Enable FEATURE_MULTICASTSTUB_AS_IL for Windows x86 #104192

huoyaoyuan · 2024-06-29T11:10:10Z

Local benchmark result for Action<int> on x86:

Method	Job	Toolchain	Mean	Error	StdDev	Ratio
SingleArg	Job-OSFVND	\x86-3-TuneBranchPrediction\corerun.exe	3.464 ns	0.0127 ns	0.0119 ns	0.95
SingleArg	Job-KBQPJV	\x86-main\corerun.exe	3.661 ns	0.0222 ns	0.0208 ns	1.00

ManyCast	Job-OSFVND	\x86-3-TuneBranchPrediction\corerun.exe	8.188 ns	0.0944 ns	0.0883 ns	0.97
ManyCast	Job-KBQPJV	\x86-main\corerun.exe	8.483 ns	0.0888 ns	0.0831 ns	1.00

ManyArg_RetFPU	Job-OSFVND	\x86-3-TuneBranchPrediction\corerun.exe	5.566 ns	0.0338 ns	0.0316 ns	0.96
ManyArg_RetFPU	Job-KBQPJV	\x86-main\corerun.exe	5.771 ns	0.0212 ns	0.0177 ns	1.00

huoyaoyuan · 2024-06-29T11:11:56Z

@EgorBot -intel -amd -arm64

using BenchmarkDotNet.Attributes;
using System;

namespace BenchmarkGround
{
    public struct GCStruct
    {
        public object a, b, c, d, e, f, h, g;
    }

    public class Bench
    {
        private readonly object obj = new object();

        private readonly Guid guid = Guid.NewGuid();

        private readonly GCStruct gcStruct = new GCStruct
        {
            a = new object(),
            g = new object()
        };

        private readonly Action<int> singleArg = (Action<int>)delegate { } + delegate { };

        private readonly Action<int> manyCast = (Action<int>)delegate { }
        +
            delegate { }
        + delegate { }
        + delegate { }
        + delegate { }
        + delegate { };

        private readonly Func<int, double, Guid, GCStruct, object, double> manyArg_RetFPU = (Func<int, double, Guid, GCStruct, object, double>)
            delegate { return 123.456; }
        + delegate { return 654.321; };

        [Benchmark]
        public void SingleArg() => singleArg(42);

        [Benchmark]
        public void ManyCast() => manyCast(42);

        [Benchmark]
        public double ManyArg_RetFPU() => manyArg_RetFPU(42, 123.0, guid, gcStruct, obj);
    }
}

dotnet-policy-service · 2024-06-29T11:14:38Z

Tagging subscribers to this area: @mangod9
See info in area-owners.md if you want to be subscribed.

EgorBot · 2024-06-29T11:54:29Z

Benchmark results on Intel

BenchmarkDotNet v0.13.12, Ubuntu 22.04.4 LTS (Jammy Jellyfish)
Intel Xeon Platinum 8370C CPU 2.80GHz, 1 CPU, 8 logical and 4 physical cores
  Job-FAGYWG : .NET 9.0.0 (42.42.42.42424), X64 RyuJIT AVX-512F+CD+BW+DQ+VL+VBMI
  Job-LQQTSO : .NET 9.0.0 (42.42.42.42424), X64 RyuJIT AVX-512F+CD+BW+DQ+VL+VBMI

Method	Toolchain	Mean	Error	Ratio
SingleArg	Main	8.084 ns	0.0012 ns	1.00
SingleArg	PR	6.078 ns	0.0012 ns	0.75

ManyCast	Main	19.581 ns	0.0313 ns	1.00
ManyCast	PR	14.887 ns	0.1134 ns	0.76

ManyArg_RetFPU	Main	14.538 ns	0.0340 ns	1.00
ManyArg_RetFPU	PR	11.450 ns	0.0147 ns	0.79

BDN_Artifacts.zip

EgorBot · 2024-06-29T11:56:24Z

Benchmark results on Arm64

BenchmarkDotNet v0.13.12, Ubuntu 22.04.4 LTS (Jammy Jellyfish)
Unknown processor
  Job-LOUVQC : .NET 9.0.0 (42.42.42.42424), Arm64 RyuJIT AdvSIMD
  Job-TLYWZG : .NET 9.0.0 (42.42.42.42424), Arm64 RyuJIT AdvSIMD

