google_hash 0.6.2 → 0.7.0

data/ext/{sparsehash-1.5.2 → sparsehash-1.8.1}/src/google/sparse_hash_set

@@ -40,10 +40,15 @@
 // because all iterators for sets are const (you obviously can't
 // change the key, and for sets there is no value).
 //
-// We adhere mostly to the STL semantics for hash-set.  One important
-// exception is that insert() invalidates iterators entirely.  On the
-// plus side, though, delete() doesn't invalidate iterators at all, or
-// even change the ordering of elements.
+// We adhere mostly to the STL semantics for hash-map.  One important
+// exception is that insert() may invalidate iterators entirely -- STL
+// semantics are that insert() may reorder iterators, but they all
+// still refer to something valid in the hashtable.  Not so for us.
+// Likewise, insert() may invalidate pointers into the hashtable.
+// (Whether insert invalidates iterators and pointers depends on
+// whether it results in a hashtable resize).  On the plus side,
+// delete() doesn't invalidate iterators or pointers at all, or even
+// change the ordering of elements.
 //
 // Here are a few "power user" tips:
 //
@@ -62,19 +67,20 @@
 //         Setting the minimum load factor to 0.0 guarantees that
 //         the hash table will never shrink.
 //
-// Guide to what kind of hash_set to use:
-//   (1) dense_hash_set: fastest, uses the most memory
+// Roughly speaking:
+//   (1) dense_hash_set: fastest, uses the most memory unless entries are small
 //   (2) sparse_hash_set: slowest, uses the least memory
-//   (3) hash_set /unordered_set (STL): in the middle
+//   (3) hash_set / unordered_set (STL): in the middle
+//
 // Typically I use sparse_hash_set when I care about space and/or when
 // I need to save the hashtable on disk.  I use hash_set otherwise.  I
 // don't personally use dense_hash_set ever; some people use it for
 // small sets with lots of lookups.
 //
-// - dense_hash_set has, typically, a factor of 2 memory overhead (if your
-//   data takes up X bytes, the hash_set uses X more bytes in overhead).
-// - sparse_hash_set has about 2 bits overhead per entry.
-// - sparse_hash_map can be 3-7 times slower than the others for lookup and,
+// - dense_hash_set has, typically, about 78% memory overhead (if your
+//   data takes up X bytes, the hash_set uses .78X more bytes in overhead).
+// - sparse_hash_set has about 4 bits overhead per entry.
+// - sparse_hash_set can be 3-7 times slower than the others for lookup and,
 //   especially, inserts.  See time_hash_map.cc for details.
 //
 // See /usr/(local/)?doc/sparsehash-*/sparse_hash_set.html
@@ -90,6 +96,7 @@
 #include <memory>                    // for alloc<>
 #include <utility>                   // for pair<>
 #include HASH_FUN_H                  // defined in config.h
+#include <google/sparsehash/libc_allocator_with_realloc.h>
 #include <google/sparsehash/sparsehashtable.h>
 
 _START_GOOGLE_NAMESPACE_
@@ -99,7 +106,7 @@ using STL_NAMESPACE::pair;
 template <class Value,
           class HashFcn = SPARSEHASH_HASH<Value>,  // defined in sparseconfig.h
           class EqualKey = STL_NAMESPACE::equal_to<Value>,
-          class Alloc = STL_NAMESPACE::allocator<Value> >
+          class Alloc = libc_allocator_with_realloc<Value> >
 class sparse_hash_set {
  private:
   // Apparently identity is not stl-standard, so we define our own
@@ -113,9 +120,8 @@ class sparse_hash_set {
     }
   };
 
-  // The actual data
-  typedef sparse_hashtable<Value, Value, HashFcn,
-                           Identity, SetKey, EqualKey, Alloc> ht;
+  typedef sparse_hashtable<Value, Value, HashFcn, Identity, SetKey,
+                           EqualKey, Alloc> ht;
   ht rep;
 
  public:
@@ -148,7 +154,7 @@ class sparse_hash_set {
 
 
   // Accessor functions
-  // TODO(csilvers): implement Alloc get_allocator() const;
+  allocator_type get_allocator() const    { return rep.get_allocator(); }
   hasher hash_funct() const               { return rep.hash_funct(); }
   hasher hash_function() const            { return hash_funct(); }  // tr1 name
   key_equal key_eq() const                { return rep.key_eq(); }
@@ -157,15 +163,18 @@ class sparse_hash_set {
   // Constructors
   explicit sparse_hash_set(size_type expected_max_items_in_table = 0,
                            const hasher& hf = hasher(),
-                           const key_equal& eql = key_equal())
-    : rep(expected_max_items_in_table, hf, eql) { }
+                           const key_equal& eql = key_equal(),
+                           const allocator_type& alloc = allocator_type())
+      : rep(expected_max_items_in_table, hf, eql, Identity(), SetKey(), alloc) {
+  }
 
   template <class InputIterator>
   sparse_hash_set(InputIterator f, InputIterator l,
                   size_type expected_max_items_in_table = 0,
                   const hasher& hf = hasher(),
-                  const key_equal& eql = key_equal())
-      : rep(expected_max_items_in_table, hf, eql) {
+                  const key_equal& eql = key_equal(),
+                  const allocator_type& alloc = allocator_type())
+      : rep(expected_max_items_in_table, hf, eql, Identity(), SetKey(), alloc) {
     rep.insert(f, l);
   }
   // We use the default copy constructor
@@ -212,7 +221,7 @@ class sparse_hash_set {
   }
   // Deprecated; use min_load_factor() or max_load_factor() instead.
   void set_resizing_parameters(float shrink, float grow) {
-    return rep.set_resizing_parameters(shrink, grow);
+    rep.set_resizing_parameters(shrink, grow);
   }
 
   void resize(size_type hint)         { rep.resize(hint); }
@@ -245,6 +254,7 @@ class sparse_hash_set {
   // time goes on, or get rid of it entirely to be insert-only.
   void set_deleted_key(const key_type& key)   { rep.set_deleted_key(key); }
   void clear_deleted_key()                    { rep.clear_deleted_key(); }
+  key_type deleted_key() const                { return rep.deleted_key(); }
 
   // These are standard
   size_type erase(const key_type& key)               { return rep.erase(key); }

data/ext/{sparsehash-1.5.2 → sparsehash-1.8.1}/src/google/sparsehash/densehashtable.h

@@ -57,19 +57,19 @@
 // <google/dense_hash_map> or <google/dense_hash_set> instead.
 
