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 "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                 if (!bitset_is_set(allocatable_regs, idx) || !arch_reg_out_is_allocatable(irn, arch_register_for_index(cls, idx))) {
 159                         /* constrained */
 160                         vector_set(costs_vector, idx, INF_COSTS);
 161                         cntConstrains++;
 162                 }
 163         }
 164
 165         /* add vector to pbqp node */
 166         add_node_costs(pbqp_inst, get_irn_idx(irn), costs_vector);
 167         pbqp_alloc_env->restr_nodes[get_irn_idx(irn)] = cntConstrains;
 168 }
 169
 170 static void insert_ife_edge(be_pbqp_alloc_env_t *pbqp_alloc_env, ir_node *src_node, ir_node *trg_node)
 171 {
 172         pbqp_t                      *pbqp                = pbqp_alloc_env->pbqp_inst;
 173         const arch_register_class_t *cls                 = pbqp_alloc_env->cls;
 174         pbqp_matrix_t               *ife_matrix_template = pbqp_alloc_env->ife_matrix_template;
 175         unsigned                    *restr_nodes         = pbqp_alloc_env->restr_nodes;
 176
 177         if (get_edge(pbqp, get_irn_idx(src_node), get_irn_idx(trg_node)) == NULL) {
 178
 179                 /* increase ife edge counter */
 180                 pbqp_alloc_env->ife_edge_num[get_irn_idx(src_node)]++;
 181                 pbqp_alloc_env->ife_edge_num[get_irn_idx(trg_node)]++;
 182
 183 #if DO_USEFUL_OPT || USE_BIPARTIT_MATCHING
 184                 /* do useful optimization to speed up pbqp solving (we can do this because we know our matrix) */
 185                 if (get_free_regs(restr_nodes, cls, src_node) == 1 && get_free_regs(restr_nodes, cls, trg_node) == 1) {
 186                         assert(vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs) !=
 187                                vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs) &&
 188                                "Interfering nodes must not have the same register!");
 189                         return;
 190                 }
 191                 if (get_free_regs(restr_nodes, cls, src_node) == 1 || get_free_regs(restr_nodes, cls, trg_node) == 1) {
 192                         if (get_free_regs(restr_nodes, cls, src_node) == 1) {
 193                                 unsigned idx = vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs);
 194                                 vector_set(get_node(pbqp, get_irn_idx(trg_node))->costs, idx, INF_COSTS);
 195                         }
 196                         else {
 197                                 unsigned idx = vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs);
 198                                 vector_set(get_node(pbqp, get_irn_idx(src_node))->costs, idx, INF_COSTS);
 199                         }
 200                         return;
 201                 }
 202 #endif
 203                 /* insert interference edge */
 204                 insert_edge(pbqp, src_node, trg_node, ife_matrix_template);
 205         }
 206 }
 207
 208 static void insert_afe_edge(be_pbqp_alloc_env_t *pbqp_alloc_env, ir_node *src_node, ir_node *trg_node, int pos)
 209 {
 210         pbqp_t                      *pbqp        = pbqp_alloc_env->pbqp_inst;
 211         const arch_register_class_t *cls         = pbqp_alloc_env->cls;
 212         unsigned                    *restr_nodes = pbqp_alloc_env->restr_nodes;
 213         pbqp_matrix_t               *afe_matrix  = pbqp_matrix_alloc(pbqp, arch_register_class_n_regs(cls), arch_register_class_n_regs(cls));
 214         unsigned                     colors_n    = arch_register_class_n_regs(cls);
 215
 216         if (get_edge(pbqp, get_irn_idx(src_node), get_irn_idx(trg_node)) == NULL) {
 217                 if (use_exec_freq) {
 218                         /* get exec_freq for copy_block */
 219                         ir_node       *root_bl   = get_nodes_block(src_node);
 220                         ir_node       *copy_bl   = is_Phi(src_node) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
 221                         ir_exec_freq  *exec_freq = be_get_irg_exec_freq(pbqp_alloc_env->irg);
 222                         unsigned long  res       = get_block_execfreq_ulong(exec_freq, copy_bl);
 223
 224                         /* create afe-matrix */
 225                         unsigned row, col;
 226                         for (row = 0; row < colors_n; row++) {
 227                                 for (col = 0; col < colors_n; col++) {
