core: optimize struct layouts, eliminate branches, and fix event loop dispatch

- Fix Wayland event loop lockout by properly decoupling read arming and adding poll-backed dispatching
- Support standalone display lifecycle without requiring caller to manage a loop
- Eliminate alignment holes across yasl_surface, yasl_display, seat, touch_point, and conn_drag using pahole
- Replace branching state/button decoders with compile-time locked LUTs and branchless bitwise operations
- Accelerate UTF-8 validation with 64-bit word scanning for ASCII runs

6 files changed, 269 insertions, 123 deletions

diff --git a/src/wayland/buffer.c b/src/wayland/buffer.c
index 492cfde..0d639e6 100644
--- a/src/wayland/buffer.c
+++ b/src/wayland/buffer.c
@@ -305,13 +305,20 @@ static void arm_frame(struct yasl_surface *w)
     }
 }
 
+static const int8_t free_buf_lut[3][8] = {
+    [0] = { 1, 1, 2, 2, 1, 1, 0, -1 },
+    [1] = { 0, 2, 0, 2, 0, 1, 0, -1 },
+    [2] = { 0, 1, 0, 2, 0, 1, 0, -1 },
+};
+
 int buffer_pick_free_index(const struct buffer *bufs, int cur)
 {
-    if (!bufs[cur ^ 1].busy) return cur ^ 1;
-    for (int i = 0; i < NBUF; i++)
-        if (i != cur && i != (cur ^ 1) && !bufs[i].busy) return i;
-    if (!bufs[cur].busy) return cur;
-    return -1;
+    unsigned busy = ((unsigned)bufs[0].busy) |
+                    (((unsigned)bufs[1].busy) << 1) |
+                    (((unsigned)bufs[2].busy) << 2);
+    unsigned c = (unsigned)cur;
+    if (unlikely(c >= 3)) c = 0;
+    return free_buf_lut[c][busy];
 }
 
 struct latency_feedback {
@@ -523,14 +530,12 @@ void yasl_surface_set_title(struct yasl_surface *w, const char *title)
 
 uint32_t yasl_surface_state(const struct yasl_surface *w)
 {
-    if (!w) return 0;
-    uint32_t s = 0;
-    if (w->fullscreen) s |= YASL_STATE_FULLSCREEN;
-    if (w->maximized)  s |= YASL_STATE_MAXIMIZED;
-    if (w->activated)  s |= YASL_STATE_ACTIVATED;
-    if (w->suspended)  s |= YASL_STATE_SUSPENDED;
-    if (!w->server_decorated) s |= YASL_STATE_CSD;
-    return s;
+    if (unlikely(!w)) return 0;
+    return ((uint32_t)w->fullscreen        * YASL_STATE_FULLSCREEN) |
+           ((uint32_t)w->maximized         * YASL_STATE_MAXIMIZED)  |
+           ((uint32_t)w->activated         * YASL_STATE_ACTIVATED)  |
+           ((uint32_t)w->suspended         * YASL_STATE_SUSPENDED)  |
+           ((uint32_t)!w->server_decorated * YASL_STATE_CSD);
 }
 
 struct yasl_timing yasl_surface_timing(const struct yasl_surface *w)
diff --git a/src/wayland/internal.h b/src/wayland/internal.h
index dc85943..79fd88d 100644
--- a/src/wayland/internal.h
+++ b/src/wayland/internal.h
@@ -89,14 +89,21 @@ struct drag_icon {
     int fd;
 };
 
+#ifndef likely
+#define likely(x)   __builtin_expect(!!(x), 1)
+#endif
+#ifndef unlikely
+#define unlikely(x) __builtin_expect(!!(x), 0)
+#endif
+
 struct conn_drag {
     struct wl_data_source *source;
     struct xdg_toplevel_drag_v1 *toplevel;
     struct drag_icon icon;
     const struct yasl_drag_listener *listener;
     void *ud;
-    uint32_t scale_120;
     struct seat *seat;
+    uint32_t scale_120;
     bool dropped;
     bool carrying;
 };
@@ -105,17 +112,19 @@ struct conn_drag {
 
 struct touch_point {
     struct yasl_surface *surface;
-    int32_t id;
     double x, y;
+    int32_t id;
     bool active;
-    bool down, moved, up;
+    bool down;
+    bool moved;
+    bool up;
 };
 
