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 #ifdef FIRM_KAPS
  32
  33 #include "debug.h"
  34 #include "error.h"
  35
  36 #include "irdom.h"
  37 #include "irdump.h"
  38 #include "iredges_t.h"
  39 #include "irprintf.h"
  40 #include "irgwalk.h"
  41 #include "time.h"
  42
  43 /* libfirm/ir/adt includes */
  44 #include "bipartite.h"
  45
  46 /* libfirm/ir/be includes */
  47 #include "bearch.h"
  48 #include "beirg.h"
  49 #include "besched.h"
  50 #include "bemodule.h"
  51 #include "bechordal_common.h"
  52 #include "bechordal.h"
  53 #include "bechordal_t.h"
  54 #include "beinsn_t.h"
  55 #include "benode.h"
  56 #include "belive.h"
  57 #include "belive_t.h"
  58 #include "beutil.h"
  59 #include "plist.h"
  60 #include "pqueue.h"
  61 #include "becopyopt.h"
  62
  63 /* pbqp includes */
  64 #include "kaps.h"
  65 #include "matrix.h"
  66 #include "vector.h"
  67 #include "vector_t.h"
  68 #include "heuristical_co.h"
  69 #include "heuristical_co_ld.h"
  70 #include "pbqp_t.h"
  71 #include "html_dumper.h"
  72 #include "pbqp_node_t.h"
  73 #include "pbqp_node.h"
  74 #include "pbqp_edge_t.h"
  75
  76 #define TIMER                 0
  77 #define PRINT_RPEO            0
  78 #define USE_BIPARTIT_MATCHING 0
  79 #define DO_USEFUL_OPT         1
  80
  81
  82 static int use_exec_freq     = true;
  83 static int use_late_decision = false;
  84
  85 typedef struct _be_pbqp_alloc_env_t {
  86         pbqp                        *pbqp_inst;                 /**< PBQP instance for register allocation */
  87         ir_graph                    *irg;                       /**< The graph under examination. */
  88         const arch_register_class_t *cls;                               /**< Current processed register class */
  89         be_lv_t                     *lv;
  90         bitset_t                    *ignored_regs;
  91         pbqp_matrix                 *ife_matrix_template;
  92         pbqp_matrix                 *aff_matrix_template;
  93         plist_t                     *rpeo;
  94         unsigned                    *restr_nodes;
  95         unsigned                    *ife_edge_num;
  96         be_chordal_env_t            *env;
  97 } be_pbqp_alloc_env_t;
  98
  99
 100 #define is_Reg_Phi(irn)                                        (is_Phi(irn) && mode_is_data(get_irn_mode(irn)))
 101 #define get_Perm_src(irn)                                      (get_irn_n(get_Proj_pred(irn), get_Proj_proj(irn)))
 102 #define is_Perm_Proj(irn)                                      (is_Proj(irn) && be_is_Perm(get_Proj_pred(irn)))
 103 #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)))
 104 #define get_free_regs(restr_nodes, cls, irn)                   (arch_register_class_n_regs(cls) - restr_nodes[get_irn_idx(irn)])
 105
 106 static inline int is_2addr_code(const arch_register_req_t *req)
 107 {
 108         return (req->type & arch_register_req_type_should_be_same) != 0;
 109 }
 110
 111 static const lc_opt_table_entry_t options[] = {
 112         LC_OPT_ENT_BOOL("exec_freq", "use exec_freq",  &use_exec_freq),
 113         LC_OPT_ENT_BOOL("late_decision", "use late decision for register allocation",  &use_late_decision),
 114         LC_OPT_LAST
 115 };
 116
 117 #if KAPS_DUMP
 118 static FILE *my_open(const be_chordal_env_t *env, const char *prefix, const char *suffix)
 119 {
 120         FILE       *result;
 121         char        buf[1024];
 122         size_t      i;
 123         size_t      n;
 124         char       *tu_name;
 125         const char *cup_name = be_get_irg_main_env(env->irg)->cup_name;
 126
 127         n = strlen(cup_name);
 128         tu_name = XMALLOCN(char, n + 1);
 129         strcpy(tu_name, cup_name);
 130         for (i = 0; i < n; ++i)
 131                 if (tu_name[i] == '.')
