Prune in single pass to speed up. Add playout threshold to limit pruning time.

[pachi/derm.git] / uct / tree.c

#include <assert.h>
#include <math.h>
#include <stddef.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

#define DEBUG
#include "board.h"
#include "debug.h"
#include "engine.h"
#include "move.h"
#include "playout.h"
#include "tactics.h"
#include "timeinfo.h"
#include "uct/internal.h"
#include "uct/prior.h"
#include "uct/tree.h"


/* Allocate one node in the fast_alloc mode. The returned node
 * is _not_ initialized. Returns NULL if not enough memory.
 * This function may be called by multiple threads in parallel. */
static struct tree_node *
tree_fast_alloc_node(struct tree *t)
{
        assert(t->nodes != NULL);
        struct tree_node *n = NULL;
        unsigned long old_size =__sync_fetch_and_add(&t->nodes_size, sizeof(*n));

        /* The test below works even if max_tree_size is not a
         * multiple of the node size because tree_init() allocates
         * space for an extra node. */
        if (old_size < t->max_tree_size)
                n = (struct tree_node *)(t->nodes + old_size);
        return n;
}

/* Allocate and initialize a node. Returns NULL (fast_alloc mode)
 * or exits the main program if not enough memory.
 * This function may be called by multiple threads in parallel. */
static struct tree_node *
tree_init_node(struct tree *t, coord_t coord, int depth)
{
        struct tree_node *n;
        if (t->nodes) {
                n = tree_fast_alloc_node(t);
                if (!n) return n;
                memset(n, 0, sizeof(*n));
        } else {
                n = calloc(1, sizeof(*n));
                if (!n) {
                        fprintf(stderr, "tree_init_node(): OUT OF MEMORY\n");
                        exit(1);
                }
                __sync_fetch_and_add(&t->nodes_size, sizeof(*n));
        }
        n->coord = coord;
        n->depth = depth;
        volatile static long c = 1000000;
        n->hash = __sync_fetch_and_add(&c, 1);
        if (depth > t->max_depth)
                t->max_depth = depth;
        return n;
}

/* Create a tree structure. Pre-allocate all nodes if max_tree_size is > 0. */
struct tree *
tree_init(struct board *board, enum stone color, unsigned long max_tree_size)
{
        struct tree *t = calloc(1, sizeof(*t));
        t->board = board;
        t->max_tree_size = max_tree_size;
        if (max_tree_size != 0) {
                /* Allocate one extra node, max_tree_size may not be multiple of node size. */
                t->nodes = malloc(max_tree_size + sizeof(struct tree_node));
                /* The nodes buffer doesn't need initialization. This is currently
                 * done by tree_init_node to spread the load. Doing a memset for the
                 * entire buffer here would be too slow for large trees (>10 GB). */
                if (!t->nodes) {
                        fprintf(stderr, "tree_init(): OUT OF MEMORY\n");
                        exit(1);
                }
        }
        /* The root PASS move is only virtual, we never play it. */
        t->root = tree_init_node(t, pass, 0);
        t->root_symmetry = board->symmetry;
        t->root_color = stone_other(color); // to research black moves, root will be white
        return t;
}


/* This function may be called by multiple threads in parallel on the
 * same tree, but not on node n. n may be detached from the tree but
 * must have been created in this tree originally.
 * It returns the remaining size of the tree after n has been freed. */
static unsigned long
tree_done_node(struct tree *t, struct tree_node *n)
{
        struct tree_node *ni = n->children;
        while (ni) {
                struct tree_node *nj = ni->sibling;
                tree_done_node(t, ni);
                ni = nj;
        }
        free(n);
        unsigned long old_size = __sync_fetch_and_sub(&t->nodes_size, sizeof(*n));
        return old_size - sizeof(*n);
}

struct subtree_ctx {
        struct tree *t;
        struct tree_node *n;
};

/* Worker thread for tree_done_node_detached(). Only for fast_alloc=false. */
static void *
tree_done_node_worker(void *ctx_)
{
        struct subtree_ctx *ctx = ctx_;
        char *str = coord2str(ctx->n->coord, ctx->t->board);

        unsigned long tree_size = tree_done_node(ctx->t, ctx->n);
        if (!tree_size)
                free(ctx->t);
        if (DEBUGL(2))
                fprintf(stderr, "done freeing node at %s, tree size %lu\n", str, tree_size);
        free(str);
        free(ctx);
        return NULL;
}

