relationalai 0.13.5__py3-none-any.whl → 1.0.0a2__py3-none-any.whl

v0/relationalai/semantics/metamodel/dependency.py

@@ -0,0 +1,862 @@
+"""
+    Support for dependency analysis of metamodel IRs.
+"""
+from __future__ import annotations
+from dataclasses import dataclass, field
+from typing import Iterable, Optional
+from itertools import count
+from more_itertools import peekable
+from v0.relationalai.semantics.metamodel import builtins, ir, helpers, visitor
+from v0.relationalai.semantics.metamodel.util import OrderedSet, ordered_set
+
+
+#--------------------------------------------------
+# Public API
+#--------------------------------------------------
+
+@dataclass
+class DependencyInfo():
+    """
+    Represents the result of performing binding and dependency analysis on a metamodel tree.
+
+    All dicts are keyed by task ids.
+    """
+    # input vars for a task
+    input_bindings: dict[int, OrderedSet[ir.Var]] = field(default_factory=dict)
+    # output vars for a task
+    output_bindings: dict[int, OrderedSet[ir.Var]] = field(default_factory=dict)
+    # clusters of dependency that each task participates on
+    dependency_clusters: dict[int, Cluster] = field(default_factory=dict)
+    # keep track of the parents of each task
+    parent: dict[int, ir.Task] = field(default_factory=dict)
+    # keep track of replacements that were made during a rewrite
+    replacements: dict[int, ir.Task] = field(default_factory=dict)
+
+    def task_inputs(self, node: ir.Task) -> Optional[OrderedSet[ir.Var]]:
+        """ The input variables for this task, if any. """
+        if node.id in self.input_bindings:
+            return self.input_bindings[node.id]
+        return None
+
+    def has_inputs(self, node: ir.Task):
+        return node.id in self.input_bindings
+
+    def task_outputs(self, node: ir.Task) -> Optional[OrderedSet[ir.Var]]:
+        """ The output variables for this task, if any. """
+        if node.id in self.output_bindings:
+            return self.output_bindings[node.id]
+        return None
+
+    def has_outputs(self, node: ir.Task):
+        return node.id in self.output_bindings
+
+    def task_dependencies(self, task: ir.Task) -> OrderedSet[ir.Task]:
+        """
+        All dependencies for this task, if any. This includes tasks in outer contexts,
+        not only siblings.
+        """
+
+        deps = ordered_set()
+        cluster = self.dependency_clusters.get(task.id, None)
+        parent = self.parent.get(task.id)
+
+        if cluster:
+            self._collect_deps(cluster, deps)
+        while parent:
+            cluster = self.dependency_clusters.get(parent.id, None)
+            if cluster:
+                self._collect_deps(cluster, deps)
+                parent = self.parent[cluster.content[0].id]
+            else:
+                parent = None
+        return self._with_replacements(deps)
+
+    def local_dependencies(self, task: ir.Task) -> Optional[OrderedSet[ir.Task]]:
+        """ Similar to task_dependencies but returns only dependencies that are siblings. """
+
+        if task.id in self.dependency_clusters:
+            deps = ordered_set()
+            self._collect_deps(self.dependency_clusters[task.id], deps)
+            return self._with_replacements(deps)
+        return None
+
+    def replaced(self, original: ir.Task, replacement: ir.Task):
+        """
+        Inform that, during some pass, this original task was replaced with this replacement
+        task. This affects the info of .task_dependencies(...) since it will answer with
+        the replacement when the original is in the set of dependencies.
+        """
+        self.replacements[original.id] = replacement
+
+    #
+    # Implementation details
+    #
+    def _collect_deps(self, cluster: Cluster, deps: OrderedSet[ir.Task]):
+            queue = []
+            # start with the cluster dependencies, because cluster represents the task we
+            # care about
+            queue.extend(cluster.dependencies)
+            seen = set()
+            while queue:
+                cluster = queue.pop()
+                if cluster.id in seen:
+                    continue
+                seen.add(cluster.id)
+                deps.update(cluster.content)
+                queue.extend(cluster.dependencies)
+
+    def _with_replacements(self, deps):
+        # Return deps with all tasks that need replacements (because they are in the
+        # replacements dict) replaced.
