feat: add incremental caching

[TimothyMakkison]

Modify CompilationAnalyzer to return an equatable model, skipping the final build and emit stage, improving performance and making the IDE more responsive.

• Removed public properties from CompilationAnalyzer.
• State is stored in CompilationModel converting types into equivalent structurally equatable types.
• Incremental pipeline can cache CompilationModel and skip the emit stage in the future if nothing changes.
• Added incremental generator tests.

Note that I had to add Mediator.SourceGenerator.Roslyn41 to use WithTrackingName for the tests. This is pretty fragile as changes to Mediator.SourceGenerator.Roslyn40 will not be detected by the tests. This could be solved with compile constant preproccessor sections, although I'm not confident I can add this.

[martinothamar]

Awesome work! In my experimental branch I think I removed the Roslyn38 project, was trying to simplify. Do you know what version of the SDK is needed for 4.1?

[martinothamar]

Thanks for your patience, sorry about the delay here!

A couple of comments, otherwise this looks awesome 🥇
I must admit I don't understand all these Roslyn APIs related to incremental compilation, but you bring benchmarks (which I ran) and tests so LGTM 😄

BenchmarkDotNet=v0.13.1, OS=pop 22.04
AMD Ryzen 5 5600X, 1 CPU, 12 logical and 6 physical cores
.NET SDK=6.0.417
  [Host] : .NET 6.0.25 (6.0.2523.51912), X64 RyuJIT

Job=InProcess  Toolchain=InProcessEmitToolchain  

Method	Mean	Error	StdDev	Gen 0	Gen 1	Allocated
Cached	13.69 ms	0.053 ms	0.047 ms	265.6250	78.1250	22 MB
Compile	14.89 ms	0.056 ms	0.049 ms	265.6250	265.6250	23 MB
LargeCached	87.59 ms	0.327 ms	0.290 ms	333.3333	166.6667	38 MB
LargeCompile	154.20 ms	0.807 ms	0.755 ms	750.0000	250.0000	77 MB

[TimothyMakkison]

Thanks for the review, I'll resolve the issues when I have more time over the weekend 👍

I must admit I don't understand all these Roslyn APIs related to incremental compilation

TLDR, each step in a incremental generator (Select, Where, SelectMany), caches the input and output of the previous runs step. By using a custom comparer or object implementing IEquality we can use this behaviour to skip a section of the pipeline if the input object is the same as the last run.

In roslyn equivalent syntax (ie SyntaxNode, InvocationExpressionSyntax) is the same between runs whereas each compilation object will always be different between runs (ITypeSymbol, ISymbol, Compilation) and should not be cached for both performance and logic reasons.

In our case I created a parse step that produces a cacheable object, this lets us skip the output stage.

[martinothamar]

Thanks! Makes sense

[TimothyMakkison]

Happy new years and thanks for the review, I think I've made all the relevant changes 👍

• I wen with your suggestion and moved TryParseConfiguration to before the array declaration
• Added support for Use correct accessibility modifer based on message implementing IRequest etc #131
• RequestMessageHandlerWrapper is still a little janky. My ide doesn't show linting if the file is in the Models folder, but will lint and show references when moved out. Still no idea why this is happening. It still occurs on a different computer with a fresh install 😩
• I'm a little concerned that the tests didn't fail when I broke diagnostics or namespace configuration. I'm planning on adding some tests to SnapshotTests. Should I add them in this pr or a different one? This pr is getting a little bloated. See feat: add namespace and diagnostics test #135

Edit: just ran the benchmarks and got drastically different results to yours. Moving TryParseConfiguration doesn't cause this as I get the same results when they are moved back.

Method	Mean	Error	StdDev	Median	Gen 0	Gen 1	Allocated
Cached	5.296 ms	0.0548 ms	0.0512 ms	5.299 ms	507.8125	156.2500	5 MB
Compile	7.333 ms	0.3460 ms	0.9982 ms	6.686 ms	593.7500	203.1250	6 MB
LargeCached	70.923 ms	0.4671 ms	0.4141 ms	70.743 ms	3875.0000	1500.0000	36 MB
LargeCompile	134.006 ms	0.8707 ms	0.8144 ms	134.052 ms	7500.0000	2500.0000	83 MB

[martinothamar]

finally merging this, thanks for the great work! 😄