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

   1 /*
   2  * Copyright (C) 1995-2011 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       Preference Guided Register Assignment
  23  * @author      Matthias Braun
  24  * @date        14.2.2009
  25  * @version     $Id$
  26  *
  27  * The idea is to allocate registers in 2 passes:
  28  * 1. A first pass to determine "preferred" registers for live-ranges. This
  29  *    calculates for each register and each live-range a value indicating
  30  *    the usefulness. (You can roughly think of the value as the negative
  31  *    costs needed for copies when the value is in the specific registers...)
  32  *
  33  * 2. Walk blocks and assigns registers in a greedy fashion. Preferring
  34  *    registers with high preferences. When register constraints are not met,
  35  *    add copies and split live-ranges.
  36  *
  37  * TODO:
  38  *  - make use of free registers in the permute_values code
  39  */
  40 #include "config.h"
  41
  42 #include <float.h>
  43 #include <stdbool.h>
  44 #include <math.h>
  45 #include "lpp.h"
  46
  47 #include "error.h"
  48 #include "execfreq.h"
  49 #include "ircons.h"
  50 #include "irdom.h"
  51 #include "iredges_t.h"
  52 #include "irgraph_t.h"
  53 #include "irgwalk.h"
  54 #include "irnode_t.h"
  55 #include "irprintf.h"
  56 #include "irdump.h"
  57 #include "obst.h"
  58 #include "raw_bitset.h"
  59 #include "unionfind.h"
  60 #include "pdeq.h"
  61 #include "hungarian.h"
  62
  63 #include "beabi.h"
  64 #include "bechordal_t.h"
  65 #include "be.h"
  66 #include "beirg.h"
  67 #include "belive_t.h"
  68 #include "bemodule.h"
  69 #include "benode.h"
  70 #include "bera.h"
  71 #include "besched.h"
  72 #include "bespill.h"
  73 #include "bespillutil.h"
  74 #include "beverify.h"
  75 #include "beutil.h"
  76 #include "bestack.h"
  77
  78 #define USE_FACTOR                     1.0f
  79 #define DEF_FACTOR                     1.0f
  80 #define NEIGHBOR_FACTOR                0.2f
  81 #define AFF_SHOULD_BE_SAME             0.5f
  82 #define AFF_PHI                        1.0f
  83 #define SPLIT_DELTA                    1.0f
  84 #define MAX_OPTIMISTIC_SPLIT_RECURSION 0
  85
  86 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  87
  88 static struct obstack               obst;
  89 static ir_graph                    *irg;
  90 static const arch_register_class_t *cls;
  91 static be_lv_t                     *lv;
  92 static const ir_exec_freq          *execfreqs;
  93 static unsigned                     n_regs;
  94 static unsigned                    *normal_regs;
  95 static int                         *congruence_classes;
  96 static ir_node                    **block_order;
  97 static size_t                       n_block_order;
  98 static int                          create_preferences        = true;
  99 static int                          create_congruence_classes = true;
 100 static int                          propagate_phi_registers   = true;
 101
 102 static const lc_opt_table_entry_t options[] = {
 103         LC_OPT_ENT_BOOL("prefs", "use preference based coloring", &create_preferences),
 104         LC_OPT_ENT_BOOL("congruences", "create congruence classes", &create_congruence_classes),
 105         LC_OPT_ENT_BOOL("prop_phi", "propagate phi registers", &propagate_phi_registers),
 106         LC_OPT_LAST
 107 };
 108
 109 /** currently active assignments (while processing a basic block)
 110  * maps registers to values(their current copies) */
 111 static ir_node **assignments;
 112
 113 /**
 114  * allocation information: last_uses, register preferences
 115  * the information is per firm-node.
 116  */
 117 struct allocation_info_t {
 118         unsigned  last_uses[2];   /**< bitset indicating last uses (input pos) */
 119         ir_node  *current_value;  /**< copy of the value that should be used */
 120         ir_node  *original_value; /**< for copies point to original value */
 121         float     prefs[0];       /**< register preferences */
 122 };
 123 typedef struct allocation_info_t allocation_info_t;
 124
 125 /** helper datastructure used when sorting register preferences */
 126 struct reg_pref_t {
 127         unsigned num;
 128         float    pref;
 129 };
 130 typedef struct reg_pref_t reg_pref_t;
 131
 132 /** per basic-block information */
 133 struct block_info_t {
 134         bool     processed;       /**< indicate whether block is processed */
 135         ir_node *assignments[0];  /**< register assignments at end of block */
 136 };
 137 typedef struct block_info_t block_info_t;
 138
 139 /**
 140  * Get the allocation info for a node.
 141  * The info is allocated on the first visit of a node.
 142  */
 143 static allocation_info_t *get_allocation_info(ir_node *node)
 144 {
 145         allocation_info_t *info = (allocation_info_t*)get_irn_link(node);
 146         if (info == NULL) {
 147                 info = OALLOCFZ(&obst, allocation_info_t, prefs, n_regs);
 148                 info->current_value  = node;
 149                 info->original_value = node;
 150                 set_irn_link(node, info);
 151         }
 152
 153         return info;
 154 }
 155
 156 static allocation_info_t *try_get_allocation_info(const ir_node *node)
 157 {
 158         return (allocation_info_t*) get_irn_link(node);
 159 }
 160
 161 /**
 162  * Get allocation information for a basic block
 163  */
 164 static block_info_t *get_block_info(ir_node *block)
 165 {
 166         block_info_t *info = (block_info_t*)get_irn_link(block);
 167
 168         assert(is_Block(block));
 169         if (info == NULL) {
 170                 info = OALLOCFZ(&obst, block_info_t, assignments, n_regs);
 171                 set_irn_link(block, info);
 172         }
 173
 174         return info;
 175 }
 176
 177 /**
 178  * Link the allocation info of a node to a copy.
 179  * Afterwards, both nodes uses the same allocation info.
 180  * Copy must not have an allocation info assigned yet.
 181  *
 182  * @param copy   the node that gets the allocation info assigned
 183  * @param value  the original node
 184  */
 185 static void mark_as_copy_of(ir_node *copy, ir_node *value)
 186 {
 187         ir_node           *original;
 188         allocation_info_t *info      = get_allocation_info(value);
 189         allocation_info_t *copy_info = get_allocation_info(copy);
 190
 191         /* find original value */
 192         original = info->original_value;
 193         if (original != value) {
 194                 info = get_allocation_info(original);
 195         }
 196
 197         assert(info->original_value == original);
 198         info->current_value = copy;
 199
 200         /* the copy should not be linked to something else yet */
 201         assert(copy_info->original_value == copy);
 202         copy_info->original_value = original;
 203
 204         /* copy over allocation preferences */
 205         memcpy(copy_info->prefs, info->prefs, n_regs * sizeof(copy_info->prefs[0]));
 206 }
 207
 208 /**
 209  * Calculate the penalties for every register on a node and its live neighbors.
 210  *
 211  * @param live_nodes  the set of live nodes at the current position, may be NULL
 212  * @param penalty     the penalty to subtract from
 213  * @param limited     a raw bitset containing the limited set for the node
 214  * @param node        the node
 215  */
 216 static void give_penalties_for_limits(const ir_nodeset_t *live_nodes,
 217                                       float penalty, const unsigned* limited,
 218                                       ir_node *node)
 219 {
 220         ir_nodeset_iterator_t iter;
 221         unsigned              r;
 222         size_t                n_allowed;
 223         allocation_info_t     *info = get_allocation_info(node);
 224         ir_node               *neighbor;
 225
 226         /* give penalty for all forbidden regs */
 227         for (r = 0; r < n_regs; ++r) {
 228                 if (rbitset_is_set(limited, r))
 229                         continue;
 230
 231                 info->prefs[r] -= penalty;
 232         }
 233
 234         /* all other live values should get a penalty for allowed regs */
 235         if (live_nodes == NULL)
 236                 return;
 237
 238         penalty   *= NEIGHBOR_FACTOR;
 239         n_allowed  = rbitset_popcount(limited, n_regs);
 240         if (n_allowed > 1) {
 241                 /* only create a very weak penalty if multiple regs are allowed */
 242                 penalty = (penalty * 0.8f) / n_allowed;
 243         }
 244         foreach_ir_nodeset(live_nodes, neighbor, iter) {
 245                 allocation_info_t *neighbor_info;
 246
 247                 /* TODO: if op is used on multiple inputs we might not do a
 248                  * continue here */
 249                 if (neighbor == node)
 250                         continue;
 251
 252                 neighbor_info = get_allocation_info(neighbor);
 253                 for (r = 0; r < n_regs; ++r) {
 254                         if (!rbitset_is_set(limited, r))
 255                                 continue;
 256
 257                         neighbor_info->prefs[r] -= penalty;
 258                 }
 259         }
 260 }
 261
 262 /**
 263  * Calculate the preferences of a definition for the current register class.
 264  * If the definition uses a limited set of registers, reduce the preferences
 265  * for the limited register on the node and its neighbors.
