relationalai 1.0.0a1__py3-none-any.whl → 1.0.0a2__py3-none-any.whl

relationalai/semantics/metamodel/typer.py

@@ -81,17 +81,21 @@ class Typer:
 # Propagation Network
 #--------------------------------------------------
 
-# The core idea of the typer is to build a propagation network where nodes
-# are vars, fields, or overloaded lookups/updates/aggregates. The intuition
-# is that _all_ types in the IR ultimately flow from relation fields, so if
-# we figure those out we just need to propagate their types to unknown vars, which
-# may then flow into other fields and so on.
-
-# This means the network only needs to contain nodes that either directly flow into
-# an abstract node or are themselves abstract. We need to track overloads because
-# their arguments effectively act like abstract vars until we've resolved the final types.
-
-Node = Union[mm.Var, mm.Field, mm.Literal, mm.Lookup, mm.Update, mm.Aggregate]
+# The core idea of the typer is to build a propagation network that represents data
+# dependencies between nodes. Nodes can be variables, literals, fields and tasks that refer
+# to relations (lookups, aggregates, updates). An edge from node A to node B means that
+# in order to resolve references and types for node B we need to know the type of node A.
+#
+# After building the network, we start with roots that have known types (literals, fields
+# loaded from previous analysis and vars with concrete types) and propagate their types via
+# the edges to other nodes. It is possible that the network contains cycles, so we iterate
+# on the work list until fixpoint is reached, i.e. when we cannot propagate any new type.
+#
+# During the propagation, we are only gathering information about the types and references
+# of nodes. Once we reach a fixpoint, we do a final pass to rewrite the model with the new
+# type information, which may include adding casts to convert between types where needed.
+#
+Node = Union[mm.Var, mm.Field, mm.Literal, mm.Lookup, mm.Aggregate, mm.Update]
 
 class PropagationNetwork():
     def __init__(self, model: mm.Model):
@@ -104,22 +108,12 @@ class PropagationNetwork():
         # map from unresolved placeholder relations to their potential target replacements
         self.potential_targets: dict[mm.Relation, list[mm.Relation]] = {}
 
-        # track the set of nodes that represent entry points into the network
-        self.roots = OrderedSet()
-        # we separately want to track nodes that were loaded from a previous run
-        # so that even if we have edges to them, we _still_ consider them roots
-        # and properly propagate types from them at the beginning
+        # nodes loaded from previous analysis, which we use as roots to start propagation
         self.loaded_roots = set()
 
         # edges in the propagation network, from one node to potentially many
         self.edges:dict[Node, OrderedSet[Node]] = defaultdict(lambda: OrderedSet())
         self.back_edges:dict[Node, OrderedSet[Node]] = defaultdict(lambda: OrderedSet())
-        # all nodes that are the target of an edge (to find roots)
-        self.has_incoming = set()
-
-        # type requirements: for a var with abstract declared type, the set of fields that
-        # it must match the type of because it flows into them
-        self.type_requirements:dict[mm.Var, OrderedSet[mm.Field]] = defaultdict(lambda: OrderedSet())
 
         # all errors collected during inference
         self.errors:list[TyperError] = []
@@ -127,22 +121,22 @@ class PropagationNetwork():
         # overloads resolved for a lookup/update/aggregate, by node id. This is only for
         # relations that declare overloads
         self.resolved_overload:dict[int, mm.Overload] = {}
+
         # placeholders resolved for a lookup, by node id. This is only for relations that
         # are placeholders (i.e. only Any fields) and will be replaced by references to
         # these concrete relations. E.g. a query for "name(Any, Any)" may be replaced by
         # the union of "name(Dog, String)" and name(Cat, String)".
         self.resolved_placeholder:dict[int, list[mm.Relation]] = {}
+
         # for a given lookup/update/aggregate that involves numbers, the specific number
         # type resolved for it.
         self.resolved_number:dict[int, mm.NumberType] = {}
-        # keep track of nodes already resolved to avoid re-resolving
-        self.resolved_nodes:set[int] = set()
 
     #--------------------------------------------------
     # Error reporting
     #--------------------------------------------------
 
-    def type_mismatch(self, node: Node, expected: mm.Type, actual: mm.Type):
+    def type_mismatch(self, node: Node|mm.Update, expected: mm.Type, actual: mm.Type):
         self.errors.append(TypeMismatch(node, expected, actual))
 
     def invalid_type(self, node: Node, type: mm.Type):
@@ -153,6 +147,8 @@ class PropagationNetwork():
         self.errors.append(UnresolvedOverload(node, [self.resolve(a) for a in node.args]))
 
     def unresolved_type(self, node: Node):
+        # TODO - this is not being used yet, we need a pass at the end to check for any node
+        # that we could not resolve
         self.errors.append(UnresolvedType(node))
 
     def has_errors(self, node: Node) -> bool:
@@ -166,15 +162,8 @@ class PropagationNetwork():
     #--------------------------------------------------
 
     def add_edge(self, source: Node, target: Node):
-        # manage roots
-        if target in self.roots and target not in self.loaded_roots:
-            self.roots.remove(target)
-        if source not in self.has_incoming:
-            self.roots.add(source)
-        # register edge
         self.edges[source].add(target)
         self.back_edges[target].add(source)
-        self.has_incoming.add(target)
 
     def add_resolved_type(self, node: Node, type: mm.Type):
         """ Register that this node was resolved to have this type. """
@@ -183,13 +172,8 @@ class PropagationNetwork():
         else:
             self.resolved_types[node] = type
 
