(CC: @griesemer @mdempsky @dsnet)

Possibly related to #24721.

I just ran into this: https://play.golang.org/p/rv5GDx0s02w

There seem to be three things going on

1. If you create an anonymous field with no name, it should use the type name (feature request)
2. Whether you give a name, it's not treating the field as anonymous (prints field name and type instead of just field name derived from type)
3. cmd/compile: inconsistent behaviors in judging whether or not two types are identical #24721 as @griesemer said above

(Thanks Sean for triaging the recently mentioned issue.)

The problem seems to be an inconsistency in the way the abi and reflect packages represent structs with embedded members.

Consider this playground example where the reflected field types are equivalent, but the native type has no field name and the reflect type has the same name as its embedded type:

Note the types of the fields are exactly equivalent according to reflect:
anon: {Name:A PkgPath: Type:main.A Tag: Offset:0 Index:[0] Anonymous:true}, {Name:B PkgPath: Type:main.B Tag: Offset:16 Index:[1] Anonymous:true}
dyn:  {Name:A PkgPath: Type:main.A Tag: Offset:0 Index:[0] Anonymous:true}, {Name:B PkgPath: Type:main.B Tag: Offset:16 Index:[1] Anonymous:true}

Though the printed value looks the same, the printed type is different and they do not compare equal:
anon: {A:{AField:hello} B:{BField:world}} :: struct { main.A; main.B }
dyn:  {A:{AField:hello} B:{BField:world}} :: struct { A main.A; B main.B }
equal?: false

When the reflect package calls unsafe_New, for some reason the fact that the field is embedded is lost and the field is given an explicit name, though it retains its embedded status. This asymmetry is problematic because a type like this is unnameable natively, and also because reflect can't create a struct with embedded members.

I tried investigating further but I ran into a wall because I can't tell where unsafe_New is implemented; this name is not associated with a definition anywhere in the go source.

The declaration of unsafe_New in the reflect package says

// implemented in package runtime

//go:noescape
func unsafe_New(*abi.Type) unsafe.Pointer

Searching for unsafe_New in the runtime package turns up

//go:linkname reflect_unsafe_New reflect.unsafe_New
func reflect_unsafe_New(typ *_type) unsafe.Pointer {
	return mallocgc(typ.Size_, typ, true)
}

That is the implementation of reflect.unsafe_New.

That said I think that unsafe_New is a red herring here. unsafe_New takes a type descriptor as an argument. The issue here is that the type descriptors are different. That is, the algorithm used by StructOf to construct the type descriptor does not match the algorithm used by the compiler.

Change https://go.dev/cl/567897 mentions this issue: reflect: omit anonymous field name from StructOf type string

That is, the algorithm used by StructOf to construct the type descriptor does not match the algorithm used by the compiler.

Yes this is what I was trying to point out, though I misattributed the issue to a downstream effect (unsafe_New) instead of walking back farther to the root cause cause (StructOf). I had forgotten that the runtime sometimes implements internal functions for packages so I did not consider looking elsewhere, and I agree that unsafe_New is a red herring. Thank you for explaining anyway, and for reading through the inaccurate terminology to decipher what I meant.

Now I'm curious about the algorithm that the compiler uses to generate type descriptors.

What do you think of the format of this playground example, where it prints out a description of the issue inline with the code that demonstrates it?

The playground example was helpful, thanks. I sent a fix for the problem.