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  *
  26  * The idea is to allocate registers in 2 passes:
  27  * 1. A first pass to determine "preferred" registers for live-ranges. This
  28  *    calculates for each register and each live-range a value indicating
  29  *    the usefulness. (You can roughly think of the value as the negative
  30  *    costs needed for copies when the value is in the specific registers...)
  31  *
  32  * 2. Walk blocks and assigns registers in a greedy fashion. Preferring
  33  *    registers with high preferences. When register constraints are not met,
  34  *    add copies and split live-ranges.
  35  *
  36  * TODO:
  37  *  - make use of free registers in the permute_values code
  38  */
  39 #include "config.h"
  40
  41 #include <float.h>
  42 #include <stdbool.h>
  43 #include <math.h>
  44 #include "lpp.h"
  45
  46 #include "error.h"
  47 #include "execfreq.h"
  48 #include "ircons.h"
  49 #include "irdom.h"
  50 #include "iredges_t.h"
  51 #include "irgraph_t.h"
  52 #include "irgwalk.h"
  53 #include "irnode_t.h"
  54 #include "irprintf.h"
  55 #include "irdump.h"
  56 #include "irtools.h"
  57 #include "util.h"
  58 #include "obst.h"
  59 #include "raw_bitset.h"
  60 #include "unionfind.h"
  61 #include "pdeq.h"
  62 #include "hungarian.h"
  63
  64 #include "beabi.h"
  65 #include "bechordal_t.h"
  66 #include "be.h"
  67 #include "beirg.h"
  68 #include "belive_t.h"
  69 #include "bemodule.h"
  70 #include "benode.h"
  71 #include "bera.h"
  72 #include "besched.h"
  73 #include "bespill.h"
  74 #include "bespillutil.h"
  75 #include "beverify.h"
  76 #include "beutil.h"
  77 #include "bestack.h"
  78
  79 #define USE_FACTOR                     1.0f
  80 #define DEF_FACTOR                     1.0f
  81 #define NEIGHBOR_FACTOR                0.2f
  82 #define AFF_SHOULD_BE_SAME             0.5f
  83 #define AFF_PHI                        1.0f
  84 #define SPLIT_DELTA                    1.0f
  85 #define MAX_OPTIMISTIC_SPLIT_RECURSION 0
  86
  87 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  88
  89 static struct obstack               obst;
  90 static ir_graph                    *irg;
  91 static const arch_register_class_t *cls;
  92 static be_lv_t                     *lv;
  93 static const ir_exec_freq          *execfreqs;
  94 static unsigned                     n_regs;
  95 static unsigned                    *normal_regs;
  96 static int                         *congruence_classes;
  97 static ir_node                    **block_order;
  98 static size_t                       n_block_order;
  99 static int                          create_preferences        = true;
 100 static int                          create_congruence_classes = true;
 101 static int                          propagate_phi_registers   = true;
 102
 103 static const lc_opt_table_entry_t options[] = {
 104         LC_OPT_ENT_BOOL("prefs", "use preference based coloring", &create_preferences),
 105         LC_OPT_ENT_BOOL("congruences", "create congruence classes", &create_congruence_classes),
 106         LC_OPT_ENT_BOOL("prop_phi", "propagate phi registers", &propagate_phi_registers),
 107         LC_OPT_LAST
 108 };
 109
 110 /** currently active assignments (while processing a basic block)
 111  * maps registers to values(their current copies) */
 112 static ir_node **assignments;
 113
 114 /**
 115  * allocation information: last_uses, register preferences
 116  * the information is per firm-node.
 117  */
 118 struct allocation_info_t {
 119         unsigned  last_uses[2];   /**< bitset indicating last uses (input pos) */
 120         ir_node  *current_value;  /**< copy of the value that should be used */
 121         ir_node  *original_value; /**< for copies point to original value */
 122         float     prefs[0];       /**< register preferences */
 123 };
 124 typedef struct allocation_info_t allocation_info_t;
 125
 126 /** helper datastructure used when sorting register preferences */
 127 struct reg_pref_t {
 128         unsigned num;
 129         float    pref;
 130 };
 131 typedef struct reg_pref_t reg_pref_t;
 132
 133 /** per basic-block information */
 134 struct block_info_t {
 135         bool     processed;       /**< indicate whether block is processed */
 136         ir_node *assignments[0];  /**< register assignments at end of block */
 137 };
 138 typedef struct block_info_t block_info_t;
 139
 140 /**
 141  * Get the allocation info for a node.
 142  * The info is allocated on the first visit of a node.
 143  */
 144 static allocation_info_t *get_allocation_info(ir_node *node)
 145 {
 146         allocation_info_t *info = (allocation_info_t*)get_irn_link(node);
 147         if (info == NULL) {
 148                 info = OALLOCFZ(&obst, allocation_info_t, prefs, n_regs);
 149                 info->current_value  = node;
 150                 info->original_value = node;
 151                 set_irn_link(node, info);
 152         }
 153
 154         return info;
 155 }
 156
 157 static allocation_info_t *try_get_allocation_info(const ir_node *node)
 158 {
 159         return (allocation_info_t*) get_irn_link(node);
 160 }
 161
 162 /**
 163  * Get allocation information for a basic block
 164  */
 165 static block_info_t *get_block_info(ir_node *block)
 166 {
 167         block_info_t *info = (block_info_t*)get_irn_link(block);
 168
 169         assert(is_Block(block));
 170         if (info == NULL) {
 171                 info = OALLOCFZ(&obst, block_info_t, assignments, n_regs);
 172                 set_irn_link(block, info);
 173         }
 174
 175         return info;
 176 }
 177
 178 /**
 179  * Link the allocation info of a node to a copy.
 180  * Afterwards, both nodes uses the same allocation info.
 181  * Copy must not have an allocation info assigned yet.
 182  *
 183  * @param copy   the node that gets the allocation info assigned
 184  * @param value  the original node
 185  */
 186 static void mark_as_copy_of(ir_node *copy, ir_node *value)
 187 {
 188         ir_node           *original;
 189         allocation_info_t *info      = get_allocation_info(value);
 190         allocation_info_t *copy_info = get_allocation_info(copy);
 191
 192         /* find original value */
 193         original = info->original_value;
 194         if (original != value) {
 195                 info = get_allocation_info(original);
 196         }
 197
 198         assert(info->original_value == original);
 199         info->current_value = copy;
 200
 201         /* the copy should not be linked to something else yet */
 202         assert(copy_info->original_value == copy);
 203         copy_info->original_value = original;
 204
 205         /* copy over allocation preferences */
 206         memcpy(copy_info->prefs, info->prefs, n_regs * sizeof(copy_info->prefs[0]));
 207 }
 208
 209 /**
 210  * Calculate the penalties for every register on a node and its live neighbors.
 211  *
 212  * @param live_nodes  the set of live nodes at the current position, may be NULL
 213  * @param penalty     the penalty to subtract from
 214  * @param limited     a raw bitset containing the limited set for the node
 215  * @param node        the node
 216  */
 217 static void give_penalties_for_limits(const ir_nodeset_t *live_nodes,
 218                                       float penalty, const unsigned* limited,
 219                                       ir_node *node)
 220 {
 221         ir_nodeset_iterator_t iter;
 222         unsigned              r;
 223         size_t                n_allowed;
 224         allocation_info_t     *info = get_allocation_info(node);
 225         ir_node               *neighbor;
 226
 227         /* give penalty for all forbidden regs */
 228         for (r = 0; r < n_regs; ++r) {
 229                 if (rbitset_is_set(limited, r))
 230                         continue;
 231
 232                 info->prefs[r] -= penalty;
 233         }
 234
 235         /* all other live values should get a penalty for allowed regs */
 236         if (live_nodes == NULL)
 237                 return;
 238
 239         penalty   *= NEIGHBOR_FACTOR;
 240         n_allowed  = rbitset_popcount(limited, n_regs);
 241         if (n_allowed > 1) {
 242                 /* only create a very weak penalty if multiple regs are allowed */
 243                 penalty = (penalty * 0.8f) / n_allowed;
 244         }
 245         foreach_ir_nodeset(live_nodes, neighbor, iter) {
 246                 allocation_info_t *neighbor_info;
 247
 248                 /* TODO: if op is used on multiple inputs we might not do a
 249                  * continue here */
 250                 if (neighbor == node)
 251                         continue;
 252
 253                 neighbor_info = get_allocation_info(neighbor);
 254                 for (r = 0; r < n_regs; ++r) {
 255                         if (!rbitset_is_set(limited, r))
 256                                 continue;
 257
 258                         neighbor_info->prefs[r] -= penalty;
 259                 }
 260         }
 261 }
 262
 263 /**
 264  * Calculate the preferences of a definition for the current register class.
 265  * If the definition uses a limited set of registers, reduce the preferences
 266  * for the limited register on the node and its neighbors.
 267  *
 268  * @param live_nodes  the set of live nodes at the current node
 269  * @param weight      the weight
 270  * @param node        the current node
 271  */
 272 static void check_defs(const ir_nodeset_t *live_nodes, float weight,
 273                        ir_node *node)
 274 {
 275         const arch_register_req_t *req = arch_get_irn_register_req(node);
 276         if (req->type & arch_register_req_type_limited) {
 277                 const unsigned *limited = req->limited;
 278                 float           penalty = weight * DEF_FACTOR;
 279                 give_penalties_for_limits(live_nodes, penalty, limited, node);
 280         }
 281
 282         if (req->type & arch_register_req_type_should_be_same) {
 283                 ir_node           *insn  = skip_Proj(node);
 284                 allocation_info_t *info  = get_allocation_info(node);
 285                 int                arity = get_irn_arity(insn);
 286                 int                i;
 287
 288                 float factor = 1.0f / rbitset_popcount(&req->other_same, arity);
 289                 for (i = 0; i < arity; ++i) {
 290                         ir_node           *op;
 291                         unsigned           r;
 292                         allocation_info_t *op_info;
 293
 294                         if (!rbitset_is_set(&req->other_same, i))
 295                                 continue;
 296
 297                         op = get_irn_n(insn, i);
 298
 299                         /* if we the value at the should_be_same input doesn't die at the
 300                          * node, then it is no use to propagate the constraints (since a
 301                          * copy will emerge anyway) */
 302                         if (ir_nodeset_contains(live_nodes, op))
 303                                 continue;
 304
 305                         op_info = get_allocation_info(op);
 306                         for (r = 0; r < n_regs; ++r) {
 307                                 op_info->prefs[r] += info->prefs[r] * factor;
 308                         }
 309                 }
 310         }
 311 }
 312
 313 /**
 314  * Walker: Runs an a block calculates the preferences for any
 315  * node and every register from the considered register class.
