Merge branch 'master' of https://github.com/redpony/cdec

24 files changed, 473 insertions, 214 deletions

diff --git a/klm/lm/builder/lmplz_main.cc b/klm/lm/builder/lmplz_main.cc
index 1e086dcc..c87abdb8 100644
--- a/klm/lm/builder/lmplz_main.cc
+++ b/klm/lm/builder/lmplz_main.cc
@@ -52,13 +52,14 @@ int main(int argc, char *argv[]) {
       std::cerr << 
         "Builds unpruned language models with modified Kneser-Ney smoothing.\n\n"
         "Please cite:\n"
-        "@inproceedings{kenlm,\n"
-        "author    = {Kenneth Heafield},\n"
-        "title     = {{KenLM}: Faster and Smaller Language Model Queries},\n"
-        "booktitle = {Proceedings of the Sixth Workshop on Statistical Machine Translation},\n"
-        "month     = {July}, year={2011},\n"
-        "address   = {Edinburgh, UK},\n"
-        "publisher = {Association for Computational Linguistics},\n"
+        "@inproceedings{Heafield-estimate,\n"
+        "  author = {Kenneth Heafield and Ivan Pouzyrevsky and Jonathan H. Clark and Philipp Koehn},\n"
+        "  title = {Scalable Modified {Kneser-Ney} Language Model Estimation},\n"
+        "  year = {2013},\n"
+        "  month = {8},\n"
+        "  booktitle = {Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics},\n"
+        "  address = {Sofia, Bulgaria},\n"
+        "  url = {http://kheafield.com/professional/edinburgh/estimate\\_paper.pdf},\n"
         "}\n\n"
         "Provide the corpus on stdin.  The ARPA file will be written to stdout.  Order of\n"
         "the model (-o) is the only mandatory option.  As this is an on-disk program,\n"
diff --git a/klm/lm/builder/ngram.hh b/klm/lm/builder/ngram.hh
index 2984ed0b..f5681516 100644
--- a/klm/lm/builder/ngram.hh
+++ b/klm/lm/builder/ngram.hh
@@ -53,7 +53,7 @@ class NGram {
     Payload &Value() { return *reinterpret_cast<Payload *>(end_); }
 
     uint64_t &Count() { return Value().count; }
-    const uint64_t Count() const { return Value().count; }
+    uint64_t Count() const { return Value().count; }
 
     std::size_t Order() const { return end_ - begin_; }
 
diff --git a/klm/lm/model.cc b/klm/lm/model.cc
index a40fd2fb..a26654a6 100644
--- a/klm/lm/model.cc
+++ b/klm/lm/model.cc
@@ -304,5 +304,26 @@ template class GenericModel<trie::TrieSearch<SeparatelyQuantize, trie::DontBhiks
 template class GenericModel<trie::TrieSearch<SeparatelyQuantize, trie::ArrayBhiksha>, SortedVocabulary>;
 
 } // namespace detail
+
+base::Model *LoadVirtual(const char *file_name, const Config &config, ModelType model_type) {
+  RecognizeBinary(file_name, model_type);
+  switch (model_type) {
+    case PROBING:
+      return new ProbingModel(file_name, config);
+    case REST_PROBING:
+      return new RestProbingModel(file_name, config);
+    case TRIE:
+      return new TrieModel(file_name, config);
+    case QUANT_TRIE:
+      return new QuantTrieModel(file_name, config);
+    case ARRAY_TRIE:
+      return new ArrayTrieModel(file_name, config);
+    case QUANT_ARRAY_TRIE:
+      return new QuantArrayTrieModel(file_name, config);
+    default:
+      UTIL_THROW(FormatLoadException, "Confused by model type " << model_type);
+  }
+}
+
 } // namespace ngram
 } // namespace lm
diff --git a/klm/lm/model.hh b/klm/lm/model.hh
index 13ff864e..60f55110 100644
--- a/klm/lm/model.hh
+++ b/klm/lm/model.hh
@@ -153,6 +153,11 @@ LM_NAME_MODEL(QuantArrayTrieModel, detail::GenericModel<trie::TrieSearch<Separat
 typedef ::lm::ngram::ProbingVocabulary Vocabulary;
 typedef ProbingModel Model;
 
+/* Autorecognize the file type, load, and return the virtual base class.  Don't
+ * use the virtual base class if you can avoid it.  Instead, use the above
+ * classes as template arguments to your own virtual feature function.*/
+base::Model *LoadVirtual(const char *file_name, const Config &config = Config(), ModelType if_arpa = PROBING);
+
 } // namespace ngram
 } // namespace lm
 
diff --git a/klm/lm/search_hashed.cc b/klm/lm/search_hashed.cc
index 2d6f15b2..62275d27 100644
--- a/klm/lm/search_hashed.cc
+++ b/klm/lm/search_hashed.cc
@@ -54,7 +54,7 @@ template <class Weights> class ActivateUnigram {
     Weights *modify_;
 };
 
-// Find the lower order entry, inserting blanks along the way as necessary.  
+// Find the lower order entry, inserting blanks along the way as necessary.
 template <class Value> void FindLower(
     const std::vector<uint64_t> &keys,
     typename Value::Weights &unigram,
@@ -64,7 +64,7 @@ template <class Value> void FindLower(
   typename Value::ProbingEntry entry;
   // Backoff will always be 0.0.  We'll get the probability and rest in another pass.
   entry.value.backoff = kNoExtensionBackoff;
-  // Go back and find the longest right-aligned entry, informing it that it extends left.  Normally this will match immediately, but sometimes SRI is dumb.  
+  // Go back and find the longest right-aligned entry, informing it that it extends left.  Normally this will match immediately, but sometimes SRI is dumb.
   for (int lower = keys.size() - 2; ; --lower) {
     if (lower == -1) {
       between.push_back(&unigram);
@@ -77,11 +77,11 @@ template <class Value> void FindLower(
   }
 }
 
-// Between usually has  single entry, the value to adjust.  But sometimes SRI stupidly pruned entries so it has unitialized blank values to be set here.  
+// Between usually has  single entry, the value to adjust.  But sometimes SRI stupidly pruned entries so it has unitialized blank values to be set here.
 template <class Added, class Build> void AdjustLower(
     const Added &added,
     const Build &build,
-    std::vector<typename Build::Value::Weights *> &between, 
+    std::vector<typename Build::Value::Weights *> &between,
     const unsigned int n,
     const std::vector<WordIndex> &vocab_ids,
     typename Build::Value::Weights *unigrams,
@@ -93,14 +93,14 @@ template <class Added, class Build> void AdjustLower(
   }
   typedef util::ProbingHashTable<typename Value::ProbingEntry, util::IdentityHash> Middle;
   float prob = -fabs(between.back()->prob);
-  // Order of the n-gram on which probabilities are based.  
+  // Order of the n-gram on which probabilities are based.
   unsigned char basis = n - between.size();
   assert(basis != 0);
   typename Build::Value::Weights **change = &between.back();
   // Skip the basis.
   --change;
   if (basis == 1) {
-    // Hallucinate a bigram based on a unigram's backoff and a unigram probability.  
+    // Hallucinate a bigram based on a unigram's backoff and a unigram probability.
     float &backoff = unigrams[vocab_ids[1]].backoff;
     SetExtension(backoff);
     prob += backoff;
@@ -128,14 +128,14 @@ template <class Added, class Build> void AdjustLower(
   typename std::vector<typename Value::Weights *>::const_iterator i(between.begin());
   build.MarkExtends(**i, added);
   const typename Value::Weights *longer = *i;
-  // Everything has probability but is not marked as extending.  
+  // Everything has probability but is not marked as extending.
   for (++i; i != between.end(); ++i) {
     build.MarkExtends(**i, *longer);
     longer = *i;
   }
 }
 