/* Asynchronously free the subtree of nodes rooted at n. If the tree becomes
 * empty free the tree also.  Only for fast_alloc=false. */
static void
tree_done_node_detached(struct tree *t, struct tree_node *n)
{
        if (n->u.playouts < 1000) { // no thread for small tree
                if (!tree_done_node(t, n))
                        free(t);
                return;
        }
        pthread_attr_t attr;
        pthread_attr_init(&attr);
        pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);

        pthread_t thread;
        struct subtree_ctx *ctx = malloc(sizeof(struct subtree_ctx));
        if (!ctx) {
                fprintf(stderr, "tree_done_node_detached(): OUT OF MEMORY\n");
                exit(1);
        }
        ctx->t = t;
        ctx->n = n;
        pthread_create(&thread, &attr, tree_done_node_worker, ctx);
        pthread_attr_destroy(&attr);
}

void
tree_done(struct tree *t)
{
        if (t->chchvals) free(t->chchvals);
        if (t->chvals) free(t->chvals);
        if (t->nodes) {
                free(t->nodes);
                free(t);
        } else if (!tree_done_node(t, t->root)) {
                free(t);
                /* A tree_done_node_worker might still be running on this tree but
                 * it will free the tree later. It is also freeing nodes faster than
                 * we will create new ones. */
        }
}


static void
tree_node_dump(struct tree *tree, struct tree_node *node, int l, int thres)
{
        for (int i = 0; i < l; i++) fputc(' ', stderr);
        int children = 0;
        for (struct tree_node *ni = node->children; ni; ni = ni->sibling)
                children++;
        /* We use 1 as parity, since for all nodes we want to know the
         * win probability of _us_, not the node color. */
        fprintf(stderr, "[%s] %f %% %d [prior %f %% %d amaf %f %% %d]; hints %x; %d children <%"PRIhash">\n",
                coord2sstr(node->coord, tree->board),
                tree_node_get_value(tree, 1, node->u.value), node->u.playouts,
                tree_node_get_value(tree, 1, node->prior.value), node->prior.playouts,
                tree_node_get_value(tree, 1, node->amaf.value), node->amaf.playouts,
                node->hints, children, node->hash);

        /* Print nodes sorted by #playouts. */

        struct tree_node *nbox[1000]; int nboxl = 0;
        for (struct tree_node *ni = node->children; ni; ni = ni->sibling)
                if (ni->u.playouts > thres)
                        nbox[nboxl++] = ni;

        while (true) {
                int best = -1;
                for (int i = 0; i < nboxl; i++)
                        if (nbox[i] && (best < 0 || nbox[i]->u.playouts > nbox[best]->u.playouts))
                                best = i;
                if (best < 0)
                        break;
                tree_node_dump(tree, nbox[best], l + 1, /* node->u.value < 0.1 ? 0 : */ thres);
                nbox[best] = NULL;
        }
}

void
tree_dump_chval(struct tree *tree, struct move_stats *v)
{
        for (int y = board_size(tree->board) - 2; y > 1; y--) {
                for (int x = 1; x < board_size(tree->board) - 1; x++) {
                        coord_t c = coord_xy(tree->board, x, y);
                        fprintf(stderr, "%.2f%%%05d  ", v[c].value, v[c].playouts);
                }
                fprintf(stderr, "\n");
        }
}

void
tree_dump(struct tree *tree, int thres)
{
        if (thres && tree->root->u.playouts / thres > 100) {
                /* Be a bit sensible about this; the opening book can create
                 * huge dumps at first. */
                thres = tree->root->u.playouts / 100 * (thres < 1000 ? 1 : thres / 1000);
        }
        fprintf(stderr, "(UCT tree; root %s; extra komi %f)\n",
                stone2str(tree->root_color), tree->extra_komi);
        tree_node_dump(tree, tree->root, 0, thres);

