nsz Git - libfirm/blob - ir/be/bepbqpcoloring.c

   1 /*
   2  * This file is part of libFirm.
   3  * Copyright (C) 2012 University of Karlsruhe.
   4  */
   5
   6 /**
   7  * @file
   8  * @brief       PBQP based register allocation.
   9  * @author      Thomas Bersch
  10  * @date        27.11.2009
  11  */
  12
  13 /* miscellaneous includes */
  14 #include "config.h"
  15
  16 #include "debug.h"
  17 #include "error.h"
  18
  19 #include "irdom.h"
  20 #include "irdump.h"
  21 #include "iredges_t.h"
  22 #include "irprintf.h"
  23 #include "irgwalk.h"
  24 #include "irtools.h"
  25 #include "time.h"
  26 #include "execfreq_t.h"
  27 #include "bipartite.h"
  28
  29 /* libfirm/ir/be includes */
  30 #include "bearch.h"
  31 #include "beirg.h"
  32 #include "besched.h"
  33 #include "bemodule.h"
  34 #include "bechordal_common.h"
  35 #include "bechordal.h"
  36 #include "bechordal_t.h"
  37 #include "becopyopt_t.h"
  38 #include "beinsn_t.h"
  39 #include "benode.h"
  40 #include "belive.h"
  41 #include "belive_t.h"
  42 #include "beutil.h"
  43 #include "plist.h"
  44 #include "pqueue.h"
  45
  46 /* pbqp includes */
  47 #include "kaps.h"
  48 #include "matrix.h"
  49 #include "vector.h"
  50 #include "vector_t.h"
  51 #include "heuristical_co.h"
  52 #include "heuristical_co_ld.h"
  53 #include "pbqp_t.h"
  54 #include "html_dumper.h"
  55 #include "pbqp_node_t.h"
  56 #include "pbqp_node.h"
  57 #include "pbqp_edge_t.h"
  58
  59 #define TIMER                 0
  60 #define PRINT_RPEO            0
  61 #define USE_BIPARTIT_MATCHING 0
  62 #define DO_USEFUL_OPT         1
  63
  64
  65 static int use_exec_freq     = true;
  66 static int use_late_decision = false;
  67
  68 typedef struct be_pbqp_alloc_env_t {
  69         pbqp_t                      *pbqp_inst;         /**< PBQP instance for register allocation */
  70         ir_graph                    *irg;               /**< The graph under examination. */
  71         const arch_register_class_t *cls;               /**< Current processed register class */
  72         be_lv_t                     *lv;
  73         bitset_t              const *allocatable_regs;
  74         pbqp_matrix_t               *ife_matrix_template;
  75         pbqp_matrix_t               *aff_matrix_template;
  76         plist_t                     *rpeo;
  77         unsigned                    *restr_nodes;
  78         unsigned                    *ife_edge_num;
  79         ir_execfreq_int_factors      execfreq_factors;
  80         be_chordal_env_t            *env;
  81 } be_pbqp_alloc_env_t;
  82
  83
  84 #define insert_edge(pbqp, src_node, trg_node, template_matrix) (add_edge_costs(pbqp, get_irn_idx(src_node), get_irn_idx(trg_node), pbqp_matrix_copy(pbqp, template_matrix)))
  85 #define get_free_regs(restr_nodes, cls, irn)                   (arch_register_class_n_regs(cls) - restr_nodes[get_irn_idx(irn)])
  86
  87 static const lc_opt_table_entry_t options[] = {
  88         LC_OPT_ENT_BOOL("exec_freq", "use exec_freq",  &use_exec_freq),
  89         LC_OPT_ENT_BOOL("late_decision", "use late decision for register allocation",  &use_late_decision),
  90         LC_OPT_LAST
  91 };
  92
  93 #if KAPS_DUMP
  94 static FILE *my_open(const be_chordal_env_t *env, const char *prefix, const char *suffix)
  95 {
  96         FILE       *result;
  97         char        buf[1024];
  98         size_t      i;
  99         size_t      n;
 100         char       *tu_name;
 101         const char *cup_name = be_get_irg_main_env(env->irg)->cup_name;
 102
 103         n = strlen(cup_name);
 104         tu_name = XMALLOCN(char, n + 1);
 105         strcpy(tu_name, cup_name);
 106         for (i = 0; i < n; ++i)
 107                 if (tu_name[i] == '.')