 132                         tu_name[i] = '_';
 133
 134         ir_snprintf(buf, sizeof(buf), "%s%s_%F_%s%s", prefix, tu_name, env->irg, env->cls->name, suffix);
 135         xfree(tu_name);
 136         result = fopen(buf, "wt");
 137         if (result == NULL) {
 138                 panic("Couldn't open '%s' for writing.", buf);
 139         }
 140
 141         return result;
 142 }
 143 #endif
 144
 145
 146 static void create_pbqp_node(be_pbqp_alloc_env_t *pbqp_alloc_env, ir_node *irn)
 147 {
 148         const arch_register_class_t *cls = pbqp_alloc_env->cls;
 149         pbqp     *pbqp_inst              = pbqp_alloc_env->pbqp_inst;
 150         bitset_t *ignored_regs           = pbqp_alloc_env->ignored_regs;
 151         unsigned  colors_n               = arch_register_class_n_regs(cls);
 152         unsigned  cntConstrains          = 0;
 153
 154         /* create costs vector depending on register constrains */
 155         struct vector *costs_vector = vector_alloc(pbqp_inst, colors_n);
 156
 157         /* set costs depending on register constrains */
 158         unsigned idx;
 159         for (idx = 0; idx < colors_n; idx++) {
 160                 if (bitset_is_set(ignored_regs, idx) || !arch_reg_out_is_allocatable(irn, arch_register_for_index(cls, idx))) {
 161                         /* constrained */
 162                         vector_set(costs_vector, idx, INF_COSTS);
 163                         cntConstrains++;
 164                 }
 165         }
 166
 167         /* add vector to pbqp node */
 168         add_node_costs(pbqp_inst, get_irn_idx(irn), costs_vector);
 169         pbqp_alloc_env->restr_nodes[get_irn_idx(irn)] = cntConstrains;
 170 }
 171
 172 static void insert_ife_edge(be_pbqp_alloc_env_t *pbqp_alloc_env, ir_node *src_node, ir_node *trg_node)
 173 {
 174         pbqp                        *pbqp                = pbqp_alloc_env->pbqp_inst;
 175         const arch_register_class_t *cls                 = pbqp_alloc_env->cls;
 176         pbqp_matrix                 *ife_matrix_template = pbqp_alloc_env->ife_matrix_template;
 177         unsigned                    *restr_nodes         = pbqp_alloc_env->restr_nodes;
 178
 179         if (get_edge(pbqp, get_irn_idx(src_node), get_irn_idx(trg_node)) == NULL) {
 180
 181                 /* increase ife edge counter */
 182                 pbqp_alloc_env->ife_edge_num[get_irn_idx(src_node)]++;
 183                 pbqp_alloc_env->ife_edge_num[get_irn_idx(trg_node)]++;
 184
 185 #if DO_USEFUL_OPT || USE_BIPARTIT_MATCHING
 186                 /* do useful optimization to speed up pbqp solving (we can do this because we know our matrix) */
 187                 if (get_free_regs(restr_nodes, cls, src_node) == 1 && get_free_regs(restr_nodes, cls, trg_node) == 1) {
 188                         assert(vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs) !=
 189                                vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs) &&
 190                                "Interfering nodes must not have the same register!");
 191                         return;
 192                 }
 193                 if (get_free_regs(restr_nodes, cls, src_node) == 1 || get_free_regs(restr_nodes, cls, trg_node) == 1) {
 194                         if (get_free_regs(restr_nodes, cls, src_node) == 1) {
 195                                 unsigned idx = vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs);
 196                                 vector_set(get_node(pbqp, get_irn_idx(trg_node))->costs, idx, INF_COSTS);
 197                         }
 198                         else {
 199                                 unsigned idx = vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs);
 200                                 vector_set(get_node(pbqp, get_irn_idx(src_node))->costs, idx, INF_COSTS);
 201                         }
 202                         return;
 203                 }
 204 #endif
 205                 /* insert interference edge */
 206                 insert_edge(pbqp, src_node, trg_node, ife_matrix_template);
 207         }
 208 }
 209
 210 static void insert_afe_edge(be_pbqp_alloc_env_t *pbqp_alloc_env, ir_node *src_node, ir_node *trg_node, int pos)
 211 {
 212         pbqp                        *pbqp        = pbqp_alloc_env->pbqp_inst;
 213         const arch_register_class_t *cls         = pbqp_alloc_env->cls;
 214         unsigned                    *restr_nodes = pbqp_alloc_env->restr_nodes;
 215         pbqp_matrix                 *afe_matrix  = pbqp_matrix_alloc(pbqp, arch_register_class_n_regs(cls), arch_register_class_n_regs(cls));
 216         unsigned                     colors_n    = arch_register_class_n_regs(cls);
 217
 218         if (get_edge(pbqp, get_irn_idx(src_node), get_irn_idx(trg_node)) == NULL) {
 219                 if (use_exec_freq) {
 220                         /* get exec_freq for copy_block */
 221                         ir_node       *root_bl   = get_nodes_block(src_node);
 222                         ir_node       *copy_bl   = is_Phi(src_node) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
 223                         ir_exec_freq  *exec_freq = be_get_irg_exec_freq(pbqp_alloc_env->irg);