        if (DEBUGL(3) && tree->chvals) {
                fprintf(stderr, "children stats:\n");
                tree_dump_chval(tree, tree->chvals);
                fprintf(stderr, "grandchildren stats:\n");
                tree_dump_chval(tree, tree->chchvals);
        }
}


static char *
tree_book_name(struct board *b)
{
        static char buf[256];
        if (b->handicap > 0) {
                sprintf(buf, "uctbook-%d-%02.01f-h%d.pachitree", b->size - 2, b->komi, b->handicap);
        } else {
                sprintf(buf, "uctbook-%d-%02.01f.pachitree", b->size - 2, b->komi);
        }
        return buf;
}

static void
tree_node_save(FILE *f, struct tree_node *node, int thres)
{
        bool save_children = node->u.playouts >= thres;

        if (!save_children)
                node->is_expanded = 0;

        fputc(1, f);
        fwrite(((void *) node) + offsetof(struct tree_node, depth),
               sizeof(struct tree_node) - offsetof(struct tree_node, depth),
               1, f);

        if (save_children) {
                for (struct tree_node *ni = node->children; ni; ni = ni->sibling)
                        tree_node_save(f, ni, thres);
        } else {
                if (node->children)
                        node->is_expanded = 1;
        }

        fputc(0, f);
}

void
tree_save(struct tree *tree, struct board *b, int thres)
{
        char *filename = tree_book_name(b);
        FILE *f = fopen(filename, "wb");
        if (!f) {
                perror("fopen");
                return;
        }
        tree_node_save(f, tree->root, thres);
        fputc(0, f);
        fclose(f);
}


void
tree_node_load(FILE *f, struct tree_node *node, int *num)
{
        (*num)++;

        fread(((void *) node) + offsetof(struct tree_node, depth),
               sizeof(struct tree_node) - offsetof(struct tree_node, depth),
               1, f);

        /* Keep values in sane scale, otherwise we start overflowing. */
#define MAX_PLAYOUTS    10000000
        if (node->u.playouts > MAX_PLAYOUTS) {
                node->u.playouts = MAX_PLAYOUTS;
        }
        if (node->amaf.playouts > MAX_PLAYOUTS) {
                node->amaf.playouts = MAX_PLAYOUTS;
        }
        memcpy(&node->pamaf, &node->amaf, sizeof(node->amaf));
        memcpy(&node->pu, &node->u, sizeof(node->u));

        struct tree_node *ni = NULL, *ni_prev = NULL;
        while (fgetc(f)) {
                ni_prev = ni; ni = calloc(1, sizeof(*ni));
                if (!node->children)
                        node->children = ni;
                else
                        ni_prev->sibling = ni;
                ni->parent = node;
                tree_node_load(f, ni, num);
        }
}

void
tree_load(struct tree *tree, struct board *b)
{
        char *filename = tree_book_name(b);
        FILE *f = fopen(filename, "rb");
        if (!f)
                return;

        fprintf(stderr, "Loading opening book %s...\n", filename);

        int num = 0;
        if (fgetc(f))
                tree_node_load(f, tree->root, &num);
        fprintf(stderr, "Loaded %d nodes.\n", num);

        fclose(f);
}


static struct tree_node *
tree_node_copy(struct tree_node *node)
{
        struct tree_node *n2 = malloc(sizeof(*n2));
        *n2 = *node;
        if (!node->children)
                return n2;
        struct tree_node *ni = node->children;
        struct tree_node *ni2 = tree_node_copy(ni);
        n2->children = ni2; ni2->parent = n2;
        while ((ni = ni->sibling)) {
                ni2->sibling = tree_node_copy(ni);
                ni2 = ni2->sibling; ni2->parent = n2;
        }
        return n2;
}

struct tree *
tree_copy(struct tree *tree)
{
        assert(!tree->nodes);
        struct tree *t2 = malloc(sizeof(*t2));
        *t2 = *tree;
        t2->root = tree_node_copy(tree->root);
        return t2;
}

