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

   1 /*
   2  * This file is part of libFirm.
   3  * Copyright (C) 2012 University of Karlsruhe.
   4  */
   5
   6 /**
   7  * @file
   8  * @brief       PBQP based register allocation.
   9  * @author      Thomas Bersch
  10  * @date        27.11.2009
  11  */
  12
  13 /* miscellaneous includes */
  14 #include "config.h"
  15
  16 #include "debug.h"
  17 #include "error.h"
  18
  19 #include "irdom.h"
  20 #include "irdump.h"
  21 #include "iredges_t.h"
  22 #include "irprintf.h"
  23 #include "irgwalk.h"
  24 #include "irtools.h"
  25 #include "time.h"
  26 #include "execfreq_t.h"
  27 #include "bipartite.h"
  28
  29 /* libfirm/ir/be includes */
  30 #include "bearch.h"
  31 #include "beirg.h"
  32 #include "besched.h"
  33 #include "bemodule.h"
  34 #include "bechordal_common.h"
  35 #include "bechordal.h"
  36 #include "bechordal_t.h"
  37 #include "beinsn_t.h"
  38 #include "benode.h"
  39 #include "belive.h"
  40 #include "belive_t.h"
  41 #include "beutil.h"
  42 #include "plist.h"
  43 #include "pqueue.h"
  44 #include "becopyopt.h"
  45
  46 /* pbqp includes */
  47 #include "kaps.h"
  48 #include "matrix.h"
  49 #include "vector.h"
  50 #include "vector_t.h"
  51 #include "heuristical_co.h"
  52 #include "heuristical_co_ld.h"
  53 #include "pbqp_t.h"
  54 #include "html_dumper.h"
  55 #include "pbqp_node_t.h"
  56 #include "pbqp_node.h"
  57 #include "pbqp_edge_t.h"
  58
  59 #define TIMER                 0
  60 #define PRINT_RPEO            0
  61 #define USE_BIPARTIT_MATCHING 0
  62 #define DO_USEFUL_OPT         1
  63
  64
  65 static int use_exec_freq     = true;
  66 static int use_late_decision = false;
  67
  68 typedef struct be_pbqp_alloc_env_t {
  69         pbqp_t                      *pbqp_inst;         /**< PBQP instance for register allocation */
  70         ir_graph                    *irg;               /**< The graph under examination. */
  71         const arch_register_class_t *cls;               /**< Current processed register class */
  72         be_lv_t                     *lv;
  73         bitset_t                    *allocatable_regs;
  74         pbqp_matrix_t               *ife_matrix_template;
  75         pbqp_matrix_t               *aff_matrix_template;
  76         plist_t                     *rpeo;
  77         unsigned                    *restr_nodes;
  78         unsigned                    *ife_edge_num;
  79         ir_execfreq_int_factors      execfreq_factors;
  80         be_chordal_env_t            *env;
  81 } be_pbqp_alloc_env_t;
  82
  83
  84 #define is_Reg_Phi(irn)                                        (is_Phi(irn) && mode_is_data(get_irn_mode(irn)))
  85 #define get_Perm_src(irn)                                      (get_irn_n(get_Proj_pred(irn), get_Proj_proj(irn)))
  86 #define is_Perm_Proj(irn)                                      (is_Proj(irn) && be_is_Perm(get_Proj_pred(irn)))
  87 #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)))
  88 #define get_free_regs(restr_nodes, cls, irn)                   (arch_register_class_n_regs(cls) - restr_nodes[get_irn_idx(irn)])
  89
  90 static const lc_opt_table_entry_t options[] = {
  91         LC_OPT_ENT_BOOL("exec_freq", "use exec_freq",  &use_exec_freq),
  92         LC_OPT_ENT_BOOL("late_decision", "use late decision for register allocation",  &use_late_decision),
  93         LC_OPT_LAST
  94 };
  95
  96 #if KAPS_DUMP
  97 static FILE *my_open(const be_chordal_env_t *env, const char *prefix, const char *suffix)
  98 {
  99         FILE       *result;
 100         char        buf[1024];
 101         size_t      i;
 102         size_t      n;
 103         char       *tu_name;
 104         const char *cup_name = be_get_irg_main_env(env->irg)->cup_name;
 105
 106         n = strlen(cup_name);
 107         tu_name = XMALLOCN(char, n + 1);
 108         strcpy(tu_name, cup_name);
 109         for (i = 0; i < n; ++i)
 110                 if (tu_name[i] == '.')