 266  *
 267  * @param live_nodes  the set of live nodes at the current node
 268  * @param weight      the weight
 269  * @param node        the current node
 270  */
 271 static void check_defs(const ir_nodeset_t *live_nodes, float weight,
 272                        ir_node *node)
 273 {
 274         const arch_register_req_t *req = arch_get_register_req_out(node);
 275         if (req->type & arch_register_req_type_limited) {
 276                 const unsigned *limited = req->limited;
 277                 float           penalty = weight * DEF_FACTOR;
 278                 give_penalties_for_limits(live_nodes, penalty, limited, node);
 279         }
 280
 281         if (req->type & arch_register_req_type_should_be_same) {
 282                 ir_node           *insn  = skip_Proj(node);
 283                 allocation_info_t *info  = get_allocation_info(node);
 284                 int                arity = get_irn_arity(insn);
 285                 int                i;
 286
 287                 float factor = 1.0f / rbitset_popcount(&req->other_same, arity);
 288                 for (i = 0; i < arity; ++i) {
 289                         ir_node           *op;
 290                         unsigned           r;
 291                         allocation_info_t *op_info;
 292
 293                         if (!rbitset_is_set(&req->other_same, i))
 294                                 continue;
 295
 296                         op = get_irn_n(insn, i);
 297
 298                         /* if we the value at the should_be_same input doesn't die at the
 299                          * node, then it is no use to propagate the constraints (since a
 300                          * copy will emerge anyway) */
 301                         if (ir_nodeset_contains(live_nodes, op))
 302                                 continue;
 303
 304                         op_info = get_allocation_info(op);
 305                         for (r = 0; r < n_regs; ++r) {
 306                                 op_info->prefs[r] += info->prefs[r] * factor;
 307                         }
 308                 }
 309         }
 310 }
 311
 312 /**
 313  * Walker: Runs an a block calculates the preferences for any
 314  * node and every register from the considered register class.
 315  */
 316 static void analyze_block(ir_node *block, void *data)
 317 {
 318         float         weight = (float)get_block_execfreq(execfreqs, block);
 319         ir_nodeset_t  live_nodes;
 320         ir_node      *node;
 321         (void) data;
 322
 323         ir_nodeset_init(&live_nodes);
 324         be_liveness_end_of_block(lv, cls, block, &live_nodes);
 325
 326         sched_foreach_reverse(block, node) {
 327                 allocation_info_t *info;
 328                 int                i;
 329                 int                arity;
 330
 331                 if (is_Phi(node))
 332                         break;
 333
 334                 if (create_preferences) {
 335                         ir_node *value;
 336                         be_foreach_definition(node, cls, value,
 337                                 check_defs(&live_nodes, weight, value);
 338                         );
 339                 }
 340
 341                 /* mark last uses */
 342                 arity = get_irn_arity(node);
 343
 344                 /* the allocation info node currently only uses 1 unsigned value
 345                    to mark last used inputs. So we will fail for a node with more than
 346                    32 inputs. */
 347                 if (arity >= (int) sizeof(info->last_uses) * 8) {
 348                         panic("Node with more than %d inputs not supported yet",
 349                                         (int) sizeof(info->last_uses) * 8);
 350                 }
 351
 352                 info = get_allocation_info(node);
 353                 for (i = 0; i < arity; ++i) {
 354                         ir_node                   *op  = get_irn_n(node, i);
 355                         const arch_register_req_t *req = arch_get_register_req_out(op);
 356                         if (req->cls != cls)
 357                                 continue;
 358
 359                         /* last usage of a value? */
 360                         if (!ir_nodeset_contains(&live_nodes, op)) {
 361                                 rbitset_set(info->last_uses, i);
 362                         }
 363                 }
 364
 365                 be_liveness_transfer(cls, node, &live_nodes);
 366
 367                 if (create_preferences) {
 368                         /* update weights based on usage constraints */
 369                         for (i = 0; i < arity; ++i) {
 370                                 const arch_register_req_t *req;
 371                                 const unsigned            *limited;
 372                                 ir_node                   *op = get_irn_n(node, i);
 373
 374                                 if (!arch_irn_consider_in_reg_alloc(cls, op))
 375                                         continue;
 376
 377                                 req = arch_get_register_req(node, i);
 378                                 if (!(req->type & arch_register_req_type_limited))
 379                                         continue;
 380
 381                                 limited = req->limited;
 382                                 give_penalties_for_limits(&live_nodes, weight * USE_FACTOR,
 383                                                           limited, op);
 384                         }
 385                 }
 386         }
 387
 388         ir_nodeset_destroy(&live_nodes);
 389 }
 390
 391 static void congruence_def(ir_nodeset_t *live_nodes, const ir_node *node)
 392 {
 393         const arch_register_req_t *req = arch_get_register_req_out(node);
 394
 395         /* should be same constraint? */
 396         if (req->type & arch_register_req_type_should_be_same) {
 397                 const ir_node *insn  = skip_Proj_const(node);
 398                 int      arity = get_irn_arity(insn);
 399                 int      i;
 400                 unsigned node_idx = get_irn_idx(node);
 401                 node_idx          = uf_find(congruence_classes, node_idx);
 402
 403                 for (i = 0; i < arity; ++i) {
 404                         ir_node               *live;
 405                         ir_node               *op;
 406                         int                    op_idx;
 407                         ir_nodeset_iterator_t  iter;
 408                         bool                   interferes = false;
 409
 410                         if (!rbitset_is_set(&req->other_same, i))
 411                                 continue;
 412
 413                         op     = get_irn_n(insn, i);
 414                         op_idx = get_irn_idx(op);
 415                         op_idx = uf_find(congruence_classes, op_idx);
 416
 417                         /* do we interfere with the value */
 418                         foreach_ir_nodeset(live_nodes, live, iter) {
 419                                 int lv_idx = get_irn_idx(live);
 420                                 lv_idx     = uf_find(congruence_classes, lv_idx);
 421                                 if (lv_idx == op_idx) {
 422                                         interferes = true;
 423                                         break;
 424                                 }
 425                         }
 426                         /* don't put in same affinity class if we interfere */
 427                         if (interferes)
 428                                 continue;
 429
 430                         node_idx = uf_union(congruence_classes, node_idx, op_idx);
 431                         DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
 432                             node, op));
 433                         /* one should_be_same is enough... */
 434                         break;
 435                 }
 436         }
 437 }
 438
 439 static void create_congruence_class(ir_node *block, void *data)
 440 {
 441         ir_nodeset_t  live_nodes;
 442         ir_node      *node;
 443
 444         (void) data;
 445         ir_nodeset_init(&live_nodes);
 446         be_liveness_end_of_block(lv, cls, block, &live_nodes);
 447
 448         /* check should be same constraints */
 449         sched_foreach_reverse(block, node) {
 450                 ir_node *value;
 451                 if (is_Phi(node))
 452                         break;
 453
 454                 be_foreach_definition(node, cls, value,
 455                         congruence_def(&live_nodes, value);
 456                 );
 457                 be_liveness_transfer(cls, node, &live_nodes);
 458         }
 459
 460         /* check phi congruence classes */
 461         sched_foreach_reverse_from(node, node) {
 462                 int i;
 463                 int arity;
 464                 int node_idx;
 465                 assert(is_Phi(node));
 466
 467                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 468                         continue;
 469
 470                 node_idx = get_irn_idx(node);
 471                 node_idx = uf_find(congruence_classes, node_idx);
 472
 473                 arity = get_irn_arity(node);
 474                 for (i = 0; i < arity; ++i) {
 475                         bool                  interferes = false;
 476                         ir_nodeset_iterator_t iter;
 477                         unsigned              r;
 478                         int                   old_node_idx;
 479                         ir_node              *live;
 480                         ir_node              *phi;
 481                         allocation_info_t    *head_info;
 482                         allocation_info_t    *other_info;
 483                         ir_node              *op     = get_Phi_pred(node, i);
 484                         int                   op_idx = get_irn_idx(op);
 485                         op_idx = uf_find(congruence_classes, op_idx);
 486
 487                         /* do we interfere with the value */
 488                         foreach_ir_nodeset(&live_nodes, live, iter) {
 489                                 int lv_idx = get_irn_idx(live);
 490                                 lv_idx     = uf_find(congruence_classes, lv_idx);
 491                                 if (lv_idx == op_idx) {
 492                                         interferes = true;
 493                                         break;
 494                                 }
 495                         }
 496                         /* don't put in same affinity class if we interfere */
 497                         if (interferes)
 498                                 continue;
 499                         /* any other phi has the same input? */
 500                         sched_foreach(block, phi) {
 501                                 ir_node *oop;
 502                                 int      oop_idx;
 503                                 if (!is_Phi(phi))
 504                                         break;
 505                                 if (!arch_irn_consider_in_reg_alloc(cls, phi))
 506                                         continue;
 507                                 oop = get_Phi_pred(phi, i);
 508                                 if (oop == op)
 509                                         continue;
 510                                 oop_idx = get_irn_idx(oop);
 511                                 oop_idx = uf_find(congruence_classes, oop_idx);
 512                                 if (oop_idx == op_idx) {
 513                                         interferes = true;
 514                                         break;
 515                                 }
 516                         }
 517                         if (interferes)
 518                                 continue;
 519
 520                         /* merge the 2 congruence classes and sum up their preferences */
 521                         old_node_idx = node_idx;
 522                         node_idx = uf_union(congruence_classes, node_idx, op_idx);
 523                         DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
 524                             node, op));
 525
 526                         old_node_idx = node_idx == old_node_idx ? op_idx : old_node_idx;
 527                         head_info  = get_allocation_info(get_idx_irn(irg, node_idx));
 528                         other_info = get_allocation_info(get_idx_irn(irg, old_node_idx));
 529                         for (r = 0; r < n_regs; ++r) {
 530                                 head_info->prefs[r] += other_info->prefs[r];
 531                         }
 532                 }
 533         }
 534 }
 535
 536 static void set_congruence_prefs(ir_node *node, void *data)
 537 {
 538         allocation_info_t *info;
 539         allocation_info_t *head_info;
 540         unsigned node_idx = get_irn_idx(node);
 541         unsigned node_set = uf_find(congruence_classes, node_idx);
 542
 543         (void) data;
 544
 545         /* head of congruence class or not in any class */
 546         if (node_set == node_idx)
 547                 return;
 548
 549         if (!arch_irn_consider_in_reg_alloc(cls, node))
 550                 return;
 551
 552         head_info = get_allocation_info(get_idx_irn(irg, node_set));
 553         info      = get_allocation_info(node);
 554
 555         memcpy(info->prefs, head_info->prefs, n_regs * sizeof(info->prefs[0]));
 556 }
 557
 558 static void combine_congruence_classes(void)
 559 {
 560         size_t n = get_irg_last_idx(irg);
 561         congruence_classes = XMALLOCN(int, n);
 562         uf_init(congruence_classes, n);
 563
 564         /* create congruence classes */
 565         irg_block_walk_graph(irg, create_congruence_class, NULL, NULL);
 566         /* merge preferences */
 567         irg_walk_graph(irg, set_congruence_prefs, NULL, NULL);
 568         free(congruence_classes);
 569 }
 570
 571
 572
 573 /**
 574  * Assign register reg to the given node.
