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

   1 /*
   2  * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
   3  *
   4  * This file is part of libFirm.
   5  *
   6  * This file may be distributed and/or modified under the terms of the
   7  * GNU General Public License version 2 as published by the Free Software
   8  * Foundation and appearing in the file LICENSE.GPL included in the
   9  * packaging of this file.
  10  *
  11  * Licensees holding valid libFirm Professional Edition licenses may use
  12  * this file in accordance with the libFirm Commercial License.
  13  * Agreement provided with the Software.
  14  *
  15  * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
  16  * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  17  * PURPOSE.
  18  */
  19
  20 /**
  21  * @file
  22  * @brief       PBQP based register allocation.
  23  * @author      Thomas Bersch
  24  * @date        27.11.2009
  25  * @version     $Id: bechordal.c 26750 2009-11-27 09:37:43Z bersch $
  26  */
  27
  28 /* miscellaneous includes */
  29 #include "config.h"
  30
  31 #include "debug.h"
  32 #include "error.h"
  33
  34 #include "irdom.h"
  35 #include "irdump.h"
  36 #include "iredges_t.h"
  37 #include "irprintf.h"
  38 #include "irgwalk.h"
  39 #include "irtools.h"
  40 #include "time.h"
  41
  42 /* libfirm/ir/adt includes */
  43 #include "bipartite.h"
  44
  45 /* libfirm/ir/be includes */
  46 #include "bearch.h"
  47 #include "beirg.h"
  48 #include "besched.h"
  49 #include "bemodule.h"
  50 #include "bechordal_common.h"
  51 #include "bechordal.h"
  52 #include "bechordal_t.h"
  53 #include "beinsn_t.h"
  54 #include "benode.h"
  55 #include "belive.h"
  56 #include "belive_t.h"
  57 #include "beutil.h"
  58 #include "plist.h"
  59 #include "pqueue.h"
  60 #include "becopyopt.h"
  61
  62 /* pbqp includes */
  63 #include "kaps.h"
  64 #include "matrix.h"
  65 #include "vector.h"
  66 #include "vector_t.h"
  67 #include "heuristical_co.h"
  68 #include "heuristical_co_ld.h"
  69 #include "pbqp_t.h"
  70 #include "html_dumper.h"
  71 #include "pbqp_node_t.h"
  72 #include "pbqp_node.h"
  73 #include "pbqp_edge_t.h"
  74
  75 #define TIMER                 0
  76 #define PRINT_RPEO            0
  77 #define USE_BIPARTIT_MATCHING 0
  78 #define DO_USEFUL_OPT         1
  79
  80
  81 static int use_exec_freq     = true;
  82 static int use_late_decision = false;
  83
  84 typedef struct be_pbqp_alloc_env_t {
  85         pbqp_t                      *pbqp_inst;         /**< PBQP instance for register allocation */
  86         ir_graph                    *irg;               /**< The graph under examination. */
  87         const arch_register_class_t *cls;               /**< Current processed register class */
  88         be_lv_t                     *lv;
  89         bitset_t                    *allocatable_regs;
  90         pbqp_matrix_t               *ife_matrix_template;
  91         pbqp_matrix_t               *aff_matrix_template;
  92         plist_t                     *rpeo;
  93         unsigned                    *restr_nodes;
  94         unsigned                    *ife_edge_num;
  95         be_chordal_env_t            *env;
  96 } be_pbqp_alloc_env_t;
  97
  98
  99 #define is_Reg_Phi(irn)                                        (is_Phi(irn) && mode_is_data(get_irn_mode(irn)))
 100 #define get_Perm_src(irn)                                      (get_irn_n(get_Proj_pred(irn), get_Proj_proj(irn)))
 101 #define is_Perm_Proj(irn)                                      (is_Proj(irn) && be_is_Perm(get_Proj_pred(irn)))
 102 #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)))
 103 #define get_free_regs(restr_nodes, cls, irn)                   (arch_register_class_n_regs(cls) - restr_nodes[get_irn_idx(irn)])
 104
 105 static inline int is_2addr_code(const arch_register_req_t *req)
 106 {
 107         return (req->type & arch_register_req_type_should_be_same) != 0;
 108 }
 109
 110 static const lc_opt_table_entry_t options[] = {
 111         LC_OPT_ENT_BOOL("exec_freq", "use exec_freq",  &use_exec_freq),
 112         LC_OPT_ENT_BOOL("late_decision", "use late decision for register allocation",  &use_late_decision),
 113         LC_OPT_LAST
 114 };
 115
 116 #if KAPS_DUMP
 117 static FILE *my_open(const be_chordal_env_t *env, const char *prefix, const char *suffix)
 118 {
 119         FILE       *result;
 120         char        buf[1024];
 121         size_t      i;
 122         size_t      n;
 123         char       *tu_name;
 124         const char *cup_name = be_get_irg_main_env(env->irg)->cup_name;
 125
 126         n = strlen(cup_name);
 127         tu_name = XMALLOCN(char, n + 1);
 128         strcpy(tu_name, cup_name);
 129         for (i = 0; i < n; ++i)
 130                 if (tu_name[i] == '.')
