[Mpdreamz]

Benchmarks of reading, writing, parsing and comparing, with allocations, on System.Collections.Immutable and CoreLib (35,751 .api lines), using BenchmarkDotNet's MemoryDiagnoser. The profiles come from dotnet-trace.

CoreLib	Before	After
Read	407 ms, 832 MB	108 ms, 109 MB
Write	19 ms, 25 MB	19 ms, 25 MB
Parse	160 ms, 120 MB	122 ms, 72 MB
Compare with its own .api file	39 s, 4.4 GB	20 ms, 23 MB

On System.Collections.Immutable everything takes 2 ms or less: read 1.8 ms, write 0.45 ms, parse 1.2 ms, compare 0.36 ms.

What was slow

• Compare:

  • For every member and parameter it wrote both types out as C# to see whether they differed.
  • It built each member's report path from rendered parameter types, changed or not.
  • Writing a type looked at every member of the type it was in, so the cost grew with the square of the member count: AdvSimd, with 1,772 members, took 5.3 s alone.

  Now it compares the models structurally (Same), and writes C# and builds paths only for a change. It pairs members by a structural identity instead of a string key, checks attributes and constraints on the models first, and caches one ApiDeclarationWriter per type.

• Read:

  • Decoding an attribute's enum argument, such as [EditorBrowsable(EditorBrowsableState.Never)], scanned every type in the assembly.
  • Every attribute check built the attribute type's name again.

  Both are looked up once now. Nullable attributes are read from their blob instead of decoded the general way, primitive types and empty lists are shared, and names are compared in place.

• Parse:

  • Each line was checked for a type declaration twice.
  • Every word was allocated twice, once to peek and once to take.
  • Each member got a new scope.

What's left

GC pauses are 45–60% of the remaining time. Reading or parsing builds a model that survives, and background GC handles that poorly. With DOTNET_gcConcurrent=0, CoreLib parses in 61 ms and reads in 105 ms. The tool now sets <ConcurrentGarbageCollection>false</ConcurrentGarbageCollection>. A library can't choose its host's GC mode, so library callers keep the numbers in the table.

Correctness

• The CsCheck properties from Check the reader, file and comparison with CsCheck, and fix what it found #101 and Generate generic math and modern interface members, and keep nullability out of paths #102 pass on this branch, including one soak run with 2,000 libraries per property and 800 pairs, so the faster paths give the same answers.
• The Roslyn oracle, the round-trip and comparison snapshots, and the 630 tool tests pass.
• api/ gains the new public ApiDeclarationWriter, which verifyapi caught and the regenerated files record.

Run the benchmarks with dotnet run -c Release --project benchmarks/Nullean.PublicApi.Benchmarks -- --filter '*' --inProcess. They run in-process because BenchmarkDotNet 0.15.8 doesn't recognise net11.0 when it launches a separate process.

🤖 Generated with Claude Code