 struct seat {
+    /* 8-byte aligned pointers and 64-bit values */
     struct yasl_display *conn;
     struct seat *next;
     struct wl_seat *proxy;
-    uint32_t name;
     struct wl_keyboard *keyboard;
     struct wl_pointer *pointer;
     struct wl_touch *touch;
@@ -128,59 +137,63 @@ struct seat {
     struct tablet_pad *tablet_pads;
     struct zwp_text_input_v3 *text_input;
     struct clipboard *clipboard;
-
     struct xkb_keymap *xkb_keymap;
     struct xkb_state *xkb_state;
-    xkb_mod_mask_t input_mod_masks[8];
-    xkb_mod_index_t mod_shift, mod_alt, mod_ctrl, mod_super;
-    bool input_mods_ready;
     uint64_t repeat_started_ns;
-    int repeat_fd, repeat_watch;
-    int32_t repeat_rate, repeat_delay;
-    uint32_t repeat_key;
-    bool repeat_held;
-
     struct yasl_surface *pointer_focus;
     struct yasl_surface *keyboard_focus;
     struct yasl_surface *tablet_focus;
     struct yasl_surface *ime_focus;
     struct yasl_surface *focus_leaving;
-    uint32_t input_serial;
-    uint32_t press_serial;
     struct yasl_surface *press_surface;
+    struct yasl_surface *ptr_left;
+    struct tablet_tool *tool;
+    struct touch_point *touch_primary;
+    uint64_t pointer_buttons;
+    uint64_t tool_buttons;
+    double axis_v, axis_h, frame_v;
+    double touch_scroll_y, touch_scroll_rest;
+
+    /* Embedded structs */
+    xkb_mod_mask_t input_mod_masks[8];
+    struct yasl_vector2i pointer_pos;
+    struct yasl_vector2i ptr_left_pos;
+    struct yasl_vector2i tool_pos;
+    struct touch_point touch_points[SEAT_TOUCH_POINTS];
+    struct press_queue ptr_presses;
+    struct press_queue tool_presses;
 
+    /* 4-byte integers */
+    uint32_t name;
+    xkb_mod_index_t mod_shift, mod_alt, mod_ctrl, mod_super;
+    int repeat_fd, repeat_watch;
+    int32_t repeat_rate, repeat_delay;
+    uint32_t repeat_key;
+    uint32_t input_serial;
+    uint32_t press_serial;
     uint32_t pointer_serial;
     int cursor_shape;
-    struct yasl_vector2i pointer_pos;
-    uint64_t pointer_buttons;
-    double axis_v, axis_h, frame_v;
     int frame_v120;
     int axis_discrete_v, axis_discrete_h;
     int axis_value120_v, axis_value120_h;
-    bool axis_has_v120_v, axis_has_v120_h;
+    struct yasl_pointer_pen pen;
+    int tool_wheel;
+
+    /* 1-byte booleans grouped tightly */
+    bool input_mods_ready;
+    bool repeat_held;
+    bool axis_has_v120_v;
+    bool axis_has_v120_h;
     bool axis_smooth;
     bool motion_pending;
-    struct press_queue ptr_presses;
-    struct yasl_surface *ptr_left;
-    struct yasl_vector2i ptr_left_pos;
     bool ptr_left_motion;
-
-    struct tablet_tool *tool;
-    struct press_queue tool_presses;
-    struct yasl_vector2i tool_pos;
-    uint64_t tool_buttons;
-    struct yasl_pointer_pen pen;
-    int tool_wheel;
     bool tool_motion;
-
-    struct touch_point touch_points[SEAT_TOUCH_POINTS];
-    struct touch_point *touch_primary;
     bool touch_pressed;
     bool touch_scrolling;
-    double touch_scroll_y, touch_scroll_rest;
 };
 
