C@tannergooding Sorry this took so long this ended up as a 75Gb DotTrace snapshot so I had to move it to my main pc from my laptop, but here are the main hot paths.
C
CIt should be noted if I only factor in runnning threads I get this
CHere are the hotspots as well
C
TSo looks like there's just a lot of cursor objects being created, which is causing a ton of time
Tand then quite a lot of time spent in cleaning up translation units, which needs to be figured out if that's contention or something else
TSortedDictionary is not fast 
I save a significant amount just by switching to
Dictionary instead-- the former is basically a wrapper over
SortedSet<KVP<TK, TV>>T@Curin, when you have some time could you see if https://github.com/dotnet/ClangSharp/pull/619 (which is merged into dotnet/main) mitigates some of the perf and memory issues you saw?
I can provide binaries if that helps, otherwise you can just pull latest
I can provide binaries if that helps, otherwise you can just pull latest
main and build for releaseGitHub
This targets a few different hot spots found in the profiler.
Tyou might also explicitly try the NAOT build since many header files are small and we're not likely to hit steady state throughput with the JIT
(
(
dotnet publish -c Release /p:PublishAot=true -r win-x64 -f net10.0)TI'm expecting you should see somewhere around 15% perf improvement and 20-30% memory improvement
which is what I saw for TerraFX locally
which is what I saw for TerraFX locally
Tfew outliers, but in general it takes about 1s to fully process a given file and emit bindings
Tthe average is 2.16s, mainly due to the outliers
while the median is 1.08s (i.e. 50% of values take less, 50% take more)
while the median is 1.08s (i.e. 50% of values take less, 50% take more)
Tthe winrt files really take a lot, with most taking 6.2s or more