 108                         tu_name[i] = '_';
 109
 110         ir_snprintf(buf, sizeof(buf), "%s%s_%F_%s%s", prefix, tu_name, env->irg, env->cls->name, suffix);
 111         xfree(tu_name);
 112         result = fopen(buf, "wt");
 113         if (result == NULL) {
 114                 panic("Couldn't open '%s' for writing.", buf);
 115         }
 116
 117         return result;
 118 }
 119 #endif
 120
 121
 122 static void create_pbqp_node(be_pbqp_alloc_env_t *pbqp_alloc_env, ir_node *irn)
 123 {
 124         arch_register_class_t const *const cls              = pbqp_alloc_env->cls;
 125         pbqp_t                      *const pbqp_inst        = pbqp_alloc_env->pbqp_inst;
 126         bitset_t              const *const allocatable_regs = pbqp_alloc_env->allocatable_regs;
 127         unsigned                     const colors_n         = arch_register_class_n_regs(cls);
 128         unsigned                           cntConstrains    = 0;
 129
 130         /* create costs vector depending on register constrains */
 131         vector_t *costs_vector = vector_alloc(pbqp_inst, colors_n);
 132
 133         /* set costs depending on register constrains */
 134         unsigned idx;
 135         for (idx = 0; idx < colors_n; idx++) {
 136                 const arch_register_req_t *req = arch_get_irn_register_req(irn);
 137                 const arch_register_t     *reg = arch_register_for_index(cls, idx);
 138                 if (!bitset_is_set(allocatable_regs, idx)
 139                     || !arch_reg_is_allocatable(req, reg)) {
 140                         /* constrained */
 141                         vector_set(costs_vector, idx, INF_COSTS);
 142                         cntConstrains++;
 143                 }
 144         }
 145
 146         /* add vector to pbqp node */
 147         add_node_costs(pbqp_inst, get_irn_idx(irn), costs_vector);
 148         pbqp_alloc_env->restr_nodes[get_irn_idx(irn)] = cntConstrains;
 149 }
 150
 151 static void insert_ife_edge(be_pbqp_alloc_env_t *pbqp_alloc_env, ir_node *src_node, ir_node *trg_node)
 152 {
 153         pbqp_t                      *pbqp                = pbqp_alloc_env->pbqp_inst;
 154         const arch_register_class_t *cls                 = pbqp_alloc_env->cls;
 155         pbqp_matrix_t               *ife_matrix_template = pbqp_alloc_env->ife_matrix_template;
 156         unsigned                    *restr_nodes         = pbqp_alloc_env->restr_nodes;
 157
 158         if (get_edge(pbqp, get_irn_idx(src_node), get_irn_idx(trg_node)) == NULL) {
 159
 160                 /* increase ife edge counter */
 161                 pbqp_alloc_env->ife_edge_num[get_irn_idx(src_node)]++;
 162                 pbqp_alloc_env->ife_edge_num[get_irn_idx(trg_node)]++;
 163
 164 #if DO_USEFUL_OPT || USE_BIPARTIT_MATCHING
 165                 /* do useful optimization to speed up pbqp solving (we can do this because we know our matrix) */
 166                 if (get_free_regs(restr_nodes, cls, src_node) == 1 && get_free_regs(restr_nodes, cls, trg_node) == 1) {
 167                         assert(vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs) !=
 168                                vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs) &&
 169                                "Interfering nodes must not have the same register!");
 170                         return;
 171                 }
 172                 if (get_free_regs(restr_nodes, cls, src_node) == 1 || get_free_regs(restr_nodes, cls, trg_node) == 1) {
 173                         if (get_free_regs(restr_nodes, cls, src_node) == 1) {
 174                                 unsigned idx = vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs);
 175                                 vector_set(get_node(pbqp, get_irn_idx(trg_node))->costs, idx, INF_COSTS);
 176                         }
 177                         else {
 178                                 unsigned idx = vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs);
 179                                 vector_set(get_node(pbqp, get_irn_idx(src_node))->costs, idx, INF_COSTS);
 180                         }
 181                         return;
 182                 }
 183 #endif
 184                 /* insert interference edge */
 185                 insert_edge(pbqp, src_node, trg_node, ife_matrix_template);
 186         }
 187 }
 188
 189 static void insert_afe_edge(be_pbqp_alloc_env_t *pbqp_alloc_env, ir_node *src_node, ir_node *trg_node, int pos)
 190 {
 191         pbqp_t                      *pbqp        = pbqp_alloc_env->pbqp_inst;