 224                         unsigned long  res       = get_block_execfreq_ulong(exec_freq, copy_bl);
 225
 226                         /* create afe-matrix */
 227                         unsigned row, col;
 228                         for (row = 0; row < colors_n; row++) {
 229                                 for (col = 0; col < colors_n; col++) {
 230                                         if (row != col)
 231                                                 pbqp_matrix_set(afe_matrix, row, col, (num)res);
 232                                 }
 233                         }
 234                 }
 235                 else {
 236                         afe_matrix = pbqp_alloc_env->aff_matrix_template;
 237                 }
 238 #if DO_USEFUL_OPT || USE_BIPARTIT_MATCHING
 239                 /* do useful optimization to speed up pbqp solving */
 240                 if (get_free_regs(restr_nodes, cls, src_node) == 1 && get_free_regs(restr_nodes, cls, trg_node) == 1) {
 241                         return;
 242                 }
 243                 if (get_free_regs(restr_nodes, cls, src_node) == 1 || get_free_regs(restr_nodes, cls, trg_node) == 1) {
 244                         if (get_free_regs(restr_nodes, cls, src_node) == 1) {
 245                                 unsigned regIdx = vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs);
 246                                 vector_add_matrix_col(get_node(pbqp, get_irn_idx(trg_node))->costs, afe_matrix, regIdx);
 247                         }
 248                         else {
 249                                 unsigned regIdx = vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs);
 250                                 vector_add_matrix_col(get_node(pbqp, get_irn_idx(src_node))->costs, afe_matrix, regIdx);
 251                         }
 252                         return;
 253                 }
 254 #endif
 255                 /* insert interference edge */
 256                 insert_edge(pbqp, src_node, trg_node, afe_matrix);
 257         }
 258 }
 259
 260 static void create_affinity_edges(ir_node *irn, void *env)
 261 {
 262         be_pbqp_alloc_env_t         *pbqp_alloc_env = env;
 263         const arch_register_class_t *cls            = pbqp_alloc_env->cls;
 264         const arch_register_req_t   *req            = arch_get_register_req_out(irn);
 265         unsigned                     pos;
 266         unsigned                     max;
 267
 268         if (is_Reg_Phi(irn)) { /* Phis */
 269                 for (pos = 0, max = get_irn_arity(irn); pos < max; ++pos) {
 270                         ir_node *arg = get_irn_n(irn, pos);
 271
 272                         if (!arch_irn_consider_in_reg_alloc(cls, arg))
 273                                 continue;
 274
 275                         /* no edges to itself */
 276                         if (irn == arg) {
 277                                 continue;
 278                         }
 279
 280                         insert_afe_edge(pbqp_alloc_env, irn, arg, pos);
 281                 }
 282         }
 283         else if (is_Perm_Proj(irn)) { /* Perms */
 284                 ir_node *arg = get_Perm_src(irn);
 285                 if (!arch_irn_consider_in_reg_alloc(cls, arg))
 286                         return;
 287
 288                 insert_afe_edge(pbqp_alloc_env, irn, arg, -1);
 289         }
 290         else { /* 2-address code */
 291                 if (is_2addr_code(req)) {
 292                         const unsigned other = req->other_same;
 293                         int            i;
 294
 295                         for (i = 0; 1U << i <= other; ++i) {
 296                                 if (other & (1U << i)) {
 297                                         ir_node *other = get_irn_n(skip_Proj(irn), i);
 298                                         if (!arch_irn_consider_in_reg_alloc(cls, other))
 299                                                 continue;
 300
 301                                         /* no edges to itself */
 302                                         if (irn == other) {
 303                                                 continue;
 304                                         }
 305
 306                                         insert_afe_edge(pbqp_alloc_env, irn, other, i);
 307                                 }
 308                         }
 309                 }
 310         }
 311 }
 312
 313 static void create_pbqp_coloring_instance(ir_node *block, void *data)
 314 {
 315         be_pbqp_alloc_env_t         *pbqp_alloc_env     = data;
 316         be_lv_t                     *lv                 = pbqp_alloc_env->lv;
 317         const arch_register_class_t *cls                = pbqp_alloc_env->cls;
 318         plist_t                     *rpeo               = pbqp_alloc_env->rpeo;
 319         pbqp                        *pbqp_inst          = pbqp_alloc_env->pbqp_inst;
 320         plist_t                     *temp_list          = plist_new();
 321         plist_element_t             *el;
 322         ir_node                     *irn;
 323         ir_nodeset_t                 live_nodes;
 324 #if USE_BIPARTIT_MATCHING
 325         int                         *assignment         = ALLOCAN(int, cls->n_regs);
 326 #else
