[solved] C++ Compiler Bug with ! bool initialisers ? | Tech Off

RichardRudek
So what do you expect for nothin'... :P

I think I've discovered a compiler bug.

Well, I don't know wether this is known or not, but if I could get confirmations with both the VS2005 and VS2008 c++ compilers.

I'm working my way through [the safe bool idiom article], and specifically, i found the error implementing the OperatorNot class:

// operator! version
class Testable {
  bool not_ok_;
public:
  explicit Testable(bool b=true):not_ok_(!b) {}

  bool operator!() const {
    return not_ok_;
  }
};

Notice the not (!) in the initialiser.

When run it does not initialise the not_ok member to false, when instanitating the class like:

 Testable test;

Here's the disassembly of the constructor:

public:
  explicit Testable(bool b=true):ok_(b) {}
004115F0  push        ebp  
004115F1  mov         ebp,esp 
004115F3  sub         esp,0CCh 
004115F9  push        ebx  
004115FA  push        esi  
004115FB  push        edi  
004115FC  push        ecx  
004115FD  lea         edi,[ebp-0CCh] 
00411603  mov         ecx,33h 
00411608  mov         eax,0CCCCCCCCh 
0041160D  rep stos    dword ptr es:[edi] 
0041160F  pop         ecx  
00411610  mov         dword ptr [ebp-8],ecx 
00411613  mov         eax,dword ptr [this] 
00411616  mov         cl,byte ptr [b] 
00411619  mov         byte ptr [eax],cl 
0041161B  mov         eax,dword ptr [this] 
0041161E  pop         edi  
0041161F  pop         esi  
00411620  pop         ebx  
00411621  mov         esp,ebp 
00411623  pop         ebp  
00411624  ret         4

WTF.

I've checked multiple times, and the original source does have the !.

Yet when I switch to the disassembly view, and single-step into the constructor, it shows it (the bit marked in red) without the !. The bit embolded is the actual native code setting the member.

Am I missing something ?

This is using VS2005sp1 (c1xx.dll version 14.0.50727.762)

Sven Groot
You can't have everything; after all, where would you put it?

No repro here, on either VS2005 SP1 or VS2008.

Could you post a complete, compilable example that exhibits the problem?

RichardRudek
So what do you expect for nothin'... :P

[Posted to Sandbox]

Thanks.

RichardRudek
So what do you expect for nothin'... :P

Ah, found it.

I tried to cheat by having the same class name, but defined in separate compilation units. I suddenly realised that these should have been in their own namespaces, and after doing so, the linker doesn't remove the 'redundant' classes. D'oh !

Sven Groot
You can't have everything; after all, where would you put it?

You are violating the one-definition-rule. You have two classes named Testable with external linkage but with different implementations. To allow class definitions to be placed in headers the linker will not complain if a class violates the one-definition-rule, but if all definitions are not the same the behaviour is undefined.

In this case, the linker has chosen the Testable from OperatorBool.cpp so the one from OperatorNot.cpp is also using that version.

If you wish to have different classes with the same name in files that are linked together, you must give the classes internal linkage. In C++ this is done by placing them in an anonymous namespace:

namespace
{
   class Testable
   {
      // stuff
   };
}

If I make that change your example works correctly.

EDIT: Good to know you figured it out yourself. Smiley

RichardRudek
So what do you expect for nothin'... :P

Thanks for helping with that Sven Smiley

You know, I had someone suggest to me that the reason why I'm 'more productive' programming late at night is because my brain is half-asleep... Wink