QCoreApplication::instance() is NULL

[pullmoll]

Hi forum,
after having migrated our application from Qt4 to Qt5, it works fine so far on Linux with Qt5.

On Windows with qt-opensource-windows-x86-msvc2010_opengl-5.4.2 being installed, however, it does not.

The culprit is that seemingly QCoreApplication() isn't initialized, judging by the output of qDebug in main():

int main(int argc, char *argv[])
{
    QApplication app(argc, argv);
    qDebug("self=%p", app.instance());
...
}

On Linux it prints some pointer value, while on Windows it prints 0x0.

The problem became obvious when qApp->applicationDirPath() would return an empty string and print the warning QCoreApplication::applicationDirPath: Please instantiate the QApplication object first

The question is: What could possibly cause the QCoreApplication constructor to not run through QCoreApplication::init() to set self = this?

Creating a new, empty application with QtCreator and printing the app.instance() pointer works as expected, so it has to be one of the things that is in the headers included in main.cpp which causes this. The project is large and there are literally dozens of headers, if not hundreds.

Any hints would be greatly appreciated.

[SGaist]

Hi and welcome to devnet,

Do you have any static widget ?

[alex_malyu]

Is qDebug("self=%p", app.instance()); displays notification?
Simply put breeakpoint in a debugger a make sure and make sure
QApplication app(argc, argv);
line is executed before you see notification

If not you probably have QObject subclass somewhere instantiated before QApplication (global variable, etc)
If yes mostly likely libraries your application is linked to and tries to load at startup are different ( check path etc )

[SGaist]

I wonder if it can be related to this bug report although it happens on linux in that case

[pullmoll]

I have been to the hospital, thus I can answer only now.
@SGaist: static QWidgets No, I don't have static QWidgets AFAICT.
I wrote code to debug every constructor in our project to see what happens before the QApplication is instantiated and there is nothing relevant, i.e. no QObject is created before the QApplication. There are some QObject derived classes instantiated inside the private members of the QApplication derived class, though. But this should be perfectly valid, no?

@alex_malyu: The qDebug() is executed after the QApplication app(argc,argv); is done, yes. I can inspect the app in the debugger's variable view. Here is an excerpt of the debugger panel:

app	class QApplication	QApplication
  QGuiApplication	class QGuiApplication	QGuiApplication
    QCoreApplication	class QCoreApplication	QCoreApplication
      QObject	class QObject	QObject
        __vfptr	0x661787f4	__fptr() [11]
          [0]	0x65ecd020	__fptr()
          [1]	0x65eccae0	__fptr()
          [2]	0x65ed60d0	__fptr()
          [3]	0x65ed1920	__fptr()
          [4]	0x65ecf0c0	__fptr()
          [5]	0x66171a4c	__fptr()
          [6]	0x66171a52	__fptr()
          [7]	0x66171a58	__fptr()
          [8]	0x66171a5e	__fptr()
          [9]	0x66171a64	__fptr()
          [10]	0x66171a6a	__fptr()
        staticMetaObject	<Static const member value not available error>	<Static const member value not available error>
        d_ptr	class QScopedPointer<>	QScopedPointer<QObjectData,QScopedPointerDeleter<QObjectData>>
          d	0xa34b70	QObjectData *
            __vfptr	0x66178790	__fptr() [1]
            q_ptr	0x15fe60	QObject *
            parent	0x0	QObject *
            children	<3 Elemente>	QList<QObject *>
            isWidget	0	unsigned int
            blockSig	0	unsigned int
            wasDeleted	0	unsigned int
            isDeletingChildren	0	unsigned int
            sendChildEvents	1	unsigned int
            receiveChildEvents	1	unsigned int
            isWindow	0	unsigned int
            unused	24468448	unsigned int
            postedEvents	0	int
            metaObject	0x0	QDynamicMetaObjectData *
        staticQtMetaObject	<Static const member value not available error>	<Static const member value not available error>
      staticMetaObject	<Static const member value not available error>	<Static const member value not available error>
      self	<Static const member value not available error>	<Static const member value not available error>
    staticMetaObject	<Static const member value not available error>	<Static const member value not available error>
  staticMetaObject	<Static const member value not available error>	<Static const member value not available error>
argc	1	int
argv	0xbb3d18	char * *

It irritates me that the static member self value is not available according to the debugger.