 131                         tu_name[i] = '_';
 132
 133         ir_snprintf(buf, sizeof(buf), "%s%s_%F_%s%s", prefix, tu_name, env->irg, env->cls->name, suffix);
 134         xfree(tu_name);
 135         result = fopen(buf, "wt");
 136         if (result == NULL) {
 137                 panic("Couldn't open '%s' for writing.", buf);
 138         }
 139
 140         return result;
 141 }
 142 #endif
 143
 144
 145 static void create_pbqp_node(be_pbqp_alloc_env_t *pbqp_alloc_env, ir_node *irn)
 146 {
 147         const arch_register_class_t *cls = pbqp_alloc_env->cls;
 148         pbqp_t   *pbqp_inst              = pbqp_alloc_env->pbqp_inst;
 149         bitset_t *allocatable_regs       = pbqp_alloc_env->allocatable_regs;
 150         unsigned  colors_n               = arch_register_class_n_regs(cls);
 151         unsigned  cntConstrains          = 0;
 152
 153         /* create costs vector depending on register constrains */
 154         vector_t *costs_vector = vector_alloc(pbqp_inst, colors_n);
 155
 156         /* set costs depending on register constrains */
 157         unsigned idx;
 158         for (idx = 0; idx < colors_n; idx++) {
 159                 const arch_register_req_t *req = arch_get_irn_register_req(irn);
 160                 const arch_register_t     *reg = arch_register_for_index(cls, idx);
 161                 if (!bitset_is_set(allocatable_regs, idx)
 162                     || !arch_reg_is_allocatable(req, reg)) {
 163                         /* constrained */
 164                         vector_set(costs_vector, idx, INF_COSTS);
 165                         cntConstrains++;
 166                 }
 167         }
 168
 169         /* add vector to pbqp node */
 170         add_node_costs(pbqp_inst, get_irn_idx(irn), costs_vector);
 171         pbqp_alloc_env->restr_nodes[get_irn_idx(irn)] = cntConstrains;
 172 }
 173
 174 static void insert_ife_edge(be_pbqp_alloc_env_t *pbqp_alloc_env, ir_node *src_node, ir_node *trg_node)
 175 {
 176         pbqp_t                      *pbqp                = pbqp_alloc_env->pbqp_inst;
 177         const arch_register_class_t *cls                 = pbqp_alloc_env->cls;
 178         pbqp_matrix_t               *ife_matrix_template = pbqp_alloc_env->ife_matrix_template;
 179         unsigned                    *restr_nodes         = pbqp_alloc_env->restr_nodes;
 180
 181         if (get_edge(pbqp, get_irn_idx(src_node), get_irn_idx(trg_node)) == NULL) {
 182
 183                 /* increase ife edge counter */
 184                 pbqp_alloc_env->ife_edge_num[get_irn_idx(src_node)]++;
 185                 pbqp_alloc_env->ife_edge_num[get_irn_idx(trg_node)]++;
 186
 187 #if DO_USEFUL_OPT || USE_BIPARTIT_MATCHING
 188                 /* do useful optimization to speed up pbqp solving (we can do this because we know our matrix) */
 189                 if (get_free_regs(restr_nodes, cls, src_node) == 1 && get_free_regs(restr_nodes, cls, trg_node) == 1) {
 190                         assert(vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs) !=
 191                                vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs) &&
 192                                "Interfering nodes must not have the same register!");
 193                         return;
 194                 }
 195                 if (get_free_regs(restr_nodes, cls, src_node) == 1 || get_free_regs(restr_nodes, cls, trg_node) == 1) {
 196                         if (get_free_regs(restr_nodes, cls, src_node) == 1) {
 197                                 unsigned idx = vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs);
 198                                 vector_set(get_node(pbqp, get_irn_idx(trg_node))->costs, idx, INF_COSTS);
 199                         }
 200                         else {
 201                                 unsigned idx = vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs);
 202                                 vector_set(get_node(pbqp, get_irn_idx(src_node))->costs, idx, INF_COSTS);
 203                         }
 204                         return;
 205                 }
 206 #endif
 207                 /* insert interference edge */
 208                 insert_edge(pbqp, src_node, trg_node, ife_matrix_template);
 209         }
 210 }
 211
 212 static void insert_afe_edge(be_pbqp_alloc_env_t *pbqp_alloc_env, ir_node *src_node, ir_node *trg_node, int pos)
 213 {
 214         pbqp_t                      *pbqp        = pbqp_alloc_env->pbqp_inst;
 215         const arch_register_class_t *cls         = pbqp_alloc_env->cls;
 216         unsigned                    *restr_nodes = pbqp_alloc_env->restr_nodes;
 217         pbqp_matrix_t               *afe_matrix  = pbqp_matrix_alloc(pbqp, arch_register_class_n_regs(cls), arch_register_class_n_regs(cls));
 218         unsigned                     colors_n    = arch_register_class_n_regs(cls);
 219
 220         if (get_edge(pbqp, get_irn_idx(src_node), get_irn_idx(trg_node)) == NULL) {