-    def add_type_requirement(self, source: mm.Var, field: mm.Field):
-        """ Register that this var, which has an abstract declared type, must match the type
-        of this field as it flows into it.. """
-        self.type_requirements[source].add(field)
-
     #--------------------------------------------------
-    # Load previous types
+    # Load types from a previous analysis
     #--------------------------------------------------
 
     def load_types(self, type_dict: dict[Node, mm.Type]):
@@ -197,22 +181,39 @@ class PropagationNetwork():
             if isinstance(node, (mm.Field)):
                 self.add_resolved_type(node, type)
                 self.loaded_roots.add(node)
-                self.roots.add(node)
 
     #--------------------------------------------------
     # Resolve Values
     #--------------------------------------------------
 
-    def resolve(self, value: Node|mm.Value) -> mm.Type:
-        if isinstance(value, (mm.Var, mm.Field, mm.Literal)):
+    def resolve(self, value: mm.Value) -> mm.Type:
+        if isinstance(value, (mm.Var, mm.Literal)):
             return self.resolved_types.get(value) or to_type(value)
-        assert not isinstance(value, (mm.Lookup, mm.Update, mm.Aggregate)), "Should never try to resolve a task"
+        if isinstance(value, mm.Field):
+            return self.resolved_types.get(value) or value.type
+        assert not isinstance(value, (mm.Task)), "Should never try to resolve a task"
         return to_type(value)
 
     #--------------------------------------------------
     # Resolve References
     #--------------------------------------------------
 
+    def all_dependencies_resolved(self, op:mm.Lookup|mm.Aggregate):
+        """ True iff all dependencies required to resolve this reference are met. """
+        rel = get_relation(op)
+        # if this is a placeholder, eq or cast, we assume all possible args were resolved
+        if bt.is_placeholder(rel) or rel == b.core.eq or rel == b.core.cast:
+            return True
+
+        # else, find whether all back-edges were resolved
+        for node in self.back_edges[op]:
+            # cannot resolve if a required var, literal or input field is still abstract
+            if isinstance(node, (mm.Var, mm.Literal)) or (isinstance(node, mm.Field) and node.input):
+                node_type = self.resolve(node)
+                if bt.is_abstract(node_type):
+                    return False
+        return True
+
     def resolve_reference(self, op: mm.Lookup|mm.Aggregate) -> Optional[mm.Overload|list[mm.Relation]]:
         # check if all dependencies required to resolve this reference are met
         if not self.all_dependencies_resolved(op):
@@ -228,7 +229,6 @@ class PropagationNetwork():
                 if all(type_matches(arg, self.resolve(field))
                         for arg, field in zip(resolved_args, fields)):
                     matches.append(target)
-
             return matches
 
         elif relation.overloads:
@@ -243,24 +243,11 @@ class PropagationNetwork():
                     self.resolved_overload[op.id] = overload
                     return overload
             return []  # no matches found
+
         else:
             # this is a relation with type vars or numbers that needs to be specialized
             return [relation]
 
-
-    def all_dependencies_resolved(self, op:mm.Lookup|mm.Aggregate):
-        # if this is a placeholder, we need assume all possible args were resolved
-        if bt.is_placeholder(get_relation(op)):
-            return True
-        # else, find whether all back-edges were resolved
-        for node in self.back_edges[op]:
-            if isinstance(node, (mm.Var, mm.Field, mm.Literal)):
-                node_type = self.resolve(node)
-                if bt.is_abstract(node_type):
-                    return False
-        return True
-
-
     #--------------------------------------------------
     # Propagation
     #--------------------------------------------------
@@ -269,73 +256,112 @@ class PropagationNetwork():
         edges = self.edges
         work_list = []
 
-        # go through all the roots and find any that are not abstract, they'll be the first
-        # nodes to push types through the network
-        unhandled_roots = OrderedSet()
-        for node in self.roots:
-            if not isinstance(node, (mm.Var, mm.Field, mm.Literal)):
-                continue
-            node_type = self.resolve(node)
-            if not bt.is_abstract(node_type):
-                work_list.append(node)
-            else:
-                unhandled_roots.add(node)
+        # start with the loaded roots + all literals + sources of edges without back edges
+        work_list.extend(self.loaded_roots)
+        for source in self.edges.keys():
+            if isinstance(source, (mm.Literal)) or not source in self.back_edges:
+                work_list.append(source)
 
-        # push known type nodes through the edges
-        while work_list:
-            source = work_list.pop(0)
-            self.resolved_nodes.add(source.id)
-            if source in unhandled_roots:
-                unhandled_roots.remove(source)
-            source_type = self.resolve(source)
-            # check to see if the source has ended up with a set of types that
-            # aren't valid, e.g. a union of primitives
-            if invalid_type(source_type):
-                self.invalid_type(source, source_type)
-
-            # propagate our type to each outgoing edge
-            for out in edges.get(source, []):
-                # if this is an overload then we need to try and resolve it
-                if isinstance(out, (mm.Lookup, mm.Aggregate)):
-                    if not out.id in self.resolved_nodes:
-                        found = self.resolve_reference(out)
-                        if found is not None:
-                            self.resolved_nodes.add(out.id)
-                            self.propagate_reference(out, found)
-                            for arg in out.args:
-                                if arg not in work_list:
-                                    work_list.append(arg)
-                # otherwise, we just add to the outgoing node's type and if it
-                # changes we add it to the work list
-                elif start := self.resolve(out):
-                    self.add_resolved_type(out, source_type)
-                    if out not in work_list and (start != self.resolve(out) or not out.id in self.resolved_nodes):
-                        work_list.append(out)
-
-        for source in unhandled_roots:
-            self.unresolved_type(source)
-
-        # now that we've pushed all the types through the network, we need to validate
-        # that all type requirements of those nodes are met
-        for node, fields in self.type_requirements.items():
-            node_type = self.resolve(node)
-            for field in fields:
-                field_type = self.resolve(field)
-                if not type_matches(node_type, field_type) and not conversion_allowed(node_type, field_type):
-                    self.type_mismatch(node, field_type, node_type)
+        # limit the number of iterations to avoid infinite loops
+        i = 0
+        max_iterations = 100 * len(self.edges)
 