 316  */
 317 static void analyze_block(ir_node *block, void *data)
 318 {
 319         float         weight = (float)get_block_execfreq(execfreqs, block);
 320         ir_nodeset_t  live_nodes;
 321         ir_node      *node;
 322         (void) data;
 323
 324         ir_nodeset_init(&live_nodes);
 325         be_liveness_end_of_block(lv, cls, block, &live_nodes);
 326
 327         sched_foreach_reverse(block, node) {
 328                 allocation_info_t *info;
 329                 int                i;
 330                 int                arity;
 331
 332                 if (is_Phi(node))
 333                         break;
 334
 335                 if (create_preferences) {
 336                         ir_node *value;
 337                         be_foreach_definition(node, cls, value,
 338                                 check_defs(&live_nodes, weight, value);
 339                         );
 340                 }
 341
 342                 /* mark last uses */
 343                 arity = get_irn_arity(node);
 344
 345                 /* the allocation info node currently only uses 1 unsigned value
 346                    to mark last used inputs. So we will fail for a node with more than
 347                    32 inputs. */
 348                 if (arity >= (int) sizeof(info->last_uses) * 8) {
 349                         panic("Node with more than %d inputs not supported yet",
 350                                         (int) sizeof(info->last_uses) * 8);
 351                 }
 352
 353                 info = get_allocation_info(node);
 354                 for (i = 0; i < arity; ++i) {
 355                         ir_node                   *op  = get_irn_n(node, i);
 356                         const arch_register_req_t *req = arch_get_irn_register_req(op);
 357                         if (req->cls != cls)
 358                                 continue;
 359
 360                         /* last usage of a value? */
 361                         if (!ir_nodeset_contains(&live_nodes, op)) {
 362                                 rbitset_set(info->last_uses, i);
 363                         }
 364                 }
 365
 366                 be_liveness_transfer(cls, node, &live_nodes);
 367
 368                 if (create_preferences) {
 369                         /* update weights based on usage constraints */
 370                         for (i = 0; i < arity; ++i) {
 371                                 const arch_register_req_t *req;
 372                                 const unsigned            *limited;
 373                                 ir_node                   *op = get_irn_n(node, i);
 374
 375                                 if (!arch_irn_consider_in_reg_alloc(cls, op))
 376                                         continue;
 377
 378                                 req = arch_get_irn_register_req_in(node, i);
 379                                 if (!(req->type & arch_register_req_type_limited))
 380                                         continue;
 381
 382                                 limited = req->limited;
 383                                 give_penalties_for_limits(&live_nodes, weight * USE_FACTOR,
 384                                                           limited, op);
 385                         }
 386                 }
 387         }
 388
 389         ir_nodeset_destroy(&live_nodes);
 390 }
 391
 392 static void congruence_def(ir_nodeset_t *live_nodes, const ir_node *node)
 393 {
 394         const arch_register_req_t *req = arch_get_irn_register_req(node);
 395
 396         /* should be same constraint? */
 397         if (req->type & arch_register_req_type_should_be_same) {
 398                 const ir_node *insn  = skip_Proj_const(node);
 399                 int      arity = get_irn_arity(insn);
 400                 int      i;
 401                 unsigned node_idx = get_irn_idx(node);
 402                 node_idx          = uf_find(congruence_classes, node_idx);
 403
 404                 for (i = 0; i < arity; ++i) {
 405                         ir_node               *live;
 406                         ir_node               *op;
 407                         int                    op_idx;
 408                         ir_nodeset_iterator_t  iter;
 409                         bool                   interferes = false;
 410
 411                         if (!rbitset_is_set(&req->other_same, i))
 412                                 continue;
 413
 414                         op     = get_irn_n(insn, i);
 415                         op_idx = get_irn_idx(op);
 416                         op_idx = uf_find(congruence_classes, op_idx);
 417
 418                         /* do we interfere with the value */
 419                         foreach_ir_nodeset(live_nodes, live, iter) {
 420                                 int lv_idx = get_irn_idx(live);
 421                                 lv_idx     = uf_find(congruence_classes, lv_idx);
 422                                 if (lv_idx == op_idx) {
 423                                         interferes = true;
 424                                         break;
 425                                 }
 426                         }
 427                         /* don't put in same affinity class if we interfere */
 428                         if (interferes)
 429                                 continue;
 430
 431                         node_idx = uf_union(congruence_classes, node_idx, op_idx);
 432                         DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
 433                             node, op));
 434                         /* one should_be_same is enough... */
 435                         break;
 436                 }
 437         }
 438 }
 439
 440 static void create_congruence_class(ir_node *block, void *data)
 441 {
 442         ir_nodeset_t  live_nodes;
 443         ir_node      *node;
 444
 445         (void) data;
 446         ir_nodeset_init(&live_nodes);
 447         be_liveness_end_of_block(lv, cls, block, &live_nodes);
 448
 449         /* check should be same constraints */
 450         sched_foreach_reverse(block, node) {
 451                 ir_node *value;
 452                 if (is_Phi(node))
 453                         break;
 454
 455                 be_foreach_definition(node, cls, value,
 456                         congruence_def(&live_nodes, value);
 457                 );
 458                 be_liveness_transfer(cls, node, &live_nodes);
 459         }
 460
 461         /* check phi congruence classes */
 462         sched_foreach_reverse_from(node, node) {
 463                 int i;
 464                 int arity;
 465                 int node_idx;
 466                 assert(is_Phi(node));
 467
 468                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 469                         continue;
 470
 471                 node_idx = get_irn_idx(node);
 472                 node_idx = uf_find(congruence_classes, node_idx);
 473
 474                 arity = get_irn_arity(node);
 475                 for (i = 0; i < arity; ++i) {
 476                         bool                  interferes = false;
 477                         ir_nodeset_iterator_t iter;
 478                         unsigned              r;
 479                         int                   old_node_idx;
 480                         ir_node              *live;
 481                         ir_node              *phi;
 482                         allocation_info_t    *head_info;
 483                         allocation_info_t    *other_info;
 484                         ir_node              *op     = get_Phi_pred(node, i);
 485                         int                   op_idx = get_irn_idx(op);
 486                         op_idx = uf_find(congruence_classes, op_idx);
 487
 488                         /* do we interfere with the value */
 489                         foreach_ir_nodeset(&live_nodes, live, iter) {
 490                                 int lv_idx = get_irn_idx(live);
 491                                 lv_idx     = uf_find(congruence_classes, lv_idx);
 492                                 if (lv_idx == op_idx) {
 493                                         interferes = true;
 494                                         break;
 495                                 }
 496                         }
 497                         /* don't put in same affinity class if we interfere */
 498                         if (interferes)
 499                                 continue;
 500                         /* any other phi has the same input? */
 501                         sched_foreach(block, phi) {
 502                                 ir_node *oop;
 503                                 int      oop_idx;
 504                                 if (!is_Phi(phi))
 505                                         break;
 506                                 if (!arch_irn_consider_in_reg_alloc(cls, phi))
 507                                         continue;
 508                                 oop = get_Phi_pred(phi, i);
 509                                 if (oop == op)
 510                                         continue;
 511                                 oop_idx = get_irn_idx(oop);
 512                                 oop_idx = uf_find(congruence_classes, oop_idx);
 513                                 if (oop_idx == op_idx) {
 514                                         interferes = true;
 515                                         break;
 516                                 }
 517                         }
 518                         if (interferes)
 519                                 continue;
 520
 521                         /* merge the 2 congruence classes and sum up their preferences */
 522                         old_node_idx = node_idx;
 523                         node_idx = uf_union(congruence_classes, node_idx, op_idx);
 524                         DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
 525                             node, op));
 526
 527                         old_node_idx = node_idx == old_node_idx ? op_idx : old_node_idx;
 528                         head_info  = get_allocation_info(get_idx_irn(irg, node_idx));
 529                         other_info = get_allocation_info(get_idx_irn(irg, old_node_idx));
 530                         for (r = 0; r < n_regs; ++r) {
 531                                 head_info->prefs[r] += other_info->prefs[r];
 532                         }
 533                 }
 534         }
 535 }
 536
 537 static void set_congruence_prefs(ir_node *node, void *data)
 538 {
 539         allocation_info_t *info;
 540         allocation_info_t *head_info;
 541         unsigned node_idx = get_irn_idx(node);
 542         unsigned node_set = uf_find(congruence_classes, node_idx);
 543
 544         (void) data;
 545
 546         /* head of congruence class or not in any class */
 547         if (node_set == node_idx)
 548                 return;
 549
 550         if (!arch_irn_consider_in_reg_alloc(cls, node))
 551                 return;
 552
 553         head_info = get_allocation_info(get_idx_irn(irg, node_set));
 554         info      = get_allocation_info(node);
 555
 556         memcpy(info->prefs, head_info->prefs, n_regs * sizeof(info->prefs[0]));
 557 }
 558
 559 static void combine_congruence_classes(void)
 560 {
 561         size_t n = get_irg_last_idx(irg);
 562         congruence_classes = XMALLOCN(int, n);
 563         uf_init(congruence_classes, n);
 564
 565         /* create congruence classes */
 566         irg_block_walk_graph(irg, create_congruence_class, NULL, NULL);
 567         /* merge preferences */
 568         irg_walk_graph(irg, set_congruence_prefs, NULL, NULL);
 569         free(congruence_classes);
 570 }
 571
 572
 573
 574 /**
 575  * Assign register reg to the given node.
