Add StringIndex: a generic open-addressed string set

gradle/libs.versions.toml

@@ -42,6 +42,7 @@ jmh = "1.37"
 # Profiling
 jmc = "8.1.0"
 jafar = "0.16.0"
+jol = "0.17"
 
 # Web & Network
 jnr-unixsocket = "0.38.25"
@@ -125,6 +126,7 @@ instrument-java = { module = "com.datadoghq:dd-instrument-java", version.ref = "
 jmc-common = { module = "org.openjdk.jmc:common", version.ref = "jmc" }
 jmc-flightrecorder = { module = "org.openjdk.jmc:flightrecorder", version.ref = "jmc" }
 jafar-tools = { module = "io.btrace:jafar-tools", version.ref = "jafar" }
+jol-core = { module = "org.openjdk.jol:jol-core", version.ref = "jol" }
 
 # Web & Network
 okio = { module = "com.datadoghq.okio:okio", version.ref = "okio" }

internal-api/build.gradle.kts

@@ -281,6 +281,7 @@ dependencies {
   testImplementation("org.junit.vintage:junit-vintage-engine:${libs.versions.junit5.get()}")
   testImplementation(libs.commons.math)
   testImplementation(libs.bundles.mockito)
+  testImplementation(libs.jol.core)
 }
 
 jmh {

internal-api/src/jmh/java/datadog/trace/util/ImmutableMapBenchmark.java

@@ -33,11 +33,55 @@
  * 10+, falls back to the input map pre-10). {@code Map.copyOf}/{@code MapN} is the honest
  * immutable-map baseline, not {@code HashMap}.
  *
+ * <p>Also compared: {@link StringIndex} used as a string-&gt;int map — an open-addressed index plus
+ * a slot-aligned {@code int[]} of values ({@code SI_VALUES[indexOf(key)]}). {@code
+ * stringIndex_get*} goes through the instance wrapper; {@code support_get*} reads via {@code static
+ * final} arrays (the JIT folds the refs). No {@code iterate} arm — StringIndex is a lookup index,
+ * not an iteration structure; its map use case is the {@code indexOf}-&gt;parallel-array read.
+ *
  * <p>Lookups use {@code EQUAL_KEYS} (distinct String instances) to exercise {@code equals()};
  * {@code *_sameKey} variants reuse the original interned key instances to show the identity fast
  * path — which is the common tracer case, since map keys are typically interned tag-name constants.
- * (Results pending a fresh multi-JVM run — {@code Map.copyOf} only materializes the compact form on
- * Java 10+.)
+ *
+ * <p>JDK 17 results (Apple M1, quiet machine, {@code @Fork(5)}, {@code @Threads(8)}; M ops/s).
+ * {@code get} uses distinct keys (exercises {@code equals()}); {@code sameKey} reuses the interned
+ * key (the {@code ==} fast path — the common tracer case):
+ *
+ * <pre>{@code
+ * Structure              get    sameKey
+ * support (static)      1498     2081    (fastest)
+ * stringIndex (inst)    1363     1900
+ * hashMap               1216     1850
+ * linkedHashMap         1214       -
+ * tagMap                1167     1386
+ * tracerImmutableMap    1049     1364    (MapN)
+ * treeMap                656       -
+ * }</pre>
+ *
+ * <p>{@code iterate} (full traversal):
+ *
+ * <pre>{@code
+ * tagMap.forEach        148    (fastest)
+ * linkedHashMap         136
+ * tracerImmutableMap    135    (MapN)
+ * treeMap               134
+ * hashMap               104
+ * tagMap (iterator)      96
+ * }</pre>
+ *
+ * <p>Key findings:
+ *
+ * <ul>
+ *   <li>StringIndex-as-map ({@code Support}) is the fastest {@code get} — beating {@code HashMap}
+ *       and {@code Map.copyOf}/{@code MapN}, most on the interned path; the instance wrapper trails
+ *       it by ~10%. (vs {@code MapN} the edge is speed + the slot/parallel-array capability, not
+ *       footprint — see {@link ImmutableSetBenchmark}.)
+ *   <li>{@code TagMap.forEach} (148) beats its own {@code iterator} (96) by ~1.5x: TagMap's
+ *       structure makes a faithful external {@code Iterator} expensive (externalized cursor +
+ *       skip-empty + per-call re-entry + the iterator allocation) — all of which internal {@code
+ *       forEach} avoids. Traverse TagMap via {@code forEach}, never its iterator; that gap only
+ *       widens as TagMap's entry model grows.
+ * </ul>
  */
 // @Fork(5): get_tracerImmutableMap* (MapN reached via interface dispatch) is JIT-bimodal at fewer
 // forks — 5
@@ -71,12 +115,31 @@ static void fill(Map<String, Integer> map) {
     }
   }
 