+        # propagate types until we reach a fixed point
+        while work_list:
+            i += 1
+            if i > max_iterations:
+                err("Infinite Loop", "Infinite loop detected in the typer. Please, report this as a bug.")
+                break
 
-    def propagate_reference(self, task:mm.Lookup|mm.Aggregate, references:mm.Overload|list[mm.Relation]):
+            node = work_list.pop(0)
+            next = None
+            if isinstance(node, mm.Field):
+                # this is a field loaded from a previous analysis, so all we need to do is
+                # propagate its type to its output edges
+                next = edges.get(node, [])
+
+            elif isinstance(node, (mm.Lookup, mm.Aggregate)):
+                # this is a lookup/aggregate that may be overloaded or a placeholder; try
+                # to resolve its reference, i.e. determine which specific relation or
+                # overload it refers to
+                found = self.resolve_reference(node)
+                if found is not None:
+                    # if found is None it means that we need more info to resolve the
+                    # reference; otherwise it was possible to resolve it (even if to no matches)
+
+                    # keep the resolved args before propagation to see if they will change
+                    resolved_args = [self.resolve(arg) for arg in node.args]
+                    # propagate the reference resolution
+                    self.propagate_reference(node, resolved_args, found)
+                    if found:
+                        # the next nodes to process are all the outgoing edges plus any arg that
+                        # changed during propagation
+                        next = OrderedSet()
+                        next.update(edges.get(node, []))
+                        next.update([arg for idx, arg in enumerate(node.args) if isinstance(arg, mm.Var) and not (self.resolve(arg) == resolved_args[idx])])
+                    else:
+                        # the reference is unresolved, so remove from the worklist the args
+                        # because they depend on this being resolved
+                        for arg in node.args:
+                            if arg in work_list:
+                                work_list.remove(arg)
+            else:
+                assert isinstance(node, (mm.Var, mm.Literal))
+                resolved = self.resolve(node)
+                # if we ended up with an invalid type, report it on a next edge (which is
+                # where the var or literal is going to be used)
+                if invalid_type(resolved):
+                    self.invalid_type(edges.get(node, [node])[0], resolved)
+
+                # if the var is still abstract, add it to the work list to try again later
+                if isinstance(node, mm.Var) and bt.is_abstract(resolved):
+                    # but if everything that is left are vars, we cannot make progress
+                    if not all(isinstance(x, mm.Var) for x in work_list):
+                        next = [node]
+                else:
+                    # otherwise, check all outgoing edges
+                    next = []
+                    for out in edges.get(node, []):
+                        if isinstance(out, mm.Update):
+                            # this is an update to a field. We have to check that the type is
+                            # valid for the field and if the field is abstract we can refine the
+                            # type. But if the field is concrete we have to check that the type
+                            # matche or can be converted.
+                            is_population_lookup = isinstance(out.relation, mm.TypeNode)
+                            for field in get_update_fields(node, out):
+                                if not type_matches(resolved, field.type, accept_expected_super_types=is_population_lookup) and not conversion_allowed(resolved, field.type):
+                                    self.type_mismatch(out, field.type, resolved)
+                                elif bt.is_abstract(field.type):
+                                    # if the field type changed, propagate further
+                                    if resolved != self.resolve(field):
+                                        next.append(field)
+                                    self.add_resolved_type(field, resolved)
+                        else:
+                            # this is an arg flowing into a task, so resolve the task next
+                            next.append(out)
+
+            # add the new nodes to the work list
+            if next is not None:
+                for n in next:
+                    if n not in work_list:
+                        work_list.append(n)
+
+
+    def propagate_reference(self, task:mm.Lookup|mm.Aggregate, resolved_args:list[mm.Type], references:mm.Overload|list[mm.Relation]):
         # TODO: distinguish between overloads and placeholders better when raising errors
         if not references:
             return self.unresolved_overload(task)
 
-        resolved_args = [self.resolve(arg) for arg in task.args]
-
-        # we need to determine the final types of our args by taking all the references
-        # and adding the type of their fields back to the args.
         relation = get_relation(task)
 
+        if relation == b.core.cast:
+            # cast is special cased: we just propagate the type to the target
+            cast_type, target = task.args[0], task.args[2]
+            assert isinstance(cast_type, mm.Type)
+            assert isinstance(target, mm.Var)
+            self.add_resolved_type(target, cast_type)
+            return
+
         if bt.is_placeholder(relation):
             assert(references and isinstance(references, list))
             # we've resolved the placeholder, so store that
@@ -357,22 +383,25 @@ class PropagationNetwork():
                 # number specialization, so use that relation's field types
                 types = list([self.resolve(f) for f in relation.fields])
 
-            # if our overload preserves types, we check to see if there's a preserved
-            # output type given the inputs and if so, shadow the field's type with the
-            # preserved type
+            # if our overload preserves types, we check to see if there's a preserved output
+            # type given the inputs and if so, shadow the field type with the preserved type.
+            # this is only attempted if all input types match the field types, i.e. no
+            # conversions are needed
             resolved_fields = types
-            if bt.is_function(relation) and len(set(resolved_fields)) == 1:
+            if bt.is_function(relation) and len(set(resolved_fields)) == 1 and not relation in self.NON_TYPE_PRESERVERS and\
+                all(type_matches(arg_type, field_type) for arg_type, field_type, field in zip(resolved_args, types, relation.fields) if field.input):
+
                 input_types = set([arg_type for field, arg_type
                                     in zip(relation.fields, resolved_args) if field.input])
                 if out_type := self.try_preserve_type(input_types):
                     resolved_fields = [field_type if field.input else out_type
                                         for field, field_type in zip(relation.fields, types)]
 
