(core) Replace compute stack and frames with _current_node and _current_row_id

Summary:
This is an attempt to optimise Engine._use_node. It doesn't seem to actually improve overall performance significantly, but it shouldn't make it worse, and I think it's an improvement to the code.

It turns out that there's no need to track a stack of compute frames any more. The only time we get close to nested evaluation, we set allow_evaluation=False to prevent it actually happening. So there's only one 'frame' during actual evaluation, which means we can get rid of the concept of frames entirely. This allows simplifying the code and letting the computer do less work in general.

Test Plan: this

Reviewers: dsagal

Reviewed By: dsagal

Subscribers: dsagal

Differential Revision: https://phab.getgrist.com/D3310

sandbox/grist/engine.py

@@ -129,18 +129,6 @@ class Engine(object):
         containing several categories of DocActions, including the results of computations.
   """
 
-  class ComputeFrame(object):
-    """
-    Represents the node and ID of the value currently being recomputed. There is a stack of
-    ComputeFrames, because during computation we may access other out-of-date nodes, and need to
-    recompute those first.
-      compute_frame.current_row_id gets set to each record ID as we go through them.
-    """
-    def __init__(self, node):
-      self.node = node
-      self.current_row_id = None
-
-
   def __init__(self):
     # The document data, including logic (formulas), and metadata (tables prefixed with "_grist_").
     self.tables = {}                # Maps table IDs (or names) to Table objects.
@@ -193,8 +181,9 @@ class Engine(object):
     # The lists of actions of different kinds, built up while applying an action.
     self.out_actions = action_obj.ActionGroup()
 
-    # Stack of compute frames.
+    # What's currently being computed
-    self._compute_stack = []
+    self._current_node = None
+    self._current_row_id = None
 
     # Certain recomputations are triggered by a particular doc action. This keep track of it.
     self._triggering_doc_action = None
@@ -478,35 +467,15 @@ class Engine(object):
     for node, dirty_rows in six.iteritems(self.recompute_map):
       log.debug("  Node %s: %s" % (node, dirty_rows))
 
-  @contextlib.contextmanager
-  def open_compute_frame(self, node):
-    """
-    Use as: `with open_compute_frame(node) as frame:`. This automatically maintains the stack of
-    ComputeFrames, pushing and popping reliably.
-    """
-    frame = Engine.ComputeFrame(node)
-    self._compute_stack.append(frame)
-    try:
-      yield frame
-    finally:
-      self._compute_stack.pop()
-
-  def get_current_frame(self):
-    """
-    Returns the compute frame currently being computed, or None if there isn't one.
-    """
-    return self._compute_stack[-1] if self._compute_stack and self._compute_stack[-1].node else None
-
   def _use_node(self, node, relation, row_ids=[]):
     # This is used whenever a formula accesses any part of any record. It's hot code, and
     # it's worth optimizing.
 
-    if self._compute_stack and self._compute_stack[-1].node:
+    if self._current_node:
       # Add an edge to indicate that the node being computed depends on the node passed in.
       # Note that during evaluation, we only *add* dependencies. We *remove* them by clearing them
       # whenever ALL rows for a node are invalidated (on schema changes and reloads).
-      current_node = self._compute_stack[-1].node
+      edge = (self._current_node, node, relation)
-      edge = (current_node, node, relation)
       if edge not in self._recompute_edge_set:
         self.dep_graph.add_edge(*edge)
         self._recompute_edge_set.add(edge)
@@ -694,7 +663,7 @@ class Engine(object):
     col = table.get_column(col_id)
     checkpoint = self._get_undo_checkpoint()
     try:
-      result = self._recompute_one_cell(None, table, col, row_id)
+      result = self._recompute_one_cell(table, col, row_id)
       # If the error is gone for a trigger formula
       if col.has_formula() and not col.is_formula():
         if not isinstance(result, objtypes.RaisedException):
@@ -778,14 +747,14 @@ class Engine(object):
 
     require_rows = sorted(require_rows or [])
 
