Improve the performance of the type loader through various tweaks #85743

davidwrighton · 2023-05-04T00:14:25Z

Disable type validation by default in ReadyToRun images
- This is potentially a behavior changing modification, and so maybe we shouldn't do this; however it produces some nice wins
- I've built a type validation layer into the compiler to validate types against (most of) the checks that are removed. As part of building this, I've found a few small bugs in the typesystem. This validation layer isn't quite complete, and will be optional, although it will be enabled by default in crossgen2.
Add new metadata to R2R images for finding generic parameter and nested class information
- These lookups are extremely common during type loading.
Store the index of the MethodDesc within the MethodDescChunk as well as the chunk index.
- Used to improve the performance of finding the temporary entry point, which is a surprisingly common operation during type load. Its converted from an O(N^2) algorithm to O(1)
Remove an old ngen'ism (CanShareVTableChunk) Since we don't have ngen images, we can ALWAYS share vtable chunks

Overall, this set of changes produces the following performance impact

Powershell startup

Initial time to launch powershell 100 times¹ - 22869ms
With all fixes enabled - 19966ms
Without enabling skipping type validation - 21240ms

Additionally, while this change does increase the size of the r2r compiled files, the total size of the dlls associated with powershell goes from 242,357,432 bytes to 242,564,792 bytes, which is a fairly minimal increase.

These powershell startup numbers are done by taking an official build of powershell targetting an older version of .NET, replacing the version of .NET with a locally built one, and then running crossgen2 on all binaries of powershell. This isn't likely how Powershell actually produces its builds, but should be representative of the sort of performance wins that can be had with these changes. ↩

- Technically, this is a breaking change, so I've provided a means for disabling the type validation skip - The model is that the C# compile won't get these details wrong, so disable the checks when run through crossgen2. The idea is that we'll get these things checked during normal, non-R2R usage of the app, and publish won't check these details.

…with R2R optimized forms

…tTemporaryEntryPoint *much* faster

…mpilation

- it was only needed to support NGen

…rides only

…tored slots anymore

- Notably, the recursive stuff now works - Also fix a bug in constraint checking involving open types in the type system

AaronRobinsonMSFT · 2023-05-20T15:30:01Z

src/coreclr/inc/readytorun.h

+    MethodIsGenericMap          = 121, // Added in V9.0
+    EnclosingTypeMap               = 122, // Added in V9.0
+    TypeGenericInfoMap          = 123, // Added in V9.0


Suggested change

MethodIsGenericMap = 121, // Added in V9.0

EnclosingTypeMap = 122, // Added in V9.0

TypeGenericInfoMap = 123, // Added in V9.0

MethodIsGenericMap = 121, // Added in V9.0

EnclosingTypeMap = 122, // Added in V9.0

TypeGenericInfoMap = 123, // Added in V9.0

src/coreclr/inc/readytorun.h

src/coreclr/vm/readytoruninfo.h

src/coreclr/vm/readytoruninfo.cpp

davidwrighton · 2023-05-30T23:28:04Z

/azp run runtime-coreclr crossgen2

Fix command line parameter to be more reasonable, and allow logging on command Fix the rest of issues noted in crossgen2 testing

docs/design/coreclr/botr/readytorun-format.md

src/coreclr/inc/readytorun.h

MichalStrehovsky · 2023-06-02T00:56:24Z

src/coreclr/tools/Common/TypeSystem/Common/SignatureVariable.cs

+            flags |= TypeFlags.HasGenericVarianceComputed;
+            flags |= TypeFlags.HasFinalizerComputed;
+            flags |= TypeFlags.AttributeCacheComputed;


Isn't this a bug if someone asks for these? Signature variables don't have "a finalizer" but also if someone is asking for that, maybe it's a bug. I've ran into the assert not setting this flag produces a 100% of times this was a bug in my code that I was happy to find so easily.

Unfortunately, this work operates almost entirely over uninstantiated types and methods, and has a few spots where it actually needs to use these things. However, it might be the case that I can only set a minimal number of flags. I enabled them all as it seemed like the right thing to do, but I believe the only 1 I really needed to work was the attribute cache.