 327         unsigned                    *restr_nodes        = pbqp_alloc_env->restr_nodes;
 328         pqueue_t                    *restr_nodes_queue  = new_pqueue();
 329         pqueue_t                    *queue              = new_pqueue();
 330         plist_t                     *sorted_list        = plist_new();
 331         ir_node                     *last_element       = NULL;
 332 #endif
 333
 334         /* first, determine the pressure */
 335         /* (this is only for compatibility with copymin optimization, it's not needed for pbqp coloring) */
 336         create_borders(block, pbqp_alloc_env->env);
 337
 338         /* calculate living nodes for the first step */
 339         ir_nodeset_init(&live_nodes);
 340         be_liveness_end_of_block(lv, cls, block, &live_nodes);
 341
 342         /* create pbqp nodes, interference edges and reverse perfect elimination order */
 343         sched_foreach_reverse(block, irn) {
 344                 ir_node               *live;
 345                 ir_nodeset_iterator_t  iter;
 346
 347                 if (get_irn_mode(irn) == mode_T) {
 348                         const ir_edge_t *edge;
 349                         foreach_out_edge(irn, edge) {
 350                                 ir_node *proj = get_edge_src_irn(edge);
 351                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 352                                         continue;
 353
 354                                 /* create pbqp source node if it dosn't exist */
 355                                 if (get_node(pbqp_inst, get_irn_idx(proj)) == NULL) {
 356                                         create_pbqp_node(pbqp_alloc_env, proj);
 357                                 }
 358
 359                                 /* create nodes and interference edges */
 360                                 foreach_ir_nodeset(&live_nodes, live, iter) {
 361                                         /* create pbqp source node if it dosn't exist */
 362                                         if (get_node(pbqp_inst, get_irn_idx(live)) == NULL) {
 363                                                 create_pbqp_node(pbqp_alloc_env, live);
 364                                         }
 365
 366                                         /* no edges to itself */
 367                                         if (proj == live) {
 368                                                 continue;
 369                                         }
 370
 371                                         insert_ife_edge(pbqp_alloc_env, proj, live);
 372                                 }
 373                         }
 374                 }
 375                 else {
 376                         if (arch_irn_consider_in_reg_alloc(cls, irn)) {
 377                                 /* create pbqp source node if it dosn't exist */
 378                                 if (get_node(pbqp_inst, get_irn_idx(irn)) == NULL) {
 379                                         create_pbqp_node(pbqp_alloc_env, irn);
 380                                 }
 381
 382                                 /* create nodes and interference edges */
 383                                 foreach_ir_nodeset(&live_nodes, live, iter) {
 384                                         /* create pbqp source node if it dosn't exist */
 385                                         if (get_node(pbqp_inst, get_irn_idx(live)) == NULL) {
 386                                                 create_pbqp_node(pbqp_alloc_env, live);
 387                                         }
 388
 389                                         /* no edges to itself */
 390                                         if (irn == live) {
 391                                                 continue;
 392                                         }
 393
 394                                         /* insert interference edge */
 395                                         insert_ife_edge(pbqp_alloc_env, irn, live);
 396                                 }
 397                         }
 398                 }
 399
 400                 /* get living nodes for next step */
 401                 if (!is_Phi(irn)) {
 402                         be_liveness_transfer(cls, irn, &live_nodes);
 403                 }
 404
 405 #if USE_BIPARTIT_MATCHING
 406                 if (get_irn_mode(irn) == mode_T) {
 407                         unsigned     clique_size         = 0;
 408                         unsigned     n_alloc             = 0;
 409                         pbqp_node   *clique[cls->n_regs];
 410                         bipartite_t *bp                  = bipartite_new(cls->n_regs, cls->n_regs);
 411
 412                         /* add all proj after a perm to clique */
 413                         const ir_edge_t *edge;
 414                         foreach_out_edge(irn, edge) {
 415                                 ir_node *proj = get_edge_src_irn(edge);
 416
 417                                 /* ignore node if it is not necessary for register allocation */
 418                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 419                                         continue;
 420
 421                                 /* insert pbqp node into temp rpeo list of this block */
