Optimize metadata reading for extension methods

[DedSec256]

Currently, finding out whether a type in an IL assembly has Extension attribute requires reading all attributes from the metadata and constructing a representation of ILAttribute for each one. This process is time-consuming and involves unnecessary work. It would be beneficial to have a more efficient method for determining if a type can contain extension methods.

Furthermore, ILTypeDef constructor is a public API but lacks the ability to pass lazy computation of custom attributes. In contrast, the internal constructor in FCS supports this feature.

This PR proposes:

• The ability to pass lazy computation of custom attributes to ILTypeDef public constructor
• The ability to specify the presence of Extension attribute in ILTypeDef (and some refactoring around isKnownToBeAttribute flag)
• Default implementation for lightweight detection of Extension attribute presence

This solution will allow for more efficient filtering of ILTypeDef that definitely does not contain extension methods, both for users of the public API and within the FCS implementation.

[github-actions]

❗ Release notes required

---

✅ Found changes and release notes in following paths:

Change path	Release notes path	Description
src/Compiler	docs/release-notes/.FSharp.Compiler.Service/8.0.400.md

[T-Gro]

Still a WIP, or ready for review? I can see CI is passing all the gates now.

[DedSec256]

I'm still doing some experiments, and it will be ready for review soon

[DedSec256]

@T-Gro, @psfinaki, I think it's ready for review.

Lazy evaluation of attributes (our own implementation inside ReSharper.FSharp) and the presence of extension methods flag allowed us to eliminate unnecessary attribute calculations (from 30 seconds to 1 second), reduce the ILTypeDef creation time by 30 seconds, and the overall analysis time of the F# file inside the Rider monorepository from 70 seconds to 40 seconds.

[vzarytovskii]

This needs release notes and a rough estimation how does it affect startup time/runtime and memory consumtion.

[DedSec256]

The main idea of PR is to spend a little more time searching for a possible ExtensionAttribute in typeDefReader, while reducing the number of searches with a full read of all type attributes later.
Theoretically, this does not affect the startup time, increases the time and memory of typeDefReader calls, and reduces the time and memory of IsTyconRefUsedForCSharpStyleExtensionMembers calls.

Benchmarks were conducted for some scenarios

• For the FCS benchmark, the following code was used to analyze files

[<MemoryDiagnoser>]typeBench()=let mutablefiles=[||]let mutablesources=[||]let mutableoptions= Unchecked.defaultof<_>let mutablechecker= Unchecked.defaultof<_>[<GlobalSetup>]memberthis.Setup()=letotherOptions=[|(* options with referenced dlls *)|]
files <-[|(* files *)|]
sources <- files |> Array.map (File.ReadAllText >> SourceText.ofString)
options <-{
ProjectFileName =// some .fsproj
ProjectId = None
SourceFiles = f
OtherOptions = otherOptions
ReferencedProjects =[||]
IsIncompleteTypeCheckEnvironment =false
UseScriptResolutionRules =false
LoadTime = DateTime.UtcNow
UnresolvedReferences = None
OriginalLoadReferences =[]
Stamp = None
}
checker <- FSharpChecker.Create(projectCacheSize =200)[<IterationSetup>]memberthis.Invalidate()=
checker.InvalidateAll()
checker.ClearLanguageServiceRootCachesAndCollectAndFinalizeAllTransients()[<Benchmark>]member_.TypeCheckFile()=for file, source in Seq.zip files sources dolet_,checkResult =
checker.ParseAndCheckFileInProject(file,0, source, options)|> Async.RunSynchronously
match checkResult with| FSharpCheckFileAnswer.Succeeded(checkResult)when checkResult.Diagnostics.Length <>0-> failwithf $"Errors: %A{checkResult.Diagnostics}"| FSharpCheckFileAnswer.Aborted -> failwithf "Aborted"|_->()

• System specification

BenchmarkDotNet v0.13.10, Windows 11 (10.0.22631.3737/23H2/2023Update/SunValley3)
12th Gen Intel Core i9-12900H, 1 CPU, 20 logical and 14 physical cores
.NET SDK 9.0.100-preview.5.24307.3
[Host] : .NET 8.0.1 (8.0.123.58001), X64 RyuJIT AVX2 DEBUG
Job-WKOUUV : .NET 8.0.1 (8.0.123.58001), X64 RyuJIT AVX2

InvocationCount=1 UnrollFactor=1

Analyzing a single file in the Giraffe framework

300 dll references required for project analysis were taken

The averaged results of launches look as follows.

