Replacing Malloc and strdup in scan helper function with FileEntry and path arenas

lf_mpmc.h

@@ -36,6 +36,8 @@ typedef struct {
   CACHE_ALIGN atomic_size_t head;
   CACHE_ALIGN atomic_size_t tail;
 
+  CACHE_ALIGN atomic_size_t work_count;
+
   size_t capacity;
   size_t mask;
 
@@ -91,6 +93,7 @@ static void mpmc_init(MPMCQueue *q, size_t max_capacity) {
 
   atomic_init(&q->head, 0);
   atomic_init(&q->tail, 0);
+  atomic_init(&q->work_count, 0);
 
   plat_sem_init(&q->items_sem, 0);
 }
@@ -138,6 +141,7 @@ static void mpmc_commit_more(MPMCQueue *q) {
 /* ----------------------------------------------------------- */
 /* PUSH */
 /* ----------------------------------------------------------- */
+// Does not increment work
 static void mpmc_push(MPMCQueue *q, void *item) {
   MPMCSlot *slot;
   size_t pos;
@@ -184,8 +188,55 @@ static void mpmc_push(MPMCQueue *q, void *item) {
   plat_sem_post(&q->items_sem, 1);
 }
 
+// Increment work
+static void mpmc_push_work(MPMCQueue *q, void *item) {
+  MPMCSlot *slot;
+  size_t pos;
+
+  for (;;) {
+
+    pos = atomic_load_explicit(&q->tail, memory_order_relaxed);
+
+    // ensure the slot is committed BEFORE accessing it
+    size_t committed =
+        atomic_load_explicit(&q->committed, memory_order_relaxed);
+
+    if (unlikely(pos >= committed)) {
+      mpmc_commit_more(q);
+      continue;
+    }
+
+    slot = &q->slots[pos & q->mask];
+
+    size_t seq = atomic_load_explicit(&slot->seq, memory_order_acquire);
+    intptr_t diff = (intptr_t)seq - (intptr_t)pos;
+
+    if (likely(diff == 0)) {
+
+      if (atomic_compare_exchange_weak_explicit(&q->tail, &pos, pos + 1,
+                                                memory_order_relaxed,
+                                                memory_order_relaxed))
+        break;
+
+    } else if (diff < 0) { // queue actually full
+
+      Sleep(1000);
+
+    } else { // waiting to grow
+
+      Sleep(0);
+    }
+  }
+
+  slot->data = item;
+
+  atomic_store_explicit(&slot->seq, pos + 1, memory_order_release);
+
+  atomic_fetch_add(&q->work_count, 1);
+  plat_sem_post(&q->items_sem, 1);
+}
 /* ----------------------------------------------------------- */
-/* POP (blocking with semaphore) */
+/* POP */
 /* ----------------------------------------------------------- */
 static void *mpmc_pop(MPMCQueue *q) {
 
@@ -213,7 +264,7 @@ static void *mpmc_pop(MPMCQueue *q) {
 
     } else { // slot is still transitioning (written by another thread)
       if (++spins > 10) {
-        SwitchToThread(); // yield CPU
+        Sleep(0); // yield CPU
         spins = 0;
       } else {
         cpu_pause();
@@ -231,48 +282,48 @@ static void *mpmc_pop(MPMCQueue *q) {
 /* ----------------------------------------------------------- */
 /* TRY POP (non blocking) */
 /* ----------------------------------------------------------- */
-static b32 mpmc_try_pop(MPMCQueue *q, void **out) {
-
-  if (!plat_sem_trywait(&q->items_sem))
-    return false;
-
-  MPMCSlot *slot;
-  size_t pos;
-
-  int spins = 0;
-
-  for (;;) {
-
-    pos = atomic_load_explicit(&q->head, memory_order_relaxed);
-    slot = &q->slots[pos & q->mask];
-
-    size_t seq = atomic_load_explicit(&slot->seq, memory_order_acquire);
-    intptr_t diff = (intptr_t)seq - (intptr_t)(pos + 1);
-
-    if (likely(diff == 0)) {
-
-      if (atomic_compare_exchange_weak_explicit(&q->head, &pos, pos + 1,
-                                                memory_order_relaxed,
-                                                memory_order_relaxed))
-        break;
-
-    } else {
-
-      if (++spins > 10) {
-        SwitchToThread();
-        spins = 0;
-      } else {
-        cpu_pause();
-      }
-    }
-  }
-
-  *out = slot->data;
-
-  atomic_store_explicit(&slot->seq, pos + q->capacity, memory_order_release);
-
-  return true;
-}
+// static b32 mpmc_try_pop(MPMCQueue *q, void **out) {
+//
+//   if (!plat_sem_trywait(&q->items_sem))
+//     return false;
+//
+//   MPMCSlot *slot;
+//   size_t pos;
+//
+//   int spins = 0;
+//
+//   for (;;) {
+//
+//     pos = atomic_load_explicit(&q->head, memory_order_relaxed);
+//     slot = &q->slots[pos & q->mask];
+//
+//     size_t seq = atomic_load_explicit(&slot->seq, memory_order_acquire);
+//     intptr_t diff = (intptr_t)seq - (intptr_t)(pos + 1);
+//
+//     if (likely(diff == 0)) {
+//
+//       if (atomic_compare_exchange_weak_explicit(&q->head, &pos, pos + 1,
+//                                                 memory_order_relaxed,
+//                                                 memory_order_relaxed))
+//         break;
+//
+//     } else {
+//
+//       if (++spins > 10) {
+//         SwitchToThread();
+//         spins = 0;
+//       } else {
+//         cpu_pause();
+//       }
+//     }
+//   }
+//
+//   *out = slot->data;
+//
+//   atomic_store_explicit(&slot->seq, pos + q->capacity, memory_order_release);
+//
+//   return true;
+// }
 
 /* ----------------------------------------------------------- */
 /* PUSH POISON */
@@ -288,6 +339,16 @@ static void mpmc_producers_finished(MPMCQueue *q, u8 consumer_count) {
   }
 }
 
+/* ----------------------------------------------------------- */
+/* Done */
+/* ----------------------------------------------------------- */
+static void mpmc_task_done(MPMCQueue *q, u8 consumer_count) {
+  size_t prev = atomic_fetch_sub(&q->work_count, 1);
+  if (prev == 1) {
+    mpmc_producers_finished(q, consumer_count);
+  }
+}
+
 /* ----------------------------------------------------------- */
 /* MPMC Cleanup */
 /* ----------------------------------------------------------- */