 111                         tu_name[i] = '_';
 112
 113         ir_snprintf(buf, sizeof(buf), "%s%s_%F_%s%s", prefix, tu_name, env->irg, env->cls->name, suffix);
 114         xfree(tu_name);
 115         result = fopen(buf, "wt");
 116         if (result == NULL) {
 117                 panic("Couldn't open '%s' for writing.", buf);
 118         }
 119
 120         return result;
 121 }
 122 #endif
 123
 124
 125 static void create_pbqp_node(be_pbqp_alloc_env_t *pbqp_alloc_env, ir_node *irn)
 126 {
 127         const arch_register_class_t *cls = pbqp_alloc_env->cls;
 128         pbqp_t   *pbqp_inst              = pbqp_alloc_env->pbqp_inst;
 129         bitset_t *allocatable_regs       = pbqp_alloc_env->allocatable_regs;
 130         unsigned  colors_n               = arch_register_class_n_regs(cls);
 131         unsigned  cntConstrains          = 0;
 132
 133         /* create costs vector depending on register constrains */
 134         vector_t *costs_vector = vector_alloc(pbqp_inst, colors_n);
 135
 136         /* set costs depending on register constrains */
 137         unsigned idx;
 138         for (idx = 0; idx < colors_n; idx++) {
 139                 const arch_register_req_t *req = arch_get_irn_register_req(irn);
 140                 const arch_register_t     *reg = arch_register_for_index(cls, idx);
 141                 if (!bitset_is_set(allocatable_regs, idx)
 142                     || !arch_reg_is_allocatable(req, reg)) {
 143                         /* constrained */
 144                         vector_set(costs_vector, idx, INF_COSTS);
 145                         cntConstrains++;
 146                 }
 147         }
 148
 149         /* add vector to pbqp node */
 150         add_node_costs(pbqp_inst, get_irn_idx(irn), costs_vector);
 151         pbqp_alloc_env->restr_nodes[get_irn_idx(irn)] = cntConstrains;
 152 }
 153
 154 static void insert_ife_edge(be_pbqp_alloc_env_t *pbqp_alloc_env, ir_node *src_node, ir_node *trg_node)
 155 {
 156         pbqp_t                      *pbqp                = pbqp_alloc_env->pbqp_inst;
 157         const arch_register_class_t *cls                 = pbqp_alloc_env->cls;
 158         pbqp_matrix_t               *ife_matrix_template = pbqp_alloc_env->ife_matrix_template;
 159         unsigned                    *restr_nodes         = pbqp_alloc_env->restr_nodes;
 160
 161         if (get_edge(pbqp, get_irn_idx(src_node), get_irn_idx(trg_node)) == NULL) {
 162
 163                 /* increase ife edge counter */
 164                 pbqp_alloc_env->ife_edge_num[get_irn_idx(src_node)]++;
 165                 pbqp_alloc_env->ife_edge_num[get_irn_idx(trg_node)]++;
 166
 167 #if DO_USEFUL_OPT || USE_BIPARTIT_MATCHING
 168                 /* do useful optimization to speed up pbqp solving (we can do this because we know our matrix) */
 169                 if (get_free_regs(restr_nodes, cls, src_node) == 1 && get_free_regs(restr_nodes, cls, trg_node) == 1) {
 170                         assert(vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs) !=
 171                                vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs) &&
 172                                "Interfering nodes must not have the same register!");
 173                         return;
 174                 }
 175                 if (get_free_regs(restr_nodes, cls, src_node) == 1 || get_free_regs(restr_nodes, cls, trg_node) == 1) {
 176                         if (get_free_regs(restr_nodes, cls, src_node) == 1) {
 177                                 unsigned idx = vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs);
 178                                 vector_set(get_node(pbqp, get_irn_idx(trg_node))->costs, idx, INF_COSTS);
 179                         }
 180                         else {
 181                                 unsigned idx = vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs);
 182                                 vector_set(get_node(pbqp, get_irn_idx(src_node))->costs, idx, INF_COSTS);
 183                         }
 184                         return;
 185                 }
 186 #endif
 187                 /* insert interference edge */
 188                 insert_edge(pbqp, src_node, trg_node, ife_matrix_template);
 189         }
 190 }
 191
 192 static void insert_afe_edge(be_pbqp_alloc_env_t *pbqp_alloc_env, ir_node *src_node, ir_node *trg_node, int pos)
 193 {
 194         pbqp_t                      *pbqp        = pbqp_alloc_env->pbqp_inst;
 195         const arch_register_class_t *cls         = pbqp_alloc_env->cls;