-            # TODO - we also need to make sure the type vars are constently resolved here
-            # i.e. if types contain typevars, check that the args that are bound to those
-            # typevars are consistent
-            if b.core.Number in types or (b.core.TypeVar in types and any(bt.is_number(t) for t in resolved_args)):
-                # this overload contains generic numbers or typevars bound to numbers, so
+
+            # eq is special cased because we don't want to specialize a number for it, as it
+            # can be just comparing numbers of different types.
+            if relation != b.core.eq and (b.core.Number in types or (b.core.TypeVar in types and any(bt.is_number(t) for t in resolved_args))):
+                # this relation contains generic numbers or typevars bound to numbers, so
                 # find which specific type of number to use given the arguments being passed
                 number, resolved_fields = self.specialize_number(relation, resolved_fields, resolved_args)
                 self.resolved_number[task.id] = number
@@ -380,16 +409,62 @@ class PropagationNetwork():
                 for field, field_type, arg in zip(relation.fields, resolved_fields, task.args):
                     if not field.input and isinstance(arg, mm.Var):
                         self.add_resolved_type(arg, field_type)
+
+            elif b.core.TypeVar in types:
+                # this relation contains type vars, so we have to make sure that all args
+                # bound to the same type var are consistent
+
+                # find which arg is bound to the type var and check that they are all consistent
+                typevar_type = None
+                for arg_type, field_type in zip(resolved_args, types):
+                    if field_type == b.core.TypeVar:
+                        if typevar_type is None:
+                            typevar_type = arg_type
+                        else:
+                            typevar_type = merge_types(typevar_type, arg_type)
+                assert typevar_type is not None
+
+                # compute the final arg types by replacing type vars with the typevar_type
+                computed_arg_types = []
+                for arg_type, field_type in zip(resolved_args, types):
+                    # check that the arg type matches the type var type
+                    if not type_matches(typevar_type, arg_type) and not conversion_allowed(arg_type, typevar_type):
+                        self.type_mismatch(task, typevar_type, arg_type)
+                    if field_type == b.core.TypeVar:
+                        computed_arg_types.append(typevar_type)
+                    else:
+                        computed_arg_types.append(arg_type)
+
+                # if no mismatches were found, propagate the computed arg types back to the args
+                if len(computed_arg_types) == len(task.args):
+                    for computed_type, arg, arg_type in zip(computed_arg_types, task.args, resolved_args):
+                        # TODO: we could allow for non-numeric/string literals because this means
+                        # the literal is being used as a value type, but the backend emitters
+                        # would have to deal with that.
+                        if isinstance(arg, mm.Var) or (isinstance(arg, mm.Literal) and (bt.is_numeric(computed_type) or computed_type == b.core.String)):
+                            self.add_resolved_type(arg, computed_type)
+
             else:
-                for field_type, arg, arg_type in zip(resolved_fields, task.args, resolved_args):
-                    if bt.is_abstract(arg_type) and isinstance(arg, mm.Var):
+                # no typevar or number specialization shenanigans, just propagate field types to args
+                for field, field_type, arg, arg_type in zip(relation.fields, resolved_fields, task.args, resolved_args):
+                    # add a resolved type if we learned more about the arg's type
+                    if isinstance(arg, mm.Var) and (bt.is_abstract(arg_type) or bt.extends(field_type, arg_type)):
                         self.add_resolved_type(arg, field_type)
+                    elif not type_matches(arg_type, field_type) and not conversion_allowed(arg_type, field_type):
+                        self.type_mismatch(task, field_type, arg_type)
+
 
+    # we try to preserve types for relations that are functions (i.e. potentially multiple
+    # input but a single output) and where all types are the same. However, there are some
+    # exceptions to this rule, e.g. range() always returns int regardless of whether the
+    # input type is an int or a value type that extends int.
+    NON_TYPE_PRESERVERS = [
+        b.common.range
+    ]
 
     def try_preserve_type(self, types:set[mm.Type]) -> Optional[mm.Type]:
-        # we keep the input type as the output type if either all inputs
-        # are the exact same type or there's one nominal and its base primitive
-        # type, e.g. USD + Decimal
+        # we keep the input type as the output type if either all inputs are the exact same
+        # type or there's one nominal and its base primitive type, e.g. USD + Decimal
         if len(types) == 1:
             return next(iter(types))
         if len(types) == 2:
@@ -411,13 +486,15 @@ class PropagationNetwork():
 
     def specialize_number(self, op, field_types:list[mm.Type], arg_types:list[mm.Type]) -> Tuple[mm.NumberType, list[mm.Type]]:
         """
-        Find the number type to use for an overload that has Number in its field_types,
-        and which is being referred to with these arg_types.
+        Find the number type to use for an overload that has Number in its field_types, and
+        which is being referred to with these arg_types.
 