 192         const arch_register_class_t *cls         = pbqp_alloc_env->cls;
 193         unsigned                    *restr_nodes = pbqp_alloc_env->restr_nodes;
 194         pbqp_matrix_t               *afe_matrix  = pbqp_matrix_alloc(pbqp, arch_register_class_n_regs(cls), arch_register_class_n_regs(cls));
 195         unsigned                     colors_n    = arch_register_class_n_regs(cls);
 196
 197         if (get_edge(pbqp, get_irn_idx(src_node), get_irn_idx(trg_node)) == NULL) {
 198                 if (use_exec_freq) {
 199                         /* get exec_freq for copy_block */
 200                         ir_node *root_bl = get_nodes_block(src_node);
 201                         ir_node *copy_bl = is_Phi(src_node) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
 202                         int      res     = get_block_execfreq_int(&pbqp_alloc_env->execfreq_factors, copy_bl);
 203
 204                         /* create afe-matrix */
 205                         unsigned row, col;
 206                         for (row = 0; row < colors_n; row++) {
 207                                 for (col = 0; col < colors_n; col++) {
 208                                         if (row != col)
 209                                                 pbqp_matrix_set(afe_matrix, row, col, (num)res);
 210                                 }
 211                         }
 212                 }
 213                 else {
 214                         afe_matrix = pbqp_alloc_env->aff_matrix_template;
 215                 }
 216 #if DO_USEFUL_OPT || USE_BIPARTIT_MATCHING
 217                 /* do useful optimization to speed up pbqp solving */
 218                 if (get_free_regs(restr_nodes, cls, src_node) == 1 && get_free_regs(restr_nodes, cls, trg_node) == 1) {
 219                         return;
 220                 }
 221                 if (get_free_regs(restr_nodes, cls, src_node) == 1 || get_free_regs(restr_nodes, cls, trg_node) == 1) {
 222                         if (get_free_regs(restr_nodes, cls, src_node) == 1) {
 223                                 unsigned regIdx = vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs);
 224                                 vector_add_matrix_col(get_node(pbqp, get_irn_idx(trg_node))->costs, afe_matrix, regIdx);
 225                         }
 226                         else {
 227                                 unsigned regIdx = vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs);
 228                                 vector_add_matrix_col(get_node(pbqp, get_irn_idx(src_node))->costs, afe_matrix, regIdx);
 229                         }
 230                         return;
 231                 }
 232 #endif
 233                 /* insert interference edge */
 234                 insert_edge(pbqp, src_node, trg_node, afe_matrix);
 235         }
 236 }
 237
 238 static void create_affinity_edges(ir_node *irn, void *env)
 239 {
 240         be_pbqp_alloc_env_t         *pbqp_alloc_env = (be_pbqp_alloc_env_t*)env;
 241         const arch_register_class_t *cls            = pbqp_alloc_env->cls;
 242         const arch_register_req_t   *req            = arch_get_irn_register_req(irn);
 243         unsigned                     pos;
 244         unsigned                     max;
 245
 246         if (is_Phi(irn)) { /* Phis */
 247                 for (pos = 0, max = get_irn_arity(irn); pos < max; ++pos) {
 248                         ir_node *arg = get_irn_n(irn, pos);
 249
 250                         if (!arch_irn_consider_in_reg_alloc(cls, arg))
 251                                 continue;
 252
 253                         /* no edges to itself */
 254                         if (irn == arg) {
 255                                 continue;
 256                         }
 257
 258                         insert_afe_edge(pbqp_alloc_env, irn, arg, pos);
 259                 }
 260         }
 261         else if (is_Perm_Proj(irn)) { /* Perms */
 262                 ir_node *arg = get_Perm_src(irn);
 263                 if (!arch_irn_consider_in_reg_alloc(cls, arg))
 264                         return;
 265
 266                 insert_afe_edge(pbqp_alloc_env, irn, arg, -1);
 267         } else if (arch_register_req_is(req, should_be_same)) {
 268                 const unsigned other = req->other_same;
 269                 int            i;
 270
 271                 for (i = 0; 1U << i <= other; ++i) {
 272                         if (other & (1U << i)) {
 273                                 ir_node *other = get_irn_n(skip_Proj(irn), i);
