perf: optimize hot-path loops, data structures, and cache locality

3 files changed, 49 insertions, 22 deletions

diff --git a/src/common/backend.c b/src/common/backend.c
index 4c755bd..c8d2caf 100644
--- a/src/common/backend.c
+++ b/src/common/backend.c
@@ -217,9 +217,13 @@ bool yasl_surface_present(struct yasl_surface *s, const struct yasl_framebuffer
         fb->size.x <= 0 || fb->size.y <= 0 ||
         damage->n < -1 || damage->n > YASL_DAMAGE_MAX)
         return false;
-    struct yasl_damage_list clipped = {.n = damage->n};
-    if (damage->n >= 0) {
+    struct yasl_damage_list clipped;
+    clipped.n = damage->n;
+    if (damage->n < 0) {
+        clipped.bounds = damage->bounds;
+    } else {
         clipped.n = 0;
+        clipped.bounds = (struct yasl_rect2i){0};
         for (int i = 0; i < damage->n; i++) {
             struct yasl_rect2i r = damage->rects[i];
             if (r.width < 0 || r.height < 0) return false;
diff --git a/src/common/input.c b/src/common/input.c
index eac4d1a..ff578e3 100644
--- a/src/common/input.c
+++ b/src/common/input.c
@@ -53,8 +53,13 @@ bool input_queue_push(struct input_queue *q, const struct yasl_input_event *ev)
             goto failed;
         struct yasl_input_event *events = malloc(capacity * sizeof(*events));
         if (!events) goto failed;
-        for (size_t i = 0; i < q->count; i++)
-            events[i] = q->events[(q->head + i) % q->capacity];
+        if (q->events && q->count > 0) {
+            size_t first = q->capacity - q->head;
+            if (first > q->count) first = q->count;
+            memcpy(events, q->events + q->head, first * sizeof(*events));
+            if (q->count > first)
+                memcpy(events + first, q->events, (q->count - first) * sizeof(*events));
+        }
         free(q->events);
         q->events = events;
         q->capacity = capacity;
diff --git a/src/common/loop.c b/src/common/loop.c
index a63df16..18d62b1 100644
--- a/src/common/loop.c
+++ b/src/common/loop.c
@@ -25,9 +25,9 @@ struct yasl_fd_source {
 struct yasl_timer_source {
     yasl_timer_cb fn;
     void *ud;
-    yasl_timer_id id;
     uint64_t interval_ns;
     uint64_t deadline_ns;
+    yasl_timer_id id;
     bool deleted;
 };
 
@@ -330,13 +330,18 @@ int yasl_loop_step(struct yasl_loop *l, int timeout_ms)
         for (int i = 0; i < count; i++) {
             if (pcopy[i].revents) {
                 short revents = pcopy[i].revents;
-                // Verify that the source was not deleted by an earlier callback in this step
+                // Fast-path: check if source is still at index i before linear scan
                 bool valid = false;
-                for (int j = 0; j < l->count; j++) {
-                    if (l->fds[j].id == scopy[i].id) {
-                        l->pollfds[j].revents = 0;
-                        valid = true;
-                        break;
+                if (i < l->count && l->fds[i].id == scopy[i].id) {
+                    l->pollfds[i].revents = 0;
+                    valid = true;
+                } else {
+                    for (int j = 0; j < l->count; j++) {
+                        if (l->fds[j].id == scopy[i].id) {
+                            l->pollfds[j].revents = 0;
+                            valid = true;
+                            break;
+                        }
                     }
                 }
                 if (valid && scopy[i].fn) {
@@ -367,18 +372,31 @@ int yasl_loop_step(struct yasl_loop *l, int timeout_ms)
             uint64_t next_interval = timer_src.fn(ud, tid, interval);
             handled++;
 
-            for (int j = 0; j < l->timer_count; j++) {
-                if (l->timers[j].id == tid) {
-                    if (next_interval == 0 || l->timers[j].deleted) {
-                        l->timers[j].deleted = true;
-                    } else {
-                        l->timers[j].interval_ns = next_interval;
-                        if (l->timers[j].deadline_ns + next_interval > now)
-                            l->timers[j].deadline_ns += next_interval;
-                        else
-                            l->timers[j].deadline_ns = now + next_interval;
+            // Fast-path: timer is usually still at index i unless an earlier timer was deleted
+            if (i < l->timer_count && l->timers[i].id == tid) {
+                if (next_interval == 0 || l->timers[i].deleted) {
+                    l->timers[i].deleted = true;
+                } else {
+                    l->timers[i].interval_ns = next_interval;
+                    if (l->timers[i].deadline_ns + next_interval > now)
+                        l->timers[i].deadline_ns += next_interval;
+                    else
+                        l->timers[i].deadline_ns = now + next_interval;
+                }
+            } else {
+                for (int j = 0; j < l->timer_count; j++) {
+                    if (l->timers[j].id == tid) {
+                        if (next_interval == 0 || l->timers[j].deleted) {
+                            l->timers[j].deleted = true;
+                        } else {
+                            l->timers[j].interval_ns = next_interval;
+                            if (l->timers[j].deadline_ns + next_interval > now)
+                                l->timers[j].deadline_ns += next_interval;
+                            else
+                                l->timers[j].deadline_ns = now + next_interval;
+                        }
+                        break;
                     }
-                    break;
                 }
             }
         }