 // You can change the following below:
-// HT_OCCUPANCY_FLT      -- how full before we double size
-// HT_EMPTY_FLT          -- how empty before we halve size
+// HT_OCCUPANCY_PCT      -- how full before we double size
+// HT_EMPTY_PCT          -- how empty before we halve size
 // HT_MIN_BUCKETS        -- default smallest bucket size
 //
-// You can also change enlarge_resize_percent (which defaults to
-// HT_OCCUPANCY_FLT), and shrink_resize_percent (which defaults to
-// HT_EMPTY_FLT) with set_resizing_parameters().
+// You can also change enlarge_factor (which defaults to
+// HT_OCCUPANCY_PCT), and shrink_factor (which defaults to
+// HT_EMPTY_PCT) with set_resizing_parameters().
 //
 // How to decide what values to use?
-// shrink_resize_percent's default of .4 * OCCUPANCY_FLT, is probably good.
+// shrink_factor's default of .4 * OCCUPANCY_PCT, is probably good.
 // HT_MIN_BUCKETS is probably unnecessary since you can specify
 // (indirectly) the starting number of buckets at construct-time.
-// For enlarge_resize_percent, you can use this chart to try to trade-off
+// For enlarge_factor, you can use this chart to try to trade-off
 // expected lookup time to the space taken up.  By default, this
 // code uses quadratic probing, though you can change it to linear
 // via _JUMP below if you really want to.
@@ -79,7 +79,7 @@
 // Quadratic collision resolution   1 - ln(1-L) - L/2    1/(1-L) - L - ln(1-L)
 // Linear collision resolution     [1+1/(1-L)]/2         [1+1/(1-L)2]/2
 //
-// -- enlarge_resize_percent --         0.10  0.50  0.60  0.75  0.80  0.90  0.99
+// -- enlarge_factor --           0.10  0.50  0.60  0.75  0.80  0.90  0.99
 // QUADRATIC COLLISION RES.
 //    probes/successful lookup    1.05  1.44  1.62  2.01  2.21  2.85  5.11
 //    probes/unsuccessful lookup  1.11  2.19  2.82  4.64  5.81  11.4  103.6
@@ -93,19 +93,23 @@
 // The probing method
 // Linear probing
 // #define JUMP_(key, num_probes)    ( 1 )
-// Quadratic-ish probing
+// Quadratic probing
 #define JUMP_(key, num_probes)    ( num_probes )
 
 
 #include <google/sparsehash/sparseconfig.h>
-#include <assert.h>
 #include <stdio.h>
+#include <assert.h>
 #include <stdlib.h>             // for abort()
 #include <algorithm>            // For swap(), eg
+#include <stdexcept>            // For length_error
 #include <iostream>             // For cerr
 #include <memory>               // For uninitialized_fill, uninitialized_copy
 #include <utility>              // for pair<>
 #include <iterator>             // for facts about iterator tags
+#include <limits>               // for numeric_limits<>
+#include <google/sparsehash/libc_allocator_with_realloc.h>
+#include <google/sparsehash/hashtable-common.h>
 #include <google/type_traits.h> // for true_type, integral_constant, etc.
 
 _START_GOOGLE_NAMESPACE_
@@ -125,7 +129,7 @@ using STL_NAMESPACE::pair;
 //         with key == deleted_key or key == empty_key.
 // EqualKey: Given two Keys, says whether they are the same (that is,
 //           if they are both associated with the same Value).
-// Alloc: STL allocator to use to allocate memory.  Currently ignored.
+// Alloc: STL allocator to use to allocate memory.
 
 template <class Value, class Key, class HashFcn,
           class ExtractKey, class SetKey, class EqualKey, class Alloc>
@@ -140,16 +144,19 @@ struct dense_hashtable_const_iterator;
 // We're just an array, but we need to skip over empty and deleted elements
 template <class V, class K, class HF, class ExK, class SetK, class EqK, class A>
 struct dense_hashtable_iterator {
+ private:
+  typedef typename A::template rebind<V>::other value_alloc_type;
+
  public:
   typedef dense_hashtable_iterator<V,K,HF,ExK,SetK,EqK,A>       iterator;
   typedef dense_hashtable_const_iterator<V,K,HF,ExK,SetK,EqK,A> const_iterator;
 
   typedef STL_NAMESPACE::forward_iterator_tag iterator_category;
   typedef V value_type;
-  typedef ptrdiff_t difference_type;
-  typedef size_t size_type;
-  typedef V& reference;                // Value
-  typedef V* pointer;
+  typedef typename value_alloc_type::difference_type difference_type;
+  typedef typename value_alloc_type::size_type size_type;
+  typedef typename value_alloc_type::reference reference;
+  typedef typename value_alloc_type::pointer pointer;
 
   // "Real" constructor and default constructor
   dense_hashtable_iterator(const dense_hashtable<V,K,HF,ExK,SetK,EqK,A> *h,
@@ -190,16 +197,19 @@ struct dense_hashtable_iterator {
 // Now do it all again, but with const-ness!
 template <class V, class K, class HF, class ExK, class SetK, class EqK, class A>
 struct dense_hashtable_const_iterator {
+ private:
+  typedef typename A::template rebind<V>::other value_alloc_type;
+
  public:
   typedef dense_hashtable_iterator<V,K,HF,ExK,SetK,EqK,A>       iterator;
   typedef dense_hashtable_const_iterator<V,K,HF,ExK,SetK,EqK,A> const_iterator;
 
   typedef STL_NAMESPACE::forward_iterator_tag iterator_category;
   typedef V value_type;
-  typedef ptrdiff_t difference_type;
-  typedef size_t size_type;
-  typedef const V& reference;                // Value
-  typedef const V* pointer;
+  typedef typename value_alloc_type::difference_type difference_type;
+  typedef typename value_alloc_type::size_type size_type;
+  typedef typename value_alloc_type::const_reference reference;
+  typedef typename value_alloc_type::const_pointer pointer;
 
   // "Real" constructor and default constructor
   dense_hashtable_const_iterator(
@@ -243,18 +253,22 @@ struct dense_hashtable_const_iterator {
 template <class Value, class Key, class HashFcn,
           class ExtractKey, class SetKey, class EqualKey, class Alloc>
 class dense_hashtable {
+ private:
+  typedef typename Alloc::template rebind<Value>::other value_alloc_type;
+
  public:
   typedef Key key_type;
   typedef Value value_type;
   typedef HashFcn hasher;
   typedef EqualKey key_equal;
-
-  typedef size_t            size_type;
-  typedef ptrdiff_t         difference_type;
-  typedef value_type*       pointer;
-  typedef const value_type* const_pointer;
-  typedef value_type&       reference;
-  typedef const value_type& const_reference;
+  typedef Alloc allocator_type;
+
+  typedef typename value_alloc_type::size_type size_type;
+  typedef typename value_alloc_type::difference_type difference_type;
+  typedef typename value_alloc_type::reference reference;
+  typedef typename value_alloc_type::const_reference const_reference;
+  typedef typename value_alloc_type::pointer pointer;
+  typedef typename value_alloc_type::const_pointer const_pointer;
   typedef dense_hashtable_iterator<Value, Key, HashFcn,
                                    ExtractKey, SetKey, EqualKey, Alloc>
   iterator;
@@ -270,24 +284,23 @@ class dense_hashtable {
   // How full we let the table get before we resize, by default.
   // Knuth says .8 is good -- higher causes us to probe too much,
   // though it saves memory.
-  static const float HT_OCCUPANCY_FLT; // = 0.5;
+  static const int HT_OCCUPANCY_PCT; // = 50 (out of 100)
 
   // How empty we let the table get before we resize lower, by default.
   // (0.0 means never resize lower.)
-  // It should be less than OCCUPANCY_FLT / 2 or we thrash resizing
-  static const float HT_EMPTY_FLT; // = 0.4 * HT_OCCUPANCY_FLT
+  // It should be less than OCCUPANCY_PCT / 2 or we thrash resizing
+  static const int HT_EMPTY_PCT; // = 0.4 * HT_OCCUPANCY_PCT;
 
   // Minimum size we're willing to let hashtables be.
   // Must be a power of two, and at least 4.
   // Note, however, that for a given hashtable, the initial size is a
   // function of the first constructor arg, and may be >HT_MIN_BUCKETS.
-  static const size_t HT_MIN_BUCKETS = 4;
+  static const size_type HT_MIN_BUCKETS = 4;
 