         Return a tuple where the first element is the specialized number type, and the second
         element is a new list that contains the same types as field_types but with
         Number replaced by this specialized number.
         """
+        # special case a few operators according to Snowflake's rules in
+        # https://docs.snowflake.com/en/sql-reference/operators-arithmetic#scale-and-precision-in-arithmetic-operations
         if op == b.core.div:
             # see https://docs.snowflake.com/en/sql-reference/operators-arithmetic#division
             numerator, denominator = get_number_type(arg_types[0]), get_number_type(arg_types[1])
@@ -436,29 +513,21 @@ class PropagationNetwork():
             # TODO!! - implement proper avg specialization
             pass
 
+        # fall back to the the current specialization policy, which is to select the number
+        # with largest scale and, if there multiple with the largest scale, the one with the
+        # largest precision. This is safe because when converting a number to the
+        # specialized number, we never truncate fractional digits (because we selected the
+        # largest scale) and, if the non-fractional digits are too large to fit the
+        # specialized number, we will have a runtime overflow, which should alert the user
+        # of the problem.
         number = None
         for arg_type in arg_types:
             x = bt.get_number_supertype(arg_type)
             if isinstance(x, mm.NumberType):
-                # the current specialization policy is to select the number with largest
-                # scale and, if there multiple with the largest scale, the one with the
-                # largest precision. This is safe because when converting a number to the
-                # specialized number, we never truncate fractional digits (because we
-                # selected the largest scale) and, if the non-fractional digits are too
-                # large to fit the specialized number, we will have a runtime overflow,
-                # which should alert the user of the problem.
-                #
-                # In the future we can implement more complex policies. For example,
-                # snowflake has well documented behavior for how the output of operations
-                # behave in face of different number types, and we may use that:
-                # https://docs.snowflake.com/en/sql-reference/operators-arithmetic#scale-and-precision-in-arithmetic-operations
                 if number is None or x.scale > number.scale or (x.scale == number.scale and x.precision > number.precision):
                     number = x
-        if number is None:
-            number = b.core.DefaultNumber
-        # assert(isinstance(number, mm.NumberType))
-        return number, [number if bt.is_number(field_type) else field_type
-                for field_type in field_types]
+        assert(number is not None)
+        return number, [number if bt.is_number(t) else t for t in field_types]
 
 
     #--------------------------------------------------
@@ -467,7 +536,6 @@ class PropagationNetwork():
 
     # draw the network as a mermaid graph for the debugger
     def to_mermaid(self, max_edges=500) -> str:
-
         # add links for edges while collecting nodes
         nodes = OrderedSet()
         link_strs = []
@@ -481,22 +549,12 @@ class PropagationNetwork():
                 link_strs.append(f"n{src.id} --> n{dst.id}")
             if len(link_strs) > max_edges:
                 break
-        # type requirements
-        for src, dsts in self.type_requirements.items():
-            nodes.add(src)
-            for dst in dsts:
-                if len(link_strs) > max_edges:
-                    break
-                nodes.add(dst)
-                link_strs.append(f"n{src.id} -.-> n{dst.id}")
-            if len(link_strs) > max_edges:
-                break
 
         def type_span(t:mm.Type) -> str:
             type_str = t.name if isinstance(t, mm.ScalarType) else str(t)
             return f"<span style='color:cyan;'>{type_str.strip()}</span>"
 
-        def reference_span(rel:mm.Relation, arg_types:list[mm.Type], root:str) -> str:
+        def reference_span(rel:mm.Relation, arg_types:list[mm.Type]) -> str:
             args = []
             for field, arg_type in zip(rel.fields, arg_types):
                 field_type = self.resolve(field)
@@ -506,22 +564,21 @@ class PropagationNetwork():
                     args.append(type_span(field_type))
                 else:
                     args.append(type_span(arg_type))
-            return f'{rel.name}{root}({", ".join(args)})'
+            return f'{rel.name}({", ".join(args)})'
 
         resolved = self.resolved_types
         node_strs = []
         for node in nodes:
             klass = ""
-            root = "(*)" if node in self.roots else ""
             if isinstance(node, mm.Var):
                 ir_type = resolved.get(node) or self.resolve(node)
                 type_str = type_span(ir_type)
-                label = f'(["{node.name}{root}:{type_str}"])'
+                label = f'(["{node.name}:{type_str}"])'
             elif isinstance(node, mm.Literal):
                 ir_type = resolved.get(node) or self.resolve(node)
                 type_str = type_span(ir_type)
                 klass = ":::literal"
-                label = f'[/"{node.value}{root}: {type_str}"\\]'
+                label = f'(["{node.value}: {type_str}"])'
             elif isinstance(node, mm.Field):
                 ir_type = resolved.get(node) or self.resolve(node)
                 type_str = type_span(ir_type)
@@ -529,19 +586,22 @@ class PropagationNetwork():
                 rel = node._relation
                 if rel is not None:
                     rel = str(node._relation)
-                    label = f'{{{{"{node.name}{root}:{type_str}\nfrom {rel}"}}}}'
+                    label = f'[/"{node.name}:{type_str}\nfrom {rel}"\\]'
                 else:
-                    label = f'{{{{"{node.name}{root}:\n{type_str}"}}}}'
-            elif isinstance(node, (mm.Lookup, mm.Update, mm.Aggregate)):
+                    label = f'[/"{node.name}:\n{type_str}"\\]'
+            elif isinstance(node, (mm.Lookup, mm.Aggregate, mm.Update)):
                 arg_types = [self.resolve(arg) for arg in node.args]
                 if node.id in self.resolved_placeholder:
                     overloads = self.resolved_placeholder[node.id]
-                    content = "<br/>".join([reference_span(o, arg_types, root) for o in overloads])
+                    content = "<br/>".join([reference_span(o, arg_types) for o in overloads])
+                else:
+                    content = reference_span(get_relation(node), arg_types)
+                if isinstance(node, mm.Update):
+                    klass = ":::update"
+                    label = f'{{{{"{content}"}}}}'
                 else:
-                    content = reference_span(get_relation(node), arg_types, root)
-                label = f'[/"{content}"/]'
-            # elif isinstance(node, mm.Relation):
-            #     label = f'[("{node}")]'
+                    klass = ":::reference"
+                    label = f'[/"{content}"/]'
             else:
                 raise NotImplementedError(f"Unknown node type: {type(node)}")
             if self.has_errors(node):
@@ -555,6 +615,7 @@ class PropagationNetwork():
             flowchart TD
                 linkStyle default stroke:#666
                 classDef field fill:#245,stroke:#478
+                classDef update fill:#245,stroke:#478
                 classDef literal fill:#452,stroke:#784
                 classDef error fill:#624,stroke:#945,color:#f9a
                 classDef default stroke:#444,stroke-width:2px, font-size:12px
@@ -598,121 +659,42 @@ class Analyzer(Walker):
         rel = node.relation
         self.compute_potential_targets(rel)
 