 196         unsigned                    *restr_nodes = pbqp_alloc_env->restr_nodes;
 197         pbqp_matrix_t               *afe_matrix  = pbqp_matrix_alloc(pbqp, arch_register_class_n_regs(cls), arch_register_class_n_regs(cls));
 198         unsigned                     colors_n    = arch_register_class_n_regs(cls);
 199
 200         if (get_edge(pbqp, get_irn_idx(src_node), get_irn_idx(trg_node)) == NULL) {
 201                 if (use_exec_freq) {
 202                         /* get exec_freq for copy_block */
 203                         ir_node *root_bl = get_nodes_block(src_node);
 204                         ir_node *copy_bl = is_Phi(src_node) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
 205                         int      res     = get_block_execfreq_int(&pbqp_alloc_env->execfreq_factors, copy_bl);
 206
 207                         /* create afe-matrix */
 208                         unsigned row, col;
 209                         for (row = 0; row < colors_n; row++) {
 210                                 for (col = 0; col < colors_n; col++) {
 211                                         if (row != col)
 212                                                 pbqp_matrix_set(afe_matrix, row, col, (num)res);
 213                                 }
 214                         }
 215                 }
 216                 else {
 217                         afe_matrix = pbqp_alloc_env->aff_matrix_template;
 218                 }
 219 #if DO_USEFUL_OPT || USE_BIPARTIT_MATCHING
 220                 /* do useful optimization to speed up pbqp solving */
 221                 if (get_free_regs(restr_nodes, cls, src_node) == 1 && get_free_regs(restr_nodes, cls, trg_node) == 1) {
 222                         return;
 223                 }
 224                 if (get_free_regs(restr_nodes, cls, src_node) == 1 || get_free_regs(restr_nodes, cls, trg_node) == 1) {
 225                         if (get_free_regs(restr_nodes, cls, src_node) == 1) {
 226                                 unsigned regIdx = vector_get_min_index(get_node(pbqp, get_irn_idx(src_node))->costs);
 227                                 vector_add_matrix_col(get_node(pbqp, get_irn_idx(trg_node))->costs, afe_matrix, regIdx);
 228                         }
 229                         else {
 230                                 unsigned regIdx = vector_get_min_index(get_node(pbqp, get_irn_idx(trg_node))->costs);
 231                                 vector_add_matrix_col(get_node(pbqp, get_irn_idx(src_node))->costs, afe_matrix, regIdx);
 232                         }
 233                         return;
 234                 }
 235 #endif
 236                 /* insert interference edge */
 237                 insert_edge(pbqp, src_node, trg_node, afe_matrix);
 238         }
 239 }
 240
 241 static void create_affinity_edges(ir_node *irn, void *env)
 242 {
 243         be_pbqp_alloc_env_t         *pbqp_alloc_env = (be_pbqp_alloc_env_t*)env;
 244         const arch_register_class_t *cls            = pbqp_alloc_env->cls;
 245         const arch_register_req_t   *req            = arch_get_irn_register_req(irn);
 246         unsigned                     pos;
 247         unsigned                     max;
 248
 249         if (is_Reg_Phi(irn)) { /* Phis */
 250                 for (pos = 0, max = get_irn_arity(irn); pos < max; ++pos) {
 251                         ir_node *arg = get_irn_n(irn, pos);
 252
 253                         if (!arch_irn_consider_in_reg_alloc(cls, arg))
 254                                 continue;
 255
 256                         /* no edges to itself */
 257                         if (irn == arg) {
 258                                 continue;
 259                         }
 260
 261                         insert_afe_edge(pbqp_alloc_env, irn, arg, pos);
 262                 }
 263         }
 264         else if (is_Perm_Proj(irn)) { /* Perms */
 265                 ir_node *arg = get_Perm_src(irn);
 266                 if (!arch_irn_consider_in_reg_alloc(cls, arg))
 267                         return;
 268
 269                 insert_afe_edge(pbqp_alloc_env, irn, arg, -1);
 270         } else if (arch_register_req_is(req, should_be_same)) {
 271                 const unsigned other = req->other_same;
 272                 int            i;
 273
 274                 for (i = 0; 1U << i <= other; ++i) {
 275                         if (other & (1U << i)) {
 276                                 ir_node *other = get_irn_n(skip_Proj(irn), i);
 277                                 if (!arch_irn_consider_in_reg_alloc(cls, other))
