IncrementalValuesProvider.GroupBy

[eiriktsarpalis]

IncrementalValuesProvider.GroupBy

[sharwell]

Incremental generators include many steps in the graph. Each step is cached depending on its own inputs and outputs. Most changes to a project result in at least one step somewhere running. As best I can tell, the proposed GroupBy operation would not inherently make anything cacheable that is not already possible to cache today using Collect and SelectMany. It's also possible that Collect().Select() has good caching characteristics. The link I gave above includes a step with that form.

[jamescarterbellMSFT]

Am I correct in understanding that calling .Collect() makes an IncrementalGenerator non-cache friendly? I think I may be running into this issue when I collect to do some deduping.

[sharwell]

You can call Collect() and then split the items back apart using SelectMany(), and the only part that runs each time is the operation within the Collect(). Here's an example:
https://github.com/dotnet/roslyn-analyzers/blob/7455f63411369a962f769361d1a979a547756ada/src/Microsoft.CodeAnalysis.ResxSourceGenerator/Microsoft.CodeAnalysis.ResxSourceGenerator/AbstractResxGenerator.cs#L131-L175

GroupBy could be efficiently implemented by using Collect() followed by SelectMany, where each element of the SelectMany is a group.

[jamescarterbellMSFT]

So provider.Collect().Select() that outputs an immutable array provider would not be cacheable, but a provider.Collect().SelectMany() would produce something cacheable?

[CyrusNajmabadi]

.Collect is fine to call, you just need to provide a comparer to properly compare the arrays it has inside of it. Note: i consider it bad that this is needed, but c'est la vie for now. But as long as you give an appropriate comparer, things will be incrementally safe and sound.

[CyrusNajmabadi]

@chsienki I think we really need to invest in:

1. a comparer for comparing ImmutableArrays.
2. having anything we produce that returns a IncrementalValueProvider<ImmutableArray<TSource>> use teh comparer.

There's really no reason at all to have no comparer here, and it means a collect call will break incrementality unless a user knows to go fix this up.

[sharwell]

@CyrusNajmabadi I'm not sure that aspect is a problem today. I've seen many source generators opt for EquatableArray<T> instead of ImmutableArray<T> in their data models. It's not clear how the current behavior of Collect() would have any negative impact on this.

Providing a default comparer for ImmutableArray<T> would not address cases where EquatableArray<T> is needed inside of record types, and if you already use EquatableArray<T> in the model it's not clear why ImmutableArray<T> would be needed separately.

[CyrusNajmabadi]

@sharwell Collect doesn't use that though, ti's just:

        public static IncrementalValueProvider<ImmutableArray<TSource>> Collect<TSource>(this IncrementalValuesProvider<TSource> source) => new IncrementalValueProvider<ImmutableArray<TSource>>(new BatchNode<TSource>(source.Node), source.CatchAnalyzerExceptions);

So you get an IVP (not an IVsP) point at an ImmutableArray without a proper comparer for the immutable array itself. .Collect doesn't let the user provide the final array the values are collected into. It chooses that it is IA<>, and you need to know to provide an appropriate comparer if you want proper incrementality at that pipeline step.

[eiriktsarpalis]

You can call Collect() and then split the items back apart using SelectMany(), and the only part that runs each time is the operation within the Collect(). Here's an example:

So if I'm understanding the semantics of SelectMany correctly, it is forking one IVP into an IVsP of independently cached values? I think that could work for my use case.

[sharwell]

.Collect doesn't let the user provide the final array the values are collected into

That should be fine. If none of the inputs to the Collect operation changed, Collect should be reusing the old cached ImmutableArray<> output instead of creating a new one. The only time Collect produces a new ImmutableArray<> to compare with the old cached value is when one or more elements in that array no longer compare element-wise equal with the old values.

and you need to know to provide an appropriate comparer if you want proper incrementality at that pipeline step

There is no public API to alter the comparer for Collect, but that's fine because the default comparer is correct for the semantics of this operation.

[CyrusNajmabadi]

There is. You call . WithComparer afterwards

[CyrusNajmabadi]

The only time Collect produces a new ImmutableArray<> to compare with the old cached value is when one or more elements in that array no longer compare element-wise equal with the old values.

Right. That's the exact problem. You get a different input, but you end up creating conceptually the same output (which is very very possible). As far as the driver is concerned, because it's a new IA, it is different, and it will not stop the incremental pipeline.