 422                                 plist_insert_front(temp_list, get_node(pbqp_inst, get_irn_idx(proj)));
 423
 424                                 if(is_Perm_Proj(proj)) {
 425                                         /* add proj to clique */
 426                                         pbqp_node *clique_member = get_node(pbqp_inst,proj->node_idx);
 427                                         vector    *costs         = clique_member->costs;
 428                                         unsigned   idx           = 0;
 429
 430                                         clique[clique_size] = clique_member;
 431
 432                                         for(idx = 0; idx < costs->len; idx++) {
 433                                                 if(costs->entries[idx].data != INF_COSTS) {
 434                                                         bipartite_add(bp, clique_size, idx);
 435                                                 }
 436                                         }
 437
 438                                         /* increase node counter */
 439                                         clique_size++;
 440                                         n_alloc++;
 441                                 }
 442                         }
 443
 444                         if(clique_size > 0) {
 445                                 plist_element_t *listElement;
 446                                 foreach_plist(temp_list, listElement) {
 447                                         pbqp_node *clique_candidate  = listElement->data;
 448                                         unsigned   idx               = 0;
 449                                         bool       isMember          = true;
 450
 451                                         /* clique size not bigger then register class size */
 452                                         if(clique_size >= cls->n_regs) break;
 453
 454                                         for(idx = 0; idx < clique_size; idx++) {
 455                                                 pbqp_node *member = clique[idx];
 456
 457                                                 if(member == clique_candidate) {
 458                                                         isMember = false;
 459                                                         break;
 460                                                 }
 461
 462                                                 if(get_edge(pbqp_inst, member->index, clique_candidate->index) == NULL && get_edge(pbqp_inst, clique_candidate->index, member->index) == NULL) {
 463                                                         isMember = false;
 464                                                         break;
 465                                                 }
 466                                         }
 467
 468                                         /* goto next list element if current node is not a member of the clique */
 469                                         if(!isMember) { continue; }
 470
 471                                         /* add candidate to clique */
 472                                         clique[clique_size] = clique_candidate;
 473
 474                                         vector *costs = clique_candidate->costs;
 475                                         for(idx = 0; idx < costs->len; idx++) {
 476                                                 if(costs->entries[idx].data != INF_COSTS) {
 477                                                         bipartite_add(bp, clique_size, idx);
 478                                                 }
 479                                         }
 480
 481                                         /* increase node counter */
 482                                         clique_size++;
 483                                 }
 484                         }
 485
 486                         /* solve bipartite matching */
 487                         bipartite_matching(bp, assignment);
 488
 489                         /* assign colors */
 490                         unsigned nodeIdx = 0;
 491                         for(nodeIdx = 0; nodeIdx < clique_size; nodeIdx++) {
 492                                 vector *costs = clique[nodeIdx]->costs;
 493                                 int     idx;
 494                                 for(idx = 0; idx < (int)costs->len; idx++) {
 495                                         if(assignment[nodeIdx] != idx) {
 496                                                 costs->entries[idx].data = INF_COSTS;
 497                                         }
 498                                 }
 499                                 assert(assignment[nodeIdx] >= 0 && "there must have been a register assigned (node not register pressure faithful?)");
 500                         }
 501
 502                         /* free memory */
 503                         bipartite_free(bp);
 504                 }
 505                 else {
 506                         if (arch_irn_consider_in_reg_alloc(cls, irn)) {
 507                                 plist_insert_front(temp_list, get_node(pbqp_inst, get_irn_idx(irn)));
 508                         }
 509                 }