 228                                         if (row != col)
 229                                                 pbqp_matrix_set(afe_matrix, row, col, (num)res);
 230                                 }
 231                         }
 232                 }
 233                 else {
 234                         afe_matrix = pbqp_alloc_env->aff_matrix_template;
 235                 }
 236 #if DO_USEFUL_OPT || USE_BIPARTIT_MATCHING
 237                 /* do useful optimization to speed up pbqp solving */
 238                 if (get_free_regs(restr_nodes, cls, src_node) == 1 && get_free_regs(restr_nodes, cls, trg_node) == 1) {
 239                         return;
 240                 }
 241                 if (get_free_regs(restr_nodes, cls, src_node) == 1 || get_free_regs(restr_nodes, cls, trg_node) == 1) {
 242                         if (get_free_regs(restr_nodes, cls, src_node) == 1) {
 243                                 unsigned regIdx = vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs);
 244                                 vector_add_matrix_col(get_node(pbqp, get_irn_idx(trg_node))->costs, afe_matrix, regIdx);
 245                         }
 246                         else {
 247                                 unsigned regIdx = vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs);
 248                                 vector_add_matrix_col(get_node(pbqp, get_irn_idx(src_node))->costs, afe_matrix, regIdx);
 249                         }
 250                         return;
 251                 }
 252 #endif
 253                 /* insert interference edge */
 254                 insert_edge(pbqp, src_node, trg_node, afe_matrix);
 255         }
 256 }
 257
 258 static void create_affinity_edges(ir_node *irn, void *env)
 259 {
 260         be_pbqp_alloc_env_t         *pbqp_alloc_env = (be_pbqp_alloc_env_t*)env;
 261         const arch_register_class_t *cls            = pbqp_alloc_env->cls;
 262         const arch_register_req_t   *req            = arch_get_register_req_out(irn);
 263         unsigned                     pos;
 264         unsigned                     max;
 265
 266         if (is_Reg_Phi(irn)) { /* Phis */
 267                 for (pos = 0, max = get_irn_arity(irn); pos < max; ++pos) {
 268                         ir_node *arg = get_irn_n(irn, pos);
 269
 270                         if (!arch_irn_consider_in_reg_alloc(cls, arg))
 271                                 continue;
 272
 273                         /* no edges to itself */
 274                         if (irn == arg) {
 275                                 continue;
 276                         }
 277
 278                         insert_afe_edge(pbqp_alloc_env, irn, arg, pos);
 279                 }
 280         }
 281         else if (is_Perm_Proj(irn)) { /* Perms */
 282                 ir_node *arg = get_Perm_src(irn);
 283                 if (!arch_irn_consider_in_reg_alloc(cls, arg))
 284                         return;
 285
 286                 insert_afe_edge(pbqp_alloc_env, irn, arg, -1);
 287         }
 288         else { /* 2-address code */
 289                 if (is_2addr_code(req)) {
 290                         const unsigned other = req->other_same;
 291                         int            i;
 292
 293                         for (i = 0; 1U << i <= other; ++i) {
 294                                 if (other & (1U << i)) {
 295                                         ir_node *other = get_irn_n(skip_Proj(irn), i);
 296                                         if (!arch_irn_consider_in_reg_alloc(cls, other))
 297                                                 continue;
 298
 299                                         /* no edges to itself */
 300                                         if (irn == other) {
 301                                                 continue;
 302                                         }
 303
 304                                         insert_afe_edge(pbqp_alloc_env, irn, other, i);
 305                                 }
 306                         }
 307                 }
 308         }
 309 }
 310
 311 static void create_pbqp_coloring_instance(ir_node *block, void *data)
 312 {
 313         be_pbqp_alloc_env_t         *pbqp_alloc_env     = (be_pbqp_alloc_env_t*)data;
 314         be_lv_t                     *lv                 = pbqp_alloc_env->lv;
 315         const arch_register_class_t *cls                = pbqp_alloc_env->cls;
 316         plist_t                     *rpeo               = pbqp_alloc_env->rpeo;