 576  *
 577  * @param node  the node
 578  * @param reg   the register
 579  */
 580 static void use_reg(ir_node *node, const arch_register_t *reg)
 581 {
 582         unsigned r = arch_register_get_index(reg);
 583         assignments[r] = node;
 584         arch_set_irn_register(node, reg);
 585 }
 586
 587 static void free_reg_of_value(ir_node *node)
 588 {
 589         const arch_register_t *reg;
 590         unsigned               r;
 591
 592         if (!arch_irn_consider_in_reg_alloc(cls, node))
 593                 return;
 594
 595         reg        = arch_get_irn_register(node);
 596         r          = arch_register_get_index(reg);
 597         /* assignment->value may be NULL if a value is used at 2 inputs
 598            so it gets freed twice. */
 599         assert(assignments[r] == node || assignments[r] == NULL);
 600         assignments[r] = NULL;
 601 }
 602
 603 /**
 604  * Compare two register preferences in decreasing order.
 605  */
 606 static int compare_reg_pref(const void *e1, const void *e2)
 607 {
 608         const reg_pref_t *rp1 = (const reg_pref_t*) e1;
 609         const reg_pref_t *rp2 = (const reg_pref_t*) e2;
 610         if (rp1->pref < rp2->pref)
 611                 return 1;
 612         if (rp1->pref > rp2->pref)
 613                 return -1;
 614         return 0;
 615 }
 616
 617 static void fill_sort_candidates(reg_pref_t *regprefs,
 618                                  const allocation_info_t *info)
 619 {
 620         unsigned r;
 621
 622         for (r = 0; r < n_regs; ++r) {
 623                 float pref = info->prefs[r];
 624                 regprefs[r].num  = r;
 625                 regprefs[r].pref = pref;
 626         }
 627         /* TODO: use a stable sort here to avoid unnecessary register jumping */
 628         qsort(regprefs, n_regs, sizeof(regprefs[0]), compare_reg_pref);
 629 }
 630
 631 static bool try_optimistic_split(ir_node *to_split, ir_node *before,
 632                                  float pref, float pref_delta,
 633                                  unsigned *forbidden_regs, int recursion)
 634 {
 635         const arch_register_t *from_reg;
 636         const arch_register_t *reg;
 637         ir_node               *original_insn;
 638         ir_node               *block;
 639         ir_node               *copy;
 640         unsigned               r = 0;
 641         unsigned               from_r;
 642         unsigned               i;
 643         allocation_info_t     *info = get_allocation_info(to_split);
 644         reg_pref_t            *prefs;
 645         float                  delta = 0;
 646         float                  split_threshold;
 647
 648         (void) pref;
 649
 650         /* stupid hack: don't optimisticallt split don't spill nodes...
 651          * (so we don't split away the values produced because of
 652          *  must_be_different constraints) */
 653         original_insn = skip_Proj(info->original_value);
 654         if (arch_get_irn_flags(original_insn) & arch_irn_flags_dont_spill)
 655                 return false;
 656
 657         from_reg        = arch_get_irn_register(to_split);
 658         from_r          = arch_register_get_index(from_reg);
 659         block           = get_nodes_block(before);
 660         split_threshold = (float)get_block_execfreq(execfreqs, block) * SPLIT_DELTA;
 661
 662         if (pref_delta < split_threshold*0.5)
 663                 return false;
 664
 665         /* find the best free position where we could move to */
 666         prefs = ALLOCAN(reg_pref_t, n_regs);
 667         fill_sort_candidates(prefs, info);
 668         for (i = 0; i < n_regs; ++i) {
 669                 float apref;
 670                 float apref_delta;
 671                 bool  res;
 672                 bool  old_source_state;
 673
 674                 /* we need a normal register which is not an output register
 675                    an different from the current register of to_split */
 676                 r = prefs[i].num;
 677                 if (!rbitset_is_set(normal_regs, r))
 678                         continue;
 679                 if (rbitset_is_set(forbidden_regs, r))
 680                         continue;
 681                 if (r == from_r)
 682                         continue;
 683
 684                 /* is the split worth it? */
 685                 delta = pref_delta + prefs[i].pref;
 686                 if (delta < split_threshold) {
 687                         DB((dbg, LEVEL_3, "Not doing optimistical split of %+F (depth %d), win %f too low\n",
 688                                 to_split, recursion, delta));
 689                         return false;
 690                 }
 691
 692                 /* if the register is free then we can do the split */
 693                 if (assignments[r] == NULL)
 694                         break;
 695
 696                 /* otherwise we might try recursively calling optimistic_split */
 697                 if (recursion+1 > MAX_OPTIMISTIC_SPLIT_RECURSION)
 698                         continue;
 699
 700                 apref        = prefs[i].pref;
 701                 apref_delta  = i+1 < n_regs ? apref - prefs[i+1].pref : 0;
 702                 apref_delta += pref_delta - split_threshold;
 703
 704                 /* our source register isn't a useful destination for recursive
 705                    splits */
 706                 old_source_state = rbitset_is_set(forbidden_regs, from_r);
 707                 rbitset_set(forbidden_regs, from_r);
 708                 /* try recursive split */
 709                 res = try_optimistic_split(assignments[r], before, apref,
 710                                            apref_delta, forbidden_regs, recursion+1);
 711                 /* restore our destination */
 712                 if (old_source_state) {
 713                         rbitset_set(forbidden_regs, from_r);
 714                 } else {
 715                         rbitset_clear(forbidden_regs, from_r);
 716                 }
 717
 718                 if (res)
 719                         break;
 720         }
 721         if (i >= n_regs)
 722                 return false;
 723
 724         reg  = arch_register_for_index(cls, r);
 725         copy = be_new_Copy(block, to_split);
 726         mark_as_copy_of(copy, to_split);
 727         /* hacky, but correct here */
 728         if (assignments[arch_register_get_index(from_reg)] == to_split)
 729                 free_reg_of_value(to_split);
 730         use_reg(copy, reg);
 731         sched_add_before(before, copy);
 732
 733         DB((dbg, LEVEL_3,
 734             "Optimistic live-range split %+F move %+F(%s) -> %s before %+F (win %f, depth %d)\n",
 735             copy, to_split, from_reg->name, reg->name, before, delta, recursion));
 736         return true;
 737 }
 738
 739 /**
 740  * Determine and assign a register for node @p node
 741  */
 742 static void assign_reg(const ir_node *block, ir_node *node,
 743                        unsigned *forbidden_regs)
 744 {
 745         const arch_register_t     *final_reg;
 746         allocation_info_t         *info;
 747         const arch_register_req_t *req;
 748         reg_pref_t                *reg_prefs;
 749         ir_node                   *in_node;
 750         unsigned                   r;
 751         const unsigned            *allowed_regs;
 752         unsigned                   final_reg_index = 0;
 753
 754         assert(!is_Phi(node));
 755         /* preassigned register? */
 756         final_reg = arch_get_irn_register(node);
 757         if (final_reg != NULL) {
 758                 DB((dbg, LEVEL_2, "Preassignment %+F -> %s\n", node, final_reg->name));
 759                 use_reg(node, final_reg);
 760                 return;
 761         }
 762
 763         req = arch_get_irn_register_req(node);
 764         /* ignore reqs must be preassigned */
 765         assert (! (req->type & arch_register_req_type_ignore));
 766
 767         /* give should_be_same boni */
 768         info    = get_allocation_info(node);
 769         in_node = skip_Proj(node);
 770         if (req->type & arch_register_req_type_should_be_same) {
 771                 float weight = (float)get_block_execfreq(execfreqs, block);
 772                 int   arity  = get_irn_arity(in_node);
 773                 int   i;
 774
 775                 assert(arity <= (int) sizeof(req->other_same) * 8);
 776                 for (i = 0; i < arity; ++i) {
 777                         ir_node               *in;
 778                         const arch_register_t *reg;
 779                         unsigned               reg_index;
 780                         if (!rbitset_is_set(&req->other_same, i))
 781                                 continue;
 782
 783                         in  = get_irn_n(in_node, i);
 784                         reg = arch_get_irn_register(in);
 785                         assert(reg != NULL);
 786                         reg_index = arch_register_get_index(reg);
 787
 788                         /* if the value didn't die here then we should not propagate the
 789                          * should_be_same info */
 790                         if (assignments[reg_index] == in)
 791                                 continue;
 792
 793                         info->prefs[reg_index] += weight * AFF_SHOULD_BE_SAME;
 794                 }
 795         }
 796
 797         /* create list of register candidates and sort by their preference */
 798         DB((dbg, LEVEL_2, "Candidates for %+F:", node));
 799         reg_prefs = ALLOCAN(reg_pref_t, n_regs);
 800         fill_sort_candidates(reg_prefs, info);
 801         for (r = 0; r < n_regs; ++r) {
 802                 unsigned num = reg_prefs[r].num;
 803                 const arch_register_t *reg;
 804
 805                 if (!rbitset_is_set(normal_regs, num))
 806                         continue;
 807                 reg = arch_register_for_index(cls, num);
 808                 DB((dbg, LEVEL_2, " %s(%f)", reg->name, reg_prefs[r].pref));
 809         }
 810         DB((dbg, LEVEL_2, "\n"));
 811
 812         allowed_regs = normal_regs;
 813         if (req->type & arch_register_req_type_limited) {
 814                 allowed_regs = req->limited;
 815         }
 816
 817         for (r = 0; r < n_regs; ++r) {
 818                 float   pref, delta;
 819                 ir_node *before;
 820                 bool    res;
 821
 822                 final_reg_index = reg_prefs[r].num;
 823                 if (!rbitset_is_set(allowed_regs, final_reg_index))
 824                         continue;
 825                 /* alignment constraint? */
 826                 if (req->width > 1 && (req->type & arch_register_req_type_aligned)
 827                                 && (final_reg_index % req->width) != 0)
 828                         continue;
 829
 830                 if (assignments[final_reg_index] == NULL)
 831                         break;
 832                 pref   = reg_prefs[r].pref;
 833                 delta  = r+1 < n_regs ? pref - reg_prefs[r+1].pref : 0;
 834                 before = skip_Proj(node);
 835                 res    = try_optimistic_split(assignments[final_reg_index], before,
 836                                               pref, delta, forbidden_regs, 0);
 837                 if (res)
 838                         break;
 839         }
 840         if (r >= n_regs) {
 841                 /* the common reason to hit this panic is when 1 of your nodes is not
 842                  * register pressure faithful */
 843                 panic("No register left for %+F\n", node);
 844         }
 845
 846         final_reg = arch_register_for_index(cls, final_reg_index);
 847         DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, final_reg->name));
 848         use_reg(node, final_reg);
 849 }
 850
 851 /**
 852  * Add an permutation in front of a node and change the assignments
 853  * due to this permutation.