What's the stack that leads to this being needed? This really feels like something where the only right answer is to unask the question.

MichalStrehovsky · 2023-06-02T01:00:08Z

...r/nativeaot/System.Private.TypeLoader/src/Internal/TypeSystem/RuntimeGenericParameterDesc.cs

+        internal void SetAssociatedTypeOrMethod(TypeSystemEntity associatedTypeOrMethod)
+        {
+            _associatedTypeOrMethod = associatedTypeOrMethod;
+        }


Can we do this in a constructor? This is already a RuntimeXXXDesc so I wouldn't mind if the constructor just understands a ReadOnlySpan of Runtime.GenericVariance.

Refactor requested and implemented.

MichalStrehovsky · 2023-06-02T01:06:35Z

src/coreclr/tools/Common/TypeSystem/Ecma/EcmaGenericParameter.cs

+                TypeSystemEntity tse = Volatile.Read(ref _associatedTypeOrMethod);
+                if (tse == null)
+                {
+                    GenericParameter parameter = _module.MetadataReader.GetGenericParameter(_handle);
+
+                    tse = (TypeSystemEntity)_module.GetObject(parameter.Parent);
+                    Volatile.Write(ref _associatedTypeOrMethod, tse);
+                }
+                return tse;


I don't think we need the Volatile here.

If this is perf critical, we can use the same pattern as BaseType:

runtime/src/coreclr/tools/Common/TypeSystem/Ecma/EcmaType.cs

Lines 174 to 182 in 8ee61fe

public override DefType BaseType

{

get

{

if (_baseType == this)

return InitializeBaseType();

return _baseType;

}

}

If it's not perf critical maybe we don't even need to cache it.

MichalStrehovsky · 2023-06-02T01:08:16Z

src/coreclr/tools/ILVerification/InstantiatedGenericParameter.cs

@@ -76,6 +76,8 @@ private InstantiatedGenericParameter(GenericParameterDesc genericParam)

        public override GenericConstraints Constraints => _genericParam.Constraints;

+        public override TypeSystemEntity AssociatedTypeOrMethod => _genericParam.AssociatedTypeOrMethod;


I don't know much about the ILVerify codebase, but doesn't this need InstantiateSignature?

Maybe? I don't know that codebase either, but it doesn't seem unreasonable.

…ssues

src/coreclr/vm/readytoruninfo.h

AaronRobinsonMSFT

The VM side looks good to me.

davidwrighton · 2023-06-09T23:08:23Z

@trylek Could you take a look at the Crossgen2 side of these changes?

trylek · 2023-06-09T23:16:59Z

...ols/aot/ILCompiler.ReadyToRun/Compiler/DependencyAnalysis/ReadyToRun/EnclosingTypeMapNode.cs

+            builder.AddSymbol(this);
+
+            // This map is only valid for assemblies with <= 0xFFFE types defined within
+            Debug.Assert(_metadata.TypeDefinitions.Count <= 0xFFFE);


Nit - this constant appears multiple times in the code, could we make it a symbolic constant? A sufficiently expressive name like MaxTypeCountSupportingEnclosingTypeMap could also make some of the code comments obsolete / self-explanatory.

trylek · 2023-06-09T23:23:00Z

...coreclr/tools/aot/ILCompiler.ReadyToRun/Compiler/DependencyAnalysis/ReadyToRun/HeaderNode.cs

+                {
+                    if (factory.OptimizationFlags.TypeValidation == TypeValidationRule.AutomaticWithLogging)
+                    {
+                        foreach (string reason in _shouldAddSkipTypeValidationFlag.Result.reasons)


I might be missing part of the intended logic here - according to the naming I would expect _shouldAddSkipTypeValidationFlag to represent some form of vetting whether we can drop the type validation and demonstrate the rationale if we do; why do we dump the reasons in the else block instead?

