[LLVMdev] Question about changes to llvm::Argument::addAttr(AttributeSet AS) API

[David Chisnall]

On 6 Feb 2013, at 07:50, Bill Wendling wrote:

> On Feb 4, 2013, at 11:54 PM, David Chisnall <David.Chisnall at cl.cam.ac.uk> wrote:
> 
>> On 5 Feb 2013, at 01:32, Bill Wendling wrote:
>> 
>>> No. It hasn't been written up. We typically don't do write-ups for API changes. However, we do list the thing we do change in the ReleaseNotes (these changes haven't made it there though).
>> 
>> The attributes API has undergone a horrendous amount of churn over the last few months, both before and after the 3.2 release.  I've lost track of the number of times I've rewritten code interacting with this API over the past year and I now have a mess of #ifdefs just to support 3.1, 3.2 and trunk.  Each time, the only warning I got was that my code stopped building with trunk, and the only way of knowing how to rewrite it was to read the attributes code.
>> 
> I'm sorry about the problems. It's a continuing problem when working with a project where the APIs are not set in stone and are subject to change at a moment's notice.

APIs in LLVM are not set in stone because we want to be able to evolve them as the requirements change.  That should not be a license to constantly rewrite APIs with large numbers of consumers, but simply an indication that if the choice is between fixing the code and maintaining backwards compatibility then developers are free to fix the code.  That doesn't mean that we should never think about the cost to downstream developers when we churn APIs, it just means that we shouldn't be held in a straitjacket by these concerns.  

And the cost of changing an API is that it MUST be properly documented.  Whenever you make the choice to break an API, you are saying 'I am saving my time by not writing compatibility interfaces, and everyone else must rewrite their code to support my changes.'  If you then compound this by saying 'oh, and if you want to know how to rewrite your code, then read my code, my time is too valuable to waste documenting my work' then you do not encourage people to depend on LLVM.

>> It would be really great if whoever is responsible could step back from their code editor for a few minutes and write some documentation on how to migrate from each iteration to the next.
>> 
> That would be more trouble than it's worth because the API is changing quite rapidly right now. A lot of it hasn't solidified just yet. Basically, once a migration doc was written it would be immediately out of date.

This comment is indicative of a code first, think second mentality.  If the APIs are changing so quickly that even a high-level overview and broad migration guide would be out of date before it's finished then it seems pretty clear that the design stage was skipped entirely.  While this is fine for code on the edges (e.g. a back-end under development, where changes don't affect any other users), it generates a lot of problems of people when it is code that is central to LLVM.  

Effectively, here, you are saying that your time is far more valuable than everyone else's.  You can, I hope, understand why this is not an opinion shared by those whose time you are spending to save some of your own.  When you save 10 minutes of your time by not writing docs, you then make every user of this API (which means the author every out-of-tree front end, a lot of out-of-tree passes, and some back ends) spend at least twice that time each reading the code to try to figure out what the new equivalent of the old API is.  

>> One of the goals of LLVM is to be a set of reusable libraries and this goal is not met by gratuitous API churn with no accompanying documentation.  If we can't have a sane deprecation strategy or an automated migration tool then please can we at least have a token attempt at documentation?  The AttributeSet documentation is just embarrassing.  For example, the document for the class is:
>> 
>>> This class manages the ref count for the opaque AttributeSetImpl
>>> object and provides accessors for it
>> 
>> Great.  It's a wrapper around an opaque type that isn't documented in the public headers.  What is an AttributeSetImpl?  What do I use it for?  How does it relate to the last three classes that had similar functionality but different names?
> 
> It's an opaque object. It shouldn't concern people using the AttributeSet class. It's not documented in public headers for that very reason.

You are missing my point.  The ONLY documentation describing AttributeSet tells me exactly one thing: that it is a wrapper around an AttributeSetImpl.  If AttributeSetImpl it is an opaque object that users should not need to know about then why does the only public documentation for AttributeSet mention it at all?  The documentation should tell me what an attribute set is (apparently, from looking over the code, it is an ordered collection of attributes indexed by parameter number?), not give me an implementation detail that apparently users shouldn't need to be aware of as the sole documentation.

>> I can tell it uses the pImpl pattern by just looking at the class definition.  That is not what the documentation should be telling me.  The rest of this file is a case study in how not to write helpful documentation.  This wouldn't be quite so bad if it were a class that's fairly obscure, but this is a class that is part of the core IR that pretty much every part of the pipeline needs to interact with.  (And that's ignoring the fact that it is very confusing for an ordered collection that allows duplicates to be called a set).
>> 
> Welcome to living on the top-of-tree! :-)

Living on the top of tree is unavoidable for most LLVM downstream consumers.  Since we have no deprecation policy, you either incrementally make changes following trunk, or you find that you need to make massive changes when a new release is branched.  We encourage people to follow top of tree to minimise surprises and to give feedback on evolving APIs, so saying 'well, we're just going to churn the API's a lot, sucks to be you' to all of our downstream consumers is a very long way from ideal.  

Take a look over the list archives for the last month for the number of questions from people using LLVM releases as old as 2.8 and you'll see what happens when we make it hard for people to follow trunk.

> We are in between releases. It's expected that the APIs will be unstable. I wrote (several times) that the attributes were going to be rewritten. That implies a lot of things, including that everything about attributes and how you use them will change. Where applicable, there will be an auto-upgrade done. (One requirement is that the old bitcode files will continue to be parsed correctly (at least until release 4.0).) And yes the documentation is also in flux. It has been improving over time.

The APIs haven't just changed in trunk, they've changed with each release.  I have some code that works with 3.1, 3.2 and trunk, and did work with 3.0 until I deleted that code to try to make it a bit cleaner.  Every single one of these has very different APIs for accomplishing exactly the same thing.  This is not rapid evolution to meet changing requirements, this is gratuitous API churn caused by coding first, rather than stopping to design the APIs properly in the first place.

Yes, it's great that you're rewriting the attributes interface now to be stable and extensible in the longer term, but the fact that this is now the fourth incompatible iteration of the APIs in as many releases implies that something is badly wrong with the design process.  

> The thing you quoted above is perhaps the worst comment from the Attributes.h file, and isn't indicative of the rest of the comments in there, which have been changing during the rewrite.

No indeed.  Most of the methods are completely lacking comments and the file contains no explanation of the relationship between attributes, functions, parameters, and attribute sets.

> -bw
> 
> P.S. AttributeSets don't allow duplicates. :)

Really?  In that case I'm using it wrongly.  Which I suspected anyway, because it is almost completely undocumented.  Are the indexes not parameter indexes?  If not, then why are they exposed at all?  

> Using a development branch and then slamming those changes into trunk is at odds with the llvm style incremental development philosophy. Living on ToT isn't easy. No one ever said it would be. The changes that are being complained about have only been happening for about a week now. And they're more stable day by day.

No, they have been happening for at least a year.  If you are honestly unaware of this, then I suggest that you stop hacking on the attributes until you're familiar with the various iterations that those APIs have gone through over the past two years.  

David