[LLVMdev] Using annotation attributes

[Chris Lattner]

On Jun 25, 2008, at 6:33 AM, Matthijs Kooijman wrote:
> Hi all,

Howdy Matthijs,

> I've also been developing an interest in using IR annotations for my  
> compiler.
> Some discussion with Bart turns out that he has implemented some  
> code to parse
> the llvm.globals.annotations array, but in no way integrated or  
> reusable.
> We've spent some thought about how this could be done properly,  
> which I will
> share here.

Ok, cool.  Annotations are tricky to do right :)

> Firstly, however, I was wondering about the format of the
> llvm.globals.annotations array. It does not seem to be defined in  
> the LLVM
> language reference, shouldn't it be? It's name suggests that it is a  
> reserved
> variable name with a fixed type (similar to intrinsic functions?).

Yes, we should document it.  It is a convention established by the  
__builtin_annotate function in the c compilers.  We should standardize  
it and document it.

> Furthermore, it seems that the AnnotationManager that is currently  
> implemented
> is capable of keeping a list of Annotations for any Annotatable  
> (currently
> only Function). These annotations are kept in memory only and really  
> have
> nothing to do at all with the annotations in the IR.

Yes, this is a really old mechanism that we should rip out.   
MachineFunction should be moved to be an analysis that is preserved as  
an actual part of the passmanager, instead of being a thing we tack  
onto the Function object.  We have killed all uses of this old  
annotation mechanism except MachineFunction.

> Still, it seems that using the AnnotationManager to make the IR  
> annotations
> accessible seems like a decent approach.

I agree that *having* an annotationmanager makes sense, but the  
existing one should die and be replaced. :)

> The way I see this is having some pass, or probably the assembly  
> reader or the
> AnnotationManager itself, parsing the llvm.global.annotations  
> variable and
> adding annotations to the corresponding GlobalValues. This would  
> just leave the
> annotations in the IR as well, so that transformation passes would  
> properly
> preserve them (and, just like debug info, sometimes be prevented from
> modifying some annotated global values unless they are taught how to  
> preserve
> the annotations).

Makes sense.  This is similar to how the MachineDebugInfo stuff  
deserializes debug info out of the LLVM IR and presents it for easy  
consumption of the code generator.

> By using a subclass of Annotation (say, GlobalAnnotation) we can  
> distinguish
> between annotations that are (or should be) in the IR and (the  
> existing)
> annotations that should be in memory only. This would also allow for  
> newly
> added annotations to be immediately be added to the IR, ensuring  
> that the
> AnnotationManager's view remains consistent with the IR.

I think we need to distinguish between two forms of annotation:

1. there are some "annotations" like "readonly", "nounwind", etc that  
are baked into the LLVM IR and are/should be documented in LangRef.

2. There are annotations that are really "cheap extensions" of the  
LLVM IR that are either experimental, very domain specific, or that  
are just metadata about the code.

For #1, the current "parameter attributes" we have work reasonable  
well, and Devang is actually cooking up a proposal to extend them a  
bit (to fix some issues with LTO).  #2 is something that llvm.annotate  
handles reasonable well, but I agree it would be great to have a nice  
interface to update/read them.

The advantage of #1 is that the compiler as a whole knows about the  
attributes, but this means that adding one is "hard".  The advantage  
of #2 is that they are easy to add, but they have limitations and can  
impact codegen (e.g. they disable IPO in some cases).

> A problem I could imagine using this approach would be name  
> conflicts. Since
> any annotation name could come from the IR, these could conflict by  
> the other
> names already in use (such as "CodeGen::MachineFunction" IIRC). This  
> could be
> solved by using a "GlobalAnnotation::" prefix for the name, or  
> something
> similar.

It could also be served by making them completely string based, and  
just provide a simple string interface?  That way you don't need  
classes for each attribute.

-Chris