+        if any([dep.id in self.replacements for dep in deps]):
+            # Only allocate and compute replacements if there's a dep that was replaced
+            info = ordered_set()
+            for dep in deps:
+                info.add(self.replacements.get(dep.id, dep))
+            return info
+        return deps
+
+
+def analyze_bindings(task: ir.Task) -> DependencyInfo:
+    """
+    Perform just the binding analysis, skipping the dependency analysis. This is useful
+    for passes that only require knowing the input/output sets for nodes, but do not need
+    the more expensive dependency information.
+
+    The returned DependencyInfo object will only have input and output bindings filled.
+    """
+    binding = BindingAnalysis()
+    task.accept(binding)
+    return binding.info
+
+def analyze(task: ir.Task) -> DependencyInfo:
+    """
+    Perform binding and dependency analysis on this task.
+
+    The returned DependencyInfo object will have all dictionaries filled.
+    """
+    # first perform binding analysis to get inputs and outputs
+    info = analyze_bindings(task)
+
+    # TODO - if the toplevel task needs inputs, that's a groundness error
+
+    # now perform the dependency analysis
+    dependency = DependencyAnalysis(info)
+    task.accept(dependency)
+    return info
+
+#--------------------------------------------------
+# Dependency Analysis
+#--------------------------------------------------
+
+# id generator for Clusters
+_global_id = peekable(count(0))
+def next_id():
+    return next(_global_id)
+
+class Cluster():
+    def __init__(self, info: DependencyInfo, task: ir.Task):
+        """ Create a cluster starting with only this task. """
+        self.id = next_id()
+        self.info = info
+        # exists and lookups for nullary relations are always required (i.e. everything
+        # should depend on them)
+        self.required = isinstance(task, ir.Exists) or (isinstance(task, ir.Lookup) and not task.args)
+        # this is a binders cluster, which is a candidate to being merged
+        self.mergeable = not self.required and isinstance(task, helpers.BINDERS)
+        # this is a cluster that will only hold an effect
+        self.effectful = isinstance(task, helpers.EFFECTS)
+        # this is a cluster that will only hold a composite
+        self.composite = isinstance(task, helpers.COMPOSITES)
+        # content is either a single task or a set of tasks
+        self.content: OrderedSet[ir.Task] = ordered_set(task)
+        # combined inputs and outputs for all tasks in the cluster
+        self.inputs: OrderedSet[ir.Var] = OrderedSet.from_iterable(info.input_bindings.get(task.id))
+        self.outputs: OrderedSet[ir.Var] = OrderedSet.from_iterable(info.output_bindings.get(task.id))
+        # eventually we will compute dependencies between clusters
+        self.dependencies: OrderedSet[Cluster] = ordered_set()
+
+    def __str__(self) -> str:
+        if isinstance(self.content, ir.Task):
+            return str(self.content.id)
+        else:
+            return ', '.join(str(node.id) for node in self.content)
+
+    def __eq__(self, other):
+        return isinstance(other, Cluster) and other.id == self.id
+
+    def __hash__(self):
+        return hash(self.id)
+
+    def depends_on(self, other: Cluster):
+        """ Assert that this cluster depends on the other cluster. """
+        if self in other.dependencies:
+            # prevent cycles caused by bugs, like the union hoisting pets in the
+            # relationship7 test
+            print("Warning: there is a cycle in the dependency graph. This is likely a bug.")