 510 #else
 511                 /* order nodes for perfect elimination order */
 512                 if (get_irn_mode(irn) == mode_T) {
 513                         bool             allHaveIFEdges = true;
 514                         const ir_edge_t *edge;
 515
 516                         foreach_out_edge(irn, edge) {
 517                                 ir_node *proj = get_edge_src_irn(edge);
 518                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 519                                         continue;
 520
 521                                 /* insert proj node into priority queue (descending by the number of interference edges) */
 522                                 if (get_free_regs(restr_nodes, cls, proj) <= 4) {
 523                                         pqueue_put(restr_nodes_queue, proj, pbqp_alloc_env->ife_edge_num[get_irn_idx(proj)]);
 524                                 }
 525                                 else {
 526                                         pqueue_put(queue, proj, pbqp_alloc_env->ife_edge_num[get_irn_idx(proj)]);
 527                                 }
 528
 529                                 /* skip last step if there is no last_element */
 530                                 if(last_element == NULL)
 531                                         continue;
 532
 533                                 /* check if proj has an if edge to last_element (at this time pbqp contains only if edges) */
 534                                 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) {
 535                                         allHaveIFEdges = false; /* there is no if edge between proj and last_element */
 536                                 }
 537                         }
 538
 539                         if(last_element != NULL && allHaveIFEdges) {
 540                                 if (get_free_regs(restr_nodes, cls, last_element) <= 4) {
 541                                         pqueue_put(restr_nodes_queue, last_element, pbqp_alloc_env->ife_edge_num[get_irn_idx(last_element)]);
 542                                 }
 543                                 else {
 544                                         pqueue_put(queue, last_element, pbqp_alloc_env->ife_edge_num[get_irn_idx(last_element)]);
 545                                 }
 546                                 plist_erase(temp_list, plist_find_value(temp_list, get_node(pbqp_inst, last_element->node_idx)));
 547                                 last_element = NULL;
 548                         }
 549
 550                         /* first insert all restricted proj nodes */
 551                         while (!pqueue_empty(restr_nodes_queue)) {
 552                                 plist_insert_front(sorted_list, get_node(pbqp_inst, get_irn_idx(pqueue_pop_front(restr_nodes_queue))));
 553                         }
 554
 555                         /* insert proj nodes descending by their number of interference edges */
 556                         while (!pqueue_empty(queue)) {
 557                                 plist_insert_front(sorted_list, get_node(pbqp_inst, get_irn_idx(pqueue_pop_front(queue))));
 558                         }
 559
 560                         /* invert sorted list */
 561                         foreach_plist(sorted_list, el) {
 562                                 plist_insert_front(temp_list, el->data);
 563                         }
 564
 565                         plist_clear(sorted_list);
 566
 567                 }
 568                 else {
 569                         if (arch_irn_consider_in_reg_alloc(cls, irn)) {
 570                                 // remember last colorable node
 571                                 last_element = irn;
 572                                 plist_insert_front(temp_list, get_node(pbqp_inst, get_irn_idx(irn)));
 573                         }
 574                         else {
 575                                 // node not colorable, so ignore it
 576                                 last_element = NULL;
 577                         }
 578                 }
 579 #endif
 580         }
 581
 582         /* add the temp rpeo list of this block to the global reverse perfect elimination order list*/
 583         foreach_plist(temp_list, el) {
 584                 plist_insert_back(rpeo, el->data);
 585         }
 586
 587         /* free reserved memory */
 588         ir_nodeset_destroy(&live_nodes);
 589         plist_free(temp_list);
 590 #if USE_BIPARTIT_MATCHING
 591 #else
 592         plist_free(sorted_list);
 593         del_pqueue(queue);
 594         del_pqueue(restr_nodes_queue);
 595 #endif
 596 }
 597
 598 static void insert_perms(ir_node *block, void *data)
 599 {
 600         /*
 601          * Start silent in the start block.
 602          * The silence remains until the first barrier is seen.
 603          * Each other block is begun loud.
 604          */
 605         be_chordal_env_t *env    = data;
 606         ir_node          *irn;
 607         int               silent = block == get_irg_start_block(get_irn_irg(block));
 608
 609         /*
 610          * If the block is the start block search the barrier and
 611          * start handling constraints from there.
