add known forth for 2D vector ElementLinks

src/uproot/interpretation/known_forth/__init__.py

@@ -0,0 +1,21 @@
+# BSD 3-Clause License; see https://github.com/scikit-hep/uproot5/blob/main/LICENSE
+
+"""
+This module defines known forth code for types it is known a priori.
+"""
+from __future__ import annotations
+
+
+def known_forth_of(branch):
+    from uproot.interpretation.known_forth.atlas.element_link import (
+        vector_vector_element_link,
+    )
+
+    if (
+        # len(branch.branches) == 0 # don't understand why this goes nuts
+        # and branch.has_member("fClassName")
+        branch.has_member("fClassName")
+    ):
+        typename = branch.member("fClassName").replace(" ", "")
+        if typename.startswith("vector<vector<ElementLink<DataVector<xAOD::"):
+            return vector_vector_element_link

src/uproot/interpretation/known_forth/atlas/__init__.py

@@ -0,0 +1,5 @@
+# BSD 3-Clause License; see https://github.com/scikit-hep/uproot5/blob/main/LICENSE
+
+"""
+This module defines ATLAS specific known forth code
+"""

src/uproot/interpretation/known_forth/atlas/element_link.py

@@ -0,0 +1,57 @@
+# BSD 3-Clause License; see https://github.com/scikit-hep/uproot5/blob/main/LICENSE
+
+"""
+This module defines known forth code for some ElementLink data types in ATLAS (D)AODs
+"""
+
+# TODO: delay import?
+from __future__ import annotations
+
+from awkward.forms import ListOffsetForm, NumpyForm, RecordForm
+
+vector_vector_element_link = (
+    """
+input stream
+input byteoffsets
+input bytestops
+output node1-offsets int64
+output node2-offsets int64
+output node3-data uint32
+output node4-data uint32
+
+0 node1-offsets <- stack
+0 node2-offsets <- stack
+
+0 do
+    byteoffsets I-> stack
+    stream seek
+    6 stream skip
+    stream !I-> stack
+    dup node1-offsets +<- stack
+    0 do
+        stream !I-> stack
+        dup node2-offsets +<- stack
+        0 do
+            20 stream skip
+            stream !I-> node3-data
+            stream !I-> node4-data
+        loop
+    loop
+loop
+""",
+    ListOffsetForm(
+        "i64",
+        ListOffsetForm(
+            "i64",
+            RecordForm(
+                [
+                    NumpyForm("uint32", form_key="node3"),
+                    NumpyForm("uint32", form_key="node4"),
+                ],
+                ["m_persKey", "m_persIndex"],
+            ),
+            form_key="node2",
+        ),
+        form_key="node1",
+    ),
+)

src/uproot/interpretation/objects.py

@@ -28,6 +28,7 @@
 
 import uproot
 import uproot._awkwardforth
+from uproot.interpretation.known_forth import known_forth_of
 
 
 class AsObjects(uproot.interpretation.Interpretation):
@@ -45,14 +46,20 @@ class AsObjects(uproot.interpretation.Interpretation):
     :ref:`uproot.interpretation.objects.AsObjects.simplify` attempts to
     replace this interpretation with a faster-to-read equivalent, but not all
     data types can be simplified.
+
+    # TODO: known_forth can define forth code and forms for special cases that will be picked up here as well
     """
 
     def __init__(self, model, branch=None):
         self._model = model
         self._branch = branch
-        self._form = None
         self._forth = True
-        self._complete_forth_code = None
+        known_forth = known_forth_of(branch)
+        if known_forth is not None:
+            self._complete_forth_code, self._form = known_forth
+        else:
+            self._complete_forth_code = None
+            self._form = None
         self._forth_lock = threading.Lock()
 
     @property
@@ -122,6 +129,8 @@ def awkward_form(
         tobject_header=False,
         breadcrumbs=(),
     ):
+        if self._form is not None:  # TODO: is this really fine?
+            return self._form
         context = self._make_context(
             context, index_format, header, tobject_header, breadcrumbs
         )