/* Copy the subtree rooted at node: all nodes at or below depth
 * or with at least threshold playouts. Only for fast_alloc.
 * The code is destructive on src, and the order of nodes is changed.
 * Returns the copy of node in the destination tree, or NULL
 * if we could not copy it. */
static struct tree_node *
tree_prune(struct tree *dest, struct tree *src, struct tree_node *node,
           int threshold, int depth)
{
        assert(dest->nodes && node);
        struct tree_node *n2 = tree_fast_alloc_node(dest);
        assert(n2);
        *n2 = *node;
        if (n2->depth > dest->max_depth)
                dest->max_depth = n2->depth;
        n2->children = NULL;
        n2->is_expanded = false;

        if (node->depth >= depth && node->u.playouts < threshold)
                return n2;
        /* For deep nodes with many playouts, we must copy all children,
         * even those with zero playouts, because partially expanded
         * nodes are not supported. Considering them as fully expanded
         * would degrade the playing strength. The only exception is
         * when dest becomes full, but this should never happen in practice
         * if threshold is chosen to limit the number of nodes traversed. */
        struct tree_node *ni = node->children;
        while (ni && unlikely(dest->nodes_size < dest->max_tree_size)) {
                struct tree_node *ni2 = tree_prune(dest, src, ni, threshold, depth);
                assert(ni2);
                ni2->sibling = n2->children;
                n2->children = ni2;
                ni2->parent = n2;
                ni = ni->sibling;
        }
        if (n2->children && !ni) {
                n2->is_expanded = true;
        } else {
                n2->children = NULL; // avoid partially expanded nodes
        }
        return n2;
}

/* The following constants are used for garbage collection of nodes.
 * A tree is considered large if the top node has >= 40K playouts.
 * For such trees, we copy deep nodes only if they have >= 40 playouts.
 * These constants define how much time we're willing to spend
 * scanning the source tree when promoting a move. The values 40K
 * and 40 makes worst case pruning in about 3s for 20 GB ram, and this
 * is only for long thinking time (>1M playouts). For fast games the
 * trees don't grow large. For small ram or fast game we copy the
 * entire tree.  These values do not degrade playing strength and are
 * necessary to avoid losing on time; increasing MIN_DEEP_PLAYOUTS
 * or decreasing LARGE_TREE_PLAYOUTS will make the program faster but
 * playing worse. */
#define LARGE_TREE_PLAYOUTS 40000
#define MIN_DEEP_PLAYOUTS 40

/* Free all the tree, keeping only the subtree rooted at node.
 * Prune the subtree if necessary to fit in max_size bytes or
 * to save time scanning the tree.
 * Returns the moved node. Only for fast_alloc. */
static struct tree_node *
tree_garbage_collect(struct tree *tree, unsigned long max_size, struct tree_node *node)
{
        assert(tree->nodes && !node->parent && !node->sibling);
        double start_time = time_now();

        struct tree *temp_tree = tree_init(tree->board,  tree->root_color, max_size);
        temp_tree->nodes_size = 0; // We do not want the dummy pass node
        struct tree_node *temp_node;

        /* Find the maximum depth at which we can copy all nodes. */
        int max_nodes = 1;
        for (struct tree_node *ni = node->children; ni; ni = ni->sibling)
                max_nodes++;
        unsigned long nodes_size = max_nodes * sizeof(*node);
        int max_depth = node->depth;
        while (nodes_size < max_size && max_nodes > 1) {
                max_nodes--;
                nodes_size += max_nodes * nodes_size;
                max_depth++;
        }

        /* Copy all nodes for small trees. For large trees, copy all nodes
         * with depth <= max_depth, and all nodes with at least MIN_DEEP_PLAYOUTS.
         * Avoiding going too deep (except for nodes with many playouts) is mostly
         * to save time scanning the source tree. It can take over 20s to traverse
         * completely a large source tree (20 GB) even without copying because
         * the traversal is not friendly at all with the memory cache. */
        if (node->u.playouts < LARGE_TREE_PLAYOUTS) {
                temp_node = tree_prune(temp_tree, tree, node, 0, max_depth + 20);
        } else {
                temp_node = tree_prune(temp_tree, tree, node, MIN_DEEP_PLAYOUTS, max_depth);
        }
        assert(temp_node);

