static int
scanline_intersects_line(f32 y, struct v2f p0, struct v2f p1, f32 lasty, f32 *vx)
{
    bool goes_up = (p0.y > p1.y);
    f32 t1 = (y - p0.y) / (p1.y - p0.y); /* NOTE(aolo2): no horizontal lines by design */
    
    if (t1 == 0) {
        f32 x1 = p0.x + (p1.x - p0.x) * t1;
        if ((lasty < p0.y) && (p0.y > p1.y)) return(0);
        if ((lasty > p0.y) && (p0.y < p1.y)) return(0);
        int result = (goes_up ? 1 : -1);
        *vx = x1;
        return(result);
    }
    
    if (0 < t1 && t1 < 1.0f) {
        f32 x1 = p0.x + (p1.x - p0.x) * t1;
        *vx = x1;
        int result = (goes_up ? 1 : -1);
        return(result);
    }
    
    return(0);
}


static int
intersect_glyph(struct line_contour *lines, f32 y, struct intersection *intersections)
{
    int nints = 0;
    
    for (int c = 0; c < lines->ncontours; ++c) {
        int from = lines->from[c];
        int to = lines->from[c + 1];
        
        for (int i = from; i < to; ++i) {
            int lasti = (i > from ? i - 1 : to - 1);
            f32 lasty = lines->data[lasti].a.y;
            f32 vx;
            int r = scanline_intersects_line(y, lines->data[i].a, lines->data[i].b, lasty, &vx);
            
            if (r) {
                intersections[nints].x = vx;
                intersections[nints].dir = r;
                ++nints;
            }
        }
    }
    
    return(nints);
}


static void
sort_intersections(struct intersection *intersections, int size)
{
    bool swapped = true;
    
    while (swapped) {
        swapped = false;
        for (int i = 0; i < size - 1; ++i) {
            f32 x1 = intersections[i].x;
            f32 x2 = intersections[i + 1].x;
            if (x1 > x2) {
                struct intersection tmp = intersections[i];
                intersections[i] = intersections[i + 1];
                intersections[i + 1] = tmp;
                swapped = true;
            }
        }
    }
}

static void
render_glyph(struct glyph g, int px_size, struct line_contour *lines,
             f32 scale, u32 *pixels, int width, int at_x, int at_y)
{
    int oversample_y = 4;
    if (px_size <= 12) {
        oversample_y = 8;
    }
    
    f32 oversample_step = 1.0f / (oversample_y + 1);
    f32 oversample_norm = 1.0f / oversample_y;
    
    int gwidth = ceil_f32((g.xmax - g.xmin + g.lsb) * scale);
    int gheight = px_size;
    
    struct intersection *intersections = malloc(lines->from[lines->ncontours] * sizeof(struct intersection));
    
    f32 *accumulator = calloc(1, gwidth * gheight * sizeof(f32));
    
    for (int y = 0; y < gheight; ++y) {
        for (int yy = 1; yy <= oversample_y; ++yy) {
            u32 ncross = intersect_glyph(lines, y + oversample_step * yy, intersections);
            if (ncross) {
                sort_intersections(intersections, ncross);
                
                int state = 0;
                
                for (u32 i = 0; i < ncross - 1; ++i) {
                    struct intersection inter = intersections[i];
                    struct intersection next_inter = intersections[i + 1];
                    
                    state += inter.dir;
                    
                    if (state != 0) {
                        f32 x0 = inter.x;
                        f32 x1 = next_inter.x;
                        
                        int x_from = x0;
                        int x_to = (x1 > gwidth - 1 ? gwidth - 1 : x1);
                        
                        f32 start_brightness = (x_from + 1 - x0);
                        f32 end_brightness = (x1 - x_to);
                        
                        for (int x = x_from + 1; x < x_to; ++x) {
                            accumulator[y * gwidth + x] += oversample_norm;
                        }
                        
                        accumulator[y * gwidth + x_from] += start_brightness * oversample_norm;
                        accumulator[y * gwidth + x_to] += end_brightness * oversample_norm;
                    }
                }
            }
        }
    }
    
    for (int y = 0; y < gheight; ++y) {
        for (int x = 0; x < gwidth; ++x) {
            //printf(" %.2f", accumulator[y * gwidth + x]);
            u32 brightness = clamp_u32(accumulator[y * gwidth + x] * 255.99f, 255);
            u32 white = 0xFF000000 | brightness << 16 | brightness << 8 | brightness;
            
            if (brightness > 0) {
                pixels[(at_y + (gheight - 1 - y)) * width + (at_x + x)] = white;
            } else {
                //pixels[(at_y + (gheight - 1 - y)) * width + (at_x + x)] = 0xFFFF0000;
            }
        }
    }
    
    free(accumulator);
    
    //exit(1);
    
    free(intersections);
}

static void
outline_to_lines(struct glyph g, f32 scale, int max_descent, struct line_contour *dest, int *cnt) 
{
    int nlines = 0;
    int points_from = 0;
    int curve_segments = 5;
    
    for (int c = 0; c < g.ncontours; ++c) {
        for (int p = points_from; p < g.end_pts_of_contours[c] + 1; ++p) {
            struct glyph_point gp = g.points[p];
            
            int nexti = (p + 1 < g.end_pts_of_contours[c] + 1 ? p + 1 : points_from);
            struct glyph_point nextgp = g.points[nexti];
            
            if (p == points_from && !gp.on_curve) {
                continue;
            }
            
            f32 x1 = (gp.x + g.lsb) * scale;
            f32 y1 = (gp.y - g.baseline - max_descent) * scale;
            
            f32 x2 = (nextgp.x + g.lsb) * scale;
            f32 y2 = (nextgp.y - g.baseline - max_descent) * scale;
            
            if (nextgp.on_curve) {
                if (gp.y != nextgp.y) {
                    if (dest->data) {
                        dest->data[nlines].a = (struct v2f) { x1, y1 };
                        dest->data[nlines].b = (struct v2f) { x2, y2 };
                    }
                    ++nlines;
                }
            } else {
                int nextnexti = (nexti + 1 < g.end_pts_of_contours[c] + 1 ? nexti + 1 : points_from);
                struct glyph_point nextnextgp = g.points[nextnexti];
                
                f32 x3 = (nextnextgp.x + g.lsb) * scale;
                f32 y3 = (nextnextgp.y - g.baseline - max_descent) * scale;
                
                /* P(t) = P0*t^2 + P1*2*t*(1-t) + P2*(1-t)^2 */
                f32 t_step = 1.0f / curve_segments;
                f32 x_prev = x1;
                f32 y_prev = y1;
                
                /* s = 1 for exact beginning */
                for (int s = 1; s <= curve_segments; ++s) {
                    f32 t_now = t_step * s;
                    f32 x_now;
                    f32 y_now;
                    
                    if (s < curve_segments) {
                        x_now = x3 * t_now * t_now 
                            + x2 * 2.0f * t_now * (1.0f - t_now) 
                            + x1 * (1.0f - t_now) * (1.0f - t_now);
                        
                        y_now = y3 * t_now * t_now 
                            + y2 * 2.0f * t_now * (1.0f - t_now) 
                            + y1 * (1.0f - t_now) * (1.0f - t_now);
                    } else {
                        /* For exact match between neighbours */
                        x_now = x3;
                        y_now = y3;
                    }
                    
                    if (abs_f32(y_now - y_prev) > F32EPS) {
                        if (dest->data) {
                            dest->data[nlines].a = (struct v2f) { x_prev, y_prev };
                            dest->data[nlines].b = (struct v2f) { x_now, y_now };
                        }
                        ++nlines;
                    }
                    
                    x_prev = x_now;
                    y_prev = y_now;
                }
                
                ++p;
            }
        }
        
        dest->from[c + 1] = nlines;
        
        points_from = g.end_pts_of_contours[c] + 1;
    }
    
    dest->ncontours = g.ncontours;
    
    if (cnt) {
        *cnt = nlines;
    }
}


static void
render_utf_string(struct ttf_font font, int px_size, u32 *pixels, u32 width, wchar_t *string, int at_x, int at_y)
{
    u32 offset_x = at_x;
    u32 offset_y = at_y;
    f32 scale = (f32) px_size / ((f32) (font.hhea.ascent - font.hhea.descent));
    
    u32 len = wcslen(string);
    
    for (u32 i = 0; i < len; ++i) {
        u16 codepoint = string[i];
        if (codepoint != ' ') {
            struct glyph g = get_outline(&font, codepoint);
            
            //exit(0);
            
            struct line_contour lines = { 0 };
            int nlines = 0;
            outline_to_lines(g, scale, font.hhea.descent, &lines, &nlines);
            lines.data = malloc(nlines * sizeof(struct line));
            outline_to_lines(g, scale, font.hhea.descent, &lines, 0);
            
            render_glyph(g, px_size, &lines, scale, pixels, width, offset_x, offset_y);
            
            offset_x += ceil_f32(scale * g.advance);
            
            free(lines.data);
        } else {
            int space_advance = get_codepoint_width(&font, scale, codepoint);
            offset_x += space_advance;
        }
        
        //XPutImage(display, window, default_gc, xwindow_buffer, 0, 0, 0, 0, width, height);
        //sleep(1);
    }
}