+            # k = ','.join(str(x.id) for x in self.content)
+            # v = ','.join(str(x.id) for x in other.content)
+            # print(f"{k} --> {v}")
+        else:
+            self.dependencies.add(other)
+
+    def shares_variable(self, other: Cluster):
+        """ Returns True iff this cluster and the other cluster have at least one var in common. """
+        return (
+            any(i in other.inputs or i in other.outputs for i in self.inputs) or
+            any(o in other.inputs or o in other.outputs for o in self.outputs)
+        )
+
+    def try_merge(self, other: Cluster, hoisted_vars: OrderedSet[ir.Var]):
+        """
+        Verify that this cluster and the other cluster can be merged. If so, merge the
+        other cluster into this cluster and return True. Otherwise, return False and the
+        clusters are left unmodified.
+        """
+
+        # 1. only mergeable
+        if not (self.mergeable and other.mergeable):
+            return False
+
+        # 2. share some variable
+        if not self.shares_variable(other):
+            return False
+
+        # 3. all inputs are covered by outputs within the cluster
+        if not all(v in self.outputs or v in other.outputs for v in self.inputs):
+            return False
+        if not all(v in self.outputs or v in other.outputs for v in other.inputs):
+            return False
+
+        # 4. if there are hoisted vars in context, we can only merge clusters that bind the
+        # same hoisted vars
+        if hoisted_vars:
+            for v in hoisted_vars:
+                # if self binds v and other does not bind v, they can't be merged
+                if v in self.outputs or v in self.inputs:
+                    if v not in other.outputs or v not in other.inputs:
+                        return False
+                # the other way around
+                if v in other.outputs or v in other.inputs:
+                    if v not in self.outputs or v not in self.inputs:
+                        return False
+
+        # ok, can merge
+        self.merge(other)
+        return True
+
+    def try_merge_group(self, others: list[Cluster]):
+        """
+        Verify that this cluster and the other clusters can all be merged together. If so,
+        merge the other clusters into this cluster and return True. Otherwise, return False
+        and all clusters are left unmodified.
+        """
+        assert(len(others) > 0)
+
+        # 1. only mergeable
+        if not (self.mergeable and all(o.mergeable for o in others)):
+            return False
+
+        # 2. share some variable
+        if not self.shares_variable(others[0]):
+            return False
+
+        # 3. all inputs are covered by outputs within the newly formed cluster
+        if len(others) == 1:
+            others_outputs = others[0].outputs
+        else:
+            others_outputs = ordered_set()
+            [others_outputs.update(o.outputs) for o in others]
+        if not all(v in self.outputs or v in others_outputs for v in self.inputs):
+            return False
+        for other in others:
+            if not all(v in self.outputs or v in others_outputs for v in other.inputs):
+                return False
+
+        # ok, can merge
+        self.merge(others)
+        return True
+
+    def merge(self, other: Cluster|Iterable[Cluster]):
+        """
+        Merge the other cluster(s) into this one. This assumes that the merge makes sense.
+        In general, we should use try_merge or try_merge_group instead.
+        """
+        # merge the other cluster's content, inputs and outputs
+        if isinstance(other, Cluster):
+            self.content.update(other.content)
+            self.inputs.update(other.inputs)
+            self.outputs.update(other.outputs)
+            # update dependencies, ensuring we remove the other from the self
+            self.dependencies.update(other.dependencies)
+            self.dependencies.remove(other)
+        else:
+            for o in other:
+                self.content.update(o.content)
+                self.inputs.update(o.inputs)
+                self.outputs.update(o.outputs)
+                self.dependencies.update(o.dependencies)
+                self.dependencies.remove(o)
+
+
+class DependencyAnalysis(visitor.Visitor):
+    """
+    A visitor to perform dependency analysis in logicals and store the result in a
+    DependencyInfo object.
+
+    The dependency analysis is performed for Logical nodes in 3 steps:
+
+    1. form clusters of children that are mutually dependent, because they are binders (like
+    lookups and aggregates) that have variables in common.
+
+    2. compute dependencies between the clusters based on input and outputs.
+
+    3. attempt to merge clusters that could form larger clusters.
