Break up prepare_aaci logic

Now we convert the ACI into trees of opaque types, then flatten the tree
into a map and a list of function specs, and only then dereference the
types in the function specs down to our accelerated annotated types.

src/hz.erl

@@ -1392,99 +1392,108 @@ prepare_contract(File) ->
     end.
 
 prepare_aaci(ACI) ->
-    Types = lists:foldl(fun prepare_namespace_types/2, #{}, ACI),
+    % We want to take the types represented by the ACI, things like N1.T(N2.T),
+    % and dereference them down to concrete types like
+    % {tuple, [integer, string]}. Our type dereferencing algorithms
+    % shouldn't act directly on the JSON-based structures that the compiler
+    % gives us, though, though, so before we do the analysis, we should strip
+    % the ACI down to a list of 'opaque' type defintions and function specs.
+    {Name, OpaqueSpecs, TypeDefs} = convert_aci_types(ACI),
 
+    % Now that we have the opaque types, we can dereference the function specs
+    % down to the concrete types they actually represent.
+    Specs = expand_contract_specs(OpaqueSpecs, TypeDefs, #{}),
+
+    {aaci, Name, Specs, TypeDefs}.
+
+expand_contract_specs([], _Types, Specs) ->
+    Specs;
+expand_contract_specs([{Name, ArgsOpaque, ResultOpaque} | Rest], Types, Specs) ->
+    {ok, Args} = flatten_opaque_types(ArgsOpaque, Types, []),
+    {ok, Result} = flatten_opaque_type(ResultOpaque, Types),
+    NewSpecs = maps:put(Name, {Args, Result}, Specs),
+    expand_contract_specs(Rest, Types, NewSpecs).
+
+convert_aci_types(ACI) ->
+    % Find the main contract, so we can get the specifications of its
+    % entrypoints.
     [{NameBin, SpecDefs}] =
         [{N, F}
          || #{contract := #{kind      := contract_main,
                             functions := F,
                             name      := N}} <- ACI],
     Name = binary_to_list(NameBin),
-    Specs = simplify_specs(SpecDefs, #{}, Types),
+    % Turn these specifications into opaque types that we can reason about.
-    {aaci, Name, Specs, Types}.
+    Specs = lists:map(fun convert_function_spec/1, SpecDefs),
 
-prepare_namespace_types(#{namespace := NS}, Types) ->
+    % These specifications can reference other type definitions from the main
-    prepare_namespace_types2(NS, false, Types);
+    % contract and any other namespaces, so extract these types and convert
-prepare_namespace_types(#{contract := NS}, Types) ->
+    % them too.
-    prepare_namespace_types2(NS, true, Types).
+    TypeDefTree = lists:map(fun convert_namespace_typedefs/1, ACI),
+    % The tree structure of the ACI naturally leads to a tree of opaque types,
+    % but we want a map, so flatten it out before we continue.
+    TypeDefMap = collect_opaque_types(TypeDefTree, #{}),
 
-prepare_namespace_types2(NS, IsContract, Types) ->
+    % This is all the information we actually need from the ACI, the rest is
+    % just pre-compute and acceleration.
+    {Name, Specs, TypeDefMap}.
+
+convert_function_spec(#{name := NameBin, arguments := Args, returns := Result}) ->
+    Name = binary_to_list(NameBin),
+    ArgTypes = lists:map(fun convert_arg/1, Args),
+    ResultType = opaque_type([], Result),
+    {Name, ArgTypes, ResultType}.
+
+convert_arg(#{name := NameBin, type := TypeDef}) ->
+    Name = binary_to_list(NameBin),
+    {ok, Type} = opaque_type([], TypeDef),
+    {Name, Type}.
+
+convert_namespace_typedefs(#{namespace := NS}) ->
+    convert_namespace_typedefs2(NS, false);
+convert_namespace_typedefs(#{contract := NS}) ->
+    convert_namespace_typedefs2(NS, true).
+
+convert_namespace_typedefs2(NS, IsContract) ->
     TypeDefs = maps:get(typedefs, NS),
     NameBin = maps:get(name, NS),
     Name = binary_to_list(NameBin),
-    Types2 = case IsContract of
+    ContractAsType = case IsContract of
-                 true ->
+                         true  -> {Name, [], contract};
-                     maps:put(Name, {[], contract}, Types);
+                         false -> []
-                 false ->
-                     Types
                      end,
-    Types3 = case maps:find(state, NS) of
+    State = case maps:find(state, NS) of
                 {ok, StateDefACI} ->
                     StateDefOpaque = opaque_type([], StateDefACI),
-                     maps:put(Name ++ ".state", {[], StateDefOpaque}, Types2);
+                    {Name ++ ".state", [], StateDefOpaque};
                 error ->
-                     Types2
+                    []
             end,
-    simplify_typedefs(TypeDefs, Types3, Name ++ ".").
+    ExplicitTypeDefs = convert_explicit_typedefs(TypeDefs, Name ++ ".", []),
+    % Throw all the weird sources of types into one messy deeplist.
+    [ContractAsType, State, ExplicitTypeDefs].
 