 #define YASL_QUEUE_CAP 256
+#define YASL_QUEUE_MASK (YASL_QUEUE_CAP - 1)
 struct input_queue {
     struct yasl_input_event events[YASL_QUEUE_CAP];
     int head;
@@ -190,17 +203,17 @@ struct input_queue {
 
 static inline void input_queue_push(struct input_queue *q, const struct yasl_input_event *ev)
 {
-    if (q->count >= YASL_QUEUE_CAP) return;
+    if (unlikely(q->count >= YASL_QUEUE_CAP)) return;
     q->events[q->tail] = *ev;
-    q->tail = (q->tail + 1) % YASL_QUEUE_CAP;
+    q->tail = (q->tail + 1) & YASL_QUEUE_MASK;
     q->count++;
 }
 
 static inline bool input_queue_pop(struct input_queue *q, struct yasl_input_event *ev)
 {
-    if (q->count <= 0) return false;
+    if (unlikely(q->count <= 0)) return false;
     *ev = q->events[q->head];
-    q->head = (q->head + 1) % YASL_QUEUE_CAP;
+    q->head = (q->head + 1) & YASL_QUEUE_MASK;
     q->count--;
     return true;
 }
@@ -211,6 +224,7 @@ static inline void input_queue_clear(struct input_queue *q)
 }
 
 struct yasl_display {
+    /* 8-byte pointers & 64-bit masks */
     struct wl_display *wl;
     struct wl_registry *registry;
     struct wl_compositor *compositor;
@@ -228,7 +242,6 @@ struct yasl_display {
     struct wp_cursor_shape_manager_v1 *cursor_shape_mgr;
     struct wl_cursor_theme *cursor_theme;
     const char *cursor_theme_name;
-    int cursor_size, cursor_theme_scale;
     struct zwp_tablet_manager_v2 *tablet_mgr;
     struct zwp_text_input_manager_v3 *text_input_mgr;
     struct wp_tearing_control_manager_v1 *tearing_mgr;
@@ -240,20 +253,28 @@ struct yasl_display {
     const struct yasl_display_cbs *cbs;
     void *ud;
     char *app_id;
-    clockid_t pres_clock;
-    bool pres_clock_known;
     struct xdg_toplevel_drag_manager_v1 *toplevel_drag_mgr;
     struct conn_drag drag;
     struct yasl_surface *surfaces;
     struct yasl_surface *paste_target;
+    struct yasl_surface *dead;
+
+    /* 4-byte integers */
+    int cursor_size;
+    int cursor_theme_scale;
+    clockid_t pres_clock;
     int watch;
     int depth;
-    struct yasl_surface *dead;
+
+    /* 1-byte booleans */
+    bool owns_loop;
+    bool pres_clock_known;
     bool reading;
     bool lost;
 };
 
 struct yasl_surface {
+    /* 8-byte pointers & 64-bit values */
     struct yasl_display *conn;
     struct yasl_surface *conn_next;
     struct yasl_surface *dead_next;
@@ -263,16 +284,11 @@ struct yasl_surface {
     struct xdg_toplevel *toplevel;
     struct xdg_popup *popup;
     struct zxdg_toplevel_decoration_v1 *toplevel_decor;
-    bool server_decorated;
     struct wp_fractional_scale_v1 *frac_scale;
     struct wp_viewport *viewport;
-
     char *ime_preedit;
     char *ime_commit;
     char *ime_shown;
-    int ime_cursor_begin, ime_cursor_end;
-    int ime_shown_begin, ime_shown_end;
-    struct yasl_rect2i ime_caret;
     struct seat *ime_seat;
     struct wp_tearing_control_v1 *tearing_control;
     struct latency_feedback *latency_feedbacks;
@@ -282,18 +298,23 @@ struct yasl_surface {
     const struct yasl_surface_cbs *cbs;
     void *ud;
     uint64_t frames_skipped;
+    uint64_t outputs;
+    double scale_factor;
 
+    /* Sub-structures */
+    struct yasl_rect2i ime_caret;
+    struct yasl_rect2i content;
     struct yasl_frame_clock clock;
-    struct input_queue input;
-
     struct buffer buffers[NBUF];
     struct yasl_frame frame;
     struct yasl_damage_list damage;
-    bool frame_active;
-    bool frame_presented;
+    int ime_cursor_begin;
+    struct input_queue input;
 
+    /* 4-byte integers */
+    int ime_cursor_end;
+    int ime_shown_begin, ime_shown_end;
     int floating_width, floating_height;
-    struct yasl_rect2i content;
     int cur;
     int width, height;
     int logical_width, logical_height;
@@ -302,9 +323,11 @@ struct yasl_surface {
     uint32_t scale_120;
     uint32_t pending_scale_120;
     uint32_t pending_serial;
-    double scale_factor;
-    uint64_t outputs;
 