        /* Now copy back to original tree. */
        tree->nodes_size = 0;
        tree->max_depth = 0;
        struct tree_node *new_node = tree_prune(tree, temp_tree, temp_node, 0, temp_tree->max_depth);

        if (DEBUGL(1)) {
                double now = time_now();
                static double prev_time;
                if (!prev_time) prev_time = start_time;
                fprintf(stderr,
                        "tree pruned in %0.6g s, prev %0.3g s ago, dest depth %d wanted %d,"
                        " max_size %lu, pruned size %lu, playouts %d\n",
                        now - start_time, start_time - prev_time, temp_tree->max_depth, max_depth,
                        max_size, temp_tree->nodes_size, new_node->u.playouts);
                prev_time = start_time;
        }
        assert(tree->nodes_size == temp_tree->nodes_size);
        assert(tree->max_depth == temp_tree->max_depth);
        tree_done(temp_tree);
        return new_node;
}


static void
tree_node_merge(struct tree_node *dest, struct tree_node *src)
{
        /* Do not merge nodes that weren't touched at all. */
        assert(dest->pamaf.playouts == src->pamaf.playouts);
        assert(dest->pu.playouts == src->pu.playouts);
        if (src->amaf.playouts - src->pamaf.playouts == 0
            && src->u.playouts - src->pu.playouts == 0) {
                return;
        }

        dest->hints |= src->hints;

        /* Merge the children, both are coord-sorted lists. */
        struct tree_node *di = dest->children, **dref = &dest->children;
        struct tree_node *si = src->children, **sref = &src->children;
        while (di && si) {
                if (di->coord != si->coord) {
                        /* src has some extra items or misses di */
                        struct tree_node *si2 = si->sibling;
                        while (si2 && di->coord != si2->coord) {
                                si2 = si2->sibling;
                        }
                        if (!si2)
                                goto next_di; /* src misses di, move on */
                        /* chain the extra [si,si2) items before di */
                        (*dref) = si;
                        while (si->sibling != si2) {
                                si->parent = dest;
                                si = si->sibling;
                        }
                        si->parent = dest;
                        si->sibling = di;
                        si = si2;
                        (*sref) = si;
                }
                /* Matching nodes - recurse... */
                tree_node_merge(di, si);
                /* ...and move on. */
                sref = &si->sibling; si = si->sibling;
next_di:
                dref = &di->sibling; di = di->sibling;
        }
        if (si) {
                /* Some outstanding nodes are left on src side, rechain
                 * them to dst. */
                (*dref) = si;
                while (si) {
                        si->parent = dest;
                        si = si->sibling;
                }
                (*sref) = NULL;
        }

        /* Priors should be constant. */
        assert(dest->prior.playouts == src->prior.playouts && dest->prior.value == src->prior.value);

        stats_merge(&dest->amaf, &src->amaf);
        stats_merge(&dest->u, &src->u);
}

/* Merge two trees built upon the same board. Note that the operation is
 * destructive on src. */
void
tree_merge(struct tree *dest, struct tree *src)
{
        /* Not suitable for fast_alloc which reorders children. */
        assert(!dest->nodes);

        if (src->max_depth > dest->max_depth)
                dest->max_depth = src->max_depth;
        tree_node_merge(dest->root, src->root);
}


static void
tree_node_normalize(struct tree_node *node, int factor)
{
        for (struct tree_node *ni = node->children; ni; ni = ni->sibling)
                tree_node_normalize(ni, factor);

#define normalize(s1, s2, t) node->s2.t = node->s1.t + (node->s2.t - node->s1.t) / factor;
        normalize(pamaf, amaf, playouts);
        memcpy(&node->pamaf, &node->amaf, sizeof(node->amaf));

        normalize(pu, u, playouts);
        memcpy(&node->pu, &node->u, sizeof(node->u));
#undef normalize
}

/* Normalize a tree, dividing the amaf and u values by given
 * factor; otherwise, simulations run in independent threads
 * two trees built upon the same board. To correctly handle
 * results taken from previous simulation run, they are backed
 * up in tree. */
void
tree_normalize(struct tree *tree, int factor)
{
        tree_node_normalize(tree->root, factor);
}


/* Tree symmetry: When possible, we will localize the tree to a single part
 * of the board in tree_expand_node() and possibly flip along symmetry axes
 * to another part of the board in tree_promote_at(). We follow b->symmetry
 * guidelines here. */