 575  *
 576  * @param node  the node
 577  * @param reg   the register
 578  */
 579 static void use_reg(ir_node *node, const arch_register_t *reg)
 580 {
 581         unsigned r = arch_register_get_index(reg);
 582         assignments[r] = node;
 583         arch_set_irn_register(node, reg);
 584 }
 585
 586 static void free_reg_of_value(ir_node *node)
 587 {
 588         const arch_register_t *reg;
 589         unsigned               r;
 590
 591         if (!arch_irn_consider_in_reg_alloc(cls, node))
 592                 return;
 593
 594         reg        = arch_get_irn_register(node);
 595         r          = arch_register_get_index(reg);
 596         /* assignment->value may be NULL if a value is used at 2 inputs
 597            so it gets freed twice. */
 598         assert(assignments[r] == node || assignments[r] == NULL);
 599         assignments[r] = NULL;
 600 }
 601
 602 /**
 603  * Compare two register preferences in decreasing order.
 604  */
 605 static int compare_reg_pref(const void *e1, const void *e2)
 606 {
 607         const reg_pref_t *rp1 = (const reg_pref_t*) e1;
 608         const reg_pref_t *rp2 = (const reg_pref_t*) e2;
 609         if (rp1->pref < rp2->pref)
 610                 return 1;
 611         if (rp1->pref > rp2->pref)
 612                 return -1;
 613         return 0;
 614 }
 615
 616 static void fill_sort_candidates(reg_pref_t *regprefs,
 617                                  const allocation_info_t *info)
 618 {
 619         unsigned r;
 620
 621         for (r = 0; r < n_regs; ++r) {
 622                 float pref = info->prefs[r];
 623                 regprefs[r].num  = r;
 624                 regprefs[r].pref = pref;
 625         }
 626         /* TODO: use a stable sort here to avoid unnecessary register jumping */
 627         qsort(regprefs, n_regs, sizeof(regprefs[0]), compare_reg_pref);
 628 }
 629
 630 static bool try_optimistic_split(ir_node *to_split, ir_node *before,
 631                                  float pref, float pref_delta,
 632                                  unsigned *forbidden_regs, int recursion)
 633 {
 634         const arch_register_t *from_reg;
 635         const arch_register_t *reg;
 636         ir_node               *original_insn;
 637         ir_node               *block;
 638         ir_node               *copy;
 639         unsigned               r = 0;
 640         unsigned               from_r;
 641         unsigned               i;
 642         allocation_info_t     *info = get_allocation_info(to_split);
 643         reg_pref_t            *prefs;
 644         float                  delta = 0;
 645         float                  split_threshold;
 646
 647         (void) pref;
 648
 649         /* stupid hack: don't optimisticallt split don't spill nodes...
 650          * (so we don't split away the values produced because of
 651          *  must_be_different constraints) */
 652         original_insn = skip_Proj(info->original_value);
 653         if (arch_irn_get_flags(original_insn) & arch_irn_flags_dont_spill)
 654                 return false;
 655
 656         from_reg        = arch_get_irn_register(to_split);
 657         from_r          = arch_register_get_index(from_reg);
 658         block           = get_nodes_block(before);
 659         split_threshold = (float)get_block_execfreq(execfreqs, block) * SPLIT_DELTA;
 660
 661         if (pref_delta < split_threshold*0.5)
 662                 return false;
 663
 664         /* find the best free position where we could move to */
 665         prefs = ALLOCAN(reg_pref_t, n_regs);
 666         fill_sort_candidates(prefs, info);
 667         for (i = 0; i < n_regs; ++i) {
 668                 float apref;
 669                 float apref_delta;
 670                 bool  res;
 671                 bool  old_source_state;
 672
 673                 /* we need a normal register which is not an output register
 674                    an different from the current register of to_split */
 675                 r = prefs[i].num;
 676                 if (!rbitset_is_set(normal_regs, r))
 677                         continue;
 678                 if (rbitset_is_set(forbidden_regs, r))
 679                         continue;
 680                 if (r == from_r)
 681                         continue;
 682
 683                 /* is the split worth it? */
 684                 delta = pref_delta + prefs[i].pref;
 685                 if (delta < split_threshold) {
 686                         DB((dbg, LEVEL_3, "Not doing optimistical split of %+F (depth %d), win %f too low\n",
 687                                 to_split, recursion, delta));
 688                         return false;
 689                 }
 690
 691                 /* if the register is free then we can do the split */
 692                 if (assignments[r] == NULL)
 693                         break;
 694
 695                 /* otherwise we might try recursively calling optimistic_split */
 696                 if (recursion+1 > MAX_OPTIMISTIC_SPLIT_RECURSION)
 697                         continue;
 698
 699                 apref        = prefs[i].pref;
 700                 apref_delta  = i+1 < n_regs ? apref - prefs[i+1].pref : 0;
 701                 apref_delta += pref_delta - split_threshold;
 702
 703                 /* our source register isn't a useful destination for recursive
 704                    splits */
 705                 old_source_state = rbitset_is_set(forbidden_regs, from_r);
 706                 rbitset_set(forbidden_regs, from_r);
 707                 /* try recursive split */
 708                 res = try_optimistic_split(assignments[r], before, apref,
 709                                            apref_delta, forbidden_regs, recursion+1);
 710                 /* restore our destination */
 711                 if (old_source_state) {
 712                         rbitset_set(forbidden_regs, from_r);
 713                 } else {
 714                         rbitset_clear(forbidden_regs, from_r);
 715                 }
 716
 717                 if (res)
 718                         break;
 719         }
 720         if (i >= n_regs)
 721                 return false;
 722
 723         reg  = arch_register_for_index(cls, r);
 724         copy = be_new_Copy(cls, block, to_split);
 725         mark_as_copy_of(copy, to_split);
 726         /* hacky, but correct here */
 727         if (assignments[arch_register_get_index(from_reg)] == to_split)
 728                 free_reg_of_value(to_split);
 729         use_reg(copy, reg);
 730         sched_add_before(before, copy);
 731
 732         DB((dbg, LEVEL_3,
 733             "Optimistic live-range split %+F move %+F(%s) -> %s before %+F (win %f, depth %d)\n",
 734             copy, to_split, from_reg->name, reg->name, before, delta, recursion));
 735         return true;
 736 }
 737
 738 /**
 739  * Determine and assign a register for node @p node
 740  */
 741 static void assign_reg(const ir_node *block, ir_node *node,
 742                        unsigned *forbidden_regs)
 743 {
 744         const arch_register_t     *final_reg;
 745         allocation_info_t         *info;
 746         const arch_register_req_t *req;
 747         reg_pref_t                *reg_prefs;
 748         ir_node                   *in_node;
 749         unsigned                   r;
 750         const unsigned            *allowed_regs;
 751         unsigned                   final_reg_index = 0;
 752
 753         assert(!is_Phi(node));
 754         /* preassigned register? */
 755         final_reg = arch_get_irn_register(node);
 756         if (final_reg != NULL) {
 757                 DB((dbg, LEVEL_2, "Preassignment %+F -> %s\n", node, final_reg->name));
 758                 use_reg(node, final_reg);
 759                 return;
 760         }
 761
 762         req = arch_get_register_req_out(node);
 763         /* ignore reqs must be preassigned */
 764         assert (! (req->type & arch_register_req_type_ignore));
 765
 766         /* give should_be_same boni */
 767         info    = get_allocation_info(node);
 768         in_node = skip_Proj(node);
 769         if (req->type & arch_register_req_type_should_be_same) {
 770                 float weight = (float)get_block_execfreq(execfreqs, block);
 771                 int   arity  = get_irn_arity(in_node);
 772                 int   i;
 773
 774                 assert(arity <= (int) sizeof(req->other_same) * 8);
 775                 for (i = 0; i < arity; ++i) {
 776                         ir_node               *in;
 777                         const arch_register_t *reg;
 778                         unsigned               reg_index;
 779                         if (!rbitset_is_set(&req->other_same, i))
 780                                 continue;
 781
 782                         in  = get_irn_n(in_node, i);
 783                         reg = arch_get_irn_register(in);
 784                         assert(reg != NULL);
 785                         reg_index = arch_register_get_index(reg);
 786
 787                         /* if the value didn't die here then we should not propagate the
 788                          * should_be_same info */
 789                         if (assignments[reg_index] == in)
 790                                 continue;
 791
 792                         info->prefs[reg_index] += weight * AFF_SHOULD_BE_SAME;
 793                 }
 794         }
 795
 796         /* create list of register candidates and sort by their preference */
 797         DB((dbg, LEVEL_2, "Candidates for %+F:", node));
 798         reg_prefs = ALLOCAN(reg_pref_t, n_regs);
 799         fill_sort_candidates(reg_prefs, info);
 800         for (r = 0; r < n_regs; ++r) {
 801                 unsigned num = reg_prefs[r].num;
 802                 const arch_register_t *reg;
 803
 804                 if (!rbitset_is_set(normal_regs, num))
 805                         continue;
 806                 reg = arch_register_for_index(cls, num);
 807                 DB((dbg, LEVEL_2, " %s(%f)", reg->name, reg_prefs[r].pref));
 808         }
 809         DB((dbg, LEVEL_2, "\n"));
 810
 811         allowed_regs = normal_regs;
 812         if (req->type & arch_register_req_type_limited) {
 813                 allowed_regs = req->limited;
 814         }
 815
 816         for (r = 0; r < n_regs; ++r) {
 817                 float   pref, delta;
 818                 ir_node *before;
 819                 bool    res;
 820
 821                 final_reg_index = reg_prefs[r].num;
 822                 if (!rbitset_is_set(allowed_regs, final_reg_index))
 823                         continue;
 824                 /* alignment constraint? */
 825                 if (req->width > 1 && (req->type & arch_register_req_type_aligned)
 826                                 && (final_reg_index % req->width) != 0)
 827                         continue;
 828
 829                 if (assignments[final_reg_index] == NULL)
 830                         break;
 831                 pref   = reg_prefs[r].pref;
 832                 delta  = r+1 < n_regs ? pref - reg_prefs[r+1].pref : 0;
 833                 before = skip_Proj(node);
 834                 res    = try_optimistic_split(assignments[final_reg_index], before,
 835                                               pref, delta, forbidden_regs, 0);
 836                 if (res)
 837                         break;
 838         }
 839         if (r >= n_regs) {
 840                 /* the common reason to hit this panic is when 1 of your nodes is not
 841                  * register pressure faithful */
 842                 panic("No register left for %+F\n", node);
 843         }
 844
 845         final_reg = arch_register_for_index(cls, final_reg_index);
 846         DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, final_reg->name));
 847         use_reg(node, final_reg);
 848 }
 849
 850 /**
 851  * Add an permutation in front of a node and change the assignments
 852  * due to this permutation.
