google_hash 0.6.2 → 0.7.0

data/ext/sparsehash-1.8.1/src/google/sparsehash/hashtable-common.h

@@ -0,0 +1,178 @@
+// Copyright (c) 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// ---
+// Author: Giao Nguyen
+
+#ifndef UTIL_GTL_HASHTABLE_COMMON_H_
+#define UTIL_GTL_HASHTABLE_COMMON_H_
+
+#include <assert.h>
+
+// Settings contains parameters for growing and shrinking the table.
+// It also packages zero-size functor (ie. hasher).
+
+template<typename Key, typename HashFunc,
+         typename SizeType, int HT_MIN_BUCKETS>
+class sh_hashtable_settings : public HashFunc {
+ public:
+  typedef Key key_type;
+  typedef HashFunc hasher;
+  typedef SizeType size_type;
+
+ public:
+  sh_hashtable_settings(const hasher& hf,
+                        const float ht_occupancy_flt,
+                        const float ht_empty_flt)
+      : hasher(hf),
+        enlarge_threshold_(0),
+        shrink_threshold_(0),
+        consider_shrink_(false),
+        use_empty_(false),
+        use_deleted_(false),
+        num_ht_copies_(0) {
+    set_enlarge_factor(ht_occupancy_flt);
+    set_shrink_factor(ht_empty_flt);
+  }
+
+  size_type hash(const key_type& v) const {
+    return hasher::operator()(v);
+  }
+
+  float enlarge_factor() const {
+    return enlarge_factor_;
+  }
+  void set_enlarge_factor(float f) {
+    enlarge_factor_ = f;
+  }
+  float shrink_factor() const {
+    return shrink_factor_;
+  }
+  void set_shrink_factor(float f) {
+    shrink_factor_ = f;
+  }
+
+  size_type enlarge_threshold() const {
+    return enlarge_threshold_;
+  }
+  void set_enlarge_threshold(size_type t) {
+    enlarge_threshold_ = t;
+  }
+  size_type shrink_threshold() const {
+    return shrink_threshold_;
+  }
+  void set_shrink_threshold(size_type t) {
+    shrink_threshold_ = t;
+  }
+
+  size_type enlarge_size(size_type x) const {
+    return static_cast<size_type>(x * enlarge_factor_);
+  }
+  size_type shrink_size(size_type x) const {
+    return static_cast<size_type>(x * shrink_factor_);
+  }
+
+  bool consider_shrink() const {
+    return consider_shrink_;
+  }
+  void set_consider_shrink(bool t) {
+    consider_shrink_ = t;
+  }
+
+  bool use_empty() const {
+    return use_empty_;
+  }
+  void set_use_empty(bool t) {
+    use_empty_ = t;
+  }
+
+  bool use_deleted() const {
+    return use_deleted_;
+  }
+  void set_use_deleted(bool t) {
+    use_deleted_ = t;
+  }
+
+  size_type num_ht_copies() const {
+    return static_cast<size_type>(num_ht_copies_);
+  }
+  void inc_num_ht_copies() {
+    ++num_ht_copies_;
+  }
+
+  // Reset the enlarge and shrink thresholds
+  void reset_thresholds(int num_buckets) {
+    set_enlarge_threshold(enlarge_size(num_buckets));
+    set_shrink_threshold(shrink_size(num_buckets));
+    // whatever caused us to reset already considered
+    set_consider_shrink(false);
+  }
+
+  // Caller is resposible for calling reset_threshold right after
+  // set_resizing_parameters.
+  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
+    set_shrink_factor(shrink);
+    set_enlarge_factor(grow);
+  }
+
+  // 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_buckets(size_type num_elts, size_type min_buckets_wanted) {
+    float enlarge = enlarge_factor();
+    size_type sz = HT_MIN_BUCKETS;             // min buckets allowed
+    while ( sz < min_buckets_wanted ||
+            num_elts >= static_cast<size_type>(sz * enlarge) ) {
+      // This just prevents overflowing size_type, since sz can exceed
+      // max_size() here.
+      if (static_cast<size_type>(sz * 2) < sz) {
+        throw std::length_error("resize overflow");  // protect against overflow
+      }
+      sz *= 2;
+    }
+    return sz;
+  }
+
+ private:
+  size_type enlarge_threshold_;  // table.size() * enlarge_factor
+  size_type shrink_threshold_;   // table.size() * shrink_factor
+  float enlarge_factor_;         // how full before resize
+  float shrink_factor_;          // how empty before resize
+  // consider_shrink=true if we should try to shrink before next insert
+  bool consider_shrink_;
+  bool use_empty_;    // used only by densehashtable, not sparsehashtable
+  bool use_deleted_;  // false until delkey has been set
+  // num_ht_copies is a counter incremented every Copy/Move
+  unsigned int num_ht_copies_;
+};
+
+#endif  // UTIL_GTL_HASHTABLE_COMMON_H_