-// Continue marking lower entries even they know that they extend left.  This is used for upper/lower bounds.  
+// Continue marking lower entries even they know that they extend left.  This is used for upper/lower bounds.
 template <class Build> void MarkLower(
     const std::vector<uint64_t> &keys,
     const Build &build,
@@ -144,15 +144,15 @@ template <class Build> void MarkLower(
     int start_order,
     const typename Build::Value::Weights &longer) {
   if (start_order == 0) return;
-  typename util::ProbingHashTable<typename Build::Value::ProbingEntry, util::IdentityHash>::MutableIterator iter;
-  // Hopefully the compiler will realize that if MarkExtends always returns false, it can simplify this code.  
+  // Hopefully the compiler will realize that if MarkExtends always returns false, it can simplify this code.
   for (int even_lower = start_order - 2 /* index in middle */; ; --even_lower) {
     if (even_lower == -1) {
       build.MarkExtends(unigram, longer);
       return;
     }
-    middle[even_lower].UnsafeMutableFind(keys[even_lower], iter);
-    if (!build.MarkExtends(iter->value, longer)) return;
+    if (!build.MarkExtends(
+          middle[even_lower].UnsafeMutableMustFind(keys[even_lower])->value,
+          longer)) return;
   }
 }
 
@@ -168,7 +168,6 @@ template <class Build, class Activate, class Store> void ReadNGrams(
     Store &store,
     PositiveProbWarn &warn) {
   typedef typename Build::Value Value;
-  typedef util::ProbingHashTable<typename Value::ProbingEntry, util::IdentityHash> Middle;
   assert(n >= 2);
   ReadNGramHeader(f, n);
 
@@ -186,7 +185,7 @@ template <class Build, class Activate, class Store> void ReadNGrams(
     for (unsigned int h = 1; h < n - 1; ++h) {
       keys[h] = detail::CombineWordHash(keys[h-1], vocab_ids[h+1]);
     }
-    // Initially the sign bit is on, indicating it does not extend left.  Most already have this but there might +0.0.  
+    // Initially the sign bit is on, indicating it does not extend left.  Most already have this but there might +0.0.
     util::SetSign(entry.value.prob);
     entry.key = keys[n-2];
 
@@ -203,7 +202,7 @@ template <class Build, class Activate, class Store> void ReadNGrams(
 
 } // namespace
 namespace detail {
- 
+
 template <class Value> uint8_t *HashedSearch<Value>::SetupMemory(uint8_t *start, const std::vector<uint64_t> &counts, const Config &config) {
   std::size_t allocated = Unigram::Size(counts[0]);
   unigram_ = Unigram(start, counts[0], allocated);
diff --git a/klm/lm/search_hashed.hh b/klm/lm/search_hashed.hh
index 00595796..9d067bc2 100644
--- a/klm/lm/search_hashed.hh
+++ b/klm/lm/search_hashed.hh
@@ -71,7 +71,7 @@ template <class Value> class HashedSearch {
     static const bool kDifferentRest = Value::kDifferentRest;
     static const unsigned int kVersion = 0;
 
-    // TODO: move probing_multiplier here with next binary file format update.  
+    // TODO: move probing_multiplier here with next binary file format update.
     static void UpdateConfigFromBinary(int, const std::vector<uint64_t> &, Config &) {}
 
     static uint64_t Size(const std::vector<uint64_t> &counts, const Config &config) {
@@ -102,14 +102,9 @@ template <class Value> class HashedSearch {
       return ret;
     }
 
-#pragma GCC diagnostic ignored "-Wuninitialized"
     MiddlePointer Unpack(uint64_t extend_pointer, unsigned char extend_length, Node &node) const {
       node = extend_pointer;
-      typename Middle::ConstIterator found;
-      bool got = middle_[extend_length - 2].Find(extend_pointer, found);
-      assert(got);
-      (void)got;
-      return MiddlePointer(found->value);
+      return MiddlePointer(middle_[extend_length - 2].MustFind(extend_pointer)->value);
     }
 
     MiddlePointer LookupMiddle(unsigned char order_minus_2, WordIndex word, Node &node, bool &independent_left, uint64_t &extend_pointer) const {
@@ -126,14 +121,14 @@ template <class Value> class HashedSearch {
     }
 
     LongestPointer LookupLongest(WordIndex word, const Node &node) const {
-      // Sign bit is always on because longest n-grams do not extend left.  
+      // Sign bit is always on because longest n-grams do not extend left.
       typename Longest::ConstIterator found;
       if (!longest_.Find(CombineWordHash(node, word), found)) return LongestPointer();
       return LongestPointer(found->value.prob);
     }
 
-    // Generate a node without necessarily checking that it actually exists.  
-    // Optionally return false if it's know to not exist.  
+    // Generate a node without necessarily checking that it actually exists.
+    // Optionally return false if it's know to not exist.
     bool FastMakeNode(const WordIndex *begin, const WordIndex *end, Node &node) const {
       assert(begin != end);
       node = static_cast<Node>(*begin);
@@ -144,7 +139,7 @@ template <class Value> class HashedSearch {
     }
 
   private:
-    // Interpret config's rest cost build policy and pass the right template argument to ApplyBuild.  
+    // Interpret config's rest cost build policy and pass the right template argument to ApplyBuild.
     void DispatchBuild(util::FilePiece &f, const std::vector<uint64_t> &counts, const Config &config, const ProbingVocabulary &vocab, PositiveProbWarn &warn);
 
     template <class Build> void ApplyBuild(util::FilePiece &f, const std::vector<uint64_t> &counts, const ProbingVocabulary &vocab, PositiveProbWarn &warn, const Build &build);
@@ -153,7 +148,7 @@ template <class Value> class HashedSearch {
       public:
         Unigram() {}
 