 854  *
 855  * To understand this imagine a permutation like this:
 856  *
 857  * 1 -> 2
 858  * 2 -> 3
 859  * 3 -> 1, 5
 860  * 4 -> 6
 861  * 5
 862  * 6
 863  * 7 -> 7
 864  *
 865  * First we count how many destinations a single value has. At the same time
 866  * we can be sure that each destination register has at most 1 source register
 867  * (it can have 0 which means we don't care what value is in it).
 868  * We ignore all fulfilled permuations (like 7->7)
 869  * In a first pass we create as much copy instructions as possible as they
 870  * are generally cheaper than exchanges. We do this by counting into how many
 871  * destinations a register has to be copied (in the example it's 2 for register
 872  * 3, or 1 for the registers 1,2,4 and 7).
 873  * We can then create a copy into every destination register when the usecount
 874  * of that register is 0 (= noone else needs the value in the register).
 875  *
 876  * After this step we should only have cycles left. We implement a cyclic
 877  * permutation of n registers with n-1 transpositions.
 878  *
 879  * @param live_nodes   the set of live nodes, updated due to live range split
 880  * @param before       the node before we add the permutation
 881  * @param permutation  the permutation array indices are the destination
 882  *                     registers, the values in the array are the source
 883  *                     registers.
 884  */
 885 static void permute_values(ir_nodeset_t *live_nodes, ir_node *before,
 886                            unsigned *permutation)
 887 {
 888         unsigned  *n_used = ALLOCANZ(unsigned, n_regs);
 889         ir_node   *block;
 890         unsigned   r;
 891
 892         /* determine how often each source register needs to be read */
 893         for (r = 0; r < n_regs; ++r) {
 894                 unsigned  old_reg = permutation[r];
 895                 ir_node  *value;
 896
 897                 value = assignments[old_reg];
 898                 if (value == NULL) {
 899                         /* nothing to do here, reg is not live. Mark it as fixpoint
 900                          * so we ignore it in the next steps */
 901                         permutation[r] = r;
 902                         continue;
 903                 }
 904
 905                 ++n_used[old_reg];
 906         }
 907
 908         block = get_nodes_block(before);
 909
 910         /* step1: create copies where immediately possible */
 911         for (r = 0; r < n_regs; /* empty */) {
 912                 ir_node *copy;
 913                 ir_node *src;
 914                 const arch_register_t *reg;
 915                 unsigned               old_r = permutation[r];
 916
 917                 /* - no need to do anything for fixed points.
 918                    - we can't copy if the value in the dest reg is still needed */
 919                 if (old_r == r || n_used[r] > 0) {
 920                         ++r;
 921                         continue;
 922                 }
 923
 924                 /* create a copy */
 925                 src  = assignments[old_r];
 926                 copy = be_new_Copy(block, src);
 927                 sched_add_before(before, copy);
 928                 reg = arch_register_for_index(cls, r);
 929                 DB((dbg, LEVEL_2, "Copy %+F (from %+F, before %+F) -> %s\n",
 930                     copy, src, before, reg->name));
 931                 mark_as_copy_of(copy, src);
 932                 use_reg(copy, reg);
 933
 934                 if (live_nodes != NULL) {
 935                         ir_nodeset_insert(live_nodes, copy);
 936                 }
 937
 938                 /* old register has 1 user less, permutation is resolved */
 939                 assert(arch_register_get_index(arch_get_irn_register(src)) == old_r);
 940                 permutation[r] = r;
 941
 942                 assert(n_used[old_r] > 0);
 943                 --n_used[old_r];
 944                 if (n_used[old_r] == 0) {
 945                         if (live_nodes != NULL) {
 946                                 ir_nodeset_remove(live_nodes, src);
 947                         }
 948                         free_reg_of_value(src);
 949                 }
 950
 951                 /* advance or jump back (if this copy enabled another copy) */
 952                 if (old_r < r && n_used[old_r] == 0) {
 953                         r = old_r;
 954                 } else {
 955                         ++r;
 956                 }
 957         }
 958
 959         /* at this point we only have "cycles" left which we have to resolve with
 960          * perm instructions
 961          * TODO: if we have free registers left, then we should really use copy
 962          * instructions for any cycle longer than 2 registers...
 963          * (this is probably architecture dependent, there might be archs where
 964          *  copies are preferable even for 2-cycles) */
 965
 966         /* create perms with the rest */
 967         for (r = 0; r < n_regs; /* empty */) {
 968                 const arch_register_t *reg;
 969                 unsigned  old_r = permutation[r];
 970                 unsigned  r2;
 971                 ir_node  *in[2];
 972                 ir_node  *perm;
 973                 ir_node  *proj0;
 974                 ir_node  *proj1;
 975
 976                 if (old_r == r) {
 977                         ++r;
 978                         continue;
 979                 }
 980
 981                 /* we shouldn't have copies from 1 value to multiple destinations left*/
 982                 assert(n_used[old_r] == 1);
 983
 984                 /* exchange old_r and r2; after that old_r is a fixed point */
 985                 r2 = permutation[old_r];
 986
 987                 in[0] = assignments[r2];
 988                 in[1] = assignments[old_r];
 989                 perm = be_new_Perm(cls, block, 2, in);
 990                 sched_add_before(before, perm);
 991                 DB((dbg, LEVEL_2, "Perm %+F (perm %+F,%+F, before %+F)\n",
 992                     perm, in[0], in[1], before));
 993
 994                 proj0 = new_r_Proj(perm, get_irn_mode(in[0]), 0);
 995                 mark_as_copy_of(proj0, in[0]);
 996                 reg = arch_register_for_index(cls, old_r);
 997                 use_reg(proj0, reg);
 998
 999                 proj1 = new_r_Proj(perm, get_irn_mode(in[1]), 1);
1000                 mark_as_copy_of(proj1, in[1]);
1001                 reg = arch_register_for_index(cls, r2);
1002                 use_reg(proj1, reg);
1003
1004                 /* 1 value is now in the correct register */
1005                 permutation[old_r] = old_r;
1006                 /* the source of r changed to r2 */
1007                 permutation[r] = r2;
1008
1009                 /* if we have reached a fixpoint update data structures */
1010                 if (live_nodes != NULL) {
1011                         ir_nodeset_remove(live_nodes, in[0]);
1012                         ir_nodeset_remove(live_nodes, in[1]);
1013                         ir_nodeset_remove(live_nodes, proj0);
1014                         ir_nodeset_insert(live_nodes, proj1);
1015                 }
1016         }
1017
1018 #ifdef DEBUG_libfirm
1019         /* now we should only have fixpoints left */
1020         for (r = 0; r < n_regs; ++r) {
1021                 assert(permutation[r] == r);
1022         }
1023 #endif
1024 }
1025
1026 /**
1027  * Free regs for values last used.