 221                 if (use_exec_freq) {
 222                         /* get exec_freq for copy_block */
 223                         ir_node       *root_bl   = get_nodes_block(src_node);
 224                         ir_node       *copy_bl   = is_Phi(src_node) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
 225                         ir_exec_freq  *exec_freq = be_get_irg_exec_freq(pbqp_alloc_env->irg);
 226                         unsigned long  res       = get_block_execfreq_ulong(exec_freq, copy_bl);
 227
 228                         /* create afe-matrix */
 229                         unsigned row, col;
 230                         for (row = 0; row < colors_n; row++) {
 231                                 for (col = 0; col < colors_n; col++) {
 232                                         if (row != col)
 233                                                 pbqp_matrix_set(afe_matrix, row, col, (num)res);
 234                                 }
 235                         }
 236                 }
 237                 else {
 238                         afe_matrix = pbqp_alloc_env->aff_matrix_template;
 239                 }
 240 #if DO_USEFUL_OPT || USE_BIPARTIT_MATCHING
 241                 /* do useful optimization to speed up pbqp solving */
 242                 if (get_free_regs(restr_nodes, cls, src_node) == 1 && get_free_regs(restr_nodes, cls, trg_node) == 1) {
 243                         return;
 244                 }
 245                 if (get_free_regs(restr_nodes, cls, src_node) == 1 || get_free_regs(restr_nodes, cls, trg_node) == 1) {
 246                         if (get_free_regs(restr_nodes, cls, src_node) == 1) {
 247                                 unsigned regIdx = vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs);
 248                                 vector_add_matrix_col(get_node(pbqp, get_irn_idx(trg_node))->costs, afe_matrix, regIdx);
 249                         }
 250                         else {
 251                                 unsigned regIdx = vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs);
 252                                 vector_add_matrix_col(get_node(pbqp, get_irn_idx(src_node))->costs, afe_matrix, regIdx);
 253                         }
 254                         return;
 255                 }
 256 #endif
 257                 /* insert interference edge */
 258                 insert_edge(pbqp, src_node, trg_node, afe_matrix);
 259         }
 260 }
 261
 262 static void create_affinity_edges(ir_node *irn, void *env)
 263 {
 264         be_pbqp_alloc_env_t         *pbqp_alloc_env = (be_pbqp_alloc_env_t*)env;
 265         const arch_register_class_t *cls            = pbqp_alloc_env->cls;
 266         const arch_register_req_t   *req            = arch_get_irn_register_req(irn);
 267         unsigned                     pos;
 268         unsigned                     max;
 269
 270         if (is_Reg_Phi(irn)) { /* Phis */
 271                 for (pos = 0, max = get_irn_arity(irn); pos < max; ++pos) {
 272                         ir_node *arg = get_irn_n(irn, pos);
 273
 274                         if (!arch_irn_consider_in_reg_alloc(cls, arg))
 275                                 continue;
 276
 277                         /* no edges to itself */
 278                         if (irn == arg) {
 279                                 continue;
 280                         }
 281
 282                         insert_afe_edge(pbqp_alloc_env, irn, arg, pos);
 283                 }
 284         }
 285         else if (is_Perm_Proj(irn)) { /* Perms */
 286                 ir_node *arg = get_Perm_src(irn);
 287                 if (!arch_irn_consider_in_reg_alloc(cls, arg))
 288                         return;
 289
 290                 insert_afe_edge(pbqp_alloc_env, irn, arg, -1);
 291         }
 292         else { /* 2-address code */
 293                 if (is_2addr_code(req)) {
 294                         const unsigned other = req->other_same;
 295                         int            i;
 296
 297                         for (i = 0; 1U << i <= other; ++i) {
 298                                 if (other & (1U << i)) {
 299                                         ir_node *other = get_irn_n(skip_Proj(irn), i);
 300                                         if (!arch_irn_consider_in_reg_alloc(cls, other))
 301                                                 continue;
 302
 303                                         /* no edges to itself */
 304                                         if (irn == other) {
 305                                                 continue;
 306                                         }
 307
 308                                         insert_afe_edge(pbqp_alloc_env, irn, other, i);
 309                                 }
 310                         }
 311                 }
 312         }
 313 }
 314