-        Unigram(void *start, uint64_t count, std::size_t /*allocated*/) : 
+        Unigram(void *start, uint64_t count, std::size_t /*allocated*/) :
           unigram_(static_cast<typename Value::Weights*>(start))
 #ifdef DEBUG
          ,  count_(count)
diff --git a/klm/lm/search_trie.hh b/klm/lm/search_trie.hh
index 1264baf5..60be416b 100644
--- a/klm/lm/search_trie.hh
+++ b/klm/lm/search_trie.hh
@@ -41,7 +41,8 @@ template <class Quant, class Bhiksha> class TrieSearch {
     static void UpdateConfigFromBinary(int fd, const std::vector<uint64_t> &counts, Config &config) {
       Quant::UpdateConfigFromBinary(fd, counts, config);
       util::AdvanceOrThrow(fd, Quant::Size(counts.size(), config) + Unigram::Size(counts[0]));
-      Bhiksha::UpdateConfigFromBinary(fd, config);
+      // Currently the unigram pointers are not compresssed, so there will only be a header for order > 2.
+      if (counts.size() > 2) Bhiksha::UpdateConfigFromBinary(fd, config);
     }
 
     static uint64_t Size(const std::vector<uint64_t> &counts, const Config &config) {
diff --git a/klm/lm/state.hh b/klm/lm/state.hh
index d8e6c132..a6b9accb 100644
--- a/klm/lm/state.hh
+++ b/klm/lm/state.hh
@@ -91,7 +91,7 @@ inline uint64_t hash_value(const Left &left) {
 }
 
 struct ChartState {
-  bool operator==(const ChartState &other) {
+  bool operator==(const ChartState &other) const {
     return (right == other.right) && (left == other.left);
   }
 
diff --git a/klm/lm/virtual_interface.hh b/klm/lm/virtual_interface.hh
index 6a5a0196..17f064b2 100644
--- a/klm/lm/virtual_interface.hh
+++ b/klm/lm/virtual_interface.hh
@@ -6,6 +6,7 @@
 #include "util/string_piece.hh"
 
 #include <string>
+#include <string.h>
 
 namespace lm {
 namespace base {
@@ -119,7 +120,9 @@ class Model {
 
     size_t StateSize() const { return state_size_; }
     const void *BeginSentenceMemory() const { return begin_sentence_memory_; }
+    void BeginSentenceWrite(void *to) const { memcpy(to, begin_sentence_memory_, StateSize()); }
     const void *NullContextMemory() const { return null_context_memory_; }
+    void NullContextWrite(void *to) const { memcpy(to, null_context_memory_, StateSize()); }
 
     // Requires in_state != out_state
     virtual float Score(const void *in_state, const WordIndex new_word, void *out_state) const = 0;
diff --git a/klm/lm/vocab.hh b/klm/lm/vocab.hh
index 3902f117..226ae438 100644
--- a/klm/lm/vocab.hh
+++ b/klm/lm/vocab.hh
@@ -25,7 +25,7 @@ uint64_t HashForVocab(const char *str, std::size_t len);
 inline uint64_t HashForVocab(const StringPiece &str) {
   return HashForVocab(str.data(), str.length());
 }
-class ProbingVocabularyHeader;
+struct ProbingVocabularyHeader;
 } // namespace detail
 
 class WriteWordsWrapper : public EnumerateVocab {
diff --git a/klm/search/Makefile.am b/klm/search/Makefile.am
index 03554276..b8c8a050 100644
--- a/klm/search/Makefile.am
+++ b/klm/search/Makefile.am
@@ -1,23 +1,10 @@
 noinst_LIBRARIES = libksearch.a
 
 libksearch_a_SOURCES = \
-  applied.hh \
-  config.hh \
-  context.hh \
-  dedupe.hh \
-  edge.hh \
-  edge_generator.hh \
-  header.hh \
-  nbest.hh \
-  rule.hh \
-  types.hh \
-  vertex.hh \
-  vertex_generator.hh \
   edge_generator.cc \
 	nbest.cc \
   rule.cc \
-  vertex.cc \
-  vertex_generator.cc
+  vertex.cc
 
-AM_CPPFLAGS = -W -Wall -Wno-sign-compare -I$(top_srcdir)/klm
+AM_CPPFLAGS = -W -Wall -Wno-sign-compare $(GTEST_CPPFLAGS) -I..
 
diff --git a/klm/search/context.hh b/klm/search/context.hh
index 08f21bbf..c3c8e53b 100644
--- a/klm/search/context.hh
+++ b/klm/search/context.hh
@@ -12,16 +12,6 @@ class ContextBase {
   public:
     explicit ContextBase(const Config &config) : config_(config) {}
 
-    VertexNode *NewVertexNode() {
-      VertexNode *ret = vertex_node_pool_.construct();
-      assert(ret);
-      return ret;
-    }
-
-    void DeleteVertexNode(VertexNode *node) {
-      vertex_node_pool_.destroy(node);
-    }
-
     unsigned int PopLimit() const { return config_.PopLimit(); }
 
     Score LMWeight() const { return config_.LMWeight(); }
@@ -29,8 +19,6 @@ class ContextBase {
     const Config &GetConfig() const { return config_; }
 
   private:
-    boost::object_pool<VertexNode> vertex_node_pool_;
-
     Config config_;
 };
 
diff --git a/klm/search/edge_generator.cc b/klm/search/edge_generator.cc
index eacf5de5..dd9d61e4 100644
--- a/klm/search/edge_generator.cc
+++ b/klm/search/edge_generator.cc
@@ -54,20 +54,20 @@ template <class Model> PartialEdge EdgeGenerator::Pop(Context<Model> &context) {
   Arity victim = 0;
   Arity victim_completed;
   Arity incomplete;
+  unsigned char lowest_niceness = 255;
   // Select victim or return if complete.   
   {
     Arity completed = 0;
-    unsigned char lowest_length = 255;
     for (Arity i = 0; i != arity; ++i) {
       if (top_nt[i].Complete()) {
         ++completed;
-      } else if (top_nt[i].Length() < lowest_length) {
-        lowest_length = top_nt[i].Length();
+      } else if (top_nt[i].Niceness() < lowest_niceness) {
+        lowest_niceness = top_nt[i].Niceness();
         victim = i;
         victim_completed = completed;
       }
     }
-    if (lowest_length == 255) {
+    if (lowest_niceness == 255) {
       return top;
     }
     incomplete = arity - completed;
@@ -92,10 +92,14 @@ template <class Model> PartialEdge EdgeGenerator::Pop(Context<Model> &context) {
     generate_.push(alternate);
   }
 
+#ifndef NDEBUG  
+  Score before = top.GetScore();
+#endif
   // top is now the continuation.
   FastScore(context, victim, victim - victim_completed, incomplete, old_value, top);
   // TODO: dedupe?  
   generate_.push(top);
+  assert(lowest_niceness != 254 || top.GetScore() == before);
 
   // Invalid indicates no new hypothesis generated.  
   return PartialEdge();
diff --git a/klm/search/vertex.cc b/klm/search/vertex.cc
index 45842982..bf40810e 100644
--- a/klm/search/vertex.cc
+++ b/klm/search/vertex.cc
@@ -2,6 +2,8 @@
 
 #include "search/context.hh"
 
+#include <boost/unordered_map.hpp>
+
 #include <algorithm>
 #include <functional>
 
