Add JFR tag encoding sidecar to ProfiledThread

Cache pre-computed Dictionary encodings and localRootSpanId in
ProfiledThread so signal handlers read O(1) arrays instead of
scanning attrs_data and probing the hash table on every sample.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: jbachorik

ddprof-lib/src/main/cpp/context_api.cpp

@@ -136,7 +136,13 @@ void ContextApi::setFull(u64 local_root_span_id, u64 span_id, u64 trace_id_high,
         // Write trace_id + span_id to OTEP record
         OtelContexts::set(trace_id_high, trace_id_low, span_id);
 
-        // Store local_root_span_id in ProfiledThread for O(1) read by signal handler
+        // Store local_root_span_id as custom attribute at reserved index 0.
+        // Readable by external OTEL profilers and decoded by writeCurrentContext.
+        uint8_t lrs_bytes[8];
+        OtelContexts::u64ToBytes(local_root_span_id, lrs_bytes);
+        OtelContexts::setAttribute(LOCAL_ROOT_SPAN_ATTR_INDEX, (const char*)lrs_bytes, 8);
+
+        // Cache in sidecar for O(1) signal-handler reads
         thrd->setOtelLocalRootSpanId(local_root_span_id);
     } else {
         // Profiler mode: local_root_span_id maps to rootSpanId, trace_id_high ignored
@@ -185,7 +191,7 @@ bool ContextApi::get(u64& span_id, u64& root_span_id) {
             return false;
         }
 
-        // Read local_root_span_id from ProfiledThread (O(1), signal-safe)
+        // Read local_root_span_id from sidecar (O(1), no attrs_data scan)
         root_span_id = 0;
         ProfiledThread* thrd = ProfiledThread::currentSignalSafe();
         if (thrd != nullptr) {
@@ -224,12 +230,11 @@ bool ContextApi::setAttribute(uint8_t key_index, const char* value, uint8_t valu
             initializeOtelTls(thrd);
         }
 
-        // Pre-register the value string in the Dictionary from this JNI thread.
-        // The signal handler (writeCurrentContext) will later call bounded_lookup
-        // to find the encoding — by pre-registering here, the signal handler
-        // only does a read (no malloc), which is async-signal-safe.
-        Profiler::instance()->contextValueMap()->bounded_lookup(
+        // Pre-register the value string in the Dictionary and cache the encoding
+        // in the sidecar for O(1) signal-handler reads (no hash lookup needed).
+        u32 encoding = Profiler::instance()->contextValueMap()->bounded_lookup(
             value, value_len, 1 << 16);
+        thrd->setOtelTagEncoding(key_index, encoding != INT_MAX ? encoding : 0);
 
         return OtelContexts::setAttribute(key_index, value, value_len);
     } else {

ddprof-lib/src/main/cpp/context_api.h

@@ -151,6 +151,10 @@ class ContextApi {
 
     static const int MAX_ATTRIBUTE_KEYS = 32;
 
+    // Reserved attribute index for local root span ID in OTEL attrs_data.
+    // Stored as 8-byte big-endian value. Visible to external OTEL profilers.
+    static const uint8_t LOCAL_ROOT_SPAN_ATTR_INDEX = 0;
+
 private:
     static ContextStorageMode _mode;
     static bool _initialized;

ddprof-lib/src/main/cpp/flightRecorder.cpp

@@ -11,7 +11,6 @@
 #include "callTraceHashTable.h"
 #include "context.h"
 #include "context_api.h"
-#include "otel_context.h"
 #include "counters.h"
 #include "dictionary.h"
 #include "flightRecorder.h"
@@ -1489,38 +1488,12 @@ void Recording::writeCurrentContext(Buffer *buf) {
   size_t numAttrs = Profiler::instance()->numContextAttributes();
 
   if (ContextApi::getMode() == CTX_STORAGE_OTEL) {
-    // In OTEL mode: decode attrs_data from the OTEP record only.
-    u32 tag_values[DD_TAGS_CAPACITY] = {};
-
+    // In OTEL mode: read pre-computed encodings from ProfiledThread sidecar.
+    // Encodings are cached by ContextApi::setAttribute() on the JNI thread —
+    // O(1) array read, no hash lookup or attrs_data scanning.
     ProfiledThread* thrd = ProfiledThread::currentSignalSafe();
-    if (thrd != nullptr && thrd->isOtelContextInitialized()) {
-      OtelThreadContextRecord* record = thrd->getOtelContextRecord();
-      uint16_t attrs_size = 0;
-      const uint8_t* attrs = OtelContexts::readAttrsData(record, attrs_size);
-      if (attrs != nullptr && attrs_size > 0) {
-        uint16_t pos = 0;
-        while (pos + 2 <= attrs_size) {
-          uint8_t key_index = attrs[pos];
-          uint8_t val_len = attrs[pos + 1];
-          if (pos + 2 + val_len > attrs_size) break;
-
-          if (key_index < numAttrs) {
-            // Read-only lookup (size_limit=0 → for_insert=false → no malloc).
-            // O(1) amortized — hash table with 3-way associative rows.
-            // Values are pre-registered by ContextApi::setAttribute() from the JNI thread.
-            u32 encoding = Profiler::instance()->contextValueMap()->bounded_lookup(
-                (const char*)(attrs + pos + 2), val_len, 0);
-            if (encoding != INT_MAX) {
-              tag_values[key_index] = encoding;
-            }
-          }
-          pos += 2 + val_len;
-        }
-      }
-    }
-
     for (size_t i = 0; i < numAttrs; i++) {
-      buf->putVar32(tag_values[i]);
+      buf->putVar32(thrd != nullptr ? thrd->getOtelTagEncoding(i) : 0);
     }
   } else {
     // Profiler mode: read directly from Context.tags[]

