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 "irtools.h"
  58 #include "util.h"
  59 #include "obst.h"
  60 #include "raw_bitset.h"
  61 #include "unionfind.h"
  62 #include "pdeq.h"
  63 #include "hungarian.h"
  64
  65 #include "beabi.h"
  66 #include "bechordal_t.h"
  67 #include "be.h"
  68 #include "beirg.h"
  69 #include "belive_t.h"
  70 #include "bemodule.h"
  71 #include "benode.h"
  72 #include "bera.h"
  73 #include "besched.h"
  74 #include "bespill.h"
  75 #include "bespillutil.h"
  76 #include "beverify.h"
  77 #include "beutil.h"
  78 #include "bestack.h"
  79
  80 #define USE_FACTOR                     1.0f
  81 #define DEF_FACTOR                     1.0f
  82 #define NEIGHBOR_FACTOR                0.2f
  83 #define AFF_SHOULD_BE_SAME             0.5f
  84 #define AFF_PHI                        1.0f
  85 #define SPLIT_DELTA                    1.0f
  86 #define MAX_OPTIMISTIC_SPLIT_RECURSION 0
  87
  88 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
  89
  90 static struct obstack               obst;
  91 static ir_graph                    *irg;
  92 static const arch_register_class_t *cls;
  93 static be_lv_t                     *lv;
  94 static const ir_exec_freq          *execfreqs;
  95 static unsigned                     n_regs;
  96 static unsigned                    *normal_regs;
  97 static int                         *congruence_classes;
  98 static ir_node                    **block_order;
  99 static size_t                       n_block_order;
 100 static int                          create_preferences        = true;
 101 static int                          create_congruence_classes = true;
 102 static int                          propagate_phi_registers   = true;
 103
 104 static const lc_opt_table_entry_t options[] = {
 105         LC_OPT_ENT_BOOL("prefs", "use preference based coloring", &create_preferences),
 106         LC_OPT_ENT_BOOL("congruences", "create congruence classes", &create_congruence_classes),
 107         LC_OPT_ENT_BOOL("prop_phi", "propagate phi registers", &propagate_phi_registers),
 108         LC_OPT_LAST
 109 };
 110
 111 /** currently active assignments (while processing a basic block)
 112  * maps registers to values(their current copies) */
 113 static ir_node **assignments;
 114
 115 /**
 116  * allocation information: last_uses, register preferences
 117  * the information is per firm-node.
 118  */
 119 struct allocation_info_t {
 120         unsigned  last_uses[2];   /**< bitset indicating last uses (input pos) */
 121         ir_node  *current_value;  /**< copy of the value that should be used */
 122         ir_node  *original_value; /**< for copies point to original value */
 123         float     prefs[0];       /**< register preferences */
 124 };
 125 typedef struct allocation_info_t allocation_info_t;
 126
 127 /** helper datastructure used when sorting register preferences */
 128 struct reg_pref_t {
 129         unsigned num;
 130         float    pref;
 131 };
 132 typedef struct reg_pref_t reg_pref_t;
 133
 134 /** per basic-block information */
 135 struct block_info_t {
 136         bool     processed;       /**< indicate whether block is processed */
 137         ir_node *assignments[0];  /**< register assignments at end of block */
 138 };
 139 typedef struct block_info_t block_info_t;
 140
 141 /**
 142  * Get the allocation info for a node.
 143  * The info is allocated on the first visit of a node.
 144  */
 145 static allocation_info_t *get_allocation_info(ir_node *node)
 146 {
 147         allocation_info_t *info = (allocation_info_t*)get_irn_link(node);
 148         if (info == NULL) {
 149                 info = OALLOCFZ(&obst, allocation_info_t, prefs, n_regs);
 150                 info->current_value  = node;
 151                 info->original_value = node;
 152                 set_irn_link(node, info);
 153         }
 154
 155         return info;
 156 }
 157
 158 static allocation_info_t *try_get_allocation_info(const ir_node *node)
 159 {
 160         return (allocation_info_t*) get_irn_link(node);
 161 }
 162
 163 /**
 164  * Get allocation information for a basic block
 165  */
 166 static block_info_t *get_block_info(ir_node *block)
 167 {
 168         block_info_t *info = (block_info_t*)get_irn_link(block);
 169
 170         assert(is_Block(block));
 171         if (info == NULL) {
 172                 info = OALLOCFZ(&obst, block_info_t, assignments, n_regs);
 173                 set_irn_link(block, info);
 174         }
 175
 176         return info;
 177 }
 178
 179 /**
 180  * Link the allocation info of a node to a copy.
 181  * Afterwards, both nodes uses the same allocation info.
 182  * Copy must not have an allocation info assigned yet.
 183  *
 184  * @param copy   the node that gets the allocation info assigned
 185  * @param value  the original node
 186  */
 187 static void mark_as_copy_of(ir_node *copy, ir_node *value)
 188 {
 189         ir_node           *original;
 190         allocation_info_t *info      = get_allocation_info(value);
 191         allocation_info_t *copy_info = get_allocation_info(copy);
 192
 193         /* find original value */
 194         original = info->original_value;
 195         if (original != value) {
 196                 info = get_allocation_info(original);
 197         }
 198
 199         assert(info->original_value == original);
 200         info->current_value = copy;
 201
 202         /* the copy should not be linked to something else yet */
 203         assert(copy_info->original_value == copy);
 204         copy_info->original_value = original;
 205
 206         /* copy over allocation preferences */
 207         memcpy(copy_info->prefs, info->prefs, n_regs * sizeof(copy_info->prefs[0]));
 208 }
 209
 210 /**
 211  * Calculate the penalties for every register on a node and its live neighbors.
 212  *
 213  * @param live_nodes  the set of live nodes at the current position, may be NULL
 214  * @param penalty     the penalty to subtract from
 215  * @param limited     a raw bitset containing the limited set for the node
 216  * @param node        the node
 217  */
 218 static void give_penalties_for_limits(const ir_nodeset_t *live_nodes,
 219                                       float penalty, const unsigned* limited,
 220                                       ir_node *node)
 221 {
 222         ir_nodeset_iterator_t iter;
 223         unsigned              r;
 224         size_t                n_allowed;
 225         allocation_info_t     *info = get_allocation_info(node);
 226         ir_node               *neighbor;
 227
 228         /* give penalty for all forbidden regs */
 229         for (r = 0; r < n_regs; ++r) {
 230                 if (rbitset_is_set(limited, r))
 231                         continue;
 232
 233                 info->prefs[r] -= penalty;
 234         }
 235
 236         /* all other live values should get a penalty for allowed regs */
 237         if (live_nodes == NULL)
 238                 return;
 239
 240         penalty   *= NEIGHBOR_FACTOR;
 241         n_allowed  = rbitset_popcount(limited, n_regs);
 242         if (n_allowed > 1) {
 243                 /* only create a very weak penalty if multiple regs are allowed */
 244                 penalty = (penalty * 0.8f) / n_allowed;
 245         }
 246         foreach_ir_nodeset(live_nodes, neighbor, iter) {
 247                 allocation_info_t *neighbor_info;
 248
 249                 /* TODO: if op is used on multiple inputs we might not do a
 250                  * continue here */
 251                 if (neighbor == node)
 252                         continue;
 253
 254                 neighbor_info = get_allocation_info(neighbor);
 255                 for (r = 0; r < n_regs; ++r) {
 256                         if (!rbitset_is_set(limited, r))
 257                                 continue;
 258
 259                         neighbor_info->prefs[r] -= penalty;
 260                 }
 261         }
 262 }
 263
 264 /**
 265  * Calculate the preferences of a definition for the current register class.
 266  * If the definition uses a limited set of registers, reduce the preferences
 267  * for the limited register on the node and its neighbors.
 268  *
 269  * @param live_nodes  the set of live nodes at the current node
 270  * @param weight      the weight
 271  * @param node        the current node
 272  */
 273 static void check_defs(const ir_nodeset_t *live_nodes, float weight,
 274                        ir_node *node)
 275 {
 276         const arch_register_req_t *req = arch_get_irn_register_req(node);
 277         if (req->type & arch_register_req_type_limited) {
 278                 const unsigned *limited = req->limited;
 279                 float           penalty = weight * DEF_FACTOR;
 280                 give_penalties_for_limits(live_nodes, penalty, limited, node);
 281         }
 282
 283         if (req->type & arch_register_req_type_should_be_same) {
 284                 ir_node           *insn  = skip_Proj(node);
 285                 allocation_info_t *info  = get_allocation_info(node);
 286                 int                arity = get_irn_arity(insn);
 287                 int                i;
 288
 289                 float factor = 1.0f / rbitset_popcount(&req->other_same, arity);
 290                 for (i = 0; i < arity; ++i) {
 291                         ir_node           *op;
 292                         unsigned           r;
 293                         allocation_info_t *op_info;
 294
 295                         if (!rbitset_is_set(&req->other_same, i))
 296                                 continue;
 297
 298                         op = get_irn_n(insn, i);
 299
 300                         /* if we the value at the should_be_same input doesn't die at the
 301                          * node, then it is no use to propagate the constraints (since a
 302                          * copy will emerge anyway) */
 303                         if (ir_nodeset_contains(live_nodes, op))
 304                                 continue;
 305
 306                         op_info = get_allocation_info(op);
 307                         for (r = 0; r < n_regs; ++r) {
 308                                 op_info->prefs[r] += info->prefs[r] * factor;
 309                         }
 310                 }
 311         }
 312 }
 313
 314 /**
 315  * Walker: Runs an a block calculates the preferences for any
 316  * node and every register from the considered register class.