+  // StringIndex as a string->int map: an open-addressed index plus a slot-aligned int[] of values
+  // (VALUES[indexOf(key)]). support_* reads via static final arrays (JIT folds the refs to
+  // constants); stringIndex_* goes through the instance wrapper. Both share one placement --
+  // StringIndex.of and Support.create place identically -- so SI_VALUES aligns with either.
+  static final int[] SI_HASHES;
+  static final String[] SI_NAMES;
+  static final int[] SI_VALUES;
+
+  static {
+    StringIndex.Data data = StringIndex.Support.create(INSERTION_KEYS);
+    SI_HASHES = data.hashes;
+    SI_NAMES = data.names;
+    SI_VALUES = new int[SI_HASHES.length];
+    for (int i = 0; i < INSERTION_KEYS.length; ++i) {
+      SI_VALUES[StringIndex.Support.indexOf(SI_HASHES, SI_NAMES, INSERTION_KEYS[i])] = i;
+    }
+  }
+
   // Built once, never mutated -- safe to share across the reader threads.
   HashMap<String, Integer> hashMap;
   LinkedHashMap<String, Integer> linkedHashMap;
   TreeMap<String, Integer> treeMap;
   TagMap tagMap;
   Map<String, Integer> tracerImmutableMap;
+  StringIndex stringIndex;
 
   @Setup(Level.Trial)
   public void setUp() {
@@ -92,6 +155,7 @@ public void setUp() {
     }
     // JDK compact immutable map (MapN on Java 10+); the agent's actual fixed-map representation.
     tracerImmutableMap = CollectionUtils.tryMakeImmutableMap(hashMap);
+    stringIndex = StringIndex.of(INSERTION_KEYS);
   }
 
   /** Per-thread lookup cursor so each reader thread cycles keys independently. */
@@ -199,4 +263,25 @@ public void iterate_tracerImmutableMap(Blackhole blackhole) {
       blackhole.consume(entry.getValue());
     }
   }
+
+  @Benchmark
+  public int stringIndex_get(Cursor cursor) {
+    return SI_VALUES[stringIndex.indexOf(cursor.nextKey())];
+  }
+
+  @Benchmark
+  public int stringIndex_get_sameKey(Cursor cursor) {
+    return SI_VALUES[stringIndex.indexOf(cursor.nextKey(INSERTION_KEYS))];
+  }
+
+  @Benchmark
+  public int support_get(Cursor cursor) {
+    return SI_VALUES[StringIndex.Support.indexOf(SI_HASHES, SI_NAMES, cursor.nextKey())];
+  }
+
+  @Benchmark
+  public int support_get_sameKey(Cursor cursor) {
+    return SI_VALUES[
+        StringIndex.Support.indexOf(SI_HASHES, SI_NAMES, cursor.nextKey(INSERTION_KEYS))];
+  }
 }

