-
Notifications
You must be signed in to change notification settings - Fork 67
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Specify version and abi information in "linking" section header #56
base: main
Are you sure you want to change the base?
Conversation
|
||
The `abi_version` field specifies the version of the ABI in use. This is used | ||
to distinguish between incompatible versions of the same ABI. The current | ||
version of `WASM_ABI_CXX` is `1`. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'm concerned about mixing C++ in with C here. The C++ ABI is likely to evolve on its own schedule, with things like using wasm tables to implement vtables, wasm exception handling, coroutines, and it would be unfortunate to impose ABI version bumps on people using the C ABI when these things happen. So at the very least, my suggestion is to at least have a field for just the C ABI.
But also, looking ahead to exception handling -- it would be nice if C++ users that aren't using exceptions at all don't have to worry about incompatibilities when EH happens. Their code should be compatible with both setjump-longjmp EH code and wasm-EH code. And other C++ features may end up in similar predicaments. So brainstorming, what if we made the ABI field be an array of key-value pairs? The keys could be named integers (WASM_ABI_*
), and the values could be whatever makes sense for a given key. A binary could declare something like WASM_ABI_LINKING=1, WASM_ABI_C=1, WASM_ABI_C++ABI=ITANIUM, WASM_ABI_C++ABI_VERSION=1, WASM_ABI_EH=false
or something. Thoughts?
Hmm, I'd rather not increase the complexity that much. We had discussed a single version string, since main use case is simply to prevent linking incompatible modules. The scheme I'm proposing here is a little more explicit. An array if key-value pairs seems most complex.. but I could be persuaded. In terms of the C++ features like EH, I imagine that would be a bit in the ABIFlags, no? Wouldn't that work? In terms of the cost of C users of combining the C and C++ abis into one, I don't think that having to rebuild your objects when the C++ abi revs is very big cost. |
The cost isn't just rebuilding your own objects. It's also getting objects that you didn't build yourself rebuilt. Which isn't a big deal for some users, but is for others. And in a mixed-language world, it includes getting all your compilers to agree on which version to use. I also like keeping C and C++ separate because while C is well-established as a base language for communicating between different high-level languages, and between libraries that want to maintain stable interfaces (eg. LLVM's C API), C++ isn't, so I don't see a reason to give C++ a special status here. Key-value pairs would let us put a C++ ABI field in there (or two, one for language and one for standard library, if we wanted that), and also leave room for other languages to define their own ABI fields. But they are also more complex to decode and implement. |
I've been thinking about it and I think I agree with you after all :) I will update the PR. BTW I kind of like the version info being in the header of the section rather than in one a subsection. What do you think about doing this? |
I'm still having trouble seeing how this extra information would be used. The simplest thing is to say that all fields have to match exactly, but that is no better than just having a single version number (right?) So I assume we're talking about allowing objects with slightly different flag fields to link together, but I'm not sure I understand what the rules would be. Is it just ad-hoc? |
What's the target for this ABI info - is it for describing static linking, or dynamic, or both? Dynamic linking is in a sense a smaller thing, although it's bigger because it's not yet nailed down at all. Whereas if we're targetting static linking, then the ABI can just be a single string. Making the C++ ABI stable across different dynamically-linked objects is probably a stretch goal at this point, although I guess the ABI-string should be able to specify it. We aren't expecting to bump the ABI often, I wouldn't have thought... |
My understanding is that static and dynamic linking will have the same requirements. i.e.: they both need to answer the question: "Can the code in one object inter-operate with the code in another". If two DLLs written in C++ need to inter-operate then the will need to agreee on the C++ ABI. I suppose they could inter-operate via the C ABI despite being written in C++? And yes, the idea is the change these ABIs as infrequently as possible. I'm still not sure exactly how much detail we want to include there. For most existing platforms, its kind of left up to the programmer to ensure the abi matches. For example I don't think ELF includes this level of detail. Hence I decided to split of the versioning of the metadata which is less controversial. |
Fixes #54