From e97145a1ddca855f7bc4f1d0fdd3e23348f38448 Mon Sep 17 00:00:00 2001
From: Adrien Grand <jpountz@apache.org>
Date: Fri, 20 Jun 2014 12:22:31 +0000
Subject: [PATCH] LUCENE-5780: Make OrdinalMap more memory-efficient.

git-svn-id: https://svn.apache.org/repos/asf/lucene/dev/trunk@1604158 13f79535-47bb-0310-9956-ffa450edef68
---
 lucene/CHANGES.txt                               |  7 +++++--
 .../org/apache/lucene/index/MultiDocValues.java  | 16 ++++++++++++----
 2 files changed, 17 insertions(+), 6 deletions(-)

diff --git a/lucene/CHANGES.txt b/lucene/CHANGES.txt
index 1f54db01e39..6a48d22bc07 100644
--- a/lucene/CHANGES.txt
+++ b/lucene/CHANGES.txt
@@ -96,12 +96,15 @@ Other
 
 ======================= Lucene 4.10.0 ======================
 
-(No Changes)
-
 API Changes
 
 * LUCENE-5752: Simplified Automaton API to be immutable. (Mike McCandless)
 
+Optimizations
+
+* LUCENE-5780: Make OrdinalMap more memory-efficient, especially in case the
+  first segment has all values. (Adrien Grand, Robert Muir)
+
 ======================= Lucene 4.9.0 =======================
 
 Changes in Runtime Behavior
diff --git a/lucene/core/src/java/org/apache/lucene/index/MultiDocValues.java b/lucene/core/src/java/org/apache/lucene/index/MultiDocValues.java
index 23757e0197b..9ad9026f0a9 100644
--- a/lucene/core/src/java/org/apache/lucene/index/MultiDocValues.java
+++ b/lucene/core/src/java/org/apache/lucene/index/MultiDocValues.java
@@ -427,14 +427,18 @@ public class MultiDocValues {
       long globalOrd = 0;
       while (mte.next() != null) {        
         TermsEnumWithSlice matches[] = mte.getMatchArray();
+        int firstSegmentIndex = Integer.MAX_VALUE;
+        long globalOrdDelta = Long.MAX_VALUE;
         for (int i = 0; i < mte.getMatchCount(); i++) {
           int segmentIndex = matches[i].index;
           long segmentOrd = matches[i].terms.ord();
           long delta = globalOrd - segmentOrd;
-          // for each unique term, just mark the first segment index/delta where it occurs
-          if (i == 0) {
-            firstSegments.add(segmentIndex);
-            globalOrdDeltas.add(delta);
+          // We compute the least segment where the term occurs. In case the
+          // first segment contains most (or better all) values, this will
+          // help save significant memory
+          if (segmentIndex < firstSegmentIndex) {
+            firstSegmentIndex = segmentIndex;
+            globalOrdDelta = delta;
           }
           // for each per-segment ord, map it back to the global term.
           while (segmentOrds[segmentIndex] <= segmentOrd) {
@@ -443,6 +447,10 @@ public class MultiDocValues {
             segmentOrds[segmentIndex]++;
           }
         }
+        // for each unique term, just mark the first segment index/delta where it occurs
+        assert firstSegmentIndex < segmentOrds.length;
+        firstSegments.add(firstSegmentIndex);
+        globalOrdDeltas.add(globalOrdDelta);
         globalOrd++;
       }
       firstSegments.freeze();