-simplify_typedefs([], Types, _NamePrefix) ->
+% The easiest step, turn a deep list of opaque types into a map.
+collect_opaque_types([], Types) ->
     Types;
-simplify_typedefs([Next | Rest], Types, NamePrefix) ->
+collect_opaque_types([L | R], Types) ->
-    #{name := NameBin, vars := ParamDefs, typedef := T} = Next,
+    NewTypes = collect_opaque_types(L, Types),
+    collect_opaque_types(R, NewTypes);
+collect_opaque_types({Name, Params, Def}, Types) ->
+    maps:put(Name, {Params, Def}, Types).
+
+
+convert_explicit_typedefs([], _NamePrefix, Converted) ->
+    Converted;
+convert_explicit_typedefs([Next | Rest], NamePrefix, Converted) ->
+    #{name := NameBin, vars := ParamDefs, typedef := DefACI} = Next,
     Name = NamePrefix ++ binary_to_list(NameBin),
     Params = [binary_to_list(Param) || #{name := Param} <- ParamDefs],
-    Type = opaque_type(Params, T),
+    Def = opaque_type(Params, DefACI),
-    NewTypes = maps:put(Name, {Params, Type}, Types),
+    convert_explicit_typedefs(Rest, NamePrefix, [Converted, {Name, Params, Def}]).
-    simplify_typedefs(Rest, NewTypes, NamePrefix).
-
-simplify_specs([], Specs, _Types) ->
-    Specs;
-simplify_specs([Next | Rest], Specs, Types) ->
-    #{name := NameBin, arguments := ArgDefs, returns := ResultDef} = Next,
-    Name = binary_to_list(NameBin),
-    ArgTypes = [simplify_args(Arg, Types) || Arg <- ArgDefs],
-    {ok, ResultType} = type(ResultDef, Types),
-    NewSpecs = maps:put(Name, {ArgTypes, ResultType}, Specs),
-    simplify_specs(Rest, NewSpecs, Types).
-
-simplify_args(#{name := NameBin, type := TypeDef}, Types) ->
-    Name = binary_to_list(NameBin),
-    % FIXME We should make this error more informative, and continue
-    % propogating it up, so that the user can provide their own ACI and find
-    % out whether it worked or not. At that point ACI -> AACI could almost be a
-    % module or package of its own.
-    {ok, Type} = type(TypeDef, Types),
-    {Name, Type}.
-
-% Type preparation has two goals. First, we need a data structure that can be
-% traversed quickly, to take sophia-esque erlang expressions and turn them into
-% fate-esque erlang expressions that gmbytecode can serialize. Second, we need
-% partially substituted names, so that error messages can be generated for why
-% "foobar" is not valid as the third field of a `bazquux`, because the third
-% field is supposed to be `option(integer)`, not `string`.
-%
-% To achieve this we need three representations of each type expression, which
-% together form an 'annotated type'. First, we need the fully opaque name,
-% "bazquux", then we need the normalized name, which is an opaque name with the
-% bare-minimum substitution needed to make the outer-most type-constructor an
-% identifiable built-in, ADT, or record type, and then we need the flattened
-% type, which is the raw {variant, [{Name, Fields}, ...]} or
-% {record, [{Name, Type}]} expression that can be used in actual Sophia->FATE
-% coercion. The type sub-expressions in these flattened types will each be
-% fully annotated as well, i.e. they will each contain *all three* of the above
-% representations, so that coercion of subexpressions remains fast AND
-% informative.
-%
-% In a lot of cases the opaque type given will already be normalized, in which
-% case either the normalized field or the non-normalized field of an annotated
-% type can simple be the atom `already_normalized`, which means error messages
-% can simply render the normalized type expression and know that the error will
-% make sense.
-
-type(T, Types) ->
-    O = opaque_type([], T),
-    flatten_opaque_type(O, Types).
 
+% Convert an ACI type defintion/spec into the 'opaque type' representation that
+% our dereferencing algorithms can reason about.
 opaque_type(Params, NameBin) when is_binary(NameBin) ->
     Name = opaque_type_name(NameBin),
     case not is_atom(Name) and lists:member(Name, Params) of
@@ -1508,7 +1517,7 @@ opaque_type(Params, Pair) when is_map(Pair) ->
     [{Name, TypeArgs}] = maps:to_list(Pair),
     {opaque_type_name(Name), [opaque_type(Params, Arg) || Arg <- TypeArgs]}.
 
-% atoms for builtins, lists for user defined types
+% atoms for builtins, strings (lists) for user-defined types
 opaque_type_name(<<"int">>)      -> integer;
 opaque_type_name(<<"address">>)  -> address;
 opaque_type_name(<<"contract">>) -> contract;
@@ -1519,6 +1528,31 @@ opaque_type_name(<<"map">>)      -> map;
 opaque_type_name(<<"string">>)   -> string;
 opaque_type_name(Name)           -> binary_to_list(Name).
 
+% Type preparation has two goals. First, we need a data structure that can be
+% traversed quickly, to take sophia-esque erlang expressions and turn them into
+% fate-esque erlang expressions that gmbytecode can serialize. Second, we need
+% partially substituted names, so that error messages can be generated for why
+% "foobar" is not valid as the third field of a `bazquux`, because the third
+% field is supposed to be `option(integer)`, not `string`.
+%
+% To achieve this we need three representations of each type expression, which
+% together form an 'annotated type'. First, we need the fully opaque name,
+% "bazquux", then we need the normalized name, which is an opaque name with the
+% bare-minimum substitution needed to make the outer-most type-constructor an
+% identifiable built-in, ADT, or record type, and then we need the flattened
+% type, which is the raw {variant, [{Name, Fields}, ...]} or
+% {record, [{Name, Type}]} expression that can be used in actual Sophia->FATE
+% coercion. The type sub-expressions in these flattened types will each be
+% fully annotated as well, i.e. they will each contain *all three* of the above
+% representations, so that coercion of subexpressions remains fast and
+% informative.
+%
+% In a lot of cases the opaque type given will already be normalized, in which
+% case either the normalized field or the non-normalized field of an annotated
+% type can simple be the atom `already_normalized`, which means error messages
+% can simply render the normalized type expression and know that the error will
+% make sense.
+
 flatten_opaque_type(T, Types) ->
     case normalize_opaque_type(T, Types) of
         {ok, AlreadyNormalized, NOpaque, NExpanded} ->