Speed up flushing of live docs. (#14998)

This uses `Bits#applyMask` (introduced in #14134) to speed up flushing live
docs, instead of checking bits one by one.

Author: jpountz

lucene/CHANGES.txt

@@ -129,6 +129,8 @@ Optimizations
 ---------------------
 * GITHUB#15140: Optimize TopScoreDocCollector with TernaryLongHeap for improved performance over Binary-LongHeap. (Ramakrishna Chilaka)
 
+* GITHUB#14998: Speed up flushing of live docs. (Adrien Grand)
+
 Bug Fixes
 ---------------------
 * GITHUB#14161: PointInSetQuery's constructor now throws IllegalArgumentException

lucene/core/src/java/org/apache/lucene/codecs/lucene90/Lucene90LiveDocsFormat.java

@@ -138,18 +138,23 @@ public void writeLiveDocs(
   }
 
   private int writeBits(IndexOutput output, Bits bits) throws IOException {
-    int delCount = 0;
-    final int longCount = FixedBitSet.bits2words(bits.length());
-    for (int i = 0; i < longCount; ++i) {
-      long currentBits = 0;
-      for (int j = i << 6, end = Math.min(j + 63, bits.length() - 1); j <= end; ++j) {
-        if (bits.get(j)) {
-          currentBits |= 1L << j; // mod 64
-        } else {
-          delCount += 1;
-        }
+    int delCount = bits.length();
+    // Copy bits in batches of 1024 bits at once using Bits#applyMask, which is faster than checking
+    // bits one by one.
+    FixedBitSet copy = new FixedBitSet(1024);
+    for (int offset = 0; offset < bits.length(); offset += copy.length()) {
+      int numBitsToCopy = Math.min(bits.length() - offset, copy.length());
+      copy.set(0, copy.length());
+      if (numBitsToCopy < copy.length()) {
+        // Clear ghost bits
+        copy.clear(numBitsToCopy, copy.length());
+      }
+      bits.applyMask(copy, offset);
+      delCount -= copy.cardinality();
+      int longCount = FixedBitSet.bits2words(numBitsToCopy);
+      for (int i = 0; i < longCount; ++i) {
+        output.writeLong(copy.getBits()[i]);
       }
-      output.writeLong(currentBits);
     }
     return delCount;
   }