@@ -11,45 +13,193 @@ namespace search {
 
 namespace {
 
-struct GreaterByBound : public std::binary_function<const VertexNode *, const VertexNode *, bool> {
-  bool operator()(const VertexNode *first, const VertexNode *second) const {
-    return first->Bound() > second->Bound();
+const uint64_t kCompleteAdd = static_cast<uint64_t>(-1);
+
+class DivideLeft {
+  public:
+    explicit DivideLeft(unsigned char index)
+      : index_(index) {}
+
+    uint64_t operator()(const lm::ngram::ChartState &state) const {
+      return (index_ < state.left.length) ? 
+        state.left.pointers[index_] :
+        (kCompleteAdd - state.left.full);
+    }
+
+  private:
+    unsigned char index_;
+};
+
+class DivideRight {
+  public:
+    explicit DivideRight(unsigned char index)
+      : index_(index) {}
+
+    uint64_t operator()(const lm::ngram::ChartState &state) const {
+      return (index_ < state.right.length) ?
+        static_cast<uint64_t>(state.right.words[index_]) :
+        (kCompleteAdd - state.left.full);
+    }
+
+  private:
+    unsigned char index_;
+};
+
+template <class Divider> void Split(const Divider &divider, const std::vector<HypoState> &hypos, std::vector<VertexNode> &extend) {
+  // Map from divider to index in extend.
+  typedef boost::unordered_map<uint64_t, std::size_t> Lookup;
+  Lookup lookup;
+  for (std::vector<HypoState>::const_iterator i = hypos.begin(); i != hypos.end(); ++i) {
+    uint64_t key = divider(i->state);
+    std::pair<Lookup::iterator, bool> res(lookup.insert(std::make_pair(key, extend.size())));
+    if (res.second) {
+      extend.resize(extend.size() + 1);
+      extend.back().AppendHypothesis(*i);
+    } else {
+      extend[res.first->second].AppendHypothesis(*i);
+    }
   }
+  //assert((extend.size() != 1) || (hypos.size() == 1));
+}
+
+lm::WordIndex Identify(const lm::ngram::Right &right, unsigned char index) {
+  return right.words[index];
+}
+
+uint64_t Identify(const lm::ngram::Left &left, unsigned char index) {
+  return left.pointers[index];
+}
+
+template <class Side> class DetermineSame {
+  public:
+    DetermineSame(const Side &side, unsigned char guaranteed) 
+      : side_(side), guaranteed_(guaranteed), shared_(side.length), complete_(true) {}
+
+    void Consider(const Side &other) {
+      if (shared_ != other.length) {
+        complete_ = false;
+        if (shared_ > other.length)
+          shared_ = other.length;
+      }
+      for (unsigned char i = guaranteed_; i < shared_; ++i) {
+        if (Identify(side_, i) != Identify(other, i)) {
+          shared_ = i;
+          complete_ = false;
+          return;
+        }
+      }
+    }
+
+    unsigned char Shared() const { return shared_; }
+
+    bool Complete() const { return complete_; }
+
+  private:
+    const Side &side_;
+    unsigned char guaranteed_, shared_;
+    bool complete_;
 };
 
+// Custom enum to save memory: valid values of policy_.
+// Alternate and there is still alternation to do.
+const unsigned char kPolicyAlternate = 0;
+// Branch based on left state only, because right ran out or this is a left tree.
+const unsigned char kPolicyOneLeft = 1;
+// Branch based on right state only.
+const unsigned char kPolicyOneRight = 2;
+// Reveal everything in the next branch.  Used to terminate the left/right policies.
+//    static const unsigned char kPolicyEverything = 3;
+
+} // namespace
+
+namespace {
+struct GreaterByScore : public std::binary_function<const HypoState &, const HypoState &, bool> {
+  bool operator()(const HypoState &first, const HypoState &second) const {
+    return first.score > second.score;
+  }
+};
 } // namespace
 
-void VertexNode::RecursiveSortAndSet(ContextBase &context, VertexNode *&parent_ptr) {
-  if (Complete()) {
-    assert(end_);
-    assert(extend_.empty());
-    return;
+void VertexNode::FinishRoot() {
+  std::sort(hypos_.begin(), hypos_.end(), GreaterByScore());
+  extend_.clear();
+  // HACK: extend to one hypo so that root can be blank.
+  state_.left.full = false;
+  state_.left.length = 0;
+  state_.right.length = 0;
+  right_full_ = false;
+  niceness_ = 0;
+  policy_ = kPolicyAlternate;
+  if (hypos_.size() == 1) {
+    extend_.resize(1);
+    extend_.front().AppendHypothesis(hypos_.front());
+    extend_.front().FinishedAppending(0, 0);
+  }
+  if (hypos_.empty()) {
+    bound_ = -INFINITY;
+  } else {
+    bound_ = hypos_.front().score;
   }
-  if (extend_.size() == 1) {
-    parent_ptr = extend_[0];
-    extend_[0]->RecursiveSortAndSet(context, parent_ptr);
-    context.DeleteVertexNode(this);
-    return;
+}
+
+void VertexNode::FinishedAppending(const unsigned char common_left, const unsigned char common_right) {
+  assert(!hypos_.empty());
+  assert(extend_.empty());
+  bound_ = hypos_.front().score;
+  state_ = hypos_.front().state;
+  bool all_full = state_.left.full;
+  bool all_non_full = !state_.left.full;
+  DetermineSame<lm::ngram::Left> left(state_.left, common_left);
+  DetermineSame<lm::ngram::Right> right(state_.right, common_right);
+  for (std::vector<HypoState>::const_iterator i = hypos_.begin() + 1; i != hypos_.end(); ++i) {
+    all_full &= i->state.left.full;
+    all_non_full &= !i->state.left.full;
+    left.Consider(i->state.left);
+    right.Consider(i->state.right);
   }
-  for (std::vector<VertexNode*>::iterator i = extend_.begin(); i != extend_.end(); ++i) {
-    (*i)->RecursiveSortAndSet(context, *i);
+  state_.left.full = all_full && left.Complete();
+  right_full_ = all_full && right.Complete();
+  state_.left.length = left.Shared();
+  state_.right.length = right.Shared();
+
+  if (!all_full && !all_non_full) {
+    policy_ = kPolicyAlternate;
+  } else if (left.Complete()) {
+    policy_ = kPolicyOneRight;
+  } else if (right.Complete()) {
+    policy_ = kPolicyOneLeft;
+  } else {
+    policy_ = kPolicyAlternate;
   }
-  std::sort(extend_.begin(), extend_.end(), GreaterByBound());
-  bound_ = extend_.front()->Bound();
+  niceness_ = state_.left.length + state_.right.length;
 }
 