+    /* 1-byte booleans packed tightly */
+    bool server_decorated;
+    bool frame_active;
+    bool frame_presented;
     bool tiled;
     bool size_pending;
     bool sample_refresh;
diff --git a/src/wayland/keyboard.c b/src/wayland/keyboard.c
index 7283d58..6045962 100644
--- a/src/wayland/keyboard.c
+++ b/src/wayland/keyboard.c
@@ -346,13 +346,11 @@ static bool mod_index_active(const struct seat *s, xkb_mod_index_t mod)
 
 int seat_modifiers(const struct seat *s)
 {
-    if (!s || !s->xkb_state) return 0;
-    int mods = 0;
-    if (mod_index_active(s, s->mod_shift)) mods |= YASL_MOD_SHIFT;
-    if (mod_index_active(s, s->mod_alt))   mods |= YASL_MOD_ALT;
-    if (mod_index_active(s, s->mod_ctrl))  mods |= YASL_MOD_CTRL;
-    if (mod_index_active(s, s->mod_super)) mods |= YASL_MOD_SUPER;
-    return mods;
+    if (unlikely(!s || !s->xkb_state)) return 0;
+    return (mod_index_active(s, s->mod_shift) * YASL_MOD_SHIFT) |
+           (mod_index_active(s, s->mod_alt)   * YASL_MOD_ALT)   |
+           (mod_index_active(s, s->mod_ctrl)  * YASL_MOD_CTRL)  |
+           (mod_index_active(s, s->mod_super) * YASL_MOD_SUPER);
 }
 
 int yasl_display_modifiers(const struct yasl_display *c)
diff --git a/src/wayland/pointer.c b/src/wayland/pointer.c
index 37e72f3..89dacde 100644
--- a/src/wayland/pointer.c
+++ b/src/wayland/pointer.c
@@ -39,12 +39,44 @@ static struct yasl_vector2i device_from_fixed(const struct yasl_surface *w,
     };
 }
 
-static int motion_button(uint64_t buttons)
-{
-    for (int b = 0; b <= 7; b++)
-        if (buttons & (1ull << b))
-            return b;
-    return 3;
+static inline int motion_button(uint64_t buttons)
+{
+    static const uint8_t motion_btn_lut[256] = {
+        [0] = 3,
+        [1] = 0, [2] = 1, [3] = 0, [4] = 2, [5] = 0, [6] = 1, [7] = 0,
+        [8] = 3, [9] = 0, [10] = 1, [11] = 0, [12] = 2, [13] = 0, [14] = 1, [15] = 0,
+        [16] = 4, [17] = 0, [18] = 1, [19] = 0, [20] = 2, [21] = 0, [22] = 1, [23] = 0,
+        [24] = 3, [25] = 0, [26] = 1, [27] = 0, [28] = 2, [29] = 0, [30] = 1, [31] = 0,
+        [32] = 5, [33] = 0, [34] = 1, [35] = 0, [36] = 2, [37] = 0, [38] = 1, [39] = 0,
+        [40] = 3, [41] = 0, [42] = 1, [43] = 0, [44] = 2, [45] = 0, [46] = 1, [47] = 0,
+        [48] = 4, [49] = 0, [50] = 1, [51] = 0, [52] = 2, [53] = 0, [54] = 1, [55] = 0,
+        [56] = 3, [57] = 0, [58] = 1, [59] = 0, [60] = 2, [61] = 0, [62] = 1, [63] = 0,
+        [64] = 6, [65] = 0, [66] = 1, [67] = 0, [68] = 2, [69] = 0, [70] = 1, [71] = 0,
+        [72] = 3, [73] = 0, [74] = 1, [75] = 0, [76] = 2, [77] = 0, [78] = 1, [79] = 0,
+        [80] = 4, [81] = 0, [82] = 1, [83] = 0, [84] = 2, [85] = 0, [86] = 1, [87] = 0,
+        [88] = 3, [89] = 0, [90] = 1, [91] = 0, [92] = 2, [93] = 0, [94] = 1, [95] = 0,
+        [96] = 5, [97] = 0, [98] = 1, [99] = 0, [100] = 2, [101] = 0, [102] = 1, [103] = 0,
+        [104] = 3, [105] = 0, [106] = 1, [107] = 0, [108] = 2, [109] = 0, [110] = 1, [111] = 0,
+        [112] = 4, [113] = 0, [114] = 1, [115] = 0, [116] = 2, [117] = 0, [118] = 1, [119] = 0,
+        [120] = 3, [121] = 0, [122] = 1, [123] = 0, [124] = 2, [125] = 0, [126] = 1, [127] = 0,
+        [128] = 7, [129] = 0, [130] = 1, [131] = 0, [132] = 2, [133] = 0, [134] = 1, [135] = 0,
+        [136] = 3, [137] = 0, [138] = 1, [139] = 0, [140] = 2, [141] = 0, [142] = 1, [143] = 0,
+        [144] = 4, [145] = 0, [146] = 1, [147] = 0, [148] = 2, [149] = 0, [150] = 1, [151] = 0,
+        [152] = 3, [153] = 0, [154] = 1, [155] = 0, [156] = 2, [157] = 0, [158] = 1, [159] = 0,
+        [160] = 5, [161] = 0, [162] = 1, [163] = 0, [164] = 2, [165] = 0, [166] = 1, [167] = 0,
+        [168] = 3, [169] = 0, [170] = 1, [171] = 0, [172] = 2, [173] = 0, [174] = 1, [175] = 0,
+        [176] = 4, [177] = 0, [178] = 1, [179] = 0, [180] = 2, [181] = 0, [182] = 1, [183] = 0,
+        [184] = 3, [185] = 0, [186] = 1, [187] = 0, [188] = 2, [189] = 0, [190] = 1, [191] = 0,
+        [192] = 6, [193] = 0, [194] = 1, [195] = 0, [196] = 2, [197] = 0, [198] = 1, [199] = 0,
+        [200] = 3, [201] = 0, [202] = 1, [203] = 0, [204] = 2, [205] = 0, [206] = 1, [207] = 0,
+        [208] = 4, [209] = 0, [210] = 1, [211] = 0, [212] = 2, [213] = 0, [214] = 1, [215] = 0,
+        [216] = 3, [217] = 0, [218] = 1, [219] = 0, [220] = 2, [221] = 0, [222] = 1, [223] = 0,
+        [224] = 5, [225] = 0, [226] = 1, [227] = 0, [228] = 2, [229] = 0, [230] = 1, [231] = 0,
+        [232] = 3, [233] = 0, [234] = 1, [235] = 0, [236] = 2, [237] = 0, [238] = 1, [239] = 0,
+        [240] = 4, [241] = 0, [242] = 1, [243] = 0, [244] = 2, [245] = 0, [246] = 1, [247] = 0,
+        [248] = 3, [249] = 0, [250] = 1, [251] = 0, [252] = 2, [253] = 0, [254] = 1, [255] = 0,
+    };
+    return motion_btn_lut[(uint8_t)buttons];
 }
 
