JIT: Improve wrapper method inliner heuristic

[hez2010]

Extracted from #113256, with additional improvements for boxed structs return.

• Extend support for newarr, newobj and initobj
• Recognize {{ldloc.*} + box} + ret

@EgorBo

[Copilot]

Copilot reviewed 1 out of 1 changed files in this pull request and generated no comments.

[dotnet-policy-service]

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

[hez2010]

@MihuBot

[hez2010]

Diffs from MihuBot look promising to me.

[hez2010]

@MichalStrehovsky NativeAOT failures seem like an ilc preexisting bug?

[MichalStrehovsky]

@MichalStrehovsky NativeAOT failures seem like an ilc preexisting bug?

The scanner needs to precompute everything RyuJIT is going to look at. Looks like in this case RyuJIT decided to inline past a generic recursion cutoff.

Maybe a fix would be to add a call to NodeFactory.TypeSystemContext.DetectGenericCycles(caller, callee) some place like here:

runtime/src/coreclr/tools/aot/ILCompiler.Compiler/Compiler/Compilation.cs

Lines 101 to 104 in c8f32e5

	public bool CanInline(MethodDesc caller, MethodDesc callee)
	{
	return _inliningPolicy.CanInline(caller, callee);
	}

But I'm afraid RyuJIT will ask canInline with something useless like Foo<_Canon> inlining Foo<_Canon> instead of a useful thing like Foo<Gen<Gen<Gen<X>>>. If so, I don't immediately know what the fix could be, besides throwing away the managed scanner and replacing it with RyuJIT.

[EgorBo]

@MihuBot

[EgorBo]

Given it's called only here, it only supports

call
ldloc
box
ret

?

is this some specific pattern we need to support?

[hez2010]

nope. it's just that doesMethodLookLikeWrapper can return true for a single ret so that it can be reused here.

[hez2010]

Closing for now as I don't have enough time to investigate the regression currently.