Technical details discussion for SEH

Hi Jb,

It’s good to see someone step up and take a shot as this again. I dropped this because it seems to me it wasn’t a high priority task for LLVM/Clang.

Implementing SEH is more of LLVM work than Clang work.

For 32-bit SEH, there are prologue/epilogue instruction sequence to emit, setting try-level ([ebp-4]), recovering EBP ([ebp-18h]), and all these can only happen in LLVM, not Clang. In my opinion, we should implement these as LLVM intrinsics, like the gcc ones: see Also, we should emit the tables for each function, which is not a simple task either.

For 64-bit SEH, things are simpler. There are no more instructions to emit than no-SEH code. We have only two tasks: emit the table for each function, and place some code right. If I remember it correctly, __finally block code should be replicated at 2 places: one in the original function as part of the normal execution path, and other one separately if anything in __try block goes south.
As for the table part, I’ve seen some commits from Kai(as CC’ed in this email) doing it, you should ask him for details.

Thanks again.

Hi Jb, Hi Tong,

with my patch LLVM emits unwind information for Windows 64bit. The Dwarf EH encoding is used language specific data. That is the same way gcc implements SEH for Windows 64bit. (As a side note, PR18546 now contains a patch for Clang to use my patch on mingw64.)

The rational behind this approach is that Windows provides support for stack unwinding (RtlUnwind etc.) but LLVM is inherently based on Dwarf EH. This approach combine both worlds. If the personality function is tolerant enough then it should also be possible to mix exceptions.

If you want to make exception handling MS compatible then you can take my code as base. You only need to emit the MS handler data instead of the Dwarf data.

However, at least as a first approach I would recommend to implement 32bit SEH with a similar approach. It should have some similarities to SjLj exception handling - this really helps to understand what needs to be done.

"real SEH handling" requires a small function which is called to decide if the exception is handled. If yes then the handler is called. If not then stack unwinding continues (after a possible call of a cleanup handler). I found it difficult (if not impossible) to create this code based on the exception design of LLVM. IMHO, some some generalizations are required.


Can you clarify what you mean by “real SEH handling”? My company has me looking at this in the hopes that I can make LLVM capable of building windows drivers. If you mean “visual c++ style SEH”, I’m fairly sure that isn’t necessary for my purposes, but it would be nice. If you mean “something that works at all,” then your concern about generalizing LLVM exception handling probably means that I will need to do some work learning about how exception handling is currently implemented. Do you have a link to the patch your discussing, or a revision number?


Short version: “something that works at all” - yes, “visual c++ style SEH” - no.

Long version: this whole SEH mechanism breaks down to two levels, actually: operating system level and compiler level.
For the operating system level, after OS catches any software/hardware exception, it just calls RtlUnwind to walk the stack and find the first stack frame that may handle this exception and execute handler defined in that stack frame, until exception is handled or we have to abort the whole program. And when the handler is called, a pointer to language-specific exception data is provided. On the OS level, OS doesn’t care what the handler is, and is not responsible for providing the language-specific data to the exception handler; it just calls the handler.
The handler and language-specific data passed to exception handler are decided on the compiler level. For example, for C programs, the handler is _except_handler3 in CRT, for .NET program, it’s some function in .NET runtime.

LLVM can emit code that runs fine on operating system level. The handler is provided by GCC runtime, and language-specific data is Dwarf EH data. This way we can make Dwarf style C++ exception work (on top of operating system level SEH). But we cannot link llvm-generated code and VC-generated code together, because C++ exception model is different.

For VC-style SEH, it uses _except_handler3 in CRT as exception handler(for every stack frame that is compiled with SEH turned on), and use extra stack space as language-specific data structure passed to _except_handler3. This is not implemented in LLVM.

Hope this explains something for you. IMHO, writing Windows driver using an other-than-VC compiler seems a lot of trouble… Is there some feature of clang that you really want and is missing in VC?


If LLVM can emit code that runs ok from the OS, I think what I’m trying to do should be doable, if frustrating to implement. Can you point me to the code that handles this functionality? I haven’t poked around much, but I can tell from reading the exception handling spec that this isn’t the default behavior. Unless I’m seriously missing things, SEH requires placing exception handlers on the stack along with the code that they handle, the fact that everything calls into the same _except_handler3 function is simply a matter of implementation, but the OS needs the address of the first handler to be in fs:[0], which did not sound like the exception handling behavior described in the documentation.

The answer to your other question is yes, LLVM transforms!

Hi Jb,

with "real SEH handling" I mean "visual c++ style SEH".
With my patch it is possible to use the Windows SEH machinery but not with the _c_specific_handler from MSVC. You have to provide your own personality function which must work with the Dwarf encoding.

My patch is always linked to this wiki page: Just look for the section "Required source downloads".
An implementation of a personality function in D is here:
You can also look at libgcc on mingw64 and search for the implementation of __gcc_personality_seh0. That is the stuff which is supported by my patch.

I also think that the same mechanism can be implemented with 32bit SEH.