 317         pbqp_t                      *pbqp_inst          = pbqp_alloc_env->pbqp_inst;
 318         plist_t                     *temp_list          = plist_new();
 319         plist_element_t             *el;
 320         ir_node                     *irn;
 321         ir_nodeset_t                 live_nodes;
 322 #if USE_BIPARTIT_MATCHING
 323         int                         *assignment         = ALLOCAN(int, cls->n_regs);
 324 #else
 325         unsigned                    *restr_nodes        = pbqp_alloc_env->restr_nodes;
 326         pqueue_t                    *restr_nodes_queue  = new_pqueue();
 327         pqueue_t                    *queue              = new_pqueue();
 328         plist_t                     *sorted_list        = plist_new();
 329         ir_node                     *last_element       = NULL;
 330 #endif
 331
 332         /* first, determine the pressure */
 333         /* (this is only for compatibility with copymin optimization, it's not needed for pbqp coloring) */
 334         create_borders(block, pbqp_alloc_env->env);
 335
 336         /* calculate living nodes for the first step */
 337         ir_nodeset_init(&live_nodes);
 338         be_liveness_end_of_block(lv, cls, block, &live_nodes);
 339
 340         /* create pbqp nodes, interference edges and reverse perfect elimination order */
 341         sched_foreach_reverse(block, irn) {
 342                 ir_node               *live;
 343                 ir_nodeset_iterator_t  iter;
 344
 345                 if (get_irn_mode(irn) == mode_T) {
 346                         const ir_edge_t *edge;
 347                         foreach_out_edge(irn, edge) {
 348                                 ir_node *proj = get_edge_src_irn(edge);
 349                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 350                                         continue;
 351
 352                                 /* create pbqp source node if it dosn't exist */
 353                                 if (get_node(pbqp_inst, get_irn_idx(proj)) == NULL) {
 354                                         create_pbqp_node(pbqp_alloc_env, proj);
 355                                 }
 356
 357                                 /* create nodes and interference edges */
 358                                 foreach_ir_nodeset(&live_nodes, live, iter) {
 359                                         /* create pbqp source node if it dosn't exist */
 360                                         if (get_node(pbqp_inst, get_irn_idx(live)) == NULL) {
 361                                                 create_pbqp_node(pbqp_alloc_env, live);
 362                                         }
 363
 364                                         /* no edges to itself */
 365                                         if (proj == live) {
 366                                                 continue;
 367                                         }
 368
 369                                         insert_ife_edge(pbqp_alloc_env, proj, live);
 370                                 }
 371                         }
 372                 }
 373                 else {
 374                         if (arch_irn_consider_in_reg_alloc(cls, irn)) {
 375                                 /* create pbqp source node if it dosn't exist */
 376                                 if (get_node(pbqp_inst, get_irn_idx(irn)) == NULL) {
 377                                         create_pbqp_node(pbqp_alloc_env, irn);
 378                                 }
 379
 380                                 /* create nodes and interference edges */
 381                                 foreach_ir_nodeset(&live_nodes, live, iter) {
 382                                         /* create pbqp source node if it dosn't exist */
 383                                         if (get_node(pbqp_inst, get_irn_idx(live)) == NULL) {
 384                                                 create_pbqp_node(pbqp_alloc_env, live);
 385                                         }
 386
 387                                         /* no edges to itself */
 388                                         if (irn == live) {
 389                                                 continue;
 390                                         }
 391
 392                                         /* insert interference edge */
 393                                         insert_ife_edge(pbqp_alloc_env, irn, live);
 394                                 }
 395                         }
 396                 }
 397
 398                 /* get living nodes for next step */
