+
+void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
+{
+ be_ifg_t *ifg = co->cenv->ifg;
+ int *color_map = alloca(co->cls->n_regs * sizeof(color_map[0]));
+ bitset_t *adm = bitset_alloca(co->cls->n_regs);
+
+ ir_node *irn;
+ void *it, *nit;
+ int i, n, n_regs;
+
+ n_regs = 0;
+ for(i = 0; i < co->cls->n_regs; ++i) {
+ const arch_register_t *reg = &co->cls->regs[i];
+ color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_regs++;
+ }
+
+ /*
+ * n contains the first node number.
+ * the values below n are the pre-colored register nodes
+ */
+
+ it = be_ifg_nodes_iter_alloca(ifg);
+ nit = be_ifg_neighbours_iter_alloca(ifg);
+
+ n = n_regs;
+ be_ifg_foreach_node(ifg, it, irn) {
+ if(!arch_irn_is(co->aenv, irn, ignore))
+ set_irn_link(irn, INT_TO_PTR(n++));
+ }
+
+ fprintf(f, "%d %d\n", n, n_regs);
+
+ be_ifg_foreach_node(ifg, it, irn) {
+ if(!arch_irn_is(co->aenv, irn, ignore)) {
+ int idx = PTR_TO_INT(get_irn_link(irn));
+ affinity_node_t *a = get_affinity_info(co, irn);
+
+ arch_register_req_t req;
+ ir_node *adj;
+
+ arch_get_register_req(co->aenv, &req, irn, BE_OUT_POS(0));
+ if(arch_register_req_is(&req, limited)) {
+ bitset_clear_all(adm);
+ req.limited(req.limited_env, adm);
+ for(i = 0; i < co->cls->n_regs; ++i)
+ if(!bitset_is_set(adm, i) && color_map[i] >= 0)
+ fprintf(f, "%d %d -1\n", color_map[i], idx);
+
+ }
+
+
+ be_ifg_foreach_neighbour(ifg, nit, irn, adj) {
+ if(!arch_irn_is(co->aenv, adj, ignore)) {
+ int adj_idx = PTR_TO_INT(get_irn_link(adj));
+ if(idx < adj_idx)
+ fprintf(f, "%d %d -1\n", idx, adj_idx);
+ }
+ }
+
+ if(a) {
+ neighb_t *n;
+
+ co_gs_foreach_neighb(a, n) {
+ if(!arch_irn_is(co->aenv, n->irn, ignore)) {
+ int n_idx = PTR_TO_INT(get_irn_link(n->irn));
+ if(idx < n_idx)
+ fprintf(f, "%d %d %d\n", idx, n_idx, n->costs);
+ }
+ }
+ }
+ }
+ }
+}
+
+typedef struct _appel_clique_walker_t {
+ phase_t ph;
+ const copy_opt_t *co;
+ int curr_nr;
+ int node_count;
+ FILE *f;
+ int dumb;
+ int *color_map;
+ struct obstack obst;
+} appel_clique_walker_t;
+
+typedef struct _appel_block_info_t {
+ int *live_end_nr;
+ int *live_in_nr;
+ int *phi_nr;
+ ir_node **live_end;
+ ir_node **live_in;
+ ir_node **phi;
+ int n_live_end;
+ int n_live_in;
+ int n_phi;
+} appel_block_info_t;
+
+static int appel_aff_weight(const appel_clique_walker_t *env, ir_node *bl)
+{
+#if 0
+ double freq = get_block_execfreq(env->co->cenv->execfreq, bl);
+ int res = (int) freq;
+ return res == 0 ? 1 : res;
+#else
+ ir_loop *loop = get_irn_loop(bl);
+ if(loop) {
+ int d = get_loop_depth(loop);
+ return 1 + d * d;
+ }
+ return 1;
+#endif
+}
+
+static void *appel_clique_walker_irn_init(phase_t *phase, ir_node *irn, void *old)
+{
+ appel_block_info_t *res = NULL;
+
+ if(is_Block(irn)) {
+ appel_clique_walker_t *d = (void *) phase;
+ res = phase_alloc(phase, sizeof(res[0]));
+ res->phi_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
+ res->live_end_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
+ res->live_in_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in_nr));
+ res->live_end = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end));