 315 static void create_pbqp_coloring_instance(ir_node *block, void *data)
 316 {
 317         be_pbqp_alloc_env_t         *pbqp_alloc_env     = (be_pbqp_alloc_env_t*)data;
 318         be_lv_t                     *lv                 = pbqp_alloc_env->lv;
 319         const arch_register_class_t *cls                = pbqp_alloc_env->cls;
 320         plist_t                     *rpeo               = pbqp_alloc_env->rpeo;
 321         pbqp_t                      *pbqp_inst          = pbqp_alloc_env->pbqp_inst;
 322         plist_t                     *temp_list          = plist_new();
 323         plist_element_t             *el;
 324         ir_node                     *irn;
 325         ir_nodeset_t                 live_nodes;
 326 #if USE_BIPARTIT_MATCHING
 327         int                         *assignment         = ALLOCAN(int, cls->n_regs);
 328 #else
 329         unsigned                    *restr_nodes        = pbqp_alloc_env->restr_nodes;
 330         pqueue_t                    *restr_nodes_queue  = new_pqueue();
 331         pqueue_t                    *queue              = new_pqueue();
 332         plist_t                     *sorted_list        = plist_new();
 333         ir_node                     *last_element       = NULL;
 334 #endif
 335
 336         /* first, determine the pressure */
 337         /* (this is only for compatibility with copymin optimization, it's not needed for pbqp coloring) */
 338         create_borders(block, pbqp_alloc_env->env);
 339
 340         /* calculate living nodes for the first step */
 341         ir_nodeset_init(&live_nodes);
 342         be_liveness_end_of_block(lv, cls, block, &live_nodes);
 343
 344         /* create pbqp nodes, interference edges and reverse perfect elimination order */
 345         sched_foreach_reverse(block, irn) {
 346                 ir_node               *live;
 347                 ir_nodeset_iterator_t  iter;
 348
 349                 if (get_irn_mode(irn) == mode_T) {
 350                         const ir_edge_t *edge;
 351                         foreach_out_edge(irn, edge) {
 352                                 ir_node *proj = get_edge_src_irn(edge);
 353                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 354                                         continue;
 355
 356                                 /* create pbqp source node if it dosn't exist */
 357                                 if (get_node(pbqp_inst, get_irn_idx(proj)) == NULL) {
 358                                         create_pbqp_node(pbqp_alloc_env, proj);
 359                                 }
 360
 361                                 /* create nodes and interference edges */
 362                                 foreach_ir_nodeset(&live_nodes, live, iter) {
 363                                         /* create pbqp source node if it dosn't exist */
 364                                         if (get_node(pbqp_inst, get_irn_idx(live)) == NULL) {
 365                                                 create_pbqp_node(pbqp_alloc_env, live);
 366                                         }
 367
 368                                         /* no edges to itself */
 369                                         if (proj == live) {
 370                                                 continue;
 371                                         }
 372
 373                                         insert_ife_edge(pbqp_alloc_env, proj, live);
 374                                 }
 375                         }
 376                 }
 377                 else {
 378                         if (arch_irn_consider_in_reg_alloc(cls, irn)) {
 379                                 /* create pbqp source node if it dosn't exist */
 380                                 if (get_node(pbqp_inst, get_irn_idx(irn)) == NULL) {
 381                                         create_pbqp_node(pbqp_alloc_env, irn);
 382                                 }
 383
 384                                 /* create nodes and interference edges */
 385                                 foreach_ir_nodeset(&live_nodes, live, iter) {
 386                                         /* create pbqp source node if it dosn't exist */
 387                                         if (get_node(pbqp_inst, get_irn_idx(live)) == NULL) {
 388                                                 create_pbqp_node(pbqp_alloc_env, live);
 389                                         }
 390
 391                                         /* no edges to itself */
 392                                         if (irn == live) {
 393                                                 continue;
 394                                         }
 395
 396                                         /* insert interference edge */
 397                                         insert_ife_edge(pbqp_alloc_env, irn, live);
 398                                 }
 399                         }
 400                 }
