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