 274                                 if (!arch_irn_consider_in_reg_alloc(cls, other))
 275                                         continue;
 276
 277                                 /* no edges to itself */
 278                                 if (irn == other) {
 279                                         continue;
 280                                 }
 281
 282                                 insert_afe_edge(pbqp_alloc_env, irn, other, i);
 283                         }
 284                 }
 285         }
 286 }
 287
 288 static void create_pbqp_coloring_instance(ir_node *block, void *data)
 289 {
 290         be_pbqp_alloc_env_t         *pbqp_alloc_env     = (be_pbqp_alloc_env_t*)data;
 291         be_lv_t                     *lv                 = pbqp_alloc_env->lv;
 292         const arch_register_class_t *cls                = pbqp_alloc_env->cls;
 293         plist_t                     *rpeo               = pbqp_alloc_env->rpeo;
 294         pbqp_t                      *pbqp_inst          = pbqp_alloc_env->pbqp_inst;
 295         plist_t                     *temp_list          = plist_new();
 296         plist_element_t             *el;
 297         ir_nodeset_t                 live_nodes;
 298 #if USE_BIPARTIT_MATCHING
 299         int                         *assignment         = ALLOCAN(int, cls->n_regs);
 300 #else
 301         unsigned                    *restr_nodes        = pbqp_alloc_env->restr_nodes;
 302         pqueue_t                    *restr_nodes_queue  = new_pqueue();
 303         pqueue_t                    *queue              = new_pqueue();
 304         plist_t                     *sorted_list        = plist_new();
 305         ir_node                     *last_element       = NULL;
 306 #endif
 307
 308         /* first, determine the pressure */
 309         /* (this is only for compatibility with copymin optimization, it's not needed for pbqp coloring) */
 310         create_borders(block, pbqp_alloc_env->env);
 311
 312         /* calculate living nodes for the first step */
 313         ir_nodeset_init(&live_nodes);
 314         be_liveness_end_of_block(lv, cls, block, &live_nodes);
 315
 316         /* create pbqp nodes, interference edges and reverse perfect elimination order */
 317         sched_foreach_reverse(block, irn) {
 318                 be_foreach_value(irn, value,
 319                         if (!arch_irn_consider_in_reg_alloc(cls, value))
 320                                 continue;
 321
 322                         /* create pbqp source node if it dosn't exist */
 323                         if (!get_node(pbqp_inst, get_irn_idx(value)))
 324                                 create_pbqp_node(pbqp_alloc_env, value);
 325
 326                         /* create nodes and interference edges */
 327                         foreach_ir_nodeset(&live_nodes, live, iter) {
 328                                 /* create pbqp source node if it dosn't exist */
 329                                 if (!get_node(pbqp_inst, get_irn_idx(live)))
 330                                         create_pbqp_node(pbqp_alloc_env, live);
 331
 332                                 /* no edges to itself */
 333                                 if (value == live)
 334                                         continue;
 335
 336                                 insert_ife_edge(pbqp_alloc_env, value, live);
 337                         }
 338                 );
 339
 340                 /* get living nodes for next step */
 341                 if (!is_Phi(irn)) {
 342                         be_liveness_transfer(cls, irn, &live_nodes);
 343                 }
 344
 345 #if USE_BIPARTIT_MATCHING
 346                 if (get_irn_mode(irn) == mode_T) {
 347                         unsigned     clique_size         = 0;
 348                         unsigned     n_alloc             = 0;
 349                         pbqp_node   *clique[cls->n_regs];
 350                         bipartite_t *bp                  = bipartite_new(cls->n_regs, cls->n_regs);
 351
 352                         /* add all proj after a perm to clique */
 353                         foreach_out_edge(irn, edge) {
 354                                 ir_node *proj = get_edge_src_irn(edge);
 355
 356                                 /* ignore node if it is not necessary for register allocation */
 357                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 358                                         continue;
 359
 360                                 /* insert pbqp node into temp rpeo list of this block */