ddprof-lib/src/main/cpp/thread.cpp

@@ -4,6 +4,7 @@
 #include "profiler.h"
 #include "common.h"
 #include "vmStructs.h"
+#include <cstring>
 #include <time.h>
 
 pthread_key_t ProfiledThread::_tls_key;
@@ -100,11 +101,14 @@ void ProfiledThread::releaseFromBuffer() {
     _filter_slot_id = -1;
     _unwind_failures.clear();
 
+    // Reset OTEL sidecar (before freeing the record, for signal safety)
+    memset(_otel_tag_encodings, 0, sizeof(_otel_tag_encodings));
+    _otel_local_root_span_id = 0;
+
     // Free the OTEL context record (null first for signal safety)
     OtelThreadContextRecord* record = _otel_ctx_record;
     _otel_ctx_record = nullptr;
     _otel_ctx_initialized = false;
-    _otel_local_root_span_id = 0;
     delete record;
 
     // Put this ProfiledThread object back in the buffer for reuse

ddprof-lib/src/main/cpp/thread.h

@@ -6,6 +6,7 @@
 #ifndef _THREAD_H
 #define _THREAD_H
 
+#include "context.h"
 #include "os.h"
 #include "threadLocalData.h"
 #include "unwindStats.h"
@@ -74,11 +75,13 @@ class ProfiledThread : public ThreadLocalData {
   bool _crash_protection_active;
   Context* _ctx_tls_ptr;
   OtelThreadContextRecord* _otel_ctx_record;
-  u64 _otel_local_root_span_id;  // Cached local root span ID for OTEL mode (O(1) read)
+  u32 _otel_tag_encodings[DD_TAGS_CAPACITY];
+  u64 _otel_local_root_span_id;
 
   ProfiledThread(int buffer_pos, int tid)
       : ThreadLocalData(), _pc(0), _sp(0), _span_id(0), _root_span_id(0), _crash_depth(0), _buffer_pos(buffer_pos), _tid(tid), _cpu_epoch(0),
-        _wall_epoch(0), _call_trace_id(0), _recording_epoch(0), _misc_flags(0), _filter_slot_id(-1), _ctx_tls_initialized(false), _otel_ctx_initialized(false), _crash_protection_active(false), _ctx_tls_ptr(nullptr), _otel_ctx_record(nullptr), _otel_local_root_span_id(0) {};
+        _wall_epoch(0), _call_trace_id(0), _recording_epoch(0), _misc_flags(0), _filter_slot_id(-1), _ctx_tls_initialized(false), _otel_ctx_initialized(false), _crash_protection_active(false), _ctx_tls_ptr(nullptr), _otel_ctx_record(nullptr),
+        _otel_tag_encodings{}, _otel_local_root_span_id(0) {};
 
   ~ProfiledThread();  // Defined in thread.cpp (needs complete OtelThreadContextRecord type)
   void releaseFromBuffer();
@@ -208,13 +211,6 @@ class ProfiledThread : public ThreadLocalData {
     return _otel_ctx_record;
   }
 
-  inline void setOtelLocalRootSpanId(u64 id) {
-    _otel_local_root_span_id = id;
-  }
-
-  inline u64 getOtelLocalRootSpanId() {
-    return _otel_local_root_span_id;
-  }
 
   // JavaThread status cache — avoids repeated vtable checks in VMThread::isJavaThread().
   // JVMTI ThreadStart only fires for application threads, not for JVM-internal
@@ -247,6 +243,16 @@ class ProfiledThread : public ThreadLocalData {
   inline bool isCrashProtectionActive() const { return _crash_protection_active; }
   inline void setCrashProtectionActive(bool active) { _crash_protection_active = active; }
 
+  // OTEL JFR tag encoding sidecar — populated by JNI thread, read by signal handler
+  inline void setOtelTagEncoding(u32 idx, u32 val) {
+    if (idx < DD_TAGS_CAPACITY) _otel_tag_encodings[idx] = val;
+  }
+  inline u32 getOtelTagEncoding(u32 idx) const {
+    return idx < DD_TAGS_CAPACITY ? _otel_tag_encodings[idx] : 0;
+  }
+  inline void setOtelLocalRootSpanId(u64 id) { _otel_local_root_span_id = id; }
+  inline u64 getOtelLocalRootSpanId() const { return _otel_local_root_span_id; }
+
 private:
   // Atomic flag for signal handler reentrancy protection within the same thread
   // Must be atomic because a signal handler can interrupt normal execution mid-instruction,