   // By default, if you don't specify a hashtable size at
   // construction-time, we use this size.  Must be a power of two, and
   // at least HT_MIN_BUCKETS.
-  static const size_t HT_DEFAULT_STARTING_BUCKETS = 32;
-
+  static const size_type HT_DEFAULT_STARTING_BUCKETS = 32;
 
   // ITERATOR FUNCTIONS
   iterator begin()             { return iterator(this, table,
@@ -300,10 +313,7 @@ class dense_hashtable {
                                                        table+num_buckets,true);}
 
   // These come from tr1 unordered_map.  They iterate over 'bucket' n.
-  // For sparsehashtable, we could consider each 'group' to be a bucket,
-  // I guess, but I don't really see the point.  We'll just consider
-  // bucket n to be the n-th element of the sparsetable, if it's occupied,
-  // or some empty element, otherwise.
+  // We'll just consider bucket n to be the n-th element of the table.
   local_iterator begin(size_type i) {
     return local_iterator(this, table + i, table + i+1, false);
   }
@@ -324,16 +334,20 @@ class dense_hashtable {
   }
 
   // ACCESSOR FUNCTIONS for the things we templatize on, basically
-  hasher hash_funct() const { return hash; }
-  key_equal key_eq() const  { return equals; }
+  hasher hash_funct() const               { return settings; }
+  key_equal key_eq() const                { return key_info; }
+  allocator_type get_allocator() const    { return allocator; }
+
+  // Accessor function for statistics gathering.
+  int num_table_copies() const { return settings.num_ht_copies(); }
 
  private:
   // Annoyingly, we can't copy values around, because they might have
   // const components (they're probably pair<const X, Y>).  We use
   // explicit destructor invocation and placement new to get around
   // this.  Arg.
-  void set_value(value_type* dst, const value_type& src) {
-    dst->~value_type();
+  void set_value(pointer dst, const_reference src) {
+    dst->~value_type();   // delete the old value, if any
     new(dst) value_type(src);
   }
 
@@ -357,50 +371,77 @@ class dense_hashtable {
     assert(num_deleted == 0);
   }
 
+  bool test_deleted_key(const key_type& key) const {
+    // The num_deleted test is crucial for read(): after read(), the ht values
+    // are garbage, and we don't want to think some of them are deleted.
+    // Invariant: !use_deleted implies num_deleted is 0.
+    assert(settings.use_deleted() || num_deleted == 0);
+    return num_deleted > 0 && equals(key_info.delkey, key);
+  }
+
  public:
   void set_deleted_key(const key_type &key) {
     // the empty indicator (if specified) and the deleted indicator
     // must be different
-    assert(!use_empty || !equals(key, get_key(emptyval)));
+    assert((!settings.use_empty() || !equals(key, get_key(emptyval)))
+           && "Passed the empty-key to set_deleted_key");
     // It's only safe to change what "deleted" means if we purge deleted guys
     squash_deleted();
-    use_deleted = true;
-    delkey = key;
+    settings.set_use_deleted(true);
+    key_info.delkey = key;
   }
   void clear_deleted_key() {
     squash_deleted();
-    use_deleted = false;
+    settings.set_use_deleted(false);
+  }
+  key_type deleted_key() const {
+    assert(settings.use_deleted()
+           && "Must set deleted key before calling deleted_key");
+    return key_info.delkey;
   }
 
   // These are public so the iterators can use them
   // True if the item at position bucknum is "deleted" marker
   bool test_deleted(size_type bucknum) const {
-    // The num_deleted test is crucial for read(): after read(), the ht values
-    // are garbage, and we don't want to think some of them are deleted.
-    return (use_deleted && num_deleted > 0 &&
-            equals(delkey, get_key(table[bucknum])));
+    return test_deleted_key(get_key(table[bucknum]));
   }
   bool test_deleted(const iterator &it) const {
-    return (use_deleted && num_deleted > 0 &&
-            equals(delkey, get_key(*it)));
+    return test_deleted_key(get_key(*it));
   }
   bool test_deleted(const const_iterator &it) const {
-    return (use_deleted && num_deleted > 0 &&
-            equals(delkey, get_key(*it)));
+    return test_deleted_key(get_key(*it));
+  }
+
+ private:
+  // Set it so test_deleted is true.  true if object didn't used to be deleted.
+  bool set_deleted(iterator &it) {
+    assert(settings.use_deleted());
+    bool retval = !test_deleted(it);
+    // &* converts from iterator to value-type.
+    set_key(&(*it), key_info.delkey);
+    return retval;
   }
-  // Set it so test_deleted is true.  true if object didn't used to be deleted
-  // See below (at erase()) to explain why we allow const_iterators
+  // Set it so test_deleted is false.  true if object used to be deleted.
+  bool clear_deleted(iterator &it) {
+    assert(settings.use_deleted());
+    // Happens automatically when we assign something else in its place.
+    return test_deleted(it);
+  }
+
+  // We also allow to set/clear the deleted bit on a const iterator.
+  // We allow a const_iterator for the same reason you can delete a
+  // const pointer: it's convenient, and semantically you can't use
+  // 'it' after it's been deleted anyway, so its const-ness doesn't
+  // really matter.
   bool set_deleted(const_iterator &it) {
-    assert(use_deleted);             // bad if set_deleted_key() wasn't called
+    assert(settings.use_deleted());
     bool retval = !test_deleted(it);
-    // &* converts from iterator to value-type
-    set_key(const_cast<value_type*>(&(*it)), delkey);
+    set_key(const_cast<pointer>(&(*it)), key_info.delkey);
     return retval;
   }
-  // Set it so test_deleted is false.  true if object used to be deleted
+  // Set it so test_deleted is false.  true if object used to be deleted.
   bool clear_deleted(const_iterator &it) {
-    assert(use_deleted);             // bad if set_deleted_key() wasn't called
-    // happens automatically when we assign something else in its place
+    assert(settings.use_deleted());
     return test_deleted(it);
   }
 
@@ -414,58 +455,52 @@ class dense_hashtable {
   // These are public so the iterators can use them
   // True if the item at position bucknum is "empty" marker
   bool test_empty(size_type bucknum) const {
-    assert(use_empty);              // we always need to know what's empty!
+    assert(settings.use_empty());  // we always need to know what's empty!
     return equals(get_key(emptyval), get_key(table[bucknum]));
   }
   bool test_empty(const iterator &it) const {
-    assert(use_empty);              // we always need to know what's empty!
+    assert(settings.use_empty());  // we always need to know what's empty!
     return equals(get_key(emptyval), get_key(*it));
   }
   bool test_empty(const const_iterator &it) const {
-    assert(use_empty);              // we always need to know what's empty!
+    assert(settings.use_empty());  // we always need to know what's empty!
     return equals(get_key(emptyval), get_key(*it));
   }
 
  private:
-  // You can either set a range empty or an individual element
-  void set_empty(size_type bucknum) {
-    assert(use_empty);
-    set_value(&table[bucknum], emptyval);
-  }
-  void fill_range_with_empty(value_type* table_start, value_type* table_end) {
-    // Like set_empty(range), but doesn't destroy previous contents
+  void fill_range_with_empty(pointer table_start, pointer table_end) {
     STL_NAMESPACE::uninitialized_fill(table_start, table_end, emptyval);
   }
-  void set_empty(size_type buckstart, size_type buckend) {
-    assert(use_empty);
-    destroy_buckets(buckstart, buckend);
-    fill_range_with_empty(table + buckstart, table + buckend);
-  }
 