 399                 if (!is_Phi(irn)) {
 400                         be_liveness_transfer(cls, irn, &live_nodes);
 401                 }
 402
 403 #if USE_BIPARTIT_MATCHING
 404                 if (get_irn_mode(irn) == mode_T) {
 405                         unsigned     clique_size         = 0;
 406                         unsigned     n_alloc             = 0;
 407                         pbqp_node   *clique[cls->n_regs];
 408                         bipartite_t *bp                  = bipartite_new(cls->n_regs, cls->n_regs);
 409
 410                         /* add all proj after a perm to clique */
 411                         const ir_edge_t *edge;
 412                         foreach_out_edge(irn, edge) {
 413                                 ir_node *proj = get_edge_src_irn(edge);
 414
 415                                 /* ignore node if it is not necessary for register allocation */
 416                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 417                                         continue;
 418
 419                                 /* insert pbqp node into temp rpeo list of this block */
 420                                 plist_insert_front(temp_list, get_node(pbqp_inst, get_irn_idx(proj)));
 421
 422                                 if(is_Perm_Proj(proj)) {
 423                                         /* add proj to clique */
 424                                         pbqp_node *clique_member = get_node(pbqp_inst,proj->node_idx);
 425                                         vector    *costs         = clique_member->costs;
 426                                         unsigned   idx           = 0;
 427
 428                                         clique[clique_size] = clique_member;
 429
 430                                         for(idx = 0; idx < costs->len; idx++) {
 431                                                 if(costs->entries[idx].data != INF_COSTS) {
 432                                                         bipartite_add(bp, clique_size, idx);
 433                                                 }
 434                                         }
 435
 436                                         /* increase node counter */
 437                                         clique_size++;
 438                                         n_alloc++;
 439                                 }
 440                         }
 441
 442                         if(clique_size > 0) {
 443                                 plist_element_t *listElement;
 444                                 foreach_plist(temp_list, listElement) {
 445                                         pbqp_node *clique_candidate  = listElement->data;
 446                                         unsigned   idx               = 0;
 447                                         bool       isMember          = true;
 448
 449                                         /* clique size not bigger then register class size */
 450                                         if(clique_size >= cls->n_regs) break;
 451
 452                                         for(idx = 0; idx < clique_size; idx++) {
 453                                                 pbqp_node *member = clique[idx];
 454
 455                                                 if(member == clique_candidate) {
 456                                                         isMember = false;
 457                                                         break;
 458                                                 }
 459
 460                                                 if(get_edge(pbqp_inst, member->index, clique_candidate->index) == NULL && get_edge(pbqp_inst, clique_candidate->index, member->index) == NULL) {
 461                                                         isMember = false;
 462                                                         break;
 463                                                 }
 464                                         }
 465
 466                                         /* goto next list element if current node is not a member of the clique */
 467                                         if(!isMember) { continue; }
 468
 469                                         /* add candidate to clique */
 470                                         clique[clique_size] = clique_candidate;
 471
 472                                         vector *costs = clique_candidate->costs;
 473                                         for(idx = 0; idx < costs->len; idx++) {
