[cfe-dev] GSoC Proposal: Virtual Tables and RTTI in the Microsoft Visual C++ ABI

[Alexander Potapenko]

On Thu, Apr 5, 2012 at 5:29 PM, Charles Davis <cdavis at mymail.mines.edu> wrote:
> Hi,
>
> Here is another of my GSoC proposals (the first was on LLVMdev). This seems somewhat small for a GSoC project. Is there something I'm missing?
>
> Chip
>
> Project Title: Virtual Tables and RTTI in the Microsoft Visual C++ ABI
>
> Abstract:
> Two years ago, I performed work to support multiple C++ ABIs in Clang, and started work to support the Microsoft Visual C++ ABI. This project is a continuation of that work, adding support for virtual tables in the Microsoft ABI. This is necessary before certain other work (RTTI, dynamic_cast, exceptions) can be performed.
>
> Proposal:
> Two years ago, in a Summer of Code project, I added support for multiple C++ ABIs to Clang, and began work on adding a second ABI to exercise this functionality: the Microsoft Visual C++ ABI. Unfortunately, implementing the whole ABI is far too much for one project. Therefore, this project proposes implementing a very important piece: emission of virtual tables.
>
> Some pieces of this are already in place. A contributor, Dmitriy "r4start" Sokolov, kindly added some code to lay out the virtual table pointers in each class. Unfortunately, that work is unfinished, because he has been busy with his own life. This project will pick up where he left off.
>
> In the Microsoft C++ ABI, there are actually two virtual tables: a function table and a base table. The function table is used to call virtual methods; the base table is used to access virtual bases. This project will implement both kinds of v-table.
>
> One of the difficulties in this is ensuring the right table is used for a lookup. Currently, Clang still assumes it is compiling for the Itanium C++ ABI when it compiles C++ code with virtual methods/bases--which has a very different idea of how to implement support for v-bases and virtual methods. The various operations that are performed--v-table method pointer lookup, virtual base lookup--must therefore be abstracted out into the C++ ABI class I added in the first project.
>
> Another is that, for ABI compatibility, the v-table entries must be emitted in the correct order. This will require research into how the Microsoft compiler lays out v-tables. It may also require another method in the C++ ABI object to lay out the virtual tables.
>
> A third might be the use and emission of hidden "vtordisp" fields, which are used when a virtual method in a virtual base is called from a 'structor. Dmitriy has performed work in this area; his patch may help here.
>
> Like the Itanium ABI, the Microsoft ABI sometimes requires the use of virtual thunks which adjust the 'this' pointer before calling the method proper. This proposal will also implement emission of such thunks.
>
> If there is time, I may also begin work on emitting RTTI structures. These are officially undocumented, but luckily the Wine developers have worked out their definitions for their implementation of the Visual C++ runtime. We can use this to implement emitting the RTTI structures. Dmitriy also says he has patches to implement this himself, but he has not sent them to the list yet.
>
> Proposed Timeline:
> Week 1-2: Implement laying out Microsoft ABI v-tables and mangling of symbols.
> Week 3: Implement virtual method calls.
> Week 4: Implement virtual thunks.
> Week 5-6: Implement virtual base loads and conversions.
> Week 7-8: Finish vtordisp support.
> Week 9-10: Implement emission of RTTI descriptors.
> Week 11-12: Implement typeid() and dynamic_cast<> support.
>
> Contact:
> email: cdavis (at) mymail (dot) mines (dot) edu
> phone: (seven-one-nine) 963-4781
> IRC: cdavis5x at freenode, cdavis5x at oftc
>
> Interest:
> I have long been interested in figuring out how the Microsoft ABI works, and also in compiling libraries that need to use the Microsoft ABI for Wine.
>
> Use to LLVM:
> Supporting the Microsoft ABI will increase adoption of LLVM and Clang, particularly on Windows, where a certain compiler is most commonly used. In addition, supporting the Microsoft ABI will also make Clang more useful to projects like Wine that need to support the ABI, and might even spur GCC into supporting this.
>
> Experience:
> I have worked with LLVM for over two and a half years. I implemented mangler support for the Microsoft C++ ABI back in GSoC 2010. I have also implemented force_align_arg_pointer support, fixed a number of small bugs in Clang with attributes and calling conventions, and began work to support the Win64 exception-handling model. I have also contributed patches to Wine, mostly in the area of support for Mac OS.
>


I think Timur was investigating Microsoft ABI recently.