  public:
   // TODO(csilvers): change all callers of this to pass in a key instead,
   //                 and take a const key_type instead of const value_type.
-  void set_empty_key(const value_type &val) {
+  void set_empty_key(const_reference val) {
     // Once you set the empty key, you can't change it
-    assert(!use_empty);
+    assert(!settings.use_empty() && "Calling set_empty_key multiple times");
     // The deleted indicator (if specified) and the empty indicator
     // must be different.
-    assert(!use_deleted || !equals(get_key(val), delkey));
-    use_empty = true;
+    assert((!settings.use_deleted() || !equals(get_key(val), key_info.delkey))
+           && "Setting the empty key the same as the deleted key");
+    settings.set_use_empty(true);
     set_value(&emptyval, val);
 
     assert(!table);                  // must set before first use
     // num_buckets was set in constructor even though table was NULL
-    table = (value_type *) malloc(num_buckets * sizeof(*table));
+    table = allocator.allocate(num_buckets);
     assert(table);
     fill_range_with_empty(table, table + num_buckets);
   }
+  // TODO(sjackman): return a key_type rather than a value_type
+  value_type empty_key() const {
+    assert(settings.use_empty());
+    return emptyval;
+  }
 
   // FUNCTIONS CONCERNING SIZE
  public:
   size_type size() const      { return num_elements - num_deleted; }
-  // Buckets are always a power of 2
-  size_type max_size() const  { return (size_type(-1) >> 1U) + 1; }
+  size_type max_size() const  { return allocator.max_size(); }
   bool empty() const          { return size() == 0; }
   size_type bucket_count() const      { return num_buckets; }
   size_type max_bucket_count() const  { return max_size(); }
@@ -476,54 +511,56 @@ class dense_hashtable {
     return begin(i) == end(i) ? 0 : 1;
   }
 
-
-
  private:
   // Because of the above, size_type(-1) is never legal; use it for errors
   static const size_type ILLEGAL_BUCKET = size_type(-1);
 
- private:
-  // This is the smallest size a hashtable can be without being too crowded
-  // If you like, you can give a min #buckets as well as a min #elts
-  size_type min_size(size_type num_elts, size_type min_buckets_wanted) {
-    size_type sz = HT_MIN_BUCKETS;             // min buckets allowed
-    while ( sz < min_buckets_wanted || num_elts >= sz * enlarge_resize_percent )
-      sz *= 2;
-    return sz;
-  }
-
-  // Used after a string of deletes
-  void maybe_shrink() {
+  // Used after a string of deletes.  Returns true if we actually shrunk.
+  // TODO(csilvers): take a delta so we can take into account inserts
+  // done after shrinking.  Maybe make part of the Settings class?
+  bool maybe_shrink() {
     assert(num_elements >= num_deleted);
     assert((bucket_count() & (bucket_count()-1)) == 0); // is a power of two
     assert(bucket_count() >= HT_MIN_BUCKETS);
+    bool retval = false;
 
     // If you construct a hashtable with < HT_DEFAULT_STARTING_BUCKETS,
     // we'll never shrink until you get relatively big, and we'll never
     // shrink below HT_DEFAULT_STARTING_BUCKETS.  Otherwise, something
     // like "dense_hash_set<int> x; x.insert(4); x.erase(4);" will
     // shrink us down to HT_MIN_BUCKETS buckets, which is too small.
-    if (shrink_threshold > 0 &&
-        (num_elements-num_deleted) < shrink_threshold &&
-        bucket_count() > HT_DEFAULT_STARTING_BUCKETS ) {
+    const size_type num_remain = num_elements - num_deleted;
+    const size_type shrink_threshold = settings.shrink_threshold();
+    if (shrink_threshold > 0 && num_remain < shrink_threshold &&
+        bucket_count() > HT_DEFAULT_STARTING_BUCKETS) {
+      const float shrink_factor = settings.shrink_factor();
       size_type sz = bucket_count() / 2;    // find how much we should shrink
-      while ( sz > HT_DEFAULT_STARTING_BUCKETS &&
-              (num_elements - num_deleted) < sz * shrink_resize_percent )
+      while (sz > HT_DEFAULT_STARTING_BUCKETS &&
+             num_remain < sz * shrink_factor) {
         sz /= 2;                            // stay a power of 2
+      }
       dense_hashtable tmp(*this, sz);       // Do the actual resizing
       swap(tmp);                            // now we are tmp
+      retval = true;
     }
-    consider_shrink = false;                // because we just considered it
+    settings.set_consider_shrink(false);    // because we just considered it
+    return retval;
   }
 
   // We'll let you resize a hashtable -- though this makes us copy all!
   // When you resize, you say, "make it big enough for this many more elements"
-  void resize_delta(size_type delta) {
-    if ( consider_shrink )                   // see if lots of deletes happened
-      maybe_shrink();
-    if ( bucket_count() > HT_MIN_BUCKETS &&
-         (num_elements + delta) <= enlarge_threshold )
-      return;                                // we're ok as we are
+  // Returns true if we actually resized, false if size was already ok.
+  bool resize_delta(size_type delta) {
+    bool did_resize = false;
+    if ( settings.consider_shrink() ) {  // see if lots of deletes happened
+      if ( maybe_shrink() )
+        did_resize = true;
+    }
+    if (num_elements >= (STL_NAMESPACE::numeric_limits<size_type>::max)() - delta)
+      throw std::length_error("resize overflow");
+    if ( bucket_count() >= HT_MIN_BUCKETS &&
+         (num_elements + delta) <= settings.enlarge_threshold() )
+      return did_resize;                          // we're ok as we are
 
     // Sometimes, we need to resize just to get rid of all the
     // "deleted" buckets that are clogging up the hashtable.  So when
@@ -531,56 +568,48 @@ class dense_hashtable {
     // are currently taking up room).  But later, when we decide what
     // size to resize to, *don't* count deleted buckets, since they
     // get discarded during the resize.
-    const size_type needed_size = min_size(num_elements + delta, 0);
-    if ( needed_size > bucket_count() ) {      // we don't have enough buckets
-      const size_type resize_to = min_size(num_elements - num_deleted + delta,
-                                           0);
-      dense_hashtable tmp(*this, resize_to);
-      swap(tmp);                             // now we are tmp
+    const size_type needed_size = settings.min_buckets(num_elements + delta, 0);
+    if ( needed_size <= bucket_count() )      // we have enough buckets
+      return did_resize;
+
+    size_type resize_to =
+      settings.min_buckets(num_elements - num_deleted + delta, bucket_count());
+
+    if (resize_to < needed_size &&    // may double resize_to
+        resize_to < (STL_NAMESPACE::numeric_limits<size_type>::max)() / 2) {
+      // This situation means that we have enough deleted elements,
+      // that once we purge them, we won't actually have needed to
+      // grow.  But we may want to grow anyway: if we just purge one
+      // element, say, we'll have to grow anyway next time we
+      // insert.  Might as well grow now, since we're already going
+      // through the trouble of copying (in order to purge the
+      // deleted elements).
+      const size_type target =
+          static_cast<size_type>(settings.shrink_size(resize_to*2));
+      if (num_elements - num_deleted + delta >= target) {
+        // Good, we won't be below the shrink threshhold even if we double.
+        resize_to *= 2;
+      }
     }
+    dense_hashtable tmp(*this, resize_to);
+    swap(tmp);                             // now we are tmp
+    return true;
   }
 