 317  */
 318 static void analyze_block(ir_node *block, void *data)
 319 {
 320         float         weight = (float)get_block_execfreq(execfreqs, block);
 321         ir_nodeset_t  live_nodes;
 322         ir_node      *node;
 323         (void) data;
 324
 325         ir_nodeset_init(&live_nodes);
 326         be_liveness_end_of_block(lv, cls, block, &live_nodes);
 327
 328         sched_foreach_reverse(block, node) {
 329                 allocation_info_t *info;
 330                 int                i;
 331                 int                arity;
 332
 333                 if (is_Phi(node))
 334                         break;
 335
 336                 if (create_preferences) {
 337                         ir_node *value;
 338                         be_foreach_definition(node, cls, value,
 339                                 check_defs(&live_nodes, weight, value);
 340                         );
 341                 }
 342
 343                 /* mark last uses */
 344                 arity = get_irn_arity(node);
 345
 346                 /* the allocation info node currently only uses 1 unsigned value
 347                    to mark last used inputs. So we will fail for a node with more than
 348                    32 inputs. */
 349                 if (arity >= (int) sizeof(info->last_uses) * 8) {
 350                         panic("Node with more than %d inputs not supported yet",
 351                                         (int) sizeof(info->last_uses) * 8);
 352                 }
 353
 354                 info = get_allocation_info(node);
 355                 for (i = 0; i < arity; ++i) {
 356                         ir_node                   *op  = get_irn_n(node, i);
 357                         const arch_register_req_t *req = arch_get_irn_register_req(op);
 358                         if (req->cls != cls)
 359                                 continue;
 360
 361                         /* last usage of a value? */
 362                         if (!ir_nodeset_contains(&live_nodes, op)) {
 363                                 rbitset_set(info->last_uses, i);
 364                         }
 365                 }
 366
 367                 be_liveness_transfer(cls, node, &live_nodes);
 368
 369                 if (create_preferences) {
 370                         /* update weights based on usage constraints */
 371                         for (i = 0; i < arity; ++i) {
 372                                 const arch_register_req_t *req;
 373                                 const unsigned            *limited;
 374                                 ir_node                   *op = get_irn_n(node, i);
 375
 376                                 if (!arch_irn_consider_in_reg_alloc(cls, op))
 377                                         continue;
 378
 379                                 req = arch_get_irn_register_req_in(node, i);
 380                                 if (!(req->type & arch_register_req_type_limited))
 381                                         continue;
 382
 383                                 limited = req->limited;
 384                                 give_penalties_for_limits(&live_nodes, weight * USE_FACTOR,
 385                                                           limited, op);
 386                         }
 387                 }
 388         }
 389
 390         ir_nodeset_destroy(&live_nodes);
 391 }
 392
 393 static void congruence_def(ir_nodeset_t *live_nodes, const ir_node *node)
 394 {
 395         const arch_register_req_t *req = arch_get_irn_register_req(node);
 396
 397         /* should be same constraint? */
 398         if (req->type & arch_register_req_type_should_be_same) {
 399                 const ir_node *insn  = skip_Proj_const(node);
 400                 int      arity = get_irn_arity(insn);
 401                 int      i;
 402                 unsigned node_idx = get_irn_idx(node);
 403                 node_idx          = uf_find(congruence_classes, node_idx);
 404
 405                 for (i = 0; i < arity; ++i) {
 406                         ir_node               *live;
 407                         ir_node               *op;
 408                         int                    op_idx;
 409                         ir_nodeset_iterator_t  iter;
 410                         bool                   interferes = false;
 411
 412                         if (!rbitset_is_set(&req->other_same, i))
 413                                 continue;
 414
 415                         op     = get_irn_n(insn, i);
 416                         op_idx = get_irn_idx(op);
 417                         op_idx = uf_find(congruence_classes, op_idx);
 418
 419                         /* do we interfere with the value */
 420                         foreach_ir_nodeset(live_nodes, live, iter) {
 421                                 int lv_idx = get_irn_idx(live);
 422                                 lv_idx     = uf_find(congruence_classes, lv_idx);
 423                                 if (lv_idx == op_idx) {
 424                                         interferes = true;
 425                                         break;
 426                                 }
 427                         }
 428                         /* don't put in same affinity class if we interfere */
 429                         if (interferes)
 430                                 continue;
 431
 432                         node_idx = uf_union(congruence_classes, node_idx, op_idx);
 433                         DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
 434                             node, op));
 435                         /* one should_be_same is enough... */
 436                         break;
 437                 }
 438         }
 439 }
 440
 441 static void create_congruence_class(ir_node *block, void *data)
 442 {
 443         ir_nodeset_t  live_nodes;
 444         ir_node      *node;
 445
 446         (void) data;
 447         ir_nodeset_init(&live_nodes);
 448         be_liveness_end_of_block(lv, cls, block, &live_nodes);
 449
 450         /* check should be same constraints */
 451         sched_foreach_reverse(block, node) {
 452                 ir_node *value;
 453                 if (is_Phi(node))
 454                         break;
 455
 456                 be_foreach_definition(node, cls, value,
 457                         congruence_def(&live_nodes, value);
 458                 );
 459                 be_liveness_transfer(cls, node, &live_nodes);
 460         }
 461
 462         /* check phi congruence classes */
 463         sched_foreach_reverse_from(node, node) {
 464                 int i;
 465                 int arity;
 466                 int node_idx;
 467                 assert(is_Phi(node));
 468
 469                 if (!arch_irn_consider_in_reg_alloc(cls, node))
 470                         continue;
 471
 472                 node_idx = get_irn_idx(node);
 473                 node_idx = uf_find(congruence_classes, node_idx);
 474
 475                 arity = get_irn_arity(node);
 476                 for (i = 0; i < arity; ++i) {
 477                         bool                  interferes = false;
 478                         ir_nodeset_iterator_t iter;
 479                         unsigned              r;
 480                         int                   old_node_idx;
 481                         ir_node              *live;
 482                         ir_node              *phi;
 483                         allocation_info_t    *head_info;
 484                         allocation_info_t    *other_info;
 485                         ir_node              *op     = get_Phi_pred(node, i);
 486                         int                   op_idx = get_irn_idx(op);
 487                         op_idx = uf_find(congruence_classes, op_idx);
 488
 489                         /* do we interfere with the value */
 490                         foreach_ir_nodeset(&live_nodes, live, iter) {
 491                                 int lv_idx = get_irn_idx(live);
 492                                 lv_idx     = uf_find(congruence_classes, lv_idx);
 493                                 if (lv_idx == op_idx) {
 494                                         interferes = true;
 495                                         break;
 496                                 }
 497                         }
 498                         /* don't put in same affinity class if we interfere */
 499                         if (interferes)
 500                                 continue;
 501                         /* any other phi has the same input? */
 502                         sched_foreach(block, phi) {
 503                                 ir_node *oop;
 504                                 int      oop_idx;
 505                                 if (!is_Phi(phi))
 506                                         break;
 507                                 if (!arch_irn_consider_in_reg_alloc(cls, phi))
 508                                         continue;
 509                                 oop = get_Phi_pred(phi, i);
 510                                 if (oop == op)
 511                                         continue;
 512                                 oop_idx = get_irn_idx(oop);
 513                                 oop_idx = uf_find(congruence_classes, oop_idx);
 514                                 if (oop_idx == op_idx) {
 515                                         interferes = true;
 516                                         break;
 517                                 }
 518                         }
 519                         if (interferes)
 520                                 continue;
 521
 522                         /* merge the 2 congruence classes and sum up their preferences */
 523                         old_node_idx = node_idx;
 524                         node_idx = uf_union(congruence_classes, node_idx, op_idx);
 525                         DB((dbg, LEVEL_3, "Merge %+F and %+F congruence classes\n",
 526                             node, op));
 527
 528                         old_node_idx = node_idx == old_node_idx ? op_idx : old_node_idx;
 529                         head_info  = get_allocation_info(get_idx_irn(irg, node_idx));
 530                         other_info = get_allocation_info(get_idx_irn(irg, old_node_idx));
 531                         for (r = 0; r < n_regs; ++r) {
 532                                 head_info->prefs[r] += other_info->prefs[r];
 533                         }
 534                 }
 535         }
 536 }
 537
 538 static void set_congruence_prefs(ir_node *node, void *data)
 539 {
 540         allocation_info_t *info;
 541         allocation_info_t *head_info;
 542         unsigned node_idx = get_irn_idx(node);
 543         unsigned node_set = uf_find(congruence_classes, node_idx);
 544
 545         (void) data;
 546
 547         /* head of congruence class or not in any class */
 548         if (node_set == node_idx)
 549                 return;
 550
 551         if (!arch_irn_consider_in_reg_alloc(cls, node))
 552                 return;
 553
 554         head_info = get_allocation_info(get_idx_irn(irg, node_set));
 555         info      = get_allocation_info(node);
 556
 557         memcpy(info->prefs, head_info->prefs, n_regs * sizeof(info->prefs[0]));
 558 }
 559
 560 static void combine_congruence_classes(void)
 561 {
 562         size_t n = get_irg_last_idx(irg);
 563         congruence_classes = XMALLOCN(int, n);
 564         uf_init(congruence_classes, n);
 565
 566         /* create congruence classes */
 567         irg_block_walk_graph(irg, create_congruence_class, NULL, NULL);
 568         /* merge preferences */
 569         irg_walk_graph(irg, set_congruence_prefs, NULL, NULL);
 570         free(congruence_classes);
 571 }
 572
 573
 574
 575 /**
 576  * Assign register reg to the given node.
