A std::map that keep track of the order of inserti

问题:

I currently have a std::map<std::string,int> that stores an integer value to an unique string identifier, and I do look up with the string. It does mostly what I want, except for that it does not keep track of the insertion order. So when I iterate the the map to print out the values, they are sorted according to the string; but I want them to be sorted according to the order of (first) insertion.

I thought about using a vector<pair<string,int>> instead, but I need to look up the string and increment the integer values about 10,000,000 times, so I don't know whether a std::vector will be significantly slower.

Is there a way to use std::map or is there another std container that better suits my need?

[I'm on GCC 3.4, and I have probably no more than 50 pairs of values in my std::map].

Thanks.

回答1:

If you have only 50 values in std::map you could copy them to std::vector before printing out and sort via std::sort using appropriate functor.

Or you could use boost::multi_index. It allows to use several indexes. In your case it could look like the following:

struct value_t {
      string s;
      int    i;
};
struct string_tag {};
typedef multi_index_container<
    value_t,
    indexed_by<
        random_access<>, // this index represents insertion order
        hashed_unique< tag<string_tag>, member<value_t, string, &value_t::s> >
    >
> values_t;

回答2:

You might combine a std::vector with a std::tr1::unordered_map (a hash table). Here's a link to Boost's documentation for unordered_map. You can use the vector to keep track of the insertion order and the hash table to do the frequent lookups. If you're doing hundreds of thousands of lookups, the difference between O(log n) lookup for std::map and O(1) for a hash table might be significant.

std::vector<std::string> insertOrder;
std::tr1::unordered_map<std::string, long> myTable;

// Initialize the hash table and record insert order.
myTable["foo"] = 0;
insertOrder.push_back("foo");
myTable["bar"] = 0;
insertOrder.push_back("bar");
myTable["baz"] = 0;
insertOrder.push_back("baz");

/* Increment things in myTable 100000 times */

// Print the final results.
for (int i = 0; i < insertOrder.size(); ++i)
{
    const std::string &s = insertOrder[i];
    std::cout << s << ' ' << myTable[s] << '\n';
}

回答3:

Keep a parallel list<string> insertionOrder.

When it is time to print, iterate on the list and do lookups into the map.

each element in insertionOrder  // walks in insertionOrder..
    print map[ element ].second // but lookup is in map

回答4:

Tessil has a very nice implementaion of ordered map (and set) which is MIT license. You can find it here: ordered-map

Map example

#include <iostream>
#include <string>
#include <cstdlib>
#include "ordered_map.h"

int main() {
tsl::ordered_map<char, int> map = {{'d', 1}, {'a', 2}, {'g', 3}};
map.insert({'b', 4});
map['h'] = 5;
map['e'] = 6;

map.erase('a');


// {d, 1} {g, 3} {b, 4} {h, 5} {e, 6}
for(const auto& key_value : map) {
    std::cout << "{" << key_value.first << ", " << key_value.second << "}" << std::endl;
}


map.unordered_erase('b');

// Break order: {d, 1} {g, 3} {e, 6} {h, 5}
for(const auto& key_value : map) {
    std::cout << "{" << key_value.first << ", " << key_value.second << "}" << std::endl;
}
}

回答5:

If you need both lookup strategies, you will end up with two containers. You may use a vector with your actual values (ints), and put a map< string, vector< T >::difference_type> next to it, returning the index into the vector.

To complete all that, you may encapsulate both in one class.

But I believe boost has a container with multiple indices.

回答6:

You cannot do that with a map, but you could use two separate structures - the map and the vector and keep them synchronized - that is when you delete from the map, find and delete the element from the vector. Or you could create a map<string, pair<int,int>> - and in your pair store the size() of the map upon insertion to record position, along with the value of the int, and then when you print, use the position member to sort.

回答7:

This is somewhat related to Faisals answer. You can just create a wrapper class around a map and vector and easily keep them synchronized. Proper encapsulation will let you control the access method and hence which container to use... the vector or the map. This avoids using Boost or anything like that.

回答8:

Another way to implement this is with a map instead of a vector. I will show you this approach and discuss the differences:

Just create a class that has two maps behind the scenes.

#include <map>
#include <string>

using namespace std;

class SpecialMap {
  // usual stuff...

 private:
  int counter_;
  map<int, string> insertion_order_;
  map<string, int> data_;
};

You can then expose an iterator to iterator over data_ in the proper order. The way you do that is iterate through insertion_order_, and for each element you get from that iteration, do a lookup in the data_ with the value from insertion_order_

You can use the more efficient hash_map for insertion_order since you don't care about directly iterating through insertion_order_.

To do inserts, you can have a method like this:

void SpecialMap::Insert(const string& key, int value) {
  // This may be an over simplification... You ought to check
  // if you are overwriting a value in data_ so that you can update
  // insertion_order_ accordingly
  insertion_order_[counter_++] = key;
  data_[key] = value;
}

There are a lot of ways you can make the design better and worry about performance, but this is a good skeleton to get you started on implementing this functionality on your own. You can make it templated, and you might actually store pairs as values in data_ so that you can easily reference the entry in insertion_order_. But I leave these design issues as an exercise :-).