Added a comment. These are logging reasons why we aren't able to skip validation. Logging reasons we could skip would be really boring. (It would just be a list of types that are considered ok, which isn't all that useful for anybody)

...coreclr/tools/aot/ILCompiler.ReadyToRun/Compiler/DependencyAnalysis/ReadyToRun/HeaderNode.cs

trylek · 2023-06-09T23:42:28Z

src/coreclr/vm/readytoruninfo.cpp

+    return count;
+}
+
+HRESULT ReadyToRun_TypeGenericInfoMap::GetGenericArgumentCountNoThrow(mdTypeDef input, uint32_t *pCount, IMDInternalImport* pImport) const


Nit - there seems to be tons of duplication between the pairs of methods with / without the NoThrow prefix. Could we somehow consolidate them by making one call the other and just slightly tweak the outputs?

trylek · 2023-06-09T23:44:20Z

src/coreclr/vm/readytoruninfo.cpp

+        return 0;
+    }
+
+    uint8_t chunk = ((uint8_t*)&MethodCount)[((rid - 1) / 8) + 4];


Nit - if you want to be consistent with your implementation of the other tables (and possibly also make the code easier to understand), you might want to consider replacing 4 with sizeof(uint32_t) to make it obvious it represents the initial count entry in the table.

trylek

Looks great to me, thanks David for making all these optimizations!

…type_load

davidwrighton added 5 commits April 13, 2023 13:49

Replace expensive lookups of generic parameter and nested class data …

3b8143c

…with R2R optimized forms

Store index of MethodDesc as well as ChunkIndex. Makes MethodDesc::Ge…

0c88982

…tTemporaryEntryPoint *much* faster

Optimize the path for computing if a method is eligible for tiered co…

910f4c4

…mpilation

Remove CanShareVtableChunksFrom concept

418af3c

- it was only needed to support NGen

ghost assigned davidwrighton May 4, 2023

dotnet-issue-labeler bot added the area-crossgen2-coreclr label May 4, 2023

build-analysis bot mentioned this pull request May 4, 2023

Tracking issue for CI build timeouts #76454

Closed

davidwrighton added 2 commits May 8, 2023 13:21

Merge branch 'main' of github.com:dotnet/runtime into faster_type_load

c1c9983

Fix up some more issues

48dfa0e

runfoapp bot mentioned this pull request May 11, 2023

Infra improvements for Helix #68176

Closed

davidwrighton added 9 commits May 11, 2023 15:06

Bring back late virtual propagation in the presence of covariant over…

d274c51

…rides only

Check correct flag on EEClass

5bae879

Drive by fix for GetRestoredSlot. We don't need the handling of unres…

9f9e586

…tored slots anymore

Fix composite build with new tables

f5e72a6

Uniquify the mangled names

8eaf4a0

Add more __

ac7ec65

Initial pass at type skip verifation checker

1a57a0a

Fix logging and some correctness issues

5f770c6

Enable the more of type checking

52bd220

- Notably, the recursive stuff now works - Also fix a bug in constraint checking involving open types in the type system

AaronRobinsonMSFT reviewed May 20, 2023

View reviewed changes

davidwrighton added 2 commits May 30, 2023 14:23

Fix build breaks involving new feature of GenericParameterDesc

dfecc6f

Merge branch 'main' of github.com:dotnet/runtime into faster_type_load

6ef05ec

Add documentation for R2R format changes

eff81d5

Fix command line parameter to be more reasonable, and allow logging on command Fix the rest of issues noted in crossgen2 testing

dotnet deleted a comment from azure-pipelines bot Jun 1, 2023

AaronRobinsonMSFT reviewed Jun 1, 2023

View reviewed changes

docs/design/coreclr/botr/readytorun-format.md Outdated Show resolved Hide resolved

docs/design/coreclr/botr/readytorun-format.md Outdated Show resolved Hide resolved

docs/design/coreclr/botr/readytorun-format.md Outdated Show resolved Hide resolved

AaronRobinsonMSFT reviewed Jun 1, 2023

View reviewed changes

src/coreclr/inc/readytorun.h Show resolved Hide resolved

davidwrighton added 2 commits June 1, 2023 16:20

Fix nits noticed by Aaron

3b21879

Add some const correctness to the world

06b6ffd

davidwrighton marked this pull request as ready for review June 2, 2023 00:11

davidwrighton requested a review from MichalStrehovsky as a code owner June 2, 2023 00:11