 577  *
 578  * @param node  the node
 579  * @param reg   the register
 580  */
 581 static void use_reg(ir_node *node, const arch_register_t *reg)
 582 {
 583         unsigned r = arch_register_get_index(reg);
 584         assignments[r] = node;
 585         arch_set_irn_register(node, reg);
 586 }
 587
 588 static void free_reg_of_value(ir_node *node)
 589 {
 590         const arch_register_t *reg;
 591         unsigned               r;
 592
 593         if (!arch_irn_consider_in_reg_alloc(cls, node))
 594                 return;
 595
 596         reg        = arch_get_irn_register(node);
 597         r          = arch_register_get_index(reg);
 598         /* assignment->value may be NULL if a value is used at 2 inputs
 599            so it gets freed twice. */
 600         assert(assignments[r] == node || assignments[r] == NULL);
 601         assignments[r] = NULL;
 602 }
 603
 604 /**
 605  * Compare two register preferences in decreasing order.
 606  */
 607 static int compare_reg_pref(const void *e1, const void *e2)
 608 {
 609         const reg_pref_t *rp1 = (const reg_pref_t*) e1;
 610         const reg_pref_t *rp2 = (const reg_pref_t*) e2;
 611         if (rp1->pref < rp2->pref)
 612                 return 1;
 613         if (rp1->pref > rp2->pref)
 614                 return -1;
 615         return 0;
 616 }
 617
 618 static void fill_sort_candidates(reg_pref_t *regprefs,
 619                                  const allocation_info_t *info)
 620 {
 621         unsigned r;
 622
 623         for (r = 0; r < n_regs; ++r) {
 624                 float pref = info->prefs[r];
 625                 regprefs[r].num  = r;
 626                 regprefs[r].pref = pref;
 627         }
 628         /* TODO: use a stable sort here to avoid unnecessary register jumping */
 629         qsort(regprefs, n_regs, sizeof(regprefs[0]), compare_reg_pref);
 630 }
 631
 632 static bool try_optimistic_split(ir_node *to_split, ir_node *before,
 633                                  float pref, float pref_delta,
 634                                  unsigned *forbidden_regs, int recursion)
 635 {
 636         const arch_register_t *from_reg;
 637         const arch_register_t *reg;
 638         ir_node               *original_insn;
 639         ir_node               *block;
 640         ir_node               *copy;
 641         unsigned               r = 0;
 642         unsigned               from_r;
 643         unsigned               i;
 644         allocation_info_t     *info = get_allocation_info(to_split);
 645         reg_pref_t            *prefs;
 646         float                  delta = 0;
 647         float                  split_threshold;
 648
 649         (void) pref;
 650
 651         /* stupid hack: don't optimisticallt split don't spill nodes...
 652          * (so we don't split away the values produced because of
 653          *  must_be_different constraints) */
 654         original_insn = skip_Proj(info->original_value);
 655         if (arch_get_irn_flags(original_insn) & arch_irn_flags_dont_spill)
 656                 return false;
 657
 658         from_reg        = arch_get_irn_register(to_split);
 659         from_r          = arch_register_get_index(from_reg);
 660         block           = get_nodes_block(before);
 661         split_threshold = (float)get_block_execfreq(execfreqs, block) * SPLIT_DELTA;
 662
 663         if (pref_delta < split_threshold*0.5)
 664                 return false;
 665
 666         /* find the best free position where we could move to */
 667         prefs = ALLOCAN(reg_pref_t, n_regs);
 668         fill_sort_candidates(prefs, info);
 669         for (i = 0; i < n_regs; ++i) {
 670                 float apref;
 671                 float apref_delta;
 672                 bool  res;
 673                 bool  old_source_state;
 674
 675                 /* we need a normal register which is not an output register
 676                    an different from the current register of to_split */
 677                 r = prefs[i].num;
 678                 if (!rbitset_is_set(normal_regs, r))
 679                         continue;
 680                 if (rbitset_is_set(forbidden_regs, r))
 681                         continue;
 682                 if (r == from_r)
 683                         continue;
 684
 685                 /* is the split worth it? */
 686                 delta = pref_delta + prefs[i].pref;
 687                 if (delta < split_threshold) {
 688                         DB((dbg, LEVEL_3, "Not doing optimistical split of %+F (depth %d), win %f too low\n",
 689                                 to_split, recursion, delta));
 690                         return false;
 691                 }
 692
 693                 /* if the register is free then we can do the split */
 694                 if (assignments[r] == NULL)
 695                         break;
 696
 697                 /* otherwise we might try recursively calling optimistic_split */
 698                 if (recursion+1 > MAX_OPTIMISTIC_SPLIT_RECURSION)
 699                         continue;
 700
 701                 apref        = prefs[i].pref;
 702                 apref_delta  = i+1 < n_regs ? apref - prefs[i+1].pref : 0;
 703                 apref_delta += pref_delta - split_threshold;
 704
 705                 /* our source register isn't a useful destination for recursive
 706                    splits */
 707                 old_source_state = rbitset_is_set(forbidden_regs, from_r);
 708                 rbitset_set(forbidden_regs, from_r);
 709                 /* try recursive split */
 710                 res = try_optimistic_split(assignments[r], before, apref,
 711                                            apref_delta, forbidden_regs, recursion+1);
 712                 /* restore our destination */
 713                 if (old_source_state) {
 714                         rbitset_set(forbidden_regs, from_r);
 715                 } else {
 716                         rbitset_clear(forbidden_regs, from_r);
 717                 }
 718
 719                 if (res)
 720                         break;
 721         }
 722         if (i >= n_regs)
 723                 return false;
 724
 725         reg  = arch_register_for_index(cls, r);
 726         copy = be_new_Copy(block, to_split);
 727         mark_as_copy_of(copy, to_split);
 728         /* hacky, but correct here */
 729         if (assignments[arch_register_get_index(from_reg)] == to_split)
 730                 free_reg_of_value(to_split);
 731         use_reg(copy, reg);
 732         sched_add_before(before, copy);
 733
 734         DB((dbg, LEVEL_3,
 735             "Optimistic live-range split %+F move %+F(%s) -> %s before %+F (win %f, depth %d)\n",
 736             copy, to_split, from_reg->name, reg->name, before, delta, recursion));
 737         return true;
 738 }
 739
 740 /**
 741  * Determine and assign a register for node @p node
 742  */
 743 static void assign_reg(const ir_node *block, ir_node *node,
 744                        unsigned *forbidden_regs)
 745 {
 746         const arch_register_t     *final_reg;
 747         allocation_info_t         *info;
 748         const arch_register_req_t *req;
 749         reg_pref_t                *reg_prefs;
 750         ir_node                   *in_node;
 751         unsigned                   r;
 752         const unsigned            *allowed_regs;
 753         unsigned                   final_reg_index = 0;
 754
 755         assert(!is_Phi(node));
 756         /* preassigned register? */
 757         final_reg = arch_get_irn_register(node);
 758         if (final_reg != NULL) {
 759                 DB((dbg, LEVEL_2, "Preassignment %+F -> %s\n", node, final_reg->name));
 760                 use_reg(node, final_reg);
 761                 return;
 762         }
 763
 764         req = arch_get_irn_register_req(node);
 765         /* ignore reqs must be preassigned */
 766         assert (! (req->type & arch_register_req_type_ignore));
 767
 768         /* give should_be_same boni */
 769         info    = get_allocation_info(node);
 770         in_node = skip_Proj(node);
 771         if (req->type & arch_register_req_type_should_be_same) {
 772                 float weight = (float)get_block_execfreq(execfreqs, block);
 773                 int   arity  = get_irn_arity(in_node);
 774                 int   i;
 775
 776                 assert(arity <= (int) sizeof(req->other_same) * 8);
 777                 for (i = 0; i < arity; ++i) {
 778                         ir_node               *in;
 779                         const arch_register_t *reg;
 780                         unsigned               reg_index;
 781                         if (!rbitset_is_set(&req->other_same, i))
 782                                 continue;
 783
 784                         in  = get_irn_n(in_node, i);
 785                         reg = arch_get_irn_register(in);
 786                         assert(reg != NULL);
 787                         reg_index = arch_register_get_index(reg);
 788
 789                         /* if the value didn't die here then we should not propagate the
 790                          * should_be_same info */
 791                         if (assignments[reg_index] == in)
 792                                 continue;
 793
 794                         info->prefs[reg_index] += weight * AFF_SHOULD_BE_SAME;
 795                 }
 796         }
 797
 798         /* create list of register candidates and sort by their preference */
 799         DB((dbg, LEVEL_2, "Candidates for %+F:", node));
 800         reg_prefs = ALLOCAN(reg_pref_t, n_regs);
 801         fill_sort_candidates(reg_prefs, info);
 802         for (r = 0; r < n_regs; ++r) {
 803                 unsigned num = reg_prefs[r].num;
 804                 const arch_register_t *reg;
 805
 806                 if (!rbitset_is_set(normal_regs, num))
 807                         continue;
 808                 reg = arch_register_for_index(cls, num);
 809                 DB((dbg, LEVEL_2, " %s(%f)", reg->name, reg_prefs[r].pref));
 810         }
 811         DB((dbg, LEVEL_2, "\n"));
 812
 813         allowed_regs = normal_regs;
 814         if (req->type & arch_register_req_type_limited) {
 815                 allowed_regs = req->limited;
 816         }
 817
 818         for (r = 0; r < n_regs; ++r) {
 819                 float   pref, delta;
 820                 ir_node *before;
 821                 bool    res;
 822
 823                 final_reg_index = reg_prefs[r].num;
 824                 if (!rbitset_is_set(allowed_regs, final_reg_index))
 825                         continue;
 826                 /* alignment constraint? */
 827                 if (req->width > 1 && (req->type & arch_register_req_type_aligned)
 828                                 && (final_reg_index % req->width) != 0)
 829                         continue;
 830
 831                 if (assignments[final_reg_index] == NULL)
 832                         break;
 833                 pref   = reg_prefs[r].pref;
 834                 delta  = r+1 < n_regs ? pref - reg_prefs[r+1].pref : 0;
 835                 before = skip_Proj(node);
 836                 res    = try_optimistic_split(assignments[final_reg_index], before,
 837                                               pref, delta, forbidden_regs, 0);
 838                 if (res)
 839                         break;
 840         }
 841         if (r >= n_regs) {
 842                 /* the common reason to hit this panic is when 1 of your nodes is not
 843                  * register pressure faithful */
 844                 panic("No register left for %+F\n", node);
 845         }
 846
 847         final_reg = arch_register_for_index(cls, final_reg_index);
 848         DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, final_reg->name));
 849         use_reg(node, final_reg);
 850 }
 851
 852 /**
 853  * Add an permutation in front of a node and change the assignments
 854  * due to this permutation.