1028  *
1029  * @param live_nodes   set of live nodes, will be updated
1030  * @param node         the node to consider
1031  */
1032 static void free_last_uses(ir_nodeset_t *live_nodes, ir_node *node)
1033 {
1034         allocation_info_t *info      = get_allocation_info(node);
1035         const unsigned    *last_uses = info->last_uses;
1036         int                arity     = get_irn_arity(node);
1037         int                i;
1038
1039         for (i = 0; i < arity; ++i) {
1040                 ir_node *op;
1041
1042                 /* check if one operand is the last use */
1043                 if (!rbitset_is_set(last_uses, i))
1044                         continue;
1045
1046                 op = get_irn_n(node, i);
1047                 free_reg_of_value(op);
1048                 ir_nodeset_remove(live_nodes, op);
1049         }
1050 }
1051
1052 /**
1053  * change inputs of a node to the current value (copies/perms)
1054  */
1055 static void rewire_inputs(ir_node *node)
1056 {
1057         int i;
1058         int arity = get_irn_arity(node);
1059
1060         for (i = 0; i < arity; ++i) {
1061                 ir_node           *op = get_irn_n(node, i);
1062                 allocation_info_t *info = try_get_allocation_info(op);
1063
1064                 if (info == NULL)
1065                         continue;
1066
1067                 info = get_allocation_info(info->original_value);
1068                 if (info->current_value != op) {
1069                         set_irn_n(node, i, info->current_value);
1070                 }
1071         }
1072 }
1073
1074 /**
1075  * Create a bitset of registers occupied with value living through an
1076  * instruction
1077  */
1078 static void determine_live_through_regs(unsigned *bitset, ir_node *node)
1079 {
1080         const allocation_info_t *info = get_allocation_info(node);
1081         unsigned r;
1082         int i;
1083         int arity;
1084
1085         /* mark all used registers as potentially live-through */
1086         for (r = 0; r < n_regs; ++r) {
1087                 if (assignments[r] == NULL)
1088                         continue;
1089                 if (!rbitset_is_set(normal_regs, r))
1090                         continue;
1091
1092                 rbitset_set(bitset, r);
1093         }
1094
1095         /* remove registers of value dying at the instruction */
1096         arity = get_irn_arity(node);
1097         for (i = 0; i < arity; ++i) {
1098                 ir_node               *op;
1099                 const arch_register_t *reg;
1100
1101                 if (!rbitset_is_set(info->last_uses, i))
1102                         continue;
1103
1104                 op  = get_irn_n(node, i);
1105                 reg = arch_get_irn_register(op);
1106                 rbitset_clear(bitset, arch_register_get_index(reg));
1107         }
1108 }
1109
1110 static void solve_lpp(ir_nodeset_t *live_nodes, ir_node *node,
1111                       unsigned *forbidden_regs, unsigned *live_through_regs)
1112 {
1113         unsigned *forbidden_edges = rbitset_malloc(n_regs * n_regs);
1114         int      *lpp_vars        = XMALLOCNZ(int, n_regs*n_regs);
1115         int       arity           = get_irn_arity(node);
1116         int       i;
1117         unsigned  l;
1118         unsigned  r;
1119
1120         lpp_t *lpp = lpp_new("prefalloc", lpp_minimize);
1121         //lpp_set_time_limit(lpp, 20);
1122         lpp_set_log(lpp, stdout);
1123
1124         /** mark some edges as forbidden */
1125         for (i = 0; i < arity; ++i) {
1126                 ir_node                   *op = get_irn_n(node, i);
1127                 const arch_register_t     *reg;
1128                 const arch_register_req_t *req;
1129                 const unsigned            *limited;
1130                 unsigned                   current_reg;
1131
1132                 if (!arch_irn_consider_in_reg_alloc(cls, op))
1133                         continue;
1134
1135                 req = arch_get_irn_register_req_in(node, i);
1136                 if (!(req->type & arch_register_req_type_limited))
1137                         continue;
1138
1139                 limited     = req->limited;
1140                 reg         = arch_get_irn_register(op);
1141                 current_reg = arch_register_get_index(reg);
1142                 for (r = 0; r < n_regs; ++r) {
1143                         if (rbitset_is_set(limited, r))
1144                                 continue;
1145
1146                         rbitset_set(forbidden_edges, current_reg*n_regs + r);
1147                 }
1148         }
1149
1150         /* add all combinations, except for not allowed ones */
1151         for (l = 0; l < n_regs; ++l) {
1152                 if (!rbitset_is_set(normal_regs, l)) {
1153                         char name[15];
1154                         snprintf(name, sizeof(name), "%u_to_%u", l, l);
1155                         lpp_vars[l*n_regs+l] = lpp_add_var(lpp, name, lpp_binary, 1);
1156                         continue;
1157                 }
1158
1159                 for (r = 0; r < n_regs; ++r) {
1160                         if (!rbitset_is_set(normal_regs, r))
1161                                 continue;
1162                         if (rbitset_is_set(forbidden_edges, l*n_regs + r))
1163                                 continue;
1164                         /* livethrough values may not use constrained output registers */
1165                         if (rbitset_is_set(live_through_regs, l)
1166                             && rbitset_is_set(forbidden_regs, r))
1167                                 continue;
1168
1169                         char name[15];
1170                         snprintf(name, sizeof(name), "%u_to_%u", l, r);
1171
1172                         double costs = l==r ? 9 : 8;
1173                         lpp_vars[l*n_regs+r]
1174                                 = lpp_add_var(lpp, name, lpp_binary, costs);
1175                         assert(lpp_vars[l*n_regs+r] > 0);
1176                 }
1177         }
1178         /* add constraints */
1179         for (l = 0; l < n_regs; ++l) {
1180                 int constraint;
1181                 /* only 1 destination per register */
1182                 constraint = -1;
1183                 for (r = 0; r < n_regs; ++r) {
1184                         int var = lpp_vars[l*n_regs+r];
1185                         if (var == 0)
1186                                 continue;
1187                         if (constraint < 0) {
1188                                 char name[64];
1189                                 snprintf(name, sizeof(name), "%u_to_dest", l);
1190                                 constraint = lpp_add_cst(lpp, name, lpp_equal, 1);
1191                         }
1192                         lpp_set_factor_fast(lpp, constraint, var, 1);
1193                 }
1194                 /* each destination used by at most 1 value */
1195                 constraint = -1;
1196                 for (r = 0; r < n_regs; ++r) {
1197                         int var = lpp_vars[r*n_regs+l];
1198                         if (var == 0)
1199                                 continue;
1200                         if (constraint < 0) {
1201                                 char name[64];
1202                                 snprintf(name, sizeof(name), "one_to_%u", l);
1203                                 constraint = lpp_add_cst(lpp, name, lpp_less_equal, 1);
1204                         }
1205                         lpp_set_factor_fast(lpp, constraint, var, 1);
1206                 }
1207         }
1208
1209         lpp_dump_plain(lpp, fopen("lppdump.txt", "w"));
1210
1211         /* solve lpp */
1212         {
1213                 ir_graph     *irg     = get_irn_irg(node);
1214                 be_options_t *options = be_get_irg_options(irg);
1215                 unsigned     *assignment;
1216                 lpp_solve(lpp, options->ilp_server, options->ilp_solver);
1217                 if (!lpp_is_sol_valid(lpp))
1218                         panic("ilp solution not valid!");
1219
1220                 assignment = ALLOCAN(unsigned, n_regs);
1221                 for (l = 0; l < n_regs; ++l) {
1222                         unsigned dest_reg = (unsigned)-1;
1223                         for (r = 0; r < n_regs; ++r) {
1224                                 int var = lpp_vars[l*n_regs+r];
1225                                 if (var == 0)
1226                                         continue;
1227                                 double val = lpp_get_var_sol(lpp, var);
1228                                 if (val == 1) {
1229                                         assert(dest_reg == (unsigned)-1);
1230                                         dest_reg = r;
1231                                 }
1232                         }
1233                         assert(dest_reg != (unsigned)-1);
1234                         assignment[dest_reg] = l;
1235                 }
1236
1237                 fprintf(stderr, "Assignment: ");
1238                 for (l = 0; l < n_regs; ++l) {
1239                         fprintf(stderr, "%u ", assignment[l]);
1240                 }
1241                 fprintf(stderr, "\n");
1242                 fflush(stdout);
1243                 permute_values(live_nodes, node, assignment);
1244         }
1245         lpp_free(lpp);
1246 }
1247
1248 /**
1249  * Enforce constraints at a node by live range splits.
1250  *
1251  * @param  live_nodes  the set of live nodes, might be changed
1252  * @param  node        the current node
1253  */
1254 static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node,
1255                                 unsigned *forbidden_regs)
1256 {
1257         int arity = get_irn_arity(node);
1258         int i, res;
1259         hungarian_problem_t *bp;
1260         unsigned l, r;
1261         unsigned *assignment;
1262         ir_node  *value;
1263
1264         /* construct a list of register occupied by live-through values */
1265         unsigned *live_through_regs = NULL;
1266
1267         /* see if any use constraints are not met */
1268         bool double_width = false;
1269         bool good = true;
1270         for (i = 0; i < arity; ++i) {
1271                 ir_node                   *op = get_irn_n(node, i);
1272                 const arch_register_t     *reg;
1273                 const arch_register_req_t *req;
1274                 const unsigned            *limited;
1275                 unsigned                   reg_index;
1276
1277                 if (!arch_irn_consider_in_reg_alloc(cls, op))
1278                         continue;
1279
1280                 /* are there any limitations for the i'th operand? */
1281                 req = arch_get_irn_register_req_in(node, i);
1282                 if (req->width > 1)
1283                         double_width = true;
1284                 if (!(req->type & arch_register_req_type_limited))
1285                         continue;
1286
1287                 limited   = req->limited;
1288                 reg       = arch_get_irn_register(op);
1289                 reg_index = arch_register_get_index(reg);
1290                 if (!rbitset_is_set(limited, reg_index)) {
1291                         /* found an assignment outside the limited set */
1292                         good = false;
1293                         break;
1294                 }
1295         }
1296
1297         /* is any of the live-throughs using a constrained output register? */
1298         be_foreach_definition(node, cls, value,
1299                 if (req_->width > 1)
1300                         double_width = true;
1301                 if (! (req_->type & arch_register_req_type_limited))
1302                         continue;
1303                 if (live_through_regs == NULL) {
1304                         rbitset_alloca(live_through_regs, n_regs);
1305                         determine_live_through_regs(live_through_regs, node);
1306                 }
1307                 rbitset_or(forbidden_regs, req_->limited, n_regs);
1308                 if (rbitsets_have_common(req_->limited, live_through_regs, n_regs))
1309                         good = false;
1310         );
1311
1312         if (good)
1313                 return;
1314
1315         /* create these arrays if we haven't yet */
1316         if (live_through_regs == NULL) {
1317                 rbitset_alloca(live_through_regs, n_regs);
1318         }
1319
1320         if (double_width) {
1321                 /* only the ILP variant can solve this yet */
1322                 solve_lpp(live_nodes, node, forbidden_regs, live_through_regs);
1323                 return;
1324         }
1325
1326         /* at this point we have to construct a bipartite matching problem to see
1327          * which values should go to which registers
1328          * Note: We're building the matrix in "reverse" - source registers are
1329          *       right, destinations left because this will produce the solution
1330          *       in the format required for permute_values.
