omg optimize

.gitignore

@@ -9,4 +9,6 @@
 llarpd
 
 *.test
-*.bin
+*.bin
+
+callgrind.*

Makefile

@@ -65,11 +65,13 @@ $(REPO)/test/%.cpp.bin: $(REPO)/test/%.cpp
 	$(CXX) $(REQUIRED_CXXFLAGS) $< -o $<.bin $(TEST_LDFLAGS)
 	mv $<.bin $<.test
 	$<.test
+	valgrind --tool=callgrind $<.test
 
 $(REPO)/test/%.c.bin: $(REPO)/test/%.c
 	$(CC) $(REQUIRED_CFLAGS) $< -o $<.bin $(TEST_LDFLAGS)
 	mv $<.bin $<.test
 	$<.test
+	valgrind --tool=callgrind $<.test
 
 $(EXE): $(DAEMON_OBJ) $(STATIC_LIB)
 	$(CXX) $(DAEMON_OBJ) $(STATIC_LIB) $(LIB_LDFLAGS) -o $(EXE) 

llarp/ev.cpp

@@ -5,6 +5,7 @@
 #include <mutex>
 #include <queue>
 
+
 struct llarp_ev_caller {
   static void *operator new(size_t sz) {
     return llarp::Alloc<llarp_ev_caller>();
@@ -21,37 +22,34 @@ struct llarp_ev_caller {
 
   bool appendCall(void *user) {
     std::unique_lock<std::mutex> lock(access);
-    bool should = pending.size() == 0;
+    pending.emplace_back(std::move(llarp_ev_async_call{loop, this, user, this->work}));
-    llarp_ev_async_call *call =
+    return true;
-        new llarp_ev_async_call{loop, this, user, this->work};
-    pending.push(call);
-    return should;
   }
 
   bool appendManyCalls(void **users, size_t n) {
     std::unique_lock<std::mutex> lock(access);
-    bool should = pending.size() == 0;
     while (n--) {
-      pending.push(new llarp_ev_async_call{loop, this, *users, this->work});
+      pending.emplace_back(std::move(llarp_ev_async_call{loop, this, *users, this->work}));
       users++;
     }
-    return should;
+    return true;
   }
 
   void Call() {
     std::unique_lock<std::mutex> lock(access);
-    while (pending.size() > 0) {
+    auto sz = pending.size();
+    while (sz > 0) {
       auto &front = pending.front();
-      front->work(front);
+      front.work(&front);
-      pending.pop();
+      pending.pop_front();
-      delete front;
+      --sz;
     }
   }
 
   std::mutex access;
   struct llarp_ev_loop *loop;
   uv_async_t async;
-  std::queue<llarp_ev_async_call *> pending;
+  std::deque<llarp_ev_async_call> pending;
   llarp_ev_work_func work;
 };
 

test/test_async_cipher.c

@@ -35,14 +35,14 @@ int main(int argc, char *argv[]) {
   struct bench_main b_main;
   struct llarp_threadpool *tp;
 
-  llarp_mem_jemalloc();
+  llarp_mem_std();
   llarp_crypto_libsodium_init(&b_main.crypto);
   llarp_ev_loop_alloc(&b_main.ev);
 
   tp = llarp_init_threadpool(2);
 
-  b_main.num = 10000000;
+  b_main.num = 100000;
-  b_main.jobs = 10000;
+  b_main.jobs = 1000;
   b_main.completed = 0;
   llarp_sharedkey_t key;
   b_main.crypto.randbytes(key, sizeof(llarp_sharedkey_t));