 278                                         continue;
 279
 280                                 /* no edges to itself */
 281                                 if (irn == other) {
 282                                         continue;
 283                                 }
 284
 285                                 insert_afe_edge(pbqp_alloc_env, irn, other, i);
 286                         }
 287                 }
 288         }
 289 }
 290
 291 static void create_pbqp_coloring_instance(ir_node *block, void *data)
 292 {
 293         be_pbqp_alloc_env_t         *pbqp_alloc_env     = (be_pbqp_alloc_env_t*)data;
 294         be_lv_t                     *lv                 = pbqp_alloc_env->lv;
 295         const arch_register_class_t *cls                = pbqp_alloc_env->cls;
 296         plist_t                     *rpeo               = pbqp_alloc_env->rpeo;
 297         pbqp_t                      *pbqp_inst          = pbqp_alloc_env->pbqp_inst;
 298         plist_t                     *temp_list          = plist_new();
 299         plist_element_t             *el;
 300         ir_nodeset_t                 live_nodes;
 301 #if USE_BIPARTIT_MATCHING
 302         int                         *assignment         = ALLOCAN(int, cls->n_regs);
 303 #else
 304         unsigned                    *restr_nodes        = pbqp_alloc_env->restr_nodes;
 305         pqueue_t                    *restr_nodes_queue  = new_pqueue();
 306         pqueue_t                    *queue              = new_pqueue();
 307         plist_t                     *sorted_list        = plist_new();
 308         ir_node                     *last_element       = NULL;
 309 #endif
 310
 311         /* first, determine the pressure */
 312         /* (this is only for compatibility with copymin optimization, it's not needed for pbqp coloring) */
 313         create_borders(block, pbqp_alloc_env->env);
 314
 315         /* calculate living nodes for the first step */
 316         ir_nodeset_init(&live_nodes);
 317         be_liveness_end_of_block(lv, cls, block, &live_nodes);
 318
 319         /* create pbqp nodes, interference edges and reverse perfect elimination order */
 320         sched_foreach_reverse(block, irn) {
 321                 be_foreach_value(irn, value,
 322                         if (!arch_irn_consider_in_reg_alloc(cls, value))
 323                                 continue;
 324
 325                         /* create pbqp source node if it dosn't exist */
 326                         if (!get_node(pbqp_inst, get_irn_idx(value)))
 327                                 create_pbqp_node(pbqp_alloc_env, value);
 328
 329                         /* create nodes and interference edges */
 330                         foreach_ir_nodeset(&live_nodes, live, iter) {
 331                                 /* create pbqp source node if it dosn't exist */
 332                                 if (!get_node(pbqp_inst, get_irn_idx(live)))
 333                                         create_pbqp_node(pbqp_alloc_env, live);
 334
 335                                 /* no edges to itself */
 336                                 if (value == live)
 337                                         continue;
 338
 339                                 insert_ife_edge(pbqp_alloc_env, value, live);
 340                         }
 341                 );
 342
 343                 /* get living nodes for next step */
 344                 if (!is_Phi(irn)) {
 345                         be_liveness_transfer(cls, irn, &live_nodes);
 346                 }
 347
 348 #if USE_BIPARTIT_MATCHING
 349                 if (get_irn_mode(irn) == mode_T) {
 350                         unsigned     clique_size         = 0;
 351                         unsigned     n_alloc             = 0;
 352                         pbqp_node   *clique[cls->n_regs];
 353                         bipartite_t *bp                  = bipartite_new(cls->n_regs, cls->n_regs);
 354
 355                         /* add all proj after a perm to clique */
 356                         foreach_out_edge(irn, edge) {
 357                                 ir_node *proj = get_edge_src_irn(edge);
 358
 359                                 /* ignore node if it is not necessary for register allocation */
 360                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 361                                         continue;
 362
 363                                 /* insert pbqp node into temp rpeo list of this block */
 364                                 plist_insert_front(temp_list, get_node(pbqp_inst, get_irn_idx(proj)));