 853  *
 854  * To understand this imagine a permutation like this:
 855  *
 856  * 1 -> 2
 857  * 2 -> 3
 858  * 3 -> 1, 5
 859  * 4 -> 6
 860  * 5
 861  * 6
 862  * 7 -> 7
 863  *
 864  * First we count how many destinations a single value has. At the same time
 865  * we can be sure that each destination register has at most 1 source register
 866  * (it can have 0 which means we don't care what value is in it).
 867  * We ignore all fulfilled permuations (like 7->7)
 868  * In a first pass we create as much copy instructions as possible as they
 869  * are generally cheaper than exchanges. We do this by counting into how many
 870  * destinations a register has to be copied (in the example it's 2 for register
 871  * 3, or 1 for the registers 1,2,4 and 7).
 872  * We can then create a copy into every destination register when the usecount
 873  * of that register is 0 (= noone else needs the value in the register).
 874  *
 875  * After this step we should only have cycles left. We implement a cyclic
 876  * permutation of n registers with n-1 transpositions.
 877  *
 878  * @param live_nodes   the set of live nodes, updated due to live range split
 879  * @param before       the node before we add the permutation
 880  * @param permutation  the permutation array indices are the destination
 881  *                     registers, the values in the array are the source
 882  *                     registers.
 883  */
 884 static void permute_values(ir_nodeset_t *live_nodes, ir_node *before,
 885                              unsigned *permutation)
 886 {
 887         unsigned  *n_used = ALLOCANZ(unsigned, n_regs);
 888         ir_node   *block;
 889         unsigned   r;
 890
 891         /* determine how often each source register needs to be read */
 892         for (r = 0; r < n_regs; ++r) {
 893                 unsigned  old_reg = permutation[r];
 894                 ir_node  *value;
 895
 896                 value = assignments[old_reg];
 897                 if (value == NULL) {
 898                         /* nothing to do here, reg is not live. Mark it as fixpoint
 899                          * so we ignore it in the next steps */
 900                         permutation[r] = r;
 901                         continue;
 902                 }
 903
 904                 ++n_used[old_reg];
 905         }
 906
 907         block = get_nodes_block(before);
 908
 909         /* step1: create copies where immediately possible */
 910         for (r = 0; r < n_regs; /* empty */) {
 911                 ir_node *copy;
 912                 ir_node *src;
 913                 const arch_register_t *reg;
 914                 unsigned               old_r = permutation[r];
 915
 916                 /* - no need to do anything for fixed points.
 917                    - we can't copy if the value in the dest reg is still needed */
 918                 if (old_r == r || n_used[r] > 0) {
 919                         ++r;
 920                         continue;
 921                 }
 922
 923                 /* create a copy */
 924                 src  = assignments[old_r];
 925                 copy = be_new_Copy(cls, block, src);
 926                 sched_add_before(before, copy);
 927                 reg = arch_register_for_index(cls, r);
 928                 DB((dbg, LEVEL_2, "Copy %+F (from %+F, before %+F) -> %s\n",
 929                     copy, src, before, reg->name));
 930                 mark_as_copy_of(copy, src);
 931                 use_reg(copy, reg);
 932
 933                 if (live_nodes != NULL) {
 934                         ir_nodeset_insert(live_nodes, copy);
 935                 }
 936
 937                 /* old register has 1 user less, permutation is resolved */
 938                 assert(arch_register_get_index(arch_get_irn_register(src)) == old_r);
 939                 permutation[r] = r;
 940
 941                 assert(n_used[old_r] > 0);
 942                 --n_used[old_r];
 943                 if (n_used[old_r] == 0) {
 944                         if (live_nodes != NULL) {
 945                                 ir_nodeset_remove(live_nodes, src);
 946                         }
 947                         free_reg_of_value(src);
 948                 }
 949
 950                 /* advance or jump back (if this copy enabled another copy) */
 951                 if (old_r < r && n_used[old_r] == 0) {
 952                         r = old_r;
 953                 } else {
 954                         ++r;
 955                 }
 956         }
 957
 958         /* at this point we only have "cycles" left which we have to resolve with
 959          * perm instructions
 960          * TODO: if we have free registers left, then we should really use copy
 961          * instructions for any cycle longer than 2 registers...
 962          * (this is probably architecture dependent, there might be archs where
 963          *  copies are preferable even for 2-cycles) */
 964
 965         /* create perms with the rest */
 966         for (r = 0; r < n_regs; /* empty */) {
 967                 const arch_register_t *reg;
 968                 unsigned  old_r = permutation[r];
 969                 unsigned  r2;
 970                 ir_node  *in[2];
 971                 ir_node  *perm;
 972                 ir_node  *proj0;
 973                 ir_node  *proj1;
 974
 975                 if (old_r == r) {
 976                         ++r;
 977                         continue;
 978                 }
 979
 980                 /* we shouldn't have copies from 1 value to multiple destinations left*/
 981                 assert(n_used[old_r] == 1);
 982
 983                 /* exchange old_r and r2; after that old_r is a fixed point */
 984                 r2 = permutation[old_r];
 985
 986                 in[0] = assignments[r2];
 987                 in[1] = assignments[old_r];
 988                 perm = be_new_Perm(cls, block, 2, in);
 989                 sched_add_before(before, perm);
 990                 DB((dbg, LEVEL_2, "Perm %+F (perm %+F,%+F, before %+F)\n",
 991                     perm, in[0], in[1], before));
 992
 993                 proj0 = new_r_Proj(perm, get_irn_mode(in[0]), 0);
 994                 mark_as_copy_of(proj0, in[0]);
 995                 reg = arch_register_for_index(cls, old_r);
 996                 use_reg(proj0, reg);
 997
 998                 proj1 = new_r_Proj(perm, get_irn_mode(in[1]), 1);
 999                 mark_as_copy_of(proj1, in[1]);
1000                 reg = arch_register_for_index(cls, r2);
1001                 use_reg(proj1, reg);
1002
1003                 /* 1 value is now in the correct register */
1004                 permutation[old_r] = old_r;
1005                 /* the source of r changed to r2 */
1006                 permutation[r] = r2;
1007
1008                 /* if we have reached a fixpoint update data structures */
1009                 if (live_nodes != NULL) {
1010                         ir_nodeset_remove(live_nodes, in[0]);
1011                         ir_nodeset_remove(live_nodes, in[1]);
1012                         ir_nodeset_remove(live_nodes, proj0);
1013                         ir_nodeset_insert(live_nodes, proj1);
1014                 }
1015         }
1016
1017 #ifdef DEBUG_libfirm
1018         /* now we should only have fixpoints left */
1019         for (r = 0; r < n_regs; ++r) {
1020                 assert(permutation[r] == r);
1021         }
1022 #endif
1023 }
1024
1025 /**
1026  * Free regs for values last used.