 361                                 plist_insert_front(temp_list, get_node(pbqp_inst, get_irn_idx(proj)));
 362
 363                                 if(is_Perm_Proj(proj)) {
 364                                         /* add proj to clique */
 365                                         pbqp_node *clique_member = get_node(pbqp_inst,proj->node_idx);
 366                                         vector    *costs         = clique_member->costs;
 367                                         unsigned   idx           = 0;
 368
 369                                         clique[clique_size] = clique_member;
 370
 371                                         for(idx = 0; idx < costs->len; idx++) {
 372                                                 if(costs->entries[idx].data != INF_COSTS) {
 373                                                         bipartite_add(bp, clique_size, idx);
 374                                                 }
 375                                         }
 376
 377                                         /* increase node counter */
 378                                         clique_size++;
 379                                         n_alloc++;
 380                                 }
 381                         }
 382
 383                         if(clique_size > 0) {
 384                                 plist_element_t *listElement;
 385                                 foreach_plist(temp_list, listElement) {
 386                                         pbqp_node *clique_candidate  = listElement->data;
 387                                         unsigned   idx               = 0;
 388                                         bool       isMember          = true;
 389
 390                                         /* clique size not bigger then register class size */
 391                                         if(clique_size >= cls->n_regs) break;
 392
 393                                         for(idx = 0; idx < clique_size; idx++) {
 394                                                 pbqp_node *member = clique[idx];
 395
 396                                                 if(member == clique_candidate) {
 397                                                         isMember = false;
 398                                                         break;
 399                                                 }
 400
 401                                                 if(get_edge(pbqp_inst, member->index, clique_candidate->index) == NULL && get_edge(pbqp_inst, clique_candidate->index, member->index) == NULL) {
 402                                                         isMember = false;
 403                                                         break;
 404                                                 }
 405                                         }
 406
 407                                         /* goto next list element if current node is not a member of the clique */
 408                                         if(!isMember) { continue; }
 409
 410                                         /* add candidate to clique */
 411                                         clique[clique_size] = clique_candidate;
 412
 413                                         vector *costs = clique_candidate->costs;
 414                                         for(idx = 0; idx < costs->len; idx++) {
 415                                                 if(costs->entries[idx].data != INF_COSTS) {
 416                                                         bipartite_add(bp, clique_size, idx);
 417                                                 }
 418                                         }
 419
 420                                         /* increase node counter */
 421                                         clique_size++;
 422                                 }
 423                         }
 424
 425                         /* solve bipartite matching */
 426                         bipartite_matching(bp, assignment);
 427
 428                         /* assign colors */
 429                         unsigned nodeIdx = 0;
 430                         for(nodeIdx = 0; nodeIdx < clique_size; nodeIdx++) {
 431                                 vector *costs = clique[nodeIdx]->costs;
 432                                 int     idx;
 433                                 for(idx = 0; idx < (int)costs->len; idx++) {
 434                                         if(assignment[nodeIdx] != idx) {
 435                                                 costs->entries[idx].data = INF_COSTS;