internal-api/src/jmh/java/datadog/trace/util/ImmutableSetBenchmark.java

@@ -35,34 +35,57 @@
  *       ({@code ImmutableCollections.SetN}), which is what the agent actually uses for fixed config
  *       sets. Java 10+; falls back to {@code HashSet} pre-10. The realistic baseline for any
  *       flat/immutable set comparison.
+ *   <li>{@code stringIndex} — {@link StringIndex#contains} on the instance wrapper (one field load
+ *       to reach the placed arrays, then an open-addressed probe).
+ *   <li>{@code support} — the same probe via {@link StringIndex.Support#indexOf} over {@code static
+ *       final} arrays, so the JIT folds the refs to constants and there is nothing to dereference
+ *       (the hot path StringIndex recommends). The {@code stringIndex}/{@code support} pair shows
+ *       the indirection cost of the wrapper.
  * </ul>
  *
  * <p>Lookups are interned (the {@code ==} fast path where a structure has one); misses are short
  * and never present.
  *
- * <p>Java 17 results (Apple M1, {@code @Fork(2)}, {@code @Threads(8)}; M ops/s = millions):
+ * <p>JDK 17 results (Apple M1, quiet machine, {@code @Fork(5)}, {@code @Threads(8)}; M ops/s =
+ * millions):
  *
  * <pre>{@code
  * Structure              hit     miss
- * hashSet               2159     1751    (fastest)
- * tracerImmutableSet    1946     1633    (Set.copyOf / SetN)
- * array                  926      584
- * sortedArray            664      588
- * treeSet                642      593
+ * support (static)      2320     2159    (fastest)
+ * hashSet               2198     2134
+ * stringIndex (inst)    2098     1548 *  (* miss bimodal -- see caveat)
+ * tracerImmutableSet    1914     1663    (Set.copyOf / SetN)
+ * array                  941      589
+ * sortedArray            685      610
+ * treeSet                657      610
  * }</pre>
  *
  * <p>Key findings:
  *
  * <ul>
- *   <li>{@code HashSet} is fastest; {@link java.util.Set#copyOf} ({@code SetN}) trails by only ~10%
- *       on hit and ~7% on miss — and it's the compact, array-backed form the agent already uses for
- *       fixed config sets, so it's a strong default when the set is immutable.
- *   <li>{@code array} / {@code sortedArray} / {@code treeSet} cluster at ~0.6–0.9B — they scan,
- *       binary-search, or tree-walk per lookup, so they trail the hashed structures, most visibly
- *       on the miss path.
+ *   <li>The static {@code Support} path is the fastest — it beats {@code HashSet} on hit and miss
+ *       and crushes the scan/search/tree forms.
+ *   <li>{@code stringIndex} (the instance wrapper) trails {@code Support} by the field-load
+ *       indirection (~10% on hit), landing near {@code HashSet} — fine off the hot path, prefer
+ *       {@code Support} on it.
+ *   <li>{@link java.util.Set#copyOf} ({@code SetN}, the agent's compact fixed-set form) is ~1.2x
+ *       behind {@code Support} on hit but the most <i>compact</i> (~27% smaller — no cached hashes,
+ *       no 2x table). So StringIndex's edge over {@code SetN} is speed + the {@code
+ *       indexOf}-&gt;parallel-array capability, not footprint; over {@code HashSet} it wins both.
+ *   <li>{@code array} / {@code sortedArray} / {@code treeSet} trail the hashed structures, most on
+ *       miss.
  * </ul>
+ *
+ * <p><b>Caveat — the instance {@code stringIndex} miss is bimodal across forks</b> (confirmed at
+ * {@code @Fork(10)}: 6 forks fast, 4 slow, nothing between). ~60% of forks compile to a fast mode
+ * (~2000, ≈ {@code support_miss} — the wrapper indirection is then free) and ~40% to a slow mode
+ * (~1070, ~half); each fork locks one at warmup. So the {@code 1548 ±27%} above is a mode-mix, not
+ * noise. Cause: C2 hoists the instance field-loads ({@code this.hashes}/{@code names}) out of the
+ * miss-path probe loop only in the fast mode; the static {@code Support} path const-folds those
+ * refs and is never bimodal ({@code support_miss} ±0.3%). Prefer {@code Support} where miss latency
+ * matters.
  */
-@Fork(2)
+@Fork(5) // 5 forks settle the bimodal stringIndex_miss / interface-dispatch arms (see header)
 @Warmup(iterations = 2)
 @Measurement(iterations = 3)
 @Threads(8)
@@ -84,12 +107,26 @@ static String[] newMisses() {
     return misses;
   }
 
+  // StringIndex static-Support mode: the placed arrays pulled into static final fields, so the JIT
+  // folds the refs to constants and Support.indexOf has nothing to dereference (the hot path the
+  // StringIndex class Javadoc recommends). Contrast support_* (these) with stringIndex_* (the
+  // instance wrapper, one field load) to see the indirection cost.
+  static final int[] SI_HASHES;
+  static final String[] SI_NAMES;
+
+  static {
+    StringIndex.Data data = StringIndex.Support.create(STRINGS);
+    SI_HASHES = data.hashes;
+    SI_NAMES = data.names;
+  }
+
   // Built once, never mutated -- safe to share across the reader threads.
   String[] array;
   String[] sortedArray;
   HashSet<String> hashSet;
   TreeSet<String> treeSet;
   Set<String> tracerImmutableSet;
+  StringIndex stringIndex;
 
   @Setup(Level.Trial)
   public void setUp() {
@@ -99,6 +136,7 @@ public void setUp() {
     hashSet = new HashSet<>(Arrays.asList(STRINGS));
     treeSet = new TreeSet<>(Arrays.asList(STRINGS));
     tracerImmutableSet = CollectionUtils.tryMakeImmutableSet(Arrays.asList(STRINGS));
+    stringIndex = StringIndex.of(STRINGS);
   }
 
   /** Per-thread lookup cursor so each reader thread cycles keys independently. */
@@ -184,4 +222,24 @@ public boolean tracerImmutableSet_hit(Cursor cursor) {
   public boolean tracerImmutableSet_miss(Cursor cursor) {
     return tracerImmutableSet.contains(cursor.nextMiss());
   }
+
+  @Benchmark
+  public boolean stringIndex_hit(Cursor cursor) {
+    return stringIndex.contains(cursor.nextHit());
+  }
+
+  @Benchmark
+  public boolean stringIndex_miss(Cursor cursor) {
+    return stringIndex.contains(cursor.nextMiss());
+  }
+
+  @Benchmark
+  public boolean support_hit(Cursor cursor) {
+    return StringIndex.Support.indexOf(SI_HASHES, SI_NAMES, cursor.nextHit()) >= 0;
+  }
+
+  @Benchmark
+  public boolean support_miss(Cursor cursor) {
+    return StringIndex.Support.indexOf(SI_HASHES, SI_NAMES, cursor.nextMiss()) >= 0;
+  }
 }