-        # if this is a type relation, the update is asserting that the argument is of that
-        # type; so, it's fine to pass a super-type in to the population e.g. Employee(Person)
-        # should be a valid way to populate that a particular Person is also an Employee.
-        is_type_relation = isinstance(rel, mm.TypeNode)
+        # arg is flowing into a field
         for arg, field in zip(node.args, rel.fields):
-            field_type = field.type
-            arg_type = self.net.resolve(arg)
-
-            # if the arg is abstract, but the field isn't, then we need to make sure that
-            # once the arg is resolved we check that it matches the field type
-            if isinstance(arg, mm.Var) and bt.is_abstract(arg_type) and bt.is_concrete(field_type):
-                self.net.add_type_requirement(arg, field)
-
-            if bt.is_abstract(field_type) and isinstance(arg, (mm.Var, mm.Literal)):
-                # if the field is abstract, then eventually this arg will help determine
-                # the field's type, so add an edge from the arg to the field
-                self.net.add_edge(arg, field)
-            elif not type_matches(arg_type, field_type, accept_expected_super_types=is_type_relation):
-                if not conversion_allowed(arg_type, field_type):
-                    self.net.type_mismatch(node, field_type, arg_type)
+            if isinstance(arg, (mm.Var, mm.Literal)):
+                self.net.add_edge(arg, node)
 
     #--------------------------------------------------
     # Walk Lookups + Aggregates
     #--------------------------------------------------
 
-    def lookup(self, node: mm.Lookup):
-        self.compute_potential_targets(node.relation)
-        self.visit_rel_op(node)
-
-    def aggregate(self, node: mm.Aggregate):
-        self.visit_rel_op(node)
-
-    def visit_rel_op(self, node: mm.Lookup|mm.Aggregate):
-        rel = get_relation(node)
-
-        # special case eq lookups
-        if isinstance(node, mm.Lookup) and rel == b.core.eq:
-            # if both args for an eq are abstract, link them, otherwise do normal processing
-            (left, right) = node.args
-            left_type = self.net.resolve(left)
-            right_type = self.net.resolve(right)
-            if bt.is_abstract(left_type) and bt.is_abstract(right_type):
-                assert isinstance(left, mm.Var) and isinstance(right, mm.Var)
-                # if both sides are abstract, then whatever we find out about
-                # either should propagate to the other
-                self.net.add_edge(left, right)
-                self.net.add_edge(right, left)
-                return
-
-        # special case when the relation needs to be resolved as there are overloads, placeholders,
-        # type vars or it needs number specialization
-        if self.requires_resolution(rel):
-            return self.visit_unresolved_reference(node)
-
-        # if this is a population check, then it's fine to pass a subtype in to do the check
-        # e.g. Employee(Person) is a valid way to check if a person is an employee
-        is_population_lookup = isinstance(rel, mm.TypeNode)
-        for arg, field in zip(node.args, rel.fields):
-            field_type = self.net.resolve(field)
-            arg_type = self.net.resolve(arg)
-            if not type_matches(arg_type, field_type, is_population_lookup):
-                # Do not complain if we can convert the arg to the field type.
-                if not conversion_allowed(arg_type, field_type):
-                    # if the arg is a var and it matches when allowing for super types of
-                    # the expected we can expect to refine it later; but we add a type
-                    # requirement to check at the end
-                    if isinstance(arg, mm.Var) and type_matches(arg_type, field_type, True):
-                        self.net.add_type_requirement(arg, field)
-                    else:
-                        self.net.type_mismatch(node, field_type, arg_type)
-            # if we have an abstract var then this field will ultimately propagate to that
-            # var's type; also, if this is a population lookup, the type of the population
-            # being looked up will flow back to the var
-            if isinstance(arg, mm.Var):
-                if not field.input:
-                    self.net.add_edge(field, arg)
-                else:
-                    self.net.add_type_requirement(arg, field)
+    def lookup(self, task: mm.Lookup):
+        self.compute_potential_targets(task.relation)
+        self.visit_rel_op(task)
 
+    def aggregate(self, task: mm.Aggregate):
+        self.visit_rel_op(task)
 
-    def requires_resolution(self, rel: mm.Relation) -> bool:
-        # has overloads or is a placeholder relation that needs replacement
-        if rel.overloads or bt.is_placeholder(rel):
-            return True
-        # there are type vars or numbers in the fields that need specialization
-        for field in rel.fields:
-            t = self.net.resolve(field)
-            if bt.is_type_var(t) or t == b.core.Number:
-                return True
-        return False
-
+    def visit_rel_op(self, task: mm.Lookup|mm.Aggregate):
+        relation = get_relation(task)
 