Update: I suppose I should say something about efficiency of using map vs. vector for insertion_order_

• lookups directly into data, in both cases are O(1)
• inserts in the vector approach are O(1), inserts in the map approach are O(logn)
• deletes in the vector approach are O(n) because you have to scan for the item to remove. With the map approach they are O(logn).

Maybe if you are not going to use deletes as much, you should use the vector approach. The map approach would be better if you were supporting a different ordering (like priority) instead of insertion order.

回答9:

// Should be like this man!

// This maintains the complexity of insertion is O(logN) and deletion is also O(logN).

class SpecialMap {
private:
  int counter_;
  map<int, string> insertion_order_;
  map<string, int> insertion_order_reverse_look_up; // <- for fast delete
  map<string, Data> data_;
};

回答10:

What you want (without resorting to Boost) is what I call an "ordered hash", which is essentially a mashup of a hash and a linked list with string or integer keys (or both at the same time). An ordered hash maintains the order of the elements during iteration with the absolute performance of a hash.

I've been putting together a relatively new C++ snippet library that fills in what I view as holes in the C++ language for C++ library developers. Go here:

https://github.com/cubiclesoft/cross-platform-cpp

Grab:

templates/detachable_ordered_hash.cpp
templates/detachable_ordered_hash.h
templates/detachable_ordered_hash_util.h

If user-controlled data will be placed into the hash, you might also want:

security/security_csprng.cpp
security/security_csprng.h

Invoke it:

#include "templates/detachable_ordered_hash.h"
...
// The 47 is the nearest prime to a power of two
// that is close to your data size.
//
// If your brain hurts, just use the lookup table
// in 'detachable_ordered_hash.cpp'.
//
// If you don't care about some minimal memory thrashing,
// just use a value of 3.  It'll auto-resize itself.
int y;
CubicleSoft::OrderedHash<int> TempHash(47);
// If you need a secure hash (many hashes are vulnerable
// to DoS attacks), pass in two randomly selected 64-bit
// integer keys.  Construct with CSPRNG.
// CubicleSoft::OrderedHash<int> TempHash(47, Key1, Key2);
CubicleSoft::OrderedHashNode<int> *Node;
...
// Push() for string keys takes a pointer to the string,
// its length, and the value to store.  The new node is
// pushed onto the end of the linked list and wherever it
// goes in the hash.
y = 80;
TempHash.Push("key1", 5, y++);
TempHash.Push("key22", 6, y++);
TempHash.Push("key3", 5, y++);
// Adding an integer key into the same hash just for kicks.
TempHash.Push(12345, y++);
...
// Finding a node and modifying its value.
Node = TempHash.Find("key1", 5);
Node->Value = y++;
...
Node = TempHash.FirstList();
while (Node != NULL)
{
  if (Node->GetStrKey())  printf("%s => %d\n", Node->GetStrKey(), Node->Value);
  else  printf("%d => %d\n", (int)Node->GetIntKey(), Node->Value);

  Node = Node->NextList();
}

I ran into this SO thread during my research phase to see if anything like OrderedHash already existed without requiring me to drop in a massive library. I was disappointed. So I wrote my own. And now I've shared it.

回答11:

Here is solution that requires only standard template library without using boost's multiindex:
You could use std::map<std::string,int>; and vector <data>; where in map you store the index of the location of data in vector and vector stores data in insertion order. Here access to data has O(log n) complexity. displaying data in insertion order has O(n) complexity. insertion of data has O(log n) complexity.

For Example:

#include<iostream>
#include<map>
#include<vector>

struct data{
int value;
std::string s;
}

typedef std::map<std::string,int> MapIndex;//this map stores the index of data stored 
                                           //in VectorData mapped to a string              
typedef std::vector<data> VectorData;//stores the data in insertion order

void display_data_according_insertion_order(VectorData vectorData){
    for(std::vector<data>::iterator it=vectorData.begin();it!=vectorData.end();it++){
        std::cout<<it->value<<it->s<<std::endl;
    }
}
int lookup_string(std::string s,MapIndex mapIndex){
    std::MapIndex::iterator pt=mapIndex.find(s)
    if (pt!=mapIndex.end())return it->second;
    else return -1;//it signifies that key does not exist in map
}
int insert_value(data d,mapIndex,vectorData){
    if(mapIndex.find(d.s)==mapIndex.end()){
        mapIndex.insert(std::make_pair(d.s,vectorData.size()));//as the data is to be
                                                               //inserted at back 
                                                               //therefore index is
                                                               //size of vector before
                                                               //insertion
        vectorData.push_back(d);
        return 1;
    }
    else return 0;//it signifies that insertion of data is failed due to the presence
                  //string in the map and map stores unique keys
}

回答12:

One thing you need to consider is the small number of data elements you are using. It is possible that it will be faster to use just the vector. There is some overhead in the map that can cause it to be more expensive to do lookups in small data sets than the simpler vector. So, if you know that you will always be using around the same number of elements, do some benchmarking and see if the performance of the map and vector is what you really think it is. You may find the lookup in a vector with only 50 elements is near the same as the map.

回答13:

Use boost::multi_index with map and list indices.

回答14:

A map of pair (str,int) and static int that increments on insert calls indexes pairs of data. Put in a struct that can return the static int val with an index () member perhaps?