 401
 402                 /* get living nodes for next step */
 403                 if (!is_Phi(irn)) {
 404                         be_liveness_transfer(cls, irn, &live_nodes);
 405                 }
 406
 407 #if USE_BIPARTIT_MATCHING
 408                 if (get_irn_mode(irn) == mode_T) {
 409                         unsigned     clique_size         = 0;
 410                         unsigned     n_alloc             = 0;
 411                         pbqp_node   *clique[cls->n_regs];
 412                         bipartite_t *bp                  = bipartite_new(cls->n_regs, cls->n_regs);
 413
 414                         /* add all proj after a perm to clique */
 415                         const ir_edge_t *edge;
 416                         foreach_out_edge(irn, edge) {
 417                                 ir_node *proj = get_edge_src_irn(edge);
 418
 419                                 /* ignore node if it is not necessary for register allocation */
 420                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 421                                         continue;
 422
 423                                 /* insert pbqp node into temp rpeo list of this block */
 424                                 plist_insert_front(temp_list, get_node(pbqp_inst, get_irn_idx(proj)));
 425
 426                                 if(is_Perm_Proj(proj)) {
 427                                         /* add proj to clique */
 428                                         pbqp_node *clique_member = get_node(pbqp_inst,proj->node_idx);
 429                                         vector    *costs         = clique_member->costs;
 430                                         unsigned   idx           = 0;
 431
 432                                         clique[clique_size] = clique_member;
 433
 434                                         for(idx = 0; idx < costs->len; idx++) {
 435                                                 if(costs->entries[idx].data != INF_COSTS) {
 436                                                         bipartite_add(bp, clique_size, idx);
 437                                                 }
 438                                         }
 439
 440                                         /* increase node counter */
 441                                         clique_size++;
 442                                         n_alloc++;
 443                                 }
 444                         }
 445
 446                         if(clique_size > 0) {
 447                                 plist_element_t *listElement;
 448                                 foreach_plist(temp_list, listElement) {
 449                                         pbqp_node *clique_candidate  = listElement->data;
 450                                         unsigned   idx               = 0;
 451                                         bool       isMember          = true;
 452
 453                                         /* clique size not bigger then register class size */
 454                                         if(clique_size >= cls->n_regs) break;
 455
 456                                         for(idx = 0; idx < clique_size; idx++) {
 457                                                 pbqp_node *member = clique[idx];
 458
 459                                                 if(member == clique_candidate) {
 460                                                         isMember = false;
 461                                                         break;
 462                                                 }
 463
 464                                                 if(get_edge(pbqp_inst, member->index, clique_candidate->index) == NULL && get_edge(pbqp_inst, clique_candidate->index, member->index) == NULL) {
 465                                                         isMember = false;
 466                                                         break;
 467                                                 }
 468                                         }
 469
 470                                         /* goto next list element if current node is not a member of the clique */
 471                                         if(!isMember) { continue; }
 472
 473                                         /* add candidate to clique */
 474                                         clique[clique_size] = clique_candidate;
 475
 476                                         vector *costs = clique_candidate->costs;
 477                                         for(idx = 0; idx < costs->len; idx++) {
 478                                                 if(costs->entries[idx].data != INF_COSTS) {
 479                                                         bipartite_add(bp, clique_size, idx);
 480                                                 }
 481                                         }
 482
 483                                         /* increase node counter */