data/ext/sparsehash-1.8.1/src/google/sparsehash/libc_allocator_with_realloc.h

@@ -0,0 +1,121 @@
+// Copyright (c) 2010, Google Inc.
+// All rights reserved.
+// 
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+// 
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+// 
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// ---
+// Author: Guilin Chen
+
+#ifndef UTIL_GTL_LIBC_ALLOCATOR_WITH_REALLOC_H_
+#define UTIL_GTL_LIBC_ALLOCATOR_WITH_REALLOC_H_
+
+#include <google/sparsehash/sparseconfig.h>
+
+#include <stdlib.h>           // for malloc/realloc/free
+#include <stddef.h>           // for ptrdiff_t
+
+
+_START_GOOGLE_NAMESPACE_
+
+template<class T>
+class libc_allocator_with_realloc {
+ public:
+  typedef T value_type;
+  typedef size_t size_type;
+  typedef ptrdiff_t difference_type;
+
+  typedef T* pointer;
+  typedef const T* const_pointer;
+  typedef T& reference;
+  typedef const T& const_reference;
+
+  libc_allocator_with_realloc() {}
+  libc_allocator_with_realloc(const libc_allocator_with_realloc&) {}
+  ~libc_allocator_with_realloc() {}
+
+  pointer address(reference r) const  { return &r; }
+  const_pointer address(const_reference r) const  { return &r; }
+
+  pointer allocate(size_type n, const_pointer = 0) {
+    return static_cast<pointer>(malloc(n * sizeof(value_type)));
+  }
+  void deallocate(pointer p, size_type) {
+    free(p);
+  }
+  pointer reallocate(pointer p, size_type n) {
+    return static_cast<pointer>(realloc(p, n * sizeof(value_type)));
+  }
+
+  size_type max_size() const  {
+    return static_cast<size_type>(-1) / sizeof(value_type);
+  }
+
+  void construct(pointer p, const value_type& val) {
+    new(p) value_type(val);
+  }
+  void destroy(pointer p) { p->~value_type(); }
+
+  template <class U>
+  libc_allocator_with_realloc(const libc_allocator_with_realloc<U>&) {}
+
+  template<class U>
+  struct rebind {
+    typedef libc_allocator_with_realloc<U> other;
+  };
+};
+
+// libc_allocator_with_realloc<void> specialization.
+template<>
+class libc_allocator_with_realloc<void> {
+ public:
+  typedef void value_type;
+  typedef size_t size_type;
+  typedef ptrdiff_t difference_type;
+  typedef void* pointer;
+  typedef const void* const_pointer;
+
+  template<class U>
+  struct rebind {
+    typedef libc_allocator_with_realloc<U> other;
+  };
+};
+
+template<class T>
+inline bool operator==(const libc_allocator_with_realloc<T>&,
+                       const libc_allocator_with_realloc<T>&) {
+  return true;
+}
+
+template<class T>
+inline bool operator!=(const libc_allocator_with_realloc<T>&,
+                       const libc_allocator_with_realloc<T>&) {
+  return false;
+}
+
+_END_GOOGLE_NAMESPACE_
+
+#endif  // UTIL_GTL_LIBC_ALLOCATOR_WITH_REALLOC_H_

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

@@ -59,20 +59,20 @@
 // <google/sparse_hash_table> or <google/sparse_hash_set> instead.
 //
 // You can modify 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              -- smallest bucket size
 // HT_DEFAULT_STARTING_BUCKETS -- default bucket size at construct-time
 //
-// 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.
@@ -82,7 +82,7 @@
 // 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
@@ -103,21 +103,29 @@
 // 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 <algorithm>              // For swap(), eg
+#include <stdexcept>              // For length_error
 #include <iterator>               // for facts about iterator tags
+#include <limits>                 // for numeric_limits<>
 #include <utility>                // for pair<>
+#include <google/sparsehash/hashtable-common.h>
 #include <google/sparsetable>     // Since that's basically what we are
 
 _START_GOOGLE_NAMESPACE_
 
 using STL_NAMESPACE::pair;
 
+// The smaller this is, the faster lookup is (because the group bitmap is
+// smaller) and the faster insert is, because there's less to move.
+// On the other hand, there are more groups.  Since group::size_type is
+// a short, this number should be of the form 32*x + 16 to avoid waste.
+static const u_int16_t DEFAULT_GROUP_SIZE = 48;   // fits in 1.5 words
+
 // Hashtable class, used to implement the hashed associative containers
 // hash_set and hash_map.
 //
@@ -131,7 +139,7 @@ using STL_NAMESPACE::pair;
 //         with key == deleted_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>
@@ -147,17 +155,21 @@ struct sparse_hashtable_const_iterator;
 // that skips over deleted elements.
 template <class V, class K, class HF, class ExK, class SetK, class EqK, class A>
 struct sparse_hashtable_iterator {
+ private:
+  typedef typename A::template rebind<V>::other value_alloc_type;
+
  public:
   typedef sparse_hashtable_iterator<V,K,HF,ExK,SetK,EqK,A>       iterator;
   typedef sparse_hashtable_const_iterator<V,K,HF,ExK,SetK,EqK,A> const_iterator;
-  typedef typename sparsetable<V>::nonempty_iterator        st_iterator;
+  typedef typename sparsetable<V,DEFAULT_GROUP_SIZE,A>::nonempty_iterator
+      st_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
   sparse_hashtable_iterator(const sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> *h,
@@ -195,17 +207,21 @@ struct sparse_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 sparse_hashtable_const_iterator {
+ private:
+  typedef typename A::template rebind<V>::other value_alloc_type;
+
  public:
   typedef sparse_hashtable_iterator<V,K,HF,ExK,SetK,EqK,A>       iterator;
   typedef sparse_hashtable_const_iterator<V,K,HF,ExK,SetK,EqK,A> const_iterator;
-  typedef typename sparsetable<V>::const_nonempty_iterator  st_iterator;
+  typedef typename sparsetable<V,DEFAULT_GROUP_SIZE,A>::const_nonempty_iterator
+      st_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
   sparse_hashtable_const_iterator(const sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> *h,
@@ -246,16 +262,20 @@ struct sparse_hashtable_const_iterator {
 // And once again, but this time freeing up memory as we iterate
 template <class V, class K, class HF, class ExK, class SetK, class EqK, class A>
 struct sparse_hashtable_destructive_iterator {
+ private:
+  typedef typename A::template rebind<V>::other value_alloc_type;
+
  public:
   typedef sparse_hashtable_destructive_iterator<V,K,HF,ExK,SetK,EqK,A> iterator;
-  typedef typename sparsetable<V>::destructive_iterator     st_iterator;
+  typedef typename sparsetable<V,DEFAULT_GROUP_SIZE,A>::destructive_iterator
+      st_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
   sparse_hashtable_destructive_iterator(const
@@ -295,18 +315,22 @@ struct sparse_hashtable_destructive_iterator {
 template <class Value, class Key, class HashFcn,
           class ExtractKey, class SetKey, class EqualKey, class Alloc>
 class sparse_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 sparse_hashtable_iterator<Value, Key, HashFcn, ExtractKey,
                                     SetKey, EqualKey, Alloc>
   iterator;
@@ -326,22 +350,23 @@ class sparse_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; // = 0.8f;
+  static const int HT_OCCUPANCY_PCT; // = 80 (out of 100);
 
   // How empty we let the table get before we resize lower, by default.
-  // It should be less than OCCUPANCY_FLT / 2 or we thrash resizing
-  static const float HT_EMPTY_FLT; // = 0.4 * HT_OCCUPANCY_FLT;
+  // (0.0 means never resize lower.)
+  // 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 minimum size is
-  // determined by the first constructor arg, and may be >HT_MIN_BUCKETS.
-  static const size_t HT_MIN_BUCKETS = 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_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.nonempty_begin(),
@@ -399,8 +424,12 @@ class sparse_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 table.get_allocator(); }
+
+  // Accessor function for statistics gathering.
+  int num_table_copies() const { return settings.num_ht_copies(); }
 
  private:
   // We need to copy values when we set the special marker for deleted
@@ -408,7 +437,7 @@ class sparse_hashtable {
   // operator because value_type might not be assignable (it's often
   // 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) {
+  void set_value(pointer dst, const_reference src) {
     dst->~value_type();   // delete the old value, if any
     new(dst) value_type(src);
   }
@@ -419,7 +448,6 @@ class sparse_hashtable {
   // can't do a destructive copy, we make the typename private.
   enum MoveDontCopyT {MoveDontCopy, MoveDontGrow};
 
-
   // DELETE HELPER FUNCTIONS
   // This lets the user describe a key that will indicate deleted
   // table entries.  This key should be an "impossible" entry --
@@ -435,59 +463,88 @@ class sparse_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) {
     // 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 && table.test(bucknum) &&
-            equals(delkey, get_key(table.unsafe_get(bucknum))));
+    if (num_deleted == 0 || !table.test(bucknum)) return false;
+    return test_deleted_key(get_key(table.unsafe_get(bucknum)));
   }
   bool test_deleted(const iterator &it) const {
-    return (use_deleted && num_deleted > 0 &&
-            equals(delkey, get_key(*it)));
+    if (!settings.use_deleted()) return false;
+    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)));
+    if (!settings.use_deleted()) return false;
+    return test_deleted_key(get_key(*it));
   }
   bool test_deleted(const destructive_iterator &it) const {
-    return (use_deleted && num_deleted > 0 &&
-            equals(delkey, get_key(*it)));
+    if (!settings.use_deleted()) return false;
+    return test_deleted_key(get_key(*it));
+  }
+
+ private:
+  // Set it so test_deleted is true.  true if object didn't used to be deleted.
+  // TODO(csilvers): make these private (also in densehashtable.h)
+  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());   // bad if set_deleted_key() wasn't called
     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());   // bad if set_deleted_key() wasn't called
     return test_deleted(it);
   }
 
-
   // FUNCTIONS CONCERNING SIZE
+ public:
   size_type size() const      { return table.num_nonempty() - 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 table.max_size(); }
   bool empty() const                  { return size() == 0; }
   size_type bucket_count() const      { return table.size(); }
   size_type max_bucket_count() const  { return max_size(); }
@@ -497,54 +554,57 @@ class sparse_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;
-    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(table.num_nonempty() >= 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
-        && (table.num_nonempty()-num_deleted) < shrink_threshold &&
-        bucket_count() > HT_DEFAULT_STARTING_BUCKETS ) {
+    const size_type num_remain = table.num_nonempty() - 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 &&
-              (table.num_nonempty() - num_deleted) <= sz *
-              shrink_resize_percent )
+      while (sz > HT_DEFAULT_STARTING_BUCKETS &&
+             num_remain < static_cast<size_type>(sz * shrink_factor)) {
         sz /= 2;                            // stay a power of 2
+      }
       sparse_hashtable tmp(MoveDontCopy, *this, sz);
       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();
+  // 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 (table.num_nonempty() >=
+        (STL_NAMESPACE::numeric_limits<size_type>::max)() - delta)
+      throw std::length_error("resize overflow");
     if ( bucket_count() >= HT_MIN_BUCKETS &&
-         (table.num_nonempty() + delta) <= enlarge_threshold )
-      return;                                // we're ok as we are
+         (table.num_nonempty() + 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
@@ -552,13 +612,34 @@ class sparse_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(table.num_nonempty() + delta, 0);
-    if ( needed_size > bucket_count() ) {      // we don't have enough buckets
-      const size_type resize_to = min_size(table.num_nonempty() - num_deleted
-                                           + delta, 0);
-      sparse_hashtable tmp(MoveDontCopy, *this, resize_to);
-      swap(tmp);                             // now we are tmp
+    const size_type needed_size =
+        settings.min_buckets(table.num_nonempty() + delta, 0);
+    if ( needed_size <= bucket_count() )      // we have enough buckets
+      return did_resize;
+
+    size_type resize_to =
+        settings.min_buckets(table.num_nonempty() - 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 (table.num_nonempty() - num_deleted + delta >= target) {
+        // Good, we won't be below the shrink threshhold even if we double.
+        resize_to *= 2;
+      }
     }
+
+    sparse_hashtable tmp(MoveDontCopy, *this, resize_to);
+    swap(tmp);                             // now we are tmp
+    return true;
   }
 