1027  *
1028  * @param live_nodes   set of live nodes, will be updated
1029  * @param node         the node to consider
1030  */
1031 static void free_last_uses(ir_nodeset_t *live_nodes, ir_node *node)
1032 {
1033         allocation_info_t *info      = get_allocation_info(node);
1034         const unsigned    *last_uses = info->last_uses;
1035         int                arity     = get_irn_arity(node);
1036         int                i;
1037
1038         for (i = 0; i < arity; ++i) {
1039                 ir_node *op;
1040
1041                 /* check if one operand is the last use */
1042                 if (!rbitset_is_set(last_uses, i))
1043                         continue;
1044
1045                 op = get_irn_n(node, i);
1046                 free_reg_of_value(op);
1047                 ir_nodeset_remove(live_nodes, op);
1048         }
1049 }
1050
1051 /**
1052  * change inputs of a node to the current value (copies/perms)
1053  */
1054 static void rewire_inputs(ir_node *node)
1055 {
1056         int i;
1057         int arity = get_irn_arity(node);
1058
1059         for (i = 0; i < arity; ++i) {
1060                 ir_node           *op = get_irn_n(node, i);
1061                 allocation_info_t *info = try_get_allocation_info(op);
1062
1063                 if (info == NULL)
1064                         continue;
1065
1066                 info = get_allocation_info(info->original_value);
1067                 if (info->current_value != op) {
1068                         set_irn_n(node, i, info->current_value);
1069                 }
1070         }
1071 }
1072
1073 /**
1074  * Create a bitset of registers occupied with value living through an
1075  * instruction
1076  */
1077 static void determine_live_through_regs(unsigned *bitset, ir_node *node)
1078 {
1079         const allocation_info_t *info = get_allocation_info(node);
1080         unsigned r;
1081         int i;
1082         int arity;
1083
1084         /* mark all used registers as potentially live-through */
1085         for (r = 0; r < n_regs; ++r) {
1086                 if (assignments[r] == NULL)
1087                         continue;
1088                 if (!rbitset_is_set(normal_regs, r))
1089                         continue;
1090
1091                 rbitset_set(bitset, r);
1092         }
1093
1094         /* remove registers of value dying at the instruction */
1095         arity = get_irn_arity(node);
1096         for (i = 0; i < arity; ++i) {
1097                 ir_node               *op;
1098                 const arch_register_t *reg;
1099
1100                 if (!rbitset_is_set(info->last_uses, i))
1101                         continue;
1102
1103                 op  = get_irn_n(node, i);
1104                 reg = arch_get_irn_register(op);
1105                 rbitset_clear(bitset, arch_register_get_index(reg));
1106         }
1107 }
1108
1109 static void solve_lpp(ir_nodeset_t *live_nodes, ir_node *node,
1110                       unsigned *forbidden_regs, unsigned *live_through_regs)
1111 {
1112         unsigned *forbidden_edges = rbitset_malloc(n_regs * n_regs);
1113         int      *lpp_vars        = XMALLOCNZ(int, n_regs*n_regs);
1114         int       arity           = get_irn_arity(node);
1115         int       i;
1116         unsigned  l;
1117         unsigned  r;
1118
1119         lpp_t *lpp = lpp_new("prefalloc", lpp_minimize);
1120         //lpp_set_time_limit(lpp, 20);
1121         lpp_set_log(lpp, stdout);
1122
1123         /** mark some edges as forbidden */
1124         for (i = 0; i < arity; ++i) {
1125                 ir_node                   *op = get_irn_n(node, i);
1126                 const arch_register_t     *reg;
1127                 const arch_register_req_t *req;
1128                 const unsigned            *limited;
1129                 unsigned                   current_reg;
1130
1131                 if (!arch_irn_consider_in_reg_alloc(cls, op))
1132                         continue;
1133
1134                 req = arch_get_register_req(node, i);
1135                 if (!(req->type & arch_register_req_type_limited))
1136                         continue;
1137
1138                 limited     = req->limited;
1139                 reg         = arch_get_irn_register(op);
1140                 current_reg = arch_register_get_index(reg);
1141                 for (r = 0; r < n_regs; ++r) {
1142                         if (rbitset_is_set(limited, r))
1143                                 continue;
1144
1145                         rbitset_set(forbidden_edges, current_reg*n_regs + r);
1146                 }
1147         }
1148
1149         /* add all combinations, then remove not allowed ones */
1150         for (l = 0; l < n_regs; ++l) {
1151                 for (r = 0; r < n_regs; ++r) {
1152                         if (!rbitset_is_set(normal_regs, r))
1153                                 continue;
1154                         if (rbitset_is_set(forbidden_edges, l*n_regs + r))
1155                                 continue;
1156                         /* livethrough values may not use constrained output registers */
1157                         if (rbitset_is_set(live_through_regs, l)
1158                             && rbitset_is_set(forbidden_regs, r))
1159                                 continue;
1160
1161                         double costs = l==r ? 9 : 8;
1162                         lpp_vars[l*n_regs+r]
1163                                 = lpp_add_var(lpp, NULL, lpp_binary, costs);
1164                         assert(lpp_vars[l*n_regs+r] > 0);
1165                 }
1166         }
1167         /* add constraints */
1168         for (l = 0; l < n_regs; ++l) {
1169                 /* only 1 destination per register */
1170                 int constraint = lpp_add_cst(lpp, NULL, lpp_equal, 1);
1171                 for (r = 0; r < n_regs; ++r) {
1172                         int var = lpp_vars[l*n_regs+r];
1173                         if (var == 0)
1174                                 continue;
1175                         lpp_set_factor_fast(lpp, constraint, var, 1);
1176                 }
1177                 /* each destination used by at most 1 value */
1178                 constraint = lpp_add_cst(lpp, NULL, lpp_less_equal, 1);
1179         }
1180
1181         /* solve lpp */
1182         {
1183                 ir_graph     *irg     = get_irn_irg(node);
1184                 be_options_t *options = be_get_irg_options(irg);
1185                 unsigned     *assignment;
1186                 lpp_solve(lpp, options->ilp_server, options->ilp_solver);
1187                 if (!lpp_is_sol_valid(lpp))
1188                         panic("ilp solution not valid!");
1189
1190                 assignment = ALLOCAN(unsigned, n_regs);
1191                 for (l = 0; l < n_regs; ++l) {
1192                         unsigned dest_reg = (unsigned)-1;
1193                         for (r = 0; r < n_regs; ++r) {
1194                                 int var = lpp_vars[l*n_regs+r];
1195                                 if (var == 0)
1196                                         continue;
1197                                 double val = lpp_get_var_sol(lpp, var);
1198                                 if (val == 1) {
1199                                         assert(dest_reg == (unsigned)-1);
1200                                         dest_reg = r;
1201                                 }
1202                         }
1203                         assert(dest_reg != (unsigned)-1);
1204                         assignment[l] = dest_reg;
1205                 }
1206                 permute_values(live_nodes, node, assignment);
1207         }
1208         lpp_free(lpp);
1209 }
1210
1211 /**
1212  * Enforce constraints at a node by live range splits.
1213  *
1214  * @param  live_nodes  the set of live nodes, might be changed
1215  * @param  node        the current node
1216  */
1217 static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node,
1218                                 unsigned *forbidden_regs)
1219 {
1220         int arity = get_irn_arity(node);
1221         int i, res;
1222         hungarian_problem_t *bp;
1223         unsigned l, r;
1224         unsigned *assignment;
1225         ir_node  *value;
1226
1227         /* construct a list of register occupied by live-through values */
1228         unsigned *live_through_regs = NULL;
1229
1230         /* see if any use constraints are not met */
1231         bool double_width = false;
1232         bool good = true;
1233         for (i = 0; i < arity; ++i) {
1234                 ir_node                   *op = get_irn_n(node, i);
1235                 const arch_register_t     *reg;
1236                 const arch_register_req_t *req;
1237                 const unsigned            *limited;
1238                 unsigned                   reg_index;
1239
1240                 if (!arch_irn_consider_in_reg_alloc(cls, op))
1241                         continue;
1242
1243                 /* are there any limitations for the i'th operand? */
1244                 req = arch_get_register_req(node, i);
1245                 if (req->width > 1)
1246                         double_width = true;
1247                 if (!(req->type & arch_register_req_type_limited))
1248                         continue;
1249
1250                 limited   = req->limited;
1251                 reg       = arch_get_irn_register(op);
1252                 reg_index = arch_register_get_index(reg);
1253                 if (!rbitset_is_set(limited, reg_index)) {
1254                         /* found an assignment outside the limited set */
1255                         good = false;
1256                         break;
1257                 }
1258         }
1259
1260         /* is any of the live-throughs using a constrained output register? */
1261         be_foreach_definition(node, cls, value,
1262                 if (req_->width > 1)
1263                         double_width = true;
1264                 if (! (req_->type & arch_register_req_type_limited))
1265                         continue;
1266                 if (live_through_regs == NULL) {
1267                         rbitset_alloca(live_through_regs, n_regs);
1268                         determine_live_through_regs(live_through_regs, node);
1269                 }
1270                 rbitset_or(forbidden_regs, req_->limited, n_regs);
1271                 if (rbitsets_have_common(req_->limited, live_through_regs, n_regs))
1272                         good = false;
1273         );
1274
1275         if (good)
1276                 return;
1277
1278         /* create these arrays if we haven't yet */
1279         if (live_through_regs == NULL) {
1280                 rbitset_alloca(live_through_regs, n_regs);
1281         }
1282
1283         if (double_width) {
1284                 solve_lpp(live_nodes, node, forbidden_regs, live_through_regs);
1285                 return;
1286         }
1287
1288         /* at this point we have to construct a bipartite matching problem to see
1289          * which values should go to which registers
1290          * Note: We're building the matrix in "reverse" - source registers are
1291          *       right, destinations left because this will produce the solution
1292          *       in the format required for permute_values.
