Why are calls to virtual generic methods so much slower than calls to other methods?

[dczuo]

[Benchmark]
public void InvokeM()
{
	M();
}

[Benchmark]
public void InvokeVM()
{
	VM();
}

[Benchmark]
public void InvokeGM()
{
	GM<string>();
}

[Benchmark]
public void InvokeVGM()
{
	VGM<string>();
}

void M()
{
}

protected virtual void VM()
{
}

void GM<T>()
{
}

protected virtual void VGM<T>()
{
}

BenchmarkDotNet=v0.13.1, OS=Windows 10.0.19044.1566 (21H2)
Intel Core i7-3520M CPU 2.90GHz (Ivy Bridge), 1 CPU, 4 logical and 2 physical cores
.NET SDK=6.0.200
[Host] : .NET 6.0.2 (6.0.222.6406), X64 RyuJIT [AttachedDebugger]
DefaultJob : .NET 6.0.2 (6.0.222.6406), X64 RyuJIT

Method	Mean	Error	StdDev	Median	Allocated
InvokeM	0.0096 ns	0.0152 ns	0.0142 ns	0.0000 ns	-
InvokeVM	0.6099 ns	0.0160 ns	0.0150 ns	0.6129 ns	-
InvokeGM	0.0031 ns	0.0074 ns	0.0066 ns	0.0000 ns	-
InvokeVGM	12.5043 ns	0.1471 ns	0.1376 ns	12.5060 ns	-

[EgorBo]

First of all, your InvokeM and InvokeGM do not measure any call overhead - they're empty after inlining.

[EgorBo]

cc @jkotas I wonder if we can improve dictionary lookup performance, e.g. .NET Dictionary uses "Fast Mod" algorithm while native one seems to be does not.

(to avoid idiv with ~40 latency)

Also, it seems we don't hoist/CSE its arguments on jit side:

[EgorBo]

Most of the perf seems to be spent around that div:

[EgorBo]

Moved discussion to #65778

[richlander]

Please keep the discourse on-topic and productive @Xyncgas.