   // Used to actually do the rehashing when we grow/shrink a hashtable
@@ -566,16 +647,17 @@ class sparse_hashtable {
     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);
+    const size_type resize_to =
+        settings.min_buckets(ht.size(), min_buckets_wanted);
     if ( resize_to > bucket_count() ) {      // we don't have enough buckets
       table.resize(resize_to);               // sets the number of buckets
-      reset_thresholds();
+      settings.reset_thresholds(bucket_count());
     }
 
     // We use a normal iterator to get non-deleted bcks from ht
     // We could use insert() here, but since we know there are
     // no duplicates and no deleted items, we can be more efficient
-    assert( (bucket_count() & (bucket_count()-1)) == 0);      // a power of two
+    assert((bucket_count() & (bucket_count()-1)) == 0);      // a power of two
     for ( const_iterator it = ht.begin(); it != ht.end(); ++it ) {
       size_type num_probes = 0;              // how many times we've probed
       size_type bucknum;
@@ -584,10 +666,12 @@ class sparse_hashtable {
            table.test(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<>");
       }
       table.set(bucknum, *it);               // copies the value to here
     }
+    settings.inc_num_ht_copies();
   }
 
   // Implementation is like copy_from, but it destroys the table of the
@@ -598,14 +682,14 @@ class sparse_hashtable {
     clear();            // clear table, set num_deleted to 0
 
     // If we need to change the size of our table, do it now
-    size_t resize_to;
+    size_type resize_to;
     if ( mover == MoveDontGrow )
       resize_to = ht.bucket_count();         // keep same size as old ht
     else                                     // MoveDontCopy
-      resize_to = min_size(ht.size(), min_buckets_wanted);
+      resize_to = settings.min_buckets(ht.size(), min_buckets_wanted);
     if ( resize_to > bucket_count() ) {      // we don't have enough buckets
       table.resize(resize_to);               // sets the number of buckets
-      reset_thresholds();
+      settings.reset_thresholds(bucket_count());
     }
 
     // We use a normal iterator to get non-deleted bcks from ht
@@ -621,10 +705,12 @@ class sparse_hashtable {
             table.test(bucknum);                          // not empty
             bucknum = (bucknum + JUMP_(key, num_probes)) & (bucket_count()-1) ) {
         ++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<>");
       }
       table.set(bucknum, *it);               // copies the value to here
     }
+    settings.inc_num_ht_copies();
   }
 
 
@@ -634,28 +720,23 @@ class sparse_hashtable {
   // 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 > table.num_nonempty() )    // we only grow
       resize_delta(req_elements - table.num_nonempty());
   }
 
-  // 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,
@@ -665,15 +746,17 @@ class sparse_hashtable {
   explicit sparse_hashtable(size_type expected_max_items_in_table = 0,
                             const HashFcn& hf = HashFcn(),
                             const EqualKey& eql = EqualKey(),
+                            const ExtractKey& ext = ExtractKey(),
                             const SetKey& set = SetKey(),
-                            const ExtractKey& ext = ExtractKey())
-    : hash(hf), equals(eql), get_key(ext), set_key(set), num_deleted(0),
-      use_deleted(false), delkey(), enlarge_resize_percent(HT_OCCUPANCY_FLT),
-      shrink_resize_percent(HT_EMPTY_FLT),
-      table(expected_max_items_in_table == 0
-            ? HT_DEFAULT_STARTING_BUCKETS
-            : min_size(expected_max_items_in_table, 0)) {
-    reset_thresholds();
+                            const Alloc& alloc = Alloc())
+      : settings(hf),
+        key_info(ext, set, eql),
+        num_deleted(0),
+        table((expected_max_items_in_table == 0
+               ? HT_DEFAULT_STARTING_BUCKETS
+               : settings.min_buckets(expected_max_items_in_table, 0)),
+              alloc) {
+    settings.reset_thresholds(bucket_count());
   }
 