-    def visit_unresolved_reference(self, node: mm.Lookup|mm.Aggregate):
-        relation = get_relation(node)
-        # functions have their outputs determined by their inputs
-        is_function = bt.is_function(relation)
         is_placeholder = bt.is_placeholder(relation)
-        # add edges between args and the relation based on input/output
-        for field, arg in zip(relation.fields, node.args):
+        for field, arg in zip(relation.fields, task.args):
             if isinstance(arg, (mm.Var, mm.Literal)):
                 if field.input:
-                    # the arg type will flow into the input
-                    self.net.add_edge(arg, node)
+                    # we need to resolve all inputs before resolving the relation
+                    self.net.add_edge(arg, task)
                 else:
-                    if is_function:
-                        # this is an output of a function, so the field type will flow to the arg
-                        self.net.add_edge(node, arg)
+                    # placeholders also need the output to be resolved
+                    if is_placeholder:
+                        self.net.add_edge(arg, task)
                     else:
-                        if is_placeholder:
-                            self.net.add_edge(arg, node)
-                        if bt.is_abstract(field.type) and not is_placeholder:
-                            self.net.add_edge(field, node)
-
-                        if bt.is_abstract(self.net.resolve(arg)):
-                            self.net.add_edge(node, arg)
+                        # args bound to outputs can be resolved after
+                        self.net.add_edge(task, arg)
+                        # if the field is abstract, it needs to be resolved before we can
+                        # resolve the task.
+                        if bt.is_abstract(self.net.resolve(field)):
+                            self.net.add_edge(field, task)
 
 
 #--------------------------------------------------
@@ -750,7 +732,7 @@ class Replacer(Rewriter):
         # TODO - this is only modifying the relation in the model, but then we have a new
         # relation there, which is different than the object referenced by tasks.
         if node in self.net.resolved_types:
-            return mm.Field(node.name, self.net.resolved_types[node], node.input)
+            return mm.Field(node.name, self.net.resolved_types[node], node.input, _relation = node._relation)
         return node
 
     def var(self, node: mm.Var):
@@ -773,31 +755,40 @@ class Replacer(Rewriter):
 
         args = types = None
         if node.id in self.net.resolved_placeholder:
+            # placeholder resolved to multiple relations
             resolved_relations = self.net.resolved_placeholder[node.id]
             args = get_lookup_args(node, resolved_relations[0])
             types = [f.type for f in resolved_relations[0].fields]
         elif node.id in self.net.resolved_overload:
+            # overload resolved to a specific relation
             resolved_relations = [node.relation]
             types = self.net.resolved_overload[node.id].types
         else:
+            # single relation
             resolved_relations = [node.relation]
 
         if len(resolved_relations) == 1:
+            # single relation, just convert arguments
             x = self.convert_arguments(node, resolved_relations[0], args, types)
             if isinstance(x, mm.Logical) and len(x.body) == 1:
                 return x.body[0]
             else:
                 return x
 
+        # multiple relations, create a union
         branches:list = []
         for target in resolved_relations:
             args = get_lookup_args(node, target)
             types = [f.type for f in get_relation_fields(resolved_relations[0], node.relation.name)]
             # adding this logical to avoid issues in the old backend
-            branches.append(mm.Logical((self.convert_arguments(node, target, args, types=types),)))
+            branches.append(mm.Logical((self.convert_arguments(node, target, args, types=types, force_copy=True),)))
         return mm.Union(tuple(branches))
 
-    def convert_arguments(self, node: mm.Lookup|mm.Update, relation: mm.Relation, args: Iterable[mm.Value]|None=None, types: Iterable[mm.Type]|None=None) -> mm.Logical|mm.Lookup|mm.Update:
+    def convert_arguments(self, node: mm.Lookup|mm.Update, relation: mm.Relation, args: Iterable[mm.Value]|None=None, types: Iterable[mm.Type]|None=None, force_copy=False) -> mm.Logical|mm.Lookup|mm.Update:
+        """ This node was resolved to target this relation using these args, which should
+        have these types. Convert any arguments as needed and return a new node with the
+        proper relation and converted args. If multiple conversions are needed, return a
+        logical that contains all the conversion tasks plus the final node. """
         args = args or node.args
         types = types or [self.net.resolve(f) for f in relation.fields]
         number_type = self.net.resolved_number.get(node.id)
@@ -817,15 +808,20 @@ class Replacer(Rewriter):
                     final_args.append(arg)
             else:
                 final_args.append(arg)
-        if isinstance(node, mm.Lookup):
-            tasks.append(node.mut(relation = relation, args = tuple(final_args)))
+        # add the original node with the proper target relation and converted args;
+        # here we want to use mut because we keep information about nodes based on their ids
+        # (e.g. we store resolved types based on ids), so we want to keep the same id. But
+        # when a lookup is being converted as part of a union (i.e. multiple targets for a
+        # placeholder), we need to create nodes with new ids to avoid conflicts.
+        if force_copy:
+            tasks.append(node.replace(relation = relation, args = tuple(final_args)))
         else:
             tasks.append(node.mut(relation = relation, args = tuple(final_args)))
+        # if we need conversion tasks, wrap in a logical
         if len(tasks) == 1:
             return tasks[0]
         return mm.Logical(tuple(tasks))
 
-
     def visit_eq_lookup(self, node: mm.Lookup):
         (left, right) = node.args
         left_type = to_type(left)
@@ -842,7 +838,7 @@ class Replacer(Rewriter):
         elif conversion_allowed(right_type, left_type):
             final_args = [left, convert(right, left_type, tasks)]
         else:
-            self.net.type_mismatch(node, left_type, right_type)
+            # this type mismatch was reported during propagation, so just return the node
             return node
 
         tasks.append(mm.Lookup(b.core.eq, tuple(final_args)))
@@ -873,6 +869,15 @@ def get_name(type: mm.Type) -> str:
 # Type and Relation helpers
 #--------------------------------------------------
 