 436                                         }
 437                                 }
 438                                 assert(assignment[nodeIdx] >= 0 && "there must have been a register assigned (node not register pressure faithful?)");
 439                         }
 440
 441                         /* free memory */
 442                         bipartite_free(bp);
 443                 }
 444                 else {
 445                         if (arch_irn_consider_in_reg_alloc(cls, irn)) {
 446                                 plist_insert_front(temp_list, get_node(pbqp_inst, get_irn_idx(irn)));
 447                         }
 448                 }
 449 #else
 450                 /* order nodes for perfect elimination order */
 451                 if (get_irn_mode(irn) == mode_T) {
 452                         bool allHaveIFEdges = true;
 453                         foreach_out_edge(irn, edge) {
 454                                 ir_node *proj = get_edge_src_irn(edge);
 455                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 456                                         continue;
 457
 458                                 /* insert proj node into priority queue (descending by the number of interference edges) */
 459                                 if (get_free_regs(restr_nodes, cls, proj) <= 4) {
 460                                         pqueue_put(restr_nodes_queue, proj, pbqp_alloc_env->ife_edge_num[get_irn_idx(proj)]);
 461                                 }
 462                                 else {
 463                                         pqueue_put(queue, proj, pbqp_alloc_env->ife_edge_num[get_irn_idx(proj)]);
 464                                 }
 465
 466                                 /* skip last step if there is no last_element */
 467                                 if(last_element == NULL)
 468                                         continue;
 469
 470                                 /* check if proj has an if edge to last_element (at this time pbqp contains only if edges) */
 471                                 if(get_edge(pbqp_inst, proj->node_idx, last_element->node_idx) == NULL && get_edge(pbqp_inst, last_element->node_idx, proj->node_idx) == NULL) {
 472                                         allHaveIFEdges = false; /* there is no if edge between proj and last_element */
 473                                 }
 474                         }
 475
 476                         if(last_element != NULL && allHaveIFEdges) {
 477                                 if (get_free_regs(restr_nodes, cls, last_element) <= 4) {
 478                                         pqueue_put(restr_nodes_queue, last_element, pbqp_alloc_env->ife_edge_num[get_irn_idx(last_element)]);
 479                                 }
 480                                 else {
 481                                         pqueue_put(queue, last_element, pbqp_alloc_env->ife_edge_num[get_irn_idx(last_element)]);
 482                                 }
 483                                 plist_erase(temp_list, plist_find_value(temp_list, get_node(pbqp_inst, last_element->node_idx)));
 484                                 last_element = NULL;
 485                         }
 486
 487                         /* first insert all restricted proj nodes */
 488                         while (!pqueue_empty(restr_nodes_queue)) {
 489                                 ir_node *node = (ir_node*)pqueue_pop_front(restr_nodes_queue);
 490                                 plist_insert_front(sorted_list, get_node(pbqp_inst, get_irn_idx(node)));
 491                         }
 492
 493                         /* insert proj nodes descending by their number of interference edges */
 494                         while (!pqueue_empty(queue)) {
 495                                 ir_node *node = (ir_node*)pqueue_pop_front(queue);
 496                                 plist_insert_front(sorted_list, get_node(pbqp_inst, get_irn_idx(node)));
 497                         }
 498
 499                         /* invert sorted list */
 500                         foreach_plist(sorted_list, el) {
 501                                 plist_insert_front(temp_list, el->data);
 502                         }
 503
 504                         plist_clear(sorted_list);
 505
 506                 }
 507                 else {
 508                         if (arch_irn_consider_in_reg_alloc(cls, irn)) {
 509                                 // remember last colorable node
 510                                 last_element = irn;