+ res->live_in = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
+ res->phi = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
+ }
+
+ return res;
+}
+
+typedef struct _insn_list_t {
+ be_insn_t *insn;
+ struct list_head list;
+} insn_list_t;
+
+static int appel_get_live_end_nr(appel_clique_walker_t *env, ir_node *bl, ir_node *irn)
+{
+ appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
+ int i;
+
+ for(i = 0; i < bli->n_live_end; ++i)
+ if(bli->live_end[i] == irn)
+ return bli->live_end_nr[i];
+
+ return -1;
+}
+
+static int appel_dump_clique(appel_clique_walker_t *env, pset *live, ir_node *bl, int curr_nr, int start_nr)
+{
+ ir_node **live_arr = alloca(env->co->cls->n_regs * sizeof(live_arr[0]));
+ ir_node *irn;
+ int n_live;
+ int j;
+
+ n_live = 0;
+ foreach_pset(live, irn)
+ live_arr[n_live++] = irn;
+
+ /* dump the live after clique */
+ if(!env->dumb) {
+ for(j = 0; j < n_live; ++j) {
+ int k;
+
+ for(k = j + 1; k < n_live; ++k) {
+ fprintf(env->f, "%d %d -1 ", curr_nr + j, curr_nr + k);
+ }
+ fprintf(env->f, "\n");
+ }
+ }
+
+ /* dump the affinities */
+ for(j = 0; !env->dumb && j < n_live; ++j) {
+ ir_node *irn = live_arr[j];
+ int old_nr = PTR_TO_INT(get_irn_link(irn));
+
+ /* if the node was already live in the last insn dump the affinity */
+ if(old_nr > start_nr) {
+ int weight = appel_aff_weight(env, bl);
+ fprintf(env->f, "%d %d %d\n", old_nr, curr_nr + j, weight);
+ }
+ }
+
+ /* set the current numbers into the link field. */
+ for(j = 0; j < n_live; ++j) {
+ ir_node *irn = live_arr[j];
+ set_irn_link(irn, INT_TO_PTR(curr_nr + j));
+ }
+
+ return curr_nr + n_live;
+}
+
+static void appel_walker(ir_node *bl, void *data)
+{
+ appel_clique_walker_t *env = data;
+ appel_block_info_t *bli = phase_get_or_set_irn_data(&env->ph, bl);
+ struct obstack *obst = &env->obst;
+ void *base = obstack_base(obst);
+ pset *live = pset_new_ptr_default();
+
+ int n_insns = 0;
+ int n_nodes = 0;
+ int start_nr = env->curr_nr;
+ int curr_nr = start_nr;
+
+ be_insn_env_t insn_env;
+ int i, j;
+ ir_node *irn;
+ be_insn_t **insns;
+
+ insn_env.aenv = env->co->aenv;
+ insn_env.cls = env->co->cls;
+ insn_env.obst = obst;
+ insn_env.ignore_colors = env->co->cenv->ignore_colors;
+
+ /* Guess how many insns will be in this block. */
+ sched_foreach(bl, irn)
+ n_nodes++;
+
+ bli->n_phi = 0;
+ insns = malloc(n_nodes * sizeof(insns[0]));
+
+ /* Put all insns in an array. */
+ irn = sched_first(bl);
+ while(!sched_is_end(irn)) {
+ be_insn_t *insn;
+ insn = be_scan_insn(&insn_env, irn);
+ insns[n_insns++] = insn;
+ irn = insn->next_insn;
+ }
+
+ DBG((env->co->cenv->dbg, LEVEL_2, "%+F\n", bl));
+ be_liveness_end_of_block(env->co->cenv->lv, env->co->aenv, env->co->cls, bl, live);
+
+ /* Generate the bad and ugly. */
+ for(i = n_insns - 1; i >= 0; --i) {
+ be_insn_t *insn = insns[i];
+
+ /* The first live set has to be saved in the block border set. */
+ if(i == n_insns - 1) {
+ j = 0;
+ foreach_pset(live, irn) {
+ bli->live_end[j] = irn;
+ bli->live_end_nr[j] = curr_nr + j;
+ ++j;
+ }
+ bli->n_live_end = j;
+ }
+
+ if(!env->dumb) {
+ for(j = 0; j < insn->use_start; ++j) {