 612          */
 613         for (irn = sched_first(block); !sched_is_end(irn);) {
 614                 int silent_old = silent;        /* store old silent value */
 615                 if (be_is_Barrier(irn))
 616                         silent = !silent;               /* toggle silent flag */
 617
 618                 be_insn_t *insn = chordal_scan_insn(env, irn);
 619                 irn             = insn->next_insn;
 620
 621                 if (silent_old)
 622                         continue;
 623
 624                 if (!insn->has_constraints)
 625                         continue;
 626
 627                 pre_process_constraints(env, &insn);
 628         }
 629 }
 630
 631 static void be_pbqp_coloring(be_chordal_env_t *env)
 632 {
 633         ir_graph                    *irg            = env->irg;
 634         const arch_register_class_t *cls            = env->cls;
 635         be_lv_t                     *lv             = NULL;
 636         plist_element_t             *element        = NULL;
 637         unsigned                     colors_n       = arch_register_class_n_regs(cls);
 638         be_pbqp_alloc_env_t          pbqp_alloc_env;
 639         unsigned                     col;
 640         unsigned                     row;
 641
 642 #if TIMER
 643         ir_timer_t *t_ra_pbqp_alloc_create     = ir_timer_new();
 644         ir_timer_t *t_ra_pbqp_alloc_solve      = ir_timer_new();
 645         ir_timer_t *t_ra_pbqp_alloc_create_aff = ir_timer_new();
 646
 647         printf("#### ----- === Allocating registers of %s (%s) ===\n", cls->name, get_entity_name(get_irg_entity(irg)));
 648 #endif
 649         lv = be_assure_liveness(irg);
 650         be_liveness_assure_sets(lv);
 651         be_liveness_assure_chk(lv);
 652
 653         /* insert perms */
 654         assure_doms(irg);
 655         dom_tree_walk_irg(irg, insert_perms, NULL, env);
 656
 657         /* dump graph after inserting perms */
 658         if (env->opts->dump_flags & BE_CH_DUMP_CONSTR) {
 659                 char buf[256];
 660                 snprintf(buf, sizeof(buf), "-%s-constr", cls->name);
 661                 dump_ir_graph(irg, buf);
 662         }
 663
 664
 665         /* initialize pbqp allocation data structure */
 666         pbqp_alloc_env.pbqp_inst    = alloc_pbqp(get_irg_last_idx(irg));                /* initialize pbqp instance */
 667         pbqp_alloc_env.cls          = cls;
 668         pbqp_alloc_env.irg          = irg;
 669         pbqp_alloc_env.lv           = lv;
 670         pbqp_alloc_env.ignored_regs = bitset_malloc(colors_n);
 671         pbqp_alloc_env.rpeo         = plist_new();
 672         pbqp_alloc_env.restr_nodes  = XMALLOCNZ(unsigned, get_irg_last_idx(irg));
 673         pbqp_alloc_env.ife_edge_num = XMALLOCNZ(unsigned, get_irg_last_idx(irg));
 674         pbqp_alloc_env.env          = env;
 675         be_put_ignore_regs(irg, cls, pbqp_alloc_env.ignored_regs);                              /* get ignored registers */
 676
 677
 678         /* create costs matrix template for interference edges */
 679         struct pbqp_matrix *ife_matrix = pbqp_matrix_alloc(pbqp_alloc_env.pbqp_inst, colors_n, colors_n);
 680         /* set costs */
 681         for (row = 0, col = 0; row < colors_n; row++, col++)
 682                 pbqp_matrix_set(ife_matrix, row, col, INF_COSTS);
 683
 684         pbqp_alloc_env.ife_matrix_template = ife_matrix;
 685
 686
 687         if (!use_exec_freq) {
 688                 /* create costs matrix template for affinity edges */
 689                 struct pbqp_matrix *afe_matrix = pbqp_matrix_alloc(pbqp_alloc_env.pbqp_inst, colors_n, colors_n);
 690                 /* set costs */
 691                 for (row = 0; row < colors_n; row++) {
 692                         for (col = 0; col < colors_n; col++) {
 693                                 if (row != col)
 694                                         pbqp_matrix_set(afe_matrix, row, col, 2);
 695                         }
 696                 }
 697                 pbqp_alloc_env.aff_matrix_template = afe_matrix;
 698         }
 699
 700
 701         /* create pbqp instance */
 702 #if TIMER
 703         ir_timer_reset_and_start(t_ra_pbqp_alloc_create);