 855  *
 856  * To understand this imagine a permutation like this:
 857  *
 858  * 1 -> 2
 859  * 2 -> 3
 860  * 3 -> 1, 5
 861  * 4 -> 6
 862  * 5
 863  * 6
 864  * 7 -> 7
 865  *
 866  * First we count how many destinations a single value has. At the same time
 867  * we can be sure that each destination register has at most 1 source register
 868  * (it can have 0 which means we don't care what value is in it).
 869  * We ignore all fulfilled permuations (like 7->7)
 870  * In a first pass we create as much copy instructions as possible as they
 871  * are generally cheaper than exchanges. We do this by counting into how many
 872  * destinations a register has to be copied (in the example it's 2 for register
 873  * 3, or 1 for the registers 1,2,4 and 7).
 874  * We can then create a copy into every destination register when the usecount
 875  * of that register is 0 (= noone else needs the value in the register).
 876  *
 877  * After this step we should only have cycles left. We implement a cyclic
 878  * permutation of n registers with n-1 transpositions.
 879  *
 880  * @param live_nodes   the set of live nodes, updated due to live range split
 881  * @param before       the node before we add the permutation
 882  * @param permutation  the permutation array indices are the destination
 883  *                     registers, the values in the array are the source
 884  *                     registers.
 885  */
 886 static void permute_values(ir_nodeset_t *live_nodes, ir_node *before,
 887                            unsigned *permutation)
 888 {
 889         unsigned  *n_used = ALLOCANZ(unsigned, n_regs);
 890         ir_node   *block;
 891         unsigned   r;
 892
 893         /* determine how often each source register needs to be read */
 894         for (r = 0; r < n_regs; ++r) {
 895                 unsigned  old_reg = permutation[r];
 896                 ir_node  *value;
 897
 898                 value = assignments[old_reg];
 899                 if (value == NULL) {
 900                         /* nothing to do here, reg is not live. Mark it as fixpoint
 901                          * so we ignore it in the next steps */
 902                         permutation[r] = r;
 903                         continue;
 904                 }
 905
 906                 ++n_used[old_reg];
 907         }
 908
 909         block = get_nodes_block(before);
 910
 911         /* step1: create copies where immediately possible */
 912         for (r = 0; r < n_regs; /* empty */) {
 913                 ir_node *copy;
 914                 ir_node *src;
 915                 const arch_register_t *reg;
 916                 unsigned               old_r = permutation[r];
 917
 918                 /* - no need to do anything for fixed points.
 919                    - we can't copy if the value in the dest reg is still needed */
 920                 if (old_r == r || n_used[r] > 0) {
 921                         ++r;
 922                         continue;
 923                 }
 924
 925                 /* create a copy */
 926                 src  = assignments[old_r];
 927                 copy = be_new_Copy(block, src);
 928                 sched_add_before(before, copy);
 929                 reg = arch_register_for_index(cls, r);
 930                 DB((dbg, LEVEL_2, "Copy %+F (from %+F, before %+F) -> %s\n",
 931                     copy, src, before, reg->name));
 932                 mark_as_copy_of(copy, src);
 933                 use_reg(copy, reg);
 934
 935                 if (live_nodes != NULL) {
 936                         ir_nodeset_insert(live_nodes, copy);
 937                 }
 938
 939                 /* old register has 1 user less, permutation is resolved */
 940                 assert(arch_register_get_index(arch_get_irn_register(src)) == old_r);
 941                 permutation[r] = r;
 942
 943                 assert(n_used[old_r] > 0);
 944                 --n_used[old_r];
 945                 if (n_used[old_r] == 0) {
 946                         if (live_nodes != NULL) {
 947                                 ir_nodeset_remove(live_nodes, src);
 948                         }
 949                         free_reg_of_value(src);
 950                 }
 951
 952                 /* advance or jump back (if this copy enabled another copy) */
 953                 if (old_r < r && n_used[old_r] == 0) {
 954                         r = old_r;
 955                 } else {
 956                         ++r;
 957                 }
 958         }
 959
 960         /* at this point we only have "cycles" left which we have to resolve with
 961          * perm instructions
 962          * TODO: if we have free registers left, then we should really use copy
 963          * instructions for any cycle longer than 2 registers...
 964          * (this is probably architecture dependent, there might be archs where
 965          *  copies are preferable even for 2-cycles) */
 966
 967         /* create perms with the rest */
 968         for (r = 0; r < n_regs; /* empty */) {
 969                 const arch_register_t *reg;
 970                 unsigned  old_r = permutation[r];
 971                 unsigned  r2;
 972                 ir_node  *in[2];
 973                 ir_node  *perm;
 974                 ir_node  *proj0;
 975                 ir_node  *proj1;
 976
 977                 if (old_r == r) {
 978                         ++r;
 979                         continue;
 980                 }
 981
 982                 /* we shouldn't have copies from 1 value to multiple destinations left*/
 983                 assert(n_used[old_r] == 1);
 984
 985                 /* exchange old_r and r2; after that old_r is a fixed point */
 986                 r2 = permutation[old_r];
 987
 988                 in[0] = assignments[r2];
 989                 in[1] = assignments[old_r];
 990                 perm = be_new_Perm(cls, block, 2, in);
 991                 sched_add_before(before, perm);
 992                 DB((dbg, LEVEL_2, "Perm %+F (perm %+F,%+F, before %+F)\n",
 993                     perm, in[0], in[1], before));
 994
 995                 proj0 = new_r_Proj(perm, get_irn_mode(in[0]), 0);
 996                 mark_as_copy_of(proj0, in[0]);
 997                 reg = arch_register_for_index(cls, old_r);
 998                 use_reg(proj0, reg);
 999
1000                 proj1 = new_r_Proj(perm, get_irn_mode(in[1]), 1);
1001                 mark_as_copy_of(proj1, in[1]);
1002                 reg = arch_register_for_index(cls, r2);
1003                 use_reg(proj1, reg);
1004
1005                 /* 1 value is now in the correct register */
1006                 permutation[old_r] = old_r;
1007                 /* the source of r changed to r2 */
1008                 permutation[r] = r2;
1009
1010                 /* if we have reached a fixpoint update data structures */
1011                 if (live_nodes != NULL) {
1012                         ir_nodeset_remove(live_nodes, in[0]);
1013                         ir_nodeset_remove(live_nodes, in[1]);
1014                         ir_nodeset_remove(live_nodes, proj0);
1015                         ir_nodeset_insert(live_nodes, proj1);
1016                 }
1017         }
1018
1019 #ifdef DEBUG_libfirm
1020         /* now we should only have fixpoints left */
1021         for (r = 0; r < n_regs; ++r) {
1022                 assert(permutation[r] == r);
1023         }
1024 #endif
1025 }
1026
1027 /**
1028  * Free regs for values last used.