1293          */
1294         bp = hungarian_new(n_regs, n_regs, HUNGARIAN_MATCH_PERFECT);
1295
1296         /* add all combinations, then remove not allowed ones */
1297         for (l = 0; l < n_regs; ++l) {
1298                 if (!rbitset_is_set(normal_regs, l)) {
1299                         hungarian_add(bp, l, l, 1);
1300                         continue;
1301                 }
1302
1303                 for (r = 0; r < n_regs; ++r) {
1304                         if (!rbitset_is_set(normal_regs, r))
1305                                 continue;
1306                         /* livethrough values may not use constrainted output registers */
1307                         if (rbitset_is_set(live_through_regs, l)
1308                                         && rbitset_is_set(forbidden_regs, r))
1309                                 continue;
1310
1311                         hungarian_add(bp, r, l, l == r ? 9 : 8);
1312                 }
1313         }
1314
1315         for (i = 0; i < arity; ++i) {
1316                 ir_node                   *op = get_irn_n(node, i);
1317                 const arch_register_t     *reg;
1318                 const arch_register_req_t *req;
1319                 const unsigned            *limited;
1320                 unsigned                   current_reg;
1321
1322                 if (!arch_irn_consider_in_reg_alloc(cls, op))
1323                         continue;
1324
1325                 req = arch_get_register_req(node, i);
1326                 if (!(req->type & arch_register_req_type_limited))
1327                         continue;
1328
1329                 limited     = req->limited;
1330                 reg         = arch_get_irn_register(op);
1331                 current_reg = arch_register_get_index(reg);
1332                 for (r = 0; r < n_regs; ++r) {
1333                         if (rbitset_is_set(limited, r))
1334                                 continue;
1335                         hungarian_remove(bp, r, current_reg);
1336                 }
1337         }
1338
1339         //hungarian_print_cost_matrix(bp, 1);
1340         hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1341
1342         assignment = ALLOCAN(unsigned, n_regs);
1343         res = hungarian_solve(bp, assignment, NULL, 0);
1344         assert(res == 0);
1345
1346 #if 0
1347         fprintf(stderr, "Swap result:");
1348         for (i = 0; i < (int) n_regs; ++i) {
1349                 fprintf(stderr, " %d", assignment[i]);
1350         }
1351         fprintf(stderr, "\n");
1352 #endif
1353
1354         hungarian_free(bp);
1355
1356         permute_values(live_nodes, node, assignment);
1357 }
1358
1359 /** test whether a node @p n is a copy of the value of node @p of */
1360 static bool is_copy_of(ir_node *value, ir_node *test_value)
1361 {
1362         allocation_info_t *test_info;
1363         allocation_info_t *info;
1364
1365         if (value == test_value)
1366                 return true;
1367
1368         info      = get_allocation_info(value);
1369         test_info = get_allocation_info(test_value);
1370         return test_info->original_value == info->original_value;
1371 }
1372
1373 /**
1374  * find a value in the end-assignment of a basic block
1375  * @returns the index into the assignment array if found
1376  *          -1 if not found
1377  */
1378 static int find_value_in_block_info(block_info_t *info, ir_node *value)
1379 {
1380         unsigned   r;
1381         ir_node  **end_assignments = info->assignments;
1382         for (r = 0; r < n_regs; ++r) {
1383                 ir_node *a_value = end_assignments[r];
1384
1385                 if (a_value == NULL)
1386                         continue;
1387                 if (is_copy_of(a_value, value))
1388                         return (int) r;
1389         }
1390
1391         return -1;
1392 }
1393
1394 /**
1395  * Create the necessary permutations at the end of a basic block to fullfill
1396  * the register assignment for phi-nodes in the next block
1397  */
1398 static void add_phi_permutations(ir_node *block, int p)
1399 {
1400         unsigned   r;
1401         unsigned  *permutation;
1402         ir_node  **old_assignments;
1403         bool       need_permutation;
1404         ir_node   *phi;
1405         ir_node   *pred = get_Block_cfgpred_block(block, p);
1406
1407         block_info_t *pred_info = get_block_info(pred);
1408
1409         /* predecessor not processed yet? nothing to do */
1410         if (!pred_info->processed)
1411                 return;
1412
1413         permutation = ALLOCAN(unsigned, n_regs);
1414         for (r = 0; r < n_regs; ++r) {
1415                 permutation[r] = r;
1416         }
1417
1418         /* check phi nodes */
1419         need_permutation = false;
1420         phi = sched_first(block);
1421         for ( ; is_Phi(phi); phi = sched_next(phi)) {
1422                 const arch_register_t *reg;
1423                 const arch_register_t *op_reg;
1424                 int                    regn;
1425                 int                    a;
1426                 ir_node               *op;
1427
1428                 if (!arch_irn_consider_in_reg_alloc(cls, phi))
1429                         continue;
1430
1431                 op = get_Phi_pred(phi, p);
1432                 a  = find_value_in_block_info(pred_info, op);
1433                 assert(a >= 0);
1434
1435                 reg  = arch_get_irn_register(phi);
1436                 regn = arch_register_get_index(reg);
1437                 /* same register? nothing to do */
1438                 if (regn == a)
1439                         continue;
1440
1441                 op     = pred_info->assignments[a];
1442                 op_reg = arch_get_irn_register(op);
1443                 /* virtual or joker registers are ok too */
1444                 if ((op_reg->type & arch_register_type_joker)
1445                                 || (op_reg->type & arch_register_type_virtual))
1446                         continue;
1447
1448                 permutation[regn] = a;
1449                 need_permutation  = true;
1450         }
1451
1452         if (need_permutation) {
1453                 /* permute values at end of predecessor */
1454                 old_assignments = assignments;
1455                 assignments     = pred_info->assignments;
1456                 permute_values(NULL, be_get_end_of_block_insertion_point(pred),
1457                                permutation);
1458                 assignments = old_assignments;
1459         }
1460
1461         /* change phi nodes to use the copied values */
1462         phi = sched_first(block);
1463         for ( ; is_Phi(phi); phi = sched_next(phi)) {
1464                 int      a;
1465                 ir_node *op;
1466
1467                 if (!arch_irn_consider_in_reg_alloc(cls, phi))
1468                         continue;
1469
1470                 op = get_Phi_pred(phi, p);
1471
1472                 /* we have permuted all values into the correct registers so we can
1473                    simply query which value occupies the phis register in the
1474                    predecessor */
1475                 a  = arch_register_get_index(arch_get_irn_register(phi));
1476                 op = pred_info->assignments[a];
1477                 set_Phi_pred(phi, p, op);
1478         }
1479 }
1480
1481 /**
1482  * Set preferences for a phis register based on the registers used on the
1483  * phi inputs.
1484  */
1485 static void adapt_phi_prefs(ir_node *phi)
1486 {
1487         int i;
1488         int arity = get_irn_arity(phi);
1489         ir_node           *block = get_nodes_block(phi);
1490         allocation_info_t *info  = get_allocation_info(phi);
1491
1492         for (i = 0; i < arity; ++i) {
1493                 ir_node               *op  = get_irn_n(phi, i);
1494                 const arch_register_t *reg = arch_get_irn_register(op);
1495                 ir_node               *pred_block;
1496                 block_info_t          *pred_block_info;
1497                 float                  weight;
1498                 unsigned               r;
1499
1500                 if (reg == NULL)
1501                         continue;
1502                 /* we only give the bonus if the predecessor already has registers
1503                  * assigned, otherwise we only see a dummy value
1504                  * and any conclusions about its register are useless */
1505                 pred_block = get_Block_cfgpred_block(block, i);
1506                 pred_block_info = get_block_info(pred_block);
1507                 if (!pred_block_info->processed)
1508                         continue;
1509
1510                 /* give bonus for already assigned register */
1511                 weight = (float)get_block_execfreq(execfreqs, pred_block);
1512                 r      = arch_register_get_index(reg);
1513                 info->prefs[r] += weight * AFF_PHI;
1514         }
1515 }
1516
1517 /**
1518  * After a phi has been assigned a register propagate preference inputs
1519  * to the phi inputs.
1520  */
1521 static void propagate_phi_register(ir_node *phi, unsigned assigned_r)
1522 {
1523         int      i;
1524         ir_node *block = get_nodes_block(phi);
1525         int      arity = get_irn_arity(phi);
1526
1527         for (i = 0; i < arity; ++i) {
1528                 ir_node           *op         = get_Phi_pred(phi, i);
1529                 allocation_info_t *info       = get_allocation_info(op);
1530                 ir_node           *pred_block = get_Block_cfgpred_block(block, i);
1531                 unsigned           r;
1532                 float              weight
1533                         = (float)get_block_execfreq(execfreqs, pred_block) * AFF_PHI;
1534
1535                 if (info->prefs[assigned_r] >= weight)
1536                         continue;
1537
1538                 /* promote the prefered register */
1539                 for (r = 0; r < n_regs; ++r) {
1540                         if (info->prefs[r] > -weight) {
1541                                 info->prefs[r] = -weight;
1542                         }
1543                 }
1544                 info->prefs[assigned_r] = weight;
1545
1546                 if (is_Phi(op))
1547                         propagate_phi_register(op, assigned_r);
1548         }
1549 }
1550
1551 static void assign_phi_registers(ir_node *block)
1552 {
1553         int                  n_phis = 0;
1554         int                  n;
1555         int                  res;
1556         unsigned            *assignment;
1557         ir_node             *node;
1558         hungarian_problem_t *bp;
1559
1560         /* count phi nodes */
1561         sched_foreach(block, node) {
1562                 if (!is_Phi(node))
1563                         break;
1564                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1565                         continue;
1566                 ++n_phis;
1567         }
1568
1569         if (n_phis == 0)
1570                 return;
1571
1572         /* build a bipartite matching problem for all phi nodes */
1573         bp = hungarian_new(n_phis, n_regs, HUNGARIAN_MATCH_PERFECT);
1574         n  = 0;
1575         sched_foreach(block, node) {
1576                 unsigned r;
1577
1578                 allocation_info_t *info;
1579                 if (!is_Phi(node))
1580                         break;
1581                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1582                         continue;
1583
1584                 /* give boni for predecessor colorings */
1585                 adapt_phi_prefs(node);
1586                 /* add stuff to bipartite problem */
1587                 info = get_allocation_info(node);
1588                 DB((dbg, LEVEL_3, "Prefs for %+F: ", node));
1589                 for (r = 0; r < n_regs; ++r) {
1590                         float costs;
1591
1592                         if (!rbitset_is_set(normal_regs, r))
1593                                 continue;
1594
1595                         costs = info->prefs[r];
1596                         costs = costs < 0 ? -logf(-costs+1) : logf(costs+1);
1597                         costs *= 100;
1598                         costs += 10000;
1599                         hungarian_add(bp, n, r, (int)costs);
1600                         DB((dbg, LEVEL_3, " %s(%f)", arch_register_for_index(cls, r)->name,
1601                                                 info->prefs[r]));
1602                 }
1603                 DB((dbg, LEVEL_3, "\n"));
1604                 ++n;
1605         }
1606
1607         //hungarian_print_cost_matrix(bp, 7);
1608         hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1609
1610         assignment = ALLOCAN(unsigned, n_regs);
1611         res        = hungarian_solve(bp, assignment, NULL, 0);
1612         assert(res == 0);
1613
1614         /* apply results */
1615         n = 0;
1616         sched_foreach(block, node) {
1617                 unsigned               r;
1618                 const arch_register_t *reg;
1619
1620                 if (!is_Phi(node))
1621                         break;
1622                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1623                         continue;
1624
1625                 r = assignment[n++];
1626                 assert(rbitset_is_set(normal_regs, r));
1627                 reg = arch_register_for_index(cls, r);
1628                 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
1629                 use_reg(node, reg);
1630
1631                 /* adapt preferences for phi inputs */
1632                 if (propagate_phi_registers)
1633                         propagate_phi_register(node, r);
1634         }
1635 }
1636
1637 /**
1638  * Walker: assign registers to all nodes of a block that
1639  * need registers from the currently considered register class.