+ ir_node *op = insn->ops[j].carrier;
+ bitset_t *adm = insn->ops[j].regs;
+ int k;
+ int nr;
+
+ if(!insn->ops[j].has_constraints)
+ continue;
+
+ nr = 0;
+ foreach_pset(live, irn) {
+ if(irn == op) {
+ pset_break(live);
+ break;
+ }
+ ++nr;
+ }
+
+ assert(nr < pset_count(live));
+
+ for(k = 0; k < env->co->cls->n_regs; ++k) {
+ int mapped_col = env->color_map[k];
+ if(mapped_col >= 0 && !bitset_is_set(adm, k) && !bitset_is_set(env->co->cenv->ignore_colors, k))
+ fprintf(env->f, "%d %d -1\n", curr_nr + nr, mapped_col);
+ }
+ }
+ }
+
+ /* dump the clique and update the stuff. */
+ curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
+
+ /* remove all defs. */
+ for(j = 0; j < insn->use_start; ++j)
+ pset_remove_ptr(live, insn->ops[j].carrier);
+
+ if(is_Phi(insn->irn) && arch_irn_consider_in_reg_alloc(env->co->aenv, env->co->cls, insn->irn)) {
+ bli->phi[bli->n_phi] = insn->irn;
+ bli->phi_nr[bli->n_phi] = PTR_TO_INT(get_irn_link(insn->irn));
+ bli->n_phi++;
+ }
+
+ /* add all uses */
+ else
+ for(j = insn->use_start; j < insn->n_ops; ++j)
+ pset_insert_ptr(live, insn->ops[j].carrier);
+ }
+
+ /* print the start clique. */
+ curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
+
+ i = 0;
+ foreach_pset(live, irn) {
+ bli->live_in[i] = irn;
+ bli->live_in_nr[i] = PTR_TO_INT(get_irn_link(irn));
+ ++i;
+ }
+ bli->n_live_in = i;
+
+ del_pset(live);
+ free(insns);
+ obstack_free(obst, base);
+ env->curr_nr = curr_nr;
+}
+
+static void appel_inter_block_aff(ir_node *bl, void *data)
+{
+ appel_clique_walker_t *env = data;
+ appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
+
+ int i, j, n;
+
+ for(i = 0; i < bli->n_live_in; ++i) {
+ ir_node *irn = bli->live_in[i];
+
+ for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
+ ir_node *pred = get_Block_cfgpred_block(bl, j);
+ appel_block_info_t *pred_bli = phase_get_irn_data(&env->ph, pred);
+
+ int nr = appel_get_live_end_nr(env, pred, irn);
+ assert(nr >= 0);
+ fprintf(env->f, "%d %d 1\n", bli->live_in_nr[i], nr);
+ }
+ }
+
+ for(i = 0; i < bli->n_phi; ++i) {
+ ir_node *irn = bli->phi[i];
+
+ for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
+ ir_node *pred = get_Block_cfgpred_block(bl, j);
+ appel_block_info_t *pred_bli = phase_get_irn_data(&env->ph, pred);
+ ir_node *op = get_irn_n(irn, j);
+
+ int nr = appel_get_live_end_nr(env, pred, op);
+ assert(nr >= 0);
+ fprintf(env->f, "%d %d 1\n", bli->phi_nr[i], nr);
+ }
+ }
+
+}
+
+void co_dump_appel_graph_cliques(const copy_opt_t *co, FILE *f)
+{
+ int i;
+ int n_colors;
+ appel_clique_walker_t env;
+ bitset_t *adm = bitset_alloca(co->cls->n_regs);
+
+ be_liveness_recompute(co->cenv->lv);
+ obstack_init(&env.obst);
+ phase_init(&env.ph, "appel_clique_dumper", co->irg, PHASE_DEFAULT_GROWTH, appel_clique_walker_irn_init);
+ env.curr_nr = co->cls->n_regs;
+ env.co = co;
+ env.f = f;
+
+ bitset_copy(adm, co->cenv->ignore_colors);
+ bitset_flip_all(adm);
+
+ /* Make color map. */
+ env.color_map = alloca(co->cls->n_regs * sizeof(env.color_map[0]));
+ for(i = 0, n_colors = 0; i < co->cls->n_regs; ++i) {
