NativeAOT: devirtualize isinst/castclass for monomorphic cases

[EgorBo]

Closes#80828

Luckily, we already have JIT-EE API to get exact list of implementations/subclasses for given type.

interfaceIMyInterface{}// this interface has a single implementationclassProgram:IMyInterface{staticvoidMain(){Case1(newProgram());Case2(newProgram());}[MethodImpl(MethodImplOptions.NoInlining)]staticboolCase1(objecto)=>oisIMyInterface;[MethodImpl(MethodImplOptions.NoInlining)]staticIMyInterfaceCase2(objecto)=>(IMyInterface)o;}

Previous NativeAOT codegen:

; Method Program:Case1(System.Object):boolsubrsp,40movrdx,rcxlearcx,[(reloc)]call CORINFO_HELP_ISINSTANCEOFINTERFACEtestrax,rax setne almovzxrax,aladdrsp,40ret; Total bytes of code: 33; Method Program:Case2(System.Object):IMyInterfacesubrsp,40movrdx,rcxlearcx,[(reloc)]call CORINFO_HELP_CHKCASTINTERFACEnopaddrsp,40ret; Total bytes of code: 25

New NativeAOT codegen:

; Method Program:Case1(System.Object):booltestrcx,rcxje SHORT G_M13626_IG05learax,[(reloc)]cmp qword ptr [rcx],raxje SHORT G_M13626_IG05xorrcx,rcxG_M13626_IG05:xoreax,eaxtestrcx,rcx setne alret; Total bytes of code: 28; Method Program:Case2(System.Object):IMyInterfacesubrsp,40movrdx,rcxmovrax,rdxtestrax,raxje SHORT G_M12806_IG05learcx,[(reloc)]cmp qword ptr [rax],rcxje SHORT G_M12806_IG05movrcx, qword ptr [rsp+20H]call CORINFO_HELP_CHKCASTINTERFACE ;; this is used to throw InvalidCastException onlyG_M12806_IG05:nopaddrsp,40ret; Total bytes of code: 43

[ghost]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch, @kunalspathak
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Closes #80828

Luckily, we already have JIT-EE API to get exact list of implementations/subclasses for given type.

interfaceIMyInterface{}// this interface has a single implementationclassProgram:IMyInterface{staticvoidMain(){Case1(newProgram());Case2(newProgram());}[MethodImpl(MethodImplOptions.NoInlining)]staticboolCase1(objecto)=>oisIMyInterface;[MethodImpl(MethodImplOptions.NoInlining)]staticIMyInterfaceCase2(objecto)=>(IMyInterface)o;}

Previous NativeAOT codegen:

; Method Program:Case1(System.Object):boolsubrsp,40movrdx,rcxlearcx,[(reloc)]call CORINFO_HELP_ISINSTANCEOFINTERFACEtestrax,rax setne almovzxrax,aladdrsp,40ret; Total bytes of code: 33; Method Program:Case2(System.Object):IMyInterfacesubrsp,40movrdx,rcxlearcx,[(reloc)]call CORINFO_HELP_CHKCASTINTERFACEnopaddrsp,40ret; Total bytes of code: 25

New NativeAOT codegen:

; Method Program:Case1(System.Object):booltestrcx,rcxje SHORT G_M13626_IG05learax,[(reloc)]cmp qword ptr [rcx],raxje SHORT G_M13626_IG05xorrcx,rcxG_M13626_IG05:xoreax,eaxtestrcx,rcx setne alret; Total bytes of code: 28; Method Program:Case2(System.Object):IMyInterfacesubrsp,40movrdx,rcxmovrax,rdxtestrax,raxje SHORT G_M12806_IG05learcx,[(reloc)]cmp qword ptr [rax],rcxje SHORT G_M12806_IG05movrcx, qword ptr [rsp+20H]call CORINFO_HELP_CHKCASTINTERFACE ;; this is used to throw InvalidCastException onlyG_M12806_IG05:nopaddrsp,40ret; Total bytes of code: 43

Author:	EgorBo
Assignees:	-
Labels:	area-CodeGen-coreclr
Milestone:	-

[am11]

Is Case2 regression avoidable?

[AndyAyersMS]

Is Case2 regression avoidable?

The code is bigger but should also be faster, since we will rarely need to call the helper (only if the cast will throw).

Though I'm not quite sure why we're shuffling those registers about like we do...

[MichalStrehovsky]

Curious: if we had hot/cold splitting support in NativeAOT (#77583), would that place the ;; this is used to throw InvalidCastException only in the cold region, or would that be more work?

[jakobbotsch]

Is there any reason this optimization shouldn't kick in for any other potential EE that returns 1 from getExactClasses?

[EgorBo]

@jakobbotsch purely throughput reasons - 2 ifs + non inlined call getExactClasses since I don't think it will be useful outside of NativeAOT anytime soon 🙂

[jakobbotsch]

Seems a bit premature? If I remove the NAOT check and hack SPMI to always return 0 on replay of getExactClasses I see no significant TP differences. IMO, as the general rule (no need to change this now) we should not be sacrificing abstractions unless they have demonstrable impact.

[EgorBo]

no strong opinion on that, will remove as part of some next PR

[EgorBo]

Curious: if we had hot/cold splitting support in NativeAOT (#77583), would that place the ;; this is used to throw InvalidCastException only in the cold region, or would that be more work?

It doesn't look like it's handled in hot/cold splitting now, but at least it's still considered as "cold" in jit