However, I now have to deal with an entirely different area in the project and can't try more things to tackle the problem, so I thank you for your thoughts and hints for now.

[alex_malyu]

Have you checked executable dependencies?

Make sure debug version is linked against debug Qt libraries and release against release libraries.
This is important you can't mix them, cause they use different c runtime.

[pullmoll]

The executable dependencies are fine. I have clearly separated the (out of source) build directories and also all the libraries our project links against have separate debug and release versions.

Now I tried with MSVC2013 community edition and Qt5 5.5.1 to see if the problem was perhaps compiler related or fixed. Unfortunately it is not.

I wish there was a way to single step through the startup in disassembler mode to see why the generated code misses to run through the constructor QCoreApplication::QCoreApplication(...) which is where self is assigned a value.

I'm now thinking of compiling Qt5 5.5.1 myself and inserting qDebug() messages into the constructors of QApplication, QGuiApplication and QCoreApplication to see if they are executed at all.

Any further hints or ideas are still very welcome, as this problem is a real showstopper and Qt4 is no longer maintained.

[pullmoll]

@SGaist said:

I wonder if it can be related to this bug report although it happens on linux in that case

I certainly looks like it is related, because it is the same effect that I see happening.

FWIW In the very first line of code in main() of main.cpp:

• instantiating a QCoreApplication app(argc, argv); results in app.instance() being non-NULL.
• instantiating a QGuiApplication app(argc, argv); results in app.instance() being NULL.

It seems as if the QCoreApplication that QGuiApplication inherits from simply isn't initialized.

#ifdef Q_QDOC
QGuiApplication::QGuiApplication(int &argc, char **argv)
#else
QGuiApplication::QGuiApplication(int &argc, char **argv, int flags)
#endif
    : QCoreApplication(*new QGuiApplicationPrivate(argc, argv, flags))
{
    d_func()->init();

    QCoreApplicationPrivate::eventDispatcher->startingUp();
}

The parameter passed to QCoreApplication in the initializers here looks a little strange to me, because the QCoreApplication constructor expects int& argc, char ** argv, int flags. But perhaps this is just my lack of understanding of C++.

I guess this is using the copy constructor, but then there is a Q_DISABLE_COPY(QCoreApplication) in qcoreapplication.h ...

[pullmoll]

FWIW I now managed to single step into the code right from the start.

The constructor QApplication::QApplication calls the constructor code of QApplicationPrivate, which initializes self in the last step, and which looks like this in the disassembler:

        Qt5Widgets!QApplicationPrivate::QApplicationPrivate:
0x64d8c3d0                    push    esi
0x64d8c3d1  <+0x0001>         push    dword ptr [esp+10h]
0x64d8c3d5  <+0x0005>         mov     esi,ecx
0x64d8c3d7  <+0x0007>         push    dword ptr [esp+10h]
0x64d8c3db  <+0x000b>         push    dword ptr [esp+10h]
0x64d8c3df  <+0x000f>         call    dword ptr [Qt5Widgets!QVBoxLayout::tr+0x58a4 (6502fef4)]
0x64d8c3e5  <+0x0015>         mov     dword ptr [esi],offset Qt5Widgets!QApplicationPrivate::`vftable' (65031590)
0x64d8c3eb  <+0x001b>         lea     ecx,[esi+0A0h]
0x64d8c3f1  <+0x0021>         mov     eax,dword ptr [Qt5Widgets!QVBoxLayout::tr+0x3e0c (6502e45c)]
0x64d8c3f6  <+0x0026>         mov     dword ptr [esi+9Ch],eax
0x64d8c3fc  <+0x002c>         call    dword ptr [Qt5Widgets!QVBoxLayout::tr+0x3ccc (6502e31c)]
0x64d8c402  <+0x0032>         lea     ecx,[esi+0A4h]
0x64d8c408  <+0x0038>         call    dword ptr [Qt5Widgets!QVBoxLayout::tr+0x3ccc (6502e31c)]
0x64d8c40e  <+0x003e>         lea     ecx,[esi+0A8h]
0x64d8c414  <+0x0044>         call    dword ptr [Qt5Widgets!QVBoxLayout::tr+0x3ed4 (6502e524)]
0x64d8c41a  <+0x004a>         lea     ecx,[esi+0B0h]
0x64d8c420  <+0x0050>         call    dword ptr [Qt5Widgets!QVBoxLayout::tr+0x3ed4 (6502e524)]
0x64d8c426  <+0x0056>         lea     ecx,[esi+0B8h]
0x64d8c42c  <+0x005c>         call    dword ptr [Qt5Widgets!QVBoxLayout::tr+0x3ed4 (6502e524)]
0x64d8c432  <+0x0062>         mov dword ptr [esi+0C0h],0
0x64d8c43c  <+0x006c>         mov dword ptr [esi+0C4h],0
0x64d8c446  <+0x0076>         mov     dword ptr [esi+58h],1
0x64d8c44d  <+0x007d>         mov dword ptr [esi+0C8h],0
0x64d8c457  <+0x0087>         mov dword ptr [esi+0CCh],0
0x64d8c461  <+0x0091>         mov     eax,dword ptr [Qt5Widgets!QApplicationPrivate::self (65181dc0)]
0x64d8c466  <+0x0096>         test    eax,eax
0x64d8c468  <+0x0098>         cmove   eax,esi
0x64d8c46b  <+0x009b>         mov     dword ptr [Qt5Widgets!QApplicationPrivate::self (65181dc0)],eax
0x64d8c470  <+0x00a0>         mov     eax,esi
0x64d8c472  <+0x00a2>         pop     esi
0x64d8c473  <+0x00a3>         ret     0Ch

Note: the absolute address of the static member self here is 0x65181dc0.
The code then continues in the constructor - calls into QApplicationPrivate::construct() - and finally returns to my main().
There the next thing that happens is the call to QCoreApplication::instance() in my qDebug()

0x192ec2  <+0x0042>         call    dword ptr [imageviewer!_imp_?instanceQCoreApplicationSAPAV1XZ (004b7c08)

and the code there looks like this:

        Qt5Cored!QCoreApplication::instance [c:\work\build\qt5_workdir\w\s\qtbase\src\corelib\kernel\qcoreapplication.h @ 110]:
0x66b53de0                    push    ebp
0x66b53de1  <+0x0001>         mov     ebp,esp
0x66b53de3  <+0x0003>         mov     eax,dword ptr [Qt5Cored!QCoreApplication::self (67367ddc)]
0x66b53de8  <+0x0008>         pop     ebp
0x66b53de9  <+0x0009>         ret

Note: the absolute address of self here is 0x67367ddc, which is obviously entirely different from the above 0x65181dc0.

So something in my executable is really - and I mean really - borked.

What irritates me is that the debugger lists the library name Qt5Widgets!QApplicationPrivate::QApplicationPrivate and not Qt5Widgetsd!QApplicationPrivate::QApplicationPrivate (the debug version).
I would not know why this could happen.

However, if I look at the executable with depends.exe it indeed lists both, Qt5Widgets.dll and Qt5Widgetsd.dll. Now I have to track down how and why the executable refers to both, non-debug and debug versions, of Qt5Widgets.

[SGaist]

What version of Windows are you running ?

Do you have any plugin from an external source in your application ?

[ambershark]

Do you have any Qt global variables that could be initialized before your main? I had that exact problem and my issue was a global Qt variable.

I can't remember the exact type, but let's just say a QObject, that was not a pointer and therefore had initialization before the main() and QCoreApplication.

It was also in a .so outside of the main application, which may have contributed to the issue as well.

[pullmoll]

@SGaist said:

What version of Windows are you running ?

Windows 7 SP1

Do you have any plugin from an external source in your application ?

No, but we have our own QImageReader/Writer plugins. They are however not the source of the issue.

I identified the module (it's a submodules project with abt. 20 modules) which actually wants to link against both, debug and release versions of Qt5Widgets.