 365
 366                                 if(is_Perm_Proj(proj)) {
 367                                         /* add proj to clique */
 368                                         pbqp_node *clique_member = get_node(pbqp_inst,proj->node_idx);
 369                                         vector    *costs         = clique_member->costs;
 370                                         unsigned   idx           = 0;
 371
 372                                         clique[clique_size] = clique_member;
 373
 374                                         for(idx = 0; idx < costs->len; idx++) {
 375                                                 if(costs->entries[idx].data != INF_COSTS) {
 376                                                         bipartite_add(bp, clique_size, idx);
 377                                                 }
 378                                         }
 379
 380                                         /* increase node counter */
 381                                         clique_size++;
 382                                         n_alloc++;
 383                                 }
 384                         }
 385
 386                         if(clique_size > 0) {
 387                                 plist_element_t *listElement;
 388                                 foreach_plist(temp_list, listElement) {
 389                                         pbqp_node *clique_candidate  = listElement->data;
 390                                         unsigned   idx               = 0;
 391                                         bool       isMember          = true;
 392
 393                                         /* clique size not bigger then register class size */
 394                                         if(clique_size >= cls->n_regs) break;
 395
 396                                         for(idx = 0; idx < clique_size; idx++) {
 397                                                 pbqp_node *member = clique[idx];
 398
 399                                                 if(member == clique_candidate) {
 400                                                         isMember = false;
 401                                                         break;
 402                                                 }
 403
 404                                                 if(get_edge(pbqp_inst, member->index, clique_candidate->index) == NULL && get_edge(pbqp_inst, clique_candidate->index, member->index) == NULL) {
 405                                                         isMember = false;
 406                                                         break;
 407                                                 }
 408                                         }
 409
 410                                         /* goto next list element if current node is not a member of the clique */
 411                                         if(!isMember) { continue; }
 412
 413                                         /* add candidate to clique */
 414                                         clique[clique_size] = clique_candidate;
 415
 416                                         vector *costs = clique_candidate->costs;
 417                                         for(idx = 0; idx < costs->len; idx++) {
 418                                                 if(costs->entries[idx].data != INF_COSTS) {
 419                                                         bipartite_add(bp, clique_size, idx);
 420                                                 }
 421                                         }
 422
 423                                         /* increase node counter */
 424                                         clique_size++;
 425                                 }
 426                         }
 427
 428                         /* solve bipartite matching */
 429                         bipartite_matching(bp, assignment);
 430
 431                         /* assign colors */
 432                         unsigned nodeIdx = 0;
 433                         for(nodeIdx = 0; nodeIdx < clique_size; nodeIdx++) {
 434                                 vector *costs = clique[nodeIdx]->costs;
 435                                 int     idx;
 436                                 for(idx = 0; idx < (int)costs->len; idx++) {
 437                                         if(assignment[nodeIdx] != idx) {
 438                                                 costs->entries[idx].data = INF_COSTS;
 439                                         }
 440                                 }
