Rationale for Ada 2005

John Barnes
Contents   Index   References   Search   Previous   Next 

8.3 Maps

We will now turn to the maps and sets packages. We will start by considering maps which are more exciting than sets and begin with ordered maps which are a little simpler and then consider hashed maps.
Remember that a map is just a means of getting from a value of one type (the key) to another type (the element). This is not a one-one relationship. Given a key there is a unique element (if any), but several keys may correspond to the same element. A simple example is an array. This is a map from the index type to the component type. Thus if we have 
S: String := "animal";
then this provides a map from integers in the range 1 to 6 to some values of the type Character. Given an integer such as 3 there is a unique character 'i' but given a character such as 'a' there might be several corresponding integers (in this case both 1 and 5).
More interesting examples are where the set of used key values is quite sparse. For example we might have a store where various spare parts are held. The parts have a five-digit part number and there are perhaps twenty racks where they are held identified by a letter. However, only a handful of the five digit numbers are in use so it would be very wasteful to use an array with the part number as index. What we want instead is a container which holds just the pairs that matter such as (34618, 'F'), (27134, 'C') and so on. We can do this using a map. We usually refer to the pairs of values as nodes of the map.
There are two maps packages with much in common. One keeps the keys in order and the other uses a hash function. Here is the specification of the ordered maps package generally showing just those facilities common to both.
   type Key_Type is private;
   type Element_Type is private;
   with function "<" (Left, Right: Key_Type) return Boolean is <>;
   with function "=" (Left, Right: Element_Type) return Boolean is <>;
package Ada.Containers.Ordered_Maps is
   pragma Preelaborate(Ordered_Maps);
   function Equivalent_Keys(Left: Right: Key_Type) return Boolean;
The generic parameters include the ordering relationship "<" on the keys and equality for the elements.
It is assumed that the ordering relationship is well behaved in the sense that if x < y is true then y < x is false. We say that two keys x and y are equivalent if both x < y and y < x are false. In other words this defines an equivalence class on keys. The relationship must also be transitive, that is, if x < y and y < z are both true then x < z must also be true.
This concept of an equivalence relationship occurs throughout the various maps and sets. Sometimes, as here, it is defined in terms of an order but in other cases, as we shall see, it is defined by an equivalence function.
It is absolutely vital that the equivalence relations are defined properly and meet the above requirements. It is not possible for the container packages to check this and if the operations are wrong then peculiar behaviour is almost inevitable.
For the convenience of the user the function Equivalent_Keys is declared explicitly. It is equivalent to 
function Equivalent_Keys(Left, Right: Key_Type) return Boolean is
   return not (Left < Right) and not (Right < Left);
end Equivalent_Keys;
The equality operation on elements is not so demanding. It must be symmetric so that x = y and y = x are the same but transitivity is not required (although cases where it would not automatically be transitive are likely to be rare). The operation is only used for the function "=" on the containers as a whole.
Note that Find and similar operations for maps and sets work in terms of the equivalence relationship rather than equality as was the case with lists and vectors.
   type Map is tagged private;
   pragma Preelaborable_Initialization(Map);
   type Cursor is private;
   pragma Preelaborable_Initialization(Cursor);
   Empty_Map: constant Map;
   No_Element: constant Cursor;
The types Map and Cursor and constants Empty_Map and No_Element are similar to the corresponding entities in the lists and vectors containers.
function "=" (Left, Right: Map) return Boolean;
function Length(Container: Map) return Count_Type;
function Is_Empty(Container: Map) return Boolean;
procedure Clear(Container: in out Map);
These are again similar to the corresponding entities for lists. Note that two maps are said to be equal if they have the same number of nodes with equivalent keys (as defined by "<") whose corresponding elements are equal (as defined by "=").
function Key(Position: Cursor) return Key_Type;
function Element(Position: Cursor) return Element_Type;
procedure Replace_Element(
      Container: in out Map;
      Position: in Cursor;
      New_Item: in Element_Type);
procedure Query_Element(
      Position: in Cursor;
      Process: not null access procedure (Key: in Key_Type; Element: in Element_Type));
procedure Update_Element(
      Container: in out Map; Position: in Cursor;
      Process: not null access procedure (Key: in Key_Type; Element: in out Element_Type));
In this case there is a function Key as well as a function Element. But there is no procedure Replace_Key since it would not make sense to change a key without changing the element as well and this really comes down to deleting the whole node and then inserting a new one.
The procedures Query_Element and Update_Element are slightly different in that the procedure Process also takes the key as parameter as well as the element to be read or updated. Note again that the key cannot be changed. Nevertheless the value of the key is given since it might be useful in deciding how the update should be performed. Remember that we cannot get uniquely from an element to a key but only from a key to an element.
procedure Move(Target, Source: in out Map);
This moves the map from the source to the target after first clearing the target. It does not make copies of the nodes so that after the operation the source is empty and Length(Source) is zero.
procedure Insert(
      Container: in out Map;
      Key: in Key_Type;
      New_Item: in Element_Type;
      Position: out Cursor;
      Inserted: out Boolean);
procedure Insert(
      Container: in out Map;
      Key: in Key_Type;
      Position: out Cursor;
      Inserted: out Boolean);
procedure Insert(
      Container: in out Map;
      Key: in Key_Type;
      New_Item: in Element_Type);
These insert a new node into the map unless a node with an equivalent key already exists. If it does exist then the first two return with Inserted set to False and Position indicating the node whereas the third raises Constraint_Error (the element value is not changed). If a node with equivalent key is not found then a new node is created with the given key, the element value is set to New_Item when that is given and otherwise it takes its default value (if any), and Position is set when given.
Unlike vectors and lists, we do not have to say where the new node is to be inserted because of course this is an ordered map and it just goes in the correct place according to the order given by the generic parameter "<".
procedure Include(
      Container: in out Map;
      Key: in Key_Type;
      New_Item: in Element_Type);
This is somewhat like the last Insert except that if an existing node with an equivalent key is found then it is replaced (rather than raising Constraint_Error). Note that both the key and the element are updated. This is because equivalent keys might not be totally equal.
For example the key part might be a record with part number and year of introduction, thus
type Part_Key is
      Part_Number: Integer;
      Year: Integer;
   end record;
and we might define the ordering relationship to be used as the generic parameter simply in terms of the part number 
function "<" (Left, Right: Part_Key) return Boolean is
   return Left.Part_Number < Right.Part_Number;
end "<";
In this situation, the keys could match without the year component being the same and so it would need to be updated. In other words with this definition of the ordering relation, two keys are equivalent provided just the part numbers are the same.
procedure Replace(
      Container: in out Map;
      Key: in Key_Type;
      New_Item: in Element_Type);
In this case, Constraint_Error is raised if the node does not already exist. On replacement both the key and the element are updated as for Include.
Perhaps a better example of equivalent keys not being totally equal is if the key were a string. We might decide that the case of letter did not need to match in the test for equivalence but nevertheless we would probably want to update with the string as used in the parameter of Replace
procedure Exclude(Container: in out Map; Key: in Key_Type);
If there is a node with an equivalent key then it is deleted. If there is not then nothing happens.
procedure Delete(Container: in out Map; Key: in Key_Type);
procedure Delete(Container: in out Map; Position: in out Cursor);
These delete a node. In the first case if there is no such equivalent key then Constraint_Error is raised (by contrast to Exclude which remains silent in this case). In the second case if the cursor is No_Element then again Constraint_Error is raised – there is also a check to ensure that the cursor otherwise does designate a node in the correct map (remember that cursors designate both an entity and the container); if this check fails then Program_Error is raised.
Perhaps it is worth observing that Insert, Include, Replace, Exclude and Delete form a sort of progression from an operation that will insert something, through operations that might insert, will neither insert nor delete, might delete, to the final operation that will delete something. Note also that Include, Replace and Exclude do not apply to lists and vectors.
function First(Container: Map) return Cursor;
function Last(Container: Map) return Cursor;
function Next(Position: Cursor) return Cursor;
procedure Next(Position: in out Cursor);
function Find(Container: Map; Key: Key_Type) return Cursor;
function Element(Container: Map; Key: Key_Type) return Element;
function Contains(Container: Map; Key: Key_Type) return Boolean;
These should be self-evident. Unlike the operations on vectors and lists, Find logically searches the whole map and not just starting at some point (and since it searches the whole map there is no point in having Reverse_Find). (In implementation terms it won't actually search the whole map because it will be structured in a way that makes this unnecessary – as a balanced tree perhaps.) Moreover, Find uses the equivalence relation based on the "<" parameter so in the example it only has to match the part number and not the year. The function call Element(My_Map, My_Key) is equivalent to Element(Find(My_Map, My_Key)).
function Has_Element(Position: Cursor) return Boolean;
procedure Iterate(
      Container: in Map;
      Process: not null access procedure (Position: in Cursor));
These are also as for other containers.
And at last we have
   ... -- not specified by the language
end Ada.Containers.Ordered_Maps;
We have omitted to mention quite a few operations that have no equivalent in hashed maps – we will come back to these in a moment.
As an example we can make a container to hold the information concerning spare parts. We can use the type Part_Key and the function "<" as above. We can suppose that the element type is 
type Stock_Info is
      Shelf: Character range 'A' .. 'T';
      Stock: Integer;
   end record;
This gives both the shelf letter and the number in stock.
We can then declare the container thus 
package Store_Maps is
   new Ordered_Maps(
       Key_Type => Part_Key,
       Element_Type => Stock_Info,
       "<" => "<");
The_Store: Store_Maps.Map;
The last parameter could be omitted because the formal has a <> default.
We can now add items to our store by calling 
The_Store.Insert((34618, 1998), ('F', 25));
The_Store.Insert((27134, 2004), ('C', 45));
We might now have a procedure which, given a part number, checks to see if it exists and that the stock is not zero, and if so returns the shelf letter and year number and decrements the stock count.
procedure Request(
      Part: in Integer; OK: out Boolean;
      Year: out Integer; Shelf: out Character) is
   C: Cursor;
   K: Part_Key;
   E: Stock_Info;
   C := The_Store.Find((Part, 0));
   if C = No_Element then
      OK := False; return;    -- no such key
   end if;
   E := Element(C);  K := Key(C);
   Year := K.Year;  Shelf := E.Shelf;
   if E.Stock = 0 then
      OK := False; return;    -- out of stock
   end if;
   Replace_Element(C, (Shelf, E.Stock–1));
   OK := True;
end Request;
Note that we had to put a dummy year number in the call of Find. We could of course use the new <> notation for this 
C := The_Store.Find((Part, others => <>));
The reader can improve this example at leisure – by using Update_Element for example.
As another example suppose we wish to check all through the stock looking for parts whose stock is low, perhaps less than some given parameter. We can use Iterate for this as follows 
procedure Check_Stock(Low: in Integer) is
   procedure Check_It(C: in Cursor) is
      if Element(C).Stock < Low then
         -- print a message perhaps
         Put("Low stock of part ");
      end if;
   end Check_It;
end Check_Stock;
Note that this uses a so-called downward closure. The procedure Check_It has to be declared locally to Check_Stock in order to access the parameter Low. (Well you could declare it outside and copy the parameter Low to a global variable but that is just the sort of wicked thing one has to do in lesser languages (such as even Ada 95). It is not task safe for one thing.)
Another approach is to use First and Next and so on thus 
procedure Check_Stock(Low: in Integer) is
   C: Cursor := The_Store.First;
      exit when C = No_Element;
      if Element(C).Stock < Low then
         -- print a message perhaps
         Put("Low stock of part ");
      end if;
      C := The_Store.Next(C);
   end loop;
end Check_Stock;
We will now consider hashed maps. The trouble with ordered maps in general is that searching can be slow when the map has many entries. Techniques such as a binary tree can be used but even so the search time will increase at least as the logarithm of the number of entries. A better approach is to use a hash function. This will be familiar to many readers (especially those who have written compilers). The general idea is as follows.