/* This function must be thread safe, given that board b is only modified by the calling thread. */
void
tree_expand_node(struct tree *t, struct tree_node *node, struct board *b, enum stone color, struct uct *u, int parity)
{
        /* Get a Common Fate Graph distance map from parent node. */
        int distances[board_size2(b)];
        if (!is_pass(b->last_move.coord) && !is_resign(b->last_move.coord)) {
                cfg_distances(b, node->coord, distances, TREE_NODE_D_MAX);
        } else {
                // Pass or resign - everything is too far.
                foreach_point(b) { distances[c] = TREE_NODE_D_MAX + 1; } foreach_point_end;
        }

        /* Get a map of prior values to initialize the new nodes with. */
        struct prior_map map = {
                .b = b,
                .to_play = color,
                .parity = tree_parity(t, parity),
                .distances = distances,
        };
        // Include pass in the prior map.
        struct move_stats map_prior[board_size2(b) + 1]; map.prior = &map_prior[1];
        bool map_consider[board_size2(b) + 1]; map.consider = &map_consider[1];
        memset(map_prior, 0, sizeof(map_prior));
        memset(map_consider, 0, sizeof(map_consider));
        struct move pm = { .color = color };
        map.consider[pass] = true;
        foreach_point(b) {
                if (board_at(b, c) != S_NONE)
                        continue;
                pm.coord = c;
                if (!board_is_valid_move(b, &pm))
                        continue;
                map.consider[c] = true;
        } foreach_point_end;
        uct_prior(u, node, &map);

        /* Now, create the nodes. */
        struct tree_node *ni = tree_init_node(t, pass, node->depth + 1);
        /* In fast_alloc mode we might temporarily run out of nodes but
         * this should be rare if MIN_FREE_MEM_PERCENT is set correctly. */
        if (!ni) {
                node->is_expanded = false;
                return;
        }
        struct tree_node *first_child = ni;
        ni->parent = node;
        ni->prior = map.prior[pass]; ni->d = TREE_NODE_D_MAX + 1;

        /* The loop considers only the symmetry playground. */
        if (UDEBUGL(6)) {
                fprintf(stderr, "expanding %s within [%d,%d],[%d,%d] %d-%d\n",
                                coord2sstr(node->coord, b),
                                b->symmetry.x1, b->symmetry.y1,
                                b->symmetry.x2, b->symmetry.y2,
                                b->symmetry.type, b->symmetry.d);
        }
        for (int i = b->symmetry.x1; i <= b->symmetry.x2; i++) {
                for (int j = b->symmetry.y1; j <= b->symmetry.y2; j++) {
                        if (b->symmetry.d) {
                                int x = b->symmetry.type == SYM_DIAG_DOWN ? board_size(b) - 1 - i : i;
                                if (x > j) {
                                        if (UDEBUGL(7))
                                                fprintf(stderr, "drop %d,%d\n", i, j);
                                        continue;
                                }
                        }

                        coord_t c = coord_xy_otf(i, j, t->board);
                        if (!map.consider[c]) // Filter out invalid moves
                                continue;
                        assert(c != node->coord); // I have spotted "C3 C3" in some sequence...

                        struct tree_node *nj = tree_init_node(t, c, node->depth + 1);
                        if (!nj) {
                                node->is_expanded = false;
                                return;
                        }
                        nj->parent = node; ni->sibling = nj; ni = nj;

                        ni->prior = map.prior[c];
                        ni->d = distances[c];
                }
        }
        node->children = first_child; // must be done at the end to avoid race
}


static coord_t
flip_coord(struct board *b, coord_t c,
           bool flip_horiz, bool flip_vert, int flip_diag)
{
        int x = coord_x(c, b), y = coord_y(c, b);
        if (flip_diag) {
                int z = x; x = y; y = z;
        }
        if (flip_horiz) {
                x = board_size(b) - 1 - x;
        }
        if (flip_vert) {
                y = board_size(b) - 1 - y;
        }
        return coord_xy_otf(x, y, b);
}