 511                                 plist_insert_front(temp_list, get_node(pbqp_inst, get_irn_idx(irn)));
 512                         }
 513                         else {
 514                                 // node not colorable, so ignore it
 515                                 last_element = NULL;
 516                         }
 517                 }
 518 #endif
 519         }
 520
 521         /* add the temp rpeo list of this block to the global reverse perfect elimination order list*/
 522         foreach_plist(temp_list, el) {
 523                 plist_insert_back(rpeo, el->data);
 524         }
 525
 526         /* free reserved memory */
 527         ir_nodeset_destroy(&live_nodes);
 528         plist_free(temp_list);
 529 #if USE_BIPARTIT_MATCHING
 530 #else
 531         plist_free(sorted_list);
 532         del_pqueue(queue);
 533         del_pqueue(restr_nodes_queue);
 534 #endif
 535 }
 536
 537 static void insert_perms(ir_node *block, void *data)
 538 {
 539         be_chordal_env_t *env    = (be_chordal_env_t*)data;
 540
 541         sched_foreach_safe(block, irn) {
 542                 be_insn_t *insn = be_scan_insn(env, irn);
 543                 if (insn)
 544                         pre_process_constraints(env, &insn);
 545         }
 546 }
 547
 548 static void be_pbqp_coloring(be_chordal_env_t *env)
 549 {
 550         ir_graph                    *irg            = env->irg;
 551         const arch_register_class_t *cls            = env->cls;
 552         be_lv_t                     *lv             = NULL;
 553         plist_element_t             *element        = NULL;
 554         unsigned                     colors_n       = arch_register_class_n_regs(cls);
 555         be_pbqp_alloc_env_t          pbqp_alloc_env;
 556         unsigned                     col;
 557         unsigned                     row;
 558         pbqp_matrix_t               *ife_matrix;
 559         num                          solution;
 560 #if KAPS_DUMP
 561         FILE                        *file_before;
 562 #endif
 563 #if TIMER
 564         ir_timer_t *t_ra_pbqp_alloc_create     = ir_timer_new();
 565         ir_timer_t *t_ra_pbqp_alloc_solve      = ir_timer_new();
 566         ir_timer_t *t_ra_pbqp_alloc_create_aff = ir_timer_new();
 567
 568         printf("#### ----- === Allocating registers of %s (%s) ===\n", cls->name, get_entity_name(get_irg_entity(irg)));
 569 #endif
 570         be_assure_live_sets(irg);
 571         lv = be_get_irg_liveness(irg);
 572
 573         /* insert perms */
 574         assure_doms(irg);
 575         dom_tree_walk_irg(irg, insert_perms, NULL, env);
 576
 577         /* dump graph after inserting perms */
 578         if (env->opts->dump_flags & BE_CH_DUMP_CONSTR) {
 579                 char buf[256];
 580                 snprintf(buf, sizeof(buf), "%s-constr", cls->name);
 581                 dump_ir_graph(irg, buf);
 582         }
 583
 584         ir_calculate_execfreq_int_factors(&pbqp_alloc_env.execfreq_factors, irg);
 585
 586         /* initialize pbqp allocation data structure */
 587         pbqp_alloc_env.pbqp_inst        = alloc_pbqp(get_irg_last_idx(irg));  /* initialize pbqp instance */
 588         pbqp_alloc_env.cls              = cls;
 589         pbqp_alloc_env.irg              = irg;
 590         pbqp_alloc_env.lv               = lv;
 591         pbqp_alloc_env.allocatable_regs = env->allocatable_regs;
 592         pbqp_alloc_env.rpeo             = plist_new();
 593         pbqp_alloc_env.restr_nodes      = XMALLOCNZ(unsigned, get_irg_last_idx(irg));
 594         pbqp_alloc_env.ife_edge_num     = XMALLOCNZ(unsigned, get_irg_last_idx(irg));
 595         pbqp_alloc_env.env              = env;
 596
 597         /* create costs matrix template for interference edges */
 598         ife_matrix = pbqp_matrix_alloc(pbqp_alloc_env.pbqp_inst, colors_n, colors_n);
 599         /* set costs */
 600         for (row = 0, col = 0; row < colors_n; row++, col++)
 601                 pbqp_matrix_set(ife_matrix, row, col, INF_COSTS);
 602
 603         pbqp_alloc_env.ife_matrix_template = ife_matrix;
 604
 605
 606         if (!use_exec_freq) {
 607                 /* create costs matrix template for affinity edges */
 608                 pbqp_matrix_t *afe_matrix = pbqp_matrix_alloc(pbqp_alloc_env.pbqp_inst, colors_n, colors_n);
 609                 /* set costs */
 610                 for (row = 0; row < colors_n; row++) {
 611                         for (col = 0; col < colors_n; col++) {
 612                                 if (row != col)
