perf: cache GetAttributeObjectInitializer

[TimothyMakkison]

Use ConditionalWeakTable to cache GetAttributeObjectInitializer. ConditionalWeakTable uses Object.ReferenceEquals for equality and does not count as a reference to the AttributeData, therefore it won't keep the object or attached compilation alive after it becomes stale.

• On a related note, I'm pretty sure InterfaceCache should be causing a memory leak by keeping most stale Compilation objects alive. I have no idea why this hasn't become an issue.

• Note that the benchmarks aren't completely accurate for real world scenarios. While GetAttributeObjectInitializer is slow, I suspect the main issue is each method in TUnit.TestProject has at least 8 identical licensing attributes, making this a slightly misleading benchmark

Before

Method	Mean	Error	StdDev	Gen0	Gen1	Allocated
Compile	499.5 ms	19.50 ms	57.48 ms	18000.0000	4000.0000	165.61 MB

After

Method	Mean	Error	StdDev	Median	Gen0	Gen1	Allocated
Compile	174.6 ms	6.29 ms	17.96 ms	166.6 ms	10000.0000	3000.0000	96.14 MB

Attribute bloat

[TimothyMakkison]

Realised that because the compilation is always the same in the benchmark, AttributeData would always be the same, meaning that all calls were cached. I reran the benchmarks, clearing the cache each time and here is the result.

Still excellent, but I might make AttributeWriter an instance and pass it around with its own internal cache. This way we don't have to bother with ConditionalWeakTable and can just use Dictionary.

Method	Mean	Error	StdDev	Median	Gen0	Gen1	Allocated
Compile	225.6 ms	10.50 ms	29.45 ms	210.0 ms	11000.0000	4000.0000	104.79 MB

[TimothyMakkison]

Sorry, rowing right now. I had a better idea to handle caching while respecting stateless generator runs. This shouldn't be merged

[TimothyMakkison]

Instead of having a static ConditionalWeakTable, instead we make AttributeWriter an instance. This shares information between the GenerateTestMethodSource calls for a given compilation. This way the benchmarks remain accurate, we don't have to mess with weak references and we can expand this for more types. Replace the static dictionaries in InterfaceCache with this technique.

I did this previously with Mapperly, where CompilationContext contains all the shared type information for a given generator run.

[TimothyMakkison]

Is modular pipelines failing on windows just a quirk?