-  // Increase number of buckets, assuming value_type has trivial copy
-  // constructor and destructor.  (Really, we want it to have "trivial
-  // move", because that's what realloc does.  But there's no way to
-  // capture that using type_traits, so we pretend that move(x, y) is
-  // equivalent to "x.~T(); new(x) T(y);" which is pretty much
-  // correct, if a bit conservative.)
-  void expand_array(size_t resize_to, true_type) {
-    table = (value_type *) realloc(table, resize_to * sizeof(value_type));
-    assert(table);
-    fill_range_with_empty(table + num_buckets, table + resize_to);
+  // We require table be not-NULL and empty before calling this.
+  void resize_table(size_type /*old_size*/, size_type new_size,
+                    true_type) {
+    table = allocator.realloc_or_die(table, new_size);
   }
 
-  // Increase number of buckets, without special assumptions about value_type.
-  // TODO(austern): make this exception safe. Handle exceptions from
-  // value_type's copy constructor.
-  void expand_array(size_t resize_to, false_type) {
-    value_type* new_table =
-      (value_type *) malloc(resize_to * sizeof(value_type));
-    assert(new_table);
-    STL_NAMESPACE::uninitialized_copy(table, table + num_buckets, new_table);
-    fill_range_with_empty(new_table + num_buckets, new_table + resize_to);
-    destroy_buckets(0, num_buckets);
-    free(table);
-    table = new_table;
+  void resize_table(size_type old_size, size_type new_size, false_type) {
+    allocator.deallocate(table, old_size);
+    table = allocator.allocate(new_size);
   }
 
   // Used to actually do the rehashing when we grow/shrink a hashtable
   void copy_from(const dense_hashtable &ht, size_type min_buckets_wanted) {
-    clear();            // clear table, set num_deleted to 0
-
-    // If we need to change the size of our table, do it now
-    const size_type resize_to = min_size(ht.size(), min_buckets_wanted);
-    if ( resize_to > bucket_count() ) { // we don't have enough buckets
-      typedef integral_constant<bool,
-          (has_trivial_copy<value_type>::value &&
-           has_trivial_destructor<value_type>::value)>
-          realloc_ok; // we pretend mv(x,y) == "x.~T(); new(x) T(y)"
-      expand_array(resize_to, realloc_ok());
-      num_buckets = resize_to;
-      reset_thresholds();
-    }
+    clear_to_size(settings.min_buckets(ht.size(), min_buckets_wanted));
 
     // We use a normal iterator to get non-deleted bcks from ht
     // We could use insert() here, but since we know there are
@@ -594,41 +623,38 @@ class dense_hashtable {
            !test_empty(bucknum);                               // not empty
            bucknum = (bucknum + JUMP_(key, num_probes)) & bucket_count_minus_one) {
         ++num_probes;
-        assert(num_probes < bucket_count()); // or else the hashtable is full
+        assert(num_probes < bucket_count()
+               && "Hashtable is full: an error in key_equal<> or hash<>");
       }
       set_value(&table[bucknum], *it);       // copies the value to here
       num_elements++;
     }
+    settings.inc_num_ht_copies();
   }
 
   // Required by the spec for hashed associative container
  public:
   // Though the docs say this should be num_buckets, I think it's much
-  // more useful as req_elements.  As a special feature, calling with
+  // more useful as num_elements.  As a special feature, calling with
   // req_elements==0 will cause us to shrink if we can, saving space.
   void resize(size_type req_elements) {       // resize to this or larger
-    if ( consider_shrink || req_elements == 0 )
+    if ( settings.consider_shrink() || req_elements == 0 )
       maybe_shrink();
     if ( req_elements > num_elements )
-      return resize_delta(req_elements - num_elements);
+      resize_delta(req_elements - num_elements);
   }
 
-  // Get and change the value of shrink_resize_percent and
-  // enlarge_resize_percent.  The description at the beginning of this
-  // file explains how to choose the values.  Setting the shrink
-  // parameter to 0.0 ensures that the table never shrinks.
+  // Get and change the value of shrink_factor and enlarge_factor.  The
+  // description at the beginning of this file explains how to choose
+  // the values.  Setting the shrink parameter to 0.0 ensures that the
+  // table never shrinks.
   void get_resizing_parameters(float* shrink, float* grow) const {
-    *shrink = shrink_resize_percent;
-    *grow = enlarge_resize_percent;
+    *shrink = settings.shrink_factor();
+    *grow = settings.enlarge_factor();
   }
   void set_resizing_parameters(float shrink, float grow) {
-    assert(shrink >= 0.0);
-    assert(grow <= 1.0);
-    if (shrink > grow/2.0f)
-      shrink = grow / 2.0f;     // otherwise we thrash hashtable size
-    shrink_resize_percent = shrink;
-    enlarge_resize_percent = grow;
-    reset_thresholds();
+    settings.set_resizing_parameters(shrink, grow);
+    settings.reset_thresholds(bucket_count());
   }
 
   // CONSTRUCTORS -- as required by the specs, we take a size,
@@ -639,105 +665,133 @@ class dense_hashtable {
                            const HashFcn& hf = HashFcn(),
                            const EqualKey& eql = EqualKey(),
                            const ExtractKey& ext = ExtractKey(),
-                           const SetKey& set = SetKey())
-    : hash(hf), equals(eql), get_key(ext), set_key(set), num_deleted(0),
-      use_deleted(false), use_empty(false),
-      delkey(), emptyval(), enlarge_resize_percent(HT_OCCUPANCY_FLT),
-      shrink_resize_percent(HT_EMPTY_FLT), table(NULL),
-      num_buckets(expected_max_items_in_table == 0
-                  ? HT_DEFAULT_STARTING_BUCKETS
-                  : min_size(expected_max_items_in_table, 0)),
-      num_elements(0) {
+                           const SetKey& set = SetKey(),
+                           const Alloc& alloc = Alloc())
+      : settings(hf),
+        key_info(ext, set, eql),
+        allocator(alloc),
+        num_deleted(0),
+        num_elements(0),
+        num_buckets(expected_max_items_in_table == 0
+                    ? HT_DEFAULT_STARTING_BUCKETS
+                    : settings.min_buckets(expected_max_items_in_table, 0)),
+        emptyval(),
+        table(NULL) {
     // table is NULL until emptyval is set.  However, we set num_buckets
     // here so we know how much space to allocate once emptyval is set
-    reset_thresholds();
+    settings.reset_thresholds(bucket_count());
   }
 