 void seat_note_input(struct seat *s, uint32_t serial)
@@ -333,18 +365,25 @@ static void pointer_button(void *data, struct wl_pointer *pointer,
     (void)pointer; (void)time;
     struct seat *s = data;
     struct yasl_surface *w = s->pointer_focus;
-    if (!w) return;
-    int b;
-    switch (button) {
-    case BTN_LEFT: b = 0; break;
-    case BTN_MIDDLE: b = 1; break;
-    case BTN_RIGHT: b = 2; break;
-    case BTN_BACK: b = 3; break;
-    case BTN_FORWARD: b = 4; break;
-    default: return;
-    }
+    if (unlikely(!w)) return;
+
+    static const int8_t btn_map[8] = {
+        [0] = 0, /* BTN_LEFT (0x110) */
+        [1] = 2, /* BTN_RIGHT (0x111) */
+        [2] = 1, /* BTN_MIDDLE (0x112) */
+        [3] = 3, /* BTN_SIDE (0x113) */
+        [4] = 4, /* BTN_EXTRA (0x114) */
+        [5] = 4, /* BTN_FORWARD (0x115) */
+        [6] = 3, /* BTN_BACK (0x116) */
+        [7] = -1,
+    };
+    unsigned idx = button - 0x110;
+    if (unlikely(idx >= 7)) return;
+    int b = btn_map[idx];
+    if (unlikely(b < 0)) return;
+
     seat_note_input(s, serial);
-    bool release = state == WL_POINTER_BUTTON_STATE_RELEASED;
+    bool release = (state == WL_POINTER_BUTTON_STATE_RELEASED);
     if (!release) {
         s->press_serial = serial;
         s->press_surface = w;
diff --git a/src/wayland/shell.c b/src/wayland/shell.c
index ef72512..bc8083a 100644
--- a/src/wayland/shell.c
+++ b/src/wayland/shell.c
@@ -277,24 +277,25 @@ static void toplevel_configure(void *data, struct xdg_toplevel *toplevel,
     (void)toplevel;
     struct yasl_surface *w = data;
 
