JIT: Devirtualize non-shared generic virtual methods

[hez2010]

Enable devirtualization support for generic virtual methods.

When we see a base method having a method instantiation, we use FindOrCreateAssociatedMethodDesc to obtain the devirted method.

Also introduced a jit knob so that it can be turned off at any time.

AOT support is not included in this PR, which needs additional work in managed type system.

Also, if we end up with an instantiating stub (i.e. a shared generic method that requires runtime lookup), we don't have the correct generic context so we need to bail out for now.

Codegen example:

publicclassIntProcessor:VirtualGenericClass,IVritualGenericInterface{publicoverridevoidProcess<T>(Titem){Console.WriteLine(item.ToString());}}publicinterfaceIVritualGenericInterface{voidProcess<T>(Titem)whereT:notnull;}publicstaticvoidTest<T>(IVritualGenericInterfaceifce,Titem)whereT:notnull{ifce.Process(item);}publicstaticvoidTest<T>(VirtualGenericClassbaseClass,Titem)whereT:notnull{baseClass.Process(item);}staticvoidTest(){IVritualGenericInterfacei=newIntProcessor();Test(i,42);VirtualGenericClassc=newIntProcessor();Test(c,42);}

Codegen diff:

 G_M27646_IG01: ;; offset=0x0000
- push rsi
push rbx
- sub rsp, 40+ sub rsp, 32- ;; size=6 bbWeight=1 PerfScore 2.25+ ;; size=5 bbWeight=1 PerfScore 1.25-G_M27646_IG02: ;; offset=0x0006+G_M27646_IG02: ;; offset=0x0005- mov rbx, 0x7FFE0803F318 ; Program+IntProcessor+ mov rbx, 0x221CA429C38 ; 'System.Int32'
mov rcx, rbx
- call CORINFO_HELP_NEWSFAST+ call [System.Console:WriteLine(System.Object)]- mov rsi, rax+ mov ecx, 42- mov rcx, rsi+ call [System.Number:Int32ToDecStr(int):System.String]- mov rdx, 0x7FFE0803F100 ; Program+IVritualGenericInterface+ mov rcx, rax- mov r8, 0x7FFE0803F6A8 ; token handle+ call [System.Console:WriteLine(System.String)]- call CORINFO_HELP_VIRTUAL_FUNC_PTR- mov rcx, rsi- mov edx, 42- call rax
mov rcx, rbx
- call CORINFO_HELP_NEWSFAST+ call [System.Console:WriteLine(System.Object)]- mov rbx, rax+ mov ecx, 42- mov rcx, rbx+ call [System.Number:Int32ToDecStr(int):System.String]- mov rdx, 0x7FFE0803EF30 ; Program+VirtualGenericClass+ mov rcx, rax- mov r8, 0x7FFE0803F8A8 ; token handle+ call [System.Console:WriteLine(System.String)]- call CORINFO_HELP_VIRTUAL_FUNC_PTR- mov rcx, rbx- mov edx, 42- call rax
nop
- ;; size=109 bbWeight=1 PerfScore 14.00+ ;; size=69 bbWeight=1 PerfScore 20.00-G_M27646_IG03: ;; offset=0x0073+G_M27646_IG03: ;; offset=0x004A- add rsp, 40+ add rsp, 32
pop rbx
- pop rsi
ret
- ;; size=7 bbWeight=1 PerfScore 2.25+ ;; size=6 bbWeight=1 PerfScore 1.75

Contributes to #112596

cc: @dotnet/jit-contrib

[Copilot]

Pull request overview

This PR enables devirtualization support for generic virtual methods in the JIT compiler, allowing calls to generic virtual methods to be devirtualized when the exact type is known at JIT time. This optimization eliminates virtual dispatch overhead and enables further optimizations like inlining.

Key changes:

• Removes the assertion that previously blocked generic method devirtualization
• Generalizes array interface devirtualization to support generic virtual methods by renaming wasArrayInterfaceDevirt to needsMethodContext
• Adds runtime lookup support for generic method instantiation parameters