+
+    Consider the following example, where we are computing dependencies for the numbered nodes:
+
+        Logical
+    |1|    Edges(edges)
+    |2|    i(edges, i)
+    |3|    Logical ⇑[v]
+              Logical ⇑[j=None]
+                j(edges, j)
+              count([edges, j], [i], [v])
+    |4|    i < 10
+    |5|    i < v
+    |6|    Logical ⇑[v_2=None]
+              max([i], [], [v, v_2])
+    |7|    → output[i](v, v_2 as 'v2')
+
+    In step 1., we merge (1,2,4) because they have `edges` and `i` in common, but nothing
+    else. In particular, (5) is not merged as it also depends on v.
+
+    In step 2. we compute dependencies:
+      (1,2,4)
+      (3) -> (1,2,4)
+      (5) -> (1,2,4), (3)
+      (6) -> (1,2,4), (3), (5)
+      (7) -> (1,2,4), (3), (5), (6)
+
+    For this example, step 3. does not change anything.
+
+    This result can be interpreted as, in order to compute task 6, tasks (1,2,4), (3) and (5)
+    must hold. This means that a compiler pass that extracts the logical in task 6 into its
+    own top-level logical must bring these dependencies together.
+
+    To illustrate the need for step 3, consider this example:
+
+        Logical
+    |1|    Edge(edge)
+    |2|    i(edge, i)
+    |3|    Edge(edge_2)
+    |4|    j(edge_2, j)
+    |5|    i = j
+    |6|    Logical ⇑[res=None]
+            .....
+    |7|    → output[edge, j, i](i, res)
+
+    The result of step 2 is the following:
+    (1,2)
+    (3,4)
+    (5) -> (1,2), (3,4)
+    (6) -> (1,2), (3,4), (5)
+    (7) -> (1,2), (3,4), (5), (6)
+
+    Step 3 observes that task (5) is a bridge between (1,2) and (3,4), so that merging those
+    3 clusters together yields a consistent cluster. So the result is:
+
+    (1,2,3,4,5)
+    (6) -> (1,2,3,4,5)
+    (7) -> (1,2,3,4,5), (6)
+
+    """
+    def __init__(self, info: DependencyInfo):
+        self.info = info
+
+
+    def enter(self, node: ir.Node, parent: Optional[ir.Node]=None):
+        # keep track of parents of all nodes
+        if parent and isinstance(parent, ir.Task):
+            self.info.parent[node.id] = parent
+        return super().enter(node, parent)
+
+
+    def visit_logical(self, node: ir.Logical, parent: Optional[ir.Node]):
+        # quick check to see if it's worth it computing clusters at all
+        some_child_has_bindings = False
+        for child in node.body:
+            if child.id in self.info.input_bindings or child.id in self.info.output_bindings:
+                some_child_has_bindings = True
+                break
+
+        if some_child_has_bindings:
+            # print(ir.node_to_string(node, print_ids=True))
+            # compute clusters for the nodes based on inputs/outputs and shared variables
+            clusters = self.compute_clusters(node)
+            # compute the dependencies between those clusters
+            self.compute_dependencies(clusters)
+            # attempt to further merge clusters
+            self.merge_clusters(clusters)
+            # index the clusters by tasks participating in those clusters, and record it
+            self.index(clusters)
+            # self._print_debug_info(node, clusters)
+
+        return super().visit_logical(node, parent)
+
+
+    def compute_clusters(self, task: ir.Logical) -> list[Cluster]:
+        """
+        Cluster the children of the logical, storing together children that are mutually
+        dependent.