-    bool fs = false, max = false, act = false, susp = false, tiled = false;
+    uint32_t state_mask = 0;
     uint32_t *s;
     wl_array_for_each(s, states) {
-        switch (*s) {
-        case XDG_TOPLEVEL_STATE_FULLSCREEN:    fs   = true; break;
-        case XDG_TOPLEVEL_STATE_MAXIMIZED:     max  = true; break;
-        case XDG_TOPLEVEL_STATE_ACTIVATED:     act  = true; break;
-        case XDG_TOPLEVEL_STATE_SUSPENDED:     susp = true; break;
-        case XDG_TOPLEVEL_STATE_TILED_LEFT:
-        case XDG_TOPLEVEL_STATE_TILED_RIGHT:
-        case XDG_TOPLEVEL_STATE_TILED_TOP:
-        case XDG_TOPLEVEL_STATE_TILED_BOTTOM:  tiled = true; break;
-        default: break;
-        }
+        if (*s < 32)
+            state_mask |= (1u << *s);
     }
 
-    if (fs != w->fullscreen || max != w->maximized ||
-        act != w->activated || susp != w->suspended) {
+    bool max   = (state_mask & (1u << XDG_TOPLEVEL_STATE_MAXIMIZED)) != 0;
+    bool fs    = (state_mask & (1u << XDG_TOPLEVEL_STATE_FULLSCREEN)) != 0;
+    bool act   = (state_mask & (1u << XDG_TOPLEVEL_STATE_ACTIVATED)) != 0;
+    bool susp  = (state_mask & (1u << XDG_TOPLEVEL_STATE_SUSPENDED)) != 0;
+    bool tiled = (state_mask & (
+        (1u << XDG_TOPLEVEL_STATE_TILED_LEFT) |
+        (1u << XDG_TOPLEVEL_STATE_TILED_RIGHT) |
+        (1u << XDG_TOPLEVEL_STATE_TILED_TOP) |
+        (1u << XDG_TOPLEVEL_STATE_TILED_BOTTOM))) != 0;
+
+    if ((fs ^ w->fullscreen) | (max ^ w->maximized) |
+        (act ^ w->activated) | (susp ^ w->suspended)) {
         w->fullscreen = fs;
         w->maximized = max;
         w->activated = act;
diff --git a/src/wayland/surface.c b/src/wayland/surface.c
index da334e4..5138201 100644
--- a/src/wayland/surface.c
+++ b/src/wayland/surface.c
@@ -83,6 +83,14 @@ static int display_ready(int fd, short revents, void *ud)
         else if (wl_display_read_events(c->wl) < 0)
             revents |= POLLERR;
     }
+    if (revents & POLLOUT) {
+        if (wl_display_flush(c->wl) >= 0) {
+            if (c->loop && c->watch > 0)
+                yasl_loop_set_fd_events(c->loop, c->watch, POLLIN);
+        } else if (errno != EAGAIN) {
+            display_lose(c);
+        }
+    }
     if (wl_display_dispatch_pending(c->wl) < 0 ||
         (revents & (POLLERR | POLLHUP | POLLNVAL)))
         display_lose(c);
@@ -90,35 +98,92 @@ static int display_ready(int fd, short revents, void *ud)
     return 0;
 }
 
