Faster type load for scenarios made more common by generic math #54588

davidwrighton · 2021-06-23T01:40:31Z

Change interface map layout in two interesting ways

For interface maps defined in System.Private.CoreLib, rely on the C# compiler to prevent any ambiguous duplicates, and to find the full interface expansion, instead of expanding it within the type loader See code marked with #SpecialCorelibInterfaceExpansionAlgorithm

Note that this optimization is only applied for ValueTypes as the presence of inheritance makes the optimization much more complex in many cases, and isn't needed.
This optimization reduces the amount of parsing of the interface implementation table must be done for valuetypes in CoreLib. In particular, with the new interfaces that are added as part of Expose and implement generic math interfaces on core types #54650 there is a very deep interface hierarchy that requires a great deal of scanning. As those interfaces are added to all the primitive types, the impact on startup performance is significant and measurable.

For interface map expansion that follows the curiously recurring generic pattern, place the open instantiation of the type in the interface map instead of the the exact instantiation, and update all places in the runtime which consider the interface map to deal with that change (Mostly by adding special purpose logic to work with the special marker type in the interface map, but there is also logic to simply force the exact interface type to be loaded, when working with the partially loaded type is not quite good enough, or excessively complex)

This optimization reduces the set of interface types that need to be loaded if they are unused. Of particular benefit are the numerous interfaces associated with the primitive types that are added as part of Expose and implement generic math interfaces on core types #54650.

Performance of launching an extremely simple .NET process (process with empty main method). Results acquired using local testing on my developer machine, using a simple script that launches the process 500 times in a row.

Before #54650	After #54650	After #54650 and #54588 (this pr)
14.1ms	16.5ms	14.3ms

src/coreclr/vm/methodtablebuilder.cpp

tannergooding · 2021-06-23T02:14:12Z

How does this impact startup? Does it remove the regression that will get introduced?

davidwrighton · 2021-06-23T15:49:40Z

This fixes 90% of the startup slowdown introduced by the generic math feature. Of course, it also doesn't quite work as can be seen from the tests, but it looks like its a feasible direction. The other thing is, there is a runtime perf penalty which isn't shown here. Of particular concern is the change to interface casting.

trylek

Looks great to me, thank you! I have only noticed a few formatting nits, no functional issues.

trylek · 2021-06-25T21:04:06Z

src/coreclr/vm/methodtablebuilder.cpp

+                    if (uninstGenericCase && pItfPossiblyApprox->HasInstantiation() && pItfPossiblyApprox->ContainsGenericVariables())
+                    {
+                        // We allow a limited set of interface generic shapes with type variables. In particular, we require the instantiations to be exactly type variables
+                        // the current type


Nit - the two lines of this code comment seem somewhat disjoint, is there a missing "of" between the lines?

trylek · 2021-06-25T21:07:35Z

src/coreclr/vm/siginfo.cpp

@@ -989,7 +989,8 @@ TypeHandle SigPointer::GetTypeHandleThrowing(
                 BOOL                        dropGenericArgumentLevel/*=FALSE*/,
                 const Substitution *        pSubst/*=NULL*/,
                 // ZapSigContext is only set when decoding zapsigs
-                 const ZapSig::Context *     pZapSigContext) const
+                 const ZapSig::Context *     pZapSigContext,
+                 MethodTable *pMTInterfaceMapOwner) const


Nit - indentation?

trylek · 2021-06-25T21:10:11Z

src/coreclr/vm/methodtablebuilder.cpp

+    // to have found all of the interfaces that the type implements, and to place them in the interface list itself. Also
+    // we can assume no ambiguous interfaces
+    // Code related to this is marked with #SpecialCorelibInterfaceExpansionAlgorithm
+   if (!(GetModule()->IsSystem() && IsValueClass()))


Nit - indentation

1. For interface maps defined in System.Private.CoreLib, rely on the C# compiler to prevent any ambiguous duplicates, and to find the full interface expansion, instead of expanding it within the type loader See code marked with #SpecialCorelibInterfaceExpansionAlgorithm 2. For interface map expansion that follows the curiously recurring generic pattern, place the open instantiation of the type in the interface map instead of the the exact instantiation, and update all places in the runtime which consider the interface map to deal with that change (Mostly by being able to not resovle to an exact interface, but in some cases there is a helper which will do exactly that)

- Re-enable fast path for NonTrivialCasting for interface misses - Adjust SetInterface usage to only occur for scenarios where the found type if fully loaded - Only use #SpecialCorelibInterfaceExpansionAlgorithm on ValueTypes, turns out derived types have all the fun problems - Adjust casting logic to work with the approx interface and owner pair. This avoid the need to load types in CanCastTo, which is important as CanCastTo is used heavily during type load in scenario where further type loading is prohibited - Give checks for the special marker type a unique name instead of just re-using IsGenericTypeDefinition. Use IsSpecialMarkerTypeForGenericCasting instead.