+def get_update_fields(arg, update: mm.Update) -> Iterable[mm.Field]:
+    """ Get the fields of the relation being updated by this arg. Note that an arg can be
+    bound to multiple fields at the same time. """
+    if arg in update.args:
+        for x, field in zip(update.args, update.relation.fields):
+            if arg == x:
+                yield field
+    return []
+
 def get_relation_fields(relation: mm.Relation, name: str) -> Iterable[mm.Field]:
     """ Get the fields of this relation, potentially reordered to match the reading with the given name."""
     if name == relation.name:
@@ -915,18 +920,20 @@ def get_potential_targets(model: mm.Model, placeholder: mm.Relation) -> list[mm.
     return list(filter(lambda r: is_potential_target(placeholder, r), model.relations))
 
 def to_type(value: mm.Value|mm.Field|mm.Literal) -> mm.Type:
-    if isinstance(value, (mm.Var, mm.Field, mm.Literal)):
+    if isinstance(value, (mm.Var, mm.Literal)):
         return value.type
 
     if isinstance(value, mm.Type):
         return b.core.Type
 
+    if isinstance(value, mm.Field):
+        return b.core.Field
+
     if isinstance(value, tuple):
         return mm.TupleType(element_types=tuple(to_type(v) for v in value))
 
     raise TypeError(f"Cannot determine IR type for value: {value} of type {type(value).__name__}")
 
-
 def convert(value: mm.Var|mm.Literal, to_type: mm.Type, tasks: list[mm.Task]) -> mm.Value:
     # if the arg is a literal, we can just change its type
     # TODO - we may want to check that the value is actually convertible
@@ -940,8 +947,10 @@ def convert(value: mm.Var|mm.Literal, to_type: mm.Type, tasks: list[mm.Task]) ->
     tasks.append(mm.Lookup(b.core.cast, (to_type_base, value, new_value)))
     return new_value
 
-
 def conversion_allowed(from_type: mm.Type, to_type: mm.Type) -> bool:
+    # value type conversion is allowed only if the value types are related by inheritance
+    if bt.is_value_type(from_type) and bt.is_value_type(to_type) and not bt.extends(from_type, to_type):
+        return False
     # value type conversion is allowed
     x = bt.get_primitive_supertype(from_type)
     y = bt.get_primitive_supertype(to_type)
@@ -954,7 +963,7 @@ def conversion_allowed(from_type: mm.Type, to_type: mm.Type) -> bool:
 
     # a number can be converted to another number of larger scale
     if isinstance(from_type, mm.NumberType) and isinstance(to_type, mm.NumberType):
-        if to_type.scale > from_type.scale:
+        if to_type.scale >= from_type.scale:
             return True
 
     if from_type == b.core.Number and isinstance(to_type, mm.NumberType):
@@ -990,11 +999,6 @@ def type_matches(actual: mm.Type, expected: mm.Type, accept_expected_super_types
     if expected == b.core.TypeVar:
         return True
 
-    # TODO - remove this once we make them singletons per precision/scale
-    if isinstance(actual, mm.NumberType) and isinstance(expected, mm.NumberType):
-        if actual.precision == expected.precision and actual.scale == expected.scale:
-            return True
-
     # if an entity type var or any entity is expected, it matches any actual entity type
     if (expected == b.core.EntityTypeVar or bt.extends(expected, b.core.AnyEntity)) and not bt.is_primitive(actual):
         return True
@@ -1006,6 +1010,10 @@ def type_matches(actual: mm.Type, expected: mm.Type, accept_expected_super_types
     if (expected == b.core.Numeric) and bt.is_numeric(actual):
         return True
 
+    # different value types never match
+    if bt.is_value_type(actual) and bt.is_value_type(expected) and not bt.extends(actual, expected):
+        return False
+
     # if actual is scalar, any of its parents may match the expected type
     if isinstance(actual, mm.ScalarType) and any([type_matches(parent, expected) for parent in actual.super_types]):
         return True
@@ -1049,18 +1057,12 @@ def type_matches(actual: mm.Type, expected: mm.Type, accept_expected_super_types
 
     # accept tuples with a single element type to match a list with that type
     if isinstance(actual, mm.TupleType) and isinstance(expected, mm.ListType):
-        if len(set(actual.element_types)) == 1:
-            return type_matches(actual.element_types[0], expected.element_type)
+        return type_matches(actual.element_types[0], expected.element_type)
 
     # otherwise no match
     return False
 
-
-def merge_types(type1: mm.Type, type2: mm.Type) -> mm.Type:
-    if type1 == type2:
-        return type1
-    types_to_process = [type1, type2]
-
+def merge_numeric_types(type1: mm.Type, type2: mm.Type) -> Optional[mm.Type]:
     # if one of them is the abstract Number type, pick the other
     if type1 == b.core.Number and isinstance(type2, mm.NumberType):
         return type2
@@ -1077,6 +1079,24 @@ def merge_types(type1: mm.Type, type2: mm.Type) -> mm.Type:
     # if we are overriding a number with a float, pick float
     if isinstance(type1, mm.NumberType) and type2 == b.core.Float:
         return type2
+    if isinstance(type2, mm.NumberType) and type1 == b.core.Float:
+        return type1
+
+    return None
+
+def merge_types(type1: mm.Type, type2: mm.Type) -> mm.Type:
+    if type1 == type2:
+        return type1
+    if bt.is_type_var(type1):
+        return type2
+    if bt.is_type_var(type2):
+        return type1
+
+    types_to_process = [type1, type2]
+
+    numeric_merge = merge_numeric_types(type1, type2)
+    if numeric_merge is not None:
+        return numeric_merge
 
     # if one extends the other, pick the most specific one
     if bt.extends(type1, type2):
@@ -1093,6 +1113,13 @@ def merge_types(type1: mm.Type, type2: mm.Type) -> mm.Type:
         elif bt.is_primitive(type2):
             return type1
 
+    if base_primitive_type1 and base_primitive_type2:
+        numeric_merge = merge_numeric_types(base_primitive_type1, base_primitive_type2)
+        if numeric_merge == base_primitive_type1:
+            return type1
+        elif numeric_merge == base_primitive_type2:
+            return type2
+
     combined = OrderedSet()
     # Iterative flattening of union types
     while types_to_process: