[mono] When creating an array type, don't do full initialization of the element type

[lambdageek]

Consider code like this:

public class Node<T>
{
    public Node<T>[][] Children;
}


Console.WriteLine (typeof(Node<int>));

In this case, when we're JITing ldtoken class Node`1<int32> we first have to parse the type Node`1<int32> and initialize it.
When we're initializing it, we need to resolve the types of all the fields in the class in order to to figure out if its instance size.

To resolve the field types we have to parse the types of all the fields, and we eventually end up parsing the type Node`1<!0>[][] which ends up here:

runtime/src/mono/mono/metadata/metadata.c

Lines 4023 to 4027 in 558345d

	MonoType *etype = mono_metadata_parse_type_checked (m, container, 0, transient, ptr, &ptr, error);
	if (!etype)
	return FALSE;
	type->data.klass = mono_class_from_mono_type_internal (etype);

When we get to line 4027 the second time (recursively), etype is Node`1<!0> as a MonoType*. And we want to set type->data.klass to its corresponding MonoClass*. That ends up calling mono_class_from_mono_type_internal, which does this:

runtime/src/mono/mono/metadata/class.c

Lines 2214 to 2215 in 558345d

	case MONO_TYPE_SZARRAY:
	return mono_class_create_array (type->data.klass, 1);

When we call mono_class_create_array we end up here:

runtime/src/mono/mono/metadata/class-init.c

Lines 1186 to 1187 in 558345d

	if (mono_class_is_ginst (eclass))
	mono_class_init_internal (eclass);

with eclass equal to Node`1<!0>` which we try to initialize and we get a TLE because we're going to try to initialize the same generic type definition Node`1`` twice.

Compare with this other class:

public unsafe class Node2<T>
{
    public Node2<T>[] Children;
}

In this case we only end up calling mono_class_from_mono_type_internal on Node2`1<int32> (not an array). And that branch does not do any initialization - it just returns type->data.klass (for Node2`1<!0> - which is seen as a gtd at this point) or mono_class_create_generic_inst (type->data.generic_class) (for Node2`1<int32> which is seen later when ldtoken inflates the gtd)) - without any extra initialization.

---

It seems the reason we get into trouble is because mono_class_create_array does more work than other MonoClass creation functions - it tries to initialize the MonoClass a little bit (for example by setting MonoClass:has_references) which needs at least a little bit of the element class to be initialized.

But note that the code has this:

runtime/src/mono/mono/metadata/class-init.c

Lines 1186 to 1189 in 558345d

	if (mono_class_is_ginst (eclass))
	mono_class_init_internal (eclass);
	if (!eclass->size_inited)
	mono_class_setup_fields (eclass);

I feel fairly certain we don't need the full class initialization if we're going to immediately initialize the field size information.

Fixes #85821

[ghost]

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

Issue Details

---

Consider code like this:

public class Node<T>
{
    public Node<T>[][] Children;
}


Console.WriteLine (typeof(Node<int>));

In this case, when we're JITing ldtoken class Node`1<int32> we first have to parse the type Node`1<int32> and initialize it.
When we're initializing it, we need to resolve the types of all the fields in the class in order to to figure out if its instance size.

To resolve the field types we have to parse the types of all the fields, and we eventually end up parsing the type Node`1<int32>[][] which ends up here:

runtime/src/mono/mono/metadata/metadata.c

Lines 4023 to 4027 in 558345d

	MonoType *etype = mono_metadata_parse_type_checked (m, container, 0, transient, ptr, &ptr, error);
	if (!etype)
	return FALSE;
	type->data.klass = mono_class_from_mono_type_internal (etype);

When we get to line 4027 the second time (recursively), etype is Node`1<int32> as a MonoType*. And we want to set type->data.klass to its corresponding MonoClass*. That ends up callin mono_class_from_mono_type_internal, which does this:

runtime/src/mono/mono/metadata/class.c

Lines 2214 to 2215 in 558345d

	case MONO_TYPE_SZARRAY:
	return mono_class_create_array (type->data.klass, 1);

When we call mono_class_create_array we end up here:

runtime/src/mono/mono/metadata/class-init.c

Lines 1186 to 1187 in 558345d

	if (mono_class_is_ginst (eclass))
	mono_class_init_internal (eclass);

with eclass equal to Node`1<int32>` which we try to initialize and we get a TLE because we're going to try to initialize the same generic type definition Node`1`` twice.

Compare with this other class:

public unsafe class Node2<T>
{
    public Node2<T>[] Children;
}

In this case we only end up calling mono_class_from_mono_type_internal on Node2`1<int32> (not an array). And that branch does not do any initialization - it just returns type->data.klass (for Node2`1<!0> - which is seen as a gtd at this point) or mono_class_create_generic_inst (type->data.generic_class) (for Node2`1<int32> which is seen later when ldtoken inflates the gtd)) - without any extra initialization.

---

It seems the reason we get into trouble is because mono_class_create_array does more work than other MonoClass creation functions - it tries to initialize the MonoClass a little bit (for example by setting MonoClass:has_references) which needs at least a little bit of the element class to be initialized.

But note that the code has this:

runtime/src/mono/mono/metadata/class-init.c

Lines 1186 to 1189 in 558345d

	if (mono_class_is_ginst (eclass))
	mono_class_init_internal (eclass);
	if (!eclass->size_inited)
	mono_class_setup_fields (eclass);

I feel fairly certain we don't need the full class initialization if we're going to immediately initialize the field size information.

Fixes #85821

Author:	lambdageek
Assignees:	lambdageek
Labels:	area-AssemblyLoader-mono
Milestone:	-

[lambdageek]

@thaystg @vargaz PTAL