 704 #endif
 705         assure_doms(irg);
 706         dom_tree_walk_irg(irg, create_pbqp_coloring_instance , NULL, &pbqp_alloc_env);
 707 #if TIMER
 708         ir_timer_stop(t_ra_pbqp_alloc_create);
 709 #endif
 710
 711
 712         /* set up affinity edges */
 713 #if TIMER
 714         ir_timer_reset_and_start(t_ra_pbqp_alloc_create_aff);
 715 #endif
 716         foreach_plist(pbqp_alloc_env.rpeo, element) {
 717                 pbqp_node *node = element->data;
 718                 ir_node   *irn  = get_idx_irn(irg, node->index);
 719
 720                 create_affinity_edges(irn, &pbqp_alloc_env);
 721         }
 722 #if TIMER
 723         ir_timer_stop(t_ra_pbqp_alloc_create_aff);
 724 #endif
 725
 726
 727 #if KAPS_DUMP
 728         // dump graph before solving pbqp
 729         FILE *file_before = my_open(env, "", "-pbqp_coloring.html");
 730         set_dumpfile(pbqp_alloc_env.pbqp_inst, file_before);
 731 #endif
 732
 733         /* print out reverse perfect eleminiation order */
 734 #if PRINT_RPEO
 735         plist_element_t *elements;
 736         foreach_plist(pbqp_alloc_env.rpeo, elements) {
 737                 pbqp_node *node                    = elements->data;
 738                 printf(" %d(%lu);", node->index, get_idx_irn(irg, node->index)->node_nr);
 739         }
 740         printf("\n");
 741 #endif
 742
 743         /* solve pbqp instance */
 744 #if TIMER
 745         ir_timer_reset_and_start(t_ra_pbqp_alloc_solve);
 746 #endif
 747         if(use_late_decision) {
 748                 solve_pbqp_heuristical_co_ld(pbqp_alloc_env.pbqp_inst,pbqp_alloc_env.rpeo);
 749         }
 750         else {
 751                 solve_pbqp_heuristical_co(pbqp_alloc_env.pbqp_inst,pbqp_alloc_env.rpeo);
 752         }
 753 #if TIMER
 754         ir_timer_stop(t_ra_pbqp_alloc_solve);
 755 #endif
 756
 757
 758         num solution = get_solution(pbqp_alloc_env.pbqp_inst);
 759         if (solution == INF_COSTS)
 760                 panic("No PBQP solution found");
 761
 762
 763         /* assign colors */
 764         foreach_plist(pbqp_alloc_env.rpeo, element) {
 765                 pbqp_node             *node  = element->data;
 766                 ir_node               *irn   = get_idx_irn(irg, node->index);
 767                 num                    color = get_node_solution(pbqp_alloc_env.pbqp_inst, node->index);
 768                 const arch_register_t *reg   = arch_register_for_index(cls, color);
 769
 770                 arch_set_irn_register(irn, reg);
 771         }
 772
 773
 774 #if TIMER
 775         printf("PBQP alloc create:     %10.3lf msec\n",
 776                (double)ir_timer_elapsed_usec(t_ra_pbqp_alloc_create) / 1000.0);
 777         printf("PBQP alloc solve:      %10.3lf msec\n",
 778                (double)ir_timer_elapsed_usec(t_ra_pbqp_alloc_solve) / 1000.0);
 779         printf("PBQP alloc create aff: %10.3lf msec\n",
 780                (double)ir_timer_elapsed_usec(t_ra_pbqp_alloc_create_aff) / 1000.0);
 781 #endif
 782
 783
 784         /* free reserved memory */
 785 #if KAPS_DUMP
 786         fclose(file_before);
 787 #endif
 788         bitset_free(pbqp_alloc_env.ignored_regs);
 789         free_pbqp(pbqp_alloc_env.pbqp_inst);
 790         plist_free(pbqp_alloc_env.rpeo);
 791         xfree(pbqp_alloc_env.restr_nodes);
 792         xfree(pbqp_alloc_env.ife_edge_num);
 793 }
 794
 795
 796 /**
 797  * Initializes this module.
 798  */
 799 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_pbqp_coloring);
 800 void be_init_pbqp_coloring(void)
 801 {
 802         lc_opt_entry_t *be_grp       = lc_opt_get_grp(firm_opt_get_root(), "be");
 803         lc_opt_entry_t *ra_grp       = lc_opt_get_grp(be_grp, "ra");
 804         lc_opt_entry_t *chordal_grp  = lc_opt_get_grp(ra_grp, "chordal");
 805         lc_opt_entry_t *coloring_grp = lc_opt_get_grp(chordal_grp, "coloring");
 806         lc_opt_entry_t *pbqp_grp     = lc_opt_get_grp(coloring_grp, "pbqp");
 807
 808         static be_ra_chordal_coloring_t coloring = {
 809                 be_pbqp_coloring
 810         };
 811
 812         lc_opt_add_table(pbqp_grp, options);
 813         be_register_chordal_coloring("pbqp", &coloring);
 814 }
 815
 816 #endif