MichalStrehovsky reviewed Jun 2, 2023

View reviewed changes

Fix issues noted by Michal, as well as remaining constrain checking i…

df3a418

…ssues

AaronRobinsonMSFT reviewed Jun 3, 2023

View reviewed changes

src/coreclr/vm/readytoruninfo.h Show resolved Hide resolved

AaronRobinsonMSFT reviewed Jun 3, 2023

View reviewed changes

src/coreclr/vm/readytoruninfo.h Show resolved Hide resolved

AaronRobinsonMSFT reviewed Jun 3, 2023

View reviewed changes

src/coreclr/vm/readytoruninfo.h Show resolved Hide resolved

davidwrighton added 3 commits June 5, 2023 14:11

Merge branch 'main' of github.com:dotnet/runtime into faster_type_load

75144c4

Code review details

1f44575

Merge branch 'main' of github.com:dotnet/runtime into faster_type_load

e6c0b4b

build-analysis bot mentioned this pull request Jun 9, 2023

Could not load file or assembly 'Microsoft.CodeAnalysis.NetAnalyzers #84995

Closed

Merge branch 'main' into faster_type_load

0f45839

AaronRobinsonMSFT approved these changes Jun 9, 2023

View reviewed changes

davidwrighton requested a review from trylek June 9, 2023 23:08

trylek reviewed Jun 9, 2023

View reviewed changes

...coreclr/tools/aot/ILCompiler.ReadyToRun/Compiler/DependencyAnalysis/ReadyToRun/HeaderNode.cs Show resolved Hide resolved

trylek reviewed Jun 9, 2023

View reviewed changes

trylek approved these changes Jun 9, 2023

View reviewed changes

davidwrighton added 2 commits June 12, 2023 13:01

Code review from trylek

2648965

Merge branch 'main' of https://github.com/dotnet/runtime into faster_…

bc127bb

…type_load

davidwrighton merged commit 8042fac into dotnet:main Jun 13, 2023

MichalStrehovsky mentioned this pull request Jun 14, 2023

Testing CI #87536

Closed

EgorBo mentioned this pull request Jun 19, 2023

R2R known test failure #87716

Closed

MichalStrehovsky mentioned this pull request Jul 14, 2023

Remove incorrect flag setting #88878

Merged

ghost locked as resolved and limited conversation to collaborators Jul 14, 2023

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Improve the performance of the type loader through various tweaks #85743

Improve the performance of the type loader through various tweaks #85743

davidwrighton commented May 4, 2023 •

edited

Loading

AaronRobinsonMSFT May 20, 2023

davidwrighton commented May 30, 2023

MichalStrehovsky Jun 2, 2023

davidwrighton Jun 2, 2023

MichalStrehovsky Jun 3, 2023

MichalStrehovsky Jun 2, 2023

davidwrighton Jun 2, 2023

MichalStrehovsky Jun 2, 2023

MichalStrehovsky Jun 2, 2023

davidwrighton Jun 2, 2023

AaronRobinsonMSFT left a comment

davidwrighton commented Jun 9, 2023

trylek Jun 9, 2023

davidwrighton Jun 12, 2023

trylek Jun 9, 2023

davidwrighton Jun 12, 2023

trylek Jun 9, 2023

trylek Jun 9, 2023

trylek left a comment

	public override DefType BaseType
	{
	get
	{
	if (_baseType == this)
	return InitializeBaseType();
	return _baseType;
	}
	}

		@@ -76,6 +76,8 @@ private InstantiatedGenericParameter(GenericParameterDesc genericParam)

		public override GenericConstraints Constraints => _genericParam.Constraints;

		public override TypeSystemEntity AssociatedTypeOrMethod => _genericParam.AssociatedTypeOrMethod;

Improve the performance of the type loader through various tweaks #85743

Improve the performance of the type loader through various tweaks #85743

Conversation

davidwrighton commented May 4, 2023 • edited Loading

Footnotes

Choose a reason for hiding this comment

davidwrighton commented May 30, 2023

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

AaronRobinsonMSFT left a comment

Choose a reason for hiding this comment

davidwrighton commented Jun 9, 2023

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

trylek left a comment

Choose a reason for hiding this comment

davidwrighton commented May 4, 2023 •

edited

Loading