+        """
+        # create initial clusters
+        clusters:list[Cluster] = [Cluster(self.info, child) for child in task.body]
+
+        # all hoisted vars of children, used to ensure we don't merge clusters that depend
+        # on a variable hoisted by a composite
+        hoisted_vars = ordered_set()
+        for child in task.body:
+            if isinstance(child, helpers.COMPOSITES):
+                hoisted_vars.update(helpers.hoisted_vars(child.hoisted))
+
+        # iterate clustering until nothing changes
+        merging = True
+        while merging:
+            merging = False
+            cs = list(clusters)
+            while cs:
+                # last c1 merged some c2s, so cs was modified, restart
+                if merging:
+                    break
+                c1 = cs.pop()
+                for c2 in cs:
+                    if c1 is c2:
+                        continue
+                    if c1.try_merge(c2, hoisted_vars):
+                        clusters.remove(c2)
+                        merging = True
+        return clusters
+
+
+    def compute_dependencies(self, clusters: list[Cluster]):
+        """
+        Traverse the clusters finding dependencies between them, based on input and output
+        variables used by tasks within the clusters.
+        """
+        def has_dependency(c1: Cluster, c2: Cluster):
+            # c2 is a required cluster, everything depends no it
+            if c2.required:
+                return True
+            # if c1 has an effect and c2 is mergeable (basically it contains only binders)
+            # then c2 behaves like a filter, so c1 must depend on it, even if it does not
+            # have variables in common (this may bring other dependencies).
+            if c1.effectful and c2.mergeable:
+                return True
+            # if c1 has an effect and c2 is a composite without hoisted variables or with a
+            # hoisted variable that does not have a default (it is a plain var), then c2
+            # behaves like a filter and c1 depends on it.
+            if c1.effectful and c2.composite:
+                task = c2.content.some()
+                assert(isinstance(task, helpers.COMPOSITES))
+                if not task.hoisted:
+                    return True
+                for h in task.hoisted:
+                    if isinstance(h, ir.Var): # no default
+                        return True
+
+            # if c1 is a composite and c2 binds its hoisted vars, c1 can't depend on c2
+            # (dependency is the other way around)
+            if c1.composite and c1.outputs:
+                for v in c1.outputs:
+                    if c2.outputs and v in c2.outputs:
+                        return False
+                    if c2.inputs and v in c2.inputs:
+                        return False
+
+            # c1 does not depend on c2 if one of its output vars is an input to c2
+            # (dependency is the other way around)
+            if (c1.outputs and c2.inputs):
+                # optimization for any([v in c2.inputs for v in c1_outputs])):
+                for v in c1.outputs:
+                    if v in c2.inputs:
+                        return False
+
+            # c1 depends on c2 if one of its input vars is an output of c2
+            if (c1.inputs and c2.outputs):
+                # optimization for any([v in c2.outputs for v in c1_inputs])):
+                for v in c1.inputs:
+                    if v in c2.outputs:
+                        return True
+
+            # c1 is a composite with hoisted variables; it depends on c2 if c2 is a
+            # composite that does not have hoisted vars, hence behaving like a filter.
+            if c1.composite and c2.composite:
+                c1task = c1.content.some()
+                assert(isinstance(c1task, helpers.COMPOSITES))
+                if c1task.hoisted:
+                    c2task = c2.content.some()
+                    assert(isinstance(c2task, helpers.COMPOSITES))
+                    if not c2task.hoisted:
+                        return True
+            return False
+
+        cs = list(clusters)
+        while cs:
+            c1 = cs.pop()
+            for c2 in cs:
+                if c1 is c2:
+                    continue
+
+                if has_dependency(c1, c2):
+                    c1.depends_on(c2)
+                if has_dependency(c2, c1):
+                    c2.depends_on(c1)
+
+
+    def merge_clusters(self, clusters: list[Cluster]):
+        """
+        Traverse clusters trying to merge multiple clusters if they together form a larger
+        cluster.