 441                                 assert(assignment[nodeIdx] >= 0 && "there must have been a register assigned (node not register pressure faithful?)");
 442                         }
 443
 444                         /* free memory */
 445                         bipartite_free(bp);
 446                 }
 447                 else {
 448                         if (arch_irn_consider_in_reg_alloc(cls, irn)) {
 449                                 plist_insert_front(temp_list, get_node(pbqp_inst, get_irn_idx(irn)));
 450                         }
 451                 }
 452 #else
 453                 /* order nodes for perfect elimination order */
 454                 if (get_irn_mode(irn) == mode_T) {
 455                         bool allHaveIFEdges = true;
 456                         foreach_out_edge(irn, edge) {
 457                                 ir_node *proj = get_edge_src_irn(edge);
 458                                 if (!arch_irn_consider_in_reg_alloc(cls, proj))
 459                                         continue;
 460
 461                                 /* insert proj node into priority queue (descending by the number of interference edges) */
 462                                 if (get_free_regs(restr_nodes, cls, proj) <= 4) {
 463                                         pqueue_put(restr_nodes_queue, proj, pbqp_alloc_env->ife_edge_num[get_irn_idx(proj)]);
 464                                 }
 465                                 else {
 466                                         pqueue_put(queue, proj, pbqp_alloc_env->ife_edge_num[get_irn_idx(proj)]);
 467                                 }
 468
 469                                 /* skip last step if there is no last_element */
 470                                 if(last_element == NULL)
 471                                         continue;
 472
 473                                 /* check if proj has an if edge to last_element (at this time pbqp contains only if edges) */
 474                                 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) {
 475                                         allHaveIFEdges = false; /* there is no if edge between proj and last_element */
 476                                 }
 477                         }
 478
 479                         if(last_element != NULL && allHaveIFEdges) {
 480                                 if (get_free_regs(restr_nodes, cls, last_element) <= 4) {
 481                                         pqueue_put(restr_nodes_queue, last_element, pbqp_alloc_env->ife_edge_num[get_irn_idx(last_element)]);
 482                                 }
 483                                 else {
 484                                         pqueue_put(queue, last_element, pbqp_alloc_env->ife_edge_num[get_irn_idx(last_element)]);
 485                                 }
 486                                 plist_erase(temp_list, plist_find_value(temp_list, get_node(pbqp_inst, last_element->node_idx)));
 487                                 last_element = NULL;
 488                         }
 489
 490                         /* first insert all restricted proj nodes */
 491                         while (!pqueue_empty(restr_nodes_queue)) {
 492                                 ir_node *node = (ir_node*)pqueue_pop_front(restr_nodes_queue);
 493                                 plist_insert_front(sorted_list, get_node(pbqp_inst, get_irn_idx(node)));
 494                         }
 495
 496                         /* insert proj nodes descending by their number of interference edges */
 497                         while (!pqueue_empty(queue)) {
 498                                 ir_node *node = (ir_node*)pqueue_pop_front(queue);
 499                                 plist_insert_front(sorted_list, get_node(pbqp_inst, get_irn_idx(node)));
 500                         }
 501
 502                         /* invert sorted list */
 503                         foreach_plist(sorted_list, el) {
 504                                 plist_insert_front(temp_list, el->data);
 505                         }
 506
 507                         plist_clear(sorted_list);
 508
 509                 }
 510                 else {
 511                         if (arch_irn_consider_in_reg_alloc(cls, irn)) {
 512                                 // remember last colorable node
 513                                 last_element = irn;
 514                                 plist_insert_front(temp_list, get_node(pbqp_inst, get_irn_idx(irn)));
 515                         }
 516                         else {