Method	Mean	Error	StdDev	Gen0	Gen1	Gen2	Allocated
TypeCheckFile - old	587.2 ms	8.64 ms	8.08 ms	2000.0000	1000.0000	1000.0000	851.73 MB
TypeCheckFile - new	581.8 ms	6.09 ms	5.70 ms	2000.0000	1000.0000	1000.0000	819.13 MB

Sequential analysis of all files in src/Giraffe

22 files were taken

Method	Mean	Error	StdDev	Gen0	Gen1	Gen2	Allocated
TypeCheckFiles - Giraffe - old	1.185 s	0.0136 s	0.0121 s	3000.0000	2000.0000	1000.0000	1.69 GB
TypeCheckFile - Giraffe - new	1.167 s	0.0078 s	0.0069 s	3000.0000	2000.0000	1000.0000	1.63 GB

Sequential analysis of all files in ReSharper.FSharp/FSharp.Psi.Services

48 files & 471 dll references required for project analysis were taken

Method	Mean	Error	StdDev	Gen0	Gen1	Gen2	Allocated
TypeCheckFile - ReSharper - old	5.689 s	0.0472 s	0.0419 s	8000.0000	6000.0000	5000.0000	7.35 GB
TypeCheckFiles - ReSharper - new	5.635 s	0.0553 s	0.0517 s	7000.0000	5000.0000	4000.0000	7.29 GB

Unit-test for ILReading

The benchmark checks for additional expenses when called typeDefReader.
450 assemblies have been taken from the benchmark for ReSharper.FSharp/FSharp.Psi.Services, which contain approximately 1,000,000 types, 287,000 of which have extensions

The code of an existing test from CompilerServiceBenchmarks was rewritten to the following

for fileName in assemblies doletreader= OpenILModuleReader fileName readerOptions
letilModuleDef= reader.ILModuleDef
for typedef in ilModuleDef.TypeDefs.AsArray()do
typedef.IsKnownToBeAttribute |> ignore

Method	Mean	Error	StdDev	Gen0	Gen1	Gen2	Allocated
ILReading - old	377.4 ms	7.55 ms	11.06 ms	3000.0000	2000.0000	2000.0000	731.9 MB
ILReading - new	485.8 ms	9.57 ms	12.77 ms	3000.0000	2000.0000	2000.0000	749.5 MB

Profiling the ReSharper.FSharp project inside the Rider soluition

In this scenario F# projects refer directly to C# projects in the Rider solution instead of DLLs from its SDK. Based on the C# sources, the metadata required for FCS analysis is built. Lazy evaluation of custom attributes (our own implementation inside ReSharper.FSharp) for ILTypeDef and the presence of the extension methods flag allowed us to eliminate unnecessary forced attribute calculations (from 30 seconds to 1 second), reduce the ILTypeDef creation time by 30 seconds, and the overall analysis time of the F# file inside the Rider solution from 70 seconds to 40 seconds.

Notes and further work

1

On all profiles captured during the investigation, it is clear that a significant amount of time is reduced from IsTyconRefUsedForCSharpStyleExtensionMembers calculation because of CanContainExtensionMethods flag

Before:

After:

2

Full line-by-line profile for merged typeDefReader calls from sequential analysis of all files in src/Giraffe benchmark looks like

The search for the range of attributes for the type adds here 150 ms.

2.1

On all profiles captured during the investigation, it is clear that a significant amount of time is consumed by non-lazy calls to the following two functions in typeDefReader:

In a separate PR, we can think about making their computation more lazy. For example, the list of generic parameters was needed inside IsTyconRefUsedForCSharpStyleExtensionMembers, which after adding the CanContainExtensionMethods flag, probably, no longer needs to be computed eagerly

Conclusions

• The benchmarks showed that the analysis time, at the very least, did not increase. Memory consumption has slightly decreased.
• Profiling showed that the IsTyconRefUsedForCSharpStyleExtensionMembers call time has significantly decreased due to the absence of the need to read custom attributes.
• The flag CanContainExtensionMembers and the ability to lazily compute attributes in the public API of ILTypeDef can significantly reduce analysis time.
• The typeDefReader call time has naturally increased due to the search for indexes of custom attributes for each readable type.
• There is room for optimizations in the typeDefReader for reading interfaces and generic parameters.

With a larger number of types being checked for the presence of ExtensionAttribute, it is likely analysis in NameResolution should result in an overall greater gain in speed.

[DedSec256]

Note that statistical optimization for searching MethodDefParent has been added: if the index of the requested method does not fall within the range of methods of the checked type, it is not necessary to call seekReadTypeDefRowExtents (and thus read the indexes of methods of the next type)

The same optimization can be applied to fields in a separate PR

fsharp/src/Compiler/AbstractIL/ilread.fs

Lines 2926 to 2936 in 295844f

	seekReadIndexedRow (
	ctxt.getNumRows TableNames.TypeDef,
	(fun i -> i, seekReadTypeDefRowWithExtents ctxt i),
	id,
	(fun(_,((_,_,_,_,fieldsIdx,_),(endFieldsIdx,_)))->
	if endFieldsIdx <= idx then1
	elif fieldsIdx <= idx && idx < endFieldsIdx then0
	else-1),
	true,
	fst
	)