+    previous_current_node = self._current_node
     # Prevents dependency creation for non-formula nodes. A non-formula column may include a
     # formula to eval for a newly-added record. Those shouldn't create dependencies.
-    formula_node = node if col.is_formula() else None
+    self._current_node = node if col.is_formula() else None
 
     changes = None
     cleaned = []    # this lists row_ids that can be removed from dirty_rows once we are no
                     # longer iterating on it.
-    with self.open_compute_frame(formula_node) as frame:
     try:
       require_count = len(require_rows)
       for i, row_id in enumerate(itertools.chain(require_rows, dirty_rows)):
@@ -815,7 +784,7 @@ class Engine(object):
           # We figure out if we've hit a cycle here.  If so, we just let _recompute_on_cell
           # know, so it can set the cell value appropriately and do some other bookkeeping.
           cycle = required and (node, row_id) in self._locked_cells
-            value = self._recompute_one_cell(frame, table, col, row_id, cycle=cycle, node=node)
+          value = self._recompute_one_cell(table, col, row_id, cycle=cycle, node=node)
         except OrderError as e:
           if not required:
             # We're out of order, but for a cell we were evaluating opportunistically.
@@ -864,6 +833,7 @@ class Engine(object):
         dirty_rows = None
 
     finally:
+      self._current_node = previous_current_node
       for row_id in cleaned:
         # Usually dirty_rows refers to self.recompute_map[node], so this modifies both
         dirty_rows.discard(row_id)
@@ -873,7 +843,7 @@ class Engine(object):
       if not self.recompute_map[node]:
         self.recompute_map.pop(node)
 
-  def _recompute_one_cell(self, frame, table, col, row_id, cycle=False, node=None):
+  def _recompute_one_cell(self, table, col, row_id, cycle=False, node=None):
     """
     Recomputes an one formula cell and returns a value.
     The value can be:
@@ -881,8 +851,7 @@ class Engine(object):
       - exception
       - exception with details if flag include_details is set
     """
-    if frame:
+    self._current_row_id = row_id
-      frame.current_row_id = row_id
 
     # Baffling, but keeping a reference to current generated "usercode" module protects against a
     # seeming garbage-collection bug: if during formula evaluation the module gets regenerated
@@ -1325,9 +1294,9 @@ class Engine(object):
     # We normally recompute formulas before returning to the user; but some formulas are also used
     # internally in-between applying doc actions. We have this workaround to ensure that those are
     # up-to-date after each doc action. See more in comments for _bring_mlookups_up_to_date.
-    # We check _compute_stack to avoid a recursive call (happens when a formula produces an
+    # We check _in_update_loop to avoid a recursive call (happens when a formula produces an
     # action, as for derived/summary tables).
-    if not self._compute_stack:
+    if not self._in_update_loop:
       self._bring_mlookups_up_to_date(doc_action)
 
   def autocomplete(self, txt, table_id, column_id, user):

sandbox/grist/functions/info.py

@@ -577,7 +577,7 @@ def _prepare_record_dict(record, dates_as_iso=False, expand_refs=0):
   table_id = record._table.table_id
   docmodel = record._table._engine.docmodel
   columns = docmodel.get_table_rec(table_id).columns
-  frame = record._table._engine.get_current_frame()
+  current_node = record._table._engine._current_node
 
   result = {'id': int(record)}
   errors = {}
@@ -590,7 +590,7 @@ def _prepare_record_dict(record, dates_as_iso=False, expand_refs=0):
     # Avoid trying to access the cell being evaluated, since cycles get detected even if the
     # CircularRef exception is caught. TODO This is hacky, and imperfect. If another column
     # references a column containing the RECORD(rec) call, CircularRefError will still happen.
-    if frame and frame.node == (table_id, col_id):
+    if current_node == (table_id, col_id):
       continue
 
     try:

sandbox/grist/lookup.py

@@ -125,17 +125,17 @@ class BaseLookupMapColumn(column.BaseColumn):
     the current frame to the returned records. Returns an empty set if no records match.
     """
     key = tuple(_extract(val) for val in key)
-    current_frame = self._engine.get_current_frame()
+    engine = self._engine
-    if current_frame:
+    if engine._current_node:
-      rel = self._get_relation(current_frame.node)
+      rel = self._get_relation(engine._current_node)
-      rel._add_lookup(current_frame.current_row_id, key)
+      rel._add_lookup(engine._current_row_id, key)
     else:
       rel = None
 
     # The _use_node call both brings LookupMapColumn up-to-date, and creates a dependency on it.
     # Relation of None isn't valid, but it happens to be unused when there is no current_frame.
     row_ids = self._row_key_map.lookup_right(key, set())
-    self._engine._use_node(self.node, rel, row_ids)
+    engine._use_node(self.node, rel, row_ids)
     if not row_ids:
       row_ids = self._row_key_map.lookup_right(key, set())