   // As a convenience for resize(), we allow an optional second argument
   // which lets you make this new hashtable a different size than ht
   dense_hashtable(const dense_hashtable& ht,
                   size_type min_buckets_wanted = HT_DEFAULT_STARTING_BUCKETS)
-    : hash(ht.hash), equals(ht.equals),
-      get_key(ht.get_key), set_key(ht.set_key), num_deleted(0),
-      use_deleted(ht.use_deleted), use_empty(ht.use_empty),
-      delkey(ht.delkey), emptyval(ht.emptyval),
-      enlarge_resize_percent(ht.enlarge_resize_percent),
-      shrink_resize_percent(ht.shrink_resize_percent), table(NULL),
-      num_buckets(0), num_elements(0) {
-    reset_thresholds();
+      : settings(ht.settings),
+        key_info(ht.key_info),
+        allocator(ht.allocator),
+        num_deleted(0),
+        num_elements(0),
+        num_buckets(0),
+        emptyval(ht.emptyval),
+        table(NULL) {
+    if (!ht.settings.use_empty()) {
+      // If use_empty isn't set, copy_from will crash, so we do our own copying.
+      assert(ht.empty());
+      num_buckets = settings.min_buckets(ht.size(), min_buckets_wanted);
+      settings.reset_thresholds(bucket_count());
+      return;
+    }
+    settings.reset_thresholds(bucket_count());
     copy_from(ht, min_buckets_wanted);   // copy_from() ignores deleted entries
   }
 
   dense_hashtable& operator= (const dense_hashtable& ht) {
     if (&ht == this)  return *this;        // don't copy onto ourselves
-    clear();
-    hash = ht.hash;
-    equals = ht.equals;
-    get_key = ht.get_key;
-    set_key = ht.set_key;
-    use_deleted = ht.use_deleted;
-    use_empty = ht.use_empty;
-    delkey = ht.delkey;
+    if (!ht.settings.use_empty()) {
+      assert(ht.empty());
+      dense_hashtable empty_table(ht);  // empty table with ht's thresholds
+      this->swap(empty_table);
+      return *this;
+    }
+    settings = ht.settings;
+    key_info = ht.key_info;
     set_value(&emptyval, ht.emptyval);
-    enlarge_resize_percent = ht.enlarge_resize_percent;
-    shrink_resize_percent = ht.shrink_resize_percent;
-    copy_from(ht, HT_MIN_BUCKETS);         // sets num_deleted to 0 too
+    // copy_from() calls clear and sets num_deleted to 0 too
+    copy_from(ht, HT_MIN_BUCKETS);
+    // we purposefully don't copy the allocator, which may not be copyable
     return *this;
   }
 
   ~dense_hashtable() {
     if (table) {
       destroy_buckets(0, num_buckets);
-      free(table);
+      allocator.deallocate(table, num_buckets);
     }
   }
 
   // Many STL algorithms use swap instead of copy constructors
   void swap(dense_hashtable& ht) {
-    STL_NAMESPACE::swap(hash, ht.hash);
-    STL_NAMESPACE::swap(equals, ht.equals);
-    STL_NAMESPACE::swap(get_key, ht.get_key);
-    STL_NAMESPACE::swap(set_key, ht.set_key);
+    STL_NAMESPACE::swap(settings, ht.settings);
+    STL_NAMESPACE::swap(key_info, ht.key_info);
     STL_NAMESPACE::swap(num_deleted, ht.num_deleted);
-    STL_NAMESPACE::swap(use_deleted, ht.use_deleted);
-    STL_NAMESPACE::swap(use_empty, ht.use_empty);
-    STL_NAMESPACE::swap(enlarge_resize_percent, ht.enlarge_resize_percent);
-    STL_NAMESPACE::swap(shrink_resize_percent, ht.shrink_resize_percent);
-    STL_NAMESPACE::swap(delkey, ht.delkey);
+    STL_NAMESPACE::swap(num_elements, ht.num_elements);
+    STL_NAMESPACE::swap(num_buckets, ht.num_buckets);
     { value_type tmp;     // for annoying reasons, swap() doesn't work
       set_value(&tmp, emptyval);
       set_value(&emptyval, ht.emptyval);
       set_value(&ht.emptyval, tmp);
     }
     STL_NAMESPACE::swap(table, ht.table);
-    STL_NAMESPACE::swap(num_buckets, ht.num_buckets);
-    STL_NAMESPACE::swap(num_elements, ht.num_elements);
-    reset_thresholds();
-    ht.reset_thresholds();
+    settings.reset_thresholds(bucket_count());  // this also resets consider_shrink
+    ht.settings.reset_thresholds(bucket_count());
+    // we purposefully don't swap the allocator, which may not be swap-able
   }
 
-  // It's always nice to be able to clear a table without deallocating it
-  void clear() {
-    if (table)
+ private:
+  void clear_to_size(size_type new_num_buckets) {
+    if (!table) {
+      table = allocator.allocate(new_num_buckets);
+    } else {
       destroy_buckets(0, num_buckets);
-    num_buckets = min_size(0,0);          // our new size
-    reset_thresholds();
-    table = (value_type *) realloc(table, num_buckets * sizeof(*table));
+      if (new_num_buckets != num_buckets) {   // resize, if necessary
+        typedef integral_constant<bool,
+            is_same<value_alloc_type,
+                    libc_allocator_with_realloc<value_type> >::value>
+            realloc_ok;
+        resize_table(num_buckets, new_num_buckets, realloc_ok());
+      }
+    }
     assert(table);
-    fill_range_with_empty(table, table + num_buckets);
+    fill_range_with_empty(table, table + new_num_buckets);
     num_elements = 0;
     num_deleted = 0;
+    num_buckets = new_num_buckets;          // our new size
+    settings.reset_thresholds(bucket_count());
+  }
+
+ public:
+  // It's always nice to be able to clear a table without deallocating it
+  void clear() {
+    // If the table is already empty, and the number of buckets is
+    // already as we desire, there's nothing to do.
+    const size_type new_num_buckets = settings.min_buckets(0, 0);
+    if (num_elements == 0 && new_num_buckets == num_buckets) {
+      return;
+    }
+    clear_to_size(new_num_buckets);
   }
 
   // Clear the table without resizing it.
   // Mimicks the stl_hashtable's behaviour when clear()-ing in that it
   // does not modify the bucket count
   void clear_no_resize() {
-    if (table) {
-      set_empty(0, num_buckets);
+    if (num_elements > 0) {
+      assert(table);
+      destroy_buckets(0, num_buckets);
+      fill_range_with_empty(table, table + num_buckets);
     }
     // don't consider to shrink before another erase()
-    reset_thresholds();
+    settings.reset_thresholds(bucket_count());
     num_elements = 0;
     num_deleted = 0;
   }
@@ -770,7 +824,8 @@ class dense_hashtable {
       }
       ++num_probes;                        // we're doing another probe
       bucknum = (bucknum + JUMP_(key, num_probes)) & bucket_count_minus_one;
-      assert(num_probes < bucket_count()); // don't probe too many times!
+      assert(num_probes < bucket_count()
+             && "Hashtable is full: an error in key_equal<> or hash<>");
     }
   }
 