   // As a convenience for resize(), we allow an optional second argument
@@ -682,63 +765,51 @@ class sparse_hashtable {
   // into us instead of copying.
   sparse_hashtable(const sparse_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), delkey(ht.delkey),
-      enlarge_resize_percent(ht.enlarge_resize_percent),
-      shrink_resize_percent(ht.shrink_resize_percent),
-      table() {
-    reset_thresholds();
+      : settings(ht.settings),
+        key_info(ht.key_info),
+        num_deleted(0),
+        table(0, ht.get_allocator()) {
+    settings.reset_thresholds(bucket_count());
     copy_from(ht, min_buckets_wanted);   // copy_from() ignores deleted entries
   }
   sparse_hashtable(MoveDontCopyT mover, sparse_hashtable& ht,
                    size_type min_buckets_wanted = HT_DEFAULT_STARTING_BUCKETS)
-    : hash(ht.hash), equals(ht.equals), get_key(ht.get_key),
-      num_deleted(0), use_deleted(ht.use_deleted), delkey(ht.delkey),
-      enlarge_resize_percent(ht.enlarge_resize_percent),
-      shrink_resize_percent(ht.shrink_resize_percent),
-      table() {
-    reset_thresholds();
+      : settings(ht.settings),
+        key_info(ht.key_info),
+        num_deleted(0),
+        table(0, ht.get_allocator()) {
+    settings.reset_thresholds(bucket_count());
     move_from(mover, ht, min_buckets_wanted);  // ignores deleted entries
   }
 
   sparse_hashtable& operator= (const sparse_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;
-    delkey = ht.delkey;
-    copy_from(ht, HT_MIN_BUCKETS);         // sets num_deleted to 0 too
+    settings = ht.settings;
+    key_info = ht.key_info;
+    num_deleted = ht.num_deleted;
+    // 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;
   }
 
   // Many STL algorithms use swap instead of copy constructors
   void swap(sparse_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(enlarge_resize_percent, ht.enlarge_resize_percent);
-    STL_NAMESPACE::swap(shrink_resize_percent, ht.shrink_resize_percent);
-    STL_NAMESPACE::swap(delkey, ht.delkey);
     table.swap(ht.table);
-    reset_thresholds();
-    ht.reset_thresholds();
   }
 
   // It's always nice to be able to clear a table without deallocating it
   void clear() {
-    table.clear();
-    reset_thresholds();
+    if (!empty() || (num_deleted != 0)) {
+      table.clear();
+    }
+    settings.reset_thresholds(bucket_count());
     num_deleted = 0;
   }
 
-
   // LOOKUP ROUTINES
  private:
   // Returns a pair of positions: 1st where the object is, 2nd where
@@ -770,7 +841,8 @@ class sparse_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<>");
     }
   }
 
@@ -830,32 +902,58 @@ class sparse_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 deleted
+      // The set() below will undelete this object.  We just worry about stats
+      assert(num_deleted > 0);
+      --num_deleted;                // used to be, now it isn't
+    }
+    table.set(pos, obj);
+    return iterator(this, table.get_iter(pos), table.nonempty_end());
+  }
+
   // 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
-    assert(!use_deleted || !equals(get_key(obj), delkey));
+    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.get_iter(pos.first),
                                           table.nonempty_end()),
                                  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.
-        // The set() below will undelete this object.  We just worry about stats
-        assert(num_deleted > 0);
-        --num_deleted;                       // used to be, now it isn't
-      }
-      table.set(pos.second, obj);
-      return pair<iterator,bool>(iterator(this, table.get_iter(pos.second),
-                                          table.nonempty_end()),
-                                 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) {
-    resize_delta(1);  // adding an object, grow if need be
+  pair<iterator, bool> insert(const_reference obj) {
+    resize_delta(1);                      // adding an object, grow if need be
     return insert_noresize(obj);
   }
 