Method	Toolchain	Mean	Error	Ratio
SingleArg	Main	15.831 ns	0.0109 ns	1.00
SingleArg	PR	8.772 ns	0.0017 ns	0.55

ManyCast	Main	38.106 ns	0.0234 ns	1.00
ManyCast	PR	21.135 ns	0.0015 ns	0.55

ManyArg_RetFPU	Main	17.636 ns	0.0024 ns	1.00
ManyArg_RetFPU	PR	16.759 ns	0.0071 ns	0.95

BDN_Artifacts.zip

EgorBot · 2024-06-29T11:56:44Z

Benchmark results on Amd

BenchmarkDotNet v0.13.12, Ubuntu 22.04.4 LTS (Jammy Jellyfish)
AMD EPYC 7763, 1 CPU, 8 logical and 4 physical cores
  Job-DXSXUL : .NET 9.0.0 (42.42.42.42424), X64 RyuJIT AVX2
  Job-BIEGFU : .NET 9.0.0 (42.42.42.42424), X64 RyuJIT AVX2

Method	Toolchain	Mean	Error	Ratio
SingleArg	Main	13.008 ns	0.1016 ns	1.00
SingleArg	PR	6.490 ns	0.0071 ns	0.50

ManyCast	Main	100.460 ns	0.5200 ns	1.00
ManyCast	PR	18.849 ns	0.0302 ns	0.19

ManyArg_RetFPU	Main	14.358 ns	0.0464 ns	1.00
ManyArg_RetFPU	PR	12.682 ns	0.0120 ns	0.88

BDN_Artifacts.zip

jkotas · 2024-06-29T12:15:38Z

src/coreclr/vm/comdelegate.cpp

+        pCode->EmitBRTRUE(invokeTraceHelper);
+        pCode->EmitBR(debuggerCheckEnd); // Tune branch prediction to prefer non-debugging path
+
+        pCode->EmitLabel(invokeTraceHelper);
+
        pCode->EmitLoadThis();
        pCode->EmitLDLOC(dwLoopCounterNum);
        pCode->EmitCALL(METHOD__STUBHELPERS__MULTICAST_DEBUGGER_TRACE_HELPER, 2, 0);


Suggested change

pCode->EmitBRTRUE(invokeTraceHelper);

pCode->EmitBR(debuggerCheckEnd); // Tune branch prediction to prefer non-debugging path

pCode->EmitLabel(invokeTraceHelper);

pCode->EmitLoadThis();

pCode->EmitLDLOC(dwLoopCounterNum);

pCode->EmitCALL(METHOD__STUBHELPERS__MULTICAST_DEBUGGER_TRACE_HELPER, 2, 0);

pCode->EmitBRTRUE(invokeTraceHelper);

And move the debugging path to be after RET:

pCode->EmitRET(); #ifdef DEBUGGING_SUPPORTED // Emit debugging support at the end of the method for better perf pCode->EmitLabel(invokeTraceHelper); pCode->EmitLoadThis(); pCode->EmitLDLOC(dwLoopCounterNum); pCode->EmitCALL(METHOD__STUBHELPERS__MULTICAST_DEBUGGER_TRACE_HELPER, 2, 0); pCode->EmitBR(debuggerCheckEnd); #endif

This should be even better

That was exactly what I did locally, and resulted in exact same codegen with current, at least for x86. The compiled asm code block order did follow IL block order exactly.

@dotnet/jit-contrib Do you have any suggestion on this?

This code has atypical loop. It looks like the JIT tried to reorder the basic blocks to turn it a more regular loop. I am not sure whether there is anything to fix in the JIT (the JIT would have to have profile data to do better).

It may be still worth it to move TraceHelper call to be at the end in IL. It provides stronger hint about the desired code layout to the JIT and makes this optimization a bit less fragile.

Well normalizing the loop closer to a for loop results in about 5% improvement for ManyArg case, but 5% regression for ManyCast.