 484                                         clique_size++;
 485                                 }
 486                         }
 487
 488                         /* solve bipartite matching */
 489                         bipartite_matching(bp, assignment);
 490
 491                         /* assign colors */
 492                         unsigned nodeIdx = 0;
 493                         for(nodeIdx = 0; nodeIdx < clique_size; nodeIdx++) {
 494                                 vector *costs = clique[nodeIdx]->costs;
 495                                 int     idx;
 496                                 for(idx = 0; idx < (int)costs->len; idx++) {
 497                                         if(assignment[nodeIdx] != idx) {
 498                                                 costs->entries[idx].data = INF_COSTS;
 499                                         }
 500                                 }
 501                                 assert(assignment[nodeIdx] >= 0 && "there must have been a register assigned (node not register pressure faithful?)");
 502                         }
 503
 504                         /* free memory */
 505                         bipartite_free(bp);
 506                 }
 507                 else {
 508                         if (arch_irn_consider_in_reg_alloc(cls, irn)) {
 509                                 plist_insert_front(temp_list, get_node(pbqp_inst, get_irn_idx(irn)));
 510                         }
 511                 }
 512 #else
 513                 /* order nodes for perfect elimination order */
 514                 if (get_irn_mode(irn) == mode_T) {
 515                         bool             allHaveIFEdges = true;
 516                         const ir_edge_t *edge;
 517
 518                         foreach_out_edge(irn, edge) {
 519                                 ir_node *proj = get_edge_src_irn(edge);
 520                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 521                                         continue;
 522
 523                                 /* insert proj node into priority queue (descending by the number of interference edges) */
 524                                 if (get_free_regs(restr_nodes, cls, proj) <= 4) {
 525                                         pqueue_put(restr_nodes_queue, proj, pbqp_alloc_env->ife_edge_num[get_irn_idx(proj)]);
 526                                 }
 527                                 else {
 528                                         pqueue_put(queue, proj, pbqp_alloc_env->ife_edge_num[get_irn_idx(proj)]);
 529                                 }
 530
 531                                 /* skip last step if there is no last_element */
 532                                 if(last_element == NULL)
 533                                         continue;
 534
 535                                 /* check if proj has an if edge to last_element (at this time pbqp contains only if edges) */
 536                                 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) {
 537                                         allHaveIFEdges = false; /* there is no if edge between proj and last_element */
 538                                 }
 539                         }
 540
 541                         if(last_element != NULL && allHaveIFEdges) {
 542                                 if (get_free_regs(restr_nodes, cls, last_element) <= 4) {
 543                                         pqueue_put(restr_nodes_queue, last_element, pbqp_alloc_env->ife_edge_num[get_irn_idx(last_element)]);
 544                                 }
 545                                 else {
 546                                         pqueue_put(queue, last_element, pbqp_alloc_env->ife_edge_num[get_irn_idx(last_element)]);
 547                                 }
 548                                 plist_erase(temp_list, plist_find_value(temp_list, get_node(pbqp_inst, last_element->node_idx)));
 549                                 last_element = NULL;
 550                         }
 551
 552                         /* first insert all restricted proj nodes */
 553                         while (!pqueue_empty(restr_nodes_queue)) {
 554                                 ir_node *node = (ir_node*)pqueue_pop_front(restr_nodes_queue);
 555                                 plist_insert_front(sorted_list, get_node(pbqp_inst, get_irn_idx(node)));
 556                         }
 557
 558                         /* insert proj nodes descending by their number of interference edges */
 559                         while (!pqueue_empty(queue)) {
 560                                 ir_node *node = (ir_node*)pqueue_pop_front(queue);
 561                                 plist_insert_front(sorted_list, get_node(pbqp_inst, get_irn_idx(node)));
 562                         }
 563
 564                         /* invert sorted list */
 565                         foreach_plist(sorted_list, el) {