+        """
+        # iterate clustering until nothing changes
+        merging = True
+        while merging:
+            merging = False
+            cs = list(clusters)
+            while cs:
+                if merging:
+                    break
+                c = cs.pop()
+                if c.dependencies and c.mergeable:
+                    deps = list(c.dependencies)
+                    if c.try_merge_group(deps):
+                        # remove the deps from clusters
+                        for d in deps:
+                            clusters.remove(d)
+                        # rewire other clusters to the new node
+                        # this is not very efficient but should not happen often
+                        for c2 in clusters:
+                            if c2 is not c:
+                                for d in deps:
+                                    if d in c2.dependencies:
+                                        c2.dependencies.add(c)
+                                        c2.dependencies.remove(d)
+                        merging = True
+
+
+    def index(self, clusters: list[Cluster]):
+        """
+        Index clusters by task, and record in the info
+        """
+        for dep_node in clusters:
+            for n in dep_node.content:
+                self.info.dependency_clusters[n.id] = dep_node
+
+
+    def _print_debug_info(self, node, clusters: list[Cluster]):
+        # print(ir.node_to_string(node, print_ids=True))
+        print("dependencies")
+        for dep_node in clusters:
+            k = ','.join(str(x.id) for x in dep_node.content)
+            print(f"({k}) ->")
+            for v in dep_node.dependencies:
+                v = ','.join(str(x.id) for x in v.content)
+                print(f"    ({v})")
+        print()
+        print("clusters")
+        for c in clusters:
+            print(f"{c}")
+            if c.inputs:
+                print(f"    inputs: {','.join(str(v.name) for v in c.inputs)}")
+            if c.outputs:
+                print(f"    outputs: {','.join(str(v.name) for v in c.outputs)}")
+        print()
+
+
+
+class BindingAnalysis(visitor.Visitor):
+    """
+    Visitor to perform binding analysis, i.e. figure out for each task in the tree, which
+    variables it binds as input ad output.
+    """
+    def __init__(self):
+        self.info = DependencyInfo()
+        # a stack of variables grounded by the last logical being visited
+        self._grounded: list[OrderedSet[ir.Var]] = []
+
+
+    def input(self, key: ir.Task, val: Optional[ir.Var|Iterable[ir.Var]]):
+        """ Assert that this task binds this variable(s) as input. """
+        self._register(self.info.input_bindings, key, val)
+
+
+    def output(self, key: ir.Task, val: Optional[ir.Var|Iterable[ir.Var]]):
+        """ Assert that this task binds this variable(s) as output. """
+        self._register(self.info.output_bindings, key, val)
+
+
+    def _register(self, map, key: ir.Task, val: Optional[ir.Var|Iterable[ir.Var]]):
+        """ Register key.id -> val in this map, assuming the map holds ordered sets of vals. """
+        if val is None or (isinstance(val, Iterable) and not val):
+            return
+        if key.id not in map:
+            map[key.id] = ordered_set()
+        if isinstance(val, Iterable):
+            for v in val:
+                map[key.id].add(v)
+        else:
+            map[key.id].add(val)
+
+
+    #
+    # Composite tasks
+    #
+    def visit_logical(self, node: ir.Logical, parent: Optional[ir.Node]):
+        # compute variables grounded by children of this logical
+        grounds = ordered_set()
+        grounded_by_ancestors = None
+        if self._grounded:
+            # grounded variables inherited from ancestors or siblings
+            grounded_by_ancestors = self._grounded[-1]
+            grounds.update(grounded_by_ancestors)
+
+        potentially_grounded = ordered_set()
+        for child in node.body:
+            # leaf constructs that ground variables
+            if isinstance(child, ir.Lookup):
+                    # special case eq because it can be input or output
+                    # TODO: this is similar to what's done below in visit_lookup, modularize
+                    if builtins.is_eq(child.relation):
+                        x, y = child.args[0], child.args[1]
+                        # Compute input/output vars of the equality
+                        if isinstance(x, ir.Var) and not isinstance(y, ir.Var):
+                            # Variable x is potentially grounded by other expressions at
+                            # level in the Logical. If it is, then we should mark it as
+                            # input (which is done later).