1029  *
1030  * @param live_nodes   set of live nodes, will be updated
1031  * @param node         the node to consider
1032  */
1033 static void free_last_uses(ir_nodeset_t *live_nodes, ir_node *node)
1034 {
1035         allocation_info_t *info      = get_allocation_info(node);
1036         const unsigned    *last_uses = info->last_uses;
1037         int                arity     = get_irn_arity(node);
1038         int                i;
1039
1040         for (i = 0; i < arity; ++i) {
1041                 ir_node *op;
1042
1043                 /* check if one operand is the last use */
1044                 if (!rbitset_is_set(last_uses, i))
1045                         continue;
1046
1047                 op = get_irn_n(node, i);
1048                 free_reg_of_value(op);
1049                 ir_nodeset_remove(live_nodes, op);
1050         }
1051 }
1052
1053 /**
1054  * change inputs of a node to the current value (copies/perms)
1055  */
1056 static void rewire_inputs(ir_node *node)
1057 {
1058         int i;
1059         int arity = get_irn_arity(node);
1060
1061         for (i = 0; i < arity; ++i) {
1062                 ir_node           *op = get_irn_n(node, i);
1063                 allocation_info_t *info = try_get_allocation_info(op);
1064
1065                 if (info == NULL)
1066                         continue;
1067
1068                 info = get_allocation_info(info->original_value);
1069                 if (info->current_value != op) {
1070                         set_irn_n(node, i, info->current_value);
1071                 }
1072         }
1073 }
1074
1075 /**
1076  * Create a bitset of registers occupied with value living through an
1077  * instruction
1078  */
1079 static void determine_live_through_regs(unsigned *bitset, ir_node *node)
1080 {
1081         const allocation_info_t *info = get_allocation_info(node);
1082         unsigned r;
1083         int i;
1084         int arity;
1085
1086         /* mark all used registers as potentially live-through */
1087         for (r = 0; r < n_regs; ++r) {
1088                 if (assignments[r] == NULL)
1089                         continue;
1090                 if (!rbitset_is_set(normal_regs, r))
1091                         continue;
1092
1093                 rbitset_set(bitset, r);
1094         }
1095
1096         /* remove registers of value dying at the instruction */
1097         arity = get_irn_arity(node);
1098         for (i = 0; i < arity; ++i) {
1099                 ir_node               *op;
1100                 const arch_register_t *reg;
1101
1102                 if (!rbitset_is_set(info->last_uses, i))
1103                         continue;
1104
1105                 op  = get_irn_n(node, i);
1106                 reg = arch_get_irn_register(op);
1107                 rbitset_clear(bitset, arch_register_get_index(reg));
1108         }
1109 }
1110
1111 static void solve_lpp(ir_nodeset_t *live_nodes, ir_node *node,
1112                       unsigned *forbidden_regs, unsigned *live_through_regs)
1113 {
1114         unsigned *forbidden_edges = rbitset_malloc(n_regs * n_regs);
1115         int      *lpp_vars        = XMALLOCNZ(int, n_regs*n_regs);
1116         int       arity           = get_irn_arity(node);
1117         int       i;
1118         unsigned  l;
1119         unsigned  r;
1120
1121         lpp_t *lpp = lpp_new("prefalloc", lpp_minimize);
1122         //lpp_set_time_limit(lpp, 20);
1123         lpp_set_log(lpp, stdout);
1124
1125         /** mark some edges as forbidden */
1126         for (i = 0; i < arity; ++i) {
1127                 ir_node                   *op = get_irn_n(node, i);
1128                 const arch_register_t     *reg;
1129                 const arch_register_req_t *req;
1130                 const unsigned            *limited;
1131                 unsigned                   current_reg;
1132
1133                 if (!arch_irn_consider_in_reg_alloc(cls, op))
1134                         continue;
1135
1136                 req = arch_get_irn_register_req_in(node, i);
1137                 if (!(req->type & arch_register_req_type_limited))
1138                         continue;
1139
1140                 limited     = req->limited;
1141                 reg         = arch_get_irn_register(op);
1142                 current_reg = arch_register_get_index(reg);
1143                 for (r = 0; r < n_regs; ++r) {
1144                         if (rbitset_is_set(limited, r))
1145                                 continue;
1146
1147                         rbitset_set(forbidden_edges, current_reg*n_regs + r);
1148                 }
1149         }
1150
1151         /* add all combinations, except for not allowed ones */
1152         for (l = 0; l < n_regs; ++l) {
1153                 if (!rbitset_is_set(normal_regs, l)) {
1154                         char name[15];
1155                         snprintf(name, sizeof(name), "%u_to_%u", l, l);
1156                         lpp_vars[l*n_regs+l] = lpp_add_var(lpp, name, lpp_binary, 1);
1157                         continue;
1158                 }
1159
1160                 for (r = 0; r < n_regs; ++r) {
1161                         if (!rbitset_is_set(normal_regs, r))
1162                                 continue;
1163                         if (rbitset_is_set(forbidden_edges, l*n_regs + r))
1164                                 continue;
1165                         /* livethrough values may not use constrained output registers */
1166                         if (rbitset_is_set(live_through_regs, l)
1167                             && rbitset_is_set(forbidden_regs, r))
1168                                 continue;
1169
1170                         char name[15];
1171                         snprintf(name, sizeof(name), "%u_to_%u", l, r);
1172
1173                         double costs = l==r ? 9 : 8;
1174                         lpp_vars[l*n_regs+r]
1175                                 = lpp_add_var(lpp, name, lpp_binary, costs);
1176                         assert(lpp_vars[l*n_regs+r] > 0);
1177                 }
1178         }
1179         /* add constraints */
1180         for (l = 0; l < n_regs; ++l) {
1181                 int constraint;
1182                 /* only 1 destination per register */
1183                 constraint = -1;
1184                 for (r = 0; r < n_regs; ++r) {
1185                         int var = lpp_vars[l*n_regs+r];
1186                         if (var == 0)
1187                                 continue;
1188                         if (constraint < 0) {
1189                                 char name[64];
1190                                 snprintf(name, sizeof(name), "%u_to_dest", l);
1191                                 constraint = lpp_add_cst(lpp, name, lpp_equal, 1);
1192                         }
1193                         lpp_set_factor_fast(lpp, constraint, var, 1);
1194                 }
1195                 /* each destination used by at most 1 value */
1196                 constraint = -1;
1197                 for (r = 0; r < n_regs; ++r) {
1198                         int var = lpp_vars[r*n_regs+l];
1199                         if (var == 0)
1200                                 continue;
1201                         if (constraint < 0) {
1202                                 char name[64];
1203                                 snprintf(name, sizeof(name), "one_to_%u", l);
1204                                 constraint = lpp_add_cst(lpp, name, lpp_less_equal, 1);
1205                         }
1206                         lpp_set_factor_fast(lpp, constraint, var, 1);
1207                 }
1208         }
1209
1210         lpp_dump_plain(lpp, fopen("lppdump.txt", "w"));
1211
1212         /* solve lpp */
1213         {
1214                 ir_graph     *irg     = get_irn_irg(node);
1215                 be_options_t *options = be_get_irg_options(irg);
1216                 unsigned     *assignment;
1217                 lpp_solve(lpp, options->ilp_server, options->ilp_solver);
1218                 if (!lpp_is_sol_valid(lpp))
1219                         panic("ilp solution not valid!");
1220
1221                 assignment = ALLOCAN(unsigned, n_regs);
1222                 for (l = 0; l < n_regs; ++l) {
1223                         unsigned dest_reg = (unsigned)-1;
1224                         for (r = 0; r < n_regs; ++r) {
1225                                 int var = lpp_vars[l*n_regs+r];
1226                                 if (var == 0)
1227                                         continue;
1228                                 double val = lpp_get_var_sol(lpp, var);
1229                                 if (val == 1) {
1230                                         assert(dest_reg == (unsigned)-1);
1231                                         dest_reg = r;
1232                                 }
1233                         }
1234                         assert(dest_reg != (unsigned)-1);
1235                         assignment[dest_reg] = l;
1236                 }
1237
1238                 fprintf(stderr, "Assignment: ");
1239                 for (l = 0; l < n_regs; ++l) {
1240                         fprintf(stderr, "%u ", assignment[l]);
1241                 }
1242                 fprintf(stderr, "\n");
1243                 fflush(stdout);
1244                 permute_values(live_nodes, node, assignment);
1245         }
1246         lpp_free(lpp);
1247 }
1248
1249 /**
1250  * Enforce constraints at a node by live range splits.
1251  *
1252  * @param  live_nodes  the set of live nodes, might be changed
1253  * @param  node        the current node
1254  */
1255 static void enforce_constraints(ir_nodeset_t *live_nodes, ir_node *node,
1256                                 unsigned *forbidden_regs)
1257 {
1258         int arity = get_irn_arity(node);
1259         int i, res;
1260         hungarian_problem_t *bp;
1261         unsigned l, r;
1262         unsigned *assignment;
1263         ir_node  *value;
1264
1265         /* construct a list of register occupied by live-through values */
1266         unsigned *live_through_regs = NULL;
1267
1268         /* see if any use constraints are not met */
1269         bool double_width = false;
1270         bool good = true;
1271         for (i = 0; i < arity; ++i) {
1272                 ir_node                   *op = get_irn_n(node, i);
1273                 const arch_register_t     *reg;
1274                 const arch_register_req_t *req;
1275                 const unsigned            *limited;
1276                 unsigned                   reg_index;
1277
1278                 if (!arch_irn_consider_in_reg_alloc(cls, op))
1279                         continue;
1280
1281                 /* are there any limitations for the i'th operand? */
1282                 req = arch_get_irn_register_req_in(node, i);
1283                 if (req->width > 1)
1284                         double_width = true;
1285                 if (!(req->type & arch_register_req_type_limited))
1286                         continue;
1287
1288                 limited   = req->limited;
1289                 reg       = arch_get_irn_register(op);
1290                 reg_index = arch_register_get_index(reg);
1291                 if (!rbitset_is_set(limited, reg_index)) {
1292                         /* found an assignment outside the limited set */
1293                         good = false;
1294                         break;
1295                 }
1296         }
1297
1298         /* is any of the live-throughs using a constrained output register? */
1299         be_foreach_definition(node, cls, value,
1300                 if (req_->width > 1)
1301                         double_width = true;
1302                 if (! (req_->type & arch_register_req_type_limited))
1303                         continue;
1304                 if (live_through_regs == NULL) {
1305                         rbitset_alloca(live_through_regs, n_regs);
1306                         determine_live_through_regs(live_through_regs, node);
1307                 }
1308                 rbitset_or(forbidden_regs, req_->limited, n_regs);
1309                 if (rbitsets_have_common(req_->limited, live_through_regs, n_regs))
1310                         good = false;
1311         );
1312
1313         if (good)
1314                 return;
1315
1316         /* create these arrays if we haven't yet */
1317         if (live_through_regs == NULL) {
1318                 rbitset_alloca(live_through_regs, n_regs);
1319         }
1320
1321         if (double_width) {
1322                 /* only the ILP variant can solve this yet */
1323                 solve_lpp(live_nodes, node, forbidden_regs, live_through_regs);
1324                 return;
1325         }
1326
1327         /* at this point we have to construct a bipartite matching problem to see
1328          * which values should go to which registers
1329          * Note: We're building the matrix in "reverse" - source registers are
1330          *       right, destinations left because this will produce the solution
1331          *       in the format required for permute_values.
