword ptr [ebp-10h]  
00EB02FB  xor         ecx,ebp  
00EB02FD  call        @__security_check_cookie@4 (0E4E9BAh)  
00EB0302  add         esp,130h  
00EB0308  cmp         ebp,esp  
00EB030A  call        __RTC_CheckEsp (0E41C3Dh)  
00EB030F  mov         esp,ebp  
00EB0311  pop         ebp  
00EB0312  ret         4  

*******back to CreateGenericArchetype*******

005C2639  call        __RTC_CheckEsp (04F1C3Dh)  
005C263E  mov         dword ptr [ebp-1D4h],eax  
005C2644  mov         ecx,dword ptr [ebp-1D4h]  
005C264A  mov         dword ptr [ebp-1D8h],ecx  
005C2650  mov         byte ptr [ebp-4],4  
005C2654  mov         edx,dword ptr [ebp-1D8h]  
005C265A  push        edx  
005C265B  lea         ecx,[created_component]  
005C265E  call        Handle<Component>::Handle<Component> (04EA050h)

*******copy constructor disassembly from above here*******

005C2663  mov         byte ptr [ebp-4],6  
005C2667  lea         ecx,[ebp-174h]  
005C266D  call        Handle<Component>::~Handle<Component> (04F2E7Bh)  
  Handle<Object> retVal = static_handle_cast<Object>(created_component);

Unless my copy constructor being called with the "created_component" as the rhs Handle is considered returning a local variable, I can't see where it could possibly be going wrong, unless it's because the returned Handle from CreateArchetypeComponent is on the stack when it's passed to the copy constructor and is then being cleared.


Solution

  • Thanks for all the great help with this issue, but we figured out what was going on. What was happening is that in some of the Handle constructors we needed to allocate memory for the pointer that the Handle's internal pointer-to-pointer was going to be pointing at. So what was happening was that the Handle itself was using a pointer that was held on the stack, and when the rep stos opcode was called, it was clearing out said pointer.