1 files changed, 41 insertions, 9 deletions

diff --git a/src/user/app/vterm/draw.c b/src/user/app/vterm/draw.c
index 44d3ffe..1a40ede 100644
--- a/src/user/app/vterm/draw.c
+++ b/src/user/app/vterm/draw.c
@@ -1,4 +1,5 @@
 #include "vterm.h"
+#include <camellia/execbuf.h>
 #include <camellia/syscalls.h>
 #include <string.h>
 
@@ -14,19 +15,50 @@ void dirty_mark(uint32_t x, uint32_t y) {
 	if (dirty.y2 < y) dirty.y2 = y;
 }
 
+static void flush_combined(struct rect pix) {
+	size_t low  = fb.pitch * pix.y1 + 4 * pix.x1;
+	size_t high = fb.pitch * pix.y2 + 4 * pix.y2 + 4;
+
+	_syscall_write(fb_fd, fb.b + low, high - low, low, 0);
+}
+
+static void flush_split(struct rect pix) {
+	static uint64_t execbuf[EXECBUF_MAX_LEN / sizeof(uint64_t)];
+	if (7 * (pix.y2 - pix.y1) * sizeof(uint64_t) >= sizeof execbuf) {
+		flush_combined(pix);
+		return;
+	}
+
+	size_t epos = 0;
+	for (uint32_t y = pix.y1; y < pix.y2; y++) {
+		size_t low  = fb.pitch * y + 4 * pix.x1;
+		size_t high = fb.pitch * y + 4 * pix.x2 + 4;
+
+		execbuf[epos++] = EXECBUF_SYSCALL;
+		execbuf[epos++] = _SYSCALL_WRITE;
+		execbuf[epos++] = fb_fd;
+		execbuf[epos++] = fb.b + low;
+		execbuf[epos++] = high - low;
+		execbuf[epos++] = low;
+		execbuf[epos++] = 0;
+	}
+	_syscall_execbuf(execbuf, epos * sizeof(uint64_t));
+}
+
 void flush(void) {
-	size_t low, high;
 	if (~dirty.x1 == 0) return;
 
-	// still not optimal, copies 80 characters instead of 1
-	low  = fb.pitch * ( dirty.y1    * font.h) + 4 * ( dirty.x1    * font.w);
-	high = fb.pitch * ((dirty.y2+1) * font.h) + 4 * ((dirty.x2+1) * font.w) + 3;
-
-	if (~dirty.y2 == 0) high = fb.len;
-	if (high > fb.len)  high = fb.len;
+	struct rect pix;
+	pix.x1 = dirty.x1 * font.w;
+	pix.x2 = dirty.x2 * font.w + font.w - 1;
+	pix.y1 = dirty.y1 * font.h;
+	pix.y2 = dirty.y2 * font.h + font.h - 1;
+	if (pix.x2 >= fb.width)  pix.x2 = fb.width - 1;
+	if (pix.y2 >= fb.height) pix.y2 = fb.height - 1;
 
-	while (low < high)
-		low += _syscall_write(fb_fd, fb.b + low, high - low, low, 0);
+	/* the threshold is mostly arbitrary, wasn't based on any real benchmarks */
+	if (pix.x2 - pix.x1 > fb.width - 600) flush_combined(pix);
+	else flush_split(pix);
 
 	dirty.x1 = ~0; dirty.y1 = ~0;
 	dirty.x2 =  0; dirty.y2 =  0;