 566                                 plist_insert_front(temp_list, el->data);
 567                         }
 568
 569                         plist_clear(sorted_list);
 570
 571                 }
 572                 else {
 573                         if (arch_irn_consider_in_reg_alloc(cls, irn)) {
 574                                 // remember last colorable node
 575                                 last_element = irn;
 576                                 plist_insert_front(temp_list, get_node(pbqp_inst, get_irn_idx(irn)));
 577                         }
 578                         else {
 579                                 // node not colorable, so ignore it
 580                                 last_element = NULL;
 581                         }
 582                 }
 583 #endif
 584         }
 585
 586         /* add the temp rpeo list of this block to the global reverse perfect elimination order list*/
 587         foreach_plist(temp_list, el) {
 588                 plist_insert_back(rpeo, el->data);
 589         }
 590
 591         /* free reserved memory */
 592         ir_nodeset_destroy(&live_nodes);
 593         plist_free(temp_list);
 594 #if USE_BIPARTIT_MATCHING
 595 #else
 596         plist_free(sorted_list);
 597         del_pqueue(queue);
 598         del_pqueue(restr_nodes_queue);
 599 #endif
 600 }
 601
 602 static void insert_perms(ir_node *block, void *data)
 603 {
 604         be_chordal_env_t *env    = (be_chordal_env_t*)data;
 605         ir_node          *irn;
 606
 607         for (irn = sched_first(block); !sched_is_end(irn);) {
 608                 be_insn_t *insn = chordal_scan_insn(env, irn);
 609                 irn             = insn->next_insn;
 610
 611                 if (!insn->has_constraints)
 612                         continue;
 613
 614                 pre_process_constraints(env, &insn);
 615         }
 616 }
 617
 618 static void be_pbqp_coloring(be_chordal_env_t *env)
 619 {
 620         ir_graph                    *irg            = env->irg;
 621         const arch_register_class_t *cls            = env->cls;
 622         be_lv_t                     *lv             = NULL;
 623         plist_element_t             *element        = NULL;
 624         unsigned                     colors_n       = arch_register_class_n_regs(cls);
 625         be_pbqp_alloc_env_t          pbqp_alloc_env;
 626         unsigned                     col;
 627         unsigned                     row;
 628         pbqp_matrix_t               *ife_matrix;
 629         num                          solution;
 630 #if KAPS_DUMP
 631         FILE                        *file_before;
 632 #endif
 633 #if TIMER
 634         ir_timer_t *t_ra_pbqp_alloc_create     = ir_timer_new();
 635         ir_timer_t *t_ra_pbqp_alloc_solve      = ir_timer_new();
 636         ir_timer_t *t_ra_pbqp_alloc_create_aff = ir_timer_new();
 637
 638         printf("#### ----- === Allocating registers of %s (%s) ===\n", cls->name, get_entity_name(get_irg_entity(irg)));
 639 #endif
 640         lv = be_assure_liveness(irg);
 641         be_liveness_assure_sets(lv);
 642         be_liveness_assure_chk(lv);
 643
 644         /* insert perms */
 645         assure_doms(irg);
 646         dom_tree_walk_irg(irg, insert_perms, NULL, env);
 647
 648         /* dump graph after inserting perms */
 649         if (env->opts->dump_flags & BE_CH_DUMP_CONSTR) {
 650                 char buf[256];
 651                 snprintf(buf, sizeof(buf), "-%s-constr", cls->name);
 652                 dump_ir_graph(irg, buf);
 653         }
 654
 655
 656         /* initialize pbqp allocation data structure */
 657         pbqp_alloc_env.pbqp_inst        = alloc_pbqp(get_irg_last_idx(irg));  /* initialize pbqp instance */
 658         pbqp_alloc_env.cls              = cls;
 659         pbqp_alloc_env.irg              = irg;
 660         pbqp_alloc_env.lv               = lv;
 661         pbqp_alloc_env.allocatable_regs = bitset_malloc(colors_n);
 662         pbqp_alloc_env.rpeo             = plist_new();
 663         pbqp_alloc_env.restr_nodes      = XMALLOCNZ(unsigned, get_irg_last_idx(irg));
 664         pbqp_alloc_env.ife_edge_num     = XMALLOCNZ(unsigned, get_irg_last_idx(irg));
 665         pbqp_alloc_env.env              = env;
 666         be_put_allocatable_regs(irg, cls, pbqp_alloc_env.allocatable_regs);
 667
 668
 669         /* create costs matrix template for interference edges */
 670         ife_matrix = pbqp_matrix_alloc(pbqp_alloc_env.pbqp_inst, colors_n, colors_n);
 671         /* set costs */
 672         for (row = 0, col = 0; row < colors_n; row++, col++)
 673                 pbqp_matrix_set(ife_matrix, row, col, INF_COSTS);
 674
 675         pbqp_alloc_env.ife_matrix_template = ife_matrix;
 676
 677
 678         if (!use_exec_freq) {
 679                 /* create costs matrix template for affinity edges */