-void VertexNode::SortAndSet(ContextBase &context) {
-  // This is the root.  The root might be empty.  
-  if (extend_.empty()) {
-    bound_ = -INFINITY;
-    return;
+void VertexNode::BuildExtend() {
+  // Already built.
+  if (!extend_.empty()) return;
+  // Nothing to build since this is a leaf.
+  if (hypos_.size() <= 1) return;
+  bool left_branch = true;
+  switch (policy_) {
+    case kPolicyAlternate:
+      left_branch = (state_.left.length <= state_.right.length);
+      break;
+    case kPolicyOneLeft:
+      left_branch = true;
+      break;
+    case kPolicyOneRight:
+      left_branch = false;
+      break;
+  }
+  if (left_branch) {
+    Split(DivideLeft(state_.left.length), hypos_, extend_);
+  } else {
+    Split(DivideRight(state_.right.length), hypos_, extend_);
   }
-  // The root cannot be replaced.  There's always one transition.  
-  for (std::vector<VertexNode*>::iterator i = extend_.begin(); i != extend_.end(); ++i) {
-    (*i)->RecursiveSortAndSet(context, *i);
+  for (std::vector<VertexNode>::iterator i = extend_.begin(); i != extend_.end(); ++i) {
+    // TODO: provide more here for branching?
+    i->FinishedAppending(state_.left.length, state_.right.length);
   }
-  std::sort(extend_.begin(), extend_.end(), GreaterByBound());
-  bound_ = extend_.front()->Bound();
 }
 
 } // namespace search
diff --git a/klm/search/vertex.hh b/klm/search/vertex.hh
index ca9a4fcd..81c3cfed 100644
--- a/klm/search/vertex.hh
+++ b/klm/search/vertex.hh
@@ -16,59 +16,74 @@ namespace search {
 
 class ContextBase;
 
+struct HypoState {
+  History history;
+  lm::ngram::ChartState state;
+  Score score;
+};
+
 class VertexNode {
   public:
-    VertexNode() : end_() {}
-
-    void InitRoot() {
-      extend_.clear();
-      state_.left.full = false;
-      state_.left.length = 0;
-      state_.right.length = 0;
-      right_full_ = false;
-      end_ = History();
+    VertexNode() {}
+
+    void InitRoot() { hypos_.clear(); }
+
+    /* The steps of building a VertexNode:
+     * 1. Default construct.
+     * 2. AppendHypothesis at least once, possibly multiple times.
+     * 3. FinishAppending with the number of words on left and right guaranteed
+     * to be common.
+     * 4. If !Complete(), call BuildExtend to construct the extensions
+     */
+    // Must default construct, call AppendHypothesis 1 or more times then do FinishedAppending.
+    void AppendHypothesis(const NBestComplete &best) {
+      assert(hypos_.empty() || !(hypos_.front().state == *best.state));
+      HypoState hypo;
+      hypo.history = best.history;
+      hypo.state = *best.state;
+      hypo.score = best.score;
+      hypos_.push_back(hypo);
+    }
+    void AppendHypothesis(const HypoState &hypo) {
+      hypos_.push_back(hypo);
     }
 
-    lm::ngram::ChartState &MutableState() { return state_; }
-    bool &MutableRightFull() { return right_full_; }
+    // Sort hypotheses for the root.
+    void FinishRoot();
 
-    void AddExtend(VertexNode *next) {
-      extend_.push_back(next);
-    }
+    void FinishedAppending(const unsigned char common_left, const unsigned char common_right);
 
-    void SetEnd(History end, Score score) {
-      assert(!end_);
-      end_ = end;
-      bound_ = score;
-    }
-    
-    void SortAndSet(ContextBase &context);
+    void BuildExtend();
 
     // Should only happen to a root node when the entire vertex is empty.   
     bool Empty() const {
-      return !end_ && extend_.empty();
+      return hypos_.empty() && extend_.empty();
     }
 
     bool Complete() const {
-      return end_;
+      // HACK: prevent root from being complete.  TODO: allow root to be complete.
+      return hypos_.size() == 1 && extend_.empty();
     }
 
     const lm::ngram::ChartState &State() const { return state_; }
     bool RightFull() const { return right_full_; }
 
+    // Priority relative to other non-terminals.  0 is highest.
+    unsigned char Niceness() const { return niceness_; }
+
     Score Bound() const {
       return bound_;
     }
 
-    unsigned char Length() const {
-      return state_.left.length + state_.right.length;
-    }
-
     // Will be invalid unless this is a leaf.   
-    History End() const { return end_; }
+    History End() const {
+      assert(hypos_.size() == 1);
+      return hypos_.front().history;
+    }
 
-    const VertexNode &operator[](size_t index) const {
-      return *extend_[index];
+    VertexNode &operator[](size_t index) {
+      assert(!extend_.empty());
+      return extend_[index];
     }
 
     size_t Size() const {
@@ -76,22 +91,26 @@ class VertexNode {
     }
 
   private:
-    void RecursiveSortAndSet(ContextBase &context, VertexNode *&parent);
+    // Hypotheses to be split.
+    std::vector<HypoState> hypos_;
 
-    std::vector<VertexNode*> extend_;
+    std::vector<VertexNode> extend_;
 
     lm::ngram::ChartState state_;
     bool right_full_;
 
+    unsigned char niceness_;
+
+    unsigned char policy_;
+
     Score bound_;
-    History end_;
 };
 
 class PartialVertex {
   public:
     PartialVertex() {}
 
-    explicit PartialVertex(const VertexNode &back) : back_(&back), index_(0) {}
+    explicit PartialVertex(VertexNode &back) : back_(&back), index_(0) {}
 
     bool Empty() const { return back_->Empty(); }
 
@@ -100,17 +119,14 @@ class PartialVertex {
     const lm::ngram::ChartState &State() const { return back_->State(); }
     bool RightFull() const { return back_->RightFull(); }
 
-    Score Bound() const { return Complete() ? back_->Bound() : (*back_)[index_].Bound(); }
-
-    unsigned char Length() const { return back_->Length(); }
+    Score Bound() const { return index_ ? (*back_)[index_].Bound() : back_->Bound(); }
 
-    bool HasAlternative() const {
-      return index_ + 1 < back_->Size();
-    }
+    unsigned char Niceness() const { return back_->Niceness(); }
 
     // Split into continuation and alternative, rendering this the continuation.
     bool Split(PartialVertex &alternative) {
       assert(!Complete());
+      back_->BuildExtend();
       bool ret;
       if (index_ + 1 < back_->Size()) {
         alternative.index_ = index_ + 1;
@@ -129,7 +145,7 @@ class PartialVertex {
     }
 
   private:
-    const VertexNode *back_;
+    VertexNode *back_;
     unsigned int index_;
 };
 
@@ -139,10 +155,21 @@ class Vertex {
   public:
     Vertex() {}
 
-    PartialVertex RootPartial() const { return PartialVertex(root_); }
+    //PartialVertex RootFirst() const { return PartialVertex(right_); }
+    PartialVertex RootAlternate() { return PartialVertex(root_); }
+    //PartialVertex RootLast() const { return PartialVertex(left_); }
+
+    bool Empty() const {
+      return root_.Empty();
+    }
+
+    Score Bound() const {
+      return root_.Bound();
+    }
 