+                            potentially_grounded.add((child, x, x))
+                        elif not isinstance(x, ir.Var) and isinstance(y, ir.Var):
+                            potentially_grounded.add((child, y, y))
+                        elif isinstance(x, ir.Var) and isinstance(y, ir.Var):
+                            # mark as potentially grounded, if any is grounded in other atoms then we later ground both
+                            potentially_grounded.add((child, x, y))
+                    else:
+                        # grounds only outputs
+                        for idx, f in enumerate(child.relation.fields):
+                            arg = child.args[idx]
+                            if not f.input and isinstance(arg, ir.Var):
+                                grounds.add(arg)
+            elif isinstance(child, ir.Data):
+                # grounds all vars
+                grounds.update(child.vars)
+            elif isinstance(child, ir.Aggregate):
+                # grounds output args
+                for idx, f in enumerate(child.aggregation.fields):
+                    arg = child.args[idx]
+                    if not f.input and isinstance(arg, ir.Var):
+                        grounds.add(arg)
+            elif isinstance(child, ir.Rank):
+                # grounds the info
+                grounds.add(child.result)
+            elif isinstance(child, ir.Construct):
+                # grounds the output var
+                grounds.add(child.id_var)
+
+        # add child hoisted vars to grounded so that they can be picked up by the children
+        for child in node.body:
+            if isinstance(child, helpers.COMPOSITES):
+                grounds.update(helpers.hoisted_vars(child.hoisted))
+
+        # equalities where both sides are already grounded mean that both sides are input
+        for child, x, y in potentially_grounded:
+            if x in grounds and y in grounds:
+                self.input(child, x)
+                self.input(child, y)
+
+        # deal with potentially grounded vars up to a fixpoint
+        changed = True
+        while changed:
+            changed = False
+            for child, x, y in potentially_grounded:
+                if x in grounds and y not in grounds:
+                    self.input(child, x)
+                    self.output(child, y)
+                    grounds.add(y)
+                    changed = True
+                elif y in grounds and x not in grounds:
+                    self.input(child, y)
+                    self.output(child, x)
+                    grounds.add(x)
+                    changed = True
+
+        # now visit the children
+        self._grounded.append(grounds)
+        super().visit_logical(node, parent)
+        self._grounded.pop()
+
+        hoisted_vars = helpers.hoisted_vars(node.hoisted)
+        if grounded_by_ancestors:
+            # inputs to this logical: grounded by ancestor while being used by a child,
+            # excluding variables hoisted by the logical
+            vars = helpers.collect_vars(node)
+            self.input(node, (grounded_by_ancestors & vars) - hoisted_vars)
+
+        # outputs are vars declared as hoisted
+        self.output(node, hoisted_vars)
+
+
+    def visit_union(self, node: ir.Union, parent: Optional[ir.Node]):
+        # visit children first
+        super().visit_union(node, parent)
+
+        # inputs taken from all children
+        for child in node.tasks:
+            self.input(node, self.info.task_inputs(child))
+        # outputs are vars declared as hoisted
+        self.output(node, helpers.hoisted_vars(node.hoisted))
+
+
+    def visit_match(self, node: ir.Match, parent: Optional[ir.Node]):
+        # visit children first
+        super().visit_match(node, parent)
+
+        # inputs taken from all children
+        for child in node.tasks:
+            self.input(node, self.info.task_inputs(child))
+        # outputs are vars declared as hoisted
+        self.output(node, helpers.hoisted_vars(node.hoisted))
+
+
+    def visit_require(self, node: ir.Require, parent: Optional[ir.Node]):
+        # visit children first
+        super().visit_require(node, parent)
+
+        # inputs taken from the domain and all check tasks
+        self.input(node, self.info.task_inputs(node.domain))
+        for check in node.checks:
+            self.input(node, self.info.task_inputs(check.check))
+
+
+    #
+    # Logical tasks
+    #
+    def visit_not(self, node: ir.Not, parent: Optional[ir.Node]):
+        # visit children first