1332          */
1333         bp = hungarian_new(n_regs, n_regs, HUNGARIAN_MATCH_PERFECT);
1334
1335         /* add all combinations, then remove not allowed ones */
1336         for (l = 0; l < n_regs; ++l) {
1337                 if (!rbitset_is_set(normal_regs, l)) {
1338                         hungarian_add(bp, l, l, 1);
1339                         continue;
1340                 }
1341
1342                 for (r = 0; r < n_regs; ++r) {
1343                         if (!rbitset_is_set(normal_regs, r))
1344                                 continue;
1345                         /* livethrough values may not use constrainted output registers */
1346                         if (rbitset_is_set(live_through_regs, l)
1347                                         && rbitset_is_set(forbidden_regs, r))
1348                                 continue;
1349
1350                         hungarian_add(bp, r, l, l == r ? 9 : 8);
1351                 }
1352         }
1353
1354         for (i = 0; i < arity; ++i) {
1355                 ir_node                   *op = get_irn_n(node, i);
1356                 const arch_register_t     *reg;
1357                 const arch_register_req_t *req;
1358                 const unsigned            *limited;
1359                 unsigned                   current_reg;
1360
1361                 if (!arch_irn_consider_in_reg_alloc(cls, op))
1362                         continue;
1363
1364                 req = arch_get_irn_register_req_in(node, i);
1365                 if (!(req->type & arch_register_req_type_limited))
1366                         continue;
1367
1368                 limited     = req->limited;
1369                 reg         = arch_get_irn_register(op);
1370                 current_reg = arch_register_get_index(reg);
1371                 for (r = 0; r < n_regs; ++r) {
1372                         if (rbitset_is_set(limited, r))
1373                                 continue;
1374                         hungarian_remove(bp, r, current_reg);
1375                 }
1376         }
1377
1378         //hungarian_print_cost_matrix(bp, 1);
1379         hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1380
1381         assignment = ALLOCAN(unsigned, n_regs);
1382         res = hungarian_solve(bp, assignment, NULL, 0);
1383         assert(res == 0);
1384
1385 #if 0
1386         fprintf(stderr, "Swap result:");
1387         for (i = 0; i < (int) n_regs; ++i) {
1388                 fprintf(stderr, " %d", assignment[i]);
1389         }
1390         fprintf(stderr, "\n");
1391 #endif
1392
1393         hungarian_free(bp);
1394
1395         permute_values(live_nodes, node, assignment);
1396 }
1397
1398 /** test whether a node @p n is a copy of the value of node @p of */
1399 static bool is_copy_of(ir_node *value, ir_node *test_value)
1400 {
1401         allocation_info_t *test_info;
1402         allocation_info_t *info;
1403
1404         if (value == test_value)
1405                 return true;
1406
1407         info      = get_allocation_info(value);
1408         test_info = get_allocation_info(test_value);
1409         return test_info->original_value == info->original_value;
1410 }
1411
1412 /**
1413  * find a value in the end-assignment of a basic block
1414  * @returns the index into the assignment array if found
1415  *          -1 if not found
1416  */
1417 static int find_value_in_block_info(block_info_t *info, ir_node *value)
1418 {
1419         unsigned   r;
1420         ir_node  **end_assignments = info->assignments;
1421         for (r = 0; r < n_regs; ++r) {
1422                 ir_node *a_value = end_assignments[r];
1423
1424                 if (a_value == NULL)
1425                         continue;
1426                 if (is_copy_of(a_value, value))
1427                         return (int) r;
1428         }
1429
1430         return -1;
1431 }
1432
1433 /**
1434  * Create the necessary permutations at the end of a basic block to fullfill
1435  * the register assignment for phi-nodes in the next block
1436  */
1437 static void add_phi_permutations(ir_node *block, int p)
1438 {
1439         unsigned   r;
1440         unsigned  *permutation;
1441         ir_node  **old_assignments;
1442         bool       need_permutation;
1443         ir_node   *phi;
1444         ir_node   *pred = get_Block_cfgpred_block(block, p);
1445
1446         block_info_t *pred_info = get_block_info(pred);
1447
1448         /* predecessor not processed yet? nothing to do */
1449         if (!pred_info->processed)
1450                 return;
1451
1452         permutation = ALLOCAN(unsigned, n_regs);
1453         for (r = 0; r < n_regs; ++r) {
1454                 permutation[r] = r;
1455         }
1456
1457         /* check phi nodes */
1458         need_permutation = false;
1459         phi = sched_first(block);
1460         for ( ; is_Phi(phi); phi = sched_next(phi)) {
1461                 const arch_register_t *reg;
1462                 const arch_register_t *op_reg;
1463                 int                    regn;
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                 a  = find_value_in_block_info(pred_info, op);
1472                 assert(a >= 0);
1473
1474                 reg  = arch_get_irn_register(phi);
1475                 regn = arch_register_get_index(reg);
1476                 /* same register? nothing to do */
1477                 if (regn == a)
1478                         continue;
1479
1480                 op     = pred_info->assignments[a];
1481                 op_reg = arch_get_irn_register(op);
1482                 /* virtual or joker registers are ok too */
1483                 if ((op_reg->type & arch_register_type_joker)
1484                                 || (op_reg->type & arch_register_type_virtual))
1485                         continue;
1486
1487                 permutation[regn] = a;
1488                 need_permutation  = true;
1489         }
1490
1491         if (need_permutation) {
1492                 /* permute values at end of predecessor */
1493                 old_assignments = assignments;
1494                 assignments     = pred_info->assignments;
1495                 permute_values(NULL, be_get_end_of_block_insertion_point(pred),
1496                                permutation);
1497                 assignments = old_assignments;
1498         }
1499
1500         /* change phi nodes to use the copied values */
1501         phi = sched_first(block);
1502         for ( ; is_Phi(phi); phi = sched_next(phi)) {
1503                 int      a;
1504                 ir_node *op;
1505
1506                 if (!arch_irn_consider_in_reg_alloc(cls, phi))
1507                         continue;
1508
1509                 op = get_Phi_pred(phi, p);
1510
1511                 /* we have permuted all values into the correct registers so we can
1512                    simply query which value occupies the phis register in the
1513                    predecessor */
1514                 a  = arch_register_get_index(arch_get_irn_register(phi));
1515                 op = pred_info->assignments[a];
1516                 set_Phi_pred(phi, p, op);
1517         }
1518 }
1519
1520 /**
1521  * Set preferences for a phis register based on the registers used on the
1522  * phi inputs.
1523  */
1524 static void adapt_phi_prefs(ir_node *phi)
1525 {
1526         int i;
1527         int arity = get_irn_arity(phi);
1528         ir_node           *block = get_nodes_block(phi);
1529         allocation_info_t *info  = get_allocation_info(phi);
1530
1531         for (i = 0; i < arity; ++i) {
1532                 ir_node               *op  = get_irn_n(phi, i);
1533                 const arch_register_t *reg = arch_get_irn_register(op);
1534                 ir_node               *pred_block;
1535                 block_info_t          *pred_block_info;
1536                 float                  weight;
1537                 unsigned               r;
1538
1539                 if (reg == NULL)
1540                         continue;
1541                 /* we only give the bonus if the predecessor already has registers
1542                  * assigned, otherwise we only see a dummy value
1543                  * and any conclusions about its register are useless */
1544                 pred_block = get_Block_cfgpred_block(block, i);
1545                 pred_block_info = get_block_info(pred_block);
1546                 if (!pred_block_info->processed)
1547                         continue;
1548
1549                 /* give bonus for already assigned register */
1550                 weight = (float)get_block_execfreq(execfreqs, pred_block);
1551                 r      = arch_register_get_index(reg);
1552                 info->prefs[r] += weight * AFF_PHI;
1553         }
1554 }
1555
1556 /**
1557  * After a phi has been assigned a register propagate preference inputs
1558  * to the phi inputs.
1559  */
1560 static void propagate_phi_register(ir_node *phi, unsigned assigned_r)
1561 {
1562         int      i;
1563         ir_node *block = get_nodes_block(phi);
1564         int      arity = get_irn_arity(phi);
1565
1566         for (i = 0; i < arity; ++i) {
1567                 ir_node           *op         = get_Phi_pred(phi, i);
1568                 allocation_info_t *info       = get_allocation_info(op);
1569                 ir_node           *pred_block = get_Block_cfgpred_block(block, i);
1570                 unsigned           r;
1571                 float              weight
1572                         = (float)get_block_execfreq(execfreqs, pred_block) * AFF_PHI;
1573
1574                 if (info->prefs[assigned_r] >= weight)
1575                         continue;
1576
1577                 /* promote the prefered register */
1578                 for (r = 0; r < n_regs; ++r) {
1579                         if (info->prefs[r] > -weight) {
1580                                 info->prefs[r] = -weight;
1581                         }
1582                 }
1583                 info->prefs[assigned_r] = weight;
1584
1585                 if (is_Phi(op))
1586                         propagate_phi_register(op, assigned_r);
1587         }
1588 }
1589
1590 static void assign_phi_registers(ir_node *block)
1591 {
1592         int                  n_phis = 0;
1593         int                  n;
1594         int                  res;
1595         unsigned            *assignment;
1596         ir_node             *node;
1597         hungarian_problem_t *bp;
1598
1599         /* count phi nodes */
1600         sched_foreach(block, node) {
1601                 if (!is_Phi(node))
1602                         break;
1603                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1604                         continue;
1605                 ++n_phis;
1606         }
1607
1608         if (n_phis == 0)
1609                 return;
1610
1611         /* build a bipartite matching problem for all phi nodes */
1612         bp = hungarian_new(n_phis, n_regs, HUNGARIAN_MATCH_PERFECT);
1613         n  = 0;
1614         sched_foreach(block, node) {
1615                 unsigned r;
1616
1617                 allocation_info_t *info;
1618                 if (!is_Phi(node))
1619                         break;
1620                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1621                         continue;
1622
1623                 /* give boni for predecessor colorings */
1624                 adapt_phi_prefs(node);
1625                 /* add stuff to bipartite problem */
1626                 info = get_allocation_info(node);
1627                 DB((dbg, LEVEL_3, "Prefs for %+F: ", node));
1628                 for (r = 0; r < n_regs; ++r) {
1629                         float costs;
1630
1631                         if (!rbitset_is_set(normal_regs, r))
1632                                 continue;
1633
1634                         costs = info->prefs[r];
1635                         costs = costs < 0 ? -logf(-costs+1) : logf(costs+1);
1636                         costs *= 100;
1637                         costs += 10000;
1638                         hungarian_add(bp, n, r, (int)costs);
1639                         DB((dbg, LEVEL_3, " %s(%f)", arch_register_for_index(cls, r)->name,
1640                                                 info->prefs[r]));
1641                 }
1642                 DB((dbg, LEVEL_3, "\n"));
1643                 ++n;
1644         }
1645
1646         //hungarian_print_cost_matrix(bp, 7);
1647         hungarian_prepare_cost_matrix(bp, HUNGARIAN_MODE_MAXIMIZE_UTIL);
1648
1649         assignment = ALLOCAN(unsigned, n_regs);
1650         res        = hungarian_solve(bp, assignment, NULL, 0);
1651         assert(res == 0);
1652
1653         /* apply results */
1654         n = 0;
1655         sched_foreach(block, node) {
1656                 unsigned               r;
1657                 const arch_register_t *reg;
1658
1659                 if (!is_Phi(node))
1660                         break;
1661                 if (!arch_irn_consider_in_reg_alloc(cls, node))
1662                         continue;
1663
1664                 r = assignment[n++];
1665                 assert(rbitset_is_set(normal_regs, r));
1666                 reg = arch_register_for_index(cls, r);
1667                 DB((dbg, LEVEL_2, "Assign %+F -> %s\n", node, reg->name));
1668                 use_reg(node, reg);
1669
1670                 /* adapt preferences for phi inputs */
1671                 if (propagate_phi_registers)
1672                         propagate_phi_register(node, r);
1673         }
1674 }
1675
1676 /**
1677  * Walker: assign registers to all nodes of a block that
1678  * need registers from the currently considered register class.