…bugger2 * origin/main: (78 commits) Fix unreached during dump. (dotnet#54861) Fix lowering usage of an unset LSRA field. (dotnet#54731) Fix setting breakpoints on AVX 256 instructions and other 32 byte immediate instructions (dotnet#54786) Add perf_slow yaml (dotnet#54853) Faster type load for scenarios made more common by generic math (dotnet#54588) Make sure we consider buffer length when marshalling back Unicode ByValTStr fields (dotnet#54695) Add YieldProcessor implementation for arm (dotnet#54829) Remove ActiveIssue for dotnet#50968 (dotnet#54831) Enable System.Text.Json tests for Wasm AOT (dotnet#54833) Remove ActiveIssue for dotnet#51723 (dotnet#54830) Fix load exception on generic covariant return type (dotnet#54790) Obsolete X509Certificate2.PrivateKey and PublicKey.Key. (dotnet#54562) First round of converting System.Drawing.Common to COMWrappers (dotnet#54636) Fix alloc-dealloc mismatches (dotnet#54701) Add one-shot ECB methods [Mono] MSBuild Task housekeeping (dotnet#54485) Move iOS/tvOS simulator AOT imports in the Mono workload (dotnet#54821) Remove unnecessary char[] allocation from Uri.GetRelativeSerializationString (dotnet#54799) Reduce overhead of Enumerable.Chunk (dotnet#54782) Fix EnumMemberRefs always returning NULL (dotnet#54805) ...

dotnet-issue-labeler bot added the area-TypeSystem-coreclr label Jun 23, 2021

lambdageek reviewed Jun 23, 2021

View reviewed changes

src/coreclr/vm/methodtablebuilder.cpp Outdated Show resolved Hide resolved

davidwrighton changed the title ~~[WIP] Faster type load~~ Faster type load for scenarios made more common by generic math Jun 24, 2021

davidwrighton requested a review from trylek June 24, 2021 00:18

davidwrighton force-pushed the faster_type_load branch from 9aa4be4 to 3e329f9 Compare June 25, 2021 00:40

runfoapp bot mentioned this pull request Jun 25, 2021

Test failure System.Threading.Tests.PeriodicTimerTests.WaitForNextTickAsync_CanceledWaitThenWaitAgain_Succeeds #54713

Closed

trylek approved these changes Jun 25, 2021

View reviewed changes

davidwrighton added 4 commits June 25, 2021 15:12

Adjust casting for new NonTrivial cases

0f0e249

Fix build issues

a615e9a

davidwrighton force-pushed the faster_type_load branch from 3e329f9 to 1a32f13 Compare June 25, 2021 22:12

Code review comments

33774b1

davidwrighton marked this pull request as ready for review June 25, 2021 22:15

davidwrighton merged commit 3144188 into dotnet:main Jun 28, 2021

tannergooding mentioned this pull request Jun 28, 2021

Expose and implement generic math interfaces on core types #54650

Merged

jkotas mentioned this pull request Jul 7, 2021

Tests failed with error message: violates the constraint of type parameter 'T'. #54941

Closed

This was referenced Jul 8, 2021

Curiously recurring pattern when abstracted through interfaces does not work #55361

Closed

dotnet hang in JIT #55323

Closed

ghost locked as resolved and limited conversation to collaborators Jul 28, 2021

davidwrighton deleted the faster_type_load branch April 13, 2023 18:48

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Faster type load for scenarios made more common by generic math #54588

Faster type load for scenarios made more common by generic math #54588

davidwrighton commented Jun 23, 2021 •

edited

Loading

tannergooding commented Jun 23, 2021

davidwrighton commented Jun 23, 2021

trylek left a comment

trylek Jun 25, 2021

trylek Jun 25, 2021

trylek Jun 25, 2021

Faster type load for scenarios made more common by generic math #54588

Faster type load for scenarios made more common by generic math #54588

Conversation

davidwrighton commented Jun 23, 2021 • edited Loading

tannergooding commented Jun 23, 2021

davidwrighton commented Jun 23, 2021

trylek left a comment

Choose a reason for hiding this comment

trylek Jun 25, 2021

Choose a reason for hiding this comment

trylek Jun 25, 2021

Choose a reason for hiding this comment

trylek Jun 25, 2021

Choose a reason for hiding this comment

davidwrighton commented Jun 23, 2021 •

edited

Loading