Codegen for x64 now:

; Assembly listing for method System.Action:IL_STUB_MulticastDelegate_Invoke():this (FullOpts) ; Emitting BLENDED_CODE for X64 with AVX - Windows ; FullOpts code ; optimized code ; optimized using Synthesized PGO ; rsp based frame ; partially interruptible ; with Synthesized PGO: fgCalledCount is 100 ; No PGO data G_M000_IG01: ;; offset=0x0000 push rsi push rbx sub rsp, 40 mov rbx, rcx G_M000_IG02: ;; offset=0x0009 xor esi, esi cmp qword ptr [rbx+0x30], 0 jle SHORT G_M000_IG05 G_M000_IG03: ;; offset=0x0012 test dword ptr [(reloc 0x7ff824c55210)], 512 jne SHORT G_M000_IG06 G_M000_IG04: ;; offset=0x001E mov rcx, gword ptr [rbx+0x28] cmp esi, dword ptr [rcx+0x08] jae SHORT G_M000_IG07 mov rax, gword ptr [rcx+8*rsi+0x10] mov rcx, gword ptr [rax+0x08] call [rax+0x18]System.Action:Invoke():this inc esi movsxd rcx, esi cmp rcx, qword ptr [rbx+0x30] jl SHORT G_M000_IG03 G_M000_IG05: ;; offset=0x003E add rsp, 40 pop rbx pop rsi ret G_M000_IG06: ;; offset=0x0045 mov rcx, rbx mov edx, esi call System.StubHelpers.StubHelpers:MulticastDebuggerTraceHelper(System.Object,int) jmp SHORT G_M000_IG04 G_M000_IG07: ;; offset=0x0051 call CORINFO_HELP_RNGCHKFAIL int3 ; Total bytes of code 87

jkotas

LGTM otherwise. Nice simplification!

src/coreclr/vm/comdelegate.cpp

jkotas · 2024-06-29T20:08:17Z

src/coreclr/vm/comdelegate.cpp

+        pCode->EmitBRTRUE(invokeTraceHelper);
+        pCode->EmitBR(debuggerCheckEnd); // Tune branch prediction to prefer non-debugging path
+
+        pCode->EmitLabel(invokeTraceHelper);
+
        pCode->EmitLoadThis();
        pCode->EmitLDLOC(dwLoopCounterNum);
        pCode->EmitCALL(METHOD__STUBHELPERS__MULTICAST_DEBUGGER_TRACE_HELPER, 2, 0);


This code has atypical loop. It looks like the JIT tried to reorder the basic blocks to turn it a more regular loop. I am not sure whether there is anything to fix in the JIT (the JIT would have to have profile data to do better).

It may be still worth it to move TraceHelper call to be at the end in IL. It provides stronger hint about the desired code layout to the JIT and makes this optimization a bit less fragile.

src/coreclr/vm/comdelegate.cpp

jkotas · 2024-06-30T14:59:13Z

src/coreclr/vm/comdelegate.cpp


        // initialize counter
        pCode->EmitLDC(0);
        pCode->EmitSTLOC(dwLoopCounterNum);

+        // Make the shape of the loop similar to what C# compiler emits


The new shape seems to be missing the invocation of the trace helper at the end. E.g. if the invocation count is 2, the trace helper should be called 3 times. It is only called 2 times if I am reading the code correctly.

Fixed. Result codegen:

; Assembly listing for method System.Action:IL_STUB_MulticastDelegate_Invoke():this (FullOpts) ; Emitting BLENDED_CODE for X64 with AVX - Windows ; FullOpts code ; optimized code ; optimized using Synthesized PGO ; rsp based frame ; partially interruptible ; with Synthesized PGO: fgCalledCount is 100 ; No PGO data G_M000_IG01: ;; offset=0x0000 push rsi push rbx sub rsp, 40 mov rbx, rcx G_M000_IG02: ;; offset=0x0009 xor esi, esi jmp SHORT G_M000_IG04 G_M000_IG03: ;; offset=0x000D mov rcx, gword ptr [rbx+0x28] cmp esi, dword ptr [rcx+0x08] jae SHORT G_M000_IG08 mov rax, gword ptr [rcx+8*rsi+0x10] mov rcx, gword ptr [rax+0x08] call [rax+0x18]System.Action:Invoke():this inc esi G_M000_IG04: ;; offset=0x0024 test dword ptr [(reloc 0x7ff84ea05210)], 512 jne SHORT G_M000_IG06 G_M000_IG05: ;; offset=0x0030 movsxd rcx, esi cmp rcx, qword ptr [rbx+0x30] jl SHORT G_M000_IG03 jmp SHORT G_M000_IG07 G_M000_IG06: ;; offset=0x003B mov rcx, rbx mov edx, esi call System.StubHelpers.StubHelpers:MulticastDebuggerTraceHelper(System.Object,int) jmp SHORT G_M000_IG05 G_M000_IG07: ;; offset=0x0047 add rsp, 40 pop rbx pop rsi ret G_M000_IG08: ;; offset=0x004E call CORINFO_HELP_RNGCHKFAIL int3 ; Total bytes of code 84