-    History BestChild() const {
-      PartialVertex top(RootPartial());
+    History BestChild() {
+      // left_ and right_ are not set at the root.
+      PartialVertex top(RootAlternate());
       if (top.Empty()) {
         return History();
       } else {
@@ -158,6 +185,12 @@ class Vertex {
     template <class Output> friend class VertexGenerator;
     template <class Output> friend class RootVertexGenerator;
     VertexNode root_;
+
+    // These will not be set for the root vertex.
+    // Branches only on left state.
+    //VertexNode left_;
+    // Branches only on right state.
+    //VertexNode right_;
 };
 
 } // namespace search
diff --git a/klm/search/vertex_generator.hh b/klm/search/vertex_generator.hh
index 646b8189..91000012 100644
--- a/klm/search/vertex_generator.hh
+++ b/klm/search/vertex_generator.hh
@@ -4,10 +4,8 @@
 #include "search/edge.hh"
 #include "search/types.hh"
 #include "search/vertex.hh"
-#include "util/exception.hh"
 
 #include <boost/unordered_map.hpp>
-#include <boost/version.hpp>
 
 namespace lm {
 namespace ngram {
@@ -19,45 +17,25 @@ namespace search {
 
 class ContextBase;
 
-#if BOOST_VERSION > 104200
-// Parallel structure to VertexNode.  
-struct Trie {
-  Trie() : under(NULL) {}
-
-  VertexNode *under;
-  boost::unordered_map<uint64_t, Trie> extend;
-};
-
-void AddHypothesis(ContextBase &context, Trie &root, const NBestComplete &end);
-
-#endif // BOOST_VERSION
-
 // Output makes the single-best or n-best list.   
 template <class Output> class VertexGenerator {
   public:
-    VertexGenerator(ContextBase &context, Vertex &gen, Output &nbest) : context_(context), gen_(gen), nbest_(nbest) {
-      gen.root_.InitRoot();
-    }
+    VertexGenerator(ContextBase &context, Vertex &gen, Output &nbest) : context_(context), gen_(gen), nbest_(nbest) {}
 
     void NewHypothesis(PartialEdge partial) {
       nbest_.Add(existing_[hash_value(partial.CompletedState())], partial);
     }
 
     void FinishedSearch() {
-#if BOOST_VERSION > 104200
-      Trie root;
-      root.under = &gen_.root_;
+      gen_.root_.InitRoot();
       for (typename Existing::iterator i(existing_.begin()); i != existing_.end(); ++i) {
-        AddHypothesis(context_, root, nbest_.Complete(i->second));
+        gen_.root_.AppendHypothesis(nbest_.Complete(i->second));
       }
       existing_.clear();
-      root.under->SortAndSet(context_);
-#else
-      UTIL_THROW(util::Exception, "Upgrade Boost to >= 1.42.0 to use incremental search.");
-#endif
+      gen_.root_.FinishRoot();
     }
 
-    const Vertex &Generating() const { return gen_; }
+    Vertex &Generating() { return gen_; }
 
   private:
     ContextBase &context_;
@@ -84,8 +62,8 @@ template <class Output> class RootVertexGenerator {
 
     void FinishedSearch() {
       gen_.root_.InitRoot();
-      NBestComplete completed(out_.Complete(combine_));
-      gen_.root_.SetEnd(completed.history, completed.score);
+      gen_.root_.AppendHypothesis(out_.Complete(combine_));
+      gen_.root_.FinishRoot();
     }
 
   private:
diff --git a/klm/util/double-conversion/utils.h b/klm/util/double-conversion/utils.h
index 767094b8..9ccb3b65 100644
--- a/klm/util/double-conversion/utils.h
+++ b/klm/util/double-conversion/utils.h
@@ -218,7 +218,8 @@ class StringBuilder {
   // 0-characters; use the Finalize() method to terminate the string
   // instead.
   void AddCharacter(char c) {
-    ASSERT(c != '\0');
+    // I just extract raw data not a cstr so null is fine.
+    //ASSERT(c != '\0');
     ASSERT(!is_finalized() && position_ < buffer_.length());
     buffer_[position_++] = c;
   }
@@ -233,7 +234,8 @@ class StringBuilder {
   // builder. The input string must have enough characters.
   void AddSubstring(const char* s, int n) {
     ASSERT(!is_finalized() && position_ + n < buffer_.length());
-    ASSERT(static_cast<size_t>(n) <= strlen(s));
+    // I just extract raw data not a cstr so null is fine.
+    //ASSERT(static_cast<size_t>(n) <= strlen(s));
     memmove(&buffer_[position_], s, n * kCharSize);
     position_ += n;
   }
@@ -253,7 +255,8 @@ class StringBuilder {
     buffer_[position_] = '\0';
     // Make sure nobody managed to add a 0-character to the
     // buffer while building the string.
-    ASSERT(strlen(buffer_.start()) == static_cast<size_t>(position_));
+    // I just extract raw data not a cstr so null is fine.
+    //ASSERT(strlen(buffer_.start()) == static_cast<size_t>(position_));
     position_ = -1;
     ASSERT(is_finalized());
     return buffer_.start();
@@ -296,7 +299,11 @@ template <class Dest, class Source>
 inline Dest BitCast(const Source& source) {
   // Compile time assertion: sizeof(Dest) == sizeof(Source)
   // A compile error here means your Dest and Source have different sizes.
-  typedef char VerifySizesAreEqual[sizeof(Dest) == sizeof(Source) ? 1 : -1];
+  typedef char VerifySizesAreEqual[sizeof(Dest) == sizeof(Source) ? 1 : -1]
+#if __GNUC__ > 4 || __GNUC__ == 4 && __GNUC_MINOR__ >= 8
+      __attribute__((unused))
+#endif
+      ;
 
   Dest dest;
   memmove(&dest, &source, sizeof(dest));
diff --git a/klm/util/file.cc b/klm/util/file.cc
index c7d8e23b..bef04cb1 100644
--- a/klm/util/file.cc
+++ b/klm/util/file.cc
@@ -116,7 +116,7 @@ std::size_t GuardLarge(std::size_t size) {
   // The following operating systems have broken read/write/pread/pwrite that
   // only supports up to 2^31.
 #if defined(_WIN32) || defined(_WIN64) || defined(__APPLE__) || defined(OS_ANDROID)
-  return std::min(static_cast<std::size_t>(INT_MAX), size);
+  return std::min(static_cast<std::size_t>(static_cast<unsigned>(-1)), size);
 #else
   return size;
 #endif
@@ -209,7 +209,7 @@ void WriteOrThrow(int fd, const void *data_void, std::size_t size) {
 #endif
     errno = 0;
     do {
-      ret = 
+      ret =
 #if defined(_WIN32) || defined(_WIN64)
         _write
 #else
@@ -229,7 +229,7 @@ void WriteOrThrow(FILE *to, const void *data, std::size_t size) {
 }
 
 void FSyncOrThrow(int fd) {
-// Apparently windows doesn't have fsync?  
+// Apparently windows doesn't have fsync?
 #if !defined(_WIN32) && !defined(_WIN64)
   UTIL_THROW_IF_ARG(-1 == fsync(fd), FDException, (fd), "while syncing");
 #endif
@@ -248,7 +248,7 @@ template <> struct CheckOffT<8> {
 typedef CheckOffT<sizeof(off_t)>::True IgnoredType;
 #endif
 