1679  */
1680 static void allocate_coalesce_block(ir_node *block, void *data)
1681 {
1682         int            i;
1683         ir_nodeset_t   live_nodes;
1684         ir_node       *node;
1685         int            n_preds;
1686         block_info_t  *block_info;
1687         block_info_t **pred_block_infos;
1688         ir_node      **phi_ins;
1689         unsigned      *forbidden_regs; /**< collects registers which must
1690                                                 not be used for optimistic splits */
1691
1692         (void) data;
1693         DB((dbg, LEVEL_2, "* Block %+F\n", block));
1694
1695         /* clear assignments */
1696         block_info  = get_block_info(block);
1697         assignments = block_info->assignments;
1698
1699         ir_nodeset_init(&live_nodes);
1700
1701         /* gather regalloc infos of predecessor blocks */
1702         n_preds          = get_Block_n_cfgpreds(block);
1703         pred_block_infos = ALLOCAN(block_info_t*, n_preds);
1704         for (i = 0; i < n_preds; ++i) {
1705                 ir_node      *pred      = get_Block_cfgpred_block(block, i);
1706                 block_info_t *pred_info = get_block_info(pred);
1707                 pred_block_infos[i]     = pred_info;
1708         }
1709
1710         phi_ins = ALLOCAN(ir_node*, n_preds);
1711
1712         /* collect live-in nodes and preassigned values */
1713         be_lv_foreach(lv, block, be_lv_state_in, i) {
1714                 bool                       need_phi = false;
1715                 const arch_register_req_t *req;
1716                 const arch_register_t     *reg;
1717                 int                        p;
1718
1719                 node = be_lv_get_irn(lv, block, i);
1720                 req  = arch_get_irn_register_req(node);
1721                 if (req->cls != cls)
1722                         continue;
1723
1724                 if (req->type & arch_register_req_type_ignore) {
1725                         allocation_info_t *info = get_allocation_info(node);
1726                         info->current_value = node;
1727
1728                         reg = arch_get_irn_register(node);
1729                         assert(reg != NULL); /* ignore values must be preassigned */
1730                         use_reg(node, reg);
1731                         continue;
1732                 }
1733
1734                 /* check all predecessors for this value, if it is not everywhere the
1735                    same or unknown then we have to construct a phi
1736                    (we collect the potential phi inputs here) */
1737                 for (p = 0; p < n_preds; ++p) {
1738                         block_info_t *pred_info = pred_block_infos[p];
1739
1740                         if (!pred_info->processed) {
1741                                 /* use node for now, it will get fixed later */
1742                                 phi_ins[p] = node;
1743                                 need_phi   = true;
1744                         } else {
1745                                 int a = find_value_in_block_info(pred_info, node);
1746
1747                                 /* must live out of predecessor */
1748                                 assert(a >= 0);
1749                                 phi_ins[p] = pred_info->assignments[a];
1750                                 /* different value from last time? then we need a phi */
1751                                 if (p > 0 && phi_ins[p-1] != phi_ins[p]) {
1752                                         need_phi = true;
1753                                 }
1754                         }
1755                 }
1756
1757                 if (need_phi) {
1758                         ir_mode *mode = get_irn_mode(node);
1759                         ir_node *phi  = be_new_Phi(block, n_preds, phi_ins, mode, cls);
1760
1761                         DB((dbg, LEVEL_3, "Create Phi %+F (for %+F) -", phi, node));
1762 #ifdef DEBUG_libfirm
1763                         {
1764                                 int pi;
1765                                 for (pi = 0; pi < n_preds; ++pi) {
1766                                         DB((dbg, LEVEL_3, " %+F", phi_ins[pi]));
1767                                 }
1768                                 DB((dbg, LEVEL_3, "\n"));
1769                         }
1770 #endif
1771                         mark_as_copy_of(phi, node);
1772                         sched_add_after(block, phi);
1773
1774                         node = phi;
1775                 } else {
1776                         allocation_info_t *info = get_allocation_info(node);
1777                         info->current_value = phi_ins[0];
1778
1779                         /* Grab 1 of the inputs we constructed (might not be the same as
1780                          * "node" as we could see the same copy of the value in all
1781                          * predecessors */
1782                         node = phi_ins[0];
1783                 }
1784
1785                 /* if the node already has a register assigned use it */
1786                 reg = arch_get_irn_register(node);
1787                 if (reg != NULL) {
1788                         use_reg(node, reg);
1789                 }
1790
1791                 /* remember that this node is live at the beginning of the block */
1792                 ir_nodeset_insert(&live_nodes, node);
1793         }
1794
1795         rbitset_alloca(forbidden_regs, n_regs);
1796
1797         /* handle phis... */
1798         assign_phi_registers(block);
1799
1800         /* all live-ins must have a register */
1801 #ifdef DEBUG_libfirm
1802         {
1803                 ir_nodeset_iterator_t  iter;
1804                 foreach_ir_nodeset(&live_nodes, node, iter) {
1805                         const arch_register_t *reg = arch_get_irn_register(node);
1806                         assert(reg != NULL);
1807                 }
1808         }
1809 #endif
1810
1811         /* assign instructions in the block */
1812         sched_foreach(block, node) {
1813                 int arity;
1814                 ir_node *value;
1815
1816                 /* phis are already assigned */
1817                 if (is_Phi(node))
1818                         continue;
1819
1820                 rewire_inputs(node);
1821
1822                 /* enforce use constraints */
1823                 rbitset_clear_all(forbidden_regs, n_regs);
1824                 enforce_constraints(&live_nodes, node, forbidden_regs);
1825
1826                 rewire_inputs(node);
1827
1828                 /* we may not use registers used for inputs for optimistic splits */
1829                 arity = get_irn_arity(node);
1830                 for (i = 0; i < arity; ++i) {
1831                         ir_node *op = get_irn_n(node, i);
1832                         const arch_register_t *reg;
1833                         if (!arch_irn_consider_in_reg_alloc(cls, op))
1834                                 continue;
1835
1836                         reg = arch_get_irn_register(op);
1837                         rbitset_set(forbidden_regs, arch_register_get_index(reg));
1838                 }
1839
1840                 /* free registers of values last used at this instruction */
1841                 free_last_uses(&live_nodes, node);
1842
1843                 /* assign output registers */
1844                 be_foreach_definition_(node, cls, value,
1845                         assign_reg(block, value, forbidden_regs);
1846                 );
1847         }
1848
1849         ir_nodeset_destroy(&live_nodes);
1850         assignments = NULL;
1851
1852         block_info->processed = true;
1853
1854         /* permute values at end of predecessor blocks in case of phi-nodes */
1855         if (n_preds > 1) {
1856                 int p;
1857                 for (p = 0; p < n_preds; ++p) {
1858                         add_phi_permutations(block, p);
1859                 }
1860         }
1861
1862         /* if we have exactly 1 successor then we might be able to produce phi
1863            copies now */
1864         if (get_irn_n_edges_kind(block, EDGE_KIND_BLOCK) == 1) {
1865                 const ir_edge_t *edge
1866                         = get_irn_out_edge_first_kind(block, EDGE_KIND_BLOCK);
1867                 ir_node      *succ      = get_edge_src_irn(edge);
1868                 int           p         = get_edge_src_pos(edge);
1869                 block_info_t *succ_info = get_block_info(succ);
1870
1871                 if (succ_info->processed) {
1872                         add_phi_permutations(succ, p);
1873                 }
1874         }
1875 }
1876
1877 typedef struct block_costs_t block_costs_t;
1878 struct block_costs_t {
1879         float costs;   /**< costs of the block */
1880         int   dfs_num; /**< depth first search number (to detect backedges) */
1881 };
1882
1883 static int cmp_block_costs(const void *d1, const void *d2)
1884 {
1885         const ir_node       * const *block1 = (const ir_node**)d1;
1886         const ir_node       * const *block2 = (const ir_node**)d2;
1887         const block_costs_t *info1  = (const block_costs_t*)get_irn_link(*block1);
1888         const block_costs_t *info2  = (const block_costs_t*)get_irn_link(*block2);
1889         return QSORT_CMP(info2->costs, info1->costs);
1890 }
1891
1892 static void determine_block_order(void)
1893 {
1894         size_t    p;
1895         ir_node **blocklist = be_get_cfgpostorder(irg);
1896         size_t    n_blocks  = ARR_LEN(blocklist);
1897         int       dfs_num   = 0;
1898         pdeq     *worklist  = new_pdeq();