+ const arch_register_t *reg = &co->cls->regs[i];
+ env.color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_colors++;
+ }
+
+ env.dumb = 1;
+ env.curr_nr = n_colors;
+ irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
+ irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
+
+ fprintf(f, "%d %d\n", env.curr_nr, n_colors);
+
+ /* make the first k nodes interfere */
+ for(i = 0; i < n_colors; ++i) {
+ int j;
+ for(j = i + 1; j < n_colors; ++j)
+ fprintf(f, "%d %d -1 ", i, j);
+ fprintf(f, "\n");
+ }
+
+ env.dumb = 0;
+ env.curr_nr = n_colors;
+ irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
+ irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
+ irg_block_walk_graph(co->irg, appel_inter_block_aff, NULL, &env);
+ obstack_free(&env.obst, NULL);
+}
+
+/*
+ ___ _____ ____ ____ ___ _____ ____ _
+ |_ _| ___/ ___| | _ \ / _ \_ _| | _ \ _ _ _ __ ___ _ __ (_)_ __ __ _
+ | || |_ | | _ | | | | | | || | | | | | | | | '_ ` _ \| '_ \| | '_ \ / _` |
+ | || _|| |_| | | |_| | |_| || | | |_| | |_| | | | | | | |_) | | | | | (_| |
+ |___|_| \____| |____/ \___/ |_| |____/ \__,_|_| |_| |_| .__/|_|_| |_|\__, |
+ |_| |___/
+*/
+
+static const char *get_dot_color_name(int col)
+{
+ static const char *names[] = {
+ "blue",
+ "red",
+ "green",
+ "yellow",
+ "cyan",
+ "magenta",
+ "orange",
+ "chocolate",
+ "beige",
+ "navy",
+ "darkgreen",
+ "darkred",
+ "lightPink",
+ "chartreuse",
+ "lightskyblue",
+ "linen",
+ "pink",
+ "lightslateblue",
+ "mintcream",
+ "red",
+ "darkolivegreen",
+ "mediumblue",
+ "mistyrose",
+ "salmon",
+ "darkseagreen",
+ "mediumslateblue"
+ "moccasin",
+ "tomato",
+ "forestgreen",
+ "darkturquoise",
+ "palevioletred"
+ };
+
+ return col < sizeof(names)/sizeof(names[0]) ? names[col] : "white";
+}
+
+typedef struct _co_ifg_dump_t {
+ const copy_opt_t *co;
+ unsigned flags;
+} co_ifg_dump_t;
+
+static void ifg_dump_graph_attr(FILE *f, void *self)
+{
+ fprintf(f, "overlap=scale");
+}
+
+static int ifg_is_dump_node(void *self, ir_node *irn)
+{
+ co_ifg_dump_t *cod = self;
+ return !arch_irn_is(cod->co->aenv, irn, ignore);
+}
+
+static void ifg_dump_node_attr(FILE *f, void *self, ir_node *irn)
+{
+ co_ifg_dump_t *env = self;
+ const arch_register_t *reg = arch_get_irn_register(env->co->aenv, irn);
+ arch_register_req_t req;
+ int limited;
+
+ arch_get_register_req(env->co->aenv, &req, irn, BE_OUT_POS(0));
+ limited = arch_register_req_is(&req, limited);
+
+ if(env->flags & CO_IFG_DUMP_LABELS) {
+ ir_fprintf(f, "label=\"%+F", irn);
+
+ if((env->flags & CO_IFG_DUMP_CONSTR) && limited) {
+ bitset_t *bs = bitset_alloca(env->co->cls->n_regs);
+ req.limited(req.limited_env, bs);
+ ir_fprintf(f, "\\n%B", bs);
+ }
+ ir_fprintf(f, "\" ");
+ }
+
+ else
+ fprintf(f, "label=\"\"" );
+
+ if(env->flags & CO_IFG_DUMP_SHAPE)
+ fprintf(f, "shape=%s ", limited ? "diamond" : "ellipse");
+
+ if(env->flags & CO_IFG_DUMP_COLORS)
+ fprintf(f, "style=filled color=%s ", get_dot_color_name(reg->index));
+}
+
+static void ifg_dump_at_end(FILE *file, void *self)
+{
+ co_ifg_dump_t *env = self;
+ affinity_node_t *a;
+
+ co_gs_foreach_aff_node(env->co, a) {