 517                                 // node not colorable, so ignore it
 518                                 last_element = NULL;
 519                         }
 520                 }
 521 #endif
 522         }
 523
 524         /* add the temp rpeo list of this block to the global reverse perfect elimination order list*/
 525         foreach_plist(temp_list, el) {
 526                 plist_insert_back(rpeo, el->data);
 527         }
 528
 529         /* free reserved memory */
 530         ir_nodeset_destroy(&live_nodes);
 531         plist_free(temp_list);
 532 #if USE_BIPARTIT_MATCHING
 533 #else
 534         plist_free(sorted_list);
 535         del_pqueue(queue);
 536         del_pqueue(restr_nodes_queue);
 537 #endif
 538 }
 539
 540 static void insert_perms(ir_node *block, void *data)
 541 {
 542         be_chordal_env_t *env    = (be_chordal_env_t*)data;
 543
 544         sched_foreach_safe(block, irn) {
 545                 be_insn_t *insn = be_scan_insn(env, irn);
 546                 if (insn)
 547                         pre_process_constraints(env, &insn);
 548         }
 549 }
 550
 551 static void be_pbqp_coloring(be_chordal_env_t *env)
 552 {
 553         ir_graph                    *irg            = env->irg;
 554         const arch_register_class_t *cls            = env->cls;
 555         be_lv_t                     *lv             = NULL;
 556         plist_element_t             *element        = NULL;
 557         unsigned                     colors_n       = arch_register_class_n_regs(cls);
 558         be_pbqp_alloc_env_t          pbqp_alloc_env;
 559         unsigned                     col;
 560         unsigned                     row;
 561         pbqp_matrix_t               *ife_matrix;
 562         num                          solution;
 563 #if KAPS_DUMP
 564         FILE                        *file_before;
 565 #endif
 566 #if TIMER
 567         ir_timer_t *t_ra_pbqp_alloc_create     = ir_timer_new();
 568         ir_timer_t *t_ra_pbqp_alloc_solve      = ir_timer_new();
 569         ir_timer_t *t_ra_pbqp_alloc_create_aff = ir_timer_new();
 570
 571         printf("#### ----- === Allocating registers of %s (%s) ===\n", cls->name, get_entity_name(get_irg_entity(irg)));
 572 #endif
 573         be_assure_live_sets(irg);
 574         lv = be_get_irg_liveness(irg);
 575
 576         /* insert perms */
 577         assure_doms(irg);
 578         dom_tree_walk_irg(irg, insert_perms, NULL, env);
 579
 580         /* dump graph after inserting perms */
 581         if (env->opts->dump_flags & BE_CH_DUMP_CONSTR) {
 582                 char buf[256];
 583                 snprintf(buf, sizeof(buf), "%s-constr", cls->name);
 584                 dump_ir_graph(irg, buf);
 585         }
 586
 587         ir_calculate_execfreq_int_factors(&pbqp_alloc_env.execfreq_factors, irg);
 588
 589         /* initialize pbqp allocation data structure */
 590         pbqp_alloc_env.pbqp_inst        = alloc_pbqp(get_irg_last_idx(irg));  /* initialize pbqp instance */
 591         pbqp_alloc_env.cls              = cls;
 592         pbqp_alloc_env.irg              = irg;
 593         pbqp_alloc_env.lv               = lv;
 594         pbqp_alloc_env.allocatable_regs = bitset_malloc(colors_n);
 595         pbqp_alloc_env.rpeo             = plist_new();
 596         pbqp_alloc_env.restr_nodes      = XMALLOCNZ(unsigned, get_irg_last_idx(irg));
 597         pbqp_alloc_env.ife_edge_num     = XMALLOCNZ(unsigned, get_irg_last_idx(irg));
 598         pbqp_alloc_env.env              = env;
 599         be_put_allocatable_regs(irg, cls, pbqp_alloc_env.allocatable_regs);
 600
 601
 602         /* create costs matrix template for interference edges */
 603         ife_matrix = pbqp_matrix_alloc(pbqp_alloc_env.pbqp_inst, colors_n, colors_n);
 604         /* set costs */
 605         for (row = 0, col = 0; row < colors_n; row++, col++)
 606                 pbqp_matrix_set(ife_matrix, row, col, INF_COSTS);
 607
 608         pbqp_alloc_env.ife_matrix_template = ife_matrix;
 609
 610
 611         if (!use_exec_freq) {
 612                 /* create costs matrix template for affinity edges */
 613                 pbqp_matrix_t *afe_matrix = pbqp_matrix_alloc(pbqp_alloc_env.pbqp_inst, colors_n, colors_n);
 614                 /* set costs */
 615                 for (row = 0; row < colors_n; row++) {
 616                         for (col = 0; col < colors_n; col++) {
 617                                 if (row != col)
 618                                         pbqp_matrix_set(afe_matrix, row, col, 2);
