// Not clipped
void draw_hline(unsigned x0, unsigned x1, unsigned y, u32 color_x4) {
	if (x0 > x1) {
		unsigned tmp = x0;
		x0 = x1;
		x1 = tmp;
	}
	
	u32 *dest = (u32*) get_framebuffer() + (y << 6) - (y << 2) + (x0 >> 2);
	unsigned shift = (x0 & 3) << 3, mask, value;

	// Small line
	if ((x1 >> 2) == (x0 >> 2)) {
		mask = 0xffffffffU << shift;
		shift = 32 - ((x1 & 3) << 3);
		mask &= 0xffffffffU >> shift;
		*dest = (*dest & ~mask) | (color_x4 & mask);
		return;
	}
	
	// First block of 4 pixels
	if (x0 & 3) {
		mask = ~(0xffffffffU << shift);
		value = color_x4 << shift;
		*dest = (*dest & mask) | value;
		dest++;
		x0 += 4 - (x0 & 3);
	}
	
	// Loop (could be unrolled)
	unsigned count = (x1 - x0) >> 2;
	while (count-- > 0)
		*dest++ = color_x4;

	// Last block
	if (x1 & 3) {
		shift = 32 - ((x1 & 3) << 3);
		mask = ~(0xffffffffU >> shift);
		value = color_x4 >> shift;
		*dest = (*dest & mask) | value;
	}
}