Implement spend_cpu_cycles in assembly

The compiler was unrolling it previously, so we're doing assembly now for consistency.

src/config.hpp

@@ -79,8 +79,8 @@ struct SubscriptionConfig {
  */
 struct BenchmarkConfig {
   /// CPU-intensive loop iterations for /ok requests in resolve stage
-  /// 0 = health check only, 7000 = default benchmark load (650ns, 1M req/s)
-  int ok_resolve_iterations = 7000;
+  /// 0 = health check only, 4000 = default benchmark load (740ns, 1M req/s)
+  int ok_resolve_iterations = 0;
 };
 
 /**

src/cpu_work.cpp

@@ -1,10 +1,61 @@
 #include "cpu_work.hpp"
 
-void spend_cpu_cycles(int iterations) {
-  // Perform CPU-intensive work that cannot be optimized away
-  // Use inline assembly to prevent compiler optimization
-  for (int i = 0; i < iterations; ++i) {
-    // Simple loop with inline assembly barrier to prevent optimization
-    asm volatile("");
-  }
-}
+#if defined(__x86_64__) || defined(__amd64__)
+// x86-64 file-scoped assembly implementation
+#ifdef __APPLE__
+asm(".text\n"
+    ".globl _spend_cpu_cycles\n"
+    "_spend_cpu_cycles:\n"
+    "    test %edi, %edi\n" // Test if iterations <= 0
+    "    jle .L_end\n"      // Jump to end if <= 0
+    ".L_loop:\n"            // Loop start
+    "    dec %edi\n"        // Decrement iterations
+    "    jnz .L_loop\n"     // Jump back if not zero
+    ".L_end:\n"             // End
+    "    ret\n"             // Return
+);
+#else
+asm(".text\n"
+    ".globl spend_cpu_cycles\n"
+    ".type spend_cpu_cycles, @function\n"
+    "spend_cpu_cycles:\n"
+    "    test %edi, %edi\n" // Test if iterations <= 0
+    "    jle .L_end\n"      // Jump to end if <= 0
+    ".L_loop:\n"            // Loop start
+    "    dec %edi\n"        // Decrement iterations
+    "    jnz .L_loop\n"     // Jump back if not zero
+    ".L_end:\n"             // End
+    "    ret\n"             // Return
+    ".size spend_cpu_cycles, .-spend_cpu_cycles\n");
+#endif
+
+#elif defined(__aarch64__)
+// ARM64 file-scoped assembly implementation
+#ifdef __APPLE__
+asm(".text\n"
+    ".globl _spend_cpu_cycles\n"
+    "_spend_cpu_cycles:\n"
+    "    cmp w0, wzr\n"     // Compare iterations with zero
+    "    b.le .L_end\n"     // Branch to end if <= 0
+    ".L_loop:\n"            // Loop start
+    "    subs w0, w0, #1\n" // Decrement iterations and set flags
+    "    b.ne .L_loop\n"    // Branch back if not zero
+    ".L_end:\n"             // End
+    "    ret\n"             // Return
+);
+#else
+asm(".text\n"
+    ".globl spend_cpu_cycles\n"
+    ".type spend_cpu_cycles, %function\n"
+    "spend_cpu_cycles:\n"
+    "    cmp w0, wzr\n"     // Compare iterations with zero
+    "    b.le .L_end\n"     // Branch to end if <= 0
+    ".L_loop:\n"            // Loop start
+    "    subs w0, w0, #1\n" // Decrement iterations and set flags
+    "    b.ne .L_loop\n"    // Branch back if not zero
+    ".L_end:\n"             // End
+    "    ret\n"             // Return
+    ".size spend_cpu_cycles, spend_cpu_cycles\n");
+#endif
+
+#endif

src/cpu_work.hpp

@@ -1,5 +1,7 @@
 #pragma once
 
+extern "C" {
+
 /**
  * @brief Perform CPU-intensive work for benchmarking and health check purposes.
  *
@@ -10,6 +12,7 @@
  * @param iterations Number of loop iterations to perform
  */
 void spend_cpu_cycles(int iterations);
+}
 
 /**
  * @brief Default CPU work iterations for health check benchmarking.
@@ -18,4 +21,4 @@ void spend_cpu_cycles(int iterations);
  * /ok health check resolve stage. This value provides 650ns of CPU work
  * and achieves 1M requests/second throughput through the 4-stage pipeline.
  */
-constexpr int DEFAULT_HEALTH_CHECK_ITERATIONS = 7000;
+constexpr int DEFAULT_HEALTH_CHECK_ITERATIONS = 4000;

src/http_handler.cpp

@@ -79,6 +79,11 @@ void HttpHandler::on_connection_closed(Connection &conn) {
 void HttpHandler::on_write_buffer_drained(
     std::unique_ptr<Connection> &conn_ptr) {
   // Reset arena after all messages have been written for the next request
+  auto *state = static_cast<HttpConnectionState *>(conn_ptr->user_data);
+  if (state) {
+    TRACE_EVENT("http", "reply",
+                perfetto::Flow::Global(state->http_request_id));
+  }
   on_connection_closed(*conn_ptr);
   conn_ptr->reset();
   on_connection_established(*conn_ptr);

src/http_handler.hpp

@@ -88,7 +88,7 @@ struct HttpHandler : ConnectionHandler {
     resolveThread = std::thread{[this]() {
       pthread_setname_np(pthread_self(), "txn-resolve");
       for (;;) {
-        auto guard = commitPipeline.acquire<1, 0>();
+        auto guard = commitPipeline.acquire<1, 0>(/*maxBatch*/ 1);
         if (process_resolve_batch(guard.batch)) {
           return; // Shutdown signal received
         }