+ const arch_register_t *ar = arch_get_irn_register(env->co->aenv, a->irn);
+ unsigned aidx = get_irn_idx(a->irn);
+ neighb_t *n;
+
+ co_gs_foreach_neighb(a, n) {
+ const arch_register_t *nr = arch_get_irn_register(env->co->aenv, n->irn);
+ unsigned nidx = get_irn_idx(n->irn);
+
+ if(aidx < nidx) {
+ const char *color = nr == ar ? "blue" : "red";
+ fprintf(file, "\tn%d -- n%d [weight=0.01 ", aidx, nidx);
+ if(env->flags & CO_IFG_DUMP_LABELS)
+ fprintf(file, "label=\"%d\" ", n->costs);
+ if(env->flags & CO_IFG_DUMP_COLORS)
+ fprintf(file, "color=%s ", color);
+ else
+ fprintf(file, "style=dashed");
+ fprintf(file, "];\n");
+ }
+ }
+ }
+}
+
+
+static be_ifg_dump_dot_cb_t ifg_dot_cb = {
+ ifg_is_dump_node,
+ ifg_dump_graph_attr,
+ ifg_dump_node_attr,
+ NULL,
+ NULL,
+ ifg_dump_at_end
+};
+
+
+
+void co_dump_ifg_dot(const copy_opt_t *co, FILE *f, unsigned flags)
+{
+ co_ifg_dump_t cod;
+
+ cod.co = co;
+ cod.flags = flags;
+ be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &cod);
+}
+
+
+void co_solve_park_moon(copy_opt_t *opt)
+{
+
+}
+
+static int void_algo(copy_opt_t *co)
+{
+ return 0;
+}
+
+/*
+ _ _ _ _ _
+ / \ | | __ _ ___ _ __(_) |_| |__ _ __ ___ ___
+ / _ \ | |/ _` |/ _ \| '__| | __| '_ \| '_ ` _ \/ __|
+ / ___ \| | (_| | (_) | | | | |_| | | | | | | | \__ \
+ /_/ \_\_|\__, |\___/|_| |_|\__|_| |_|_| |_| |_|___/
+ |___/
+*/
+
+static co_algo_t *algos[] = {
+ void_algo,
+ co_solve_heuristic,
+ co_solve_heuristic_new,
+ co_solve_heuristic_java,
+#ifdef WITH_ILP
+ co_solve_ilp2
+#endif
+};
+
+/*
+ __ __ _ ____ _
+ | \/ | __ _(_)_ __ | _ \ _ __(_)_ _____ _ __
+ | |\/| |/ _` | | '_ \ | | | | '__| \ \ / / _ \ '__|
+ | | | | (_| | | | | | | |_| | | | |\ V / __/ |
+ |_| |_|\__,_|_|_| |_| |____/|_| |_| \_/ \___|_|
+
+*/
+
+void co_driver(be_chordal_env_t *cenv)
+{
+ copy_opt_t *co;
+ co_algo_t *algo_func;
+
+ if(algo < 0 || algo >= CO_ALGO_LAST)
+ return;
+
+ co = new_copy_opt(cenv, cost_func);
+ co_build_ou_structure(co);
+ co_build_graph_structure(co);
+ if(do_stats) {
+ ir_printf("%30F %10s %7d%7d%7d%7d", cenv->irg, cenv->cls->name,
+ co_get_max_copy_costs(co), co_get_copy_costs(co),
+ co_get_inevit_copy_costs(co), co_get_lower_bound(co));
+ }
+
+ /* Dump the interference graph in Appel's format. */
+ if(dump_flags & DUMP_APPEL) {
+ FILE *f = be_chordal_open(cenv, "", ".apl");
+ co_dump_appel_graph(co, f);
+ fclose(f);
+ }
+
+ if(dump_flags & DUMP_BEFORE) {
+ FILE *f = be_chordal_open(cenv, "", "-before.dot");
+ co_dump_ifg_dot(co, f, style_flags);
+ fclose(f);
+ }
+
+ algo_func = algos[algo];
+ algo_func(co);
+
+ if(dump_flags & DUMP_AFTER) {
+ FILE *f = be_chordal_open(cenv, "", "-after.dot");
+ co_dump_ifg_dot(co, f, style_flags);
+ fclose(f);
+ }
+
+ if(do_stats) {
+ int optimizable_costs = co_get_max_copy_costs(co) - co_get_lower_bound(co);
+ int remaining = co_get_copy_costs(co);
+ int evitable = remaining - co_get_lower_bound(co);
+
+ if(optimizable_costs > 0)
+ printf("%5d %5.2f\n", remaining, (evitable * 100.0) / optimizable_costs);
+ else
+ printf("%5d %5s\n", remaining, "-");
+ }
+
+ co_free_graph_structure(co);
+ co_free_ou_structure(co);
+ free_copy_opt(co);
+}