 474                                                 if(costs->entries[idx].data != INF_COSTS) {
 475                                                         bipartite_add(bp, clique_size, idx);
 476                                                 }
 477                                         }
 478
 479                                         /* increase node counter */
 480                                         clique_size++;
 481                                 }
 482                         }
 483
 484                         /* solve bipartite matching */
 485                         bipartite_matching(bp, assignment);
 486
 487                         /* assign colors */
 488                         unsigned nodeIdx = 0;
 489                         for(nodeIdx = 0; nodeIdx < clique_size; nodeIdx++) {
 490                                 vector *costs = clique[nodeIdx]->costs;
 491                                 int     idx;
 492                                 for(idx = 0; idx < (int)costs->len; idx++) {
 493                                         if(assignment[nodeIdx] != idx) {
 494                                                 costs->entries[idx].data = INF_COSTS;
 495                                         }
 496                                 }
 497                                 assert(assignment[nodeIdx] >= 0 && "there must have been a register assigned (node not register pressure faithful?)");
 498                         }
 499
 500                         /* free memory */
 501                         bipartite_free(bp);
 502                 }
 503                 else {
 504                         if (arch_irn_consider_in_reg_alloc(cls, irn)) {
 505                                 plist_insert_front(temp_list, get_node(pbqp_inst, get_irn_idx(irn)));
 506                         }
 507                 }
 508 #else
 509                 /* order nodes for perfect elimination order */
 510                 if (get_irn_mode(irn) == mode_T) {
 511                         bool             allHaveIFEdges = true;
 512                         const ir_edge_t *edge;
 513
 514                         foreach_out_edge(irn, edge) {
 515                                 ir_node *proj = get_edge_src_irn(edge);
 516                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 517                                         continue;
 518
 519                                 /* insert proj node into priority queue (descending by the number of interference edges) */
 520                                 if (get_free_regs(restr_nodes, cls, proj) <= 4) {
 521                                         pqueue_put(restr_nodes_queue, proj, pbqp_alloc_env->ife_edge_num[get_irn_idx(proj)]);
 522                                 }
 523                                 else {
 524                                         pqueue_put(queue, proj, pbqp_alloc_env->ife_edge_num[get_irn_idx(proj)]);
 525                                 }
 526
 527                                 /* skip last step if there is no last_element */
 528                                 if(last_element == NULL)
 529                                         continue;
 530
 531                                 /* check if proj has an if edge to last_element (at this time pbqp contains only if edges) */
 532                                 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) {
 533                                         allHaveIFEdges = false; /* there is no if edge between proj and last_element */
 534                                 }
 535                         }
 536
 537                         if(last_element != NULL && allHaveIFEdges) {
 538                                 if (get_free_regs(restr_nodes, cls, last_element) <= 4) {
 539                                         pqueue_put(restr_nodes_queue, last_element, pbqp_alloc_env->ife_edge_num[get_irn_idx(last_element)]);
 540                                 }
 541                                 else {
 542                                         pqueue_put(queue, last_element, pbqp_alloc_env->ife_edge_num[get_irn_idx(last_element)]);
 543                                 }
 544                                 plist_erase(temp_list, plist_find_value(temp_list, get_node(pbqp_inst, last_element->node_idx)));
 545                                 last_element = NULL;
 546                         }
 547
 548                         /* first insert all restricted proj nodes */
 549                         while (!pqueue_empty(restr_nodes_queue)) {
 550                                 ir_node *node = (ir_node*)pqueue_pop_front(restr_nodes_queue);
 551                                 plist_insert_front(sorted_list, get_node(pbqp_inst, get_irn_idx(node)));
 552                         }
 553
 554                         /* insert proj nodes descending by their number of interference edges */
 555                         while (!pqueue_empty(queue)) {
 556                                 ir_node *node = (ir_node*)pqueue_pop_front(queue);
 557                                 plist_insert_front(sorted_list, get_node(pbqp_inst, get_irn_idx(node)));
 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         be_chordal_env_t *env    = (be_chordal_env_t*)data;
 601         ir_node          *irn;
 602
 603         for (irn = sched_first(block); !sched_is_end(irn);) {
 604                 be_insn_t *insn = chordal_scan_insn(env, irn);
 605                 irn             = insn->next_insn;
 606
 607                 if (!insn->has_constraints)
 608                         continue;
 609
 610                 pre_process_constraints(env, &insn);
 611         }
 612 }
 613
 614 static void be_pbqp_coloring(be_chordal_env_t *env)
 615 {
 616         ir_graph                    *irg            = env->irg;
 617         const arch_register_class_t *cls            = env->cls;
 618         be_lv_t                     *lv             = NULL;
 619         plist_element_t             *element        = NULL;
 620         unsigned                     colors_n       = arch_register_class_n_regs(cls);
 621         be_pbqp_alloc_env_t          pbqp_alloc_env;
 622         unsigned                     col;
 623         unsigned                     row;
 624         pbqp_matrix_t               *ife_matrix;
 625         num                          solution;
 626 #if KAPS_DUMP
 627         FILE                        *file_before;
 628 #endif
 629 #if TIMER
 630         ir_timer_t *t_ra_pbqp_alloc_create     = ir_timer_new();
 631         ir_timer_t *t_ra_pbqp_alloc_solve      = ir_timer_new();
 632         ir_timer_t *t_ra_pbqp_alloc_create_aff = ir_timer_new();
 633
 634         printf("#### ----- === Allocating registers of %s (%s) ===\n", cls->name, get_entity_name(get_irg_entity(irg)));
 635 #endif
 636         lv = be_assure_liveness(irg);
 637         be_liveness_assure_sets(lv);
 638         be_liveness_assure_chk(lv);
 639
 640         /* insert perms */
 641         assure_doms(irg);
 642         dom_tree_walk_irg(irg, insert_perms, NULL, env);
 643
 644         /* dump graph after inserting perms */
 645         if (env->opts->dump_flags & BE_CH_DUMP_CONSTR) {
 646                 char buf[256];
 647                 snprintf(buf, sizeof(buf), "-%s-constr", cls->name);
 648                 dump_ir_graph(irg, buf);
 649         }
 650
 651
 652         /* initialize pbqp allocation data structure */
 653         pbqp_alloc_env.pbqp_inst        = alloc_pbqp(get_irg_last_idx(irg));  /* initialize pbqp instance */
 654         pbqp_alloc_env.cls              = cls;
 655         pbqp_alloc_env.irg              = irg;
 656         pbqp_alloc_env.lv               = lv;
 657         pbqp_alloc_env.allocatable_regs = bitset_malloc(colors_n);
 658         pbqp_alloc_env.rpeo             = plist_new();
 659         pbqp_alloc_env.restr_nodes      = XMALLOCNZ(unsigned, get_irg_last_idx(irg));
 660         pbqp_alloc_env.ife_edge_num     = XMALLOCNZ(unsigned, get_irg_last_idx(irg));
 661         pbqp_alloc_env.env              = env;
 662         be_put_allocatable_regs(irg, cls, pbqp_alloc_env.allocatable_regs);
 663
 664
 665         /* create costs matrix template for interference edges */
 666         ife_matrix = pbqp_matrix_alloc(pbqp_alloc_env.pbqp_inst, colors_n, colors_n);
 667         /* set costs */
 668         for (row = 0, col = 0; row < colors_n; row++, col++)
 669                 pbqp_matrix_set(ife_matrix, row, col, INF_COSTS);
 670
 671         pbqp_alloc_env.ife_matrix_template = ife_matrix;
 672
 673
 674         if (!use_exec_freq) {
 675                 /* create costs matrix template for affinity edges */
 676                 pbqp_matrix_t *afe_matrix = pbqp_matrix_alloc(pbqp_alloc_env.pbqp_inst, colors_n, colors_n);
 677                 /* set costs */
 678                 for (row = 0; row < colors_n; row++) {
 679                         for (col = 0; col < colors_n; col++) {
 680                                 if (row != col)
 681                                         pbqp_matrix_set(afe_matrix, row, col, 2);
 682                         }
 683                 }
 684                 pbqp_alloc_env.aff_matrix_template = afe_matrix;
 685         }
 686
 687
 688         /* create pbqp instance */
 689 #if TIMER
 690         ir_timer_reset_and_start(t_ra_pbqp_alloc_create);
 691 #endif
 692         assure_doms(irg);
 693         dom_tree_walk_irg(irg, create_pbqp_coloring_instance , NULL, &pbqp_alloc_env);
 694 #if TIMER
 695         ir_timer_stop(t_ra_pbqp_alloc_create);
 696 #endif
 697
 698
 699         /* set up affinity edges */
 700 #if TIMER
 701         ir_timer_reset_and_start(t_ra_pbqp_alloc_create_aff);
 702 #endif
 703         foreach_plist(pbqp_alloc_env.rpeo, element) {
 704                 pbqp_node_t *node = (pbqp_node_t*)element->data;
 705                 ir_node     *irn  = get_idx_irn(irg, node->index);
 706
 707                 create_affinity_edges(irn, &pbqp_alloc_env);
 708         }
 709 #if TIMER
 710         ir_timer_stop(t_ra_pbqp_alloc_create_aff);
 711 #endif
 712
 713
 714 #if KAPS_DUMP
 715         // dump graph before solving pbqp
 716         file_before = my_open(env, "", "-pbqp_coloring.html");
 717         set_dumpfile(pbqp_alloc_env.pbqp_inst, file_before);
 718 #endif
 719
 720         /* print out reverse perfect elimination order */
 721 #if PRINT_RPEO
 722         {
 723                 plist_element_t *elements;
 724                 foreach_plist(pbqp_alloc_env.rpeo, elements) {
 725                         pbqp_node_t *node = elements->data;
 726                         printf(" %d(%ld);", node->index, get_idx_irn(irg, node->index)->node_nr);
 727                 }
 728                 printf("\n");
 729         }
 730 #endif
 731
 732         /* solve pbqp instance */
 733 #if TIMER
 734         ir_timer_reset_and_start(t_ra_pbqp_alloc_solve);
 735 #endif
 736         if(use_late_decision) {
 737                 solve_pbqp_heuristical_co_ld(pbqp_alloc_env.pbqp_inst,pbqp_alloc_env.rpeo);
 738         }
 739         else {
 740                 solve_pbqp_heuristical_co(pbqp_alloc_env.pbqp_inst,pbqp_alloc_env.rpeo);
 741         }
 742 #if TIMER
 743         ir_timer_stop(t_ra_pbqp_alloc_solve);
 744 #endif
 745
 746
 747         solution = get_solution(pbqp_alloc_env.pbqp_inst);
 748         if (solution == INF_COSTS)
 749                 panic("No PBQP solution found");
 750
 751
 752         /* assign colors */
 753         foreach_plist(pbqp_alloc_env.rpeo, element) {
 754                 pbqp_node_t           *node  = (pbqp_node_t*)element->data;
 755                 ir_node               *irn   = get_idx_irn(irg, node->index);
 756                 num                    color = get_node_solution(pbqp_alloc_env.pbqp_inst, node->index);
 757                 const arch_register_t *reg   = arch_register_for_index(cls, color);
 758
 759                 arch_set_irn_register(irn, reg);
 760         }
 761
 762
 763 #if TIMER
 764         printf("PBQP alloc create:     %10.3lf msec\n",
 765                (double)ir_timer_elapsed_usec(t_ra_pbqp_alloc_create) / 1000.0);
 766         printf("PBQP alloc solve:      %10.3lf msec\n",
 767                (double)ir_timer_elapsed_usec(t_ra_pbqp_alloc_solve) / 1000.0);
 768         printf("PBQP alloc create aff: %10.3lf msec\n",
 769                (double)ir_timer_elapsed_usec(t_ra_pbqp_alloc_create_aff) / 1000.0);
 770 #endif
 771
 772
 773         /* free reserved memory */
 774 #if KAPS_DUMP
 775         fclose(file_before);
 776 #endif
 777         bitset_free(pbqp_alloc_env.allocatable_regs);
 778         free_pbqp(pbqp_alloc_env.pbqp_inst);
 779         plist_free(pbqp_alloc_env.rpeo);
 780         xfree(pbqp_alloc_env.restr_nodes);
 781         xfree(pbqp_alloc_env.ife_edge_num);
 782 }
 783
 784
 785 /**
 786  * Initializes this module.
 787  */
 788 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_pbqp_coloring)
 789 void be_init_pbqp_coloring(void)
 790 {
 791         lc_opt_entry_t *be_grp       = lc_opt_get_grp(firm_opt_get_root(), "be");
 792         lc_opt_entry_t *ra_grp       = lc_opt_get_grp(be_grp, "ra");
 793         lc_opt_entry_t *chordal_grp  = lc_opt_get_grp(ra_grp, "chordal");
 794         lc_opt_entry_t *coloring_grp = lc_opt_get_grp(chordal_grp, "coloring");
 795         lc_opt_entry_t *pbqp_grp     = lc_opt_get_grp(coloring_grp, "pbqp");
 796
 797         static be_ra_chordal_coloring_t coloring = {
 798                 be_pbqp_coloring
 799         };
 800
 801         lc_opt_add_table(pbqp_grp, options);
 802         be_register_chordal_coloring("pbqp", &coloring);
 803 }