@@ -829,36 +884,63 @@ class dense_hashtable {
 
   // INSERTION ROUTINES
  private:
+  // Private method used by insert_noresize and find_or_insert.
+  iterator insert_at(const_reference obj, size_type pos) {
+    if (size() >= max_size())
+      throw std::length_error("insert overflow");
+    if ( test_deleted(pos) ) {      // just replace if it's been del.
+      // shrug: shouldn't need to be const.
+      const_iterator delpos(this, table + pos, table + num_buckets, false);
+      clear_deleted(delpos);
+      assert( num_deleted > 0);
+      --num_deleted;                // used to be, now it isn't
+    } else {
+      ++num_elements;               // replacing an empty bucket
+    }
+    set_value(&table[pos], obj);
+    return iterator(this, table + pos, table + num_buckets, false);
+  }
+
   // If you know *this is big enough to hold obj, use this routine
-  pair<iterator, bool> insert_noresize(const value_type& obj) {
+  pair<iterator, bool> insert_noresize(const_reference obj) {
     // First, double-check we're not inserting delkey or emptyval
-    assert(!use_empty || !equals(get_key(obj), get_key(emptyval)));
-    assert(!use_deleted || !equals(get_key(obj), delkey));
+    assert((!settings.use_empty() || !equals(get_key(obj), get_key(emptyval)))
+           && "Inserting the empty key");
+    assert((!settings.use_deleted() || !equals(get_key(obj), key_info.delkey))
+           && "Inserting the deleted key");
     const pair<size_type,size_type> pos = find_position(get_key(obj));
     if ( pos.first != ILLEGAL_BUCKET) {      // object was already there
       return pair<iterator,bool>(iterator(this, table + pos.first,
                                           table + num_buckets, false),
                                  false);          // false: we didn't insert
     } else {                                 // pos.second says where to put it
-      if ( test_deleted(pos.second) ) {      // just replace if it's been del.
-        const_iterator delpos(this, table + pos.second,              // shrug:
-                              table + num_buckets, false);// shouldn't need const
-        clear_deleted(delpos);
-        assert( num_deleted > 0);
-        --num_deleted;                       // used to be, now it isn't
-      } else {
-        ++num_elements;                      // replacing an empty bucket
-      }
-      set_value(&table[pos.second], obj);
-      return pair<iterator,bool>(iterator(this, table + pos.second,
-                                          table + num_buckets, false),
-                                 true);           // true: we did insert
+      return pair<iterator,bool>(insert_at(obj, pos.second), true);
     }
   }
 
+  // Specializations of insert(it, it) depending on the power of the iterator:
+  // (1) Iterator supports operator-, resize before inserting
+  template <class ForwardIterator>
+  void insert(ForwardIterator f, ForwardIterator l, STL_NAMESPACE::forward_iterator_tag) {
+    size_t dist = STL_NAMESPACE::distance(f, l);
+    if (dist >= (std::numeric_limits<size_type>::max)())
+      throw std::length_error("insert-range overflow");
+    resize_delta(static_cast<size_type>(dist));
+    for ( ; dist > 0; --dist, ++f) {
+      insert_noresize(*f);
+    }
+  }
+
+  // (2) Arbitrary iterator, can't tell how much to resize
+  template <class InputIterator>
+  void insert(InputIterator f, InputIterator l, STL_NAMESPACE::input_iterator_tag) {
+    for ( ; f != l; ++f)
+      insert(*f);
+  }
+
  public:
   // This is the normal insert routine, used by the outside world
-  pair<iterator, bool> insert(const value_type& obj) {
+  pair<iterator, bool> insert(const_reference obj) {
     resize_delta(1);                      // adding an object, grow if need be
     return insert_noresize(obj);
   }
@@ -870,59 +952,80 @@ class dense_hashtable {
     insert(f, l, typename STL_NAMESPACE::iterator_traits<InputIterator>::iterator_category());
   }
 
-  // Iterator supports operator-, resize before inserting
-  template <class ForwardIterator>
-  void insert(ForwardIterator f, ForwardIterator l,
-              STL_NAMESPACE::forward_iterator_tag) {
-    size_type n = STL_NAMESPACE::distance(f, l);   // TODO(csilvers): standard?
-    resize_delta(n);
-    for ( ; n > 0; --n, ++f)
-      insert_noresize(*f);
-  }
-
-  // Arbitrary iterator, can't tell how much to resize
-  template <class InputIterator>
-  void insert(InputIterator f, InputIterator l,
-              STL_NAMESPACE::input_iterator_tag) {
-    for ( ; f != l; ++f)
-      insert(*f);
+  // This is public only because dense_hash_map::operator[] uses it.
+  // It does the minimal amount of work to implement operator[].
+  template <class DataType>
+  DataType& find_or_insert(const key_type& key) {
+    // First, double-check we're not inserting emptykey or delkey
+    assert((!settings.use_empty() || !equals(key, get_key(emptyval)))
+           && "Inserting the empty key");
+    assert((!settings.use_deleted() || !equals(key, key_info.delkey))
+           && "Inserting the deleted key");
+    const pair<size_type,size_type> pos = find_position(key);
+    if ( pos.first != ILLEGAL_BUCKET) {  // object was already there
+      return table[pos.first].second;
+    } else if (resize_delta(1)) {        // needed to rehash to make room
+      // Since we resized, we can't use pos, so recalculate where to insert.
+      return insert_noresize(value_type(key, DataType())).first->second;
+    } else {                             // no need to rehash, insert right here
+      return insert_at(value_type(key, DataType()), pos.second)->second;
+    }
   }
 
-
   // DELETION ROUTINES
   size_type erase(const key_type& key) {
-    // First, double-check we're not trying to erase delkey or emptyval
-    assert(!use_empty || !equals(key, get_key(emptyval)));
-    assert(!use_deleted || !equals(key, delkey));
+    // First, double-check we're not trying to erase delkey or emptyval.
+    assert((!settings.use_empty() || !equals(key, get_key(emptyval)))
+           && "Erasing the empty key");
+    assert((!settings.use_deleted() || !equals(key, key_info.delkey))
+           && "Erasing the deleted key");
     const_iterator pos = find(key);   // shrug: shouldn't need to be const
     if ( pos != end() ) {
       assert(!test_deleted(pos));  // or find() shouldn't have returned it
       set_deleted(pos);
       ++num_deleted;
-      consider_shrink = true;      // will think about shrink after next insert
+      settings.set_consider_shrink(true); // will think about shrink after next insert
       return 1;                    // because we deleted one thing
     } else {
       return 0;                    // because we deleted nothing
     }
   }
 
-  // This is really evil: really it should be iterator, not const_iterator.
-  // But...the only reason keys are const is to allow lookup.
-  // Since that's a moot issue for deleted keys, we allow const_iterators
-  void erase(const_iterator pos) {
+  // We return the iterator past the deleted item.
+  void erase(iterator pos) {
     if ( pos == end() ) return;    // sanity check
     if ( set_deleted(pos) ) {      // true if object has been newly deleted
       ++num_deleted;
-      consider_shrink = true;      // will think about shrink after next insert
+      settings.set_consider_shrink(true); // will think about shrink after next insert
+    }
+  }
+
+  void erase(iterator f, iterator l) {
+    for ( ; f != l; ++f) {
+      if ( set_deleted(f)  )       // should always be true
+        ++num_deleted;
     }
+    settings.set_consider_shrink(true); // will think about shrink after next insert
   }
 
+  // We allow you to erase a const_iterator just like we allow you to
+  // erase an iterator.  This is in parallel to 'delete': you can delete
+  // a const pointer just like a non-const pointer.  The logic is that
+  // you can't use the object after it's erased anyway, so it doesn't matter
+  // if it's const or not.
+  void erase(const_iterator pos) {
+    if ( pos == end() ) return;    // sanity check
+    if ( set_deleted(pos) ) {      // true if object has been newly deleted
+      ++num_deleted;
+      settings.set_consider_shrink(true); // will think about shrink after next insert
+    }
+  }
   void erase(const_iterator f, const_iterator l) {
     for ( ; f != l; ++f) {
       if ( set_deleted(f)  )       // should always be true
         ++num_deleted;
     }
-    consider_shrink = true;        // will think about shrink after next insert
+    settings.set_consider_shrink(true);   // will think about shrink after next insert
   }
 
