[LLVMdev] [RFC] Developer Policy for LLVM C API

[Jim Grosbach]

> On Jul 20, 2015, at 1:08 PM, Eric Christopher <echristo at gmail.com> wrote:
> 
>> 
>> Part (most?) of the point of having a stable API is as a way of decoupling the development processes of two separate projects (modulo well-documented release-to-release updating). Requiring our users to add tests in our tree doesn't really achieve much decoupling.
> 
> I’m not sure there is much “coupling” here. The point is that we expose a C API that is supposed to be stable but is not well tested. And some part of the C API is just a wrapper around the C++ and hasn’t really been designed to be “stable” in time.
> It seems also that we don’t really know what part of the C API really needs to be stable and is important for the users, so I read Pete’s proposal as “let’s collect the current use-cases and make them tests in LLVM, so that we define what is part of the stable C API and so that we won’t (inadvertently) break valid use cases".
> 
> 
> Yeah, this is just terrible though for all of the reasons I raised in my email and as you even say here "And some part of the C API is just a wrapper around the C++ and hasn’t really been designed to be “stable” in time."
>  
> 
> As of having this in-tree or out-of-tree, I’m not sure about that and there is a trade-off.
> 
> 
> I'm regretting ever saying "out of tree" here as I don't think it's the main issue, rather the splitting of the "bindings" style of api that we see a lot of in the C API directory and the more solid ones that we see from libclang and liblto.

I agree with this. There’s a lot of muddy waters surrounding the C API right now. “Stable API” vs. “bindings” is a great clarifying question. Out-of-tree or not doesn’t matter too much to me, personally. It was a useful turn-of-phrase to jump-start this aspect of the conversation, but pretty much all of the deeper questions apply equally regardless of which implementation is chosen.

I think you, and others who’ve observed the same basic thing, have hit it on the head that our current C API has morphed in definition over time from a very small, restricted API for very specific use-cases to something broader and more general purpose. It’s done so in a gradual and mostly organic way along the lines of “Hey, I need this widget and can’t get to it from C, so I’ll add an API to get to it.” As a consequence, it’s also morphed from a very (intentionally) stable API to something a bit more fluid. Exactly how fluid is unclear, thus the various breakages, discussions about how to deal with them, etc.. We started with something that was very much a stable API, but now we have some sort of weird mix of API and bindings, with no clear delineation of which is which or even which is intended to be which.

I see value in having C APIs for both classes of use cases, but they should be separated out from one another. I strongly agree that especially for the stable API, we need very robust test cases which test not only that the API is maintained from release to release, but that we have a well defined, well documented and well tested set of semantics for that API. Basically, we need to stop doing this by half measures and actually have a real stable API.

To get there, we’ll need to have two sets of headers, one for bindings and one for APIs. I suggest we start by cloning what exists today for both, with the bindings API moving to a new location. Then the bindings side can start changing (or not) as much or as little as those consumers want without muddying the stable API waters. On the stable API side, we can work on a strategy for deprecating those APIs we currently have that aren’t a great fit for a stable API and replacing them with new ones (if appropriate).

-Jim

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150720/24565095/attachment.html>