1640  */
1641 static void allocate_coalesce_block(ir_node *block, void *data)
1642 {
1643         int            i;
1644         ir_nodeset_t   live_nodes;
1645         ir_node       *node;
1646         int            n_preds;
1647         block_info_t  *block_info;
1648         block_info_t **pred_block_infos;
1649         ir_node      **phi_ins;
1650         unsigned      *forbidden_regs; /**< collects registers which must
1651                                                 not be used for optimistic splits */
1652
1653         (void) data;
1654         DB((dbg, LEVEL_2, "* Block %+F\n", block));
1655
1656         /* clear assignments */
1657         block_info  = get_block_info(block);
1658         assignments = block_info->assignments;
1659
1660         ir_nodeset_init(&live_nodes);
1661
1662         /* gather regalloc infos of predecessor blocks */
1663         n_preds          = get_Block_n_cfgpreds(block);
1664         pred_block_infos = ALLOCAN(block_info_t*, n_preds);
1665         for (i = 0; i < n_preds; ++i) {
1666                 ir_node      *pred      = get_Block_cfgpred_block(block, i);
1667                 block_info_t *pred_info = get_block_info(pred);
1668                 pred_block_infos[i]     = pred_info;
1669         }
1670
1671         phi_ins = ALLOCAN(ir_node*, n_preds);
1672
1673         /* collect live-in nodes and preassigned values */
1674         be_lv_foreach(lv, block, be_lv_state_in, i) {
1675                 bool                       need_phi = false;
1676                 const arch_register_req_t *req;
1677                 const arch_register_t     *reg;
1678                 int                        p;
1679
1680                 node = be_lv_get_irn(lv, block, i);
1681                 req  = arch_get_register_req_out(node);
1682                 if (req->cls != cls)
1683                         continue;
1684
1685                 if (req->type & arch_register_req_type_ignore) {
1686                         allocation_info_t *info = get_allocation_info(node);
1687                         info->current_value = node;
1688
1689                         reg = arch_get_irn_register(node);
1690                         assert(reg != NULL); /* ignore values must be preassigned */
1691                         use_reg(node, reg);
1692                         continue;
1693                 }
1694
1695                 /* check all predecessors for this value, if it is not everywhere the
1696                    same or unknown then we have to construct a phi
1697                    (we collect the potential phi inputs here) */
1698                 for (p = 0; p < n_preds; ++p) {
1699                         block_info_t *pred_info = pred_block_infos[p];
1700
1701                         if (!pred_info->processed) {
1702                                 /* use node for now, it will get fixed later */
1703                                 phi_ins[p] = node;
1704                                 need_phi   = true;
1705                         } else {
1706                                 int a = find_value_in_block_info(pred_info, node);
1707
1708                                 /* must live out of predecessor */
1709                                 assert(a >= 0);
1710                                 phi_ins[p] = pred_info->assignments[a];
1711                                 /* different value from last time? then we need a phi */
1712                                 if (p > 0 && phi_ins[p-1] != phi_ins[p]) {
1713                                         need_phi = true;
1714                                 }
1715                         }
1716                 }
1717
1718                 if (need_phi) {
1719                         ir_mode *mode = get_irn_mode(node);
1720                         ir_node *phi  = be_new_Phi(block, n_preds, phi_ins, mode, cls);
1721
1722                         DB((dbg, LEVEL_3, "Create Phi %+F (for %+F) -", phi, node));
1723 #ifdef DEBUG_libfirm
1724                         {
1725                                 int pi;
1726                                 for (pi = 0; pi < n_preds; ++pi) {
1727                                         DB((dbg, LEVEL_3, " %+F", phi_ins[pi]));
1728                                 }
1729                                 DB((dbg, LEVEL_3, "\n"));
1730                         }
1731 #endif
1732                         mark_as_copy_of(phi, node);
1733                         sched_add_after(block, phi);
1734
1735                         node = phi;
1736                 } else {
1737                         allocation_info_t *info = get_allocation_info(node);
1738                         info->current_value = phi_ins[0];
1739
1740                         /* Grab 1 of the inputs we constructed (might not be the same as
1741                          * "node" as we could see the same copy of the value in all
1742                          * predecessors */
1743                         node = phi_ins[0];
1744                 }
1745
1746                 /* if the node already has a register assigned use it */
1747                 reg = arch_get_irn_register(node);
1748                 if (reg != NULL) {
1749                         use_reg(node, reg);
1750                 }
1751
1752                 /* remember that this node is live at the beginning of the block */
1753                 ir_nodeset_insert(&live_nodes, node);
1754         }
1755
1756         rbitset_alloca(forbidden_regs, n_regs);
1757
1758         /* handle phis... */
1759         assign_phi_registers(block);
1760
1761         /* all live-ins must have a register */
1762 #ifdef DEBUG_libfirm
1763         {
1764                 ir_nodeset_iterator_t  iter;
1765                 foreach_ir_nodeset(&live_nodes, node, iter) {
1766                         const arch_register_t *reg = arch_get_irn_register(node);
1767                         assert(reg != NULL);
1768                 }
1769         }
1770 #endif
1771
1772         /* assign instructions in the block */
1773         sched_foreach(block, node) {
1774                 int arity;
1775                 ir_node *value;
1776
1777                 /* phis are already assigned */
1778                 if (is_Phi(node))
1779                         continue;
1780
1781                 rewire_inputs(node);
1782
1783                 /* enforce use constraints */
1784                 rbitset_clear_all(forbidden_regs, n_regs);
1785                 enforce_constraints(&live_nodes, node, forbidden_regs);
1786
1787                 rewire_inputs(node);
1788
1789                 /* we may not use registers used for inputs for optimistic splits */
1790                 arity = get_irn_arity(node);
1791                 for (i = 0; i < arity; ++i) {
1792                         ir_node *op = get_irn_n(node, i);
1793                         const arch_register_t *reg;
1794                         if (!arch_irn_consider_in_reg_alloc(cls, op))
1795                                 continue;
1796
1797                         reg = arch_get_irn_register(op);
1798                         rbitset_set(forbidden_regs, arch_register_get_index(reg));
1799                 }
1800
1801                 /* free registers of values last used at this instruction */
1802                 free_last_uses(&live_nodes, node);
1803
1804                 /* assign output registers */
1805                 be_foreach_definition_(node, cls, value,
1806                         assign_reg(block, value, forbidden_regs);
1807                 );
1808         }
1809
1810         ir_nodeset_destroy(&live_nodes);
1811         assignments = NULL;
1812
1813         block_info->processed = true;
1814
1815         /* permute values at end of predecessor blocks in case of phi-nodes */
1816         if (n_preds > 1) {
1817                 int p;
1818                 for (p = 0; p < n_preds; ++p) {
1819                         add_phi_permutations(block, p);
1820                 }
1821         }
1822
1823         /* if we have exactly 1 successor then we might be able to produce phi
1824            copies now */
1825         if (get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) == 1) {
1826                 const ir_edge_t *edge
1827                         = get_irn_out_edge_first_kind(block, EDGE_KIND_BLOCK);
1828                 ir_node      *succ      = get_edge_src_irn(edge);
1829                 int           p         = get_edge_src_pos(edge);
1830                 block_info_t *succ_info = get_block_info(succ);
1831
1832                 if (succ_info->processed) {
1833                         add_phi_permutations(succ, p);
1834                 }
1835         }
1836 }
1837
1838 typedef struct block_costs_t block_costs_t;
1839 struct block_costs_t {
1840         float costs;   /**< costs of the block */
1841         int   dfs_num; /**< depth first search number (to detect backedges) */
1842 };
1843
1844 static int cmp_block_costs(const void *d1, const void *d2)
1845 {
1846         const ir_node       * const *block1 = (const ir_node**)d1;
1847         const ir_node       * const *block2 = (const ir_node**)d2;
1848         const block_costs_t *info1  = (const block_costs_t*)get_irn_link(*block1);
1849         const block_costs_t *info2  = (const block_costs_t*)get_irn_link(*block2);
1850         return QSORT_CMP(info2->costs, info1->costs);
1851 }
1852
1853 static void determine_block_order(void)
1854 {
1855         size_t    p;
1856         ir_node **blocklist = be_get_cfgpostorder(irg);
1857         size_t    n_blocks  = ARR_LEN(blocklist);
1858         int       dfs_num   = 0;
1859         pdeq     *worklist  = new_pdeq();