 
@@ -962,12 +1065,12 @@ class dense_hashtable {
 
   bool read_metadata(FILE *fp) {
     num_deleted = 0;            // since we got rid before writing
-    assert(use_empty);          // have to set this before calling us
-    if (table)  free(table);    // we'll make our own
+    assert(settings.use_empty() && "empty_key not set for read_metadata");
+    if (table)  allocator.deallocate(table, num_buckets);  // we'll make our own
     // TODO: read magic number
     // TODO: read num_buckets
-    reset_thresholds();
-    table = (value_type *) malloc(num_buckets * sizeof(*table));
+    settings.reset_thresholds(bucket_count());
+    table = allocator.allocate(num_buckets);
     assert(table);
     fill_range_with_empty(table, table + num_buckets);
     // TODO: read num_elements
@@ -1001,35 +1104,114 @@ class dense_hashtable {
   }
 
  private:
-  // The actual data
-  hasher hash;                      // required by hashed_associative_container
-  key_equal equals;
-  ExtractKey get_key;
-  SetKey set_key;
-  size_type num_deleted;        // how many occupied buckets are marked deleted
-  bool use_deleted;                          // false until delkey has been set
-  bool use_empty;                          // you must do this before you start
-  // TODO(csilvers): make a pointer, and get rid of use_deleted (benchmark!)
-  key_type delkey;                           // which key marks deleted entries
-  value_type emptyval;                        // which key marks unused entries
-  float enlarge_resize_percent;                       // how full before resize
-  float shrink_resize_percent;                       // how empty before resize
-  size_type shrink_threshold;            // num_buckets * shrink_resize_percent
-  size_type enlarge_threshold;          // num_buckets * enlarge_resize_percent
-  value_type *table;
-  size_type num_buckets;
-  size_type num_elements;
-  bool consider_shrink;   // true if we should try to shrink before next insert
+  template <class A>
+  class alloc_impl : public A {
+   public:
+    typedef typename A::pointer pointer;
+    typedef typename A::size_type size_type;
+
+    // Convert a normal allocator to one that has realloc_or_die()
+    alloc_impl(const A& a) : A(a) { }
+
+    // realloc_or_die should only be used when using the default
+    // allocator (libc_allocator_with_realloc).
+    pointer realloc_or_die(pointer ptr, size_type n) {
+      fprintf(stderr, "realloc_or_die is only supported for "
+                      "libc_allocator_with_realloc");
+      exit(1);
+      return NULL;
+    }
+  };
+
+  // A template specialization of alloc_impl for
+  // libc_allocator_with_realloc that can handle realloc_or_die.
+  template <class A>
+  class alloc_impl<libc_allocator_with_realloc<A> >
+      : public libc_allocator_with_realloc<A> {
+   public:
+    typedef typename libc_allocator_with_realloc<A>::pointer pointer;
+    typedef typename libc_allocator_with_realloc<A>::size_type size_type;
+
+    alloc_impl(const libc_allocator_with_realloc<A>& a)
+        : libc_allocator_with_realloc<A>(a) { }
+
+    pointer realloc_or_die(pointer ptr, size_type n) {
+      pointer retval = this->reallocate(ptr, n);
+      if (retval == NULL) {
+        // We really should use PRIuS here, but I don't want to have to add
+        // a whole new configure option, with concomitant macro namespace
+        // pollution, just to print this (unlikely) error message.  So I cast.
+        fprintf(stderr, "sparsehash: FATAL ERROR: failed to reallocate "
+                "%lu elements for ptr %p",
+                static_cast<unsigned long>(n), ptr);
+        exit(1);
+      }
+      return retval;
+    }
+  };
+
+  // Package functors with another class to eliminate memory needed for
+  // zero-size functors.  Since ExtractKey and hasher's operator() might
+  // have the same function signature, they must be packaged in
+  // different classes.
+  struct Settings :
+      sh_hashtable_settings<key_type, hasher, size_type, HT_MIN_BUCKETS> {
+    explicit Settings(const hasher& hf)
+        : sh_hashtable_settings<key_type, hasher, size_type, HT_MIN_BUCKETS>(
+            hf, HT_OCCUPANCY_PCT / 100.0f, HT_EMPTY_PCT / 100.0f) {}
+  };
+
+  // Packages ExtractKey and SetKey functors.
+  class KeyInfo : public ExtractKey, public SetKey, public key_equal {
+   public:
+    KeyInfo(const ExtractKey& ek, const SetKey& sk, const key_equal& eq)
+        : ExtractKey(ek),
+          SetKey(sk),
+          key_equal(eq) {
+    }
+    const key_type get_key(const_reference v) const {
+      return ExtractKey::operator()(v);
+    }
+    void set_key(pointer v, const key_type& k) const {
+      SetKey::operator()(v, k);
+    }
+    bool equals(const key_type& a, const key_type& b) const {
+      return key_equal::operator()(a, b);
+    }
+
+    // Which key marks deleted entries.
+    // TODO(csilvers): make a pointer, and get rid of use_deleted (benchmark!)
+    typename remove_const<key_type>::type delkey;
+  };
 
-  void reset_thresholds() {
-    enlarge_threshold = static_cast<size_type>(num_buckets
-                                               * enlarge_resize_percent);
-    shrink_threshold = static_cast<size_type>(num_buckets
-                                              * shrink_resize_percent);
-    consider_shrink = false;   // whatever caused us to reset already considered
+  // Utility functions to access the templated operators
+  size_type hash(const key_type& v) const {
+    return settings.hash(v);
+  }
+  bool equals(const key_type& a, const key_type& b) const {
+    return key_info.equals(a, b);
+  }
+  const key_type get_key(const_reference v) const {
+    return key_info.get_key(v);
   }
+  void set_key(pointer v, const key_type& k) const {
+    key_info.set_key(v, k);
+  }
+
+ private:
+  // Actual data
+  Settings settings;
+  KeyInfo key_info;
+  alloc_impl<value_alloc_type> allocator;
+
+  size_type num_deleted;  // how many occupied buckets are marked deleted
+  size_type num_elements;
+  size_type num_buckets;
+  value_type emptyval;    // which key marks unused entries
+  pointer table;
 };
 
+
 // We need a global swap as well
 template <class V, class K, class HF, class ExK, class SetK, class EqK, class A>
 inline void swap(dense_hashtable<V,K,HF,ExK,SetK,EqK,A> &x,
@@ -1041,7 +1223,7 @@ inline void swap(dense_hashtable<V,K,HF,ExK,SetK,EqK,A> &x,
 
 template <class V, class K, class HF, class ExK, class SetK, class EqK, class A>
 const typename dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::size_type
-dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::ILLEGAL_BUCKET;
+  dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::ILLEGAL_BUCKET;
 
 // How full we let the table get before we resize.  Knuth says .8 is
 // good -- higher causes us to probe too much, though saves memory.
@@ -1049,13 +1231,14 @@ dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::ILLEGAL_BUCKET;
 // more space (a trade-off densehashtable explicitly chooses to make).
 // Feel free to play around with different values, though.
 template <class V, class K, class HF, class ExK, class SetK, class EqK, class A>
-const float dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_OCCUPANCY_FLT = 0.5f;
+const int dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_OCCUPANCY_PCT = 50;
 
 // How empty we let the table get before we resize lower.
-// It should be less than OCCUPANCY_FLT / 2 or we thrash resizing
+// It should be less than OCCUPANCY_PCT / 2 or we thrash resizing
 template <class V, class K, class HF, class ExK, class SetK, class EqK, class A>
-const float dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_EMPTY_FLT
-    = 0.4f * dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_OCCUPANCY_FLT;
+const int dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_EMPTY_PCT
+  = static_cast<int>(0.4 *
+                     dense_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_OCCUPANCY_PCT);
 
 _END_GOOGLE_NAMESPACE_