-// Can't we all just get along?  
+// Can't we all just get along?
 void InternalSeek(int fd, int64_t off, int whence) {
   if (
 #if defined(_WIN32) || defined(_WIN64)
@@ -457,9 +457,9 @@ bool TryName(int fd, std::string &out) {
   std::ostringstream convert;
   convert << fd;
   name += convert.str();
-  
+
   struct stat sb;
-  if (-1 == lstat(name.c_str(), &sb)) 
+  if (-1 == lstat(name.c_str(), &sb))
     return false;
   out.resize(sb.st_size + 1);
   ssize_t ret = readlink(name.c_str(), &out[0], sb.st_size + 1);
@@ -471,7 +471,7 @@ bool TryName(int fd, std::string &out) {
   }
   out.resize(ret);
   // Don't use the non-file names.
-  if (!out.empty() && out[0] != '/') 
+  if (!out.empty() && out[0] != '/')
     return false;
   return true;
 #endif
diff --git a/klm/util/pool.cc b/klm/util/pool.cc
index 429ba158..db72a8ec 100644
--- a/klm/util/pool.cc
+++ b/klm/util/pool.cc
@@ -25,7 +25,9 @@ void Pool::FreeAll() {
 }
 
 void *Pool::More(std::size_t size) {
-  std::size_t amount = std::max(static_cast<size_t>(32) << free_list_.size(), size);
+  // Double until we hit 2^21 (2 MB).  Then grow in 2 MB blocks. 
+  std::size_t desired_size = static_cast<size_t>(32) << std::min(static_cast<std::size_t>(16), free_list_.size());
+  std::size_t amount = std::max(desired_size, size);
   uint8_t *ret = static_cast<uint8_t*>(MallocOrThrow(amount));
   free_list_.push_back(ret);
   current_ = ret + size;
diff --git a/klm/util/probing_hash_table.hh b/klm/util/probing_hash_table.hh
index 57866ff9..51a2944d 100644
--- a/klm/util/probing_hash_table.hh
+++ b/klm/util/probing_hash_table.hh
@@ -109,9 +109,20 @@ template <class EntryT, class HashT, class EqualT = std::equal_to<typename Entry
         if (equal_(got, key)) { out = i; return true; }
         if (equal_(got, invalid_)) return false;
         if (++i == end_) i = begin_;
-      }   
+      }
+    }
+
+    // Like UnsafeMutableFind, but the key must be there.
+    template <class Key> MutableIterator UnsafeMutableMustFind(const Key key) {
+       for (MutableIterator i(begin_ + (hash_(key) % buckets_));;) {
+        Key got(i->GetKey());
+        if (equal_(got, key)) { return i; }
+        assert(!equal_(got, invalid_));
+        if (++i == end_) i = begin_;
+      }
     }
 
+
     template <class Key> bool Find(const Key key, ConstIterator &out) const {
 #ifdef DEBUG
       assert(initialized_);
@@ -124,6 +135,16 @@ template <class EntryT, class HashT, class EqualT = std::equal_to<typename Entry
       }    
     }
 
+    // Like Find but we're sure it must be there.
+    template <class Key> ConstIterator MustFind(const Key key) const {
+      for (ConstIterator i(begin_ + (hash_(key) % buckets_));;) {
+        Key got(i->GetKey());
+        if (equal_(got, key)) { return i; }
+        assert(!equal_(got, invalid_));
+        if (++i == end_) i = begin_;
+      }
+    }
+
     void Clear() {
       Entry invalid;
       invalid.SetKey(invalid_);
diff --git a/klm/util/proxy_iterator.hh b/klm/util/proxy_iterator.hh
index 121a45fa..0ee1716f 100644
--- a/klm/util/proxy_iterator.hh
+++ b/klm/util/proxy_iterator.hh
@@ -6,11 +6,11 @@
 
 /* This is a RandomAccessIterator that uses a proxy to access the underlying
  * data.  Useful for packing data at bit offsets but still using STL
- * algorithms.  
+ * algorithms.
  *
  * Normally I would use boost::iterator_facade but some people are too lazy to
  * install boost and still want to use my language model.  It's amazing how
- * many operators an iterator has. 
+ * many operators an iterator has.
  *
  * The Proxy needs to provide:
  *   class InnerIterator;
@@ -22,15 +22,15 @@
  *   operator<(InnerIterator)
  *   operator+=(std::ptrdiff_t)
  *   operator-(InnerIterator)
- * and of course whatever Proxy needs to dereference it.  
+ * and of course whatever Proxy needs to dereference it.
  *
- * It's also a good idea to specialize std::swap for Proxy.  
+ * It's also a good idea to specialize std::swap for Proxy.
  */
 
 namespace util {
 template <class Proxy> class ProxyIterator {
   private:
-    // Self.  
+    // Self.
     typedef ProxyIterator<Proxy> S;
     typedef typename Proxy::InnerIterator InnerIterator;
 
@@ -38,16 +38,21 @@ template <class Proxy> class ProxyIterator {
     typedef std::random_access_iterator_tag iterator_category;
     typedef typename Proxy::value_type value_type;
     typedef std::ptrdiff_t difference_type;
-    typedef Proxy reference;
+    typedef Proxy & reference;
     typedef Proxy * pointer;
 
     ProxyIterator() {}
 
-    // For cast from non const to const.  
+    // For cast from non const to const.
     template <class AlternateProxy> ProxyIterator(const ProxyIterator<AlternateProxy> &in) : p_(*in) {}
     explicit ProxyIterator(const Proxy &p) : p_(p) {}
 
-    // p_'s operator= does value copying, but here we want iterator copying.  
+    // p_'s swap does value swapping, but here we want iterator swapping
+    friend inline void swap(ProxyIterator<Proxy> &first, ProxyIterator<Proxy> &second) {
+      swap(first.I(), second.I());
+    }
+
+    // p_'s operator= does value copying, but here we want iterator copying.
     S &operator=(const S &other) {
       I() = other.I();
       return *this;
@@ -72,8 +77,8 @@ template <class Proxy> class ProxyIterator {
 
     std::ptrdiff_t operator-(const S &other) const { return I() - other.I(); }
 
-    Proxy operator*() { return p_; }
-    const Proxy operator*() const { return p_; }
+    Proxy &operator*() { return p_; }
+    const Proxy &operator*() const { return p_; }
     Proxy *operator->() { return &p_; }
     const Proxy *operator->() const { return &p_; }
     Proxy operator[](std::ptrdiff_t amount) const { return *(*this + amount); }
diff --git a/klm/util/sized_iterator.hh b/klm/util/sized_iterator.hh
index cf998953..dce8f229 100644
--- a/klm/util/sized_iterator.hh
+++ b/klm/util/sized_iterator.hh
@@ -36,6 +36,11 @@ class SizedInnerIterator {
     void *Data() { return ptr_; }
     std::size_t EntrySize() const { return size_; }
 