JIT reorders the debugging block in front of the ret. Benchmark shows no regression. It's just a bit strange with the unconditional jumps.

huoyaoyuan added 7 commits June 29, 2024 15:40

Enable FEATURE_MULTICASTSTUB_AS_IL on win-x86

5ec5198

Unconditionally enable at use sites

a8225a4

Remove dead code

70c5e19

Delete feature flag

c2bbd66

Skip MulticastDebuggerTraceHelper when no debugger

49bd3b2

Reduce register pressure

e6ca8d5

Tune branch prediction

c50c39c

dotnet-issue-labeler bot added the area-VM-coreclr label Jun 29, 2024

dotnet-policy-service bot added the community-contribution Indicates that the PR has been added by a community member label Jun 29, 2024

jkotas reviewed Jun 29, 2024

View reviewed changes

ifdef format

5964165

jkotas reviewed Jun 29, 2024

View reviewed changes

huoyaoyuan added 4 commits June 30, 2024 14:29

Typo

03e8811

Simplify register saving

0744c51

Normalize loop

a5ec68a

Hot-cold splitting

b492c11

jkotas reviewed Jun 30, 2024

View reviewed changes

Fix trace helper

4c83681

build-analysis bot mentioned this pull request Jun 30, 2024

Test failure: GC\\Scenarios\\FinalizeTimeout\\FinalizeTimeout\\FinalizeTimeout.cmd #103874

Closed

jkotas approved these changes Jul 1, 2024

View reviewed changes

jkotas merged commit 3dd3488 into dotnet:main Jul 1, 2024
86 of 89 checks passed

jkotas mentioned this pull request Jul 1, 2024

[NativeAOT] Optimize multicast delegate thunk #104219

Closed

huoyaoyuan deleted the multicast-stub-x86 branch July 1, 2024 04:39

github-actions bot locked and limited conversation to collaborators Jul 31, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Enable FEATURE_MULTICASTSTUB_AS_IL for Windows x86 #104192

Enable FEATURE_MULTICASTSTUB_AS_IL for Windows x86 #104192

huoyaoyuan commented Jun 29, 2024

huoyaoyuan commented Jun 29, 2024

dotnet-policy-service bot commented Jun 29, 2024

EgorBot commented Jun 29, 2024

EgorBot commented Jun 29, 2024

EgorBot commented Jun 29, 2024

jkotas Jun 29, 2024

huoyaoyuan Jun 29, 2024 •

edited

Loading

jkotas Jun 29, 2024

huoyaoyuan Jun 30, 2024 •

edited

Loading

huoyaoyuan Jun 30, 2024

jkotas left a comment

jkotas Jun 29, 2024

jkotas Jun 30, 2024

huoyaoyuan Jun 30, 2024

Enable FEATURE_MULTICASTSTUB_AS_IL for Windows x86 #104192

Enable FEATURE_MULTICASTSTUB_AS_IL for Windows x86 #104192

Conversation

huoyaoyuan commented Jun 29, 2024

huoyaoyuan commented Jun 29, 2024

dotnet-policy-service bot commented Jun 29, 2024

EgorBot commented Jun 29, 2024

EgorBot commented Jun 29, 2024

EgorBot commented Jun 29, 2024

jkotas Jun 29, 2024

Choose a reason for hiding this comment

huoyaoyuan Jun 29, 2024 • edited Loading

Choose a reason for hiding this comment

jkotas Jun 29, 2024

Choose a reason for hiding this comment

huoyaoyuan Jun 30, 2024 • edited Loading

Choose a reason for hiding this comment

huoyaoyuan Jun 30, 2024

Choose a reason for hiding this comment

jkotas left a comment

Choose a reason for hiding this comment

jkotas Jun 29, 2024

Choose a reason for hiding this comment

jkotas Jun 30, 2024

Choose a reason for hiding this comment

huoyaoyuan Jun 30, 2024

Choose a reason for hiding this comment

huoyaoyuan Jun 29, 2024 •

edited

Loading

huoyaoyuan Jun 30, 2024 •

edited

Loading