1331          */
1332         bp = hungarian_new(n_regs, n_regs, HUNGARIAN_MATCH_PERFECT);
1333
1334         /* add all combinations, then remove not allowed ones */
1335         for (l = 0; l < n_regs; ++l) {
1336                 if (!rbitset_is_set(normal_regs, l)) {
1337                         hungarian_add(bp, l, l, 1);
1338                         continue;
1339                 }
1340
1341                 for (r = 0; r < n_regs; ++r) {
1342                         if (!rbitset_is_set(normal_regs, r))
1343                                 continue;
1344                         /* livethrough values may not use constrainted output registers */
1345                         if (rbitset_is_set(live_through_regs, l)
1346                                         && rbitset_is_set(forbidden_regs, r))
1347                                 continue;
1348
1349                         hungarian_add(bp, r, l, l == r ? 9 : 8);
1350                 }
1351         }
1352
1353         for (i = 0; i < arity; ++i) {
1354                 ir_node                   *op = get_irn_n(node, i);
1355                 const arch_register_t     *reg;
1356                 const arch_register_req_t *req;
1357                 const unsigned            *limited;
1358                 unsigned                   current_reg;
1359
1360                 if (!arch_irn_consider_in_reg_alloc(cls, op))
1361                         continue;
1362
1363                 req = arch_get_irn_register_req_in(node, i);
1364                 if (!(req->type & arch_register_req_type_limited))
1365                         continue;
1366
1367                 limited     = req->limited;
1368                 reg         = arch_get_irn_register(op);
1369                 current_reg = arch_register_get_index(reg);
1370                 for (r = 0; r < n_regs; ++r) {
1371                         if (rbitset_is_set(limited, r))
1372                                 continue;
1373                         hungarian_remove(bp, r, current_reg);
1374                 }
1375         }
1376
1377         //hungarian_print_cost_matrix(bp, 1);
1378         hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1379
1380         assignment = ALLOCAN(unsigned, n_regs);
1381         res = hungarian_solve(bp, assignment, NULL, 0);
1382         assert(res == 0);
1383
1384 #if 0
1385         fprintf(stderr, "Swap result:");
1386         for (i = 0; i < (int) n_regs; ++i) {
1387                 fprintf(stderr, " %d", assignment[i]);
1388         }
1389         fprintf(stderr, "\n");
1390 #endif
1391
1392         hungarian_free(bp);
1393
1394         permute_values(live_nodes, node, assignment);
1395 }
1396
1397 /** test whether a node @p n is a copy of the value of node @p of */
1398 static bool is_copy_of(ir_node *value, ir_node *test_value)
1399 {
1400         allocation_info_t *test_info;
1401         allocation_info_t *info;
1402
1403         if (value == test_value)
1404                 return true;
1405
1406         info      = get_allocation_info(value);
1407         test_info = get_allocation_info(test_value);
1408         return test_info->original_value == info->original_value;
1409 }
1410
1411 /**
1412  * find a value in the end-assignment of a basic block
1413  * @returns the index into the assignment array if found
1414  *          -1 if not found
1415  */
1416 static int find_value_in_block_info(block_info_t *info, ir_node *value)
1417 {
1418         unsigned   r;
1419         ir_node  **end_assignments = info->assignments;
1420         for (r = 0; r < n_regs; ++r) {
1421                 ir_node *a_value = end_assignments[r];
1422
1423                 if (a_value == NULL)
1424                         continue;
1425                 if (is_copy_of(a_value, value))
1426                         return (int) r;
1427         }
1428
1429         return -1;
1430 }
1431
1432 /**
1433  * Create the necessary permutations at the end of a basic block to fullfill
1434  * the register assignment for phi-nodes in the next block
1435  */
1436 static void add_phi_permutations(ir_node *block, int p)
1437 {
1438         unsigned   r;
1439         unsigned  *permutation;
1440         ir_node  **old_assignments;
1441         bool       need_permutation;
1442         ir_node   *phi;
1443         ir_node   *pred = get_Block_cfgpred_block(block, p);
1444
1445         block_info_t *pred_info = get_block_info(pred);
1446
1447         /* predecessor not processed yet? nothing to do */
1448         if (!pred_info->processed)
1449                 return;
1450
1451         permutation = ALLOCAN(unsigned, n_regs);
1452         for (r = 0; r < n_regs; ++r) {
1453                 permutation[r] = r;
1454         }
1455
1456         /* check phi nodes */
1457         need_permutation = false;
1458         phi = sched_first(block);
1459         for ( ; is_Phi(phi); phi = sched_next(phi)) {
1460                 const arch_register_t *reg;
1461                 const arch_register_t *op_reg;
1462                 int                    regn;
1463                 int                    a;
1464                 ir_node               *op;
1465
1466                 if (!arch_irn_consider_in_reg_alloc(cls, phi))
1467                         continue;
1468
1469                 op = get_Phi_pred(phi, p);
1470                 a  = find_value_in_block_info(pred_info, op);
1471                 assert(a >= 0);
1472
1473                 reg  = arch_get_irn_register(phi);
1474                 regn = arch_register_get_index(reg);
1475                 /* same register? nothing to do */
1476                 if (regn == a)
1477                         continue;
1478
1479                 op     = pred_info->assignments[a];
1480                 op_reg = arch_get_irn_register(op);
1481                 /* virtual or joker registers are ok too */
1482                 if ((op_reg->type & arch_register_type_joker)
1483                                 || (op_reg->type & arch_register_type_virtual))
1484                         continue;
1485
1486                 permutation[regn] = a;
1487                 need_permutation  = true;
1488         }
1489
1490         if (need_permutation) {
1491                 /* permute values at end of predecessor */
1492                 old_assignments = assignments;
1493                 assignments     = pred_info->assignments;
1494                 permute_values(NULL, be_get_end_of_block_insertion_point(pred),
1495                                permutation);
1496                 assignments = old_assignments;
1497         }
1498
1499         /* change phi nodes to use the copied values */
1500         phi = sched_first(block);
1501         for ( ; is_Phi(phi); phi = sched_next(phi)) {
1502                 int      a;
1503                 ir_node *op;
1504
1505                 if (!arch_irn_consider_in_reg_alloc(cls, phi))
1506                         continue;
1507
1508                 op = get_Phi_pred(phi, p);
1509
1510                 /* we have permuted all values into the correct registers so we can
1511                    simply query which value occupies the phis register in the
1512                    predecessor */
1513                 a  = arch_register_get_index(arch_get_irn_register(phi));
1514                 op = pred_info->assignments[a];
1515                 set_Phi_pred(phi, p, op);
1516         }
1517 }
1518
1519 /**
1520  * Set preferences for a phis register based on the registers used on the
1521  * phi inputs.
1522  */
1523 static void adapt_phi_prefs(ir_node *phi)
1524 {
1525         int i;
1526         int arity = get_irn_arity(phi);
1527         ir_node           *block = get_nodes_block(phi);
1528         allocation_info_t *info  = get_allocation_info(phi);
1529
1530         for (i = 0; i < arity; ++i) {
1531                 ir_node               *op  = get_irn_n(phi, i);
1532                 const arch_register_t *reg = arch_get_irn_register(op);
1533                 ir_node               *pred_block;
1534                 block_info_t          *pred_block_info;
1535                 float                  weight;
1536                 unsigned               r;
1537
1538                 if (reg == NULL)
1539                         continue;
1540                 /* we only give the bonus if the predecessor already has registers
1541                  * assigned, otherwise we only see a dummy value
1542                  * and any conclusions about its register are useless */
1543                 pred_block = get_Block_cfgpred_block(block, i);
1544                 pred_block_info = get_block_info(pred_block);
1545                 if (!pred_block_info->processed)
1546                         continue;
1547
1548                 /* give bonus for already assigned register */
1549                 weight = (float)get_block_execfreq(execfreqs, pred_block);
1550                 r      = arch_register_get_index(reg);
1551                 info->prefs[r] += weight * AFF_PHI;
1552         }
1553 }
1554
1555 /**
1556  * After a phi has been assigned a register propagate preference inputs
1557  * to the phi inputs.