+    friend inline void swap(SizedInnerIterator &first, SizedInnerIterator &second) {
+      std::swap(first.ptr_, second.ptr_);
+      std::swap(first.size_, second.size_);
+    }
+
   private:
     uint8_t *ptr_;
     std::size_t size_;
@@ -64,9 +69,19 @@ class SizedProxy {
     const void *Data() const { return inner_.Data(); }
     void *Data() { return inner_.Data(); }
 
+  /**
+     // TODO: this (deep) swap was recently added. why? if any std heap sort etc
+     // algs are using swap, that's going to be worse performance than using
+     // =. i'm not sure why we *want* a deep swap. if C++11 compilers are
+     // choosing between move constructor and swap, then we'd better implement a
+     // (deep) move constructor. it may also be that this is moot since i made
+     // ProxyIterator a reference and added a shallow ProxyIterator swap? (I
+     // need Ken or someone competent to judge whether that's correct also. -
+     // let me know at graehl@gmail.com
+  */
     friend void swap(SizedProxy &first, SizedProxy &second) {
       std::swap_ranges(
-          static_cast<char*>(first.inner_.Data()), 
+          static_cast<char*>(first.inner_.Data()),
           static_cast<char*>(first.inner_.Data()) + first.inner_.EntrySize(),
           static_cast<char*>(second.inner_.Data()));
     }
@@ -87,7 +102,7 @@ typedef ProxyIterator<SizedProxy> SizedIterator;
 
 inline SizedIterator SizedIt(void *ptr, std::size_t size) { return SizedIterator(SizedProxy(ptr, size)); }
 
-// Useful wrapper for a comparison function i.e. sort.  
+// Useful wrapper for a comparison function i.e. sort.
 template <class Delegate, class Proxy = SizedProxy> class SizedCompare : public std::binary_function<const Proxy &, const Proxy &, bool> {
   public:
     explicit SizedCompare(const Delegate &delegate = Delegate()) : delegate_(delegate) {}
@@ -106,7 +121,7 @@ template <class Delegate, class Proxy = SizedProxy> class SizedCompare : public
     }
 
     const Delegate &GetDelegate() const { return delegate_; }
-    
+
   private:
     const Delegate delegate_;
 };
diff --git a/klm/util/stream/chain.hh b/klm/util/stream/chain.hh
index 154b9b33..0cc83a85 100644
--- a/klm/util/stream/chain.hh
+++ b/klm/util/stream/chain.hh
@@ -122,7 +122,7 @@ class Chain {
       threads_.push_back(new Thread(Complete(), kRecycle));
     }
 
-    Chain &operator>>(const Recycler &recycle) {
+    Chain &operator>>(const Recycler &) {
       CompleteLoop();
       return *this;
     }
diff --git a/klm/util/usage.cc b/klm/util/usage.cc
index ad4dc7b4..8db375e1 100644
--- a/klm/util/usage.cc
+++ b/klm/util/usage.cc
@@ -5,51 +5,95 @@
 #include <fstream>
 #include <ostream>
 #include <sstream>
+#include <set>
+#include <string>
 
 #include <string.h>
 #include <ctype.h>
 #if !defined(_WIN32) && !defined(_WIN64)
 #include <sys/resource.h>
 #include <sys/time.h>
+#include <time.h>
 #include <unistd.h>
 #endif
 
 namespace util {
 
-namespace {
 #if !defined(_WIN32) && !defined(_WIN64)
+namespace {
+
+// On Mac OS X, clock_gettime is not implemented.
+// CLOCK_MONOTONIC is not defined either.
+#ifdef __MACH__
+#define CLOCK_MONOTONIC 0
+
+int clock_gettime(int clk_id, struct timespec *tp) {
+  struct timeval tv;
+  gettimeofday(&tv, NULL);
+  tp->tv_sec = tv.tv_sec;
+  tp->tv_nsec = tv.tv_usec * 1000;
+  return 0;
+}
+#endif // __MACH__
+
 float FloatSec(const struct timeval &tv) {
   return static_cast<float>(tv.tv_sec) + (static_cast<float>(tv.tv_usec) / 1000000.0);
 }
-#endif
+float FloatSec(const struct timespec &tv) {
+  return static_cast<float>(tv.tv_sec) + (static_cast<float>(tv.tv_nsec) / 1000000000.0);
+}
 
 const char *SkipSpaces(const char *at) {
-  for (; *at == ' '; ++at) {}
+  for (; *at == ' ' || *at == '\t'; ++at) {}
   return at;
 }
+
+class RecordStart {
+  public:
+    RecordStart() {
+      clock_gettime(CLOCK_MONOTONIC, &started_);
+    }
+
+    const struct timespec &Started() const {
+      return started_;
+    }
+
+  private:
+    struct timespec started_;
+};
+
+const RecordStart kRecordStart;
 } // namespace
+#endif
 
 void PrintUsage(std::ostream &out) {
 #if !defined(_WIN32) && !defined(_WIN64)
+  // Linux doesn't set memory usage in getrusage :-(
+  std::set<std::string> headers;
+  headers.insert("VmPeak:");
+  headers.insert("VmRSS:");
+  headers.insert("Name:");
+
+  std::ifstream status("/proc/self/status", std::ios::in);
+  std::string header, value;
+  while ((status >> header) && getline(status, value)) {
+    if (headers.find(header) != headers.end()) {
+      out << header << SkipSpaces(value.c_str()) << '\t';
+    }
+  }
+
   struct rusage usage;
-  if (getrusage(RUSAGE_SELF, &usage)) {
+  if (getrusage(RUSAGE_CHILDREN, &usage)) {
     perror("getrusage");
     return;
   }
-  out << "user\t" << FloatSec(usage.ru_utime) << "\nsys\t" << FloatSec(usage.ru_stime) << '\n';
-  out << "CPU\t" << (FloatSec(usage.ru_utime) + FloatSec(usage.ru_stime)) << '\n';
-  // Linux doesn't set memory usage :-(.  
-  std::ifstream status("/proc/self/status", std::ios::in);
-  std::string line;
-  while (getline(status, line)) {
-    if (!strncmp(line.c_str(), "VmRSS:\t", 7)) {
-      out << "RSSCur\t" << SkipSpaces(line.c_str() + 7) << '\n';
-      break;
-    } else if (!strncmp(line.c_str(), "VmPeak:\t", 8)) {
-      out << "VmPeak\t" << SkipSpaces(line.c_str() + 8) << '\n';
-    }
-  }
-  out << "RSSMax\t" << usage.ru_maxrss << " kB" << '\n';
+  out << "RSSMax:" << usage.ru_maxrss << " kB" << '\t';
+  out << "user:" << FloatSec(usage.ru_utime) << "\tsys:" << FloatSec(usage.ru_stime) << '\t';
+  out << "CPU:" << (FloatSec(usage.ru_utime) + FloatSec(usage.ru_stime));
+
+  struct timespec current;
+  clock_gettime(CLOCK_MONOTONIC, &current);
+  out << "\treal:" << (FloatSec(current) - FloatSec(kRecordStart.Started())) << '\n';
 #endif
 }