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