First I found, by running depends.exe, that it seemed to be QWidget::winId() that was linked imported from Qt5Widgets instead of Qt5Widgetsd. I disabled its use and now this specific module does not any longer want to import from Qt5Widgets (release). The depends.exe tool lists the symbols which are imported from a specific DLL.

However I still have the same crash. I'm investigating and will certainly find the source of the problem sooner or later. The dependency walker is a great help in this case (http://www.dependencywalker.com/)

[pullmoll]

@ambershark said:

Do you have any Qt global variables that could be initialized before your main? I had that exact problem and my issue was a global Qt variable.

I can't remember the exact type, but let's just say a QObject, that was not a pointer and therefore had initialization before the main() and QCoreApplication.

It was also in a .so outside of the main application, which may have contributed to the issue as well.

Hi ambershark,

as I described above, I already tracked down the reason why the executable doesn't work: linking against both, debug and release version of Qt5Gui and Qt5Widgets. Now I only need to find the reason why it happens, which is much easier now the core problem was tracked down.

[pullmoll]

News: I can't seem to get rid of the dependencies to Qt5Widgets, Qt5Gui and Qt5Core in my executable, even after excluding a lot of submodules and their headers which could have been the cause. The symbols listed in depends.exe for the release DLLs are not giving me any hint where the import results from.

Now comes my general question: what could possibly cause a debug build to refer to the release versions of the Qt5 DLLs? As far as I understand Qt, it would have to be some header file included without QT_DEBUG being defined. Is there perhaps a chance to closely inspect the generated Makefiles to see why and where this is happening?

Despite the fact I now know the reason for the problem, I can't find the cause, i.e. why does my executable being built in debug mode want to link against Qt5 release DLLs. I'm stumped.

[sierdzio]

Add

QT -=core gui widgets

to your .pro file?

[pullmoll]

@sierdzio said:

Add

QT -=core gui widgets

to your .pro file?

I first thought you were kidding me, but this even works for my main *.pro file.
This seems a little.. umm.. counter-intuitive and the opposite of what the the help for the QT qmake variable suggests, no!?

[sierdzio]

Not sure what you mean.

By default, QT contains both core and gui, ensuring that standard GUI applications can be built without further configuration.

So, if you don't want to link them, remove them. As I understand from (brief, I admit) look at your previous comments, you aim to remove all traces of Qt libs from the project, to pinpoint the place where the release libs are injected.

I know this might be a stupid approach, but when everything else fails... :-)

[pullmoll]

@sierdzio said:

Not sure what you mean.

By default, QT contains both core and gui, ensuring that standard GUI applications can be built without further configuration.

Yeah, sure, this has been the case since Qt4 AFAICT. I mean that removing "core gui widgets" from a main *.pro file where QMainWindow and QWigets etc. are used is counter-intuitive.

@sierdzio said:

So, if you don't want to link them, remove them. As I understand from (brief, I admit) look at your previous comments, you aim to remove all traces of Qt libs from the project, to pinpoint the place where the release libs are injected.

This is a good idea at least! I would never have thought of trying to remove everything from QT execpt what is really needed (e.g. xml, network, ...).

I know this might be a stupid approach, but when everything else fails... :-)

Thank you for the hint. It may really help me find the bad guy :-)

[pullmoll]

The issue is solved!

After an odyssey of trial and error, the culprit is or was static libraries!

Within our project there were two submodules built as static libraries (CONFIG += static resources create_prl) and linked to several other submodules. The static libraries were built with the same project settings, i.e. DEBUG enabled, and still they were the root cause of my problems.

It has to be either the link_prl in the project, or generally linking against static libraries which causes our project to import the unwanted non-debug Qt5-DLLs in a debug build.

After converting the static libraries to DLLs with export symbols now everything works as expected.

I don't know if this problem and its workaround should be considered a Qt5 bug, or if it is my fault because of not having read some documentation part mentioning this specific fact?

Thanks for all your help and cheers!
Jürgen

[sierdzio]

Nice.

You might want to ask this on Qt development mailing list, the Qt core developers could probably comment on that weird release symbol linking.