We define a function which takes a key and returns some value in a given range. In the case of the Ada containers it has to return a value of the modular type Hash_Type which is declared in the root package Ada.Containers. We could then convert this value onto a range representing an index into an array whose size corresponds to the capacity of the map. This index value is the preferred place to store the entry. If there already is an entry at this place (because some other key has hashed to the same value) then a number of approaches are possible. One way is to create a list of entries with the same index value (often called buckets); another way is simply to put it in the next available slot. The details don't matter. But the overall effect is that provided the map is not too full and the hash function is good then we can find an entry almost immediately more or less irrespective of the size of the map.
So as users all we have to do is to define a suitable hash function. It should give a good spread of values across the range of Hash_Type for the population of keys, it should avoid clustering and above all for a given key it must always return the same hash value. A good discussion on hash functions by Knuth will be found in [10].
Defining good hash functions needs care. In the case of the part numbers we might multiply the part number by some obscure prime number and then truncate the result down to the modular type Hash_Type. The author hesitates to give an example but perhaps 
function Part_Hash(P: Part_Key) return Hash_Type is
   M31: constant := 2**31–1;    -- a nice Mersenne prime
   return Hash_Type(P.Part_Number) * M31;
end Part_Hash;
On reflection that's probably a very bad prime to use because it is so close to half of 2**32 a typical value of Hash_Type'Last+1. Of course it doesn't have to be prime but simply relatively prime to it such as 5**13. Knuth suggests dividing the range by the golden number τ = (√5+1)/2 = 1.618... and then taking the nearest number relatively prime which is in fact simply the nearest odd number (in this case it is 2654435769).
Here is a historic interlude. Marin Mersenne (1588-1648) was a Franciscan monk who lived in Paris. He studied numbers of the form Mp = 2p – 1 where p is prime. A lot of these are themselves prime. Mersenne gave a list of those up to 257 which he said were prime (namely 2, 3, 5, 7, 13, 17, 19, 31, 67, 127, 257). It was not until 1947 that it was finally settled that he got some of them wrong (61, 89, and 107 are also prime but 67 and 257 are not). At the time of writing there are 44 known Mersenne primes and the largest which is also the largest known prime number is M32582657 – see www.mersenne.org.
The specification of the hashed maps package is very similar to that for ordered maps. It starts
   type Key_Type is private;
   type Element_Type is private;
   with function Hash(Key: Key_Type) return Hash_Type;
   with function Equivalent_Keys(Left, Right: Key_Type) return Boolean;
   with function "=" (Left, Right: Element_Type) return Boolean is <>;
package Ada.Containers.Hashed_Maps is
   pragma Preelaborate(Hashed_Maps);
The differences from the ordered maps package are that there is an extra generic parameter Hash giving the hash function and the ordering parameter "<" has been replaced by the function Equivalent_Keys. It is this function that defines the equivalence relationship for hashed maps; it is important that Equivalent_Keys(X, Y) is always the same as Equivalent_Keys(Y, X). Moreover if X and Y are equivalent and Y and Z are equivalent then X and Z must also be equivalent.
Note that the function Equivalent_Keys in the ordered maps package discussed above corresponds to the formal generic parameter of the same name in this hashed maps package. This should make it easier to convert between the two forms of packages.
Returning to our example, if we now write 
function Equivalent_Parts(Left, Right: Part_Key) return Boolean is
   return Left.Part_Number = Right.Part_Number;
end Equivalent_Parts;
then we can instantiate the hashed maps package as follows 
package Store_Maps is new Hashed_Maps(
      Key_Type => Part_Key,
      Element_Type => Stock_Info,
      Hash => Part_Hash,
      Equivalent_Keys => Equivalent_Parts);
The_Store: Store_Maps.Map;
and then the rest of our example will be exactly as before. It is thus easy to convert from an ordered map to a hashed map and vice versa provided of course that we only use the facilities common to both.
We will finish this discussion of maps by briefly considering the additional facilities in the two packages.