 619                         }
 620                 }
 621                 pbqp_alloc_env.aff_matrix_template = afe_matrix;
 622         }
 623
 624
 625         /* create pbqp instance */
 626 #if TIMER
 627         ir_timer_reset_and_start(t_ra_pbqp_alloc_create);
 628 #endif
 629         assure_doms(irg);
 630         dom_tree_walk_irg(irg, create_pbqp_coloring_instance , NULL, &pbqp_alloc_env);
 631 #if TIMER
 632         ir_timer_stop(t_ra_pbqp_alloc_create);
 633 #endif
 634
 635
 636         /* set up affinity edges */
 637 #if TIMER
 638         ir_timer_reset_and_start(t_ra_pbqp_alloc_create_aff);
 639 #endif
 640         foreach_plist(pbqp_alloc_env.rpeo, element) {
 641                 pbqp_node_t *node = (pbqp_node_t*)element->data;
 642                 ir_node     *irn  = get_idx_irn(irg, node->index);
 643
 644                 create_affinity_edges(irn, &pbqp_alloc_env);
 645         }
 646 #if TIMER
 647         ir_timer_stop(t_ra_pbqp_alloc_create_aff);
 648 #endif
 649
 650
 651 #if KAPS_DUMP
 652         // dump graph before solving pbqp
 653         file_before = my_open(env, "", "-pbqp_coloring.html");
 654         set_dumpfile(pbqp_alloc_env.pbqp_inst, file_before);
 655 #endif
 656
 657         /* print out reverse perfect elimination order */
 658 #if PRINT_RPEO
 659         {
 660                 plist_element_t *elements;
 661                 foreach_plist(pbqp_alloc_env.rpeo, elements) {
 662                         pbqp_node_t *node = elements->data;
 663                         printf(" %d(%ld);", node->index, get_idx_irn(irg, node->index)->node_nr);
 664                 }
 665                 printf("\n");
 666         }
 667 #endif
 668
 669         /* solve pbqp instance */
 670 #if TIMER
 671         ir_timer_reset_and_start(t_ra_pbqp_alloc_solve);
 672 #endif
 673         if(use_late_decision) {
 674                 solve_pbqp_heuristical_co_ld(pbqp_alloc_env.pbqp_inst,pbqp_alloc_env.rpeo);
 675         }
 676         else {
 677                 solve_pbqp_heuristical_co(pbqp_alloc_env.pbqp_inst,pbqp_alloc_env.rpeo);
 678         }
 679 #if TIMER
 680         ir_timer_stop(t_ra_pbqp_alloc_solve);
 681 #endif
 682
 683
 684         solution = get_solution(pbqp_alloc_env.pbqp_inst);
 685         if (solution == INF_COSTS)
 686                 panic("No PBQP solution found");
 687
 688
 689         /* assign colors */
 690         foreach_plist(pbqp_alloc_env.rpeo, element) {
 691                 pbqp_node_t           *node  = (pbqp_node_t*)element->data;
 692                 ir_node               *irn   = get_idx_irn(irg, node->index);
 693                 num                    color = get_node_solution(pbqp_alloc_env.pbqp_inst, node->index);
 694                 const arch_register_t *reg   = arch_register_for_index(cls, color);
 695
 696                 arch_set_irn_register(irn, reg);
 697         }
 698
 699
 700 #if TIMER
 701         printf("PBQP alloc create:     %10.3lf msec\n",
 702                (double)ir_timer_elapsed_usec(t_ra_pbqp_alloc_create) / 1000.0);
 703         printf("PBQP alloc solve:      %10.3lf msec\n",
 704                (double)ir_timer_elapsed_usec(t_ra_pbqp_alloc_solve) / 1000.0);
 705         printf("PBQP alloc create aff: %10.3lf msec\n",
 706                (double)ir_timer_elapsed_usec(t_ra_pbqp_alloc_create_aff) / 1000.0);
 707 #endif
 708
 709
 710         /* free reserved memory */
 711 #if KAPS_DUMP
 712         fclose(file_before);
 713 #endif
 714         bitset_free(pbqp_alloc_env.allocatable_regs);
 715         free_pbqp(pbqp_alloc_env.pbqp_inst);
 716         plist_free(pbqp_alloc_env.rpeo);
 717         xfree(pbqp_alloc_env.restr_nodes);
 718         xfree(pbqp_alloc_env.ife_edge_num);
 719 }
 720
 721
 722 /**
 723  * Initializes this module.
 724  */
 725 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_pbqp_coloring)
 726 void be_init_pbqp_coloring(void)
 727 {
 728         lc_opt_entry_t *be_grp       = lc_opt_get_grp(firm_opt_get_root(), "be");
 729         lc_opt_entry_t *ra_grp       = lc_opt_get_grp(be_grp, "ra");
 730         lc_opt_entry_t *chordal_grp  = lc_opt_get_grp(ra_grp, "chordal");
 731         lc_opt_entry_t *coloring_grp = lc_opt_get_grp(chordal_grp, "coloring");
 732         lc_opt_entry_t *pbqp_grp     = lc_opt_get_grp(coloring_grp, "pbqp");
 733
 734         static be_ra_chordal_coloring_t coloring = {
 735                 be_pbqp_coloring
 736         };
 737
 738         lc_opt_add_table(pbqp_grp, options);
 739         be_register_chordal_coloring("pbqp", &coloring);
 740 }