1558  */
1559 static void propagate_phi_register(ir_node *phi, unsigned assigned_r)
1560 {
1561         int      i;
1562         ir_node *block = get_nodes_block(phi);
1563         int      arity = get_irn_arity(phi);
1564
1565         for (i = 0; i < arity; ++i) {
1566                 ir_node           *op         = get_Phi_pred(phi, i);
1567                 allocation_info_t *info       = get_allocation_info(op);
1568                 ir_node           *pred_block = get_Block_cfgpred_block(block, i);
1569                 unsigned           r;
1570                 float              weight
1571                         = (float)get_block_execfreq(execfreqs, pred_block) * AFF_PHI;
1572
1573                 if (info->prefs[assigned_r] >= weight)
1574                         continue;
1575
1576                 /* promote the prefered register */
1577                 for (r = 0; r < n_regs; ++r) {
1578                         if (info->prefs[r] > -weight) {
1579                                 info->prefs[r] = -weight;
1580                         }
1581                 }
1582                 info->prefs[assigned_r] = weight;
1583
1584                 if (is_Phi(op))
1585                         propagate_phi_register(op, assigned_r);
1586         }
1587 }
1588
1589 static void assign_phi_registers(ir_node *block)
1590 {
1591         int                  n_phis = 0;
1592         int                  n;
1593         int                  res;
1594         unsigned            *assignment;
1595         ir_node             *node;
1596         hungarian_problem_t *bp;
1597
1598         /* count phi nodes */
1599         sched_foreach(block, node) {
1600                 if (!is_Phi(node))
1601                         break;
1602                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1603                         continue;
1604                 ++n_phis;
1605         }
1606
1607         if (n_phis == 0)
1608                 return;
1609
1610         /* build a bipartite matching problem for all phi nodes */
1611         bp = hungarian_new(n_phis, n_regs, HUNGARIAN_MATCH_PERFECT);
1612         n  = 0;
1613         sched_foreach(block, node) {
1614                 unsigned r;
1615
1616                 allocation_info_t *info;
1617                 if (!is_Phi(node))
1618                         break;
1619                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1620                         continue;
1621
1622                 /* give boni for predecessor colorings */
1623                 adapt_phi_prefs(node);
1624                 /* add stuff to bipartite problem */
1625                 info = get_allocation_info(node);
1626                 DB((dbg, LEVEL_3, "Prefs for %+F: ", node));
1627                 for (r = 0; r < n_regs; ++r) {
1628                         float costs;
1629
1630                         if (!rbitset_is_set(normal_regs, r))
1631                                 continue;
1632
1633                         costs = info->prefs[r];
1634                         costs = costs < 0 ? -logf(-costs+1) : logf(costs+1);
1635                         costs *= 100;
1636                         costs += 10000;
1637                         hungarian_add(bp, n, r, (int)costs);
1638                         DB((dbg, LEVEL_3, " %s(%f)", arch_register_for_index(cls, r)->name,
1639                                                 info->prefs[r]));
1640                 }
1641                 DB((dbg, LEVEL_3, "\n"));
1642                 ++n;
1643         }
1644
1645         //hungarian_print_cost_matrix(bp, 7);
1646         hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1647
1648         assignment = ALLOCAN(unsigned, n_regs);
1649         res        = hungarian_solve(bp, assignment, NULL, 0);
1650         assert(res == 0);
1651
1652         /* apply results */
1653         n = 0;
1654         sched_foreach(block, node) {
1655                 unsigned               r;
1656                 const arch_register_t *reg;
1657
1658                 if (!is_Phi(node))
1659                         break;
1660                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1661                         continue;
1662
1663                 r = assignment[n++];
1664                 assert(rbitset_is_set(normal_regs, r));
1665                 reg = arch_register_for_index(cls, r);
1666                 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
1667                 use_reg(node, reg);
1668
1669                 /* adapt preferences for phi inputs */
1670                 if (propagate_phi_registers)
1671                         propagate_phi_register(node, r);
1672         }
1673 }
1674
1675 /**
1676  * Walker: assign registers to all nodes of a block that
1677  * need registers from the currently considered register class.
1678  */
1679 static void allocate_coalesce_block(ir_node *block, void *data)
1680 {
1681         int            i;
1682         ir_nodeset_t   live_nodes;
1683         ir_node       *node;
1684         int            n_preds;
1685         block_info_t  *block_info;
1686         block_info_t **pred_block_infos;
1687         ir_node      **phi_ins;
1688         unsigned      *forbidden_regs; /**< collects registers which must
1689                                                 not be used for optimistic splits */
1690
1691         (void) data;
1692         DB((dbg, LEVEL_2, "* Block %+F\n", block));
1693
1694         /* clear assignments */
1695         block_info  = get_block_info(block);
1696         assignments = block_info->assignments;
1697
1698         ir_nodeset_init(&live_nodes);
1699
1700         /* gather regalloc infos of predecessor blocks */
1701         n_preds          = get_Block_n_cfgpreds(block);
1702         pred_block_infos = ALLOCAN(block_info_t*, n_preds);
1703         for (i = 0; i < n_preds; ++i) {
1704                 ir_node      *pred      = get_Block_cfgpred_block(block, i);
1705                 block_info_t *pred_info = get_block_info(pred);
1706                 pred_block_infos[i]     = pred_info;
1707         }
1708
1709         phi_ins = ALLOCAN(ir_node*, n_preds);
1710
1711         /* collect live-in nodes and preassigned values */
1712         be_lv_foreach(lv, block, be_lv_state_in, i) {
1713                 bool                       need_phi = false;
1714                 const arch_register_req_t *req;
1715                 const arch_register_t     *reg;
1716                 int                        p;
1717
1718                 node = be_lv_get_irn(lv, block, i);
1719                 req  = arch_get_irn_register_req(node);
1720                 if (req->cls != cls)
1721                         continue;
1722
1723                 if (req->type & arch_register_req_type_ignore) {
1724                         allocation_info_t *info = get_allocation_info(node);
1725                         info->current_value = node;
1726
1727                         reg = arch_get_irn_register(node);
1728                         assert(reg != NULL); /* ignore values must be preassigned */
1729                         use_reg(node, reg);
1730                         continue;
1731                 }
1732
1733                 /* check all predecessors for this value, if it is not everywhere the
1734                    same or unknown then we have to construct a phi
1735                    (we collect the potential phi inputs here) */
1736                 for (p = 0; p < n_preds; ++p) {
1737                         block_info_t *pred_info = pred_block_infos[p];
1738
1739                         if (!pred_info->processed) {
1740                                 /* use node for now, it will get fixed later */
1741                                 phi_ins[p] = node;
1742                                 need_phi   = true;
1743                         } else {
1744                                 int a = find_value_in_block_info(pred_info, node);
1745
1746                                 /* must live out of predecessor */
1747                                 assert(a >= 0);
1748                                 phi_ins[p] = pred_info->assignments[a];
1749                                 /* different value from last time? then we need a phi */
1750                                 if (p > 0 && phi_ins[p-1] != phi_ins[p]) {
1751                                         need_phi = true;
1752                                 }
1753                         }
1754                 }
1755
1756                 if (need_phi) {
1757                         ir_mode *mode = get_irn_mode(node);
1758                         ir_node *phi  = be_new_Phi(block, n_preds, phi_ins, mode, cls);
1759
1760                         DB((dbg, LEVEL_3, "Create Phi %+F (for %+F) -", phi, node));
1761 #ifdef DEBUG_libfirm
1762                         {
1763                                 int pi;
1764                                 for (pi = 0; pi < n_preds; ++pi) {
1765                                         DB((dbg, LEVEL_3, " %+F", phi_ins[pi]));
1766                                 }
1767                                 DB((dbg, LEVEL_3, "\n"));
1768                         }
1769 #endif
1770                         mark_as_copy_of(phi, node);
1771                         sched_add_after(block, phi);
1772
1773                         node = phi;
1774                 } else {
1775                         allocation_info_t *info = get_allocation_info(node);
1776                         info->current_value = phi_ins[0];
1777
1778                         /* Grab 1 of the inputs we constructed (might not be the same as
1779                          * "node" as we could see the same copy of the value in all
1780                          * predecessors */
1781                         node = phi_ins[0];
1782                 }
1783
1784                 /* if the node already has a register assigned use it */
1785                 reg = arch_get_irn_register(node);
1786                 if (reg != NULL) {
1787                         use_reg(node, reg);
1788                 }
1789
1790                 /* remember that this node is live at the beginning of the block */
1791                 ir_nodeset_insert(&live_nodes, node);
1792         }
1793
1794         rbitset_alloca(forbidden_regs, n_regs);
1795
1796         /* handle phis... */
1797         assign_phi_registers(block);
1798
1799         /* all live-ins must have a register */
1800 #ifdef DEBUG_libfirm
1801         {
1802                 ir_nodeset_iterator_t  iter;
1803                 foreach_ir_nodeset(&live_nodes, node, iter) {
1804                         const arch_register_t *reg = arch_get_irn_register(node);
1805                         assert(reg != NULL);
1806                 }
1807         }
1808 #endif
1809
1810         /* assign instructions in the block */
1811         sched_foreach(block, node) {
1812                 int arity;
1813                 ir_node *value;
1814
1815                 /* phis are already assigned */
1816                 if (is_Phi(node))
1817                         continue;
1818
1819                 rewire_inputs(node);
1820
1821                 /* enforce use constraints */
1822                 rbitset_clear_all(forbidden_regs, n_regs);
1823                 enforce_constraints(&live_nodes, node, forbidden_regs);
1824
1825                 rewire_inputs(node);
1826
1827                 /* we may not use registers used for inputs for optimistic splits */
1828                 arity = get_irn_arity(node);
1829                 for (i = 0; i < arity; ++i) {
1830                         ir_node *op = get_irn_n(node, i);
1831                         const arch_register_t *reg;
1832                         if (!arch_irn_consider_in_reg_alloc(cls, op))
1833                                 continue;
1834
1835                         reg = arch_get_irn_register(op);
1836                         rbitset_set(forbidden_regs, arch_register_get_index(reg));
1837                 }
1838
1839                 /* free registers of values last used at this instruction */