1860         ir_node **order     = XMALLOCN(ir_node*, n_blocks);
1861         size_t    order_p   = 0;
1862
1863         /* clear block links... */
1864         for (p = 0; p < n_blocks; ++p) {
1865                 ir_node *block = blocklist[p];
1866                 set_irn_link(block, NULL);
1867         }
1868
1869         /* walk blocks in reverse postorder, the costs for each block are the
1870          * sum of the costs of its predecessors (excluding the costs on backedges
1871          * which we can't determine) */
1872         for (p = n_blocks; p > 0;) {
1873                 block_costs_t *cost_info;
1874                 ir_node *block = blocklist[--p];
1875
1876                 float execfreq   = (float)get_block_execfreq(execfreqs, block);
1877                 float costs      = execfreq;
1878                 int   n_cfgpreds = get_Block_n_cfgpreds(block);
1879                 int   p2;
1880                 for (p2 = 0; p2 < n_cfgpreds; ++p2) {
1881                         ir_node       *pred_block = get_Block_cfgpred_block(block, p2);
1882                         block_costs_t *pred_costs = (block_costs_t*)get_irn_link(pred_block);
1883                         /* we don't have any info for backedges */
1884                         if (pred_costs == NULL)
1885                                 continue;
1886                         costs += pred_costs->costs;
1887                 }
1888
1889                 cost_info          = OALLOCZ(&obst, block_costs_t);
1890                 cost_info->costs   = costs;
1891                 cost_info->dfs_num = dfs_num++;
1892                 set_irn_link(block, cost_info);
1893         }
1894
1895         /* sort array by block costs */
1896         qsort(blocklist, n_blocks, sizeof(blocklist[0]), cmp_block_costs);
1897
1898         ir_reserve_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1899         inc_irg_block_visited(irg);
1900
1901         for (p = 0; p < n_blocks; ++p) {
1902                 ir_node       *block = blocklist[p];
1903                 if (Block_block_visited(block))
1904                         continue;
1905
1906                 /* continually add predecessors with highest costs to worklist
1907                  * (without using backedges) */
1908                 do {
1909                         block_costs_t *info       = (block_costs_t*)get_irn_link(block);
1910                         ir_node       *best_pred  = NULL;
1911                         float          best_costs = -1;
1912                         int            n_cfgpred  = get_Block_n_cfgpreds(block);
1913                         int            i;
1914
1915                         pdeq_putr(worklist, block);
1916                         mark_Block_block_visited(block);
1917                         for (i = 0; i < n_cfgpred; ++i) {
1918                                 ir_node       *pred_block = get_Block_cfgpred_block(block, i);
1919                                 block_costs_t *pred_info  = (block_costs_t*)get_irn_link(pred_block);
1920
1921                                 /* ignore backedges */
1922                                 if (pred_info->dfs_num > info->dfs_num)
1923                                         continue;
1924
1925                                 if (info->costs > best_costs) {
1926                                         best_costs = info->costs;
1927                                         best_pred  = pred_block;
1928                                 }
1929                         }
1930                         block = best_pred;
1931                 } while (block != NULL && !Block_block_visited(block));
1932
1933                 /* now put all nodes in the worklist in our final order */
1934                 while (!pdeq_empty(worklist)) {
1935                         ir_node *pblock = (ir_node*)pdeq_getr(worklist);
1936                         assert(order_p < n_blocks);
1937                         order[order_p++] = pblock;
1938                 }
1939         }
1940         assert(order_p == n_blocks);
1941         del_pdeq(worklist);
1942
1943         ir_free_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1944
1945         DEL_ARR_F(blocklist);
1946
1947         obstack_free(&obst, NULL);
1948         obstack_init(&obst);
1949
1950         block_order   = order;
1951         n_block_order = n_blocks;
1952 }
1953
1954 /**
1955  * Run the register allocator for the current register class.
1956  */
1957 static void be_pref_alloc_cls(void)
1958 {
1959         size_t i;
1960
1961         lv = be_assure_liveness(irg);
1962         be_liveness_assure_sets(lv);
1963
1964         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
1965
1966         DB((dbg, LEVEL_2, "=== Allocating registers of %s ===\n", cls->name));
1967
1968         be_clear_links(irg);
1969
1970         irg_block_walk_graph(irg, NULL, analyze_block, NULL);
1971         if (create_congruence_classes)
1972                 combine_congruence_classes();
1973
1974         for (i = 0; i < n_block_order; ++i) {
1975                 ir_node *block = block_order[i];
1976                 allocate_coalesce_block(block, NULL);
1977         }
1978
1979         ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
1980 }
1981
1982 static void dump(int mask, ir_graph *irg, const char *suffix)
1983 {
1984         if (be_get_irg_options(irg)->dump_flags & mask)
1985                 dump_ir_graph(irg, suffix);
1986 }
1987
1988 /**
1989  * Run the spiller on the current graph.
1990  */
1991 static void spill(void)
1992 {
1993         /* make sure all nodes show their real register pressure */
1994         be_timer_push(T_RA_CONSTR);
1995         be_pre_spill_prepare_constr(irg, cls);
1996         be_timer_pop(T_RA_CONSTR);
1997
1998         dump(DUMP_RA, irg, "-spillprepare");
1999
2000         /* spill */
2001         be_timer_push(T_RA_SPILL);
2002         be_do_spill(irg, cls);
2003         be_timer_pop(T_RA_SPILL);
2004
2005         be_timer_push(T_RA_SPILL_APPLY);
2006         check_for_memory_operands(irg);
2007         be_timer_pop(T_RA_SPILL_APPLY);
2008
2009         dump(DUMP_RA, irg, "-spill");
2010 }
2011
2012 /**
2013  * The pref register allocator for a whole procedure.
2014  */
2015 static void be_pref_alloc(ir_graph *new_irg)
2016 {
2017         const arch_env_t *arch_env = be_get_irg_arch_env(new_irg);
2018         int   n_cls                = arch_env->n_register_classes;
2019         int   c;
2020
2021         obstack_init(&obst);
2022
2023         irg       = new_irg;
2024         execfreqs = be_get_irg_exec_freq(irg);
2025
2026         /* determine a good coloring order */
2027         determine_block_order();
2028
2029         for (c = 0; c < n_cls; ++c) {
2030                 cls = &arch_env->register_classes[c];
2031                 if (arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
2032                         continue;
2033
2034                 stat_ev_ctx_push_str("regcls", cls->name);
2035
2036                 n_regs      = arch_register_class_n_regs(cls);
2037                 normal_regs = rbitset_malloc(n_regs);
2038                 be_set_allocatable_regs(irg, cls, normal_regs);
2039
2040                 spill();
2041
2042                 /* verify schedule and register pressure */
2043                 be_timer_push(T_VERIFY);
2044                 if (be_get_irg_options(irg)->verify_option == BE_VERIFY_WARN) {
2045                         be_verify_schedule(irg);
2046                         be_verify_register_pressure(irg, cls);
2047                 } else if (be_get_irg_options(irg)->verify_option == BE_VERIFY_ASSERT) {
2048                         assert(be_verify_schedule(irg) && "Schedule verification failed");
2049                         assert(be_verify_register_pressure(irg, cls)
2050                                 && "Register pressure verification failed");
2051                 }
2052                 be_timer_pop(T_VERIFY);
2053
2054                 be_timer_push(T_RA_COLOR);
2055                 be_pref_alloc_cls();
2056                 be_timer_pop(T_RA_COLOR);
2057
2058                 /* we most probably constructed new Phis so liveness info is invalid
2059                  * now */
2060                 /* TODO: test liveness_introduce */
2061                 be_liveness_invalidate(lv);
2062                 free(normal_regs);
2063
2064                 stat_ev_ctx_pop("regcls");
2065         }
2066
2067         be_timer_push(T_RA_SPILL_APPLY);
2068         be_abi_fix_stack_nodes(irg);
2069         be_timer_pop(T_RA_SPILL_APPLY);
2070
2071         be_timer_push(T_VERIFY);
2072         if (be_get_irg_options(irg)->verify_option == BE_VERIFY_WARN) {
2073                 be_verify_register_allocation(irg);
2074         } else if (be_get_irg_options(irg)->verify_option == BE_VERIFY_ASSERT) {
2075                 assert(be_verify_register_allocation(irg)
2076                        && "Register allocation invalid");
2077         }
2078         be_timer_pop(T_VERIFY);
2079
2080         obstack_free(&obst, NULL);
2081 }
2082
2083 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_pref_alloc)
2084 void be_init_pref_alloc(void)
2085 {
2086         static be_ra_t be_ra_pref = {
2087                 be_pref_alloc,
2088         };
2089         lc_opt_entry_t *be_grp              = lc_opt_get_grp(firm_opt_get_root(), "be");
2090         lc_opt_entry_t *prefalloc_group = lc_opt_get_grp(be_grp, "prefalloc");
2091         lc_opt_add_table(prefalloc_group, options);
2092
2093         be_register_allocator("pref", &be_ra_pref);
2094         FIRM_DBG_REGISTER(dbg, "firm.be.prefalloc");
2095 }