@@ -866,66 +964,102 @@ class sparse_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 sparse_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 delkey
+    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.get_iter(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 erasing delkey
-    assert(!use_deleted || !equals(key, delkey));
+    // First, double-check we're not erasing delkey.
+    assert((!settings.use_deleted() || !equals(key, key_info.delkey))
+           && "Erasing the deleted key");
+    assert(!settings.use_deleted() || !equals(key, key_info.delkey));
     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
+      // will think about shrink after next insert
+      settings.set_consider_shrink(true);
       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
+      // will think about shrink after next insert
+      settings.set_consider_shrink(true);
+    }
+  }
+
+  void erase(iterator f, iterator l) {
+    for ( ; f != l; ++f) {
+      if ( set_deleted(f)  )       // should always be true
+        ++num_deleted;
     }
+    // will think about shrink after next insert
+    settings.set_consider_shrink(true);
   }
 
+  // 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;
+      // will think about shrink after next insert
+      settings.set_consider_shrink(true);
+    }
+  }
   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
+    // will think about shrink after next insert
+    settings.set_consider_shrink(true);
   }
 
 
   // COMPARISON
   bool operator==(const sparse_hashtable& ht) const {
-    // We really want to check that the hash functions are the same
-    // but alas there's no way to do this.  We just hope.
-    return ( num_deleted == ht.num_deleted && table == ht.table );
+    if (size() != ht.size()) {
+      return false;
+    } else if (this == &ht) {
+      return true;
+    } else {
+      // Iterate through the elements in "this" and see if the
+      // corresponding element is in ht
+      for ( const_iterator it = begin(); it != end(); ++it ) {
+        const_iterator it2 = ht.find(get_key(*it));
+        if ((it2 == ht.end()) || (*it != *it2)) {
+          return false;
+        }
+      }
+      return true;
+    }
   }
   bool operator!=(const sparse_hashtable& ht) const {
     return !(*this == ht);
@@ -946,7 +1080,7 @@ class sparse_hashtable {
   bool read_metadata(FILE *fp) {
     num_deleted = 0;            // since we got rid before writing
     bool result = table.read_metadata(fp);
-    reset_thresholds();
+    settings.reset_thresholds(bucket_count());
     return result;
   }
 
@@ -961,31 +1095,67 @@ class sparse_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
-  // TODO(csilvers): make a pointer, and get rid of use_deleted (benchmark!)
-  key_type delkey;                           // which key marks deleted entries
-  float enlarge_resize_percent;                       // how full before resize
-  float shrink_resize_percent;                       // how empty before resize
-  size_type shrink_threshold;           // table.size() * shrink_resize_percent
-  size_type enlarge_threshold;         // table.size() * enlarge_resize_percent
-  sparsetable<value_type> table;      // holds num_buckets and num_elements too
-  bool consider_shrink;   // true if we should try to shrink before next insert
-
-  void reset_thresholds() {
-    enlarge_threshold = static_cast<size_type>(table.size()
-                                               * enlarge_resize_percent);
-    shrink_threshold = static_cast<size_type>(table.size()
-                                              * shrink_resize_percent);
-    consider_shrink = false;   // whatever caused us to reset already considered
+  // Table is the main storage class.
+  typedef sparsetable<value_type, DEFAULT_GROUP_SIZE, value_alloc_type> Table;
+
+  // Package templated functors with the other types to eliminate memory
+  // needed for storing these zero-size operators.  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) {}
+  };
+
+  // KeyInfo stores delete key and packages zero-size functors:
+  // ExtractKey and SetKey.
+  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;
+  };
+
+  // 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;
+  size_type num_deleted;   // how many occupied buckets are marked deleted
+  Table table;     // holds num_buckets and num_elements too
 };
 
+
 // 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(sparse_hashtable<V,K,HF,ExK,SetK,EqK,A> &x,
@@ -1002,13 +1172,14 @@ const typename sparse_hashtable<V,K,HF,ExK,SetK,EqK,A>::size_type
 // 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
 template <class V, class K, class HF, class ExK, class SetK, class EqK, class A>
-const float sparse_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_OCCUPANCY_FLT = 0.8f;
+const int sparse_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_OCCUPANCY_PCT = 80;
 
 // 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 sparse_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_EMPTY_FLT = 0.4f *
-sparse_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_OCCUPANCY_FLT;
+const int sparse_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_EMPTY_PCT
+  = static_cast<int>(0.4 *
+                     sparse_hashtable<V,K,HF,ExK,SetK,EqK,A>::HT_OCCUPANCY_PCT);
 
 _END_GOOGLE_NAMESPACE_