Thread overview
Bug on C/C++ library to D, parameters suddenly becomming null or invalid
4 days ago
Hipreme
4 days ago
Avrina
2 days ago
Hipreme
2 days ago
Mike Parker
4 days ago
I found some really strange problem happening when trying to create a binding between
cimgui (https://github.com/cimgui/cimgui), which is a automatically generated binding from the main C++ project (https://github.com/ocornut/imgui)

While trying to create a binding, everything worked until I wanted multiviewports.

The problem was caused by functions pointers members from a struct.

Inside the cpp code, I saw that the calling didn't change the viewport instance, but when arriving at the D code, the instance would be transformed in invalid.


The reference code I'm saying is:

Take a look onto

https://github.com/ocornut/imgui/blob/682249396f02b8c21e5ff333ab4a1969c89387ad/imgui.cpp#L11120

This line executes just correctly the viewport I just created on newviewport call from my D code.


While that line:

https://github.com/ocornut/imgui/blob/682249396f02b8c21e5ff333ab4a1969c89387ad/imgui.cpp#L11134

References some of the same viewports, but it strangely became invalid memory, while the line from getminimized are all working.

Things got even stranger when I compiled a DLL for not having to use my SDL implementation, only the OpenGL which I can't change because the OpenGL loader is the bindbc one.

This is a image from my debugger, where the program had break:

https://user-images.githubusercontent.com/10136262/96338382-79a9f700-1042-11eb-8f0f-f0d8155cd82e.png


As you can see, viewport is 0x000000 while void* param has some value.
While the c++ code did pass a viewport to the first parameter, as seem in:

https://github.com/ocornut/imgui/blob/682249396f02b8c21e5ff333ab4a1969c89387ad/imgui.cpp#L11634

This bug was mentioned in the cimgui repo, take a look as the reference:
https://github.com/cimgui/cimgui/issues/156

I solved this bug by casting void* param to ImGuiViewport, so, the first parameter was actually null while the second would be actually the first parameter.

The func prototype is correctly declared, although is a struct with a member callback which is assigned inside some D function.
It is a lot to read, but this bug is really strange for me and I don't know what to do else
4 days ago
On Saturday, 17 October 2020 at 18:03:45 UTC, Hipreme wrote:
> I found some really strange problem happening when trying to create a binding between
> cimgui (https://github.com/cimgui/cimgui), which is a automatically generated binding from the main C++ project (https://github.com/ocornut/imgui)
>
> While trying to create a binding, everything worked until I wanted multiviewports.
>
> The problem was caused by functions pointers members from a struct.
>
> Inside the cpp code, I saw that the calling didn't change the viewport instance, but when arriving at the D code, the instance would be transformed in invalid.
>
>
> The reference code I'm saying is:
>
> Take a look onto
>
> https://github.com/ocornut/imgui/blob/682249396f02b8c21e5ff333ab4a1969c89387ad/imgui.cpp#L11120
>
> This line executes just correctly the viewport I just created on newviewport call from my D code.
>
>
> While that line:
>
> https://github.com/ocornut/imgui/blob/682249396f02b8c21e5ff333ab4a1969c89387ad/imgui.cpp#L11134
>
> References some of the same viewports, but it strangely became invalid memory, while the line from getminimized are all working.
>
> Things got even stranger when I compiled a DLL for not having to use my SDL implementation, only the OpenGL which I can't change because the OpenGL loader is the bindbc one.
>
> This is a image from my debugger, where the program had break:
>
> https://user-images.githubusercontent.com/10136262/96338382-79a9f700-1042-11eb-8f0f-f0d8155cd82e.png
>
>
> As you can see, viewport is 0x000000 while void* param has some value.
> While the c++ code did pass a viewport to the first parameter, as seem in:
>
> https://github.com/ocornut/imgui/blob/682249396f02b8c21e5ff333ab4a1969c89387ad/imgui.cpp#L11634
>
> This bug was mentioned in the cimgui repo, take a look as the reference:
> https://github.com/cimgui/cimgui/issues/156
>
> I solved this bug by casting void* param to ImGuiViewport, so, the first parameter was actually null while the second would be actually the first parameter.
>
> The func prototype is correctly declared, although is a struct with a member callback which is assigned inside some D function.
> It is a lot to read, but this bug is really strange for me and I don't know what to do else



You have to look at more than the declaration. The ABI depends on the types being passed as well. Unfortunately C++/C aren't direct translations with D and you hav to sometimes do hacks so that a D type is passed the same way as a C++ type. I'd investigate whether the type is POD or not. By the looks of it, it is being passed as a register.
2 days ago
On Saturday, 17 October 2020 at 20:43:03 UTC, Avrina wrote:
> On Saturday, 17 October 2020 at 18:03:45 UTC, Hipreme wrote:
>> [...]
>
>
>
> You have to look at more than the declaration. The ABI depends on the types being passed as well. Unfortunately C++/C aren't direct translations with D and you hav to sometimes do hacks so that a D type is passed the same way as a C++ type. I'd investigate whether the type is POD or not. By the looks of it, it is being passed as a register.


So, I have finally solved the issue:

I asked for help to the binding mantainer and it seems that when the C++ file gets compiled on MSVC it turns into a non-POD struct, while on MinGW and GCC it turns into a POD, so it was quite strange, it was solved by creating a function on the C side which would receive a pointer to this POD-mutating struct instead of returning and assigning its value on this _out parameter.

The void* being the real parameter was related for some missing extern(System)
2 days ago
On Monday, 19 October 2020 at 17:49:24 UTC, Hipreme wrote:
> when the C++ file gets compiled on MSVC it turns into a non-POD struct, while on MinGW and GCC it turns into a POD, so it was quite strange,

As a general rule, you shouldn't mix binaries from MinGW and MSVC. Inconsistencies like this are bound to arise. DMD is implemeted to match MSVC. It's fine when loading C dlls dynamically, but as soon as you bring a linker or the C++ API into the picture, you'll hit issues eventually.