-void yasl_display_dispatch(struct yasl_display *c)
-{
-    if (!c || c->lost || c->reading)
-        return;
-    display_enter(c);
-    if (wl_display_dispatch_pending(c->wl) < 0)
-        display_lose(c);
-    display_leave(c);
-}
-
-static void display_arm(struct yasl_display *c)
+static bool display_arm(struct yasl_display *c)
 {
+    bool had_pending = false;
     while (!c->reading && wl_display_prepare_read(c->wl) != 0) {
+        had_pending = true;
         if (wl_display_dispatch_pending(c->wl) < 0) {
             display_lose(c);
-            return;
+            return false;
         }
     }
     c->reading = true;
     short events = POLLIN;
     if (wl_display_flush(c->wl) < 0) {
         if (errno != EAGAIN) {
+            display_cancel_read(c);
             display_lose(c);
-            return;
+            return false;
         }
         events |= POLLOUT;
     }
-    if (c->loop)
+    if (c->loop && c->watch > 0)
         yasl_loop_set_fd_events(c->loop, c->watch, events);
+    return had_pending;
+}
+
+static bool display_has_immediate_work(const struct yasl_display *c)
+{
+    if (!c || c->lost) return false;
+    for (struct yasl_surface *w = c->surfaces; w; w = w->conn_next) {
+        if (w->closed) continue;
+        if (w->input.count > 0)
+            return true;
+        if (w->want_draw && !w->frame_cb && w->configured)
+            return true;
+    }
+    return false;
+}
+
+int yasl_display_dispatch_timeout(struct yasl_display *c, int timeout_ms)
+{
+    if (!c || c->lost)
+        return -1;
+
+    display_enter(c);
+
+    if (wl_display_dispatch_pending(c->wl) < 0) {
+        display_lose(c);
+        display_leave(c);
+        return -1;
+    }
+
+    bool had_pending = display_arm(c);
+    if (c->lost) {
+        display_leave(c);
+        return -1;
+    }
+
+    if (had_pending && timeout_ms != 0 && display_has_immediate_work(c)) {
+        timeout_ms = 0;
+    }
+
+    int n = 0;
+    if (c->loop) {
+        n = yasl_loop_step(c->loop, timeout_ms);
+    }
+
+    if (c->reading) {
+        display_cancel_read(c);
+    }
+
+    if (wl_display_dispatch_pending(c->wl) < 0) {
+        display_lose(c);
+        display_leave(c);
+        return -1;
+    }
+
+    display_leave(c);
+    return n;
+}
+
+void yasl_display_dispatch(struct yasl_display *c)
+{
+    if (!c || c->lost)
+        return;
+    int timeout = display_has_immediate_work(c) ? 0 : -1;
+    yasl_display_dispatch_timeout(c, timeout);
 }
 
 void yasl_display_flush(struct yasl_display *c)
@@ -130,8 +195,10 @@ void yasl_display_flush(struct yasl_display *c)
         next = w->conn_next;
         surface_draw_frame(w);
     }
-    if (!c->lost)
-        display_arm(c);
+    if (!c->lost && c->wl) {
+        if (wl_display_flush(c->wl) < 0 && errno != EAGAIN)
+            display_lose(c);
+    }
     display_leave(c);
 }
 
@@ -163,6 +230,10 @@ static void display_free(struct yasl_display *c)
     if (c->compositor) wl_compositor_destroy(c->compositor);
     if (c->registry) wl_registry_destroy(c->registry);
     if (c->wl) wl_display_disconnect(c->wl);
+    if (c->owns_loop && c->loop) {
+        yasl_loop_destroy(c->loop);
+        c->loop = NULL;
+    }
     free(c->app_id);
     free(c);
 }
@@ -171,7 +242,13 @@ struct yasl_display *yasl_display_open(struct yasl_loop *loop, const struct yasl
 {
     struct yasl_display *c = calloc(1, sizeof(*c));
     if (!c) return NULL;
-    c->loop = loop;
+    if (!loop) {
+        c->loop = yasl_loop_create();
+        c->owns_loop = true;
+    } else {
+        c->loop = loop;
+    }
+    if (!c->loop) goto err;
     if (spec) {
         c->cbs = spec->cbs;
         c->ud = spec->ud;
@@ -189,7 +266,7 @@ struct yasl_display *yasl_display_open(struct yasl_loop *loop, const struct yasl
 
     surface_init_cursor(c);
     if (c->loop) {
-        c->watch = yasl_loop_add_fd(loop, wl_display_get_fd(c->wl), POLLIN, display_ready, c);
+        c->watch = yasl_loop_add_fd(c->loop, wl_display_get_fd(c->wl), POLLIN, display_ready, c);
         if (c->watch <= 0) { c->watch = 0; goto err; }
     }
     if (display_input_init(c) < 0)
@@ -371,7 +448,10 @@ struct yasl_display *yasl_surface_display(const struct yasl_surface *w)
 
 void yasl_surface_schedule_redraw(struct yasl_surface *w)
 {
-    if (w) w->want_draw = true;
+    if (!w || w->closed) return;
+    w->want_draw = true;
+    if (w->conn && w->conn->loop)
+        yasl_loop_kick(w->conn->loop);
 }
 
 struct yasl_vector2i yasl_surface_size(const struct yasl_surface *w)