+        super().visit_not(node, parent)
+
+        # not gets the inputs from its child
+        self.input(node, self.info.task_inputs(node.task))
+
+
+    def visit_exists(self, node: ir.Exists, parent: Optional[ir.Node]):
+        # visit children first
+        super().visit_exists(node, parent)
+
+        # exists variables are local, so they are ignored
+        self.input(node, self.info.task_inputs(node.task))
+
+
+    #
+    # Leaf tasks
+    #
+    def visit_data(self, node: ir.Data, parent: Optional[ir.Node]):
+        # data outputs all its variables
+        for v in helpers.vars(node.vars):
+            self.output(node, v)
+
+        return super().visit_data(node, parent)
+
+
+    def visit_update(self, node: ir.Update, parent: Optional[ir.Node]):
+        # register variables being used as arguments to the update, it's always considered an input
+        for v in helpers.vars(node.args):
+            self.input(node, v)
+        return super().visit_update(node, parent)
+
+
+    def visit_lookup(self, node: ir.Lookup, parent: Optional[ir.Node]):
+        def register(node, field, arg):
+            if isinstance(arg, ir.Var):
+                if field.input:
+                    self.input(node, arg)
+                else:
+                    self.output(node, arg)
+
+        if builtins.is_eq(node.relation):
+            # Most cases are covered already at the parent level if the equality is part of
+            # a Logical. The remaining cases are when the equality is a child of a
+            # non-Logical, or if its variables are not ground elsewhere in the Logical.
+            if self.info.task_inputs(node) or self.info.task_outputs(node):
+                # already covered
+                pass
+            else:
+                x, y = node.args[0], node.args[1]
+                grounds = self._grounded[-1] if self._grounded else ordered_set()
+                if isinstance(x, ir.Var):
+                    if x in grounds:
+                        self.input(node, x)
+                    else:
+                        self.output(node, x)
+                if isinstance(y, ir.Var):
+                    if y in grounds:
+                        self.input(node, y)
+                    else:
+                        self.output(node, y)
+        else:
+            # register variables depending on the input flag of the relation bound to the lookup
+            for idx, f in enumerate(node.relation.fields):
+                arg = node.args[idx]
+                if isinstance(arg, Iterable):
+                    # deal with ListType fields that pack arguments in a tuple
+                    for element in arg:
+                        register(node, f, element)
+                else:
+                    register(node, f, arg)
+        return super().visit_lookup(node, parent)
+
+
+    def visit_output(self, node: ir.Output, parent: Optional[ir.Node]):
+        # register variables being output, they always considered an input to the task
+        for v in helpers.output_vars(node.aliases):
+            self.input(node, v)
+        # also register keys as inputs
+        self.input(node, node.keys)
+        return super().visit_output(node, parent)
+
+
+    def visit_construct(self, node: ir.Construct, parent: Optional[ir.Node]):
+        # values are inputs, id_var is an output
+        for v in helpers.vars(node.values):
+            self.input(node, v)
+        self.output(node, node.id_var)
+
+
+    def visit_aggregate(self, node: ir.Aggregate, parent: Optional[ir.Node]):
+        # register projection and group as inputs
+        for v in node.projection:
+            self.input(node, v)
+        for v in node.group:
+            self.input(node, v)
+
+        # register variables depending on the input flag of the aggregation relation
+        for idx, f in enumerate(node.aggregation.fields):
+            arg = node.args[idx]
+            if isinstance(arg, ir.Var):
+                if f.input:
+                    self.input(node, arg)
+                else:
+                    self.output(node, arg)
+        return super().visit_aggregate(node, parent)
+
+
+    def visit_rank(self, node: ir.Rank, parent: Optional[ir.Node]):
+        # register projection and group as inputs
+        for v in node.projection:
+            self.input(node, v)
+        for v in node.group:
+            self.input(node, v)
+        for v in node.args:
+            self.input(node, v)
+
+        self.output(node, node.result)
+        return super().visit_rank(node, parent)