 613                                         pbqp_matrix_set(afe_matrix, row, col, 2);
 614                         }
 615                 }
 616                 pbqp_alloc_env.aff_matrix_template = afe_matrix;
 617         }
 618
 619
 620         /* create pbqp instance */
 621 #if TIMER
 622         ir_timer_reset_and_start(t_ra_pbqp_alloc_create);
 623 #endif
 624         assure_doms(irg);
 625         dom_tree_walk_irg(irg, create_pbqp_coloring_instance , NULL, &pbqp_alloc_env);
 626 #if TIMER
 627         ir_timer_stop(t_ra_pbqp_alloc_create);
 628 #endif
 629
 630
 631         /* set up affinity edges */
 632 #if TIMER
 633         ir_timer_reset_and_start(t_ra_pbqp_alloc_create_aff);
 634 #endif
 635         foreach_plist(pbqp_alloc_env.rpeo, element) {
 636                 pbqp_node_t *node = (pbqp_node_t*)element->data;
 637                 ir_node     *irn  = get_idx_irn(irg, node->index);
 638
 639                 create_affinity_edges(irn, &pbqp_alloc_env);
 640         }
 641 #if TIMER
 642         ir_timer_stop(t_ra_pbqp_alloc_create_aff);
 643 #endif
 644
 645
 646 #if KAPS_DUMP
 647         // dump graph before solving pbqp
 648         file_before = my_open(env, "", "-pbqp_coloring.html");
 649         set_dumpfile(pbqp_alloc_env.pbqp_inst, file_before);
 650 #endif
 651
 652         /* print out reverse perfect elimination order */
 653 #if PRINT_RPEO
 654         {
 655                 plist_element_t *elements;
 656                 foreach_plist(pbqp_alloc_env.rpeo, elements) {
 657                         pbqp_node_t *node = elements->data;
 658                         printf(" %d(%ld);", node->index, get_idx_irn(irg, node->index)->node_nr);
 659                 }
 660                 printf("\n");
 661         }
 662 #endif
 663
 664         /* solve pbqp instance */
 665 #if TIMER
 666         ir_timer_reset_and_start(t_ra_pbqp_alloc_solve);
 667 #endif
 668         if(use_late_decision) {
 669                 solve_pbqp_heuristical_co_ld(pbqp_alloc_env.pbqp_inst,pbqp_alloc_env.rpeo);
 670         }
 671         else {
 672                 solve_pbqp_heuristical_co(pbqp_alloc_env.pbqp_inst,pbqp_alloc_env.rpeo);
 673         }
 674 #if TIMER
 675         ir_timer_stop(t_ra_pbqp_alloc_solve);
 676 #endif
 677
 678
 679         solution = get_solution(pbqp_alloc_env.pbqp_inst);
 680         if (solution == INF_COSTS)
 681                 panic("No PBQP solution found");
 682
 683
 684         /* assign colors */
 685         foreach_plist(pbqp_alloc_env.rpeo, element) {
 686                 pbqp_node_t           *node  = (pbqp_node_t*)element->data;
 687                 ir_node               *irn   = get_idx_irn(irg, node->index);
 688                 num                    color = get_node_solution(pbqp_alloc_env.pbqp_inst, node->index);
 689                 const arch_register_t *reg   = arch_register_for_index(cls, color);
 690
 691                 arch_set_irn_register(irn, reg);
 692         }
 693
 694
 695 #if TIMER
 696         printf("PBQP alloc create:     %10.3lf msec\n",
 697                (double)ir_timer_elapsed_usec(t_ra_pbqp_alloc_create) / 1000.0);
 698         printf("PBQP alloc solve:      %10.3lf msec\n",
 699                (double)ir_timer_elapsed_usec(t_ra_pbqp_alloc_solve) / 1000.0);
 700         printf("PBQP alloc create aff: %10.3lf msec\n",
 701                (double)ir_timer_elapsed_usec(t_ra_pbqp_alloc_create_aff) / 1000.0);
 702 #endif
 703
 704
 705         /* free reserved memory */
 706 #if KAPS_DUMP
 707         fclose(file_before);
 708 #endif
 709         free_pbqp(pbqp_alloc_env.pbqp_inst);
 710         plist_free(pbqp_alloc_env.rpeo);
 711         xfree(pbqp_alloc_env.restr_nodes);
 712         xfree(pbqp_alloc_env.ife_edge_num);
 713 }
 714
 715
 716 /**
 717  * Initializes this module.
 718  */
 719 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_pbqp_coloring)
 720 void be_init_pbqp_coloring(void)
 721 {
 722         lc_opt_entry_t *be_grp       = lc_opt_get_grp(firm_opt_get_root(), "be");
 723         lc_opt_entry_t *ra_grp       = lc_opt_get_grp(be_grp, "ra");
 724         lc_opt_entry_t *chordal_grp  = lc_opt_get_grp(ra_grp, "chordal");
 725         lc_opt_entry_t *coloring_grp = lc_opt_get_grp(chordal_grp, "coloring");
 726         lc_opt_entry_t *pbqp_grp     = lc_opt_get_grp(coloring_grp, "pbqp");
 727
 728         static be_ra_chordal_coloring_t coloring = {
 729                 be_pbqp_coloring
 730         };
 731
 732         lc_opt_add_table(pbqp_grp, options);
 733         be_register_chordal_coloring("pbqp", &coloring);
 734 }