1899         ir_node **order     = XMALLOCN(ir_node*, n_blocks);
1900         size_t    order_p   = 0;
1901
1902         /* clear block links... */
1903         for (p = 0; p < n_blocks; ++p) {
1904                 ir_node *block = blocklist[p];
1905                 set_irn_link(block, NULL);
1906         }
1907
1908         /* walk blocks in reverse postorder, the costs for each block are the
1909          * sum of the costs of its predecessors (excluding the costs on backedges
1910          * which we can't determine) */
1911         for (p = n_blocks; p > 0;) {
1912                 block_costs_t *cost_info;
1913                 ir_node *block = blocklist[--p];
1914
1915                 float execfreq   = (float)get_block_execfreq(execfreqs, block);
1916                 float costs      = execfreq;
1917                 int   n_cfgpreds = get_Block_n_cfgpreds(block);
1918                 int   p2;
1919                 for (p2 = 0; p2 < n_cfgpreds; ++p2) {
1920                         ir_node       *pred_block = get_Block_cfgpred_block(block, p2);
1921                         block_costs_t *pred_costs = (block_costs_t*)get_irn_link(pred_block);
1922                         /* we don't have any info for backedges */
1923                         if (pred_costs == NULL)
1924                                 continue;
1925                         costs += pred_costs->costs;
1926                 }
1927
1928                 cost_info          = OALLOCZ(&obst, block_costs_t);
1929                 cost_info->costs   = costs;
1930                 cost_info->dfs_num = dfs_num++;
1931                 set_irn_link(block, cost_info);
1932         }
1933
1934         /* sort array by block costs */
1935         qsort(blocklist, n_blocks, sizeof(blocklist[0]), cmp_block_costs);
1936
1937         ir_reserve_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1938         inc_irg_block_visited(irg);
1939
1940         for (p = 0; p < n_blocks; ++p) {
1941                 ir_node       *block = blocklist[p];
1942                 if (Block_block_visited(block))
1943                         continue;
1944
1945                 /* continually add predecessors with highest costs to worklist
1946                  * (without using backedges) */
1947                 do {
1948                         block_costs_t *info       = (block_costs_t*)get_irn_link(block);
1949                         ir_node       *best_pred  = NULL;
1950                         float          best_costs = -1;
1951                         int            n_cfgpred  = get_Block_n_cfgpreds(block);
1952                         int            i;
1953
1954                         pdeq_putr(worklist, block);
1955                         mark_Block_block_visited(block);
1956                         for (i = 0; i < n_cfgpred; ++i) {
1957                                 ir_node       *pred_block = get_Block_cfgpred_block(block, i);
1958                                 block_costs_t *pred_info  = (block_costs_t*)get_irn_link(pred_block);
1959
1960                                 /* ignore backedges */
1961                                 if (pred_info->dfs_num > info->dfs_num)
1962                                         continue;
1963
1964                                 if (info->costs > best_costs) {
1965                                         best_costs = info->costs;
1966                                         best_pred  = pred_block;
1967                                 }
1968                         }
1969                         block = best_pred;
1970                 } while (block != NULL && !Block_block_visited(block));
1971
1972                 /* now put all nodes in the worklist in our final order */
1973                 while (!pdeq_empty(worklist)) {
1974                         ir_node *pblock = (ir_node*)pdeq_getr(worklist);
1975                         assert(order_p < n_blocks);
1976                         order[order_p++] = pblock;
1977                 }
1978         }
1979         assert(order_p == n_blocks);
1980         del_pdeq(worklist);
1981
1982         ir_free_resources(irg, IR_RESOURCE_BLOCK_VISITED);
1983
1984         DEL_ARR_F(blocklist);
1985
1986         obstack_free(&obst, NULL);
1987         obstack_init(&obst);
1988
1989         block_order   = order;
1990         n_block_order = n_blocks;
1991 }
1992
1993 /**
1994  * Run the register allocator for the current register class.
1995  */
1996 static void be_pref_alloc_cls(void)
1997 {
1998         size_t i;
1999
2000         lv = be_assure_liveness(irg);
2001         be_liveness_assure_sets(lv);
2002
2003         ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
2004
2005         DB((dbg, LEVEL_2, "=== Allocating registers of %s ===\n", cls->name));
2006
2007         be_clear_links(irg);
2008
2009         irg_block_walk_graph(irg, NULL, analyze_block, NULL);
2010         if (create_congruence_classes)
2011                 combine_congruence_classes();
2012
2013         for (i = 0; i < n_block_order; ++i) {
2014                 ir_node *block = block_order[i];
2015                 allocate_coalesce_block(block, NULL);
2016         }
2017
2018         ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
2019 }
2020
2021 static void dump(int mask, ir_graph *irg, const char *suffix)
2022 {
2023         if (be_get_irg_options(irg)->dump_flags & mask)
2024                 dump_ir_graph(irg, suffix);
2025 }
2026
2027 /**
2028  * Run the spiller on the current graph.
2029  */
2030 static void spill(void)
2031 {
2032         /* make sure all nodes show their real register pressure */
2033         be_timer_push(T_RA_CONSTR);
2034         be_pre_spill_prepare_constr(irg, cls);
2035         be_timer_pop(T_RA_CONSTR);
2036
2037         dump(DUMP_RA, irg, "spillprepare");
2038
2039         /* spill */
2040         be_timer_push(T_RA_SPILL);
2041         be_do_spill(irg, cls);
2042         be_timer_pop(T_RA_SPILL);
2043
2044         be_timer_push(T_RA_SPILL_APPLY);
2045         check_for_memory_operands(irg);
2046         be_timer_pop(T_RA_SPILL_APPLY);
2047
2048         dump(DUMP_RA, irg, "spill");
2049 }
2050
2051 /**
2052  * The pref register allocator for a whole procedure.
2053  */
2054 static void be_pref_alloc(ir_graph *new_irg)
2055 {
2056         const arch_env_t *arch_env = be_get_irg_arch_env(new_irg);
2057         int   n_cls                = arch_env->n_register_classes;
2058         int   c;
2059
2060         obstack_init(&obst);
2061
2062         irg       = new_irg;
2063         execfreqs = be_get_irg_exec_freq(irg);
2064
2065         /* determine a good coloring order */
2066         determine_block_order();
2067
2068         for (c = 0; c < n_cls; ++c) {
2069                 cls = &arch_env->register_classes[c];
2070                 if (arch_register_class_flags(cls) & arch_register_class_flag_manual_ra)
2071                         continue;
2072
2073                 stat_ev_ctx_push_str("regcls", cls->name);
2074
2075                 n_regs      = arch_register_class_n_regs(cls);
2076                 normal_regs = rbitset_malloc(n_regs);
2077                 be_set_allocatable_regs(irg, cls, normal_regs);
2078
2079                 spill();
2080
2081                 /* verify schedule and register pressure */
2082                 be_timer_push(T_VERIFY);
2083                 if (be_get_irg_options(irg)->verify_option == BE_VERIFY_WARN) {
2084                         be_verify_schedule(irg);
2085                         be_verify_register_pressure(irg, cls);
2086                 } else if (be_get_irg_options(irg)->verify_option == BE_VERIFY_ASSERT) {
2087                         assert(be_verify_schedule(irg) && "Schedule verification failed");
2088                         assert(be_verify_register_pressure(irg, cls)
2089                                 && "Register pressure verification failed");
2090                 }
2091                 be_timer_pop(T_VERIFY);
2092
2093                 be_timer_push(T_RA_COLOR);
2094                 be_pref_alloc_cls();
2095                 be_timer_pop(T_RA_COLOR);
2096
2097                 /* we most probably constructed new Phis so liveness info is invalid
2098                  * now */
2099                 /* TODO: test liveness_introduce */
2100                 be_liveness_invalidate(lv);
2101                 free(normal_regs);
2102
2103                 stat_ev_ctx_pop("regcls");
2104         }
2105
2106         be_timer_push(T_RA_SPILL_APPLY);
2107         be_abi_fix_stack_nodes(irg);
2108         be_timer_pop(T_RA_SPILL_APPLY);
2109
2110         be_timer_push(T_VERIFY);
2111         if (be_get_irg_options(irg)->verify_option == BE_VERIFY_WARN) {
2112                 be_verify_register_allocation(irg);
2113         } else if (be_get_irg_options(irg)->verify_option == BE_VERIFY_ASSERT) {
2114                 assert(be_verify_register_allocation(irg)
2115                        && "Register allocation invalid");
2116         }
2117         be_timer_pop(T_VERIFY);
2118
2119         obstack_free(&obst, NULL);
2120 }
2121
2122 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_pref_alloc)
2123 void be_init_pref_alloc(void)
2124 {
2125         static be_ra_t be_ra_pref = {
2126                 be_pref_alloc,
2127         };
2128         lc_opt_entry_t *be_grp              = lc_opt_get_grp(firm_opt_get_root(), "be");
2129         lc_opt_entry_t *prefalloc_group = lc_opt_get_grp(be_grp, "prefalloc");
2130         lc_opt_add_table(prefalloc_group, options);
2131
2132         be_register_allocator("pref", &be_ra_pref);
2133         FIRM_DBG_REGISTER(dbg, "firm.be.prefalloc");
2134 }