static void
tree_fix_node_symmetry(struct board *b, struct tree_node *node,
                       bool flip_horiz, bool flip_vert, int flip_diag)
{
        if (!is_pass(node->coord))
                node->coord = flip_coord(b, node->coord, flip_horiz, flip_vert, flip_diag);

        for (struct tree_node *ni = node->children; ni; ni = ni->sibling)
                tree_fix_node_symmetry(b, ni, flip_horiz, flip_vert, flip_diag);
}

static void
tree_fix_symmetry(struct tree *tree, struct board *b, coord_t c)
{
        if (is_pass(c))
                return;

        struct board_symmetry *s = &tree->root_symmetry;
        int cx = coord_x(c, b), cy = coord_y(c, b);

        /* playground   X->h->v->d normalization
         * :::..        .d...
         * .::..        v....
         * ..:..        .....
         * .....        h...X
         * .....        .....  */
        bool flip_horiz = cx < s->x1 || cx > s->x2;
        bool flip_vert = cy < s->y1 || cy > s->y2;

        bool flip_diag = 0;
        if (s->d) {
                bool dir = (s->type == SYM_DIAG_DOWN);
                int x = dir ^ flip_horiz ^ flip_vert ? board_size(b) - 1 - cx : cx;
                if (flip_vert ? x < cy : x > cy) {
                        flip_diag = 1;
                }
        }

        if (DEBUGL(4)) {
                fprintf(stderr, "%s [%d,%d -> %d,%d;%d,%d] will flip %d %d %d -> %s, sym %d (%d) -> %d (%d)\n",
                        coord2sstr(c, b),
                        cx, cy, s->x1, s->y1, s->x2, s->y2,
                        flip_horiz, flip_vert, flip_diag,
                        coord2sstr(flip_coord(b, c, flip_horiz, flip_vert, flip_diag), b),
                        s->type, s->d, b->symmetry.type, b->symmetry.d);
        }
        if (flip_horiz || flip_vert || flip_diag)
                tree_fix_node_symmetry(b, tree->root, flip_horiz, flip_vert, flip_diag);
}


static void
tree_unlink_node(struct tree_node *node)
{
        struct tree_node *ni = node->parent;
        if (ni->children == node) {
                ni->children = node->sibling;
        } else {
                ni = ni->children;
                while (ni->sibling != node)
                        ni = ni->sibling;
                ni->sibling = node->sibling;
        }
        node->sibling = NULL;
        node->parent = NULL;
}

/* Promotes the given node as the root of the tree. In the fast_alloc
 * mode, the node may be moved and some of its subtree may be pruned. */
void
tree_promote_node(struct tree *tree, struct tree_node **node)
{
        assert((*node)->parent == tree->root);
        tree_unlink_node(*node);
        if (!tree->nodes) {
                /* Freeing the rest of the tree can take several seconds on large
                 * trees, so we must do it asynchronously: */
                tree_done_node_detached(tree, tree->root);
        } else {
                unsigned long min_free_size = (MIN_FREE_MEM_PERCENT * tree->max_tree_size) / 100;
                if (tree->nodes_size >= tree->max_tree_size - min_free_size)
                        *node = tree_garbage_collect(tree, min_free_size, *node);
                /* If we still have enough free memory, we will free everything later. */
        }
        tree->root = *node;
        tree->root_color = stone_other(tree->root_color);
        board_symmetry_update(tree->board, &tree->root_symmetry, (*node)->coord);
        /* If the tree deepest node was under node, or if we called tree_garbage_collect,
         * tree->max_depth is correct. Otherwise we could traverse the tree
         * to recompute max_depth but it's not worth it: it's just for debugging
         * and soon the tree will grow and max_depth will become correct again. */
        if (tree->chchvals) { free(tree->chchvals); tree->chchvals = NULL; }
        if (tree->chvals) { free(tree->chvals); tree->chvals = NULL; }
}

bool
tree_promote_at(struct tree *tree, struct board *b, coord_t c)
{
        tree_fix_symmetry(tree, b, c);

        for (struct tree_node *ni = tree->root->children; ni; ni = ni->sibling) {
                if (ni->coord == c) {
                        tree_promote_node(tree, &ni);
                        return true;
                }
        }
        return false;
}