1840                 free_last_uses(&live_nodes, node);
1841
1842                 /* assign output registers */
1843                 be_foreach_definition_(node, cls, value,
1844                         assign_reg(block, value, forbidden_regs);
1845                 );
1846         }
1847
1848         ir_nodeset_destroy(&live_nodes);
1849         assignments = NULL;
1850
1851         block_info->processed = true;
1852
1853         /* permute values at end of predecessor blocks in case of phi-nodes */
1854         if (n_preds > 1) {
1855                 int p;
1856                 for (p = 0; p < n_preds; ++p) {
1857                         add_phi_permutations(block, p);
1858                 }
1859         }
1860
1861         /* if we have exactly 1 successor then we might be able to produce phi
1862            copies now */
1863         if (get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) == 1) {
1864                 const ir_edge_t *edge
1865                         = get_irn_out_edge_first_kind(block, EDGE_KIND_BLOCK);
1866                 ir_node      *succ      = get_edge_src_irn(edge);
1867                 int           p         = get_edge_src_pos(edge);
1868                 block_info_t *succ_info = get_block_info(succ);
1869
1870                 if (succ_info->processed) {
1871                         add_phi_permutations(succ, p);
1872                 }
1873         }
1874 }
1875
1876 typedef struct block_costs_t block_costs_t;
1877 struct block_costs_t {
1878         float costs;   /**< costs of the block */
1879         int   dfs_num; /**< depth first search number (to detect backedges) */
1880 };
1881
1882 static int cmp_block_costs(const void *d1, const void *d2)
1883 {
1884         const ir_node       * const *block1 = (const ir_node**)d1;
1885         const ir_node       * const *block2 = (const ir_node**)d2;
1886         const block_costs_t *info1  = (const block_costs_t*)get_irn_link(*block1);
1887         const block_costs_t *info2  = (const block_costs_t*)get_irn_link(*block2);
1888         return QSORT_CMP(info2->costs, info1->costs);
1889 }
1890
1891 static void determine_block_order(void)
1892 {
1893         size_t    p;
1894         ir_node **blocklist = be_get_cfgpostorder(irg);
1895         size_t    n_blocks  = ARR_LEN(blocklist);
1896         int       dfs_num   = 0;
1897         pdeq     *worklist  = new_pdeq();
1898         ir_node **order     = XMALLOCN(ir_node*, n_blocks);
1899         size_t    order_p   = 0;
1900
1901         /* clear block links... */
1902         for (p = 0; p < n_blocks; ++p) {
1903                 ir_node *block = blocklist[p];
1904                 set_irn_link(block, NULL);
1905         }
1906
1907         /* walk blocks in reverse postorder, the costs for each block are the
1908          * sum of the costs of its predecessors (excluding the costs on backedges
1909          * which we can't determine) */
1910         for (p = n_blocks; p > 0;) {
1911                 block_costs_t *cost_info;
1912                 ir_node *block = blocklist[--p];
1913
1914                 float execfreq   = (float)get_block_execfreq(execfreqs, block);
1915                 float costs      = execfreq;
1916                 int   n_cfgpreds = get_Block_n_cfgpreds(block);
1917                 int   p2;
1918                 for (p2 = 0; p2 < n_cfgpreds; ++p2) {
1919                         ir_node       *pred_block = get_Block_cfgpred_block(block, p2);
1920                         block_costs_t *pred_costs = (block_costs_t*)get_irn_link(pred_block);
1921                         /* we don't have any info for backedges */
1922                         if (pred_costs == NULL)
1923                                 continue;
1924                         costs += pred_costs->costs;
1925                 }
1926
1927                 cost_info          = OALLOCZ(&obst, block_costs_t);
1928                 cost_info->costs   = costs;
1929                 cost_info->dfs_num = dfs_num++;
1930                 set_irn_link(block, cost_info);
1931         }
1932
1933         /* sort array by block costs */
1934         qsort(blocklist, n_blocks, sizeof(blocklist[0]), cmp_block_costs);
1935
1936         ir_reserve_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1937         inc_irg_block_visited(irg);
1938
1939         for (p = 0; p < n_blocks; ++p) {
1940                 ir_node       *block = blocklist[p];
1941                 if (Block_block_visited(block))
1942                         continue;
1943
1944                 /* continually add predecessors with highest costs to worklist
1945                  * (without using backedges) */
1946                 do {
1947                         block_costs_t *info       = (block_costs_t*)get_irn_link(block);
1948                         ir_node       *best_pred  = NULL;
1949                         float          best_costs = -1;
1950                         int            n_cfgpred  = get_Block_n_cfgpreds(block);
1951                         int            i;
1952
1953                         pdeq_putr(worklist, block);
1954                         mark_Block_block_visited(block);
1955                         for (i = 0; i < n_cfgpred; ++i) {
1956                                 ir_node       *pred_block = get_Block_cfgpred_block(block, i);
1957                                 block_costs_t *pred_info  = (block_costs_t*)get_irn_link(pred_block);
1958
1959                                 /* ignore backedges */
1960                                 if (pred_info->dfs_num > info->dfs_num)
1961                                         continue;
1962
1963                                 if (info->costs > best_costs) {
1964                                         best_costs = info->costs;
1965                                         best_pred  = pred_block;
1966                                 }
1967                         }
1968                         block = best_pred;
1969                 } while (block != NULL && !Block_block_visited(block));
1970
1971                 /* now put all nodes in the worklist in our final order */
1972                 while (!pdeq_empty(worklist)) {
1973                         ir_node *pblock = (ir_node*)pdeq_getr(worklist);
1974                         assert(order_p < n_blocks);
1975                         order[order_p++] = pblock;
1976                 }
1977         }
1978         assert(order_p == n_blocks);
1979         del_pdeq(worklist);
1980
1981         ir_free_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1982
1983         DEL_ARR_F(blocklist);
1984
1985         obstack_free(&obst, NULL);
1986         obstack_init(&obst);
1987
1988         block_order   = order;
1989         n_block_order = n_blocks;
1990 }
1991
1992 /**
1993  * Run the register allocator for the current register class.
1994  */
1995 static void be_pref_alloc_cls(void)
1996 {
1997         size_t i;
1998
1999         lv = be_assure_liveness(irg);
2000         be_liveness_assure_sets(lv);
2001
2002         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
2003
2004         DB((dbg, LEVEL_2, "=== Allocating registers of %s ===\n", cls->name));
2005
2006         be_clear_links(irg);
2007
2008         irg_block_walk_graph(irg, NULL, analyze_block, NULL);
2009         if (create_congruence_classes)
2010                 combine_congruence_classes();
2011
2012         for (i = 0; i < n_block_order; ++i) {
2013                 ir_node *block = block_order[i];
2014                 allocate_coalesce_block(block, NULL);
2015         }
2016
2017         ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
2018 }
2019
2020 static void dump(int mask, ir_graph *irg, const char *suffix)
2021 {
2022         if (be_get_irg_options(irg)->dump_flags & mask)
2023                 dump_ir_graph(irg, suffix);
2024 }
2025
2026 /**
2027  * Run the spiller on the current graph.
2028  */
2029 static void spill(void)
2030 {
2031         /* make sure all nodes show their real register pressure */
2032         be_timer_push(T_RA_CONSTR);
2033         be_pre_spill_prepare_constr(irg, cls);
2034         be_timer_pop(T_RA_CONSTR);
2035
2036         dump(DUMP_RA, irg, "spillprepare");
2037
2038         /* spill */
2039         be_timer_push(T_RA_SPILL);
2040         be_do_spill(irg, cls);
2041         be_timer_pop(T_RA_SPILL);
2042
2043         be_timer_push(T_RA_SPILL_APPLY);
2044         check_for_memory_operands(irg);
2045         be_timer_pop(T_RA_SPILL_APPLY);
2046
2047         dump(DUMP_RA, irg, "spill");
2048 }
2049
2050 /**
2051  * The pref register allocator for a whole procedure.
2052  */
2053 static void be_pref_alloc(ir_graph *new_irg)
2054 {
2055         const arch_env_t *arch_env = be_get_irg_arch_env(new_irg);
2056         int   n_cls                = arch_env->n_register_classes;
2057         int   c;
2058
2059         obstack_init(&obst);
2060
2061         irg       = new_irg;
2062         execfreqs = be_get_irg_exec_freq(irg);
2063
2064         /* determine a good coloring order */
2065         determine_block_order();
2066
2067         for (c = 0; c < n_cls; ++c) {
2068                 cls = &arch_env->register_classes[c];
2069                 if (arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
2070                         continue;
2071
2072                 stat_ev_ctx_push_str("regcls", cls->name);
2073
2074                 n_regs      = arch_register_class_n_regs(cls);
2075                 normal_regs = rbitset_malloc(n_regs);
2076                 be_set_allocatable_regs(irg, cls, normal_regs);
2077
2078                 spill();
2079
2080                 /* verify schedule and register pressure */
2081                 be_timer_push(T_VERIFY);
2082                 if (be_get_irg_options(irg)->verify_option == BE_VERIFY_WARN) {
2083                         be_verify_schedule(irg);
2084                         be_verify_register_pressure(irg, cls);
2085                 } else if (be_get_irg_options(irg)->verify_option == BE_VERIFY_ASSERT) {
2086                         assert(be_verify_schedule(irg) && "Schedule verification failed");
2087                         assert(be_verify_register_pressure(irg, cls)
2088                                 && "Register pressure verification failed");
2089                 }
2090                 be_timer_pop(T_VERIFY);
2091
2092                 be_timer_push(T_RA_COLOR);
2093                 be_pref_alloc_cls();
2094                 be_timer_pop(T_RA_COLOR);
2095
2096                 /* we most probably constructed new Phis so liveness info is invalid
2097                  * now */
2098                 /* TODO: test liveness_introduce */
2099                 be_liveness_invalidate(lv);
2100                 free(normal_regs);
2101
2102                 stat_ev_ctx_pop("regcls");
2103         }
2104
2105         be_timer_push(T_RA_SPILL_APPLY);
2106         be_abi_fix_stack_nodes(irg);
2107         be_timer_pop(T_RA_SPILL_APPLY);
2108
2109         be_timer_push(T_VERIFY);
2110         if (be_get_irg_options(irg)->verify_option == BE_VERIFY_WARN) {
2111                 be_verify_register_allocation(irg);
2112         } else if (be_get_irg_options(irg)->verify_option == BE_VERIFY_ASSERT) {
2113                 assert(be_verify_register_allocation(irg)
2114                        && "Register allocation invalid");
2115         }
2116         be_timer_pop(T_VERIFY);
2117
2118         obstack_free(&obst, NULL);
2119 }
2120
2121 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_pref_alloc)
2122 void be_init_pref_alloc(void)
2123 {
2124         static be_ra_t be_ra_pref = {
2125                 be_pref_alloc,
2126         };
2127         lc_opt_entry_t *be_grp              = lc_opt_get_grp(firm_opt_get_root(), "be");
2128         lc_opt_entry_t *prefalloc_group = lc_opt_get_grp(be_grp, "prefalloc");
2129         lc_opt_add_table(prefalloc_group, options);
2130
2131         be_register_allocator("pref", &be_ra_pref);
2132         FIRM_DBG_REGISTER(dbg, "firm.be.prefalloc");
2133 }