Devirtualization isn't acting on devirtualizable calls past the first

[NinoFloris]

Ran from a VM due to dotnet/BenchmarkDotNet#1499

BenchmarkDotNet=v0.12.1, OS=Windows 10.0.17763.1039 (1809/October2018Update/Redstone5)
Intel Core i7-4980HQ CPU 2.80GHz (Haswell), 1 CPU, 4 logical and 4 physical cores
.NET Core SDK=5.0.100-preview.6.20318.15
  [Host]   : .NET Core 5.0.0 (CoreCLR 5.0.20.30506, CoreFX 5.0.20.30506), X64 RyuJIT
  ShortRun : .NET Core 5.0.0 (CoreCLR 5.0.20.30506, CoreFX 5.0.20.30506), X64 RyuJIT

Job=ShortRun  IterationCount=3  LaunchCount=1  
WarmupCount=3  

Method	Mean	Error	StdDev	Code Size
Combined	3.6617 ns	0.9868 ns	0.0541 ns	42 B
Separate	3.6139 ns	0.4095 ns	0.0224 ns	42 B
JustMe	0.0000 ns	0.0000 ns	0.0000 ns	6 B

The idea is to do a two step virtual call, once dispatching from this (Combined) and once to a separate sealed class (Separate), JustMe is the control as it devirtualizes correctly.

using System;
using System.Runtime.CompilerServices;
using BenchmarkDotNet.Attributes;
using BenchmarkDotNet.Configs;
using BenchmarkDotNet.Jobs;
using BenchmarkDotNet.Running;

namespace test2
{
    public interface IMe 
    {
        bool Hello();   
    }

    public interface IYou 
    {
        IMe You { get; }   
    }

    public sealed class Combined : IMe, IYou
    {
        public IMe You => this;
        public bool Hello() => true;
    }
    
    public sealed class MeImpl : IMe
    {
        public bool Hello() => true;
    }
    
    public sealed class YouImpl : IYou
    {
        public IMe You => new MeImpl();
    }

    [ShortRunJob, DisassemblyDiagnoser(exportGithubMarkdown: true)]
    public class Program
    {
        [Benchmark]
        public bool Combined() => new Combined().You.Hello();
        
        [Benchmark]
        public bool Separate() => new YouImpl().You.Hello();
        
        [Benchmark]
        public bool JustMe() => new MeImpl().Hello();
        
        static void Main(string[] args)
        {
            BenchmarkRunner.Run<Program>();
        }
    }
}

And the resulting asm:

.NET Core 5.0.0 (CoreCLR 5.0.20.30506, CoreFX 5.0.20.30506), X64 RyuJIT

; test2.Program.Combined()
       sub       rsp,28
       mov       rcx,offset MT_test2.Combined
       call      CORINFO_HELP_NEWSFAST
       mov       rcx,rax
       mov       rax,[7FFA0807B738]
       add       rsp,28
       jmp       rax
; Total bytes of code 36

; test2.Combined.Hello()
       mov       eax,1
       ret
; Total bytes of code 6

.NET Core 5.0.0 (CoreCLR 5.0.20.30506, CoreFX 5.0.20.30506), X64 RyuJIT

; test2.Program.Separate()
       sub       rsp,28
       mov       rcx,offset MT_test2.MeImpl
       call      CORINFO_HELP_NEWSFAST
       mov       rcx,rax
       mov       rax,[7FFA0808B7F0]
       add       rsp,28
       jmp       rax
; Total bytes of code 36

; test2.MeImpl.Hello()
       mov       eax,1
       ret
; Total bytes of code 6

.NET Core 5.0.0 (CoreCLR 5.0.20.30506, CoreFX 5.0.20.30506), X64 RyuJIT

; test2.Program.JustMe()
       mov       eax,1
       ret
; Total bytes of code 6

Expectation would be that all benchmarks compile down to mov eax, 1

category:cq
theme:devirtualization
skill-level:expert
cost:large
impact:medium

[AndyAyersMS]

It's not readily apparent from the disassembly above, but both calls are in fact devirtualized. But this happens "late" as the object being dispatched is the return value from a call, and because of this, we currently can't do inlining or box removal optimizations.

;; early attempt during importation

impDevirtualizeCall: Trying to devirtualize interface call:
    class for 'this' is IMe (attrib 00200400)
    base method is IMe::Hello
No unique implementor of interface 00000000D1FFAB1E (IMe), sorry

;; late attempt after inlining

impDevirtualizeCall: Trying to devirtualize interface call:
    class for 'this' is Combined [exact] (attrib 20000010)
    base method is IMe::Hello
--- base class is interface
    devirt to Combined::Hello -- exact
               [000009] --C-G-------              *  CALLV stub int    IMe.Hello
               [000006] ------------ this in rcx  \--*  LCL_VAR   ref    V02 tmp1         
    exact; can devirtualize
... after devirt...
               [000009] --C-G-------              *  CALL nullcheck int    Combined.Hello
               [000006] ------------ this in rcx  \--*  LCL_VAR   ref    V02 tmp1         

The calls end up getting invoked via an indirection cell which makes them look like they're possibly still virtual.

[AndyAyersMS]

Linked this into the devirtualization "todo" issue #7541. Marking as future.

[hez2010]

This has been addressed by #110827 and we should close it. (Perhaps update the todo in #7541 as all)

[AndyAyersMS]

Yep.

This was likely "fixed" by dynamic PGO a while back, eg

Runtime=.NET 9.0

Method	Job	Toolchain	Mean	Error	StdDev	Median	Ratio	RatioSD
Combined	Job-JXEEPS	net10.0	0.0007 ns	0.0027 ns	0.0026 ns	0.0000 ns	0.04	0.12
Combined	Job-SGEPJK	net9.0	0.0207 ns	0.0032 ns	0.0027 ns	0.0214 ns	1.02	0.18
Separate	Job-JXEEPS	net10.0	0.0240 ns	0.0165 ns	0.0138 ns	0.0200 ns	?	?
Separate	Job-SGEPJK	net9.0	0.0019 ns	0.0061 ns	0.0051 ns	0.0000 ns	?	?
JustMe	Job-JXEEPS	net10.0	0.0000 ns	0.0000 ns	0.0000 ns	0.0000 ns	0.000	0.00
JustMe	Job-SGEPJK	net9.0	0.0301 ns	0.0199 ns	0.0176 ns	0.0230 ns	1.511	1.82

but if you disable tiered compilation:

EnvironmentVariables=DOTNET_TieredCompilation=0 Runtime=.NET 9.0

Method	Job	Toolchain	Mean	Error	StdDev	Median	Ratio	RatioSD
Combined	Job-BRSGZU	net10.0	0.0194 ns	0.0013 ns	0.0010 ns	0.0194 ns	0.004	0.00
Combined	Job-DNGAOV	net9.0	4.9356 ns	0.1226 ns	0.1147 ns	4.8767 ns	1.000	0.03
Separate	Job-BRSGZU	net10.0	0.0179 ns	0.0031 ns	0.0026 ns	0.0173 ns	0.004	0.00
Separate	Job-DNGAOV	net9.0	4.9616 ns	0.0981 ns	0.0870 ns	4.9532 ns	1.000	0.02
JustMe	Job-BRSGZU	net10.0	0.0229 ns	0.0117 ns	0.0098 ns	0.0180 ns	1.25	0.59
JustMe	Job-DNGAOV	net9.0	0.0197 ns	0.0083 ns	0.0069 ns	0.0170 ns	1.08	0.44