 680                 pbqp_matrix_t *afe_matrix = pbqp_matrix_alloc(pbqp_alloc_env.pbqp_inst, colors_n, colors_n);
 681                 /* set costs */
 682                 for (row = 0; row < colors_n; row++) {
 683                         for (col = 0; col < colors_n; col++) {
 684                                 if (row != col)
 685                                         pbqp_matrix_set(afe_matrix, row, col, 2);
 686                         }
 687                 }
 688                 pbqp_alloc_env.aff_matrix_template = afe_matrix;
 689         }
 690
 691
 692         /* create pbqp instance */
 693 #if TIMER
 694         ir_timer_reset_and_start(t_ra_pbqp_alloc_create);
 695 #endif
 696         assure_doms(irg);
 697         dom_tree_walk_irg(irg, create_pbqp_coloring_instance , NULL, &pbqp_alloc_env);
 698 #if TIMER
 699         ir_timer_stop(t_ra_pbqp_alloc_create);
 700 #endif
 701
 702
 703         /* set up affinity edges */
 704 #if TIMER
 705         ir_timer_reset_and_start(t_ra_pbqp_alloc_create_aff);
 706 #endif
 707         foreach_plist(pbqp_alloc_env.rpeo, element) {
 708                 pbqp_node_t *node = (pbqp_node_t*)element->data;
 709                 ir_node     *irn  = get_idx_irn(irg, node->index);
 710
 711                 create_affinity_edges(irn, &pbqp_alloc_env);
 712         }
 713 #if TIMER
 714         ir_timer_stop(t_ra_pbqp_alloc_create_aff);
 715 #endif
 716
 717
 718 #if KAPS_DUMP
 719         // dump graph before solving pbqp
 720         file_before = my_open(env, "", "-pbqp_coloring.html");
 721         set_dumpfile(pbqp_alloc_env.pbqp_inst, file_before);
 722 #endif
 723
 724         /* print out reverse perfect elimination order */
 725 #if PRINT_RPEO
 726         {
 727                 plist_element_t *elements;
 728                 foreach_plist(pbqp_alloc_env.rpeo, elements) {
 729                         pbqp_node_t *node = elements->data;
 730                         printf(" %d(%ld);", node->index, get_idx_irn(irg, node->index)->node_nr);
 731                 }
 732                 printf("\n");
 733         }
 734 #endif
 735
 736         /* solve pbqp instance */
 737 #if TIMER
 738         ir_timer_reset_and_start(t_ra_pbqp_alloc_solve);
 739 #endif
 740         if(use_late_decision) {
 741                 solve_pbqp_heuristical_co_ld(pbqp_alloc_env.pbqp_inst,pbqp_alloc_env.rpeo);
 742         }
 743         else {
 744                 solve_pbqp_heuristical_co(pbqp_alloc_env.pbqp_inst,pbqp_alloc_env.rpeo);
 745         }
 746 #if TIMER
 747         ir_timer_stop(t_ra_pbqp_alloc_solve);
 748 #endif
 749
 750
 751         solution = get_solution(pbqp_alloc_env.pbqp_inst);
 752         if (solution == INF_COSTS)
 753                 panic("No PBQP solution found");
 754
 755
 756         /* assign colors */
 757         foreach_plist(pbqp_alloc_env.rpeo, element) {
 758                 pbqp_node_t           *node  = (pbqp_node_t*)element->data;
 759                 ir_node               *irn   = get_idx_irn(irg, node->index);
 760                 num                    color = get_node_solution(pbqp_alloc_env.pbqp_inst, node->index);
 761                 const arch_register_t *reg   = arch_register_for_index(cls, color);
 762
 763                 arch_set_irn_register(irn, reg);
 764         }
 765
 766
 767 #if TIMER
 768         printf("PBQP alloc create:     %10.3lf msec\n",
 769                (double)ir_timer_elapsed_usec(t_ra_pbqp_alloc_create) / 1000.0);
 770         printf("PBQP alloc solve:      %10.3lf msec\n",
 771                (double)ir_timer_elapsed_usec(t_ra_pbqp_alloc_solve) / 1000.0);
 772         printf("PBQP alloc create aff: %10.3lf msec\n",
 773                (double)ir_timer_elapsed_usec(t_ra_pbqp_alloc_create_aff) / 1000.0);
 774 #endif
 775
 776
 777         /* free reserved memory */
 778 #if KAPS_DUMP
 779         fclose(file_before);
 780 #endif
 781         bitset_free(pbqp_alloc_env.allocatable_regs);
 782         free_pbqp(pbqp_alloc_env.pbqp_inst);
 783         plist_free(pbqp_alloc_env.rpeo);
 784         xfree(pbqp_alloc_env.restr_nodes);
 785         xfree(pbqp_alloc_env.ife_edge_num);
 786 }
 787
 788
 789 /**
 790  * Initializes this module.
 791  */
 792 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_pbqp_coloring)
 793 void be_init_pbqp_coloring(void)
 794 {
 795         lc_opt_entry_t *be_grp       = lc_opt_get_grp(firm_opt_get_root(), "be");
 796         lc_opt_entry_t *ra_grp       = lc_opt_get_grp(be_grp, "ra");
 797         lc_opt_entry_t *chordal_grp  = lc_opt_get_grp(ra_grp, "chordal");
 798         lc_opt_entry_t *coloring_grp = lc_opt_get_grp(chordal_grp, "coloring");
 799         lc_opt_entry_t *pbqp_grp     = lc_opt_get_grp(coloring_grp, "pbqp");
 800
 801         static be_ra_chordal_coloring_t coloring = {
 802                 be_pbqp_coloring
 803         };
 804
 805         lc_opt_add_table(pbqp_grp, options);
 806         be_register_chordal_coloring("pbqp", &coloring);
 807 }