Reviewed changes

Copilot reviewed 13 out of 13 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
src/coreclr/vm/jitinterface.cpp	Removes assertion blocking generic method devirtualization and adds logic to handle generic virtual methods using FindOrCreateAssociatedMethodDesc
src/coreclr/inc/corinfo.h	Renames wasArrayInterfaceDevirt to needsMethodContext to generalize the field meaning
src/coreclr/jit/jitconfigvalues.h	Adds JitEnableGenericVirtualDevirtualization configuration flag to control the feature
src/coreclr/jit/gentree.h	Adds IsGenericVirtual() helper method to identify generic virtual method calls
src/coreclr/jit/gentree.cpp	Updates IsDevirtualizationCandidate() to include generic virtual methods (non-AOT only)
src/coreclr/jit/importercalls.cpp	Implements devirtualization logic for generic virtual methods including runtime lookup handling, updates comments and variable names, and introduces DEVIRT label for control flow
src/coreclr/jit/inline.h	Renames arrayInterface field to needsMethodContext in InlineCandidateInfo struct
src/coreclr/jit/indirectcalltransformer.cpp	Updates to use renamed needsMethodContext field
src/coreclr/tools/Common/JitInterface/CorInfoTypes.cs	Updates managed struct definition to match renamed field
src/coreclr/tools/Common/JitInterface/CorInfoImpl.cs	Updates to use renamed field in managed implementation
src/coreclr/tools/superpmi/superpmi-shared/agnostic.h	Updates SuperPMI data structure with renamed field
src/coreclr/tools/superpmi/superpmi-shared/methodcontext.cpp	Updates SuperPMI recording/replay to use renamed field

[hez2010]

@MihuBot

[davidwrighton]

@hez2010 sorry about the delay, I got stuck in a bunch of work just before the holidays and wasn't able to do the review before vacation began.

[jakobbotsch]

I'm guessing the crossgen issues are related, I don't see them on other PRs. Can you see if you can reproduce them?

[hez2010]

I'm guessing the crossgen issues are related, I don't see them on other PRs. Can you see if you can reproduce them?

I managed to reproduce the freebsd one on my local cross-build environment. Will investigate it soon to see whether it's related or not.

[hez2010]

We hit the assertion here:

/home/i/runtime/src/coreclr/jit/importercalls.cpp:8666
Assertion failed 'call->IsVirtualStub()' in 'System.Threading.StackHelper:CallOnEmptyStack[System.__Canon,System.__Canon,System.__Canon,System.__Canon](System.Func`4[System.__Canon,System.__Canon,System.__Canon,System.__Canon],System.__Canon,System.__Canon,System.__Canon):System.__Canon' during 'Importation' (IL size 116; hash 0x7bf56612; FullOpts)
/home/i/runtime/src/coreclr/jit/importercalls.cpp:8666
Assertion failed 'call->IsVirtualStub()' in 'System.Threading.StackHelper:CallOnEmptyStack[System.__Canon,System.__Canon](System.Func`2[System.__Canon,System.__Canon],System.__Canon):System.__Canon' during 'Importation' (IL size 102; hash 0x657768d4; FullOpts)

https://github.com/hez2010/runtime/blob/40d7569852bd8cecef706532d374ab0e8f7e53f2/src/coreclr/jit/importercalls.cpp#L8664-L8670

I think GVM applies the same with virtual stub, where in R2R mode, we might see GVM calls to non-virtuals. I think we should change the assertion to call->IsVirtualStub() || call->IsGenericVirtual(). What do you think? @jakobbotsch

[jakobbotsch]

I think GVM applies the same with virtual stub, where in R2R mode, we might see GVM calls to non-virtuals. I think we should change the assertion to call->IsVirtualStub() || call->IsGenericVirtual(). What do you think? @jakobbotsch

Seems reasonable to me.

[hez2010]

CI is green now.
By the way, do you know why the assertion message didn't show up in the CI log before? It makes it hard to spot the cause of bug without a local repro. (I guess it's because MSBuild somehow omitted all the logs from crossgen2?)

[jakobbotsch]

/azp run runtime-coreclr jitstress, runtime-coreclr outerloop

[azure-pipelines]

Azure Pipelines successfully started running 2 pipeline(s).

[jakobbotsch]

CI is green now. By the way, do you know why the assertion message didn't show up in the CI log before? It makes it hard to spot the cause of bug without a local repro. (I guess it's because MSBuild somehow omitted all the logs from crossgen2?)

I am not sure, perhaps that is something specific to how the crossgen2 testing is done for the community targets.

I kicked off some more testing. There are some known test failures in jitstress (for async), so might take some filtering once it's done to validate things.

[hez2010]

I kicked off some more testing. There are some known test failures in jitstress (for async), so might take some filtering once it's done to validate things.

All failures seem to be preexisting and unrelated.

[jakobbotsch]

/ba-g Failure is #122345 that is not being matched automatically

[jakobbotsch]

Thanks!