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

1 files changed, 25 insertions, 3 deletions

diff --git a/src/x11/window.c b/src/x11/window.c
index 5e1b612..271c128 100644
--- a/src/x11/window.c
+++ b/src/x11/window.c
@@ -1280,10 +1280,32 @@ static bool present(struct yasl_surface *base, const struct yasl_framebuffer *fb
     XImage *im = s->image;
     bool direct = im->bits_per_pixel == 32 && im->red_mask == 0xff0000 &&
                   im->green_mask == 0xff00 && im->blue_mask == 0xff;
-    for (int y = 0; y < s->size.y; y++) {
+    int y_start = 0, y_end = s->size.y;
+    int x_start = 0, x_end = s->size.x;
+    if (damage->n >= 0) {
+        y_start = s->size.y; y_end = 0;
+        x_start = s->size.x; x_end = 0;
+        for (int i = 0; i < damage->n; i++) {
+            struct yasl_rect2i r = damage->rects[i];
+            int y0 = r.y > 0 ? r.y : 0;
+            int y1 = (r.y + r.height < s->size.y) ? r.y + r.height : s->size.y;
+            int x0 = r.x > 0 ? r.x : 0;
+            int x1 = (r.x + r.width < s->size.x) ? r.x + r.width : s->size.x;
+            if (y0 < y_start) y_start = y0;
+            if (y1 > y_end) y_end = y1;
+            if (x0 < x_start) x_start = x0;
+            if (x1 > x_end) x_end = x1;
+        }
+        if (y_start >= y_end || x_start >= x_end) {
+            y_start = y_end = x_start = x_end = 0;
+        }
+    }
+    size_t copy_bytes = (size_t)(x_end - x_start) * 4;
+    for (int y = y_start; y < y_end; y++) {
         const uint32_t *row = s->pixels + (size_t)y * s->size.x;
-        if (direct) memcpy(im->data + (size_t)y * im->bytes_per_line, row, (size_t)s->size.x * 4);
-        else for (int x = 0; x < s->size.x; x++) {
+        if (direct) memcpy(im->data + (size_t)y * im->bytes_per_line + (size_t)x_start * 4,
+                           row + x_start, copy_bytes);
+        else for (int x = x_start; x < x_end; x++) {
             uint32_t p = row[x];
             XPutPixel(im, x, y, component((p >> 16) & 255, im->red_mask) |
                     component((p >> 8) & 255, im->green_mask) | component(p & 255, im->blue_mask));