Speed up drawing of printable ascii chars

kitty/line.c

@@ -647,15 +647,6 @@ line_get_char(Line *self, index_type at) {
     return ch;
 }
 
-void
-line_set_printable_ascii_chars(Line *self, unsigned int at, const uint8_t *chars, unsigned num, GPUCell g, CPUCell cc) {
-    for (unsigned i = at; i < at + num; i++, chars++) {
-        cc.ch = *chars;
-        self->gpu_cells[i] = g;
-        self->cpu_cells[i] = cc;
-    }
-}
-
 void
 line_set_char(Line *self, unsigned int at, uint32_t ch, unsigned int width, Cursor *cursor, hyperlink_id_type hyperlink_id) {
     GPUCell *g = self->gpu_cells + at;

kitty/lineops.h

@@ -86,7 +86,6 @@ void line_clear_text(Line *self, unsigned int at, unsigned int num, char_type ch
 void line_apply_cursor(Line *self, Cursor *cursor, unsigned int at, unsigned int num, bool clear_char);
 char_type line_get_char(Line *self, index_type at);
 void line_set_char(Line *, unsigned int , uint32_t , unsigned int , Cursor *, hyperlink_id_type);
-void line_set_printable_ascii_chars(Line *self, unsigned int at, const uint8_t *chars, unsigned num, GPUCell g, CPUCell cc);
 void line_right_shift(Line *, unsigned int , unsigned int );
 void line_add_combining_char(Line *, uint32_t , unsigned int );
 index_type line_url_start_at(Line *self, index_type x);

kitty/screen.c

@@ -661,7 +661,7 @@ screen_draw_printable_ascii(Screen *self, const uint8_t *chars, size_t num) {
     screen_on_input(self);
     self->last_graphic_char = chars[num-1];
     self->is_dirty = true;
-    CPUCell cc = {.hyperlink_id=self->active_hyperlink_id};
+    const CPUCell cc = {.hyperlink_id=self->active_hyperlink_id};
     GPUCell g = {.attrs=cursor_to_attrs(self->cursor, 1), .fg=self->cursor->fg & COL_MASK, .bg=self->cursor->bg & COL_MASK, .decoration_fg=self->cursor->decoration_fg & COL_MASK};
     if (OPT(underline_hyperlinks) == UNDERLINE_ALWAYS && cc.hyperlink_id) {
         g.decoration_fg = ((OPT(url_color) & COL_MASK) << 8) | 2;
@@ -671,8 +671,15 @@ screen_draw_printable_ascii(Screen *self, const uint8_t *chars, size_t num) {
 #define fill_single_line(chars, num) { \
         linebuf_init_line(self->linebuf, self->cursor->y); \
         if (self->modes.mIRM) line_right_shift(self->linebuf->line, self->cursor->x, num); \
-        line_set_printable_ascii_chars(self->linebuf->line, self->cursor->x, chars, num, g, cc); \
-        self->cursor->x += num; \
+        const unsigned limit = self->cursor->x + num; \
+        const uint8_t *p = chars; \
+        GPUCell *gp = self->linebuf->line->gpu_cells; \
+        CPUCell *cp = self->linebuf->line->cpu_cells; \
+        for (; self->cursor->x < limit; self->cursor->x++, p++) { \
+            gp[self->cursor->x] = g; \
+            cp[self->cursor->x] = cc; \
+            cp[self->cursor->x].ch = *p; \
+        } \
         if (selection_has_screen_line(&self->selections, self->cursor->y)) clear_selection(&self->selections); \
         linebuf_mark_line_dirty(self->linebuf, self->cursor->y); \
 }

kitty/vt-parser.c

@@ -94,6 +94,9 @@ _report_params(PyObject *dump_callback, id_type window_id, const char *name, int
 #define REPORT_DRAW(ch) \
     Py_XDECREF(PyObject_CallFunction(self->dump_callback, "KsC", self->window_id, "draw", ch)); PyErr_Clear();
 
+#define REPORT_DRAW_ASCII(ch, sz) \
+    Py_XDECREF(PyObject_CallFunction(self->dump_callback, "Kss#", self->window_id, "draw", ch, (Py_ssize_t)sz)); PyErr_Clear();
+
 #define REPORT_PARAMS(name, params, num, is_group, region) _report_params(self->dump_callback, self->window_id, name, params, num_params, is_group, region)
 
 #define REPORT_OSC(name, string) \
@@ -110,6 +113,7 @@ _report_params(PyObject *dump_callback, id_type window_id, const char *name, int
 #define REPORT_COMMAND(...)
 #define REPORT_VA_COMMAND(...)
 #define REPORT_DRAW(ch)
+#define REPORT_DRAW_ASCII(ch, sz)
 #define REPORT_PARAMS(...)
 #define REPORT_OSC(name, string)
 #define REPORT_OSC2(name, code, string)
@@ -272,10 +276,9 @@ dispatch_normal_mode_byte(PS *self, uint8_t ch) {
 
 static void
 dispatch_printable_ascii(PS *self, const size_t sz) {
-    for (const size_t limit = self->read.pos + sz; self->read.pos < limit; self->read.pos++) {
-        REPORT_DRAW(self->buf[self->read.pos]);
-        screen_draw(self->screen, self->buf[self->read.pos]);
-    }
+    REPORT_DRAW_ASCII(self->buf + self->read.pos, sz);
+    screen_draw_printable_ascii(self->screen, self->buf + self->read.pos, sz);
+    self->read.pos += sz;
 }
 
 static void