The ordered maps package has the following additional subprograms 
procedure Delete_First(Container: in out Map);
procedure Delete_Last(Container: in out Map);
function First_Element(Container: Map) return Element_Type;
function First_Key(Container: Map) return Key_Type;
function Last_Element(Container: Map) return Element_Type;
function Last_Key(Container: Map) return Key_Type;
function Previous(Position: Cursor) return Cursor;
procedure Previous(Position: in out Cursor);
function Floor(Container: Map; Key: Key_Type) return Cursor;
function Ceiling(Container: Map; Key: Key_Type) return Cursor;
function "<" (Left, Right: Cursor) return Boolean;
function ">" (Left, Right: Cursor) return Boolean;
function "<" (Left: Cursor; Right: Key_Type) return Boolean;
function ">" (Left: Cursor; Right: Key_Type) return Boolean;
function "<" (Left: Key_Type; Right: Cursor) return Boolean;
function ">" (Left: Key_Type; Right: Cursor) return Boolean;
procedure Reverse_Iterate(
      Container: in Map;
      Process: not null access procedure (Position: in Cursor));
These are again largely self-evident. The functions Floor and Ceiling are interesting. Floor searches for the last node whose key is not greater than Key and similarly Ceiling searches for the first node whose key is not less than Key – they return No_Element if there is no such element. The subprograms Previous are of course the opposite of Next and Reverse_Iterate is like Iterate only backwards.
The functions "<" and ">" are mostly for convenience. Thus the first is equivalent to 
function "<" (Left, Right: Cursor) return Boolean is
   return Key(Left) < Key(Right);
end "<";
Clearly these additional operations must be avoided if we wish to retain the option of converting to a hashed map later.
Hashed maps have a very important facility not in ordered maps which is the ability to specify a capacity as for the vectors package. (Underneath their skin the hashed maps are a bit like vectors whereas the ordered maps are a bit like lists.) Thus we have 
procedure Reserve_Capacity(Container: in out Map; Capacity: in Count_Type);
function Capacity(Container: Map) return Count_Type;
The behaviour is much as for vectors. We don't have to set the capacity ourselves since it will be automatically extended as necessary but it might significantly improve performance to do so. In the case of maps, increasing the capacity requires the hashing to be redone which could be quite time consuming, so if we know that our map is going to be a big one, it is a good idea to set an appropriate capacity right from the beginning. Note again that Length(M) cannot exceed Capacity(M) but might be much less.
The other additional subprograms for hashed maps are 
function Equivalent_Keys(Left, Right: Cursor) return Boolean;
function Equivalent_Keys(Left: Cursor; Right: Key_Type) return Boolean;
function Equivalent_Keys(Left: Key_Type; Right: Cursor) return Boolean;
These (like the additional "<" and ">" for ordered maps) are again mostly for convenience. The first is equivalent to 
function Equivalent_Keys(Left, Right: Cursor) return Boolean is
   return Equivalent_Keys(Key(Left), Key(Right));
end Equivalent_Keys;
Before moving on to sets it should be noticed that there are also some useful functions in the string packages. The main one is
with Ada.Containers;
function Ada.Strings.Hash(Key: String) return Containers.Hash_Type;
pragma Pure(Ada.Strings.Hash);
There is a similar function Ada.Strings.Unbounded.Hash where the parameter Key has type Unbounded_String. It simply converts the parameter to the type String and then calls Ada.Strings.Hash. There is also a generic function for bounded strings which again calls the basic function Ada.Strings.Hash. For completeness the function Ada.Strings.Fixed.Hash is a renaming of Ada.Strings.Hash.
These are provided because it is often the case that the key is a string and they save the user from devising good hash functions for strings which might cause a nasty headache.
We could for example save ourselves the worry of defining a good hash function in the above example by making the part number into a 5-character string. So we might write 
function Part_Hash(P: Part_Key) return Hash_Type is
   return Ada.Strings.Hash(P.Part_Number);
end Part_Hash;
and if this doesn't work well then we can blame the vendor.

Contents   Index   References   Search   Previous   Next 
© 2005, 2006, 2007 John Barnes Informatics.
Sponsored in part by:
The Ada Resource Association and its member companies: ARA Members AdaCore Polyspace Technologies Praxis Critical Systems IBM Rational Sofcheck and   Ada-Europe: