extern char __heap_base;

static int allocated_static;
static int allocated_dynamic;

void
free_static(void)
{
    allocated_static = 0;
}

void
free_dynamic(void)
{
    allocated_dynamic = 0;
}

void *
alloc_static(int size)
{
    void *result = &__heap_base + allocated_static;
    allocated_static += size;
    return(result);
}

void *
alloc_dynamic(int size)
{
    void *result = &__heap_base + allocated_static + allocated_dynamic;
    allocated_dynamic += size;
    return(result);
}

static int
rdp_find_max(float *coordinates, float zoom, int coords_from, 
             int segment_start, int segment_end)
{
    float EPS = 1.0 / zoom;
    
    int result = -1;
    float max_dist = 0.0f;
    
    float ax = coordinates[coords_from + segment_start * 2 + 0];
    float ay = coordinates[coords_from + segment_start * 2 + 1];
    float bx = coordinates[coords_from   + segment_end * 2 + 0];
    float by = coordinates[coords_from   + segment_end * 2 + 1];
    
    float dx = bx - ax;
    float dy = by - ay;
    
    float dist_ab = __builtin_sqrtf(dx * dx + dy * dy);
    float dir_nx = dy / dist_ab;
    float dir_ny = -dx / dist_ab;
    
    for (int i = segment_start + 1; i < segment_end; ++i) {
        float px = coordinates[coords_from + i * 2 + 0];
        float py = coordinates[coords_from + i * 2 + 1];
        
        float apx = px - ax;
        float apy = py - ay;
        
        float dist = __builtin_fabsf(apx * dir_nx + apy * dir_ny);
        
        if (dist > EPS && dist > max_dist) {
            result = i;
            max_dist = dist;
        }
    }
    
    return(result);
}

int
do_lod(int *clipped_indices, int clipped_count, float zoom, 
       int *stroke_coords_from,
       float *line_threshold, 
       float *coordinates,
       unsigned char *pressures,
       int coordinates_count)
{
    if (clipped_count == 0) {
        return(0);
    }

    int *segments_from = alloc_dynamic((clipped_count + 1) * 4);
    int *segments = alloc_dynamic(coordinates_count / 2 * 4);
    
    int segments_head = 0;
    int stack[4096]; // TODO: what's a reasonable max size for this?
    
    for (int i = 0; i < clipped_count; ++i) {
        int stroke_index = clipped_indices[i];
        
        // TODO: convert to a proper CSR, save half the memory
        int coords_from = stroke_coords_from[stroke_index];
        int coords_to = stroke_coords_from[stroke_index + 1];
        
        int point_count = (coords_to - coords_from) / 2;
        
        // Basic CSR crap
        segments_from[i] = segments_head;
        
        if (zoom < line_threshold[stroke_index]) {
            // Fast paths for collapsing to a single line segment
            segments[segments_head++] = 0;
            segments[segments_head++] = point_count - 1;
            continue;
        }
        
        int segment_count = 2;
        int stack_head = 0;
        
        segments[segments_head++] = 0;
        
        stack[stack_head++] = 0;
        stack[stack_head++] = 0;
        stack[stack_head++] = point_count - 1;
        
        while (stack_head > 0) {
            int end = stack[--stack_head];
            int start = stack[--stack_head];
            int type = stack[--stack_head];
            
            if (type == 1) {
                segments[segments_head++] = start;
            } else {
                int max = rdp_find_max(coordinates, zoom, coords_from, start, end);
                if (max != -1) {
                    segment_count += 1;
                    
                    stack[stack_head++] = 0;
                    stack[stack_head++] = max;
                    stack[stack_head++] = end;
                    
                    stack[stack_head++] = 1;
                    stack[stack_head++] = max;
                    stack[stack_head++] = -1;
                    
                    stack[stack_head++] = 0;
                    stack[stack_head++] = start;
                    stack[stack_head++] = max;
                }
            }
        }
        
        segments[segments_head++] = point_count - 1;
        
        if (segment_count == 2 && zoom > line_threshold[stroke_index]) {
            line_threshold[stroke_index] = zoom;
        }
    }

    segments_from[clipped_count] = segments_head;

    // Write actual coordinates (points) and stroke ids
    float *points = alloc_dynamic(segments_head * 2 * 4);
    int *ids = alloc_dynamic(segments_head * 4);
    unsigned char *pressures_res = alloc_dynamic(segments_head);

    int phead = 0;
    int ihead = 0;

    for (int i = 0; i < clipped_count; ++i) {
        int stroke_index = clipped_indices[i];
        
        // TODO: convert to a proper CSR, save half the memory
        int base_stroke = stroke_coords_from[stroke_index];
        int from = segments_from[i];
        int to = segments_from[i + 1];

        for (int j = from; j < to; ++j) {
            int point_index = segments[j];
            float x = coordinates[base_stroke + point_index * 2 + 0];
            float y = coordinates[base_stroke + point_index * 2 + 1];

            points[phead++] = x;
            points[phead++] = y;

            pressures_res[ihead] = pressures[base_stroke / 2 + point_index];

            if (j != to - 1) {
                ids[ihead++] = stroke_index;
            } else {
                ids[ihead++] = stroke_index | (1 << 31);
            }
        }
    }
    
    return(segments_head);
}